INTRODUCTION TO

MATHEMATICAL MODELING
 INTRODUCTION TO
MATHEMATICAL MODELING

MAYER HUMI
WORCESTER POLYTECHNIC INSTITUTE
USA
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742
© 2014 by Taylor & Francis Group, LLC
CRC Press is an imprint of Taylor & Francis Group, an Informa business
No claim to original U.S. Government works
Printed on acid-free paper
Version Date: 20161101
International Standard Book Number-13: 978-1-4987-2800-3 (Hardback)
This book contains information obtained from authentic and highly regarded sources. Reasonable efforts have been made to publish
reliable data and information, but the author and publisher cannot assume responsibility for the validity of all materials or the
consequences of their use. The authors and publishers have attempted to trace the copyright holders of all material reproduced in this
publication and apologize to copyright holders if permission to publish in this form has not been obtained. If any copyright material has not
been acknowledged please write and let us know so we may rectify in any future reprint.
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or utilized in any form by
any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and recording, or in
any information storage or retrieval system, without written permission from the publishers.
For permission to photocopy or use material electronically from this work, please access www.copyright.com
(http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-
750-8400. CCC is a not-for-profit organization that provides licenses and registration for a variety of users. For organizations that have
been granted a photocopy license by the CCC, a separate system of payment has been arranged.
Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identification and
explanation without intent to infringe.
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com
Contents

CHAPTER 1 ▪ The Process of Mathematical Modeling

1.1 What is Model Building?
1.2 Modeling Framework
1.3 Genes and Biological Reproduction
CHAPTER 2 ▪ Modeling with Ordinary Differential Equations
2.1 The Motion of a Projectile
2.1.1 Approximations and Simplifications
2.1.2 Model
2.1.3 Model Compounding
2.2 Springmass Systems
2.2.1 Data Collection
2.2.2 Approximations and Simplifications
2.2.3 Mathematical Model
2.2.4 Remarks and Refinements
2.3 Electrical Circuits
2.3.1 RLC Circuits
2.3.2 Approximations
2.4 Population Models
2.4.1 Logistic Model
2.4.2 Prototype Model
2.4.3 Data and Approximations
2.4.4 Solution of the logistic equation
2.5 Motion in a Central Force Field
2.5.1 Radial Coordinate System in R 2
2.5.2 Linear Pendulum
2.5.3 Nonlinear Pendulum
2.5.4 A Short Introduction to Elliptic Functions
2.5.5 Motion of a Pr ojectile on a Rotating Earth
2.5.6 A Particle in a Central Force Field
2.5.7 Motion of a Rocket
2.5.8 Multistage Rockets
2.5.9 Control of a Satellite in Orbit
2.6 Greenhouse Effect
2.7 Current Energy Balance of the Earth
2.7.1 Critique of the Model
 2.7.2 Humanity and Energy
CHAPTER 3 ▪ Solutions of Systems of ODEs
3.1 Review
3.1.1 Linear differential equations with constant coefficients
3.2 Review of Linear Algebra
3.2.1 Eigenvalues and Eigenvectors
3.3 Reformulation Of Systems Odes
3.4 Linear Systems With Constant Coefficients
3.5 Numerical Solution Of Initial Value Problems
3.5.1 Euler Algorithm
3.6 Finite Difference Approximations
3.6.1 Extension to Higher Dimensions
3.7 Modified Euler And Runge-Kutta Methods
3.7.1 Modified Euler Algorithm
3.7.2 Runge-Kutta Methods
3.8 Boundary Value Problems
CHAPTER 4 ▪ Stability Theory
4.1 General Introduction
4.2 Two species model
4.2.1 Steady States
4.2.2 Stability Analysis
4.3 Basic concepts
4.4 Linearizable Dynamical Systems
4.5 Linearizable systems in two dimensions
4.6 Liapounov method
4.7 Periodic SOLUTIONS(LIMIT cycles)
CHAPTER 5 ▪ Bifurcations and Chaos
5.1 Introduction
5.2 Bifurcations of Co-Dimension One
5.2.1 Trans-critical Bifurcation
5.2.2 Saddle Point Bifurcation
5.2.3 Pitchfork Bifurcation
5.2.4 Subcritical Bifurcation (Hysteresis)
5.2.5 Hopf Bifurcation
5.3 Rossler Oscillator
5.4 Lorenz Equations
5.5 Nerve Models
5.6 Miscellaneous Topics
5.6.1 Dimension
5.6.2 Liapunov Exponents
5.7 Appendix A: Derivation Of Lorenz Equations
CHAPTER 6 ▪ Perturbations
 6.1 Introduction
6.3 Regular Perturbations
6.4 Singular Perturbations
6.5 Boundary Layers
CHAPTER 7 ▪ Modeling with Partial Differential Equations
7.1 The Heat (or Diffusion) Equation
7.1.1 Burger’s Equation
7.1.2 Similarity Solutions
7.1.3 Stephan Problem(s)
7.2 Modeling Wave Phenomena
7.2.1 Nonlinear Wave Equations
7.2.2 Riemann Invariants
7.3 Shallow Water Waves
7.3.1 Tsunamis
7.4 Uniform Transmission Line
7.5 The Potential (Or Laplace) Equation
7.5.1 Kirchoff Transformation
7.6 The Continuity Equation
7.7 Electromagnetism
7.7.1 Maxwell Equations
7.7.2 Electrostatic Fields
7.7.3 Multipole Expansion
7.7.4 Magnetostatic
7.7.5 Electromagnetic Waves
7.7.6 Electromagnetic Energy and Momentum
7.7.7 Electromagnetic Potential
CHAPTER 8 ▪ Solutions of Partial Differential Equations
8.1 Method of Separation of Variables
8.1.1 Method of Separation of Variables By Example
8.1.2 Non Cartesian Coordinate Systems
8.1.3 Boundary Value Problems with General Initial Conditions
8.1.4 Boundary Value Problems with Inhomogeneous Equations
8.2 Green’S Functions
8.3 Laplace Transform
8.3.1 Basic Properties of the Laplace Transform
8.3.2 Applications to the Heat Equation
8.4 Numerical Solutions Of PDES
8.4.1 Finite Difference Schemes
8.4.2 Numerical Solutions for the Poisson Equation
8.4.2.1Other Boundary Conditions
8.4.3 Irregular Regions
8.4.4 Numerical Solutions for the Heat and Wave Equations
CHAPTER 9 ▪ Variational Principles
9.1 Extrema of Functions
9.2 Constraints and Lagrange Multipliers
9.3 Calculus of Variations
9.3.1 Natural Boundary Conditions
9.3.2 Variational Notation
9.4 Extensions
9.5 Applications
9.6 Variation with Constraints
9.7 Airplane Control, Minimum Flight Time
9.8 Applications In Elasticity
9.9 Rayleigh‐ritz Method
9.10 The Finite Element Method in 2-D
9.10.1Geometrical Triangulations
9.10.2Linear Interpolation in 2D
9.10.3Galerkin Formulation of FEM
9.11 Appendix
CHAPTER 10 ▪ Modeling Fluid Flow
10.1 Strain and Stress
10.2 Equations of Motion for Ideal Fluid
10.2.1Continuity equation
10.2.2Eulers’ equations
10.3 Navierstokes Equations
10.4 Similarity and Reynolds’ Number
10.5 Different Formulations of Navierstokes Equations
10.6 Convection and Boussinesq Approximation
10.7 Complex Variables in 2‐D Hydrodynamics
10.8 Blasius Boundary Layer Equation
10.9 Introduction to Turbulence Modeling
10.9.1Incompressible Turbulent Flow
10.9.2Modeling Eddy Viscosity
10.9.3k - ɛModel
10.9.4The Turbulent Energy Spectrum
10.10Stability of Fluid Flow
10.11Astrophysical Applications
10.11.1Derivation of the Model Equations
10.11.2Steady State Model Equations
10.11.3Physical Meaning of the Functions H(ρ),S(ρ)
10.11.4Radial Solutions for the Steady State Model
10.12Appendix A-Gauss Theorem And Its Variants
10.13Appendix B ‐Poincare Inequality And Burger’S Equation
10.14Appendix C ‐Gronwell Inequality
10.15Appendix D‐The Spectrum
CHAPTER 11 ▪ Modeling Geophysical Phenomena
11.1 Atmospheric Structure
11.2 Thermodynamics and Compressibility
11.2.1 Thermodynamic Modeling
11.2.2 Compressibility
11.3 General Circulation
11.4 Climate
CHAPTER 12 ▪ Stochastic Modeling
12.1 Introduction
12.2 Pure birth process
12.3 Kermackand mckendrick model
12.4 Queuing models
12.5 Markov chains
CHAPTER 13 ▪ Answers to Problems
 CHAPTER 1

The Process of Mathematical Modeling

CONTENTS
1.1 What is Model Building?
1.2 Modeling Framework
1.3 Genes and Biological Reproduction
1.1 What is Model Building?
Definition: modeling is the art of describing in symbolic language a real life system so that
approximately correct predictions can be made regarding the behavior or evolution of the system
under varied circumstances of interest.
We now elaborate on this definition.
First, note that in this definition “modeling” is referred to as an art. As such one cannot develop
rigid preset rules for this task. What can be done, however, is to point out a pattern that is found to be
useful in many cases and can help the practitioner to avoid many pitfalls.
Furthermore, a model is described as being able to make “correct predictions” about the system.
This usually does not mean 100% accuracy. Predictions of many models have a rather wide error
margin. The pertinent question, therefore, is whether these margins are acceptable to the user or not.
Moreover, it might turn out that several models are capable of describing the same phenomena with
different degrees of accuracy (and complexity).
Another important aspect of the definition is that a model should be “solvable.” A sophisticated but
“insolvable” model might be less useful from a practical point of view than a simple and
straightforward one which is capable of making predictions with acceptable error margins.
One should also bear in mind that every model is constructed with certain limitations on its
validity, and these should be borne in mind by the prospective user. Thus, in many practical
applications it is not that “the model is incorrect” but it is the application which Figure 1.1 violates
the basic assumptions of the model used.
Figure 1.1 A flow chart of the modeling process.

A classical example to illustrate these points is given by gravitation theory. Here we do have at
present two concurrent theories which pertain to modeling the same phenomena viz. Newton’s Law of
Gravitation and Einstein’s theory. Though it is accepted and proven that Einstein’s theory is better and
more accurate, still it is highly complex and “hard to solve.” As a result, in most terrestrial
applications we use Newton’s Law of Gravitation with “acceptable error margin.”
As to the problems which require model construction, their source and scope vary between applied
problems in life to attempts to duplicate natural phenomena and (what might seem to be) intellectual
curiosity.
Examples and Illustrations:
1. In many cases of daily life, we construct “mini-models” without even paying any attention to
these facts; e.g., “How do I get to downtown?” (by car, by bus, by subway, on foot, or
otherwise) requires a model which depends on:
The distance to downtown,
The time element (How fast do I want to get there?),
Money considerations,
Security considerations (Is it safe to ride the subway?),
Availability of means (How frequently do the buses run?),
The mood of the person.
2. Consider a truck company operating in the U.S. with “truck depots and service centers” in
some major cities.
 A major problem for such a company is how to dispatch trucks to their destinations in the
most economical way (saving gas and drivers’ time).
3. How can the wheat crop be increased to feed the growing human population?
4. What is the cause of global warming and climate change and how can these effects be
mitigated?
5. How can sound and light be recorded in a better way?
6. How can rockets be sent to the Moon or the planets?
7. Why is the sky blue?

1.2 MODELING FRAMEWORK

As we said above, model building is a creative act for which no preset rules apply. We do trace here,
however, a series of steps which hopefully will be useful in avoiding costly mistakes and around
which one can develop one’s skills in this field. We would like to stress, moreover, that model
building is a non-sequential process. In some cases, several of these steps will overlap, some might
be “missing” (i.e., not needed), and between others “loops” have to be made until one may come up
with a reasonable (and acceptable) model for the problem at hand.
We now describe these steps.
STEP 0: Set up as precisely as possible the reasons for constructing the model and its objectives.
We note here that in many projects a clear statement of these reasons and objectives might radically
influence the model to be built.
Example: The statement “Build a model to predict the weather” is a rather loose and incomplete
statement from a modeling point of view. Thus, if no statement is made for the reasons and the precise
objectives of the desired model, the problem should be considered as ill-defined. In fact, as stated
above, each of the following might be the actual objective of the model.
1. Predict the weather for the next hour.
Reason: One wants to go shopping.
Appropriate model: Just look through the window.
2. Predict the weather tomorrow.
Reason: Going on a trip one would like to know how to dress.
Appropriate model: Listen to the weather forecast on the radio or the Internet.
3. Predict the weather next winter.
Reason: Will it be a good idea to build a ski-motel during the summer?
Discussion of an appropriate model: We note, however, that even after these clarifications the
model to be built is not precisely defined since the word “weather” might have different connotations,
for example:
1. The model should predict the temperature to within 10° (should I take my sweater with
me?).
2. The model should predict the temperature, the condition of the sky, and the possibility of
showers (would it be a nice day for a trip? Should I take a raincoat?).
3. The model should predict the possibility of snow (should I take my ski gear with me?).
Another point to remember at this stage is that in many instances a reformulation of the model
objectives will be required during the process of model building. This is especially so if the original
scope of the model turns out to be too large.
Example: The objective of building a model to “cure cancer” requires many sub-models since
there are several types of cancer.
STEP 1: Study the problem as it is in real life.
Example: If one attempts to build a model to improve the performance of a production line, then it
is imperative to go to the factory and study first hand how this line operates (nothing else will do). In
many instances, one might discover that the problem of improved production depends on factors
which are independent of the line operation.
STEP 2: Data collection and analysis in real life.
At this stage, one studies the phenomenon and its behavior in real life with the objective of
identifying the major factors (i.e., causes) that influence the phenomenon.
Example 1: Analysis of car accidents.
As result of data collection and its statistical analysis, one might conclude that the main factors
which have a bearing on the frequency of car accidents are: driver, road, car, and weather. Each of
these can be further subdivided into several sub-headings; e.g., to define “driver,” we must specify
age, sex, height, sight, mental state (e.g., intoxication), etc.
We note that at this step some very crucial decisions have to be made viz. to identify those factors
that are most important to the problem at hand. For example, in analyzing car accidents, one might
make the (questionable) assumption that the height of the driver is of little importance and hence can
be ignored. Once again, it is important to keep in mind the need to strike the balance between model
simplicity (i.e., few variables and easy to solve) and effectiveness (i.e., accurate predictions).
Moreover, at this point one must also decide whether to limit the scope of the model to be built or
to make its objectives more precise.
Example 1: If we started with the objective of curing cancer, we might decide at this stage to study
only the relationship between drug X and lung cancer.
Example 2: If we wanted to study, originally, the performance of a given car, we might decide to
limit ourselves to the study of a certain component, e.g., the motor.
STEP 3. Controlled lab studies or simulations.
Studies carried out in the labs enable us to vary the factors that influence the phenomena under
study in a controlled manner and thus study the influence of each factor separately.
Example : If the strength of a certain material depends on the temperature and pressure, then lab
experiments will enable us to study the strength as a function of one variable only (for a constant
value of the other variables), something that is not easy to achieve in real life situations.
STEP 4: Construction of a conceptual qualitative model.
As a result of the studies conducted in the previous steps, one should be able to construct a
qualitative model for the phenomena at hand.
Example 1: “The accident rate depends mainly on the driver’s age where 25 seems to be the most
important in changing driving habits.”
Example 2: “Pleasure is the main motivating force in human behavior.”
Example 3: “Reliability of a given system decreases with heat and speed but increases with
weight.”
If such a qualitative model is acceptable to the user, then the author or researcher might feel no
further need for a mathematical-quantitative model (or the model might be hard to quantify, e.g.,
amount of pleasure, motivation, etc.) and therefore may proceed directly to step 8 (bypassing steps 5,
6, and 7, which are needed when a mathematical model is constructed).
In many cases, one is required at this stage to make a creative breakthrough, that is provide a new
conceptual framework for the problem under consideration and its solution.
Example 4: “The H2-molecule can be represented by a rod whose mass is concentrated at the end
points.” This is Raman Model for the H2-molecule which earned him the Nobel prize.
STEP 5: Conclusions, predictions, and recommendations that follow for the qualitative model.
Example 1: “Tomorrow will be partly cloudy.”
Example 2: “Saccharine is carcinogenic.”
Example 3: “Car X is not reliable.”
Example 4: “Mr. X has a strong personality.”
Example 5: “Travel is a pleasurable experience.”
The shortcomings of qualitative models are:
1. Models tend to be limited in scope.
2. Model reliability in making predictions is sometimes questionable.
3. Predictions are qualitative rather than quantitative.
The points of strength of these models, on the other hand, are:
1. Their meaning is clear to almost everybody.
2. They are hard to question and challenge (sometimes these models are accepted solely due to
the experience and authority of the person that suggests the model).
3. They are simple in structure.
4. They are based sometimes on long personal experience and intuition.
Example : “Buy stocks in January and sell them in April.”
While qualitative models are accepted and used extensively in the social sciences, they are rarely
acceptable in the natural sciences or engineering. In these fields, one has to perform the following
additional steps which lead to a mathematical-quantitative model and its solutions.
Remark: Not all mathematical models make quantitative predictions, e.g., in the stability theory
for solutions (or equilibrium points) of differential equations.
STEP 6: Abstraction and symbolic representation.
This step sometimes requires a lot of insight and creativity as the true variables which control the
phenomena might be masked by the data. In the majority of the cases, however, it consists of
representing the variables by symbols, naming the functions and identifying the axioms or constraints
of the model.
Example 1: “We can treat a car as a point particle whose position is given by the vector x
(abstraction and simplification).”
Example 2: “The temperature T is a polynomial function of the speed v and the time t, i.e.,
T=p(v,t)
where p is a polynomial.
Example 3: “Let the number of trucks at station A be denoted by NA. Then N must satisfy NA < N
A

where N is the total number of trucks operated by the company (symbolic representation and
constraint).”
Example 4: “For transportation purposes, it is enough to represent the U.S.
map by a set of discrete points whose location coincide with the major cities (abstraction and
simplification).”
 STEP 7: Derive the equations that govern the phenomena.
Example 1: If F is acting on point particle of mass m, and the particle acceleration is denoted by a,
then
F=ma.
This is Newton’s Second Law.
Example 2: Denoting by P, V, and T the pressure, volume, and temperature of a gas, then
PV=RT
where R is a constant. This is the Ideal Gas Law.
Remarks: Broadly, mathematical models are classified as deterministic versus stochastic (viz.
probabilistic). Another possible classification of these models is as continuous (i.e., the variables
used are continuous) and discrete. Each of these classifications has its merit within a given context.
STEP 8: Model testing.
To this end, one must solve the model equations and compare the solution with the actual data
collected in steps 2 and 3. If there is a bad fit,
i.e. non-acceptable deviations, then it will be necessary to redo steps 4, 5, and 6.
In this context, we remark that sometimes new mathematical techniques have been devised to solve
a mathematical model. If the model equations remain intractable, then some approximations to the
model equations must be made, thereby sacrificing accuracy in favor of easier computability.
Example: If the original model equations are highly nonlinear and hard to solve, then one may find
an acceptable linear approximation which might be solved easily.
STEP 9: Model limitations and constraints.
At this point, one must become clearly aware of the limitations that must be imposed on the use of
the model and the permissible range of the variables.
Example 1: One cannot use the equations of classical mechanics to predict the motion of a particle
whose speed is close to the speed of light.
Example 2: The ideal gas law is a good model for some gases but not for others.
STEP 10: Predictions and sensitivity analysis.
Once a model has been tested and found acceptable, then it can be used to make predictions.
Whenever such a prediction is found to be correct, the model is considered to be more reliable (in a
way every such prediction is a further test of the model).
One should bear in mind, however, that sensitivity analysis of many models is required before their
actual use; i.e., one has to find the extent to which the model predictions are sensitive to small
variations of the model parameters. We note that some models are “required” to be highly sensitive
while in others insensitivity to such variations is necessary (e.g., if the data contains inherent errors).
Example 1: Ballistic tables.
To construct an “exact” ballistic table, one has to know the exact atmospheric conditions, amount of
charge, geographic altitude, and state of the cannon to be fired. In field conditions, however, these
variables are known approximately at best. Hence, a good ballistic model must be somewhat
insensitive to small variations in these parameters while giving a reasonable prediction about the
range of the shot.
Example 2: Models for physical resonances.
Here, sensitivity is highly desirable especially when several such “close by” resonances are
involved.
Example 3: Chaotic systems.
 When the evolution of a system under consideration displays high sensitivity to the initial
conditions, we say that the system is “chaotic.” Under these circumstances, it is possible to make only
“short time” predictions about the state of the system. This is why weather forecasts are accurate only
for a “few days” (at best).
STEP 11: Extensions and refinements.
If a model is found to be correct in some instances but less accurate in others, then a refinement of
it is needed to take care of these exceptions.
Example: The ideal gas law needs such a refinement when the gas molecules are “large” (e.g.,
diatomic gases). The refined model is given by
(P+αV2)V=RT
where a is a parameter which depends on the gas.
STEP 12: Compounding.
Once a correct and reliable model has been established for some phenomena, then related
problems can be modeled by a process of compounding.
Example: Once the equations of motion for the spring-mass system are found, one can compound
the model to systems of several masses and springs.
Finally, we present here a schematic overview of the modeling process.

EXERCISES
E. Fermi was one of the great theoretical physicists in the twentieth century. Some of the following
“mini-model” questions are attributed to him. We offer these here to sharpen the reader’s skill in the
modeling of “real world problems.”
1. Ignoring oceans and such, how long would it take to walk entirely around the world?
2. How much water per year flows in the Mississippi river?
3. How many dump-truck loads would it take to move Mt. Washington in New Hampshire,
USA?
4. Find the dimension of a box that can contain all of the human race (five billion
approximately).
5. What is the linear velocity of Earth around the Sun?
6. How many drops of water are in the Pacific ocean?
7. How many books are in a bookstore?
8. How many atoms are in a cell?
9. How many cells are in the human body?
10. How many light-bulbs burn out in one minute throughout the world?
11. What is the actual volume of material in a solid cubic meter of metal (remember atoms are
made of nuclei and electrons)?
12. How do you buy the best car for your money?
13. How do you buy the best computer for your money?

1.3 Genes and Biological Reproduction

Most of the models that we consider in this book will use “continuous variables.” Moreover, it will
be advantageous in some cases to convert a discrete variable problem into a problem with continuous
variables. However, to illustrate the modeling process we consider in this section a model which
reduces a “continuous variable problem” into a discrete one.
Motivation and Objective:
A company grows and sells various types of beans (Lima-beans,kidney-beans etc.). It is a known
fact that green and smooth texture (Lima-)beans are preferred by the consumers and hence command a
higher price. However, on the farms the company grows beans which vary in color (white, red,
yellow, green, and anything in between) and texture (from smooth to wrinkled).
It is the objective of this project to understand this phenomenon so that the company will be able to
produce larger quantities of the desirable beans and enhance its profits.
Data from the farm:
1. Beans self fertilize.
2. Beans can be divided to a good approximation into the following groups: Green-Smooth (G-
S), Green-Wrinkled (G-W), White-Smooth (W-S), White-Wrinkled (W-W), Yellow-Smooth
(Y-S), Red-Smooth (R-S), and Red-Wrinkled (R-W).
Remarks: Note that by using the simplification in (2) above, we converted a continuous set of
variables for the color and texture into a discrete one. This, sometimes, simplifies the problem
considerably. However, sometimes the reverse is true (e.g., in population models).
Experimental Data:
Observations of bean plants grown in seclusion are being made in regard to their crop and their
descendants.

Results
1. It is possible to obtain pure lines of beans, i.e., beans which by self fertilization will always
produce descendants of the same type. However, these pure lines are prone to disease and
therefore not very desirable from a commercial point of view.
2. Sometimes, a plant from a line producing green smooth beans will produce by self
fertilization some descendants which are wrinkled, etc. (so that appearances might be
deceptive).
3. If we cross pure lines of G-S beans with G-W, we obtain first generation G-S beans only.
However, in the second generation, these G-S beans will give both G-S and G-W beans by
approximate ratio of 3:1.
Subproblem:
Build a model to explain texture only; i.e., assume all beans are Green.
Qualitative Model:
1. A bean carries entities which we shall call “genes” which determine whether it is smooth or
wrinkled. These will be denoted by S and W.
2. Each bean contains two such genes.
3. A bean is smooth if the combination of genes is SS or SW and wrinkled if WW.
4. In cross fertilization, one gene is accepted (independently) from each parent.
Remark: In such a situation where the combination SW is smooth, we shall say that S is
“dominant” with respect to W.
Mathematical Model
Let R(*, *) denote the reproduction function, i.e., the probability distribution of the descendants for
a given pair of parents
R(p1,p2)=(R1,R2,R3)
where pi and p2 represent the parents and Ri, R 2 , and R 3 are the probabilities of SS, SW, and WW
descendants respectively. If P(*) is the probability that a given bean carries a certain gene, we then
have
R1(p1,p2)=P(S|p1)P(S|p2)
R2(p1,p2)=P(S|p1)P(W|p2)+P(W|p1)P(S|p2)
R3(p1,p2)=P(W|p1)P(W|p2)
(In these equations P(S | Pi) represents the conditional probability that the parent Pi contributes the S
gene to the descendant and so on.)
Model Predictions: In the cross fertilization experiment, we started with two pure lines, i.e. pi =
SS, p2 = WW. As a result, our model predicts for first generation descendants:
R1=0R2=1R3=0,
e. all first generation beans are G-S which corresponds to the experimental results. For the
second generation, we therefore have
p1=p2=SW
and hence,
R1=1212=14,R2=12,R3=14,
i.e. 3 4 of the beans are smooth and 1 4 are wrinkled, i.e., a ratio of 3:1.
EXERCISES
1. Predict the results of cross fertilization between
SW and WW beans.
SW and SS beans.
2. (Compounding) Devise a model for beans which takes color into account.
Hint: Each bean will now have four genes; S, W for texture and C, c for color. Suppose that
beans with (SS,CC) are crossed with (ww,cc) (and C is also dominant with respect to c)
and the first generation descendants reproduce by self fertilization. Predict the results for
color and texture of the crop.
3. The following are well known facts regarding blood types in humans:
There are four (major) blood types denoted by A, B, AB and O.
Each blood cell contains two genes which determine the blood type.
The O-gene is regressive with respect to the A and B genes, i.e., AO and BO bloods
are A and B bloods respectively.
A and B genes are of “equal strength.”
In the process of reproduction, each parent donates one gene to determine the blood
type of the descendant.
Use this data to:
1. Give an explicit representation for the reproduction function of this system.
2. Predict the blood type distribution for the descendants to parents with blood types AO and
BO.
Bibliography
[1] R. Aris - Mathematical Modelling Techniques (Ferron-Pitman).
[2] B. Barnes and G.R. Fulford -Mathematical Modeling with Case Studies, 3rd Edition (CRC Press).
[3] Edward A. Bender - An Introduction to Mathematical Modeling (Dover).
[4] C. Dym - Principles of Mathematical Modeling, 2nd Edition (Academic Press).[5] Lin, C.C, and Segal, L.A.,1974, Mathematics
Applied to Deterministic Problems in the Natural Sciences, Macmillan, NY.
[6] Lindsay, R.B., and, Margenau, H.,1955, Foundation of Physics, Dover, NY.
[7] Maki, D.P., and Thompson, M., 2006,Mathematical Modeling and Computer Simulations Brooks/Cole, Belmont, CA, USA
[8] Meerschaert, M. M.,2012 - Mathematical Modeling, 4th Edition, Elsevier, Burlington, MA.
[9] Melnik, R., (Editor), 2015 Mathematical and Computational Modeling: With Applications in Natural and Social Sciences,
Engineering, and the Arts Wiley, Hoboken, NJ
[10] Noble, B.,1967, Applications of Undergraduate Mathematics in Engineering, MAA.
[11] Temam, R. and Miranville, R., 2000 Mathematical Modeling in Continuum Mechanics, Cambrige Univ. Press, Cambridge, UK.
 CHAPTER 2

Modeling with Ordinary Differential Equations

CONTENTS
2.1 The Motion of a Projectile
2.1.1 Approximations and Simplifications
2.1.2 Model
2.1.3 Model Compounding
2.2 Springmass Systems
2.2.1 Data Collection
2.2.2 Approximations and Simplifications
2.2.3 Mathematical Model
2.2.4 Remarks and Refinements
2.3 Electrical Circuits
2.3.1 RLC Circuits
2.3.2 Approximations
2.4 Population Models
2.4.1 Logistic Model
2.4.2 Prototype Model
2.4.3 Data and Approximations
2.4.4 Solution of the logistic equation
2.5 Motion in a Central Force Field
2.5.1 Radial Coordinate System in R 2
2.5.2 Linear Pendulum
2.5.3 Nonlinear Pendulum
2.5.4 A Short Introduction to Elliptic Functions
2.5.5 Motion of a Pr ojectile on a Rotating Earth
2.5.6 A Particle in a Central Force Field
2.5.7 Motion of a Rocket
2.5.8 Multistage Rockets
2.5.9 Control of a Satellite in Orbit
2.6 Greenhouse Effect
2.7 Current Energy Balance of the Earth
2.7.1 Critique of the Model
2.7.2 Humanity and Energy
The behavior and evolution of many scientific and engineering systems are described by equations
which involve unknown functions and their derivatives. These are called differential equations, and
methods for their solution play a central role in many disciplines.
Differential equations are classified as ordinary differential equations (ODEs) and partial
differential equations (PDEs). ODEs are equations which involve only one independent variable
while PDEs involve several independent variables.
To motivate the study of these equations we consider in this chapter problems in various areas
which are modeled naturally by ODEs. For some of these models a solution is possible by elementary
integration methods. For others more elaborate methods are needed.
For all the models presented in this chapter we illustrate the modeling process by adhering as
closely as possible to the modeling framework that was introduced in the previous chapter.

2.1 The Motion of a Projectile

Model Objective and Motivation: Build a prototype model which describes the motion of a small
particle in the gravity field of the Earth. Neglect all other forces and the rotation of the Earth.
This study is motivated by the fact that the motion of a projectile in the atmosphere is important in
many applications (e.g rockets, cannon shells, etc). As per usual in the modeling process we first
consider this problem in its ‘‘bare bones” setting and derive a prototype model.
Background: To derive the equations of motion for this problem we need Newton’s second law,
which states that the external force acting on a point mass is proportional to its acceleration. Thus
F=ma
where F,m, a denote respectively the force, mass, and acceleration of the particle.

2.1.1 Approximations and Simplifications

1. We assume that the speed of the particle is small compared to the speed of light. Hence
relativistic corrections to Newton’s second law can be neglected.
2. The projectile is considered to be a point particle and the motion around its center of
gravity is neglected.
3. We assume that the distance covered by the projectile is small compared to the Earth’s
radius. Consequently the spherical shape of the Earth can be neglected and we consider the
flight to be over a flat plane.
4. We neglect the variation of the gravitational force with height and location (which is due to
the fact that the Earth is not a perfect sphere). Hence we approximate g the acceleration due
to gravity by a constant.
5. We neglect the influence of the atmosphere on the motion of the projectile. These include air
drag and variations in temperature, density, and pressure.
6. We neglect the effect of the Earth’s rotation on the projectile motion.

2.1.2 Model
With the approximations delineated above it follows from Newton’s second law that the equation of
motion of the projectile is
(2.1)
md2Xdt2=-mgj
where j is a unit vector in the upward vertical direction. Since the only force acting on the projectile
is in the j-direction, we infer also that its motion is constrained to a plane. Without loss of generality
we can choose this plane to be the x-y plane with x = (x,y), (see Fig. 2.1). (Equation 2.1) is
equivalent then to two scalar equations
(2.2)
x¨=d2Xdt2=0,y¨=d2ydt2=-g

Figure 2.1 Trajectory of a projectile.

Since g is constant, we can readily integrate these equations twice to obtain

(2.3)
x˙=c1,y=-gt+c2
(2.4)
x=c1t+C3,y=-gt22+c2t+c4
where c i , i = 1, 2, 3, 4, are constants. To determine these integration constants we need some
((initial conditions” which (in this case) must specify the position and velocity of the projectile at
some (initial) time. Thus if we assume that at time t = 0 the projectile is at the origin and its velocity y
0 = ( v 0 cos θ , v 0 sin θ ) , then
(2.5)
x ( 0 ) = 0 , y ( 0 ) = 0 , x ˙ ( 0 ) = v 0 cos θ , y ˙ ( 0 ) = v 0 sin θ
To use these conditions we substitute t = 0 in (Equations 2.3), (2.4) to obtain
C3 = c 4 = 0
(2.6)
c 1 = v 0 cos θ , c 2 = v 0 sin θ
It follows then that the parametric representation of the trajectory is
(2.7)
x = ( v 0 cos θ ) t , y = - g t 2 2 + ( v 0 sin θ ) t
The nonparametric representation of the trajectory is obtained by eliminating t from (Equation 2.7).
This leads to
(2.8)
y = x tan θ - g 2 ( x v 0 cos θ ) 2
Example 2.1.1
Find the relation between the range of a projectile on Earth and the Moon if they satisfy the same
initial conditions.
Solution 2.1.1 The range of a projectile is the distance to where it returns to ground zero, i.e.,
y = 0. To find the range R e on Earth we set y = 0 in Equation (2.8) and solve for x.
We obtain
(2.9)
R e = v 0 2 sin 2 θ g e
where g e is the gravitational acceleration on Earth. On the Moon the projectile satisfies the same
equation of motion, but the gravitational acceleration is g m . Hence the range of the projectile on the
Moon is
(2.10)
R m = v 0 2 sin 2 θ g m
Therefore
(2.11)
RmRe=gegm

2.1.3 Model Compounding

We now compound the prototype model derived above by removing some of its constraints.
Example 2.1.2
Derive the equation of motion of the projectile when air resistance (drag) has to be taken into
consideration.
Solution 2.1.2 When the velocity of the projectile is not large, the drag force F d is (to a good
approximation) proportional to the velocity of the projectile
(2.12)
Fd=-αv
The necessary modifications to (Equation 2.1) are given by
(2.13)
mx¨=-mgj-αx˙
or in scalar form
(2.14)
mx¨=-αx˙
 (2.15)
my¨=-mg-αy˙
Eq. (2.14), (2.15) can be solved by direct integration
(2.16)
mx˙=-αx+c1
(2.17)
x=c2e-bt+c1mb,b=αm,α≠0
To solve for y we introduce y = u. (Equation 2.15) becomes
mu˙=-mg-αu
which then leads to
(2.18)
y˙=u=c3e-bt-gb
(2.19)
y=-1b(c3e-bt+gt)+c4
Once again we need initial conditions in order to solve for the integration constants c i , i = 1, . . ., 4.
We observe that at least formally the solution, (Equations 2.16) - (2.19), “looks” totally different from
the one obtained when α = 0 (however, see ex. 4).
Exercises
1. Derive the equations of motion for a projectile if the variation of the gravitational force with
height is to be taken into consideration.
2. Find the maximum height that a projectile will achieve as a function of the initial speed and
firing angle. If v 0 = 1 km/sec, what will be the maximal change in g along such a trajectory?
3. Solve for the constants c i , i = 1, . . ., 4, in (Equations 2.16) - (2.19) using the initial conditions
in (Equation 2.5).
4. Use the results of exercise 3 and a first order Taylor expansion for e -bt to show that as b → 0 the
solution, (Equations 2.16) - (2.19), converges to the one given by (Equation 2.7).
5. Write down the equation of motion and initial conditions for the motion of a projectile if there is
a wind blowing with velocity w = ( w 1 , w 2 ) where w 1, w 2 are constants (w 1, w 2 are the wind
components in the x, y directions). Solve your model.
6. How many firing angles can be used to achieve a given range for a projectile with initial
velocity v 0?
7. For a fixed initial speed at what firing angle will a projectile achieve its maximum range?
8. A plane with speed u is flying from city S to city N, which is at a distance d exactly north of S. A
wind of speed y is blowing in the eastern direction. Find differential equations for the position of the
plane if its pilot makes sure that the plane is always aimed towards N. (See Fig. 2.2).
Figure 2.2 A diagram for the plane position and velocity

Figure 2.3 Spring Mass system

Hint: Find differential equations for d x d t , d y d t in terms of the position ( x , y ) and u, v.

2.2 Springmass Systems

In this section we model spring‐mass systems as well as systems with torsion. Per usual we start with
a prototype problem and then compound it to model related systems.
Objective: Build a prototype model which describes the motion (in onedimension) of a mass
attached to a spring whose other end is rigidly fixed (see Fig. 3.3). Neglect gravity and all other
external forces.
Background: To model the motion of such a system one usually applies Hooke’s law. It states that
for small displacements x from the natural length L of the spring the force exerted by it is given by
F=-kx,|xL|≪ 1
where k > 0 is called the spring‐stiffness. However, to understand the limitations and approximations
made in its derivation, we state here some of the experiments needed to establish the law.

2.2.1 Data Collection

 1. Experiments to measure the force that various springs exert for positive and negative
displacements (stretchings and contractions respectively).
2. Experiments to determine to what extent the force exerted by a spring varies with its use.
3. Experiments to find the effects of environmental factors such as temperature, pressure, and
location on various springs.
4. Experiments to determine how the force that is exerted by the spring varies as a function of
the material, size of the coil, and number and diameter of the loops.
5. Experiments to find how various imperfections in the structure of a spring, e.g. variations in
the diameters of the loops and coil or deviations from circular symmetry, affect its
performance.

2.2.2 Approximations and Simplifications

As a result of the data collected in the experiments listed above, one isjustified in making the
following approximations and simplifications for metal springs:
1. Small deviations in the structure of a spring minimally affect its performance. Henceforth,
we only consider “ideal springs” which are made of homogeneous material, circular coil,
and loops whose diameters are constant.
2. Modeling with Ordinary Differential Equations blacksquare31
3. The force exerted by a spring depends very weakly on environmental factors and the number
of times that the spring is used. Hence we shall neglect the influence of these factors on the
performance of the spring. 3. For equal but opposite displacements the magnitude of the
force exerted by the spring is equal but in opposite directions.
4. For small displacements the force is proportional to the displacement with a negative
proportionality constant. The determination of this proportionality constant for a given
spring from first principles (i.e. as function of the material, number of loops, etc.) requires a
major modeling effort and is in most cases impractical.

2.2.3 Mathematical Model

Let
F = force exerted by the spring
m = mass of body attached to the spring
x = displacement of the mass (or the center of mass) from equilibrium
a = acceleration of the mass.
Using the approximations to the data introduced above we can now write for |x| < < 1 that
(2.20)
Fx=-kxHookesLaw
where k > 0 is called the stiffness of the spring.
Using Newton’s second law we find that the equation of motion for the mass attached to the spring
is given by
(2.21)
mx¨=-kx
or
(2.22)
md2Xdt2=-kx.

2.2.4 Remarks and Refinements

1. If we displace a spring from x to x + ▵ x, we infer from (Equation 2.20) that
(2.23)
ΔF=F( x+Δx) - F( x) =- kx.
This observation, that the additional force exerted by the spring due to a displacement ▵ x from x is
independent of x, is important in many applications.
1. (In preparation for 3) Let there be given an infinite series
(2.24)
p(x)=∑n=0∞anxn
which converges for |x| < R, R > 0. If p(x) is an odd function, i.e.
(2.25)
p(x)=-p(-x),|x|<R
then a 2m = 0, m = 0, 1, . . . . In fact we infer from (Equations 2.24), (2.25) that
2∑m=0∞a2mx2m=0
and since this must be true for all |x| < R, it follows that a 2m = 0. Similarly if p(x) = p(x),i.e. p(x) is
even, one infers that a 2m+1 = 0, m = 0, 1, . . .
1. In many instances engineers and scientists are called upon to solve or model systems in a
short period of time. Under these constraints it is impossible to conduct a thorough set of
experiments to establish the laws governing the system’s behavior. Instead “mathematical
approximations” must be used. We now illustrate this procedure.
Assume that the only information given about the force exerted by the spring is:
(a) = f(x), i.e. the force is a function of the displacement only.
(b) F(0) = 0
(c) (x) = - F( - x), i.e. F is an odd function of x, where F is some unknown but smooth (i.e.
analytic) function.
Since F is analytic, we can expand it in a Taylor expansion around x = 0
F(x)=F(0)+F′(0)1!x+F′′(0)2!x2+…=∑n=0∞F(n)(0)n!xn
However, from the fact that F is odd it follows (using the previous observation) that F (2m) (0) = 0,
m = 0, 1, 2, . . .. Hence
F(x)=k1x+k3x3+…
If |x| is small and k 1 is assumed to be nonzero, then we can approximate F(x) by
(2.26)
F(x)=-kx,k>0
(the sign can be determined by a simple experiment). (Equation 2.26) is called the linear (or first
order) approximation to F. It is valid when F ’(0) ≠ 0 and |x| is small. Hooke’s law can be interpreted
then as representing this approximation. Our analysis, however, goes one step beyond this law. In fact
it shows that the next order approximation to the force exerted by the spring (under present
assumptions) is not proportional to x 2 but to x 3, i.e.
(2.27)
F(x)=-kx±k3x3
Compounding: Inclusion of external forces.
If an external force besides that of the spring acts on the mass we have from Newton’s second law
(2.28)
ma=-kx+Fext
However if the external force F ext contains frictional forces F f then it is customary to separate this
force from the other external forces so that
(2.29)
ma=-kx+Fext+Ff
In three dimensions one can obtain the following data about the frictional force: F f always acts in the
direction opposite of the velocity y and is a function of y, the material and shape of the body, and the
medium in which the body moves.
For a given body moving on a uniform surface, F f = F f (v), and we infer from the “data” above
that
(2.30)
Ff(v)=-Ff(-v),v=|v|
Hence in one dimension the “first two term approximation” for F f is given by
(2.31)
Ff(v)=-bv-rv3
where b, r are positive constants.
In three dimensions the equivalent approximation for F f is
(2.32)
Ff(v)=-bv-r(v·v)v
For small | v | we therefore obtain the following equation of motion for the spring mass system in one‐
dimension.
(2.33)
ma+bv+kx=Fext
or
(2.34)
mx¨+bx˙+kx=Fext
where dots denote differentiation with respect to time (“standard” notation). The equivalent equation
of motion of this system in three dimensions is given by
(2.35)
mx¨+bx˙+kx=Fext
As a particular application of the nonlinear frictional forces given by eq. (2.31) we mention the
vibrations in the clarinet tube. Lord Rayleigh, who investigated this problem in the 19th century,
modeled these vibrations by the equation
(2.36)
mx¨+kx=bx˙-c(x˙)3,b,c>0
or equivalently
(2.37)
mx¨+[c(x˙)2-b]x˙+kx=0
This nonlinear equation is called Rayleigh equation.
Solution of (Equation 3.84) without friction.
When friction can be neglected and there are no external forces, (Equation 3.84) reduces to
(2.38)
mx¨+kx=0
In (Equation 2.38) the highest order derivative is ẍ , and therefore this is a second order differential
equation. This equation is also linear; i.e., if we consider the equation as a polynomial in x, x ˙ , ẍ ,
etc., then each term is of the first order. Also we observe that the coefficients of the equation are
constant. We now show how (Equation 2.38) can be solved by elementary techniques of integration.
Multiplying (Equation 2.38) by x ˙ and observing that x ˙ ẍ = 1 2 d d t ( x ˙ 2 ) . This yields
(2.39)
m2ddt(x˙2)+kx˙x=0
Integrating this with respect to t leads to
(2.40)
x˙+ω2x2=c2,ω=k/m
where c 2 is a constant of integration (observe that this constant must be nonnegative since the left
hand side of (2.40) is a sum of squares). Hence
(2.41)
x˙=c2-ω2x2
(Equation 2.41) can be easily integrated, and we obtain the solution in the form
(2.42)
x = A cos ( ω t + ϕ )
where A, φ are constant. Thus the general solution of (Equation 2.38) contains two arbitrary
constants. These can be determined if the initial conditions x(0), x ˙ ( 0 ) are known.
As expected, the solution, (Equation 2.42), represents vibrations with fixed amplitude as there is
no friction to damp the motion.
Related Systems:
Example 2.2.1.
Derive the equations of motion for two masses m 1, m 2 which are attached to a spring with
stiffness k as in Fig. 2.4.
Figure 2.4 Two masses attached to a spring

Solution 2.2.1 Let the distance between the center of mass of m 1, m 2 at equilibrium be L (if we
idealize the system and treat m 1, m 2 as point particles, then L is the natural length of the spring).
If these centers of mass at time t are at x 1, x 2 respectively, then either a. x 2 - x 1 - L > 0 or b. x
2 - x 1 - L < 0.
In the first case (a) the spring is stretched beyond its natural length, and hence m 1 is pulled to the
right and m 2 to the left (by Newton’s third law these two forces are equal but in opposite directions.)
Hence, using (Equation 2.23), we have
m1d2x1dt2=k(x2-x1-L)
(2.43)
m2d2x2dt2=-k(x2-x1-L)
Similarly in case (b) m 1 is pushed to the left and m 2 to the right. Since x 2 - x 1 - L < 0, we infer once
again that the equations of motion are given by (2.43). Thus the differential equations which govern
the system are the same in both cases.
We observe that this system is modeled by a system of coupled ordinary differential equations.
Example 2.2.2
Derive the equation of motion for a mass in between two springs which are attached to rigid walls
whose distance from each other is L, as shown in .
Solution 2.2.2 In problems of this type it is natural to use a coordinate system whose origin
coincides with the equilibrium position of the mass (which does not have to be calculated) and
obtain a diffe rential equation for the displacement from this position as a function of time. In fact
for such a displacement x the change in the forces acting on m is given by (using Equation (2.23))
Fig. 2.5.
Figure 2.5 A mass and two springs enclosed by rigid walls

Figure 2.6 A mass attached to a thin elastic bar

F=-k1x-k2x
(regardless of the sign of x). Hence the desired equation of motion is
(2.44)
mx¨=-(k1+k2)x
Remark 2.2.1
To evaluate the position x eq of m at equilibrium we use the fact that in this state F ext = 0. Hence if
ℓ1, ℓ2 are the natural lengths of the springs and m is treated as a point particle we have
k1(xeq-ℓ1)=k2(L-xeq-ℓ2)
(where we used a coordinate system whose origin is at the left wall of the system). However, note
again that Equation x eq is not needed for the derivation of Equation (2.44).
Example 2.2.3
Derive a model equation for the motion of a mass which is attached to a thin elastic bar and
subject to torsional forces (“twists”).
Solution 2.2.3 To model this problem one must conduct the same type of experiments and make
the same approximations as in the spring‐mass system. For small “twists, “ i.e., when the twist
angle θ is small, the elastic restoring torque due to the bar can be approximated by
(2.45)
T=-kθ
Using Newton’s second law for rotating bodies we then have
(2.46)
Iθ¨+kθ=Text
where T ext is the external torque and I is the moment of inertia of m around the axis of rotation which
is defined as
(2.47)
I=∫Vr2ρ(x)dx
Here r is the distance of x from the axis of rotation and ρ(x) is the density of the mass attached to the
bar.
When frictional forces are also present then for | θ ˙ | ≪ 1 , we have
(2.48)
Ff=-bθ˙
and the equation of motion for m becomes
(2.49)
Iθ¨+bθ˙+kθ=Text
Exercises
1. Find the differential equation which governs the motion of the system shown in Fig. 2.7:

Figure 2.7 A mass attached to two springs “in series”

Hint: Apply Newton’s second law to the massless point P at which the two springs are connected.
2. Repeat Ex. 1 for the system shown in Fig. 2.8.
Figure 2.8 A mass connected to two springs in “parallel”

Assume that m is always “gliding” on the x-axis.

3. What is the equivalent stiffness for the two springs in the systems of ex. 1,2; i.e., if one wants to
replace the two springs by one, what should its stiffness be to yield exactly the same equation of
motion for a mass m attached to it? Compare these results to the addition of resistors in series and
parallel in an electric circuit.
4. Find the moment of inertia for a thin homogeneous rod of length L and linear density ρ (i.e.
mass/unit length) which is rotating around its
(a) mid-point
(b) end point.
5. Model the following systems:
(a) Two masses attached to three elastic rods (Fig. 2.9).

Figure 2.9 Two masses attached to three elastic rods

(b) Two masses suspended vertically on springs in the gravitational field of the Earth (Fig. 2.10)
Figure 2.10 Two masses suspended vertically on springs

Hint: Derive the equation of motion in terms of the displacements x 1, x 2 of the two masses from
equilibrium as in example 2.
6. Generalize example 2.2.2 to a system of N masses with N + 1 springs as shown in Fig. 2.11.

Figure 2.11 Nmasses attached N + 1 springs enclosed by rigid walls

Figure 2.12 Components of a RLC circuit

Hint: Derive equations for the displacements x i , i = 1, . . ., N from equilibrium.

7. Simulate numerically the system in the previous exercise for n = 2. Assume k i = 1, i = 1, 2, 3,
with different mass ratios. For each mass plot x and x ˙ as a function of time. Plot also x ˙ versus x
versus x ˙ (phase-diagram). Experiment with different values of k i to evaluate the impact on the
solution.
8. Determine A, φ in (Equation 2.42) if
(a) x(0) = 1, x ˙ ( 0 ) = 0
(b) x(0) = 0, x ˙ ( 0 ) = 1 .
9. Show that the solution of (Equation 2.38) can be written as
x = A cos ω t + B sin ω t , ω 2 = k m .
10. Explain why (Equation 3.84) is linear. What is the order of this equation?
11. Show that when F ext = 0 and b ≠ 0, (Equation 3.84) represents a dissipative system. That is, the
total energy of the system
E(t)=12mv2+12kx2
is a monotonically decreasing function of time.
Solution: Multiply (Equation 3.84) by x ˙ and observe that
x¨x˙=12ddt(x˙),xx˙=12ddt(x2)
Hence
mx¨x˙+b(x)2+kxx˙=0
implies
12mddt(x˙2)+12kddt(x2)=-bx2
ddt(12mx˙2+12kx2)=-bx˙2
i.e.
dE(t)dt=-b(x˙)2≤0
i.e. E(t) is a monotonically decreasing function.
Observe that if b < 0 (“negative friction”) then E(t) is monotonically increasing.
12. Solve (Equation 3.84) with F ext = 0 and b 2 - 4mk > 0.
Hint: In some instances one can obtain solutions of differential equations by “inspection” or by
making appropriate assumptions on the form of the solution. In this case, assume that x(t) = Ae αt
where A, α are constants.
13. Use the results of ex. 12 to show that under present assumptions
(a) The equation has two independent solutions φ 1, φ 2
(b) Show that φ = c 1 φ 1 + c 2 φ 2, where c 1, c 2 are arbitrary constants, is also a solution. This is
called the superposition principle. It is true for all linear equations?
(c) What happens to x as t → ∞?
14. Solve eq. (3.84) with F ext = 0 and b 2 - 4mk < 0.
15. Consider a spring mass system with a nonlinear spring force
F=-kx-αx3
and without friction. Show that for this system
E=12mx˙2+kx22+αx44
is constant. Explain why a spring with α > 0 is referred to as “hard” while one with α < 0 is called
“soft.”
16. Show that for Rayleigh (Equation 2.37) the rate of change in the energy of the system is
negative if x ˙ 2 ≥ b / c .
Hint: Multiply this equation by x ˙

2.3 Electrical Circuits

In this section we discuss the modeling of RLC circuits. These circuits contain resistors, capacitors,
and inductances.

2.3.1 RLC Circuits

Objective: Build a model which will predict the electric current at any point of an electric circuit.
Background:
The following figure depicts the basic element of an RLC circuit and their universal representation.
Each of these components in a circuit diagram is considered pure; i.e. a resistor has a zero capacity
and inductance, etc.
The basic physical quantities of interest in an electric circuit are
1. The electric current ( x , t ) .
2. The electric potential (or voltage) V ( x , t ) . This potential is measured with respect to a fixed
reference point called “ground”‐since it is usually the Earth’s potential that is used for this purpose.
3. The electric charge Q.
 The units (in the MKS system) of the quantities introduced above are as follows:
R ‐ohms (denoted by Ω)
L - Henries
C - Farads
Q - Coulombs
i - Amperes
V - Volts
Remark
Sometimes one uses a quantity called the conductance G instead of the resistance R where G = 1/R.
The unit of conductance is called mho (ohm spelled backward).
The relationship between the components and the physical quantities in an electric circuit is as
follows:
1. The voltage drop across a resistance is related to the current passing through it by
(2.50)
V=Ri
2. The potential drop through an inductance L is given by
(2.51)
V=Ldidt
3. The total electric charge Q on a capacitor is given by
(2.52)
Q=CV
hence the “ virtual current” i in the capacitor is given by
(2.53)
i=dQdt=CdVdt
Kirchhoff’s Laws
Kirchhoff’s laws form the basis for the analysis of all electric circuits. To introduce these laws we
make the following definitions:
1. A node in an electric circuit is a juncture where current can flow along three or more paths (i.e.
a point where three or more electric wires are joined together) (see Fig. 2.13).
Figure 2.13 Kirchhoff’s first law

2. A loop in an electric circuit is a sequence of circuit elements which start and end at the same
point (i.e. a closed path).
First Kirchhoff law:
The algebraic sum of all the currents at a node is 0. For the node in Fig. 2.13 we have
i1-i2+i3=0.
By convention currents coming to the node are considered positive while those leaving it are
negative.
Second Kirchhoff law:
The algebraic sum of the voltage drops around a loop in an electric circuit is equal to the algebraic
sum of the external voltage sources in the loop.

2.3.2 Approximations
1. We assume that the resistance, capacitance, or inductance of a given electrical component is
independent of the environmental factors (such as temperature, humidity, etc.) and the previous history
of the circuit.
2. Cables connecting circuit components have zero resistance, capacity, and inductance.
3. The passage of an electric current through a cable always involves a leakage which leads to a
loss of electric energy. For short distances one can usually ignore this loss. However, over long
distances (i.e. transmission lines) one must take these losses into account.
With this data and approximations one can in principle analyze any given circuit. We present a few
examples.
Example 2.3.1
RLC Circuit.
A simple RLC circuit is illustrated in Fig. 2.14.
Figure 2.14 RLC circuit

Figure 2.15 Diagram of Wheatstone Bridgel

To solve it we first note that there are no nodes in this circuit, and therefore only Kirchhoff’s
second law applies, hence
(2.54)
e(t)=Ri+QC+Ldidt
Differentiating with respect to t and using (2.53) we obtain
(2.55)
dedt=Rdidt+1Ci+Ld2idt2
If d e d t is known, then (Equation 2.55) constitutes a second order (inhomogeneous) differential
equation with constant coefficients for the current i in the circuit.
Example 2.3.2
Wheatstone Bridge.
 The circuit shown in Fig. 2.14 is used to measure the resistance R x of a resistor by the use of two
fixed resistances R 1, R 2 and a third, variable one, R 3.
In this circuit e is a low voltage battery and G a galvanometer, i.e. an instrument to measure
currents. Once R x is inserted and the circuit is closed, the operator manipulates R 3 until the current in
G is zero (the circuit is then said to be balanced). Apply Kirchhoff’s second law to the loops ACD
and CBD. In the balanced state we obtain
(2.56)
R1i1-R2i2=0,R3i1-Rxi2=0
hence
(2.57)
R2Rx=R1R3,i.e.Rx=R2R3R1
Example 2.3.3
Multicomponent RLC Circuits
Definition 2.3.1
We say that a resistance R is equivalent to R 1 and R 2 in a given circuit if the replacement of these
resistors by R does not affect the current in the circuit.
Example 2.3.4
From Kirchhoff’s second law it is easy to see that:
1. Two resistors R 1, R 2 in series are equivalent to one resistance R = R 1 + R 2
2. Two resistors R 1, R 2 in parallel (see Fig. 2.16) are equivalent to one resistor with
1R=1R1+1R2.

Figure 2.16 Resistors in series and parallel

Resistors in parallel
Example 2.3.5
Find the equivalent resistance of the “infinite” circuit in Fig. 2.17.

Figure 2.17 Circuit with infinite resistors in parallel

 Solution 2.3.1 Since “∞ = ∞ - 1“ the equivalent resistance R 0 of AB∞ is equal to that of CD∞.
Hence the circuit AB∞ is equivalent to the one shown in Fig. 2.18 whose total equivalent
resistance must also equal R 0.

Figure 2.18 Circuits equivalent to the circuit in Fig 17

Thus we infer
R 0 = 2 R + 1 1 R + 1 R 0 i .e . R 0 = [ 1 + 3 ] R .
Example 2.3.6
DC motor.
A DC motor is an electro‐mechanical system consisting of two circuits the field circuit F, the
armament circuit A -and a shaft (Fig. 2.19.).

Figure 2.19 A diagram of a DC motor

Background:
The following information about the electro‐mechanical coupling in this circuit is needed to model
this system.
1. The motion of the shaft causes a potential drop V s in the armament circuit which is proportional
to its angular velocity ω and the field current i F
(2.58)
Vs=c1ωiF
The proportionality constant c 1 in this equation is called the electromechanical constant of the motor.
2. The torque T exerted on the shaft is proportional to i A and i F
(2.59)
T=c2iAiF
Model: A mathematical model for this system can be written down now by applying Kirchhoff’s
second law to the two circuits and Newton’s second law to the shaft whose moment of inertia we
denote by J.
(2.60)
LFdiFdt+RFiF=eF
 (2.61)
LAdiAdt+RAiA+c1ωiF=eA
(2.62)
Jdωdt=c2iAiF-c3ω
where C3 represents frictional damping which is proportional to the angular velocity of the shaft.
This is a system of three coupled nonlinear ordinary differential equations of the first order.
Example 2.3.7
Circuits with nonlinear resistors
Some electrical circuits contain “vacuum tubes” (or their solid state equivalents). For these
elements the resistance is a function of the current, e.g.,
(2.63)
R=μ(i2-1)
If we substitute this expression for the resistance in Equation. (2.55), we obtain
(2.64)
Ld2idt2+μ(i2-1)didt+1ci=dedt
Without the forcing term this is equivalent to
(2.65)
x¨+μ(x2-1)x˙+kx=0
(Equation 8.34) is called Van der Pol equation. Although this equation was originally derived to
model electrical circuits in vacuum tubes, it has been used since then to provide a basic model for the
function of nerve cells.
Exercises
1. Prove the statements made in example 2.3.4.
2. Derive model equations for the circuit in Fig. 2.20.

Figure 2.20 Generalized RLC circuit

3. What happens if the direct current source in the previous exercise is replaced by an alternating
current source ( t ) = 2 sin ( 2 t ) ?
4. Derive model equations for the circuit in the following figure (Fig 2.21).
Figure 2.21 Generalized RLC circuit

2.4 Population Models

In the first part of this section we introduce the logistic model for the population of one species and
the Lotka‐Volterra model for two interacting species. Applications to epidemics, chemical reactions,
and radioactive decay are discussed in the second part of the section.

2.4.1 Logistic Model

Objective: A company which harvests a system of artificial fish pools wants to increase the daily
catch but is afraid that this might adversely affect the future population of fish in the pools. As a result
a team of biologists and mathematicians is charged with the study and modeling of the fish population.

2.4.2 Prototype Model

Since the pools might contain many species which compete with each other for food and cannibalize
each other, it is decided to first study the population of one species of fish in an artificial pool.

2.4.3 Data and Approximations

1. For the first few generations of fish in the artificial pool the rate of increase of the fish population
(i.e. rate of birth-rate of death) was proportional to the population size at that time.
2. As the number of fish in the pool increases, the rate of increase decreases. An investigation of
the reasons for this change shows that the prime reason for this change is the quantity of food
available to the fish population.
3. Since the population of the fish, N(t), is large, we can assume that N is a continuous variable
rather than discrete.
Model:
 To begin with we assume that the rate of birth and death in the fish population is proportional to its
size and the time span over which we observe this population
(2.66)
N(t+△ t)-N(t)=αN(t)△ t-βN(t)△ t=AN(t)△ t,
where α and β are the birth and death rates respectively. Dividing by ▵ t and letting ▵ t → 0 we obtain
(2.67)
dNdt=AN(t)
Hence
(2.68)
N(t)=N(0)eAt
where N(0) is the number of fish at time 0.
It follows then that if A > 0 (rate of birth is greater than the rate of death), the population will
increase exponentially while if A < 0 it will decrease exponentially.
To model the competition for food (or resources in general) assume that a “piece of food” is
available at some location. The competition for food is represented then by the fact that several fish
come together to vie for it. Let us assume, however, that the event where three or more fish coming
together at the same time to “grab” the same piece of food is a “rare event” (the population is not
“very dense”). The competition will be represented then by a pair of fish coming together to vie for
the same piece of food. Hence we can assume that this competition is proportional to the number of
pairs in the population which is equal to N ( N - 1 ) 2 . (Equation 12.30) will be modified then as
(2.69)
N(t+△ t)-N(t)=AN(t)△ t-γN(N-1)2△ t
where γ > 0 is a constant. The minus sign in front of γ represents the adverse
effect that food competition has on the population. Combining same terms of N, renaming the
constants and letting ▵ t → 0 we finally obtain the logistic model equation for the population of one
species,
(2.70)
dNdt=aN-bN2
A more formal approach to the derivation of (Equation 12.34) is to observe that (12.30) fails to take
into account the decrease in N ˙ as N(t) increases. We can attempt to modify this equation by letting a
be a function of N, i.e.,
a = a(N).
A second order Taylor expansion of a(N) then yields
(2.71)
a(N)=a-bN+O(N2),b>0
where the minus sign in front of b in (Equation 2.71) is necessary to account for the fact that N ˙ is
decreasing as N increases.
Substituting (Equation 2.71) in (Equation 12.30) we obtain
(2.72)
dNdt=aN-bN2+O(N3)
In this equation the term - bN 2(b > 0) is interpreted as representing the competition between fish of
the same species for food.
2.4.4 Solution of the logistic equation
Although (Equation 2.72) is nonlinear, it can be solved analytically using partial fractions as follows:
dNN(a-bN)=dt
Hence
[1N+ba-bN]dN=adt
ln | N | - ln | a - b N | = a t + k
which after some algebra yields (note that N < 0 is meaningless)
(2.73)
N=aN0bN0+(a-bN0)e-at
where N 0 is the initial population of the fish in the pool. We infer from (Equation 2.73) that if a > 0
and N 0 ≠ 0 then
lim t → ∞ N ( t ) = a b .
This ratio is called the saturation level of the pool. It also represents the equilibrium state of the fish
population since N ˙ = 0 when N = a/b.
Fig. 2.22 displays the evolution of the (normalized) fish population with two different initial values
of N. The dashed line represents the equilibrium population.

Figure 2.22 Evolution of the normalized fish population in the pool as a function of time for two
initial populations

Example 2.4.1
Predator-prey ecosystem
Consider a lake in which there are two species of fish. The first of these F feeds on plants while the
other P is a predator of F. To write down a mathematical model for this ecosystem we assume that
the fish is consumed by the predator at a rate which is proportional to the population size of the
two species (which we denote also by F, P) . Hence
 (2.74)
dFdt=aF-bF2-cFP,b,c>0
As for the predator population we assume that
1. It will become extinct without its prey (rate of death will exceed rate of birth).
2. P increases at a rate which is proportional to F and P (remember as F increases food becomes
more abundant for P).
Thus we infer that
(2.75)
dPdt=-kP+eFP
The system (Equations 2.73)–(2.75) is a special case of Lotka-Volterra equations which model such
ecosystems.
Example 2.4.2
SIR epidemics model
The SIR model for the spread of disease or epidemics is due to W.O. Kermack and A. G.
McKendrick. It assumes that the size of the population remains unchanged (no births and deaths).
Since its inception in 1927 the model has been generalized in various ways. Here we consider only
the original model.
This model assumes that at time t = 0 a part of the population is infected with some infectious
disease. We wish to derive equations for the spread of this disease within the population.
To derive these equations the model divides the individuals within the population into three
groups:
1. S -individuals susceptible to the disease but not infected as of yet.
2. I -infected individuals who are free to mix in the population at large and transmit the disease.
3. R-individuals who contracted the disease but recovered and are no longer susceptible to the
disease.
Subject to the constraint
(2.76)
I+S+R=N
where N is the (fixed) size of the population.
Assuming a free mixing between the S and I groups we infer that:
2.77a
S˙=-aSI
2.77b
I˙=-bI+aSI
2.77c
R˙=cI
where a, b, c > 0.
We now use (Equations 2.77) to demonstrate how in some instances one can derive inferences
about the behavior of a system even without solving the differential equations that govern its
behavior. In particular we show that if the population size is constant then according to this model:
1. b = c.
2. aS(0) < b implies that there will be no epidemics, i.e. I(t) will decrease in time.
 3. For all times S(t) > 0, thus the population will always contain some healthy individuals.
To prove the first statement we differentiate (Equation 2.76) with respect to time
(2.78)
I˙+S˙+R˙=0
Substituting Equations (2.77a) - (2.77c) in (Equation 2.78) yields b = c.
To prove the second statement we first observe that S ˙ ( t ) ≤ 0 since S, I ≥ 0 in Equation (2.77a).
Hence S(t) is a decreasing function of time, i.e. S(t) ≤ S(0) . Now from Equation (2.77b) we have
(2.79)
I˙=(-b+aS)I
If aS(0) < b then aS(t) < b for all t. Therefore I ˙ < 0 ( i . e . I ( t ) ≤ I ( 0 ) ) and there will be no
epidemic.
Finally to prove the third statement we use the chain rule
(2.80)
dSdt=dSdRdRdt
Hence (from Equations (2.77a),(2.77b))
(2.81)
dSdR=-abS
Integrating this equation with respect to R we have
(2.82)
S(t)=S(0)e-a/bR>S(0)e-a/bN>0
which proves our statement.
Example 2.4.3
Chemical reactions
The basic law which governs the rate of a chemical reaction is the “law of mass action” which
states that the rate of a reaction is proportional to the (active) concentration of the reactants.
Thus for the reaction
X+Y→ Z
(2.83)
d[Z]dt=k[X][Y]=-d[X]dt=-d[Y]dt
where [X],[Y], [Z] stand for the (active) concentration of the corresponding chemicals.
If the initial active molar concentration of X, Y is a, b, then at time t
[X](t)=a-[Z](t)
(2.84)
[Y](t)=b-[Z](t)
since the production of one mole of Z requires one mole of X and Y. Hence
(2.85)
d[Z]dt=k(a-[Z])(b-[Z])
This differential equation can be solved by direct integration, and if a ≠ b we obtain (using the initial
condition [Z](0) = 0)
(2.86)
k=1t(a-b)lnb(a-[Z])a(b-[Z])
This relationship is usually used to determine experimentally the rate of the reaction. The solution for
concentration of [Z] as a function of time is
[ Z ] ( t ) = b exp [ k ( a - b ) ( t + c 1 ) ] - a exp [ k ( a - b ) ( t + c 1 ) ] - 1 .
If [Z](O) = 0 then
c 1 = ln a b k ( a - b ) .
Example 2.4.4
Catalytic reactions
Of particular interest from a chemical point of view are reactions where the addition of some
“catalyst” accelerates the rate of a reaction that is “slow going.” As an example we consider the
oxidation of sulfur dioxide using nitrogen dioxide as a catalyst.
(2.87)
NO2+SO2→ k1SO3+NONO+12O2→ k2NO2
Observe that the net result of this reaction is
(2.88)
SO2+12O2→ SO3
i.e. the amount of NO 2 in the chemical reactor remains unchanged. Thus a catalyst provides a path
for a desired reaction to happen which has a lower activation energy than the uncatalysed
reaction.
A model for the reactions in (Equation 2.87) consists of five coupled nonlinear differential
equations.
(2.89)
d[NO2]dt=-k1[NO2][SO2]+k2[NO]·[O2]1/2
(2.90)
d[NO]dt=-k1[NO2][SO2]-k2[NO]·[O2]1/2
(2.91)
d[SO3]dt=k[NO2]·[SO2]
(2.92)
d[SO2]dt=-k[NO2]·[SO2]
(2.93)
d[O2]dt=-k2·[NO]·[O2]1/2
Example 2.4.5
Radioactive decay
The nuclei of many isotopes are not stable and therefore decay over time. In many cases the
products of this decay are not stable themselves, and the system then consists of a chain of such
reactions. In all these reactions the decay rate is assumed to be proportional to the “population
size” viz. to the number of nuclei present.
As a particular example we consider here a chain of such reactions where N 1 decays to N 2, which
then decays to a stable nuclei N 3. Here N i , i = 1, 2, 3, represents both the nuclei and their
number.
To model these reactions we consider the time interval [ t , t + △ t ] . On this interval we have
2.94a
N1(t+△ t)N1(t)=-α1N1(t)△ t
2.94b
N2(t+△ t)N2(t)=α1N1(t)△ t-α2N2(t)△ t
2.94c
N3(t+△ t)N3(t)=α2N2(t)△ t,αi>0,i=1,2,3
In Equation (2.94a) the first term on the left hand side represents the number of N 1 nuclei which
were converted to N 2, while the second term represents the number of N 2 nuclei which decayed to
N 3.
Dividing by ▵ t and letting ▵ t → 0 we obtain the system
dN1dt=-α1N1
dN2dt=α1N1-α2N2
(2.95)
dN3dt=α2N2
This is a system of three coupled first order equations. The initial conditions for this system must
specify the number of the nuclei N i , i = 1, 2, 3 at some time t 0.
Exercises
1. Derive a model equation for a fish population which consumes plants as well as itself (Hint:
Remember that a represents the rate of birth minus the rate of death).
2. In a lake there are three species of fish X, Y, Z. X eats plants that are highly abundant. Y is a
predator of X and Z is a predator of X and Y. Derive a model for this ecosystem.
3. Derive a model for an ecosystem which consists of two species X, Y under the following
assumptions:
a. Both species compete for the same nutrient whose supply is limited.
b. There is a migration of X (from outside the ecosystem) at a rate r per unit time.
4. Explain why the coefficient of SI in Equations (2.77a), (2.77b) is the same.
5. Develop model equations for the concentrations of X,Y, and A in the reactions
(2.96)
A+X→ B+2X
(2.97)
X+Y→ B+2Y
(2.98)
A+Y→ B
Observe that the net result of these reactions is 2A → 3B.
6. Show that (Equation 2.86) is the solution of (Equation 2.85).
7. Solve (Equation 2.85) when a = b.
8. Show (by substitution) that the solution of (Equation 2.95) with the initial conditions N 1(0) = N
0, N 2(0) = N 3(0) = 0 is
(2.99)
N1(t)=N0e-α1t
(2.100)
N2(t)=λ2N0[e-α1t-e-α2t]
 (2.101)
N3(t)=N0[1-λ2e-α1t+λ1e-α2t]
where λ 1 = α 1 α 2 - α 1 , λ 2 = α 2 α 2 - α 1 .
9. Find the solution of the system, (Equation 2.95), when α 1 ≈ α 2. Assume that N 1(0) = N 0 and N
2(0) = N 3(0) = 0.
10. Let P, Q be the price and quantity of a certain fuel on the open market. A population model for
the evolution of these variables was proposed in the form
P˙=aP/Q-bP2
Q˙=cPQ-dQ2
where a, b, c, d > 0. Justify and discuss the meaning of this model.
2.5 Motion in a Central Force Field
In this section we discuss the equations of motion for a body in a central force field, i.e., when
F = f(r)r where r is the radius vector from the origin and r = |r|. Then various systems (such as the
pendulum) are considered. We begin, however, by introducing the radial coordinate system in R 2.

2.5.1 Radial Coordinate System in R 2

In this coordinate system (which should not be confused with the polar coordinate system) we attach
two perpendicular unit vectors (a frame) to any point in the plane. The first of these e r is a unit vector
along the radius vector connecting the point to the origin of some fixed Cartesian system, and the
second e θ is a unit vector orthogonal to e r in the counterclockwise direction (see Fig. 2.23).

Figure 2.23 Radial coordinate system

Using simple trigonometry we infer that the expressions of e r , e θ in Cartesian coordinates are
given by
(2.102)
e r = cos θ i + sin θ j , e θ = - sin θ i + cos θ j
where θ is the angle between the radius vector and the positive x axis.
To obtain expressions for the velocity and acceleration in this coordinate system we observe that
always
(2.103)
x=rer
Hence
(2.104)
X=r˙er+re˙
and
(2.105)
x¨=r¨er+2re˙+re¨
but from (Equation 2.102) we infer that
(2.106)
e˙=θ˙eθ,eθ=-θ˙er
(2.107)
e¨=θ¨eθ-θ˙er
Inserting Equation(2.106)‐(Equation 2.107) in (Equation 2.104)‐(Equation 2.105) we finally obtain
(2.108)
X=r˙er+rθ˙eθ
(2.109)
x¨=(r¨-rθ˙)er+(rθ¨+2r˙θ˙)eθ
Remark: It is useful in some applications to introduce radial coordinates in 3‐dimensions; i.e., attach
to each point in space a triad of orthonormal vectors. If (r, θ, ϕ) are the coordinates of a point in
spherical coordinates (where φ is the angle between the radius vector and z while θ is the azimuthal
angle) then these vectors are
(2.110)
e r = ( sin ϕ cos θ , sin ϕ sin θ , cos ϕ )
e ϕ = ( cos ϕ cos θ , cos ϕ sin θ , - sin ϕ )
e θ = ( - sin θ , cos θ , 0 )

2.5.2 Linear Pendulum

Objective: Derive the equation of motion of a system composed of a rigid (massless) rod and a mass
attached at its end in the gravitational field of the Earth.
For this system it is natural to redefine the angle θ as the angle between the radius vector and the
negative y axis (see Fig. 2.24). In this case the expressions for e r , e θ are given by
(2.111)
e r = sin θ i - cos θ j , e θ = cos θ i + sin θ j
Figure 2.24 Radial coordinate system for the pendulum

Figure 2.25 Force diagram for the linear pendulum

However the expressions for the velocity and acceleration, (Equations 2.108) ‐(2.109) remain
unchanged.
Model: Let x(t) denote the position of the center of mass of m at time t; then in radial coordinates
(2.112)
x=Ler
Since L is constant, we obtain from (Equation 2.109)
(2.113)
x¨=Le¨
x¨=L(θ¨eθ-θ˙er)
Expressing the gravitational force in radial coordinates
(2.114)
F g = - m g j = - m g sin θ e θ + m g cos θ e r
and using Newton’s second law yields
(2.115)
m x ¨ = m L ( θ ¨ e θ - θ ˙ 2 e r ) = - m g sin θ e θ + m g cos θ e r - T e r
where T is the tension in the rod. Rewriting (Equation 2.115) in components form we obtain
(2.116)
L θ ¨ = - g sin θ
(2.117)
m L θ ˙ = T - m g cos θ
The second of these equations can be considered as an equation for T while the first is the desired
equation of motion for the pendulum, i.e.
(2.118)
θ ¨ = - g L sin θ = - g L ( θ - θ 3 3 ! + … )
We observe that (Equation 2.118) is nonlinear in θ. However for small θ we can approximate
(Equation 2.118) by
(2.119)
θ¨+gLθ=0
which is formally the same equation as for the spring mass system without friction. The general
solution of this equation is θ = A cos ( ω t + ϕ ) where A, φ are integration constants and ω 2 = g L .
The period of this pendulum is P = 2 π ω .

2.5.3 Nonlinear Pendulum

We now want to treat the nonlinear (Equation 2.118) subject to the initial values θ(0) = α and θ ˙ = 0 .
To find the solution of this problem we first multiply eq. (2.118) by dθ/dt. This yields
(2.120)
1 2 d d t ( d θ d t ) 2 = - ω 2 sin ( θ ) d θ d t
where ω 2 = g L . We can integrate this equation once on the interval [ 0 , t ] to obtain (using the
initial conditions):
(2.121)
( d θ d t ) 2 = 2 ω 2 ( cos θ - cos α ) .
This can be rewritten also as
( d θ d t ) 2 = 2 ω 2 ( - 1 + cos θ + 1 - cos α )
using the identity 1 - cos ( ϕ ) = 2 sin 2 ( ϕ / 2 ) we then obtain
( d θ d t ) 2 = 4 ω 2 ( sin 2 α 2 - sin 2 θ 2 )
We now introduce a new variable ψ which is defined by sin ( ψ ) = sin ( θ / 2 ) sin ( α / 2 ) . The
expression above becomes
 (2.122)
( d θ d t ) 2 = 4 ω 2 sin 2 α 2 cos 2 ψ .
Using the definition of ψ we obtain (using implicit differentiation) for the left hand side of this
equation
( d θ d t ) 2 = 4 sin 2 α 2 cos 2 ψ c o s 2 θ 2 ( d ψ d t ) 2
(Equation 3.21) transforms therefore into
(2.123)
d ψ d t = ω cos θ 2 = ω 1 - k 2 sin 2 ψ
where = sin ( α / 2 ) . This equation can be integrated now, and we obtain the implicit solution to our
problem in the form
(2.124)
a + ω t = ∫ 0 ψ ( t ) d ψ 1 - k 2 sin 2 ψ
where a is an integration constant. The integral on the right hand side of (Equation 2.124) is an
elliptic integral of the first kind, since at time 0 sin ( ψ ) = 1 we infer that ψ(0) = π/2. Hence
a = ∫ 0 π 2 d ψ 1 - k 2 sin 2 ψ .
We can rewrite Equation(2.124) in the form
(2.125)
t = 1 ω ∫ π 2 ψ ( t ) d ψ 1 - k 2 sin 2 ψ
The pendulum will complete one period when θ first takes the value θ = α, i.e., ψ = 2 π + π 2 . Using
the following identity for elliptic functions f ( k , ϕ + n π ) = n f ( k , π ) + f ( k , ϕ ) (see next
subsection) we conclude that the period of the nonlinear pendulum is
T=2ωf(k,π)
or explicitly
T = 2 ω ∫ 0 π d ψ 1 - k 2 sin 2 ψ .

2.5.4 A Short Introduction to Elliptic Functions

Elliptic integrals of the first and second kinds are defined as
(2.126)
f ( k , ϕ ) = ∫ 0 ϕ d ψ 1 - k 2 sin 2 ψ
(2.127)
e ( k , ϕ ) = ∫ 0 ϕ 1 - k 2 sin 2 ψ d ψ
respectively. The variable φ is called the amplitude of f or e. Related to these integrals are the
functions
(2.128)
s n ( k , ϕ ) = sin ( f ( k , ϕ ) )
(2.129)
c n ( k , ϕ ) = cos ( f ( k , ϕ ) )
(2.130)
dn(k,ϕ)=1-k2sn(k,ϕ)
The expressions f ( π / 2 , k ) and e ( π / 2 , k ) are called complete elliptic integrals. The basic
properties of these integrals and functions are:
(2.131)
f(k,π)=2f(k,π/2),e(k,π)=2e(k,π/2)
(2.132)
f(k,ϕ+nπ)=nf(k,π)+f(k,ϕ)
(2.133)
e(k,ϕ+nπ)=ne(k,π)+e(k,ϕ)
(2.134)
sn2(ϕ,k)+cn2(ϕ,k)=1
(2.135)
sn(-ϕ,k)=-sn(ϕ,k),cn(-ϕ,k)=cn(ϕ,k)
Remark: A standard text which contains an extensive list of functions that appear in many applications
is
M. Abramowitz and A. Stegun-Handbook of Mathematical functions (Dover Publications).

2.5.5 Motion of a Pr ojectile on a Rotating Earth

Example 2.5.1
Derive the equations of motion of a projectile taking into account the rotation of the Earth.
Solution 2.5.1 When the rotation of the Earth is taken into consideration, it is natural to use an
inertial coordinate system fixed at the Earth’s center to describe the motion. In this coordinate
system the Earth’s rotation introduces an additional term in Newton’s second law which is called
Coriolis force. The expression for this force is
(2.136)
Fc=-2mω×x˙
where ω = ( 0 , 0 , ω ) is the Earth’s angular velocity and x = ( x , y , z ) is the position of the
particle. The equations of motion of the projectile are then
(2.137)
mx¨=mg-2mω×x˙
where in spherical coordinates,
(2.138)
g = - g ( sin ϕ cos θ , sin ϕ sin θ , cos ϕ )
For a projectile with limited range θ, φ can be considered to be constants and the equations of
motion in scalar form are
(2.139)
x ¨ = - g sin ϕ cos θ - ω y ˙
(2.140)
y ¨ = - g sin ϕ sin θ + ω x ˙
(2.141)
z ¨ = - g cos ϕ
This is a system of three coupled second order differential equations which must be solved
simultaneously for x, y, z (however, note that the equation for the z coordinate is decoupled from the
other two equations).

2.5.6 A Particle in a Central Force Field

Objective: Derive the equation for the motion of a body of mass m in a force field where the force
acting on it is F = mf(r)e r .
Solution:
From Newton’s second law we know that
(2.142)
mx¨=mf(r)er
Using (Equation 2.109) we infer that
(2.143)
r¨-rθ˙2=f(r)
(2.144)
rθ¨+2r˙θ˙=0
The second of these equations can be integrated once (after multiplication by ), and it follows that
(2.145)
r2θ˙=constant=h.
(This is the law of conservation of angular momentum.) Hence we can write formally
d2dt=hr2ddθ
and
(2.146)
d2dt2=hr2ddθ(hr2ddθ)
Substituting (Equation 9.9) and (Equation 2.146) in (Equation 2.143) we obtain
(2.147)
hr2ddθ(hr2drdθ)-h2r3=f(r)
To simplify this equation we introduce u = 1 r and note that d u d θ = - 1 r 2 d r d θ . This leads to
(2.148)
h2u2[d2udθ2+u]=-f(1u)
For the Newtonian gravitational field we have f ( r ) = - μ r 2 where μ is a constant (for the resulting
orbits see Ex. 5).

2.5.7 Motion of a Rocket

Objective: Derive a prototype model for the motion of a one stage rocket. Background: To derive the
required equation we have to use the law of momentum conservation (actually this is a restatement of
Newton’s second law). This law states that the net change in the momentum of a physical system in
time ▵ t is (approximately) equal to f ext ▵ t where f ext is the sum of the external forces acting on the
system.
 Approximations:
1. We assume that the flight of the rocket is along the radius vector of the Earth, i.e., “straight up.”
2. Neglect the existence of the Earth atmosphere
3. Neglect the rotation of the Earth.
4. Neglect the ignition stage and the stability of the rocket in flight (usually to stabilize this motion
the rocket rotates around its axis).
5. Neglect the gravitational field of the Sun and Moon.
6. Assume the Earth is a perfect sphere.
Model: The motion of the rocket is due to the fuel that is being burnt. This generates a stream of
gases which leave the rocket and thus exert a forward thrust. To a good approximation the velocity u
of these gases is constant (except at ignition). Since the only external force acting on the rocket is the
gravitational field of the Earth we obtain by applying the law of momentum conservation (see Fig.
2.26).
Momentum Momentum Momentum Momentum
of gases of of imparted
leaving rocket in+ rocket at ‐rocket at t = by gravitational
[t,t+△ t] [t + ▵ t] field in
[t,t+△ T]

Figure 2.26 Rocket at time and

Thus if v(t) is the velocity of the rocket and m(t) is its mass we have
(2.149)
[-dmdt△ t](v(t)-u)+m(t+△ t)v(t+△ t)-m(t)v(t)=-m(t)g(h(t))t
where h(t) is the altitude of the rocket at t. Dividing by ▵ t → 0 we obtain
(2.150)
ddt(m(t)v(t))=dmdt(v(t)-u)-m(t)g(h(t))
If we let g(h(t)) be a constant, then the equation of motion reduces to
(2.151)
dvdt=-u1mdmdt-g
Integrating with respect to t we obtain
(2.152)
v = v 0 + u ln m 0 m ( t ) - g t
where m 0 is the initial mass of the rocket and v 0 is its initial velocity. We see that if u ≅ 3 k m / sec
and the flight of the rocket is short, the effect of the last term is negligible and therefore
(2.153)
v = v 0 + u ln m 0 m ( t )
In a one stage rocket
(2.154)
m0=mp+mf+ms
where m p is the mass of the payload, m f the initial mass of the fuel and m s is the mass of the
structures engines and fuel containers. The final velocity of such a rocket (i.e., when the fuel is
exhausted) is given then by
(2.155)
v F = v 0 + u ln m 0 m p + m s
Usually m s m 0 ≈ 1 10 . We see therefore that even if m p = 0 and the effect of
gravity is neglected
v F = v 0 + u ln 10 .
Letting v 0 = 0 and u = 3 we obtain v F ≈ 7.2 km/sec.

2.5.8 Multistage Rockets

Objective:
A team of engineers, physicists, and mathematicians is asked to examine the desirable gross
structure of a rocket capable of putting a satellite in orbit 600 kms above the Earth surface
(concentrating on the energy aspects of the problem).
Approximations:
We make here the same approximations as in the previous section but in addition we neglect the
final stage of actually putting the satellite in orbit and assume a circular orbit for the satellite.
Reduction of the problem:
I. Determine the speed of the satellite which will keep it in orbit 600 km above ground.
II. Develop and solve a model for the motion of a multistage rocket.
III. Find out the optimal structure of the rocket to be measured by the ratio of payload
( ≡ satellite)/total weight.
We discuss each of these points separately.
I. Once the satellite is in orbit, there are two forces which act on it, viz. the gravitational force of
the Earth given by F G = μ m r 2 and the centrifugal force F c = m v 2 r , where μ is a constant
depending on the mass of the Earth and the nature of the gravitation force. We also know that on the
surface of the Earth:
mg=μm/R2
where R is the radius of the Earth (approximately 6400 km). For the satellite to stay in orbit the two
forces must cancel each other, i.e.
mv2r=μm/r2
but μ = gR 2 hence v 2 = gR 2/r for r = 6400 + 600 = 7000 km. We obtain
v ≅ 7.6 k m / s e c .
Using the results obtained in the previous section we see that using present technology it is
impossible to put a satellite in orbit using a one stage rocket. The reason for the poor performance of
this rocket is due to the fact that the engine has to carry not only the payload but also the whole
structure up to the final orbit. Ideally, however, one should be able to jettison useless weight as the
burning proceeds, i.e. keep the mass ratio between fuel and engines and containers constant.
II. Motion of a One Stage Ideal Rocket.
Let us therefore examine the performance of such a rocket. Let m(t) be the rocket mass (without the
payload) at time t. Of this the mass of the fuel is m f ( t ) = ( 1 - λ ) m ( t ) and the mass of the structure
is m s = λ m ( t ) . The ratio m f m s = 1 - λ λ will remain constant throughout the flight if λ is
independent of time. When λ is constant we have λ m f = ( 1 - λ ) m s and hence m s = 0 when the fuel
is exhausted (i.e.m f = 0) .
We now show that λ will remain constant in time if on the interval [ t , t + △ t ] the mass of the fuel
burned equals △ m f ( t ) = - ( 1 - λ ) d m d t △ t and the corresponding mass of the structure that is
jettisoned away is △ m s ( t ) = - λ d m d t △ t (remember that the total mass change of the rocket on
this time interval is - d m d t ▵ t
In fact since m f ( t ) m s ( t ) = 1 - λ λ and △ m f ( t ) △ m s ( t ) = 1 - λ λ we infer that
mf(t+△ t)ms(t+△ t)=mf(t)-△ mf(t)ms(t)-△ ms(t)=1-λλ.
The equation for the conservation of momentum will now read (neglecting gravity)
m(t)v(t)=m(t+△ t)v(t+△ t)-λdmdt△ tv(t)-(1-λ)dmdt△ t(v(t)-u)
Dividing by ▵ t and letting ▵ t → 0 we then obtain
mdvdt=-(1-λ)udmdt,
therefore
v ( t ) = ( 1 - λ ) u ln m 0 m ( t ) .
At the final state of this rocket we now have m f = m p and hence
v F = ( 1 - λ ) u ln m 0 m p .
If we assume that in order to account for air resistance, gravity, etc. we need v F = 10.5 km/sec
(rather than 7.6 km/sec) and λ ≅ 0.1 we obtain m 0/m p = 50. Thus the payload in this idealized
situation cannot exceed 1/50 of the initial mass of the whole system.
III. Optimal Design of a Multistage Rocket.
Naturally the ideal performance discussed above is impossible to implement in practice. We must
therefore approximate it by building a multistage rocket and jettison each stage when its fuel is
exhausted.
Let m i = mass of fuel and structures of the i-th stage. Of this λ m i is the mass of the structure
(engines and containers) and ( 1 - λ ) m i is fuel. (We assume same λ for all stages.)
The total mass of a three stage rocket will then be
mt=mp+m1+m2+m3.
The equations that govern the motion of this rocket during the “burning” of the first stage are the same
as for the one stage rocket. Hence the final speed of the rocket at the end of the first stage is
v 1 = u ln m 0 m p + λ m 1 + m 2 + m 3 .
At this moment the structural mass λ m 1 is dropped and the second stage is ignited. When the fuel in
the second stage is exhausted, the final speed of the rocket will be
v 2 = v 1 + u ln m p + m 2 + m 3 m p + λ m 2 + m 3
and similarly at the end of the third stage we have
v F = v 3 = v 2 + u ln m p + m 3 m p + λ m 3 .
Combining all these results we obtain
v F u = ln ( m 0 m p + λ m 1 + m 2 + m 3 ) ( m p + m 2 + m 3 m p + λ m 2 + m 3 ) × ( m p + m 3 m
p+λm3)
We now try to maximize m p /m 0 for fixed v F , u, and λ . To this end we introduce the variables
x1=m0mp+m2+m3,x2=mp+m2+m3mp+m3,X3=mp+m3mp
therefore
v F u = ln { ( x 1 1 + λ ( x 1 - 1 ) ) ( x 2 1 + λ ( x 2 - 1 ) ) ( x 3 1 + λ ( x 3 - 1 ) ) }
Since m p appears in the denominators of x 1, x 2, x 3, we would like to minimize the values of these
variables. However, by the symmetry of this expression in x 1, x 2, X3 we must have at minimum x
1 = x 2 = x 3 = x. Therefore at minimum
v F u = ln ( x 1 + λ ( x - 1 ) ) 3 and hence x = 1 - λ q - λ
where q = exp ( - v F / 3 u ) . But
m0/mp=x1x2x3=(1-λq-λ)3
which gives for the values of λ , u, v F used so far m 0/m p = 77.

2.5.9 Control of a Satellite in Orbit

Consider a satellite in a circular orbit around the Earth which is equipped with thrusters to correct
deviations in its orbit. These deviations are in general due to several “secondary effects” such as the
gravitation of the Sun and the Moon, friction with the upper atmosphere, etc. We want to derive model
equations for the motion of the satellite so that one can find out how to use the thrusters to effect the
needed corrections in orbit.
Solution:
a. General Equations of Motion:
If we assume that the satellite always remains in one plane and that the only force acting on it is the
Earth’s gravitation we deduce from (Equation 2.109) that its equations of motion are
(2.156)
r¨=rθ˙2-GMr2+u1
(2.157)
θ¨=-2r˙θ˙r+1ru2.
In these equations G is the gravitational constant, M the Earth’s mass, and u 1, u 2 are the radial and
tangential thrusts.
If initially the satellite is in a circular orbit of radius R and constant angular velocity ω then (since
in this orbit the centrifugal and gravitational forces must be equal and opposite)
(2.158)
mRω2=GMmR2i.e.R3ω2=GM
and this is also a solution of (Equations 2.156), (2.157) with u 1 = u 2 = 0.
b. Equations for Small Deviations in Orbit.
Since we consider only small deviations from the circular orbit described above, it is natural to
introduce as variables the deviations of the satellite in position and velocity from the parameters of
this orbit:
(2.159)
x1=r-RR,x2=r˙R
x3=θ-ωt,x4=θ˙-ω
Substituting these variables in each of the terms that appear in (Equations 2.156) –(2.157) yields
Rx˙2=R(1+x1)(ω+x4)2-GMR2(1+x1)2+u1
x˙4=-2Rx2(ω+x4)R(1+x1)+u2R(1+x1).
We now use the following expansions
11+x=1-x+x2
1(1+x)2=1-2x-3x2
(the second expansion is obtained by differentiating the first) and neglect all the nonlinear terms that
appear in these equations (remember: the deviations and the thrusts u 1, u 2 are assumed to be small).
Using (Equation 2.158) we obtain the following system of differential equations:
(2.160)
dxdt=Ax+1RBu
where x = x 1 x 2 x 3 x 4 , u = u 1 u 2 and
(2.161)
A=01003ω2002ω00010-2ω00,B=00100001
Exercises
1. Enumerate at least five approximations which are made in the derivation of the model equation
for the pendulum.
2. Compare the period of the nonlinear pendulum with that of the linear pendulum when θ(0) = 5,
10, 20, 85 degrees and θ ˙ ( 0 ) = 0 .
3. Use (Equation 2.115) to derive the equation of motion for a pendulum system where the rod has
been replaced by a spring (Fig. 2.27).
Figure 2.27 A Pendulum with spring attachment

Hint: The force exerted by the spring is - k(L - L 0)e r where L 0 is the natural length of the spring.
4. Simulate the system of differential equations that was obtained in the previous exercise with
proper choices for L and k (using “MATLAB” or similar). Compare the period and phase diagrams
(i.e. plot θ ˙ vs. θ) for this pendulum with those for the regular pendulum in Ex 1.
5. Show that when f ( r ) = - μ r 2 in (Equation 2.148) then
1 r = k c 2 ( 1 + e cos ( θ - θ 0 ) )
where e, θ 0 are constants. Discuss the nature of the orbits described by this equation for different
values of e.
Hint: Show that for e < 1 the orbit is an ellipse, e = 1 a parabola, and for e > 1 a hyperbola.
6. Carry out the derivation of (Equation 2.160) from (Equations 2.156)–(2.157)
7. Verify Equations(2.102)‐(2.111).
8. Develop a model for a one stage rocket with air friction. Assume F f = αm(t)v(t) [try to solve
by neglecting gravity]
Hint: d v d t = d v d h d h d t = v d v d h
9. Develop model equations for the motion of the rocket if u = u(t) where u is the speed of the
gases leaving the rocket.
10. Find the optimal value of m 0/m p for a two stage and four stage rocket. What happens when the
number of stages goes to ∞?
Solution
The ratio L = m 0 m p for n‐stage rocket is given by
L=[1-λexp(vFnu)-λ]n
To find the limit of this expression as n → ∞ we rewrite it in the form
exp { ln [ 1 - λ e x p ( v F n u ) - λ ] n } = exp { n ln [ 1 - λ e x p ( v F n u ) - λ ] }
We then have
lim n → ∞ n { ln [ 1 - λ e x p ( v F n u ) - λ ] } = lim n → ∞ ln [ 1 - λ e x p ( v F n u ) - λ ] 1 / n .
As n → ∞ both numerator and denominator in the last expression approach zero; therefore, we can
compute this limit by replacing n (a discrete variable) by x (a continuous variable) and apply
L’Hopital rule:
lim x → ∞ ln [ 1 - λ e x p ( v F x u ) - λ ] 1 / x = lim x → ∞ v F / u 1 - λ e x p ( v F x u ) = v F / u 1
-λ.
Hence
L = exp ( v F / u 1 - λ )
11. A person wants to put a satellite in orbit around the Earth at 64000 km from the Earth’s center. A
short calculation similar to the one in the text shows that the speed of the satellite in orbit should be
(approximately) 2.8 km/sec. Accordingly this person concludes that a one stage rocket is suitable for
this purpose. Explain why this conclusion is wrong.
12. Estimate the effect of gravity on the flight of a one stage rocket if its mass is decreasing linearly
with time:
m = m 0 ( 1 - 5 . 10 - 4 t )
and the flight time is a. 100 sec. b. 190sec. Hint: Use (Equation 2.152) with and without the gravity
term assuming g(h) to be a constant 0 ≤ g ≤ 9.8m/sec 2.
13. Write down model equations for the motion of a rocket which is launched vertically from under
the sea.
Hint: Both below and above sea level the forces acting on the rocket are the thrust T(h), gravity,
and the frictional forces which are proportional to the speed ( o r ( s p e e d ) 2 ) of the rocket. Since
the coefficient of friction is different for h > 0 and h < 0 we obtain two different equations for these
two ranges of h.
14. A person wants to put a satellite in circular orbit around the Earth so that the satellite will
always stay over the same point on Earth. Compute the height and velocity of this satellite (these are
called “geocentric satellites”).

2.6 Greenhouse Effect

Energy coming from the Sun to Earth is “usually” reflected back to space. Some trace gases such as
carbon dioxide (CO 2), methane (CH) and others can block part of this energy from going back to
space and thereby raise the mean temperature of the Earth. This is called the Greenhouse Effect.
There are many models that attempt to gauge the impact of the greenhouse effect on Earth’s climate.
These models differ in the number of spatial dimensions which are incorporated into the model, their
sophistication (in modeling the properties of the Earth surface and atmosphere), and resolution. At the
bottom of this “ladder” of models are the 0‐dimensional models where all variables depend only on
time, i.e., we consider only the mean of these variables as a function of time, e.g., the mean
temperature of the Earth.
In this section we consider only 0‐dimensional models and start with some qualitative examples
which give some insights about the mechanism of this effect.
Example 2.6.1
Perfect Mirror
If light is directed at a perfect mirror, then all this light energy will be reflected back and none
will be absorbed by the mirror. As a result the mirror temperature remains unchanged.
Example 2.6.2
Black Body
By definition a black body absorbs all radiation that is impinging on it. Since radiation is a form
of energy, the temperature of this body will rise. However, a black body also emits radiation (per
unit area) at a rate which is proportional to the forth power of its temperature. This is the famous
Stephan-Boltzmann law. At the equilibrium temperature of such a body the amount of energy
absorbed is equal to the amount of the energy emitted and hence
P=σT4,
where P is the incident power and T is the equilibrium temperature (usually the energy emitted
from such a body is in the infra‐red end of the spectrum, i.e. long wave length).
Example 2.6.3
Grey Body
A grey body is a body whose properties are in between the two extremes described above. That is,
such a body reflects part of the electromagnetic radiation, absorbs part of it, and then emits part
of this energy again as thermal energy. The ratio of the reflected radiation to the amount incident
upon it is called the “albedo” of the body. Usually the albedo is expressed in percentage form.
Current estimates for the albedo of Earth are in the range of 30—39 percent.
For a grey body we therefore have
(2.162)
(1-A)P=σT4
where A is the “albedo” (or grayness). For a black body A = 0.
If Earth had no atmosphere then it would be a good example of such a grey body. It would reflect
part of the Sun’s energy (this would depend strongly on the ice and snow cover) and would absorb
part of it. The part absorbed would lead to a rise in its temperature until the amount of thermal energy
radiated into space equals the amount absorbed by it. It is estimated that under these conditions the
mean temperature of the Earth would be approximately - 25∘ C (see next section). Thus Earth would
be a cold rock covered with ice and snow except (perhaps) for a small belt around the equator.
To understand the effect of Earth’s atmosphere on the mean temperature of the Earth we have to
explain first how light is absorbed and emitted by atoms (or molecules).
When it comes to the processes of absorption or emission of energy from atoms, light should be
considered as made of ‘particles’ called photons. These photons carry energy of hν where h is called
the Planck constant and ν is the electromagnetic wave frequency. An atom can absorb such a photon
only if it has two energy levels E 1, E 2 so that
(2.163)
E2-E1=hν
If such an atom is at energy level E 1, then by absorbing a photon with energy hν it gets “excited” to an
energy level E 2. Such an excited atom will then relax back to its ground energy level E 1 by releasing
a photon with energy hν (but in a direction which might be different from the one in which the original
photon was traveling). Such a photon is then “trapped” in Earth’s atmosphere and has a chance of
being reabsorbed by Earth and raising its mean temperature.
It turns out that although O 2 and N 2 are the main components of the Earth’s atmosphere, they do not
have the proper energy levels (or bands) that can absorb the thermal (long wave length) energy that is
radiated by the Earth. Therefore they play only a minor role in this process. On the other hand CO 2,
methane (CH 4), and other trace gases do have exactly the right energy bands needed for the trapping
of this energy, thereby leading to the greenhouse effect.
 To develop a 0-dimensional model for the amount of CO 2 (or CH 4) in the atmosphere it is
customary to divide the Earth‐atmosphere system into several reservoirs, and the number of these
reservoirs depends on the fine details and complexity of the model. A popular model which strikes a
balance between simplicity and complexity divides the Earth‐atmosphere system into seven
reservoirs as follows:
1. Troposphere(lower atmosphere up to ≈ 10Km above Earth surface)
2. Stratosphere and up (upper atmosphere)
3. Upper ocean layer (up to a depth of 200‐300 meters)
4. Deep ocean (below the upper Ocean layer)
5. Short lived vegetation (corn,wheat etc) and animal life
6. Long lived vegetation(trees, forests)
7. Marine life
Furthermore the model contains a forcing term due to human‐made effects (burning of fossil fuels,
paved tar roads, deforestation, etc.) and natural catastrophes (e.g. volcanic eruptions) (see Fig. 2.28).

Figure 2.28 Schematic overview of the interaction among the reservoirs

These reservoirs interact dynamically (through diffusion and turbulent mixing) in the same way as
chemical reactants do in a chemical reaction; i.e., we assume that the rate of CO 2 transfer between
the reservoirs is proportional to the concentration of CO 2 in each reservoir. However, not all the
reservoirs connect with each other. If we denote the CO 2 concentration in reservoir i by C i then
(2.164)
dCidt=∑i=17kijCj+Fi,i=1,…,7
where k ij are constants and F i is the forcing (in this model F i = 0 except for the lower atmosphere).
Thus the model depends on 49 parameters (some of which are zero). The proper estimation of these
coefficients (which remains an outstanding research issue) is one of the important steps that is needed
to make accurate predictions about the concentration of CO 2 in the atmosphere.
Once we determine the (mean) concentration of CO 2 in the lower and upper atmosphere, another
model is needed to gauge the impact of this concentration on the mean temperature of Earth (see
chapter 11).

2.7 Current Energy Balance of the Earth

The mean temperature of Earth has been “stable” for many years (with a small increase in the last
decades due to the greenhouse effect). It is reasonable therefore to assume that at the present time the
amount of energy incoming to Earth approximately equals the outgoing amount into space. In the
following we present the current estimates that lead to this ““equilibrium.”
Earth receives energy from the Sun in short and visible wavelengths ( λ ≤ 0.5 m ) and emits energy
back to space as long wave radiation ( λ ∼ 10 μ m ) . To achieve equilibrium these two processes
must balance each other (we neglect energy from geothermal sources as their impact is negligible).
Let the energy flux from the Sun be denoted by F s ( ≈ 1372W/m 2) . A fraction A(≈ 30%) of this
energy is reflected back to space by the Earth and atmosphere (this fraction is called the albedo of the
Earth).
The long wave radiation from Earth can be described by the black‐body
Stefan‐Boltzmann law
(2.1)
P=σT4,
where P is the energy flux integrated over all wave lengths emitted by such a body at temperature T
and σ the Stefan‐Boltzmann constant ( = 5.67 c d o t 10 - 8 W a t t s m 2 K 4 ) .
For balance we must have therefore
(2.2)
(1-A)FsπR2=4πR2σTe4,R=Earthsradius.
This leads to T e = 255K which is well below the freezing point and 33 o degrees lower than the
observed mean temperature of Earth T 0 = 288 o k. The difference between T e and T 0 is due to the
greenhouse effect and the structure of the atmosphere and the oceans.
We now examine these issues in greater detail. The average energy flux from the Sun to the Earth is
F ¯ s = F s 4 = 343 W a t t s m 2 sec (the factor “4” represents the ratio of the Earth cross section to
its surface area). Of this flux 20 Watts are absorbed by clouds, 48 Watts are absorbed by water,
vapor, O 3 and other aerosols, 169 Watts arrive at the surface, and the rest are reflected back to space.
As to long wave radiation flux, the Earth emits 390 W a t t s m 2 sec in these wavelengths. Of these
22 go to space, 120 are absorbed by clouds, and 248 are absorbed by water vapor, O 3, and other
aerosols. However, clouds and aerosols emit 327 W a t t m 2 sec back to Earth.
These lead to the following approximate budgets:
1. Earth‐surface energy flux W a t t m 2 sec
169 327 -390
Short wave Long wave absorption Long wave
Radiation from atmospheric sources emission
This table shows that there is a surplus of 106 W a t t m 2 sec in this flux.
To balance this energy surplus the surface transfers heat to the atmosphere through conduction at a
rate of 16 W a t t m 2 sec . Furthermore the oceans transfer 90 W a t t m 2 sec through evaporative
cooling (“latent heat”). Without these two processes the Earth surface would have had to be 50K
warmer to balance this extra energy through black body radiation.
 The budget for the atmospheric Energy flux is as follows:
2. Atmospheric flux W a t t m 2 sec
68 +368 -327 -215
Short Long wave Longwave Long
Wave absorption emission wave
absorption (from surface) to surface emission to space
According to this table there is a deficiency in the atmospheric flux which is equal to the surplus of
the Earth’s energy flux; this keeps the the Earth’s total energy in balance. This balance will be
disturbed if, due to enhanced greenhouse effect, the long wave emission to space from the atmosphere
is diminished.

2.7.1 Critique of the Model

The majority of climate scientists accept the model that man made emissions of CO 2 and other trace
gases affect the energy balance of the Earth and raise the mean temperature of the Earth. However
there are others who argue that these emissions are small when compared to other natural sources of
these gases. There is also some evidence that when these gases diffuse into the upper atmosphere they
might have a “cooling effect” by reflecting more of the Sun energy into space. Another major factor
that impacts the energy balance of the Earth is the Sun energy output which fluctuates in time. Thus the
Earth’s climate is influenced of by a large number of variables and the impact of many of these
variables is not well understood as of yet. Earth and its climate form a truly complex system due to
the interaction of many factors.

2.7.2 Humanity and Energy

Abundant energy at a reasonable cost is a key ingredient for the well being of humanity and its
technological progress. Since the advent of the industrial revolution humanity relied on fossil fuels
(coal, oil, and gas) to satisfy these needs. Ingenious technologies were developed to tap additional
sources of these fuels. For example in the last few decades energy companies were able to access oil
and gas reserves buried in the deep ocean. Fracking technology which can unlock gas reserves in
certain types of rocks was developed. Extraction of oil from tar sands became feasible.
In spite of all these advances it is obvious that these resources are finite. How long they will last is
an open question in view of the growing human population and the energy consumed per capita.
Furthermore, fossil fuels contribute to pollution, the rise in the mean temperature of the Earth, and
have an impact on the health of humans. (There is some evidence that the rise of respiratory diseases
and cancer cases is due to the extensive use of these fuels). It follows then that in order to ensure
sustained energy supply, alternative energy sources have to be used and proper technologies have to
be developed to harvest these sources at competitive prices. Foremost among these sources are the
wind, using wind turbines, and Sun energy using photovoltaic. However other sources are being
considered at the present time. These are biofuels (i.e energy from plants and algae), ocean energy
(from the tides and ocean currents), and energy from space which requires the placement of large Sun
energy collectors in geocentric orbits around the Earth and beaming this energy to Earth. However the
price per energy unit from all of these sources presently exceeds the price of fossil fuels by a wide
margin. It should be remembered, however, that fossil fuels have a hidden cost due to their impact on
the environment and human health. Furthermore, it can be expected that fossil energy prices will
skyrocket as their reserves dwindle.
In the future (and to some extent at the present time) energy supply will come from a basket of
sources. Energy policy will be required then to use these different energy sources to optimize the
following requirements,
1. Ensure sufficient energy supply to satisfy humanity needs,
2. Minimize the cost per energy unit to the consumer,
3. Minimize the impact on the environment and human health,
4. Ensure the sustainability of the energy supply.
 CHAPTER 3

Solutions of Systems of ODEs

CONTENTS
3.1 Review
3.1.1 Linear differential equations with constant coefficients
3.2 Review of Linear Algebra
3.2.1 Eigenvalues and Eigenvectors
3.3 Reformulation Of Systems Odes
3.4 Linear Systems With Constant Coefficients
3.5 Numerical Solution Of Initial Value Problems
3.5.1 Euler Algorithm
3.6 Finite Difference Approximations
3.6.1 Extension to Higher Dimensions
3.7 Modified Euler And Runge-Kutta Methods
3.7.1 Modified Euler Algorithm
3.7.2 Runge-Kutta Methods
3.8 Boundary Value Problems
Differential equations, which relate a set of unknown functions with their derivatives, play an
important role in many applications of science and engineering. In this chapter, we present some basic
analytical and numerical methods for the solution of some classes of systems of differential equations.
However, our treatment is not comprehensive. We start with a short review of some basic theory.

3.1 Review

3.1.1 Linear differential equations with constant coefficients

The general form of n-th order linear differential equation with constant coefficients is
(3.1)
any(n)+a(n-1)yn-1+…+a1y′+a0y=f(x)
where y is a function of x , { a k , k = 0 , … n } are constants, and y k = d k y d x k . When f(x) = 0,
we say that the equation is homogeneous. Otherwise, we refer to it as inhomogeneous (or non‐
homogeneous).
A unique solution of (Equation 10.1) is obtained by specifying n initial conditions (i.e. all the
conditions on the desired solution are specified at one value of x) or boundary conditions (where the
conditions are specified at different values of x). In the first case, we refer to the problem as an
“initial value problem,” and in the second case, we refer to it as a boundary value problem.”
When f(x) = 0, the solutions of (Equation 10.1) satisfy the superposition principle which states that
if y 1 and y 2 are two solutions of Equation(10.1) and c 1, c 2 are constants then
y=c1y1+c2y2
is also a solution of (Equation 10.1).
Example
The following differential equation
(3.2)
ay′′+by′+cy=f(x),y(0)=c1,y′(0)=c2
where y = y(x) and a, b, c, c 1, c 2 are constants is a second order differential equation with two
initial conditions on y at x = 0.
When (Equation 10.1) is homogeneous, the general solution of this equation y h is given by
(3.3)
yh=α1y1+α2y2+…+αnyn
where { y 1 , y 2 , … , y n } are n independent solutions of (Equation 10.1) and α 1, \ldots , α n are
constants.
When (Equation 10.1) is inhomogeneous (f(x) ≠ 0) , the general solution is obtained if one “particular
solution” y p of the whole equation is known. The general solution of the equation is given by
yG=yh+yp.
If the general solution of (Equation 10.1) is known, then the solution that satisfies a set of initial
conditions or boundary conditions can be found by solving a system of algebraic equations for the
constants α 1,....,α n .
Example 3.1.1
Here is an example of a boundary value problem:
(3.4)
y″ +y′ -2y=0,y(O)=0,y(π)=0.
Example 3.1.2
Initial value problem:
For the same differential equation as in the previous example 3.1.1 with y(0) = 1, y′(0) = 2 is an
initial value problem.
In the following, we consider only linear second order equations with constant coefficients (as in
(Equation 3.2)). However, the methods described for these equations hold for higher order equations
with constant coefficients.
(Equation 3.2) with f(x) = 0 is a homogeneous second order equation. It has two independent
solutions: y 1(x) and y 2(x). To find these two solutions we use “trial solution” (or “ansatz”) of the
form
(3.5)
y=ekx
where k is a constant to be determined. Substituting this expression in (3.2) we obtain
(3.6)
(ak2+bk+c)ekx=0.
Since e kx ≠ 0, k must be a root of the “characteristic polynomial”
(3.7)
ak2+bk+c=0.
This quadratic equation can have
1. Two distinct real roots: k 1, k 2.
2. Double real root:r.
3. Two conjugate complex roots: k 1,2 = α ± iω
We now discuss each of these possibilities separately.
1. Two distinct real roots: (b 2 - 4ac > 0)
In this case, the “trial solution” yielded two independent solutions to the homogeneous
equation. The general solution is in the form
(3.8)
yh=α1ek1x+α2ek2x.
2. Double root: (b 2 - 4ac = 0)
In this case, the characteristic (Equation 10.7) yields only one independent solution e rx
where r = - b 2 a . However, a second solution is obtained in the form xe rx and the general
solution of Equation. (3.2) is
(3.9)
yh=er x( Î ±1+Î ±2x)
3. Two complex conjugate roots
In this case, we again obtain two independent solutions; the general solution to (Equation 3.2) can
be written as
(3.10)
yh=ekx(c1eωx+c2e-ωx)
where c 1, c 2 are arbitrary constants. However,
(3.11)
e ± ω x = cos ω x ± i sin ω x .
Hence,
(3.12)
y h = e k x [ ( c 1 + c 2 ) cos ω x + i ( c 1 - c 2 ) sin ω x ]
(3.13)
= e k x ( α 1 cos ω x + α 2 sin ω x )
ω is called the natural frequency of the equation. The solution (3.12),
can be rewritten in “phase-amplitude” form. To this end we multiply and divide (3.12) by A = α 1
2 + α 2 2 . We obtain
(3.14)
y h = A e k x [ β 1 cos ω x + β 2 sin ω x ]
where β 1 = α 1 A and β 2 = α 2 A . Since β 1 2 + β 2 2 = 1 , we can find an angle φ so that
(3.15)
cos ϕ = α 1 A
and rewrite (Equation 3.14) as
(3.16)
y h = A e k x cos ( ω x - ϕ ) .
A is the amplitude and φ is the phase of the solution.

A Particular Solution
If the general solution of the homogeneous (Equation 3.2) (with f(x) = 0) is known, then a particular
solution to (Equation 10.1) (with f(x)) ≠ 0 c a n be found by “variation of coefficients.” The
following (short) table summarizes the trial function one has to use to find y p for some important
cases which appear in applications (this table can be extended). The coefficients A, B, and etc. that
appear on the right hand side of the table have to be determined by substituting the trial function in the
differential equation. As a result, one obtains a system of algebraic equations which has to be solved
for the coefficients.
Table 3.1 Finding y p
f(x) yp
ae kx sin β x , cos β x Ae kx A cos β x + B sin β x
a n x n + a n-1 x n-1 + … + a 0 Anxn+ An- 1Xn- 1+ … + A0′
An exception to this table happens when the forcing function f(x) is one of the solutions of the
corresponding homogeneous equation (this is referred to as “resonance”.
Remark
Observe that in general y p will not satisfy the initial conditions. It is only y G that has to satisfy these
conditions.
We illustrate the use of this table by examples.
Example 3.1.3
f(x) = h = constant in Equation ( 3.2 ).
In this case, y p is given by
(3.17)
y p = h c c ≠ 0 h x b + γ 1 b ≠ 0 , c = 0 h x 2 2 a + γ 1 x + γ 2 b = c = 0 ( 3.17 )
where γ 1 and γ 2 are arbitrary constants.
Example 3.1.4
Consider Equation ( 3.2 ) with b = 0, a > 0, c > 0 and
(3.18)
f ( x ) = b 1 cos ω x + b 2 sin ω x .
The general solution of the homogeneous equation is
y h = C 1 cos ν x + C 2 sin ν x
where ν = c a is referred to as the natural frequency of the equation (the frequency of the
oscillations when the forcing function f(x) = 0). If ω ≠ ν, then to find y p we try
(3.19)
y p = A cos ω x + B cos ω x .
Substituting this in Equation ( 3.2 ) we obtain an algebraic equation for A and B
(3.20)
sin ω x [ - a A ω 2 + c A - b 1 ] + cos ω x [ - a B ω 2 + c B - b 2 ] = 0 .
However, since cos ω x and sin ω x are independent functions, we infer that to satisfy this equation
each of the brackets in Equation ( 3.20 ) must be zero. Hence, we obtain the following system of
equations for A and B
(3.21)
-aω2+c00-aω2+cAB=b1b2
since ω is not the natural frequency of the system ( - aω 2 + c) ≠ 0 Equation ( 3.21 ) yields
nontrivial solutions for A and B.
Example 3.1.5
Consider (3.2) with b = 0, a > 0, c > 0 and f(x) = b 1 cos ω x + b 2 sin ω x but with ω = ν.
We refer to this situation where the forcing function has frequency equal to the natural frequency of the
system as “resonance.” (The forcing function f(x) is one of the solutions of the homogeneous equation)
In this case, we use for y p a trial solution of the form
(3.22)
y p = x ( A cos ω x + B sin ω x ) .
Substituting this expression in (Equation 3.2) we obtain
cos ω x [ ( 2 a ω B - b 1 ) + ( - a ω 2 + c ) x A ]
(3.23)
sin ω x [ ( - 2 a ω A - b 2 ) + ( - a ω 2 + c ) x B ] = 0 .
Since cos ω x and sin ω x are independent functions, each bracket in (Equation 3.23) must be zero.
Moreover, since ω is the natural frequency of the system - aω 2 + c = 0,
(3.24)
2aωB-b1=0,2aωA+b2=0.
We obtain the following solution for A and B.
A=-b22aω,B=b12aω

3.2 Review of Linear Algebra

In this section, we review some topics from linear algebra regarding eigenvalues and eigenvectors
that are needed fro the solution of systems of differential equations.

3.2.1 Eigenvalues and Eigenvectors

Definition 3.2.1
Let A be a n × n matrix with constant entries. The characteristic polynomial of A is
(3.25)
p(λ)=|A-λI|
where I is the n × n unit matrix (that is the matrix with 1 along the main diagonal and zero at all
other entries).
Definition 3.2.2
The roots of the characteristic polynomial are called the eigenvalues of A.
Thus, when λ is an eigenvalue of A, the matrix ( A - λ I ) is singular. When λ is not an eigenvalue, ( A
- λ I ) is not singular and therefore has an inverse. Now consider the following system of linear
equations
(3.26)
Ax=λx.
This system can be written also as
(3.27)
(A-λI)x=0.
It follows then that when λ is not an eigenvalue of A, then ( A - λ I ) is invertible. Therefore, the only
solution of this system is the trivial solution x = 0. However, if λ is an eigenvalue of A, the matrix ( A
- λ I ) is singular and there is a nontrivial solution of the system, (Equation 3.26). That is there exists
y ≠ 0 that satisfies this equation. We refer to the vector y as the eigenvector of A that is related to the
eigenvalue λ .
Corollary 3.2.1
If y is an eigenvector of A, then cv for all c ≠ 0 is also an eigenvector of A (with the same
eigenvalue λ ).
Therefore, an eigenvector is determined only up to a multiplicative constant.
We note that each eigenvalue has a related eigenvector. However, the number of independent
eigenvectors related to a given eigenvalue does not have to equal to its algebraic multiplicity (i.e., the
number of times it appears as a root of p ( λ ) ).
Example 3.2.1
Find the eigenvalues and eigenvectors of
A=2100-21002
Solution 3.2.1.
The characteristic polynomial of A is
(3.28)
p(λ)=2-λ100-2-λ1002-λ=(2-λ)2(-2-λ).
Hence, λ = - 2 is a simple root (i.e., multiplicity 1) and λ = 2 is a double root.
To find the corresponding eigenvectors, we solve (Equation 3.26) with each of these eigenvalues.
λ=-2
If v = x 1 x 2 x 3 , we obtain from the definition the system
2x1+x2=- 2x1
2x2+x3=- 2x2
2x3=- 2x3.

Hence, up to a multiplicative constant which can be expected in view of corollary 3.2.1,

v=1-40
\lambda = 2
Denoting once again the required eigenvector by v = x 1 x 2 x 3 , we obtain the system
2x1+x2=2x1
2x2+x3=2x2
2x3=2x3.
Hence, v = 1 0 0 . Thus, there exists only one independent eigenvector related to λ = 2 although the
multiplicity of this eigenvalue is 2.
Theorem 3.2.1
If n×n matrix A has n independent eigenvectors v 1,...,v n , then
(3.29)
M-1AM=λ1O⋱ Oλn
where M = ( v 1 , … , v n ) . Such a matrix A is said to be diagonalizable.
Example 3.2.2
Verify theorem 3.2.1 for
A=1100-2000-2
Solution 3.2.2.
The eigenvalues of A are λ = 1 and λ = - 2 (with multiplicity 2). However, contrary to example
3.2.1, A has three independent eigenvectors
v 1 = 1 0 0 , v 2 = - 1 3 0 and v 3 = 0 0 1
Hence,
M=1-10030001,M-1=11300130101
and a direct multiplication shows that
M-1AM=1000-2000-2
Exercises
1. Find the relationship between the eigenvalues of A and those of A -1 and A T (where A T is the
 transpose of A). What about the eigenvectors?
2. Let p(x) be a polynomial. Show that the eigenvalues of p(A) are given by p(λ i ),where λ i
are the eigenvalues of A. What can you say about the eigenvectors of p(A)?
Hint: Show that if v is an eigenvector of A, then it is also an eigenvector of A m .
3. Find the eigenvalues and eigenvectors of the following matrices:
a.A=1-2-43,b.A=1-2-24

c.A=210-1-2-1012,d.A=111000-1-23

3.3 Reformulation Of Systems Odes

In the following sections, we consider systems of ordinary differential equations and their solutions.
To standardize this treatment, we first show through examples how a coupled system of differential
equations of any order can always be rewritten as a system of first order equations.
Example 3.3.1
Consider the nth order equation
(3.30)
dnydxn=f(x,y,dydx,…,dn-1ydxn-1)
with the initial conditions
(3.31)
y(a)=b0,y′(a)=b1,…,y(n-1)(a)=bn-1.
This equation and the initial conditions are obviously equivalent to the system
dydx=y1
dy1dx=y2[≡ d2ydx2]
dyn-2dx=yn-1[≡ dn-1ydxn-1]
(3.32)
dyn-1dx=f(x,y,y1,…,yn-1)
with the initial conditions
(3.33)
y(a)=b0,y1(a)=b1,…,y(n-1)(a)=bn-1.
Example 3.3.2
Rewrite the system
x″ +y′+xy=0
(3.34)
y″ +x′y′+2x2y2=0
(3.35)
x(0)=0,x′(0)=1,y(0)=1,y′(0)=2
where primes denote differentiation with respect to t, as a system of first order equations.
Solution 3.3.1.
 He required system is
x′=x1,y′=y1
x1′=-y1-xy
y1′=-x1y1-2x2y2
(3.36)
x(0)=0,x1(0)=1,y(0)=1,y1(0)=2.
Example 3.3.3
Rewrite the system
(3.37)
x″ -2y′+x-y=0
(3.38)
x′-y′+2x+y=0
as a system of first order equations.
Solution 3.3.2. By introducing x ’ = z, the system becomes
(3.39)
z′-2y′+x-y=0
(3.40)
z-y′+2x+y=0.
Using (Equation 3.40) to substitute for y′ in (Equation 3.39), we finally obtain the system
(3.41)
x′=z
(3.42)
z′=3x+2z+3y
(3.43)
y′=2x+z+y.
Exercises
1. Rewrite the following as systems of first order ordinary differential equations:
(a) x ’’ + y′ - 3xy = t
y″ +x′+2x-3y=e2t
x(1)=1,x′(1)=0,y(1)=-2,y′(1)=0
(b) x+ 5y′ + xy 2 = t 3
y″ -3x″ y′-2y=0
x ( 0 ) = 1 , x ′ ( 0 ) = 0 , x ″ ( 0 ) = 6 , y ( 0 ) = 10
y′(0)=1.

3.4 Linear Systems With Constant Coefficients

In this section, we describe methods to solve the n‐dimensional system
 (3.44)
dxdt=Ax+f(t),
with initial conditions
(3.45)
x(a)=C,
where A is a matrix with constant entries
A=(aij),x(t)=x1(t)⋮ xn(t),f(t)=f1(t)⋮ fn(t).
If A has n independent eigenvectors, then x G can be found by decoupling the system to a set of n
independent equations through a change of variables.
Thus, if
M=(v1,…,vn),
where v i , i = 1,...n are the eigenvectors of A, we define
(3.46)
u=M-1X.
Substituting (Equation 3.46) in (Equation 3.44), we obtain
Mdudt=AMu+f(t).
Multiplying both sides of this equation by M -1 leads to
(3.47)
dudt=M-1AMu+M-1f(t),u(a)=M-1C
but
M-1AM=D=λ1O⋱ Oλn
and, therefore, the system (Equation 3.47) is a set of n uncoupled equations. Each equation in this
system can be solved independently of the others.
Example 3.4.1
Solve the system, Equation (3.44) if
A = 3 2 4 2 0 2 4 2 3 , f = 1 0 1 and x ( 0 ) = 1 - 1 - 1
Solution 3.4.1. The eigenvalues of A are λ 1 = - 1 , λ 2 = - 1 , and λ 3 = 8 . The corresponding
eigenvectors are
v1=-120,v2=-101,v3=212.
Hence,
M=-1-12201012,M-1=19-14-1-4-25212
and from Equation ( 3.47 ) we infer that the original system is equivalent to
ddtu1u2u3=-1000-10008u1u2u3+19-214,u(0)=19-4-7-1
The solution of this system is
u 1 ( t ) = - 2 9 ( 1 + e - t ) , u 2 ( t ) = 1 9 ( - 8 e - t + 1 ) , u 3 = - 1 18 ( 1 + e 8 t ) .
Finally, we express the solution in terms of the original variables by using equation ( 3.46 ) which
yields
x1=-u1-u2+2u3,x2=2u1+u3,x3=u2+2u3.
Another way to solve the system, (Equation 3.44), is to note that this is a linear system of equations.
Therefore, its general solution x G is given by
xG=xh+xp
where x h is the general solution of the homogeneous part of (Equation 3.44) (that is the system,
(Equation 3.44), with f = 0) and x p is a solution of (Equation 3.44).
Theorem 3.4.1
If λ is an eigenvalue of A, with a corresponding eigenvector y, then x = c v e λ t (where c is an
arbitrary constant) is a solution of the homogeneous system
(3.48)
X˙=Ax.
Proof
Substituting x = c v e λ t on the left‐hand side of (Equation 3.48), we obtain
(3.49)
X=cλveλt.
Substituting x = c v e λ t on the right‐hand side of (Equation 3.48) yields
(3.50)
Ax=A(cveλt)=ceλt(Av)=ceλt(λv).
Hence, x = c v e λ t is a solution of (Equation 3.48).
Since (Equation 3.48) is a homogeneous system of equations, the superposition principle holds.
Therefore, if all the eigenvalues of A are simple, then the general solution of (Equation 3.48) is given
by
(3.51)
xh=∑i=1ncivieλit.
To find a particular solution of the system (Equation 3.44) when f ≠ 0, we make the ansatz that
yp=Nu.
where N is the matrix
N=(x1,…,xn),
with x i = v i e λ i t . This matrix is called the Fundamental matrix of the
system, (Equation 3.44).
Since d d t x i = A x i (see (Equation 3.48)), it follows that
dNdt=(ddtx1,…,ddtxn)=AN.
Therefore, the left hand side of (Equation 3.44) becomes
dypdt=Ndudt+dNdtu=Ndudt+ANu
while the right hand side of (Equation 3.44) yields
Ayp+f=ANu+f.
Hence,
Ndudt=f,
i.e.
dudt=N-1f.
Integrating this equation we obtain
u(t)=∫N-1f(t)dt.
Thus,
yp=N∫N-1f(t)dt.
Exercises
Solve the following systems of equations:
1.  x ˙ = x + 2 y - 1
y=4x+3y+( t- 1)
x1( 0) =- 1,x2( 0) =1
x¨ +3x˙ - 4x=e2t
x( 0) =1,x˙ ( 0) =0
2.  x ¨ = 2 x + 5 y - 2 t
y=x˙ - 2y- 1
x( 0) =- 1,x˙ ( 0) =0,y( 0) =1
4.  ẍ + 2 x ˙ + 3 x = y
y¨+2y˙-4y=2x
x(0) = y(0) = 0, x ˙ ( 0 ) = 1 , y ˙ ( 0 ) = 2 .

3.5 Numerical Solution Of Initial Value Problems

To explore the ideas underlining the numerical solution of initial value problems, we consider the
following example.
Example 3.5.1
Find the value of solution to the following differential equation at x = 0.2 and x = 1:
(3.52)
y ′ = x 2 + 2 y , y ( 0 ) = 0.5 .
Solution 3.5.1. Assuming y(x) is analytic, it can be expanded in a Taylor series around x = 0:
(3.53)
y(x)=y(0)+y′(0)x+y″ (0)2!x2+⋯ .
Thus, inside the radius of convergence of this series, we can evaluate y(x) if we can compute { y ( k )
( 0 ) } k = 0 ∞ . To see how this can be carried out, we first note that y(0) = 0.5 is given. To compute
y′(0) , we use the differential equation
y′(0)=02+2y(0)=1.
To compute y′ ’(0) , y′ ’’(0) , etc., we differentiate the given differential equation
y″ (x)=2x+2y′(x)
y′′′(x)=2+2y″ (x)
y(4)(x)=2y′′′(x)
and in general
(3.54)
y(n)=2y(n-1),n≥3.
This implies that
y′ ’(0) = 2, y′ ’(0) = 6, y (4) = 12, etc.
 Thus,
y ( 0.2 ) = 0.5 + 0.2 + 0 . 2 2 + 0 . 2 3 + 1 2 0 . 2 4 + …
and we can obtain “reasonable” approximation for y(0.2) with a small number of terms in the Taylor
expansion. The situation at x = 1 is the same in principle but a larger number of terms in the Taylor
expansion is needed to obtain the same accuracy. For this particular differential equation, this is not a
problem since we have a recursive formula for y (n)(x) . However, in general, such a recursive
formula is not available and this implies that a large number of symbolic computations is necessary to
obtain a reasonable approximation for the value of the solution.

3.5.1 Euler Algorithm

From a numerical point of view, we want to utilize only the differential equation and the initial
conditions in the computation of the solution. This implies that we must circumvent the need (as
described above) for the computation of the higher order derivatives.
To see how this is done, consider the equation
(3.55)
y′=f(x,y),y(0)=y0
where the solution is desired on the interval [ a , b ] .
To carry this out, we divide the interval [ a , b ] into n (“small”) sub‐intervals of fixed length
(≡step size ) h. We note, however, that under exceptional circumstances (e.g., near a resonance where
the solution varies rapidly) a variable step size has to be used. We shall say that the differential
equation has been solved numerically if we can compute y k = y(X) at the grid points x k = a + kh,
k = 0,....,n (see Fig. 3.1). Obviously, if the value of the solution is needed at any other point x ε [ a , b
] , then interpolation can be used.

Figure 3.1 Computational grid with varying step size

To compute y 1 = y(x 1), we now employ a first order Taylor expansion around x 0 This leads to
(3.56)
y1≅ y(x0)+hy′(x0)=y0+hf(x0,y(x0))=y0+hf(x0,y0).
We now use this (approximate) value of y(x 1) as an initial value for the differential equation on the
interval [ x 1 , x 2 ] . Using the same strategy as before, we obtain
(3.57)
y2=y(x1)+hy′(x1)=y1+hf(x1,y1).
Continuing in this manner, we derive the general formula
(3.58)
yk+1=yk+hf(xk,yk),k=1,…,n-1.
This is known as the Euler algorithm for the numerical solution of differential equations.
Example 3.5.2
Use the Euler algorithm to solve
(3.59)
y′=x+y2,y(0)=0
on the interval [0, 1 ] with step size of 1 3 (see Fig. 2.2). Use linear interpolation to find an
approximate value for y(0.5)

Figure 3.2 Computational grid for example 3.4.2

Solution 3.5.2. Using Equation (3.58) we have

y1=0+13(0+02)=0
y2=0+13(13+02)=19
y 3 = 1 9 + 1 3 ( 2 3 + ( 1 9 ) 2 ) = 82 243 .
Table 3.2 Solution of example 2
x y
0 0
1/3 0
2/3 1/9
1 82/243
Usually, this is presented in table form (Table 3.2).
To compute y(0.5) , we observe that x = 0.5 is the midpoint of the interval [ 1 3 , 2 3 ] . Therefore,
if linear interpolation is used, then y(0.5) is the average of y 1 and y 2 (i.e., y ( 0.5 ) = 1 18 ) .
 It is appropriate at this point to ask about the accuracy of this scheme. It is clear that at each point
in the computation we neglected terms of order h 2 or higher in the Taylor expansion. Since there are n
steps, we can estimate the cumulative error to be of order nh 2. However, n = b - a h . We deduce then
that the cumulative error is order h. Thus, the error depends strongly on the step size chosen.
Moreover, this implies that the accuracy of the solution usually improves as h becomes smaller. In
fact, it is common practice to gauge the accuracy of the solution by applying the algorithm with step
sizes h and h 2 .
If the two solutions agree with each other on the common grid points up to, let’s say, three digits,
then we say that this is the accuracy of the numerical solution.
Remark
Observe that as h becomes smaller, the number of arithmetic operation needed to compute the solution
over the whole interval increases. As a result, the round‐off error in these computations increases.
Example 3.5.3
Solve the differential equation
(3.60)
y′=2x-y,y(0)=1
with step sizes of 1 3 and 1 6 and compare with the exact solution.
Solution 3.5.3. The differential equation
(3.61)
y′+y=2x
is a first order linear inhomogeneous equation. To solve it we must find the general solution of the
homogeneous equation
(3.62)
y′+y=0
and a particular solution of Equation ( 3.61 ). The general solution of Equation ( 3.62 ) is
(3.63)
y=Ce-x.
We now try to find a particular solution in the form
(3.64)
yp=u(x)e-x
(variations of parameters). Substituting (Equation 3.64) in (Equation 3.61) yields
u′(x)=2xex
Hence,
u(x)=(2x-2)ex
and therefore
y(x)=Ce-x+2(x-1).
To determine the constant C, we use the initial condition
y(0)=1=C-2,
i.e., C = 3.
A tabulation of this solution and the numerical solution using the Euler algorithm is given in Table
3.3.
Table 3.3 Solution of y′ + y = 2x using the Euler method:

Table 3.4 Solution of y′ + y = 2x with modified Euler method

We now show how this algorithm can be extended to systems of equations. To this end, we
consider the system
(3.65)
y′=F(x,y),y(0)=c
where
(3.66)
y=y1⋮ yn,F=f1⋮ fn
and f i = f i ( x , y 1 , … , y n ) on [ a , b ] .
As before, we subdivide the interval [ a , b ] into sub‐intervals of step size h and apply a first
order Taylor expansion of y i on each of the sub‐intervals [ x k , x k + 1 ] :
(3.67)
yi(xk+1)=yi(xk)+hyi(xk)=yi(xk)+hfi(xk,yk1,…,ykm),i=1,…,m.
This can be rewritten transparently using vector notation as
(3.68)
yk+1=y(xk+1)=yk+hF(xk,yk).
This is the vector version of (Equation 3.58).
Example 3.5.4
Solve the following system numerically by using Euler algorithm:
u′=u+2v
v′=u-v,u(0)=0,v(0)=1
on [0, 1] with h = 0.25.
Solution 3.5.4 Applying Equation (3.68), we infer that
uk+1=u(xk+1)=uk+h(uk+2vk)
vk+1=v(xk+1)=vk+h(uk-vk).
Applying this recursively, we obtain
u 1 = u ( 0.25 ) = 0 + 0.25 ( 0 + 2.1 ) = 0.5
v 1 = v ( 0.25 ) = 1 + 0.25 ( 0 - 1 ) = 0.75
u 2 = u ( 0.5 ) = 0.5 + 0.25 ( 0.5 + 2 · 0.75 ) = 1
v 2 = v ( 0.5 ) = 0.75 + 0.25 ( 0.5 - 0.75 ) = 0.6875
u 3 = u ( 0.75 ) = 1 + 0.25 ( 1 + 2 · 0.6875 ) = 1.594
v 3 = v ( 0.75 ) = 0.6875 + 0.25 ( 1 - 0.6875 ) = 0.766
u 4 = u ( 1.0 ) = 1.594 + 0.25 ( 1.594 + 2 · 0.766 ) = 2.376
v 4 = v ( 1.0 ) = 0.766 + 0.25 ( 1.594 - 0.766 ) = 0.973
Exercises
1. Solve the following equations numerically by using the Euler algorithm and compare with
the exact solution. Graph your results.
y ′ = 2 x + y o n [ 0 , 1 ] , y ( 0 ) = 1 , h = 0.2
y ′ = y - y 2 o n [ 0 , 1 ] , y ( O ) = - 1 , h = 0.2
y′ = xy —l on[1, 2], y(1) = 0.5, h = 0.2
y ′ = ( 2 + y ) y - x o n [ 0 , 1 ] , y ( 0 ) = 0 , h = 0.2
y ′ = y sin x on [ 0 , π ] , y(O) = 1, h = π 4 .
2. Solve the equations in Ex. 1 with h 2 and compare with those obtained previously.
3. Solve the equations in Ex. 1 using the Taylor series expansions. Estimate the number of
terms needed to obtain an accuracy of 10−3.
4. Solve
y ′ = sin x + y 2 , y ( 0 ) = 1

on [0, 1 ] with h = 0.25 and use interpolation to find y ( 2 3 ) . Next, solve this equation with
h = 1 3 and compare the two values obtained for y ( 2 3 ) .
5. Use a computer to solve the equation in Ex. 1 with h = 0.01 and h = 0.001 and compare
with the exact solution.
6. Solve the equation in Ex. 4 with h = 0.1 and h = 0.01 to find a better approximation for y ( 2
3).
7. Solve the following systems numerically by using the Euler algorithm
with the initial conditions u(0) = 0, v(0) = - 1 on [0,1] and h = 0.2. Can you solve these
systems analytically?
Hint: Eliminate v from the second equation by using the first equation.
u ’ = u + 2v, v ’ = 4u - 3v
u ’ = u + v, v ’ = u - v + 2x
8. Use the “extended Euler method” where
 to solve
y′ = x 2 + 2y, y(0) = 1, on [0,1]
with h = 0.2. Compare with the solution obtained from the Euler algorithm with the same
step size and the exact solution.
yk+1=yk+hyk′+h22yk″
9. Estimate the cumulative error in the “Extended Euler Method” in terms of h.
3.6 Finite Difference Approximations
Before discussing improvements to the Euler algorithm and the solution of boundary value problems,
we must first introduce the notation of O(h k ) (pronounced as “O‐big” of h k ). Moreover, we shall
derive finite difference approximations to the derivatives of a function f(x) in terms of its values at
some set of discrete points.
Definition 3.6.1
Let f(x) and g(x) be two functions which are continuous at x = 0.
We say that f is O(g) if
(3.69)
0 < | lim x → 0 f ( x ) g ( x ) | < ∞ .
Example 3.6.1
sin x = O ( x ) since
(3.70)
lim x → 0 sin x x = 1 .
Example 3.6.2
f = 104 x 3 + 3x 2 + x is O(x 3) since
(3.71)
lim x → 0 f ( x ) x 3 = 10 4
Example 3.6.3
x 4 + x 5 is not O(x 3) since
(3.72)
lim x → 0 x 4 + x 5 x 3 = 0 .
Example 3.6.4
The remainder of the Taylor expansion of f(x + h) around x
(3.73)
Rn(h)=f(x+h)-∑k=0nf(n)(x)hkk!
is O(h) if f (n+1)(x) ≠ 0.
In fact, the remainder of the Taylor expansion is
(3.74)
Rn(h)=∑k=n+1∞f(k)(x)hkk!=hn+1{f(n+1)(x)(n+1)!+hf(n+2)(x)(
n+2)!+…}
(3.75)
lim h → 0 R n ( h ) h n + 1 = f ( n + 1 ) ( x ) ( n + 1 ) ! .
Therefore, we can write
(3.76)
f(x+h)=∑k=0nf(k)(x)hk
From the definition we now infer the following:
Theorem 3.6.1
If f is O(h n ) , g is O(h m ) , and m ≤ n then
1. f ± g is O(h m )
2. f ⋅ g is O(h n+m )
3. f/g is O(h n-m ) .
Example 3.6.5
The local error in the Euler algorithm (i.e., the error in each step) is O(h 2) while the cumulative
(global) error is O(h) .
In fact, the basic approximation in the Euler method is
(3.77)
yk+1=yk+hyk′ +O(h2).
Since there are n steps and n = b - a h , the global error is 1 h O ( h 2 ) = O ( h ) .
The notation of O‐big is used to give the user a rough idea about the behavior of the error committed
by an algorithm as h → 0. This helps to compare between different algorithms and determines which
is superior.
We now introduce finite difference approximations.
If a symbolic representation of a function f(x) is not known, then we consider the function to be
“known numerically” on a domain D if its values on a grid of points in this domain are known and if
by using these values we can calculate an acceptable approximation to the value of u at any other
point in D (by interpolation). It follows, then, that in order to solve a differential equation on D, we
have to introduce a grid of points on this domain and cotruct an algorithm to calculate the values of
the unknown function u at these points. Thus, if (in one dimension) x i , i = 1,...,m, are the grid points,
then the fundamental unknowns that have to be computed are
(3.78)
(fi,Xi)
Since the differential equation and the initial (or boundary) conditions at hand are the only means to
compute these quantities, we must develop an approximation scheme for the derivatives of f in terms
of these data. To see how this can be done, consider the one‐dimensional case where, to begin with,
we assume that the x i ′ s are equispaced (See Fig. 3.3):
xi-xi-1=h
(where h is constant for all values of i).
Thus, suppose that we are given ( x i , f i ) , i = 0, \ldots , n, where x i+1 - x j = h = const. To find an
approximation for f i ′ = f ′ ( x i ) , we use the Taylor expansion of f around x i :
(3.79)
fi+1=f(xi+h)=fi+hfi′+h22fi″ +h33!fi′′′+h44!fi(4)+….
Hence,
(3.80)
fi′=fi+1-fih-h2fi″ -h23!fi′′′-…
or
(3.81)
fi′ =fi+1-fih+O(h).
This formula is called the forward difference approximation for f i ′ (as the point x i+1 is used).
Similarly, we can derive a backward difference approximation for f i ′ using
 (3.82)
fi-1=f(xi-h)=fi-hfi′+h22!fi″ -h33!fi′′′+h44!fi(4)….
This leads to
(3.83)
fi′ =fi-fi-1h+O(h).
In both of these approximations, the error is O(h) . However, we can obtain a better approximation by
using “central difference” (i.e., use an interval where x i is at the center). Thus, by subtracting
(Equation 3.82) from (Equation 3.79) we obtain
fi′=fi+1-fi-12h+2h23!fi′′′+25!h5fi(5)+…
or
(3.84)
fi′ =fi+1-fi-12h+O(h2)
We note that although this is a superior approximation, it cannot be used at the ends of the
computational interval ( i . e . , a t x 0 a n d x n ) .
Similar approximation can be derived for the second order derivatives. In fact, adding (Equation
3.79) and (Equation 3.82) leads to
(3.85)
fi″ =fi+1-2fi+fi-1h2+O(h2)
This is a central difference formula. A forward difference formula for f i ″ can be derived by using f
i+2. In fact,
(3.86)
fi+2=f(x1+2h)=fi+2hfi′+2h2fi″ +4h33fi″ +….
Subtracting (Equation 3.79) multiplied by 2 from (Equation 3.86) leads to
(3.87)
fi″ =fi+2-2fi+1+fih2+O(h).
Finite difference approximation for higher order derivatives can be obtained in a similar fashion. For
instance, we have the following central difference formula for the 4th derivative
(3.88)
fi(4)=fi+2-4fi+1+6fi-4fi-1+fi-2h4+O(h2).
Figure 3.3 Equispaced computational grid

Example 3.6.6
Suppose that the values of f at x 0, x 0 + h, and x 0 + 3h are given. Develop a finite difference
approximation for f ’(X 0) with O(h 2) error.
Solution 3.6.1. From the Taylor expansion around x 0 we have
f(x0+h)=f(x0)+hf′(x0)+h22!f″ (x0)+h33!f′′′(x0)+…
f ( x 0 + 3 h ) = f ( x 0 ) 3 h f ′ ( x 0 ) + 9 h 2 2 ! f ″ ( x 0 ) + 27 3 ! f ′ ′ ′ ( x 0 ) + …
To obtain an O(h 2) approximation for f ’(x 0), we must find a combination of f(x 0 + h) and f(x
2
0 + 3h) which cancels the h terms in these expansions. Thus,
9f(x0+h)-f(x0+3h)=8f(x0)+6hf′(x0)-3h3f′′′(x0)+…,
i.e.
f′ (x0)=9f(x0+h)-f(x0+3h)-8f(x0)h+O(h2).
When the grid points are not equispaced, the general technique described above can be easily adapted
to derive appropriate approximations. Thus, if
xi-xi-1=hi,
then
(3.89)
fi+1=f(xi+hi+1)=f(xi)+hi+1f′ (xi)+O(h2)
fi-1=f(xi-hi)=f(xi)-hif′ (xi)+O(h2),
which yields
(3.90)
fi′ =f′ (xi)=fi+1-fi-1hi+hi+1+O(h).
Similarly, for the second‐order derivative,
(3.91)
fi″ =2hifi+1+hi+1fi-1-(hi+hi+1)fi(hihi+1)(hi+hi+1)+O(h).
From (Equations 3.90) and (3.91), we infer that in general the accuracy of our difference
approximations is better with an equispaced grid. (Equations 3.90) and (3.91) or their generalizations
to higher dimensions are, therefore, useful near irregular boundaries where their use is mandatory.

3.6.1 Extension to Higher Dimensions

The formulas derived above in one dimension have a natural extension to higher dimensions. For
example, for an equispaced grid in two dimensions with step size h, we have (see Fig. 3.4)
(3.92)
ui-1,j=u(xi-h,yj)=u(xi,yj)-h(∂u∂x)ij+h22(∂2u∂x2)ij
+O(h3)
ui+1,j=u(xi+h,yj)=u(xi,yj)+h(∂u∂x)ij+h22(∂2u∂x2)ij+O(h3).

Figure 3.4 Two dimensional computational grid around u i,j

Therefore,
(3.93)
(∂u∂x)ij=ui+1,j-ui-1,j2h+O(h2)
(∂2u∂x2)ij=ui+1,j+ui-1,j-2uijh2+O(h2).
The corresponding formulas for ( ∂ u/ ∂ y) ij and ( ∂ 2 u/ ∂ y 2) ij should be obvious.
Exercises
1. Derive a backward finite difference formula for f i ’’
2. Prove that
f i ′ = - f i + 1 + 8 f i + 1 - 8 f i - 1 + f i - 2 12 h + O ( h 4 ) .
3. Prove the finite difference formula for f i ( 4 ) which is given in (Equation 3.88).
4. Derive an approximation for f ’’(X) with O(h 4) error using f i+2, f i+1, f i , f i-1 and f i-2. (Use a
symbolic computation package if available).
5. Derive a forward difference approximation for f ’’’(X) with O(h) error using f i+3, f i+2, f i+1
and f i
6. Derive a central difference approximation for f ’’’(X) with O(h 2) error.
7. Derive a central difference formula for f ’(x i ) with O(h 2) error if the grid points are not
 equispaced, i.e., h i = x i+1 - x i are not constant.

3.7 Modified Euler And Runge-Kutta Methods

It is natural to inquire as to how one can improve the Euler method. One possibility which was
discussed already is to use smaller step size h. However as h becomes smaller, the corresponding
computational effort increases rapidly; furthermore, the accumulated numerical error (due to round‐
off errors) increases. The other options available are to use either higher order Taylor expansions (or
their equivalents) or “predictor-corrector” scheme. We illustrate both of these through the modified
Euler algorithm which is a prototype for the Runge‐Kutta class of algorithms.

3.7.1 Modified Euler Algorithm

To describe this algorithm we start with a geometrical argument.
The Euler algorithm to solve
(3.94)
y′=f(x,y),y(a)=y0
applies a linear approximation to y(x) at x 1 to compute y(x i+1) .
yi+1=yi+hf(xi,yi)
This linear approximation is based on the tangent to the graph of y(x) at x i whose slope is given by y i
′ = f ( x i , y i ) (see Fig. 3.5).

Figure 3.5 Euler’s Numerical Approximation

Since the function y(x) in general is not linear, it stands to reason that a better approximation for
y(X) can be obtained if an “averaged slope” of f(x) over [ x i , x i + 1 ] is used. An appealing way to
implement this will be to average the slopes at x i and x i+1, i.e.,
(3.95)
yi+1=yi+h2[yi′+yi+1′]
The only obstacle for the use of this formula is that
(3.96)
yi+1′=f(xi+1yi+1)
This implies that in order to compute y i + 1 ′ we must know y i+1 first!
To circumvent this difficulty we can utilize the Euler algorithm to obtain a predictor for the value
of y i+1
(3.97)
yi+1(p)=yi+hf(xi,yi).
This value can be used then in eq. (3.95) to obtain “corrected” value of y i+1
(3.98)
yi+1(c)=yi+h2[f(xi,yi)+f(xi+1,yi+1(p))]
Should the difference between y (p) and y (c) be larger than some preset error tolerance ɛ, we can use y
(c) as a new y (p) and iterate the process until

(3.99)
|yi+1(p)-yi+1(c)|<ε
(Obviously we must set a limit on the number of these iterations to avoid the possibility of an infinite
loop.)
We illustrate the implementation of this algorithm through the following example.
Example 3.7.1
Solve
(3.100)
y′=2x-y,y(0)=1
with h = 0.2 on [0, 1 ]. Use error tolerance of ɛ = 0.05 and one predictorcorrector iteration.
Solution 3.7.1. Using Equation (3.97), (3.98) we obtain Ta ble 3.4 for the solution of Equation (
3.100 ). To gauge the accuracy of the numerical solution we present in the last column of this table
the values of the exact solution
(3.101)
y=3e-x+(2x-2).
We demonstrate now that the modified Euler algorithm (without iterations) is equivalent to the use of
second order Taylor expansions for y(x) on each of the sub‐intervals [ x i , x i + 1 ] . In fact according
to this scheme
(3.102)
yi+1=yi+hyi′ +h22yi″ +O(h3).
Replacing y i ″ = ( y i ′ ) ′ by its forward finite difference approximation (eq. (3.79)).
(3.103)
yi″ =(yi′ )=yi+1′ -yi′ h+O(h)
leads to
(3.104)
yi+1=yi+h{yi′ +h2[yi+1′ -yi′ h+O(h)]}
(3.105)
+o(h3)=yi+h2{yi′ +yi+1′ }+O(h3).
This also shows that the local error in the modified Euler algorithm is O(h 3) . Consequently, this
leads to the following estimate for the cumulative (global) error
(3.106)
E=nO(h3)=b-ahO(h3)=O(h2).

3.7.2 Runge-Kutta Methods

The basic idea that underlie, these methods is similar to the modified Euler algorithm. They use
“sampling” of y(x) on [ x i , x i + 1 ] to obtain an approximation for y i+1 which is equivalent to the use
of a higher order Taylor expansion. Thus we speak of second, fourth, sixth or higher order Runge‐
Kutta methods according to the number of points used.
For example to determine a fourth order Runge‐Kutta algorithm we try a prescription of the form,
(3.107)
yi+1=yi+a1k1+a2k2+a3k3+a4k4
where
k1=hf(xi,yi)
k2=hf(xi+α1h,yi+α1k1)
k3=hf(xi+α2h,yi+α2k2)
(3.108)
k4=hf(xi+α3h,yi+α3k3).
The unknowns a i , α i in these equations have to be determined so that Equation. (3.107) is equivalent
to a fourth order Taylor expansion. It can be shown (after a long algebra) that this leads to eight
equations in seven unknowns. Thus the choice of the parameters in this integration scheme is
somewhat redundant. A classical solution is in the form
yi+1=yi+16(k1+2k2+2k3+k4)
k1=hf(xi,yi),k2=hf(xi+12h,yi+12k1)
k3=hf(xi+12h,yi+12k2),
(3.109)
k4=hf(xi+h,yi+12k3).
The freedom in the choice of the parameters of these algorithms (which exists for Runge‐Kutta
algorithms of any order) can be used to impose on the solution certain physical constraints. Thus if the
differential equation under consideration describes the trajectory of a system in which energy is
conserved, then the parameters of the Runge‐Kutta algorithm can be chosen so that this conservation
law is satisfied in each of the integration steps. (These are referred to as “symplectic algorithms”.)
For more detailed treatment of the Runge‐Kutta methods the reader is referred to the literature.
Exercises
Solve the equations in exercises 1–8 using the modified Euler algorithm with one predictor‐
corrector iteration
1. y′ = x + y 2, y( - 1) = 0 on [1, 1] with h = 0.2.
2. y′ = x 2 y - x, y(0) = 1 on [0, 1] with h = 0.2.
3. y′ = xy 2, y(0) = - 1 on [0, 1] with h = 0.2.
 4. y′ = x 2 - y 2, y(0) = 0.5 on [0, 2] with h = 0.2.
5. y ′ = x sin y , y(0) = 0 on [0, 1] with h = 0.2.
6. Solve the differential equations in exercises 1–5 using the modified Euler algorithm without
iterations but with h = 0.1.
7. Use Runge‐Kutta of order four to solve exercises 1–5.
8. The following system of equations model an ecosystem of fish and sharks. Use Runge‐Kutta
of order four to solve these equations and plot the solution
u ′ = u - u 2 - 0.1 u v u ( 0 ) = 2 , v ( 0 ) = 0.1
v ′ = - v + 0.1 u v 0 ≤ t ≤ 100

(we are using here normalized units for the populations).

Hint: Solve these equations using MATLAB or similar.
9. Generalize the modified Euler algorithm to a system of equations and use it to solve the
following:
u′=u+2v+k
v′=u- 3v

u(0) = 0, v(0) = 1 on [0, 1] with h = 0.2 and 0.1

10. Solve, using different numerical methods, the following system and compare the results
u′=u2- v2+x

v′=u+v

u(0) = 1, v(0) = 0 on [0, 2]. Use different step sizes.

11. Solve
y ′ + 1000 y = 0 y ( 0 ) = 1

on [0, 100 ] using different step sizes. Compare with the exact solution.
Remark 3.7.1
Problems like this where numerical solution diverges from the correct solution are called “stiff”

3.8 Boundary Value Problems

The treatment of boundary value problems is totally different from the one for initial value problems
which was presented in previous sections. The objective of this section is to present the basic ideas
for solving boundary value problems for linear ordinary differential equations. Two prominent
methods exist in the literature to solve these problems. These are the finite differences and the finite
element methods. Both techniques reduce the original boundary value problem to a system of linear
algebraic equations, but they are conceptually different from each other. In this section we consider
only the finite difference method. The finite element method will be introduced in the chapter on
variational principles. We present the application of finite difference method to boundary value
problems through examples.
Example 3.8.1
Solve the equation for the harmonic oscillator
(3.110)
y″ +ω2y=0,ω≠0,0≤x≤1
subject to the boundary conditions y(0) = y(1) = 0 with h = 1 4 .
Solution 3.8.1. To solve this problem with h = 1 4 we must compute
y 1 = y ( 0.25 ) , y 2 = y ( 0.5 ) , y 3 = y ( 0.75 )
( y 0 = y(0) and y 4 = y(1) are given). To derive a system of linear equations for these unknowns we
apply the central difference formula for y′ ’ (Equation (3.85)) to approximate the differential
equation. At x = 0.25 we have
y 2 + y 0 - 2 y 1 0 . 25 2 + ω 2 y 1 = 0
Similarly at x = 0.5 and x = 0.75 we have
y 3 + y 1 - 2 y 2 0 . 25 2 + ω 2 y 2 = 0
y 4 + y 2 - 2 y 3 0 . 25 2 + ω 2 y 3 = 0
Since y 0 = y 4 = 0, these equations represent a system of three equations in the three unknowns y
1, y 2, y 3. In matrix form we can rewrite these equations as
ω 2 - 32 16 0 16 ω 2 - 32 16 0 16 ω 2 - 32 y 1 y 2 y 3 = 0 0 0
This is a system of homogeneous equations, and a nontrivial solution (i.e., not all y i are zero)
exists only if the coefficients matrix A of this system is singular, viz. its determinant is zero. Since
det A = ( ω 2 - 32 ) [ ( ω 2 - 32 ) 2 - 2 × 16 2 ]
we infer that a nontrivial solution to our problem exists only if
(3.111)
ω=±42,±42-2,±42+2,
e.g. for ω = 4 2 we obtain
y1=λ,y2=0,y3=-λ,-∞<λ<∞
where λ is a parameter.
It is interesting to compare this solution with the analytic one. The general solution of (Equation
3.110) is
y = A cos ω x + B sin ω x
The first boundary condition then implies
y(0)=0=A
To satisfy the second we must have
y ( 1 ) = B sin ω = 0
The solution of this equation is B = 0 (trivial solution for y) unless
(3.112)
ω=n,n=1,2,…,.
Thus the analytic solution has infinite values of ω for which a nontrivial solution y = B sin n π x
exists.
 To see the relation between the numerical and analytic solution we point out that the values of ω in
(Equation 3.111) are
ω = ± 3.061 , ± 5.657 , ± 7.391
These represent approximations to the first three values of ω in (Equation 3.112). Observe that we
can ignore the negative values of ω since they lead to the same solution for y as the positive values.
We can expect, naturally, a better agreement between the numerical and analytic solutions as h
becomes smaller.
Example 3.8.2
Solve the following boundary value problem
y″ +xy=x,0≤x≤1
subject to the conditions y(O) = 0, y ( 1 ) = 1 16 with h = 1 4 .
Solution 3.8.2. As in the previous example we use Equation (3.85) to approximate the differential
equation for y 1 = y ( 1 4 ) , y 2 = y ( 1 2 ) and y 3 = y ( 3 4 ) . This leads to
y 2 + y 0 - 2 y 1 ( 1 4 ) 2 + 0.25 y 1 = 0.25
y 3 + y 1 - 2 y 2 ( 1 4 ) 2 + 0.5 y 2 = 0.5
y 4 + y 2 - 2 y 3 ( 1 4 ) 2 + 0.75 y 3 = 0.75
where y 0 = y(0), y 4 = y(1) .
Rewriting this in matrix form we have
- 31.75 16 0 16 - 31.50 16 0 16 - 31.25 y 1 y 2 y 3 = 0.25 0.5 - 0.25 .
Since the coefficient matrix of this system is non‐singular, the system has a unique solution
y 1 = - 0.024 , y 2 = - 0.033 , y 3 = - 0.009
Example 3.8.3
Solve
(3.113)
u″ +(1+x2)u′+u=x,u(0)=0,u(1)=1
on [0, 1 ) with step size h = 0.25.
Solution 3.8.3. Since h = 0.25, the numerical solution of this problem is equivalent to the
computation of the three unknowns (see Fig. 3.6)
Figure 3.6 Computational grid for example 3.7.3

u 1 = u ( 0.25 ) , u 2 = u ( 0.5 ) , u 3 = u ( 0.75 )

(Note that u 0 = u(0) = 0 and u 4 = u(1) = 1.)
Using Equation ( 3.113 ) and the finite difference formulas in Equations (3.84) and (3.85), we
obtain the following equations at x = 0.25, 0.5 and 0.75, respectively:
(3.114)
u 0 + u 2 - 2 u 1 h 2 + [ 1 + ( 0.25 ) 2 ] u 2 - u 0 2 h + u 1 = 0.25
u 1 + u 3 - 2 u 2 h 2 + [ 1 + ( 0.5 ) 2 ] u 3 - u 1 2 h + u 2 = 0.5
u 4 + u 2 - 2 u 3 h 2 + [ 1 + ( 0.75 ) 2 ] u 4 - u 2 2 h + u 3 = 0.75
Using the boundary conditions and rearranging these equations lead to
(3.115)
- 31 u 1 + 18.125 u 2 = 0.25
13.5 u 1 - 31 u 2 + 18.5 u 3 = 0.5
12.875 u 2 - 31 u 3 = - 18.375
The solution of this system to three decimals is
u 1 = 0.385 u 2 = 0.672 u 3 = 0.872
In the previous examples we showed how to solve problems where the boundary conditions are
imposed on the unknown function. (These are referred to as ”Dirichlet Boundary Conditions”).
However boundary conditions can be imposed on the derivatives as well. These are called
“Neumann Boundary Conditions.” Furthermore, boundary conditions can involve both the value of the
unknown function and it derivatives at the boundary. These are called “Robin Boundary Conditions”
or “Boundary Conditions of the third kind.” We demonstrate the treatment of such problems through
the following example.
Example 3.8.4
A spring‐mass system is governed by the following equation
y¨+2y+y=t0≤t≤1
subject to the boundary conditions y ˙ ( 0 ) = 1 , y(1) = 0. Solve this problem with h = 0.25.
Solution 3.8.4. Since y(0) is not given, the problem involves four unknowns.
yn=y(nh),n=0,1,2,3
(y 4 = y(1) = 0) (see Fig. 3.7). Approximating the differential equation at t n = nh using (Equations
3.83), (3.85) we obtain
(3.116)
y - 1 + y 1 - 2 y 0 ( 0.25 ) 2 + 2 y 1 - y - 1 2 · 0.25 + y 0 = 0
(3.117)
y 0 + y 2 - 2 y 1 0.25 ) 2 + 2 y 2 . - y 0 20.25 = y 1 = 0.25
(3.118)
y 1 + y 3 - 2 y 2 ( 0.25 ) 2 + 2 y 3 . - y 1 20.25 = y 2 = 0.5
(3.119)
y 2 + y 4 - 2 y 3 ( 0.25 ) 2 + 2 y 4 . - y 2 20.25 = y 3 = 0.75

Figure 3.7 Computational grid for example 3.7.4

where we had to add a fifth (fictitious) point y -1 = y( - 0.25) . (Equations 3.116), (3.119) form a
system of four equations in five unknown. To add one more equation we use (Equation 3.83) to
approximate the boundary condition y ˙ ( 0 ) = 1 .
This leads to
(3.120)
y 1 - y - 1 0.5 = 1 .
Together (Equations 3.116), (3.120) form a system of five equations in five unknowns whose solution
is
y 0 = 0.619 , y 1 = 0.412 , y 2 = - 0.255 , y 3 = - 0.123 .
Exercises
1. Solve the boundary value problem given by (Equation 3.113) with h = 10−1, 10−2. What can
be said about the coefficient matrix for the resulting system of equations?
2. Compare the exact and numerical solution of
u″ +2u′+u=xu( 0) =1,u( 1) =3

with h = 10−1, 10−2.

 CHAPTER 4

Stability Theory

CONTENTS
4.1 General Introduction
4.2 Two species model
4.2.1 Steady States
4.2.2 Stability Analysis
4.3 Basic concepts
4.4 Linearizable Dynamical Systems
4.5 Linearizable systems in two dimensions
4.6 Liapounov method
4.7 Periodic SOLUTIONS(LIMIT cycles)
4.1 General Introduction
In the 19th and early 20th centuries most mathematical models for physical phenomena used linear
differential equations which can be solved by analytic techniques. If a model turned out to be
nonlinear, then some methods were used to “linearize” the model to obtain approximate solutions.
With the rapid advancement of science and engineering and the advent of computers, it was found
advantageous to consider sophisticated models where nonlinearities and their impact cannot be
ignored. However while computers can generate “heaps” of numbers, it is not easy to extract insights
about the system evolution under different conditions from this data.
Stability theory addresses the issue of model nonlinearities by making the observation that in many
models an initial state of the system will evolve into a steady state (i.e., a state that is time
independent). It was found that for many practical purposes the transient states of the system (i.e.,
those that are time dependent) are of “marginal” importance while the steady states and their
properties are of utmost practical importance.
Since most mathematical models provide only an approximation to reality, a basic issue about the
steady states is their “response” to “deviations” or perturbations of the system from the steady state.
In other words if the system “somehow” deviates from the the steady state, will it return to it or”run
away from it As an example consider a car running smoothly on a highway then hits a pothole on the
road. This will cause the body of the car to vibrate and the question arises then whether these
oscillations will “damp out” and the car return to its “normal steady state”, or a breakdown will
occur and the car settle (after a short transient state) into a new steady state (i.e., the car is not
drivable and has to be repaired) which is “far” from the original state. In other words “how stable
was the original state of the car to perturbations” The anawer to this fundamental question is the
subject of stability theory.
From another point of view one has to remember that mathematical models contain parameters that
have to be evaluated experimentally and hence, are subject to measurement errors. As a result the
actual steady state of a given system might be somewhat different from the one predicted by the model
equations. The question arises again as to whether these “small deviations” are “benevolent” (i.e.
have little or no practical impact) or “fatal.”
Historically the study of stability theory was initiated by H. Poincare and A. M. Liaponouv.
Poincare was interested in the stability of the solar system and in particular whether the Earth orbit is
stable. That is whether the Earth eventually “fall” into the Sun or recede from it. (This is still an open
question.) A. M. Liaponouv considered the stability of various mechanical systems in particular under
the influence of gravity. Today stability theory is still an important and evolving branch of science and
engineering. In this chapter we give an exposition of the basic ideas of this theory.
To motivate our study and introduce some of the basic ideas we revisit the population model for the
number of fish in a pool (see Chapter 2 Sec. 2.4) and assume that the size N(t) of the population in
this (simplified) ecological model is given by
(4.1)
dNdt=(N-1)(N-3),N(O)=N0
where the quadratic term in N represents the competition for resources (including food).
At a steady state d N d t = 0 and hence Equation (4.1) has two steady (or equilibrium) states N = 1
and N = 3.
To study the stability of these steady states we note that Equation (4.1), though nonlinear, can be
integrated by elementary methods and we obtain
(4.2)
N(t)=(N0-3)e2t-3(N0-1)(N0-3)e2t-(N0-1).
From Equation (4.2) we infer that whenever N 0 ≠ 0 the population N(t) will approach the steady state
N = 1 as t → ∞ even if N 0 is very small or large. (However, N 0 must be positive) . Thus, a (small)
perturbation of this ecological system from N = 1 will cause its state to “come back” to this
equilibrium while similar deviations from N = 3 will cause the the population to “run away” from this
steady state and converge to N = 1 as t → ∞. We conclude, therefore, that the steady state N = 1 is
(asymptotically) stable while N = 3 is unstable.
To demonstrate these results graphically we plotted in Figure 4.1 the solutions to (4.2) with
different initial conditions.

Figure 4.1 Population evolution from different initial conditions

The objective of stability theory (and phase space methods) is to derive these results about the
stability of the steady states without solving analytically or numerically the model equations under
consideration. This is a crucial feature since most current mathematical models lead to systems of
nonlinear differential equations which cannot be solved analytically in closed form. Furthermore,
numerical algorithms for the solution of these equations consume inordinate amounts of CPU time and
in many cases do not converge or converge to the wrong solution.
The underlying idea of these techniques is to consider the model equations as an algebraic
relationship between the state variables and their derivatives. Thus Equation (4.1) is viewed as an
algebraic relation between N and dN/dt. To analyze the stability of the steady states we have only to
remind ourselves (from elementary calculus) that d N d t > 0 implies that N is increasing while d N d
t < 0 implies that N is decreasing with time. Thus in the “phase space” ( N , d N d t ) Equation (4.1) is
described by Figure 4.2. (Remember that N < 0 representsin this model the population size and
therefore N < 0 is meaningless.)
Figure 4.2 stability of the steady states N = 1, N = 3. Arrows indicate the evolution of perturbations
from these states

We see from this figure that if N < 3 then d N d t < 0 and hence N is a decreasing function. On the
other hand if N > 3 then d N d t > 0 and N will increase (these facts are indicated by arrows in the
diagram). This is characteristic behavior of unstable steady state at N = 3. Similarly for the steady
state N = 1, d N d t > 0 if N < 1 (N is increasing with time) and d N d t < 0 if 1 < N < 3 (N is
decreasing with time). Hence N = 1 is stable. We point out that these results about the steady states of
this model were obtained without reference to the actual solutions of the model Equation (4.1).
Exercises
1. Solve the following equations in order to determine whether each of the steady states is
stable or unstable.
2. d N d t = a N - b N 2 , a, b > 0, N > 0
3. d N d t = - a N + b N 2 , a, b > 0, N > 0
4. d N d t = a N + b N 2 , a, b > 0 - ∞ < N < ∞
5. Use phase space techniques to deduce the stability of the steady states in Ex 1.
6. Use phase space techniques to discuss the stability of the steady states for
7. d N d t = α N + β N 2 + γ N 3 , β 2 - 4aγ > 0.
8. d N d t = ( N - α ) ( N - β ) ( N - γ ) . What happens when α = β = γ or α = β.
9. 4. A model of two societies M, N where M exploits N was suggested by May and Noy‐
Meir. According to this model if M, N are the population sizes of the two populations
respectively, then
dNft=aN(1-N/A)-BMN21+N2
where a, A, B M are constants. In this equation the first term represents the natural growth of society
N while the second represents the loss to this society due to the encounter between the two societies.
Discuss the steady states of this model and their stability when M is small, moderate, and large.
Hint: To find the steady states plot the two terms in the model equation separately. The steady states
are represented by the intersection of these two curves.

4.2 Two species model

Mathematical models which lead to systems of first order ordinary differential equations appear in
many applications. It should be observed also that when sati sodel equations contain second (or
higher) order differential equations, then these equations can be rewritten as a system of first order
equations. In particular when the model consists of one ordinary differential equation of second order,
then this equation is equivalent to a system of two first order equations.
When the model is represented by two first order equations, then phase space methods in
conjunction with graphical representation can be used to determine the stability of the steady states. In
this section we demonstrate this technique through an example.
Wearple 1:
Predator‐Prey Ecological System.
Model: In this model we consider an ecological system which consists of two apecies. The first
(Prey) consumes vegetable food and a second species (Predator) consumes vegetable and the first
species for food (e.g., fish and sharks in the ocean or humans and cows). If only one of these species
exists in the ecosystem, then its population will be governed by an equation similar to Equation(4.1).
To model the interaction between these two species we shall assume that at each encounter between
members of the two species there is a “chance” that the predator will consume the prey. Hence this
interaction can be modeled as being proportional to the size of the two populations.
Denoting the size of the two populations at time t by F(t), S(t) respectively, the ecological system
under consideration will be modeled by
(4.3)
dFdt=F(a1-b1F-c1S)
(4.4)
dSdt=S(a2-b2S+c2F)
where a i , b i , c i , i = 1, 2 are non-negative constants.

4.2.1 Steady States

At the steady states of this system d F d t = d S d t = 0 . Hence, these steady states of this model are
given by the simultaneous solutions of
(4.5)
F(a1-b1F-c1S)=0
(4.6)
S(a2-b2S+c2F)=0.
These solutions are:
1. F = S = 0
2. F = a 1 b 1 , S = 0
3. F = 0 , S = a 2 b 2
4. The intersection of the two lines
(4.7)
a1-b1F-c1S=0
(4.8)
a2-b2F+c2S=0
 If such an intersection exists in the first quadrant of the F - S (phase) plane (obviously negative
populations are meaningless).

4.2.2 Stability Analysis

To analyze the steady states with respect to their stability, using phase space techniques, we must first
find the regions in the F - S plane for which d F d t ≶ 0 , d S d t ≶ 0 . To accomplish this we note that d
F d t > 0 when
F(a1-b1F-c1S)>0.
But F ≥ 0 and, therefore, d F d t is positive of and only if
(4.9)
a1-b1F-c1S>0,F>0.
Similarly d F d t is negative if and only if
(4.10)
a1-b1F-c1S<0,F>0.
It follows then that the dividing line in the phase plane between the regions in which d F d t is
positive and negative is given by
(4.11)
a1-b1F-c1S=0.
Similar analysis for d S d t shows that
d S d t > 0 if and only if
(4.12)
a2-b2S+c2F>0,S>0.
d S d t < 0 if and only if
(4.13)
a2-b2S+c2F<0,S>0.
Therefore the two regions in which d S d t is positive or negative are separated by the line
(4.14)
a2-b2F2+c2F1=0.
For a 1 = 1, b 1 = 0.2, c 1 = 0.1, a 2 = 0.5, b 2 = 0.4 and c 2 = 0.5 we plot the lines given by
(Equations 12.9), (12.12) separately in Figures 4.3 and 4.4 respectively indicating by arrows the
regions where d F d t and d S d t are positive and negative. In Figure 4.5 we plotted these two lines
together to show by arrows the response of the system to any perturbation from the steady state
represented by the intersection of these two lines. Two trajectories starting from (3.5,4) and (2.5,3)
are also plotted in this figure. This demonstrates that any perturbation from the steady state will decay
in time and the population of the two species will return to the steady state. This implies that the two
species will coexist in this ecosystem.
Figure 4.3 Regions in phase plane of (Equations 12.1) - (12.2) where d F d t is positive and negative

Figure 4.4 Regions in phase plane of (Equations 12.1) - (12.2) where d S d t is positive and negative

Figure 4.5 Stability of the steady state represented by the intersection of the lines d F d t = d S d t = 0
Figure 4.6 lntegral Curves of the Spring-mass system without friction

Figure 4.7 Improper node (unstable)

Figure 4.8 Saddle point.

 Using similar analysis we find that
1. (0, 0) is unstable
2. [ a 1 b 1 , 0 ] is unstable in the S‐direction and hence unstable.
3. [ 0 , a 2 b 2 ] is unstable in the F‐direction and hence unstable.

4.3 Basic concepts

In the previous section of this chapter we introduced the basic concepts of stability theory in an
intuitive way. A graphical representation of these concepts is illustrated by the following examples. In
the first case a ball is placed on top of a mountain and hence in unstable steady state. In the second
case the ball is placed at the bottom of a “valley” (with ground friction. In this case the steady state is
“asymptotically stable That is, a perturbation from the steady state will decay in time, and the ball
will return to the original steady state. Finally we consider a ball on a plane (with friction) where a
perturbation from the original state will leave the ball at some distance from the original steady state
but not “far away”. (The distance remains bounded due to friction.) In this case we say that the steady
state is “stable”
In the following we reintroduce and formalize these concepts.

Figure 4.9 Proper node (unstable).

Definition 4.3.1
A system of ordinary differential equations
(4.15)
dxdt=F(x,t)
where x = ( x 1 ( t ) , … , x n ( t ) ) and
(4.16)
F(x,t)=f1(x,t)⋮ fn(x,t),x∈ Rn
is called autonomous if F(x, t) = F(x), i.e., the independent variable t does not appear explicitly in
F.
We observe that a non-autonomous system is equivalent to an autonomous system with an additional
equation. In fact if we define
(4.17)
z=xt,G(z)=F(x,t)1
then
(4.18)
dzdt=G(z)
is equivalent to (Equation 4.15)
In the following we consider only autonomous systems. Such systems are referred to as dynamical
systems.
Definition 4.3.2
The phase space of the system
(4.19)
dxdt=F(x)=f1(x)⋮ fn(x),x∈ Rn
is the space R 2n with coordinates ( x 1 , … , x n , x ˙ 1 , … x ˙ n ) ,
where x ˙ = d x i d t . Thus each point in phase space represents a unique state of the system. For a
system of particles satisfying Newton’s second law the phase space consists of all possible values of
the momentum and position variables.
Example 4.3.1
The equations of motion for a point particle of mass under the influence of an external force
F = F(x) is
(4.20)
md2xdt2=F(x),x∈ R3
which is equivalent to
(4.21)
dxdt=y,mdvdt=F(x).
Hence the phase space of Equation (4⋅ 20) is R 6 consisting of the points ( x , v ) .
Definition 4.3.3
A steady state(≡ critical point or equilibrium state) of the system (4⋅ 19) is a point x 0 such that
F(x 0) = 0.
We observe that a steady state can be isolated; i.e., there exists a neighborhood of x 0 which contains
no other steady state or it might be part of a continuous set of critical points.
Example 4.3.2
The system
(4.22)
dxdt=a1x+b1y,dydt=a2x+b2y,
has a continuous set of steady states when a 1 b 2 - b 1 a 2 = 0. In fact the steady state conditions d x
d t = d y d t = 0 lead to
(4.23)
a1x+b1y=0,a2x+b2y=0,
and these equations are linearly dependent (that is they represent the same equation) when a 1 b
2 - b 1 a 2 = 0. It follows then that any point on the line y = - a 1 x b 1 , b 1 ≠ 0 is a steady state of
the system. (When b 1 = 0, any point on the line x = 0 is a steady state.)
On the other hand this analysis shows that (0, 0) is an isolated steady state of the system represented
by Equation(4.22) when a 1 b 2 - b 1 a 2 ≠ 0, since then the coefficient matrix of (Equation 4.23) is
invertible.
In general it is rather difficult to analyze the stability of a system with a continuous set of steady
states. Therefore, in the following we restrict our discussion to systems with isolated critical points.
The following is a generalization of this example.
Theorem 4.3.1
Let F(x) be analytic and x 0 a steady state of the system represented by Equation (4 ⋅ 19). A
sufficient condition for x 0 to be isolated is that the Jacobian matrix of F(x) at x 0
(4.24)
J(x0)=[∂fi∂xj](x1)
is nonsingular.
Proof
Since F(x) is analytic, it has a Taylor series expansion that converges to F(x) . This Taylor expansion
can be used to approximate this function around the critical point x 0.
(4.25)
F(x)=F(x0)+J(x0)(x-x0)+O(|x-x02)=J(x0)(x-x0)+O(|x-x0|2).
Therefore, if another steady state exists in every neighborhood of x 0, then in the vicinity of x 0 all
higher order terms in Equation (4.25) become negligible in comparison to the linear term, and the
steady state must satisfy
(4.26)
J(x0)(x-x0)=0.
But J(x 0) is nonsingular which implies that the only solution of this system is x = x 0. Hence x 0 is
isolated. Note, however, that this theorem sets only a sufficient) but not necessary condition for x 0 to
be isolated.
Example 4.3.3
The only critical point of the system
(4.27)
dxdt=y2,dydt=x2
is (0, 0) although J(0, 0) is singular.
When a steady state is isolated, it is convenient, for conceptual and practical reasons, to translate the
steady state under consideration to the origin by the transformation w = x - x 0. The equations of the
system will become,
(4.28)
dwdt=G(w)
where G(w) = F(w + x 0)
Example 4.3.4
Consider the system
(4.29)
dxdt=4-xy,dydt=4x-y3
whose steady states are ( 2 , 2 ) and ( - 2 , - 2 ) . To translate ( 2 , 2 ) to the origin we perform the
transformation
(4.30)
u=x-2,v=y-2
and the resulting form of the system given by Equation (4 ⋅ 29), is
dudt=-2u-2v-uv
(4.31)
d v d t = 4 u - 12 v - 6 v 2 - v 3
Similarly if we wish to consider the steady state ( - 2 , - 2 ) we perform the translation
(4.32)
u=x+2,v=p+2
The resulting form of the system in Equation (4 ⋅ 29) is
dudt=2u+2v-uv
(4.33)
d v d t = 4 u - 12 v + 6 v 2 - v 3 .
Definition 4.3.4
A trajectory ( ≡ o r b i t , p a t h ) of the system given by Equation (4 ⋅ 19) is a solution of this
system x = x(t) subject to an initial condition x(0) = c.
Sometimes such a trajectory is written as x ( t , c ) .
Definition 4.3.5
Let x 0 be a steady state of the system, Equation (4 ⋅ 19) . We say that
1. x 0 is stable if for any given ɛ > 0 there exists a δ > 0 so that whenever |x(0) - x 0| < δ then
|x(t) - x 0| < ɛ for all t > 0.
2. x 0 is asymptotically stable if it satisfies the condition for being stable and in addition
|x(0) - x 0| < δ implies that
(4.34)
lim t → ∞ | x ( t ) - x 0 | = 0 .
Thus, x 0 is stable if whenever the initial state of the system is near the steady state then its trajectory
will remain in its vicinity for all times. On the other hand if these trajectories approach x 0 as t → ∞
then we say that x 0 is asymptotically stable.
Definition 4.3.6
A steady state that is neither stable nor asymptotically stable is called unstable.
Example 4.3.5
The differential equation governing the motion of the linear pendulum (without friction) which
was presented in Chapter 2, is
d2θdt2+ω2θ.
The general solution of this equation is
θ = A cos ( ω t + ϕ )
where A and φ are integration constants which are determined by the initial conditions. A small
perturbation from the steady state θ = 0, θ ˙ = 0 will take the system to a new (time dependent)
state but the “distance” between the equilibrium state and the new state will remain bounded and
does not go to 0 as time goes by. Hence the steady state is stable. On the other hand if we add
friction to this system the amplitude of the oscillations will decay to zero as t → ∞ and the
equilibrium state of the system becomes asymptotically stable.
Definition 4.3.7
The integral curves of the system represented by Equation (4 ⋅ 19) are the solutions of the system
(4.35)
dx1f1=…=dxnfn.
Thus the integral curves of Equation (4 ⋅ 19) are the trajectories of this system parametrized in
terms of the x i ′ s rather than in terms of the “extraneous” variable t.
Example 4.3.6
Find the trajectories and integral curves of
(4.36)
dxdt=αx,dydt=βy.
Solution 4.3.1 The trajectories of the system are
(4.37)
x=C1eαt,y=C2eβt
where C 1, C 2 are arbitrary constants. The corresponding integral curves are y α = Cx β where C is
a constant.
Example 4.3.7
Find the integral curves of the equations of the spring-mass system(with no friction).
Solution 4.3.2 The equation of motion for the spring mass system for small displacements from the
steady state is
(4.38)
d2xdt2+kx=0.
This is equivalent to the system
dxdt=v
(4.39)
dvdt=-kx.
The integral curves of this system satisfy
(4.40)
dxdv=v-kx
and hence
kx2+v2=C2
where C 2 is a constant. We conclude, therefore, that the integral curves of the spring‐mass system
are ellipses (see Figure 4.9). It follows from this result that the steady state (0, 0) of the system
given by Equation (4 ⋅ 39) is stable but not asymptotically stable.This result is obvious from a
physical point of view since the system contains no friction to damp the motion.
 Exercises
1. Find the form of the following systems when each of their steady states is translated to the
origin.
dxdt=x( x- 2) ( y- 2) ,dydt=( x+2) ( y- 1) 2
d x d t = sin 2 π x cos 2 π y , d y d t = ( sin π y ) 2 - 1
Find the trajectories, integral curves and stability of the critical point at the origin for the
following systems
dxdt=- x,dydt=2x- y
dxdt=- 3x+y,dydt=x- 3y
d xd t= nz- ky, d yd t= kx- mz, d zd t= my- nx
Find the integral curves and stability of the critical points for the following systems.
ẍ=x3
m ẍ + b x ˙ + k x = 0 , m, b, k > 0. What happens if b < 0?
θ ¨ + ν 2 sin θ = 0

(nonlinear pendulum). What happens when |θ| is small and it is solution to approximate sin θ
by θ.

4.4 Linearizable Dynamical Systems

When a dynamical system is nonlinear, it is impossible, in general, to find analytical expressions for
the trajectories or integral curves of the system. As a result it is, impossible to infer the stability of its
steady states directly. To overcome this issue one usually attempts to approximate the expansion of the
dynamical system in the vicinity of a steady state by a linear system of equations. When such an
approximation exists, we say that the system is linearizable (near the steady state under
consideration). This linear approximation can be used then to deduce the stability of the steady state
(with some exceptions).
Definition 4.4.1
Let x 0 be an isolated critical point of the system
(4.41)
dxdt=F(x)=f1(x)⋮ fn(x)
We say that the system, (Equation 4.41), is linearizable at x 0 if the Jacobian of F at x 0 is nonsingular,
i.e.
(4.42)
det J ( x 0 ) = | ( ∂ f i ∂ x j ( x 0 ) | ≠ 0 .
Example 4.4.1
The system
(4.43)
dxdt=(x-1)(x-3)2,dydt=y
is linearizable at the steady state (1, 0) but is not so at the steady state (3, 0).
To treat systems which are linearizable at the critical point x 0 we first translate this point to the origin
and then take the Taylor expansion of each f i (x) around the origin
 (4.44)
fi(x)=∑j=1naijxj+O(|x|2).
It follows then that linear approximation to the original system around the origin is given by
(4.45)
dxdt=Ax
where A = J(0) is a constant coefficient matrix.
We observe that (Equation 4.42) implies that not all the coefficients a i1,...,a in in the expansion of the
function f i (x) are zero. Therefore, there exists a neighborhood of 0 where the linear terms are
dominant in the corresponding equation. This can be considered as the reason for the relationship
between the stability of the linear system (Equation 4.45) at the origin and the original system.
Example 4.4.2
The system
dxdt=x+y2-6,dydt=2y-x2
has a critical point at x = 2, y = 2. The determinant of the Jacobian at this
point is
(4.46)
det J ( 1 , 1 ) = ∂ f 1 ∂ x ∂ f 1 ∂ y ∂ f 2 ∂ x ∂ f 2 ∂ y ( 2 , 2 ) = 1 4 - 4 2 ≠ 0 .
Hence this system is linearizable at this steady state. To compute the linear approximation we first
translate this steady state to the origin by the translation
(4.47)
u=x-2,v=y-2.
The system now takes the form
(4.48)
dudt=u+4v+v2,dvdt=-4u+2v-u2,
and the resulting linear approximation at the origin is
(4.49)
dudt=u+4v,dvdt=-4u+2v.
In matrix notation this takes the following form:
(4.50)
ddtuv=14-42uv.
When the system (Equation 4.41), is linearizable at the steady state x 0, it is possible to characterize
the stability of this steady state (after the translation to the origin) in terms of the eigenvalues of the
matrix A in (Equation 4.45). We start with the following theorem.
Theorem 4.4.4
The critical point at the origin of the system Equation (4 ⋅ 45), is
1. Asymptotically stable if all the eigenvalues of A have negative real parts.
2. Stable if all the eigenvalues of A have non‐positive real parts and every eigenvalue of A
which has a zero real part is a simple eigenvalue of A.
3. Unstable if (1) and (2) are false.
Example 4.4.3
The eigenvalues of the matrix A in Equation (4 ⋅ 50) are
λ ± = 3 ± i 55 2 .
Therefore this steady state an (0, 0) of Equation (4 ⋅ 49) is unstable
Example so.4.4 The critical point at the origin of the system
(4.51)
d x d t = 1 - 2 4 7 - 8 10 2 - 2 1 x
is asymptotically stable since the eigenvalues of A are - 3, - 1, - 2.
The relationship between the stability of the steady states of the systems (4.41) and (4.45) at x = 0 is
summarized by following theorem.
Theorem 4.4.2
If the steady state at the origin of the linearized system Equation (4 ⋅ 45), is asymptotically stable
or unstable, then the same is true for the steady state of the original system Equation (4 ⋅ 41).
However, if the origin is a stable steady sThe for the linearized system Equation (4 ⋅ 45), then the
stability of Equation (4 ⋅ 41) at this point is indeterminate, i.e., the stability of the original system
at this point cannot be deduced from that of the linearized system.
Example 4.4.5
The system
(4.52)
ddtxy=02-20xy-α0x2y
is almost linear at the critical point 0. Its linearization at this point is given by
(4.53)
ddtxy=02-20xy=Ax(
Since the eigenvalues of A are ± 2i, it follows from theorem 4.3.1 that (0, 0) is a stable critical point
of the system (Equation 4.53). However, it can be shown that the original system (Equation 4.52), is
asymptotically stable if α > 0 and unstable if α < 0. (See exercise 6 in Section 6).
Exercises
1. Show that the solution x = 0 of
(4.54)
akx( k) +ak- 1x( k- 1) +…+a1x′+a0x=0

is stable if and only if all the roots of the polynomial

p( λ ) =a0λ n+…+a0

have non-positive real parts and all roots with zero real parts are simple.
Hint: Solve (Equation 4.54)
algorithm is given in Table 3.3.
For the following exercises find the critical points at which the system is almost linear and
discuss their stability.
2. d x d t = ( x - 1 ) ( x - 2 ) 2 ( y - 1 ) , d y d t = ( x - 1 ) 2 ( y - 2 )
3. d x d t = y ( 1 + cos x ) , d y d t = x ( 1 + sin y )
4. ẍ + β ( x 2 - 1 ) x ˙ + k x = 0 , b > 0. (Van der Pol eq.)
 5. ẍ + β x ˙ - x + 2 x 3 = 0 (Duffin’s equation)

4.5 Linearizable systems in two dimensions

The previous section presented a general framework that relates, under proper conditions, the
stability of a steady state of a nonlinear dynamical system to its linearization at this point. However in
two dimensions it is possible carry out further classification, of these steady states in terms of their
“phase portrait”(i.e., phase diagram). In the following we describe this classification. Let the system
(4.55)
dxdt=G(x,y),dydt=F(x,y)
have an isolated critical point at the origin. If this system is linearizable and the Jacobian at (0, 0) is
nonsingular. We can take the Taylor expansion of G, F around this point and rewrite (Equation 4.55)
as
(4.56)
dxdt=ax+by+f1(x,y)
(4.57)
dydt=cx+dy+f2(x,y)
Since the Jacobian of (Equation 4.55) is nonsingular at (0, 0), ad—bc ≠ 0.Moreover, f 1, f 2 are of
order |x|2. Therefore, the linear approximation to the system given by (Equation 4.55) near the origin
is given by
(4.58)
dxdt=ax+by,dydt=cx+dy,ad-bc≡ 0.
To solve the system (Equation 4.58), we make the ansatz that the solution is of the form
(4.59)
x = A exp ( λ t ) y = B exp ( λ t )
(same λ for both x and y). Substituting (Equation 4.59) in (Equation 4.58) yields then
(4.60)
(a-λ)A+bB=0
(4.61)
cA+(d-λ)B=0.
Solutions (4.60), (4.61) form a system of linear homogeneous equations for the coefficients A, B. A
nontrivial solution for these coefficients exists if and only if the determinant of the coefficients of this
system vanishes, i.e.,
(4.62)
a-λbcd-λ=λ2-(a+d)λ+(ad-bc)=0.
It follows that a nontrivial solution exists only if λ is an eigenvalue of the coefficient matrix of the
system (Equation 4.58), and the corresponding solution of (Equation 4.62), A B , is the eigenvector
related to this eigenvalue. The stability, instability, and the phase portrait of the steady state at the
origin is determined, therefore, by the eigenvalues and eigenvectors of this coefficient matrix.
In the following we provide a classification for all these possibilities.
Case 1: (Equation 4.62) has two real unequal roots λ 1 , λ 2 of the same sign. The general solution
of the system, (Equation 4.58), is given by
(4.63)
x=A1eλ1t+A2eλ2ty=B1eλ1t+B2eλ2t.
If λ 1 , λ 2 are positive, the steady state at the Graph is unstable. If on the other hand both are
negative, the steady state is asymptotically stable. (Observe that these eigenvalues cannot be zero
since we are assuming that the coefficient matrix of (Equation 4.58) is not singular.) A typical
illustration of the integral curves of the system in this case is shown in Figure 4.7. A steady state with
the phase portrait shown in this figure is called improper node (stable or unstable).
Case 2: (Equation 4.62) has two real roots with opposite signs.
Under the assumption of this case (Equation 4.63) still represents the solution of the system,
(Equation 4.58). However, now, the steady state is always unstable since one of the eigenvalues is
positive. Its phase portrait is given in Figure 4.8. Since there is a direction in which the steady state is
stable, it is referred to as a “saddle point.”
Case 3: (Equation 4.62) has two real roots .
Since the coefficient matrix of Equation(4.62) is nonsingular (ad—bc≠ 0) it is obvious that this
root must be real and different from zero. Hence the solution of the system, (Equation 4.58), is given
by
(4.64)
x=(A1+A2t)eαt,y=(B1+B2t)eαt.
This implies that the steady state is asymptotically stable if α < 0 and unstable if α > 0. The phase
portrait of such a steady state is referred to as improper or proper node (Figures 4.7 or 4.9
respectively).
Case 4: (Equation 4.62) has complex roots with nonzero real part
s±=β±iω,β≠0.
The solution of the system, (Equation 4.58), is
x = e β t ( A 1 cos ω t + A 2 sin ω t )
(4.65)
y = e β t ( B 1 cos ω t + B 2 sin ω t ) .
The critical point is unstable if β > 0 and asymptotically stable if β < 0. The phase portrait of the
steady state is a spiral, regardless of the stability of the steady state, as shown in Figure 4.10.
Figure 4.10 Spiral (stable).

Case 5: (Equation 4.62) has pure imaginary roots

s=±iω
The solution of the system, (Equation 4.58), is represented by (Equation 4.65) with λ = 0 . The
critical point is stable and its phase portrait as shown in Figure 4.5. The steady state is called a
“center”.
As to the relationship between the phase portraits, at the origin of the system, (Equation 4.55), and
its linearization (Equation 4.58), we have the following
Theorem 4.5.1
The phase portraits at (0, 0) of the systems, Equation (4 ⋅ 55) and Equation (4 ⋅ 58) are the same
except in cases 3 and 5 where a spiral is possible as an additional phase portrait.
To motivate this result we note that the additional terms in (Equation 4.56) might destroy the exact
equality of the roots (case 3) or add to them a small real part when they are purely imaginary as in
case 5.
It should be also noted that when the roots of (Equation 4.62) are purely imaginary the stability of the
steady state at (0, 0)of the system, (Equation 4.55), cannot be deduced from those of the system,
Equation (4.gin This conforms to the statement of theorem 3.4.2 in the previous section.
Exercises
For the following systems classify and draw the phase portrait for the steady state at the origin.
1. m ẍ + b x ˙ + k x = 0 m, b, k > 0. What happens if b < 0
2. ẍ + b x ˙ + k x 2 = 0 , k > 0. Consider separately the case b > 0 and b < 0.
3. d x d t = sin ( x - y ) , d y d t = e x - y - cos ( x - y )
4. d x d t = x + y + 2 y 2 - x 2 , d y d t = y - 2 x + y 3
5. d x d t = 2 y - x 3 , d y d t = 2 x + y 2 (compare with the exact solution)
6. d x d t = y - 3 x , d y d t = e x + e y sin x - 1

4.6 Liapounov method

In the last two sections of this chapter we discussed dynamical systems which are linearizable near
an isolated steady state. We showed that under this restriction there is a close correspondence
between the stability of the steady state of the original system and its linear approximation. The
question naturally arises as what techniques can be used to determine the stability of the steady state
when the dynamical system is not linearizable.
Another aspect of stability theory which was not discussed so far is the size of the “basin of
stability” of the steady state. That is if x 0 is an asymptotically stable steady state of the system what is
the size of the “maximal perturbation” to the system which will decay in time to x 0.
To answer these questions Liapounov (direct) method is a powerful tool which has been applied
successfully in many practical applications. In the following we present the essence of this technique
and its application to gradient dynamical systems.
Definition 4.6.1
Let the function F:R n → R be defined on a domain Ω ∊ R n containing the origin and F(0) = 0.
1. F is said to be positive definite on Ω if F(x) > 0 for all x ≠ 0 in Ω.
2. F is said to be positive semidefinite on Ω if F(x) ≥ 0 for all x ∊ Ω.
In a similar fashion one can define functions which are negative definite and negative semidefinite on
Ω.
Example 4.6.1
The function
(4.66)
F(x,y,1)=x2+y2+z2
is positive definite in R 3.
Example 4.6.2
The function
(4.67)
F ( x , y ) = sin x 2 + y 2
is positive definite on the disk 0 ≤ x 2 + y 2 < π.
Example 4.6.3
The function
(4.68)
F(x,y)=(3x-y)2
is positive semidefinite in R 2 since F ( x , y ) = 0 on the line y = 3x but positive otherwise.
Example 4.6.4
Show that the function
(4.69)
F(x,y)=ax2+bxy+cy2
is positive definite in R 2 if c > 0 and 4ac - b 2 > 0.
Solution 4.6.1 By adding and subtracting b 2 4 c x 2 from the expression of F ( x , y ) we can
rewrite it as
(4.70)
F(x,y)=[b2cx+cy]2+4ac-b24cx2
which is obviously positive definite if c > 0 and 4ac - b 2 > 0.
Consider now the autonomous dynamical system
(4.71)
dxdt=G(x)=g1(x)⋮ gn(x)
which has an isolated steady state at x = 0.
Theorem 4.6.1
(Liapounov) Let F(x) be, a positive definite function with continuous derivatives on a domain Ω
containing the origin and define
(4.72)
F˙(x)=DGF(x)=gradF·G=∑i=1n∂F∂xi·gi.
If on some domain containing the origin
1. F ˙ ( x ) is negative definite then the steady state x = 0 is asymptotically stable,
2. F ˙ ( x ) is negative semidefinite then x = 0 is stable,
3. F ˙ ( x ) is positive definite then x = 0 is unstable.
Observe that this theorem makes we reference to the solution of the system, (Equation 4.71), in
contrast to the techniqueof linearization.
Example 4.6.5
Discuss the stability of the steady state x = 0 for the system
(4.73)
x˙=-2y+αx(x2+y2)=g1(x,y)
(4.74)
y˙=2x+αy(x2+y2)=g2(x,y)
where α is a parameter.
Solution 4.6.2 Let F = x 2 + y 2, then
(4.75)
F˙(x,y)=gradF·G=2α(x2+y2)2.
Therefore, Liapounov theorem implies that if α < 0, then the steady state at the origin is
asymptotically stable. If α = 0 the state is stable while if α > 0 the state is unstable.
We infer from this example that the major difficulty inherent in this method is that one has to construct
(or find) a proper positive definite function F(x) (which is refered to as Liapounov function) that
satisfies the requirements of Liapounov theorem. Several ad hoc algorithms, for particular
applications, were suggested in the literature to overcome this difficulty. However, most of the
important applications of this method remain limited to systems for which F(x) can be deduced from
physical arguments. These are gradient and conservative dynamical systems. (For many of these
systems the function F is the Hamiltonian or a function related to it.)
Definition 4.6.2
A dynamical system, Equation (4⋅ 71), is called a gradient system if there exists a (smooth)
function φ(x) so that
(4.76)
G(x)=-gradϕ(x).
φ(x) is called the potential function of the system.
Theorem 4.6.2
Let φ(x) be a potential function for the system, Equation (4⋅ 71), and let x 0 be an isolated local
minimum of φ. Then x 0 is an asymptotically stable steady state of the system, Equation (4⋅ 71).
Proof
Since φ is smooth and x 0 is a local minima of φ, there exists a neighborhood Ω of x 0 in which
(4.77)
F(x)=ϕ(x)-ϕ(x0)
is positive definite. Furthermore since x 0 is an isolated minimum in Ω, grad F= grad φ ≠ 0.
On Ω we have
(4.78)
F˙(x)=gradF·G=gradF·(-gradϕ)=-|gradF|2.
Hence F ˙ ( x ) is negative definite on some neighborhood of x 0. It follows then from Liapounov
theorem that x 0 is asymptotically stable.
Definition 4.6.3
A force field F(x) is said to be conservative if there exists a smooth function φ(x) so that
F(x)=-gradϕ(x).
In particular a mechanical dynamical system is said to be conservative if the force field acting on
the system is conservative.
Example 4.6.6
The equation of motion for a particle of mass m under the action of a force F is given by Newton’s
second law
(4.79)
md2Xdt2=F(x).
If F is conservative, then
(4.80)
md2Xdt2=-gradϕ.
Rewriting equation (4 ⋅ 80) as a system of first order equation we obtain
(4.81)
dxdt=v
(4.82)
d v d t = - 1 m grad ϕ .
It follows from Equations ( 4.81 )‐(4.82) that at a steady state ( x 0 , v 0 ) of such a particle,
v 0 = 0 , grad ϕ ( x 0 ) = 0
i. e., x 0 corresponds to a local extremum of φ.
Theorem 4.6.3
A stable steady state (x 0, 0) of a particle in a conservative force field corresponds to an isolated
local minima of φ.
Proof
The total energy of the system under consideration is
 (4.83)
E(x,v)=12mv2+ϕ(x).
If x 0 is a local minima of φ, then
F(x,v)=E(x,v)-E(x0,0)
is positive definite on some domain around (x 0, 0). Moreover, from, (Equations 4.81) - (4.82) it
follows that
(4.84)
F ˙ ( x , v ) = grad F · ( v , - 1 m grad ϕ ) = ( grad ϕ , m v ) ( v , - 1 m g r a d ϕ ) = 0 .
Hence by Liapounov theorem (x 0, 0) is a stable steady state.
We remark that theorem (4.6.3) is a special case of Lagrange theorem. This theorem states the
following
Theorem 4.6.4
Steady state (x 0, 0) of a system of particles in a conservative force field is stable if x 0 is an
isolated local minima of φ(x) .
Exercises
1. Let f(0) = 0, f(x) > 0 on the interval [ 0 , b ] and f(x) < 0 on the interval [ - b , 0 ] . Show
that
F( x,y) =12y4+∫0xf( t) dt

is positive definite on
Ω={ ( x,y) ;- b<x<b,- ∞<y<∞} .

What happens if y 4 is replaed by y 2?.

2. Show that
H( x,y) =ax2+bxy+cy2

is negative definite if c < 0 and 4ac - b 2 > 0.

3. For the system
x¨ +αf( x) =0,α>0

where f(x) satisfies the assumptions of exercise 1, show that x = 0, x ˙ = 0 is a stable steady
state.
Hint: Let F ( x , x ˙ ) = 1 2 x ˙ 2 + α ∫ 0 x f ( t ) d t .
4. Apply the results of exercise 3 to the nonlinear pendulum
x ¨ + ω 2 sin x = 0 , ω 2 = g L .
5. Show that the following two systems are stable or asymptotically stable at 0
6. d u d t = - u - 2 u v 2 , d v d t = - v - v u 2
7. d u d t = - u 3 + 2 u v 2 , d v d t = - 2 u 2 v - 4 v 3 .
Hint: Consider a Liapounov function of the form
F( x,y) =au2+bv2,a,b>0.
8. For the system
 dxdt=- ny- xf( x,y)
dydt=nx- yf( x,y) ,n=1,2,⋯

show that 0 is an asymptotically stable state when f ( x , y ) > 0 in some neighborhood of the
steady state and unstable if f ( x , y ) < 0 in some neighborhood of 0.
Hint: Note that this is a generalization of Example 4.6.5.

4.7 Periodic SOLUTIONS(LIMIT cycles)

In many important practical applications, one has to consider dynamical systems, (Equation 4.71),
with periodic motions. That is, some of the trajectories “allowed” by the system satisfy either exactly
or approximately x(t + T) = x(t), for some T > 0. In these cases the steady state of the system is
represented by a one dimensional curve in R n (usually referred to as a limit cycle) rather than a point.
Perhaps one of the most important examples of such a system relates to the motion of Earth and the
other planets around the Sun. This motion is subject to perturbations due to the gravitational field of
the other planets and other celestial bodies in the system. It follows then that the answer to the
question of stability of this periodic motion is of paramount importance to the ve + y existence of life
on Earth. However, periodic motion is important also in many mundane applications, e.g. waves of
various types (water waves, electromagnetic waves, etc). In this section we present some elementary
techniques, to determine the existence and stability of these limit cycles. However, in general the
existence and stability of limit cycles for a given dynamical system is a rather difficult mathematical
problem.
We start with, few examples.
Example 4.7.1
Consider the following almost linear system with a critical point at (0, 0):
(4.85)
dxdt=α2x-ωy-x(x2+y2)
(4.86)
dydt=ωx+α2y-y(x2+y2)
where α and ω are constants. The linearization of this system at the steady state (0, 0) is given by
prescription of the form,
(4.87)
dxdt=α2x-ωy,dydt=ωx+α2y.
The eigenvalues of the coefficient matrix for this system are α 2 ± ωi. Therefore, (0, 0) is an
unstable spiral point. Thus a trajectory originating from a point near (0, 0) will spiral away from
this point. However, we now show that these spiral trajectories remain bounded. Moreover
trajectories of the system, Equations ( 4.85 )–( 4.86 ) which originate from a point far away from
the origin are actually directed inward.
To prove these statements we recast (Equations 4.85),(4.86) using polar coordinates
(4.88)
x = r cos θ y = r sin θ
and note that
(4.89)
xdxdt+ydydt=rdrdt
(4.90)
ydxdt-xdydt=-r2dθdt.
The system, Eq U1tion (4.85), (4.86) now takes the form
(4.91)
drdt=r(α2-r2),dθdt=ω.
Therefore, if a trajectory starts from a point which satisfies r > α, then d r d t < 0 and the trajectory
will spiral inward. If on the other hand a trajectory starts from a point which satisfies r < α, then d r d
t > 0 and the trajectory will spiral outward. Moreover, for the circle r = α we have d r d t = 0 which
implies that
(4.92)
r=α=const,θ=ωt+t0
is a limit cycle of this system. TR sum it up: Trajectories of this system with initial condition r 0 > α
will spiral toward the limit cycle, and the same is true for those with r 0 < α. It follows then that this
periodic solution is, asymptotically stable limit cycle.
Example 4.7.2
Systems with variable damping
We first observe that the solution of the spring‐mass system
(4.93)
mx¨+bx˙+kx=0,m,k>0
is
x(t)=e-αt[C1eβt+C2eβt]
where
α=bm,β=b2-4kmm.
Therefore, when b is positive the motion is damped and the amplitude of x(t) will decrease in time.
However, when b < 0 (negative damping), the amplitude ofx(t) will increase in time. Thus the sign
of the frictional coefficient b determines if the motion is damped or not.
Now consider Rayleigh equation.
(4.94)
x¨-b(1-x2)x˙+kx=0,b,k>0.
In this equation the “effective” frictional coefficient is - b ( 1 - x ˙ 2 ) . Therefore when ( x ˙ ) 2 < 1
, the damping in this equation is negative but when ( x ˙ ) 2 > 1 , the “damping” is positive.
Therefore, the amplitude of the motion described by this equation will increase for small velocities
and decrease for large ones. Therefore, it stands to reason that in between these two types of
motions there will be an oscillation of constant amplitude, i.e., a periodic motion.
Here are some formal definitions and theorems regarding limit cycles.
Lemma 4.7.1
A trajectory of a dynamical system in phase space is closed if and only if it corresponds to a
periodic solution of the system.
Proof
If the solution x(t) of the dynamical system is periodic, there exist T > 0 so that x(t + T) = x(t) for all
t > 0. This obviously implies that the trajectory of the system in phase space is also closed. The
reverse is also obvious.
Definition 4.7.1
A closed trajectory X of a dynamical system which has nearby open trajectories spiraling towards
it from both the inside and outside as t → ∞ is called, asymptotically stable limit cycle. Similarly
if nearby trajectories neither approach nor recede from X, we say that it is (neutrally) stable.
Example 4.7.3
The trajectories of the linear pendulum without damping are neutrally stable.
Definition 4.7.2
A point y is a limit point of a trajectory x(t) of a dynamical system
(4.95)
dxdt=F(x)
If there exists a sequence {t n }, lim n → ∞ t n = ± ∞ so that lim x ( t n ) = y .
Definition 4.7.3
A set X is called a limit set of Equation (4 ⋅ 95) if each of its points is a limit point of some
trajectory of this system.
Example 4.7.4
An isolated asymptotically stable steady state of a dynamical system is obviously a limit point of all
nearby trajectories.
Example 4.7.5
The trajectory r = α in Example 4.7.1 is a limit set for the system, Equations (4 ⋅ 85),(4 ⋅ 86),
since each of its points is a limit point for the trajectories that spiral towards r = α.
Exercises
1. Use the same analysis as in example 4.7.2 to show (intuitively) that the Van der Pol equation
(4.96)
d2udt2+b( 1- u2) dudt+ku=0

where b and k are positive constants admits a limit cycle.

2. Show that the system
(4.97)
dudt=- v+uh( r ) ,dvdt=u+vh( r ) r 2=u2+v2

has limit cycles which correspond to the roots of the function h(r).
Hint: Use a polar representation of the system, (Equation 4.97).
3. Determine the periodic solutions and their stability for the system, (Equation 4.97) if
4. h(r) = r(r - 1)2(r - 3)(r - 4)
5. h(r) = r 2 - 4
6. h ( r ) = sin n r , n = 1, 2, . . .
7. Under what conditions on h(r), does the following system admit a limit cycle
 (4.98)
dudt=2u+v- uh( r ) dvdt=- u+2v- vh( r )

Hint: Use polar representation of the system.

8. Determine the periodic solutions and their stability for the system, (Equation 4.98), if
9. h(r) = r 2 - n 2, n = 1, 2,...
10. h ( r ) = 2 cos π r .
11. Perform a qualitative analysis similar to example 4.7.2 to show that the steady state x = 0, x
˙ = 0 of
x¨ +bx2x˙ +kx=0

is asymptotically stable if b > 0 and unstable if b < 0. Apply these results to the system,
(Equation 4.52), in Section 4.
Bibliography
[1] N.P. Bhatia, G.P. Szeg, Stability Theory of Dynamical Systems, Springer-Verlag, 2002
[2] R. Borrelli, C. Coleman, and W. Boyce, Differential Equations Laboratory Workbook: A Collection of Experiments, Explorations
and Modeling Projects for the Computer Wiley, NY, 1992.
[3] W.E. Boyce and R. Diprima, Elementary Differential Equations and Boundary Value Problems, 3rd edition, J. Wiley and Sons,
1977.
[4] Clive L. Dym - Stability Theory and Its Applications to Structural Mechanics, Dover, NY.
[5] M.W. Hirsch and S. Smale, Differential Equations, Dynamical Systems and Linear Algebra, Academic Press, 1974.
[6] R. K. Miller and A. N. Michel, Ordinary Differential Equations, Academic Press, 1982.
[7] J. Palis, Jr. and W. de Melo, Geometric Theory of Dynamical Systems, Springer-Verlag, 1982.
[8] A. Pillay - Geometric Stability Theory, Oxford University press 1996.
[9] N. Rouche, P. Habets and M. Laloy, Stability Theory by Liapounov's Direct Method, Springer-Verlag, 1977.
 CHAPTER 5

Bifurcations and Chaos

CONTENTS
5.1 Introduction
5.2 Bifurcations of Co-Dimension One
5.2.1 Trans-critical Bifurcation
5.2.2 Saddle Point Bifurcation
5.2.3 Pitchfork Bifurcation
5.2.4 Subcritical Bifurcation (Hysteresis)
5.2.5 Hopf Bifurcation
5.3 Rossler Oscillator
5.4 Lorenz Equations
5.5 Nerve Models
5.6 Miscellaneous Topics
5.6.1 Dimension
5.6.2 Liapunov Exponents
5.7 Appendix A: Derivation Of Lorenz Equations
5.1 Introduction
In the previous chapter we discussed various methods to analyze the stability of the equilibrium states
of a dynamical system when the values of the system parameters are known and fixed. The objective
of bifurcation theory is to investigate what happens to the type, number, and stability of the steady
states as a result of a continuous change in some (or all) of the system parameters. If, as a result, the
system undergoes “sudden” change in its properties or behavior, we say that the system underwent a
bifurcation.
The motivation for such analysis stems from the fact that in many real life situations the values of
the model parameters are not known accurately or might actually be (very) slowly varying functions
of time which we approximate by constants to simplify the model Equations.
Here are some examples of “real life” bifurcations.
1. Phase transitions
If we cool down water, it will remain liquid until we reach 0 o C. If we continue the cooling
process below 0 o C, the water will undergo a “ phase transition” and turn to ice. Thus the “liquid
water system” undergoes a bifurcation at 0 o C. Similarly a bifurcation occurs when the water is
heated over 100C (at which point it turns to “gas”).
2. Earthquakes
The Earth’s continents “float” on the liquid core of the Earth. When these plates collide, stress
starts to build up, and one plate starts to glide over the other. As these stresses build up, a point
is reached where the plates break and we experience an earthquake. Thus an earthquake
represents a bifurcation in the plate’s structure.
3. Buckling of columns
If we increase the vertical load on a (vertical) column, a point will be reached where the
column will buckle and break (if the “load” is a building then the building will collapse.) Under
these circumstances the column undergoes a bifurcation.
4. Magnetic hysteresis
If we place a magnet in a weak magnetic field whose direction is opposite to its magnetization,
“nothing” will happen to the magnet. However, as we increase the field intensity, there will
come a point when all of a sudden the magnet will change its poles (the north magnet pole
becomes the southern one and vice versa). Thus a bifurcation occurs. If we reverse now the
direction of the field, another bifurcation will occur and the magnetization of the magnet will
return to its original state. These changes in the magnetization of the magnet are represented by
the “hysteresis loop” (see Fig. 5.5).

5.2 Bifurcations of Co-Dimension One

Definition: Let
(5.1)
x˙=F(x,λ),xεRn,λεRm
be a dynamical system with m dynamical parameters λ = ( λ 1 , … λ m ) .
We say that a bifurcation is of co-dimension k if this is the smallest dimension of parameter space
which contains such a bifurcation (observe that there are no restrictions on n). In particular,
bifurcations of co-dimension one require only one dynamical parameter.
The steady states of the dynamical system Equation (5.1) satisfy X = 0. Therefore, to find the
bifurcation points of this system one has to study the nature of the solutions to the Equation F ( x , λ )
= 0 as λ varies. If λ 0 is a bifurcation point, then in order to study this bifurcation it is a common
practice to reduce Equation (5.1) to a “local form” by the following transformation
(5.2)
y=x-x0,μ=λ-λ0
where x 0 is a stable steady state of the system near λ 0 .
Table 5 contains a list of all co-dimension one bifurcations and “generic” dynamical systems which
contain them.
We now discuss each of these bifurcations in detail.
Table 5.1 List of Co-dimension One Bifurcations

5.2.1 Trans-critical Bifurcation

The system
(5.3)
x˙=μx-x2=x(λ-x)
has two steady states x = 0 and x = λ for all possible values of the dynamical parameter λ . The
linearization of the system around x = 0 yields
(5.4)
x˙=λx.
It is obvious then that this state is asymptotically stable when λ < 0 and unstable for λ > 0 (since x ( t
)=Ceλt).
The linearization of the system around the steady state x = λ is obtained by introducing
(5.5)
y=x-λ
and hence
(5.6)
y˙=-y(y+λ)
For the linear part we then have
(5.7)
y˙=-λy,
i.e., this state is asymptotically stable for λ > 0 and unstable for λ < 0 . The stability of these steady
states is plotted on a “bifurcation diagram” where solid and dashed lines represent asymptotically
stable and unstable states respectively (Fig. 5.1). From this diagram we see that for both states λ = 0
is a bifurcation point. A dynamical system in the stable steady state x = 0 for λ < 0 will become
unstable at λ = 0 and will “transition” to the new stable state x = λ as λ becomes positive. In this new
state the system will usually exhibit different properties than in the state x = 0. We refer to this
phenomenon as “exchange of stability” between the steady states (a phase transition is an example of
this bifurcation).

Figure 5.1 Trans-critical bifurcation

5.2.2 Saddle Point Bifurcation

This type of bifurcation has two generic representations.
A.
(5.8)
x˙=λ-x2.
The system has two steady states x = ± λ for λ > 0 and none when λ < 0 (since x is assumed to be
real). To investigate the stability of the state x = λ we introduce y = x - λ . The system then becomes
y=-2λy-y2
Using linearization around y = 0 we infer that the state x = λ is asymptotically stable. Similarly, for x
= - λ we find that the steady state is unstable. This is represented by the bifurcation diagram of Fig.
5.2.
Figure 5.2 Saddle Point Bifurcation

B.
(5.9)
x˙=λ+x2
The system has two steady states for x = ± - λ for λ < 0 and none for λ > 0 . In this case the steady
state x = - λ is unstable while x = - - λ is asymptotically stable (see Fig. 5.3).

Figure 5.3 Saddle point bifurcation

5.2.3 Pitchfork Bifurcation

The system
(5.10)
x˙=μx-x3=x(λ-x2)
has a steady state x = 0 for all λ and two additional steady states x = ± λ for λ > 0 . The linearization
around x = 0 is
(5.11)
x˙=λx
which implies that x = 0 is asymptotically stable for λ < 0 and unstable for λ > 0 . To linearize the
system around the steady state x = λ , λ > 0 , we introduce y = x - λ to obtain
(5.12)
y˙=-2λy-3λy2-y3.
The linearization of this Equation around y = 0 yields
(5.13)
y˙=-2λy,λ>0
which implies that this steady state is asymptotically stable. Similarly, one can show that x = - λ , λ >
0 is asymptotically stable.
The name of this bifurcation comes from its diagram (Fig. 5.4) which looks like a fork.

Figure 5.4 Pitchfork Bifurcation

5.2.4 Subcritical Bifurcation (Hysteresis)

To analyze the system
(5.14)
x˙=λ+x-x3
we first note that a third order polynomial always has three roots in the complex plane. These roots
can be all real or one real and two complex conjugates. The polynomial p ( x ) = λ + x - x 3 has three
real roots on the interval λ ∈ [ - 2 9 3 , 2 9 η 3 ] and only one real root outside this interval. We
conclude therefore that the system has three steady states for λ ε [ - 2 9 3 , 2 9 3 ] and only one
outside this interval (See Fig. 5.5). To analyze the stability of these steady states we let λ = 0 . At this
point the three steady states are x = 0, x = ± 1. For x = 0 the linearization of the dynamical system is x
˙ = x , and we conclude that this state is unstable. For x = 1 we introduce y = x - 1 and linearize. This
yields
 (5.15)
y˙=-2y

Figure 5.5 Subcritical Bifurcation

and hence the steady state is asymptotically stable. Similarly one can show that x = - 1 is
asymptotically stable. While this analysis has been carried out only for λ = 0 , it is possible to show
that it is valid for other values of λ in the interval [ - 2 9 3 , 2 9 3 ] (as shown in Fig. 5.5). Thus on
this interval the system has one unstable state and two asymptotically stable states.
For values of λ which are outside the interval [ - 2 9 3 , 2 9 3 ] , it is straight- forward to show that
the corresponding steady state is asymptotically stable. For example, if λ = 6 , then the corresponding
steady state’s x = 2. Applying the transformation to local form Equation (5.2) to (Equation 5.14), viz.
μ=λ-6,y=x-2
we obtain
y = μ - 11 y - 6 y 2 - y 3
The linearized form of this Equation at μ = 0 is y ˙ = - 11 y which shows that the state y = 0 (viz.
x = 2) is asymptotically stable.
If initially λ < - 2 9 3 , then the system must be in the lower (stable) state. When λ becomes larger
than - 2 9 3 , the system acquires a second stable steady state but since the lower steady state is
stable, it will remain in its original state. However, when λ becomes larger than 2 9 3 , the lower
steady state ceases to exist and the system will “jump” to the upper stable steady state and will
remain there for all λ > 2 9 3 . A similar picture unfolds when λ is decreasing.
If λ > 2 9 3 , the system will be in the upper steady state and will remain in this state until λ = - 2 9
3 when this state ceases to exist. The actual state of the system is described therefore by the hysteresis
diagram (Fig. 5.5).

5.2.5 Hopf Bifurcation

The two dimensional system
x˙=-y+λx-x(x2+y2)
 (5.16)
y˙=x+λy-y(x2+y2)
has an obvious steady state x = ( x , y ) = 0 for all λ . The existence of another attractor of dimension
one can be inferred by recasting this system in polar coordinates ( r , θ ) using:
(5.17)
rdrdt=xdxdt+ydydt
(5.18)
-r2dθdt=ydxdt-xdydt
This leads to
(5.19)
rdrdt=r2(λ-r2)
(5.20)
-r2dθdt=-r2.
or
(5.21)
drdt=r(λ-r2)
(5.22)
dθ/dt=1.
This implies that the system has a limit cycle for
(5.23)
r=x2+y2=λ,λ>0.
We discuss now the stability of these steady states.
A. x = 0
The linearization of the system around this steady state is
(5.24)
x˙=-y+λx,y˙=x+λy
or in a matrix form,
(5.25)
x˙=λ-11λx.
The eigenvalues of the coefficient matrix in Equaton (5.25) are
(5.26)
μ=λ+±i.
The general solution of this linear system is
(5.27)
x=eλt[c1v1eit+c2v2e-it]
where v 1, v 2 are the eigenvectors which correspond to the eigenvalues λ ± i and c 1, c 2 arbitrary
constants. We conclude then that the steady state x = 0 is asymptotically stable for λ < 0 and unstable
for λ > 0 .
B. To explore the stability of the limit cycle r = λ we observe that for r < λ , d r d t is positive.
However, when r > λ , d r d t is negative. We conclude therefore that the limit cycle r = λ is
asymptotically stable.
If we now plot the eigenvalues, (Equation 5.26), in the complex plane (Fig. 5.6) we can conclude
that the steady state x = 0 becomes unstable as λ becomes positive and a new asymptotically stable
limit cycle appears when these complex conjugate eigenvalues cross from the left to the right side of
the complex plane. In fact, this is the essence of the Hopf bifurcation theorem (1950) which states that
whenever the linearization of a dynamical system around a steady state has two complex conjugate
eigenvalues which behave in this manner then the original steady state becomes unstable and an
asymptotically stable limit cycle appears. Such a bifurcation is called “Hopf bifurcation.” A complete
statement of this important theorem is as follows.

Figure 5.6 Hopf bifurcation: a steady state destabilizes as two conjugate eigenvalues move into the
the right hand side of the complex plane

Figure 5.7 Trajectory of the Rossler oscillator starting near x 1 with c = 4

Figure 5.8 Same trajectory in 3D

Figure 5.9 Trajectory of the Rossler oscillator starting near x 1 with c = 6

Figure 5.10 Same trajectory in 3D

Figure 5.11 Trajectory of the Rossler oscillator starting near x 1 with c = 8.2
Figure 5.12 Same trajectory in 3D

Figure 5.13 Trajectory of the Rossler oscillator starting near x 1 with c = 15

Figure 5.14 Same rajectory in 3D

Theorem:(Hopf Theorem) Assume that the system

(5.28)
x˙=F(x,μ),x∈ Rn,λ∈ R
where
(5.29)
F(x)=f1(x)⋮ fn(x)
has a steady state ( x ( μ 0 ) , μ 0 ) and at this state
1. the Jacobian of the system
J(x(μ0))=[∂fj∂xi](x(μ0))
has a simple pair of (pure) imaginary eigenvalues ± iν and all other eigenvalues of J have negative
real part.
2. The two eigenvalues of J(x(\mu)) which are imaginary at μ = μ 0 are smooth functions of μ and
satisfy
d(Reν(μ))dμ>0atμ=μ0.
(That is, as a function of μ these two eigenvalues move into the right half of the complex plane thus
destabilizing the steady state ( x ( μ 0 ) , μ 0 ) ; this is called “ the transversality condition”.)
Then μ 0 is a bifurcation point of the system into a limit cycle.
As an example of such a bifurcation we consider the following physical system.
Example: Brusselator Reaction.
There are several chemical reactions in which the color of the reactant in the chemical reactors
changes periodically. Examples of such reactions are the Belousov-Zhabotinsky reaction, the Bray-
Liebhafsky reaction, and the Brusselator reaction. These color changes are due to the fact that the
concentration of some of the intermediate chemicals in the reaction oscillates in time.
A mathematical explanation for this “oscillatory behavior” is given by the following model for the
Brusselator reaction. In this reaction the color of the reactants in the chemical reactor alternates
periodically between blue and red. The model postulates the following sequence of (catalytic)
reactions to explain this phenomenon:
a→ X
b+X→ Y+c
(5.30)
2x+Y→ 3X
X→ d.
Here a, b, c, d are the initial and final chemicals of these reactions. Observe that the sum of all these
reactions is a + b → c + d. Thus X, Y are catalysts whose presence and concentration oscillate as is
indicated by the color of the reactants in the reactor.
In the following we use the same letter to denote the chemical and its concentration and assume that
the rate of a reaction is proportional to the concentration of the reactants. To derive Equations for the
rate of change of X, Y we observe that the production rates of X in the first and third reactions are
proportional to a and X 2 Y while its rate of loss in the second and fourth are bX and X. Hence
(5.31)
dxdt=a-(b+1)X+X2Y.
Similarly, the rate of production of Y in the first reaction is proportional to A while the loss of Y in the
third reaction is proportional to X 2 Y. Hence
(5.32)
dYdt=bX-X2Y.
Renormalizing the concentration of a to 1, we deduce that the system, (Equations 5.31) - (5.32), has
an equilibrium state X = 1, Y = b. Reducing (Equations 5.31) - (5.32) to a local form by the
transformation
(5.33)
X=1+U,Y=b+V
we obtain
(5.34)
U˙=(b-1)U+V+U2(b+V)+2UV
(5.35)
V˙=-bU-V-U2(b+V)-2UV.
(Here b plays the role of the “bifurcation parameter” λ ). Hence at the equilibrium state (0, 0) we
have
(5.36)
J(0,0,b)=b-11-b-1
whose eigenvalues are
(5.37)
ν=b-2±(b-2)2-42.
We see that when 0 < b < 2, (0, 0) is a stable spiral point. At b = 2 the Jacobian has two pure
imaginary eigenvalues which cross to the right half complex plane and all the other conditions of
Hopf bifurcation theorem hold. We conclude then that b = 2 is a Hopf bifurcation point for the system
i.e., for b > 2 the system has a stable limit cycle and the concentrations of X, Y will vary periodically
with time.
Exercises
1. Let
x˙=x(a-x-λy)
y˙=y(a-y-2x)
where a > 0 is a constant and λ is a parameter.
(a) Interpret the dynamics of the ecosystem that is described by these Equations (as a model for
interaction between species).
(b) Classify the equilibrium states and bifurcation points of this system as a function of λ
2. Classify the bifurcation of the system
x˙=x(λ+a-x)
where a is a constant and λ is a parameter.
3. Show that the following system undergoes a Hopf bifurcation
x˙=λx+y+x2y
y˙=-x+λy+y3
4. Study the stability of the steady states and bifurcation diagram for the system
x˙=x+4y2
y˙=λy+2xy+4y3
5. Draw the bifurcation diagrams for the following systems
(a) x ˙ = ( λ x - x 3 ) ( λ + 2 x - 1 )
(b) x ˙ = ( λ x - x 3 ) ( λ - x 2 )
(c) x ˙ = x ( 1 - x ) ( λ - x )
6. Show that the following system
x¨+(1+λ2+x2)x˙+2x-x3=0
has a Hopf bifurcation at λ = 0 . Draw the phase diagram of the system for different values of λ .

5.3 ROSSLER OSCILLATOR

There is no universal agreement about the definition of chaos. However, “an approximate” definition
of a chaotic system was proposed by E. Lorenz in his 1963 seminal paper as follows:
“When the present determines the future, but the approximate present does not approximately
determine the future we say that the system is chaotic.”
In other words the evolution of a chaotic system is sensitive to the values of the initial conditions,
and a small variation in these conditions leads (in the long run) to a completely different trajectory of
the system. One usually refers to this phenomenon as “sensitivity to the initial conditions.”
It was proposed that one possible route for a deterministic system to be- come chaotic is when it
undergoes a cascade of bifurcations. The Rossler Oscillator (or attractor) was discovered in an
attempt to find a “simple” system which exhibits such behavior. The system consists of the following
three Equations:
(5.38)
x˙=-y-z,y˙=x+ay,z˙=b+(x-c)z
where a, b, c are non-negative constants (or parameters). Originally this system was proposed as a
theoretical model; however, since then it found applications in the modeling of some chemical
reactions.
 We observe that although this system is nonlinear, the nonlinearity is due to a single term xz in the
third Equation. Therefore, if we let b = 0 and consider the system in the plane z = 0, we obtain a
linear system
(5.39)
x˙=-y,y˙=x+ay.
When 0 < a < 2 the solution of the system is
x ( t ) = e a t 2 [ ( - α D 1 - a D 2 2 ) sin ( α t ) + ( - a D 1 2 D 1 + D 2 α ) cos ( α t ) ]
y ( t ) = e a t 2 [ D 1 cos ( ( α t ) + D 2 sin ( α t ) ]
where α = 4 - a 2 . This implies that the steady state (0, 0) is unstable spiral. Similarly when a > 2,
(0, 0) is unstable focus.The three dimensional system, (Equation 10.89), with c 2 - 4ab > 0 has two
steady states
x1=(c-β2,-c-β2a,c-β2a)
x2=(c+β2,-c+β2a,c+β2a)
where β = c 2 - 4 a b . To determine the stability of these two steady states,
we compute the Jacobian of the system
(5.40)
J(a,b,c)=0-1-11a00zx-c.
The characteristic polynomial p ( λ ) of the Jacobian is
p(λ)=λ3-(a+x-c)λ2+[(a(x-c)+1+z]λ-(x-c)-az.
For a = 0.1, b = 0.1, c = 4, we have
x 1 = ( 0.0025 , - 0.0250 , 0.0250 ) , x 2 = ( 3.997 , - 39.975 , 39.97 )
Thus the two steady states are well separated from each other. The corresponding eigenvalues of J at
these points are
(5.41)
λ ± = 0.047 ± 0.999 i , λ 3 = - 3.992
(5.42)
λ ± = - 2.973 × 10 - 7 ± 6.401 i , λ 3 = 0.097 .
It follows then that both steady states are unstable. We now observe that eigenvalues with negative
real parts attract in the direction of the corresponding eigenvector while eigenvalues with positive
real parts repulse in that direction. For a (general) trajectory that starts near x 1 the directions along
the eigenvectors which correspond to λ ± will cause the trajectory to spiral away from x 1. However,
as the trajectory moves away from x 1 and the x - y plane, the direction along the third eigenvector
will become dominant (since the real part of λ ± is small) and the trajectory will curve back to x 1.
To illustrate the bifurcation cascade which the Rossler oscillator undergoes as the parameters in
Equation(10.89) change, we let a = b = 0.1 and allow c to change. To illustrate the results we plot
below the trajectories which start near x 1 with c = 4, 6, 8.2 and 15. For c = 4 we obtain a trajectory
which, after a transient state, settles to a one cycle trajectory. For c = 6 the system undergoes a
bifurcation and we obtain a two cycle trajectory. For c = 8.2 it is a four cycle trajectory. Finally for
c = 15 the trajectory is fully chaotic.
5.4 Lorenz Equations
Atmospheric flow in general is governed by a set of complicated nonlinear partial differential
Equations. In 1963 Lorenz “projected” these Equations (see appendix) to obtain a highly simplified
set of three ordinary differential Equations
x˙=σ(y-x)
y=rx-y-xz
(5.43)
z˙=-bz+xy.
In these Equations σ, b are positive constants and r is a (dynamical) parameter. These Equations,
which are called “Lorenz Equations,” since their discovery have played a major role in the theory of
nonlinear dynamical systems. In fact, they led to the establishment of new areas of research such as
“chaos theory,” “strange attractors,” and other related topics.
In this section we address the steady states of this system, their stability, and bifurcation as r
changes.
To obtain the steady states of (Equation 5.43) we must solve the algebraic system
(5.44)
σ(x-y)=0
(5.45)
rx-y-xz=0
(5.46)
-bz+xy=0.
Since σ ≠ 0 we obtain x = y from (Equation 5.44). Substituting for y in (Equation 5.45) we have
(5.47)
x(r-1-z)=0.
Thus either
(5.48)
x=(x,y,z)=0
or
(5.49)
x=y,z=r-1.
From (Equation 5.46) we deduce then that two other steady states exist for r ≥ 1
(5.50)
x=(±b(r-1),±b(r-1),r-1)
(Observe that for r < 1 we obtain complex values for x, y which are not admissible.) A. Stability of
the steady state x = 0
The linearization of the system, (Equation 5.43), around this steady state yields
(5.51)
x˙=-σσ0r-1000-bx.
The characteristic polynomial of the coefficient matrix in this system is
 (5.52)
p(λ)=(λ+b)[λ2+(1+σ)λ-σ(r-1)].
Hence the eigenvalues are
λ1=-b
(5.53)
λ±=-1+σ2±(1+σ2)2+σ(r-1).
We infer that if r < 1, all the eigenvalues are negative and therefore the steady state is asymptotically
stable. However, if r > 1, one of the eigenvalues is positive and therefore x = 0 is unstable.
B. The steady states
x=(±b(r-1),±b(r-1),r-1),r≥1.
Using Taylor expansion around these points one can show that in both cases the linearization of
(Equation 5.43) around these steady states is given by
(5.54)
X˙=-σσ01-1-ααα-bx,α=±b(r-1).
The characteristic polynomial of the coefficient matrix is
(5.55)
p(λ)=λ3+(1-σ+b)λ2+b(σ+r)λ+2σb(r-1)
For r = 1 the roots of this polynomial are
(5.56)
λ1=0,λ2=-(1+σ),λ3=-b.
Thus we conclude that at this point the “new” states that appear for r ≥ 1 are stable. For 0 < r - 1 ≪ 1
the first root of Equation (5.55) can be approximated by
(5.57)
λ1≈2σ(1-r)1+σ<0,
i.e., the states are asymptotically stable.
To summarize for r < 1 the steady state x = 0 is asymptotically stable, but it becomes unstable for
r > 1. At r = 1 two new steady states appear, and they are asymptotically stable. It follows then that
r = 1 is a pitchfork bifurcation point for the Lorenz system, (Equation 5.43).
To demonstrate the sensitivity to initial conditions of Lorenz (Equations 5.43) we simulated these
Equations with σ = 10, r = 28 and b = 8/3 with two sets of initial conditions ( x , y , z ) = ( 5 , 5 , 5 )
and ( x , y , z ) = ( 5 , 5.001 ; 5 ) . (The values assigned t o σ , r a n d b a r e those used by Lorenz i n
h i s 1963 paper).
Although these initial conditions are close to each other, the difference between the solutions for
the y variable over the time interval [0,100] spikes to more than 30 (see Fig. 5.15). The
corresponding relative difference is over 100%.

Figure 5.15 Difference between the y-solutions to Lorenz Equations

5.5 NERVE MODELS

A cell in the nervous system of a living organism is called a neuron. A simplified picture of a neuron
is that it is composed of a “core” (referred to as “soma”), dendrites which act as receptors for signals
from other neurons, and a “long extension” which is called an axon. The axon ends with synapses
through which signals are transmitted to other neurons. Under proper stimulation the electric potential
in the axon membrane rises above a certain threshold and the axon “fires” electrical impulses to
neighboring cells to communicate with them through the synapses. The basic model for this firing
process is due to
Hodgkin-Huxley. It consists of four ODEs and depends heavily on the modeling of the electro-
chemical properties of sodium and potassium ions in the axon. To simplify this model conceptually
FitzHugh-Nagumo introduced a model which consists of two ODEs.
(5.58)
v˙=v-13v3-w+Iext
(5.59)
W=ε(βv-γw+δ)
where v is the (axon) membrane potential, w is a recovery variable that pro- vides negative feedback
(that is provides damping to the amplitude of the membrane potential), and I ext is a control parameter
which represents the current due to the external stimulus. The parameters ɛ, β, γ, δ are constants.
These parameters can change the equilibrium (i.e. rest) state and dynamics of the model. In the
FitzHugh-Nagumo model the values of these constants are usually set to ɛ = 0.08, β = 1, γ = 0.8 and
δ = 0.7. The cubic term for v allows for regenerative self-excitation of the membrane potential.
Observe that this model neglects the spatial dependence of the signal; i.e. v, w depend only on time.
Such a model is usually referred to as “space clamped.”
At equilibrium v ˙ = 0 and w ˙ = 0 . Hence the equilibrium state is given by the intersection of the
curves
w=v-13v3+Iext,w=βv+δγ
in the phase plane (v,w). This equilibrium is unstable if I ext is strong enough. In this case the model
exhibits periodic spiking activity.
Numerical simulations of this model show that weak stimuli (small amplitude pulses of I ext ) result
in small-amplitude closed trajectories in phase space around the equilibrium point. Biologically this
corresponds to sub- threshold responses (no firing). On the other hand strong stimuli result in large-
amplitude trajectories in phase space that correspond to a firing spike. Thus the model does not have
a well-defined firing threshold.
A more general form of this model replaces (Equation 12.30) by
(5.60)
v˙=f(v)-w+Iext.
The Equations of this model are used also to model the propagation of waves in excitable media, such
as heart tissue or nerve fiber. Thus if we let v = v ( x , t ) , w = w ( x , t ) and add a diffusion term Dv
xx to (12.32) we obtain the reaction-diffusion system
(5.61)
vt=f(v)-w+I+Dvxx
(5.62)
wt=ε(βv-γw+δ)
where f(v) = v(a - v)(v - 1) . In this model a represents the threshold for excitation, and ɛ the
excitability of the medium.
 To gain some insight into the general properties of this model we let D = 1, δ = 0, and I = 0. With
this simplification the model can be rewritten as
(5.63)
vt=vxx+v(a-v)(v-1)-w
(5.64)
wt=bv-dw.
We assume that 0 < a < 1, and b, d ≥ 0.
A traveling wave solution v = f (x-ct), w = g(x - ct) of Equations(12.35) and (12.36) will have to
satisfy
(5.65)
-cf′-f′′=f(a-f)(f-1)-g,-cg′=bf-dg.
where primes denote differentiation with respect to z = x - ct. We can rewrite these Equations as a
system of first order Equations as follows:
(5.66)
f′=h,h′=-ch-f(a-f)(f-1)+g,cg′=-bf+dg
If d ≠ 0 the equilibrium points of this system are at
f=αi,h=0,g=bαid,i=1,2,3
where α i are the roots of the third order Equation
x[(a-x)(x-1)-bd]=0.
The root α 1 = 0 will be the unique real root (and hence the only equilibrium point of the system) if
(1-a)2<4bd.
On the other hand, if d = 0 then the only equilibrium state of the system is (0, 0, 0,).
The Jacobian of the system, (Equation 9.69), at (0, 0, 0) is
(5.67)
J(0,0,0)=010a-c1-b0d
and the stability of this equilibrium point will be determined by the eigenvalues of this matrix.
In the special case b = 0 the system, (Equation 9.69), decouples, and we can solve for g
g = A exp ( d z c ) .
Therefore, if we are looking only for bounded solutions of the system, we must let A = 0, i.e. g = 0.
The remaining system for (f,h) has three equilibrium points (0, 0), ( 1 , 0 ) , and ( a , 0 ) . The first
two equilibrium points are saddle- points while (a,0) is a center if c = 0 and unstable spiral or node
if c ≠ 0.

5.6 MISCELLANEOUS TOPICS

5.6.1 Dimension
We were all taught that a point has 0-dimension, a line is an entity of one dimension, a plane is two-
dimensional, and that “real space” has three dimensions.
We now try to formalize these intuitive concepts.
 Consider the unit interval [0, 1 ] in R and consider measuring sticks of length ε = 1 2 n , n = 1,2,…
. The number M(ɛ) of such sticks that is needed to cover the interval [0, 1] is 2 n . Hence
(5.68)
D = lim ε → 0 ln M ( ε ) ln ( 1 ε ) = 1 .
(Here we dropped the requirement ε = 1 2 n ). Thus the expression, (Equation 5.68), yields a number
that corresponds to our intuitive notions about the dimension of a line.
Now consider the unit square [0, 1 ] × [0, 1 ] in the plane. Instead of a measuring stick we need
here a “ measuring square” with sides of 1 2 n . When n = 1
we need 4 such squares to cover [0, 1] × [0, 1]. For n = 2 we need 16 squares, and for ε = 1 2 n we
need 22n squares. Hence
(5.69)
D = lim ε → ∞ ln M ( ε ) ln ( 1 ε ) = 2 .
In three dimensions we consider the unit cube with “measuring cubes” whose sides are of length ε = 1
2 n . With n = 1 we need 8 such cubes to cover the unit cube. With n = 2 we need 64 cubes, etc.
Hence
(5.70)
D = lim ln M ( ε ) ln ( 1 ε ) = 3 .
Thus it seems appropriate to define the dimension of a metric space as
(5.71)
D = lim ε → 0 ln M ( ε ) ln ( 1 ε )
where M(ɛ) is the number of “cubes” with side ɛ that are needed to cover the unit cube in such a
space. The dimension D defined by (Equation 5.71) is called the “Hausdorff dimension” of the space.
We now apply this definition to an “exotic” set and compute its dimension.
Example 5.6.1 Cantor Set
To define the Cantor set we consider the interval I = [0, 1]. We now define
(5.72)
C1=I-(13,23)=[0,13]∪ [23,1],
i.e. C 1 is the subset that is obtained from the unit interval by removing the middle (open) 1 3 of I.
Now define C 2 as
(5.73)
C2=C1-(19,29)-79,89,
i.e. we remove again the middle 1 3 of each of the subintervals of C 1.
Similarly we define
(5.74)
C 3 = C 2 - ( 1 27 , 2 27 ) - ( 7 27 , 8 27 ) - ( 19 27 , 20 27 ) - ( 25 27 , 26 27 )
and so on. The Cantor set is now defined as the set remaining at the “end” of such interval divisions
and subtractions, i.e. (See Fig. 5.16)
(5.75)
C=∩ i=1∞Cn.
Figure 5.16 Cantor set

To show that C is not empty (in fact it is uncountable) we use the number representation with base
3. That is, a number 0 ≤ x ≤ 1 is represented by
(5.76)
x=.a1a2…,0≤ai≤2.
(That i s , t h e a i a r e integers which can take the values 0, 1, 2) where the a i s are determined by
the relation
(5.77)
x=∑n=1∞an3n.
Remark 5.6.1 The decimal and binary representation of a number 0 ≤ x ≤ 1 is a short hand notation
for
(5.78)
x = ∑ n = 1 ∞ α n 10 n , x = ∑ n = 1 ∞ β n 2 n
respectively where α i ε { 0 , 1 … 9 } , β i ɛ{0, 1 }.
It is now clear from its construction that the Cantor set consists of all numbers in base 3 whose
representation contains only the digits {0, 2 }; i.e. a i ≠ 1 for all i.
What is the Hausdorff dimension of the Cantor set?
To answer this question we consider a yard stick of length ε = 1 3 . It is then obvious that M ( 1 3 )
= 2 . Similarly for ε = 1 9 we have M ( 1 3 2 ) = 2 2 . In general we then have M ( 1 3 n ) = 2 n and
hence
(5.79)
D = lim ln M ( ε ) ln ( 1 ε ) = ln 2 ln 3 .
Thus the Cantor set is a set with non-integer Hausdorff dimension.
Example 5.6.2 Sierpinski Triangle
To construct this set we start with an equilateral triangle S 0 with sides of length one. In the first
iteration we split it into four equal-sided triangles (as shown in Fig. 5.17) and remove the one in
the center. This yields S 1. By iterating this procedure on each of the three remaining triangles we
obtain S 2 and so on. The Sierpinski triangle is the limit set that remains after all these divisions
and removals.
Figure 5.17 First two iterations toward the creation of the Sierpinski Triangle

S = lim n → ∞ S n .
To find the Hausdorff dimension of S we note that for S 0 we have one triangle and need a
measuring stick with length ɛ = 1. For S 1 we have three triangles and ε = 1 2 . For S 2 we have nine
triangle and ε = 1 4 . Hence
D = lim ε → 0 ln M ( ε ) ln ( 1 ε ) ≡ lim n → ∞ ln 3 n ln 2 n = ln 3 ln 2 .
Thus the dimension of this “triangle” is greater than one but less than two. To compute the area of
S we observe that S 0 has area 3 4 . S 1 has an area of 3 4 . 3 4 (since 1 4 th of the area has been
removed). S 2 has an area of ( 3 4 3 4 ) · 3 4 (since we removed once again 1 4 of the area of S 1). In
general we then have
areaofSn=34(34)n
and hence
area of S = lim n → ∞ (area of S n ) = 0.

5.6.2 Liapunov Exponents

Let
(5.80)
x˙=F(x),xεRn
be a dynamical system, and let x(t) be the trajectory of the solution with initial condition x(0) = x 0.
Now consider another solution y(t) of this system with initial condition y(0) = x 0 + δx 0.
Definition:
(5.81)
α = lim t → ∞ 1 t ln | | y ( t ) - x ( t ) | | | | δ x 0 | |
is called a Liapunov exponent of the dynamical system (Equation 5.80). It is obvious from the
definition that different δx 0 might lead to different Liapunov exponents.
Intuitively, Liapunov exponents tell us how the solution of system (5.80) “reacts” to small changes
in the initial conditions. If the Liapunov exponents of the system are all negative, then small changes
in the initial conditions cause small changes in the trajectory, and one can make predictions about the
evolution of the system even if the initial conditions are not known exactly. However if at least one
Liapunov exponent of the system is positive, then (random) small changes in the initial conditions
might lead to diverging trajectories and hence no long term predictions about the evolution of the sys-
tem can be made. Thus such a system is sensitive to the values of the initial conditions and hence is
called “chaotic.”
This issue (of sensitivity to the initial conditions) is important from a practical point of view since
in most cases the initial conditions are known only with a certain margin of error. If all the Liapunov
exponents of the system are negative such errors are not significant as the approximate trajectory of
the system (using inexact initial condition) remain close to the true one. However, if at least one of the
Liapunov exponents is positive even minute errors in the initial condition will lead (in the long run)
to large predictive errors regarding the behavior of the system.
Example 5.6.3 The weather system of the Earth is chaotic. As a result one cannot make long term
predictions about its behavior.
Example 5.6.4 The Lorenz system is chaotic (see exercises).
Exercises
1. Carry out in detail the calculations needed to establish the stability of the various steady states
for the Lorenz model Equations.
2. Simulate Lorenz Equations with
σ = 10 , r = 28 , b = 8 / 3
(These were the original parameters used by Lorenz in his 1963 paper) subject to the initial
conditions x = (6, 6, 6), x = (6.01,6,6), x = (6.001, 6, 6) over the time interval [0, 250].
(a) Plot the solutions of x(t) on the same graph and compare (do the same for the solutions of y(t)
and z(t)).
(b) Plot the difference between the solutions of x(t) on the same graph and compare (do the same
for the solutions of y(t) and z(t)).
(c) Plot x vs. y and x vs. z. Explain what you see.
(d) Compute an approximation for one of the Lyaponuv exponents of the system at x = (6, 6, 6)
based on the simulations of this problem. 3. Simulate the Lorenz system for other values of r‘s e.g., 1,
5, 10, 20 and make similar plots to those in Ex. 2b. ANALYZE your results.

5.7 Appendix A: Derivation Of Lorenz Equations

Basic Model
Atmospheric convection in two dimensions is described by the following system of partial
differential Equations:
(A.1)
∂∂t∇ 2ψ=-∂(ψ,∇ 2ψ)∂(x,z)+ν∇ 4ψ+αg∂θ∂x
(A.2)
∂θ∂t=-∂(ψ,θ)∂(x,z)+△ TH∂ψ∂x+κ∇ 2θ
Subject to the stress free boundary conditions:
(A.3)
ψ=0,Δ2ψ=0,δ=0,z=0,1.
In this system ψ is the stream function and θ is the potential temperature
θ=T(p/p0)R/cp
where T is the temperature in Kelvins, p is the pressure, and p 0 is the pressure at sea level. The
constants α, ν, κ, g denote respectively the coefficients of thermal expansion, the kinematic
viscosity, the thermal conductivity, and the acceleration of gravity. H is the fluid layer thickness, and
▵ T is the temperature difference between the upper and lower surface of the fluid (which is assumed
to be held constant). We also have
(A.4)
∂(f,g)∂(x,z)=∂f∂x∂f∂z∂g∂x∂g∂z.
In his famous 1963 paper Lorenz introduced and studied a model in which the solution to (Equations
A.1)-(A.2) is approximated by a three Fourier modes
(A.5)
ψ = κ ( 1 + a 2 ) 2 a X ( t ) sin ( π a x H ) sin ( π z H )
(A.6)
θ = R c π R a { 2 Y ( t ) cos ( π a x H ) sin ( π z H ) - Z ( t ) sin ( 2 π z H ) }
where a is a parameter and
(A.7)
Ra=gαH3△ Tκν
(A.8)
Rc=π4(1+a2)3a2
are the Rayleigh number and the critical Rayleigh numbers for the flow. This led to the following
three coupled Equations for X, Y, Z.
(A.9)
X˙=-σX+σY
A. 10
Y˙=XZ+rX-Y
Z˙=XY-bZ
where
(A.10)
σ=vk,b=4(1+a2),r=RaRc.
These Equations are usually referred to as the (Lorenz model.’ Since its appearance this model and its
implications have been studied in great detail in hundreds of publications with special attention to its
bifurcations as a function of the parameters σ, r, and b. It has been recognized, however, that (as
expected) the approximation of the solution to (Equations A.1) and (A.2) which is provided by (A.4)-
(A.10) becomes ‘poor’ as r increases. This has led several authors to develop and study models with
a larger number of modes [4,21].
Higher Order Models
 In an attempt to relate the leading Lyaponuv exponent to the turbulence strength in the flow we
consider a truncated expansion of (Equations A.1)-(A.2) with 3 and 6 Fourier components. For the
six components model we let
(A.11)
ψ = 2 ( 1 + a 2 ) κ a { [ X 1 ( t ) sin ( π a x H ) + X 2 ( t ) sin ( 2 π a x H ) ]
sin ( π z H ) + X 3 ( t ) sin ( π a x H ) sin ( 2 π z H ) }
θ = 2 △ T R c π R a [ Y 1 ( t ) cos ( π a x H ) + Y 2 ( t ) cos ( 2 π a x H ) ]
(A.12)
sin ( π z H ) - △ T R a π R c Z ( t ) sin ( 2 π z H )
which leads after some lengthy algebra to the following six Equations for X 1, X 2, X 3, Y 1, Y 2, Z
(A.13)
X˙=σY1-σX1+94(a2-1)21+a2X2X3
(A.14)
X˙=2σ(1+a2)1+4a2Y2-σ(1+4a2)1+a2X2+924(1+4a2)X1X3
(A.15)
X˙=-σ(4+a2)1+a2X3-9a224(4+a2)X1X2
(A.16)
Y˙=342X3Y2-X1Z-Y1+rX1
(A.17)
Y˙=-342X3Y1-(1+4a2)1+a2Y2-2X2Z+2rX2
(A.18)
Z˙=X1Y1+2X2Y2-bZ
 CHAPTER 6

Perturbations

CONTENTS
6.1 Introduction
6.3 Regular Perturbations
6.4 Singular Perturbations
6.5 Boundary Layers
6.1 INTRODUCTION
When we model a complex system, we create at first a “prototype model.” This model takes into
account the “major features” of the system at hand and ignores other aspects which might have “small
impact” on the system evolution. In other words the prototype model reduces the system to its “bare
bones.” In many cases these prototype models lead to equations with closed form solutions, viz.
solutions in terms of analytical formulas. However, when refinements of the prototype model are
needed, new (nonlinear) terms have to be added, and in many cases the resulting model equations
have no closed form solutions. At this stage one realizes that the additional terms in the refined model
might be “small” when compared to original terms in the prototype model and attempts to take
advantage of this fact. The essence of perturbation theory and techniques is to develop methods that
yield at least approximate solutions for these refined models. As a first step in this approach one has
to rewrite the model equations in non-dimensional form. That is, the equations have to be rewritten in
a form that is independent of the physical units that are being used. In this form one can see
immediately which terms are small when compared to other terms in the equations and thereby treat
them as a “perturbation.”
It might be argued that with wide availability of fast computers and appropriate software one can
solve these complicated (nonlinear) equations (of the refined model) numerically. However, in
general, it is not straightforward to gain insights about the evolution of the system from the numerical
solution. Furthermore, the numerical solution of a complex system of nonlinear differential equations
is always a challenging problem. Algorithms to solve such a system might not converge, and even if
they converge, they might yield incorrect solutions.
In the following we present some of the basic techniques of perturbations theory, but for a more
comprehensive treatment we refer the reader to books on this topic.
When we model a physical system, the variables and the value of the coefficients that appear in the
model equations depend on the physical units being used. For example the length unit might be
centimeter, meter, kilometer, light year (i.e. the distance that light travels in one Earth year) and so on.
It is obvious then that the use of different units will impact the value of the parameters and
coefficients that appear in the model equation. In order to overcome this issue one attempts to find
“characteristic values” for the variables that appear in the equations and combinations of these values
so that the equations are expressed in terms of “pure numbers” (independent of the physical units). We
illustrate this process by a few examples.
Example 6.2.1
Consider Newton’s second law
(6.1)
md2xdt2=F.
To rewrite this law in non‐dimensional form let us assume that the characteristic values of mass \,
length, and time in this equation are M, L, and T respectively. The force term F in this equation
has a dimension of m a s s · l e n g t h t i m e 2 . Therefore, we define the following dimensionless
quantities,
(6.2)
m¯=mM,x¯=xL,t¯=tT,F¯=FT2ML.
Then the left hand side of Equation (12.1) becomes
md2xdt2=Mm-d2(Lx¯)d(T2t¯)=MLT2m-d2X¯dt¯.
At the same time the right hand side of Equation (12.1) yields
F=MLT2F¯.
Therefore, Equation (12.1) takes the following form
(6.3)
m-d2x¯dt¯=F¯.
Although Equations (12.1), (12.3) look “similar, “ the quantities in (12.3) are dimensionless (i.e.
independent of the physical units).
To see why this non‐dimensionalization process is important we the discuss the following
example.
Example 6.2.2
Suppose we consider the following modification of Equation (12.1)
(6.4)
md2xdt2=F+α(dxdt)2.
The question naturally arises as to the impact of the additional term on the solution. Is it “small”
or “large” compared to the other terms of the equatio n ? To answer this question we introduce the
same non‐dimensional quantities as in (12.2), and the additional term in Equation (12.4) becomes
(dxdt)2=(d(Lx¯)d(Tt¯))2=L2T2(dx¯dt¯)2
Therefore, Equation (12.4) leads to
(6.5)
MLT2m-d2x¯dt¯=MLT2F¯+αL2T2(dx¯dt¯)2
Hence
(6.6)
m-d2X¯dt¯=F¯+αLM(dx¯dt¯)2.
It follows then that the size of the additional term in Equation (12.4) (as compared to other terms
in this equation) is not determined by α alone but by the ratio α ¯ = α L M . If α ¯ ≪ 1 , we can
treat the additional term in Equation (12.4) as a small perturbation. On the other hand if α ¯ ≈ 1 ,
such a treatment will be inappropriate.
Remark
Another way to look on this issue is to realize that α is not a dimensionless number since each of
the terms in Equation (12.4) must have the same dimension. TherefO re α must have the dimension
of m a s s l e n g t h . Hence the dimensionless form of this constant is α L M .
Additional examples of the non‐dimensionalization process are presented in Chapters 9 and .

6.3 REGULAR PERTURBATIONS

To illustrate how the regular perturbation procedure is used to obtain an approximate solution of an
equation with a term which is “small” compared to other terms, we consider the following example.
Figure 6.1 Difference between the approximate and exact solutions of (Equation 12.23)

Example 6.3.1
Find an approximate solution to the following equation,
(6.7)
dydt+ky+εy3=0,y(0)=2.
Here we assume that the equation is written already in non‐dimensional form and ɛ ≪ 1.
Solution 6.3.1
If ɛ = 0 the solution of Equation (12.23) is
(6.8)
y(t)=2e-kt.
To find approximate solution of (12.23) when ɛ ≠ 0 we write the desired solution y ( t , ε ) in the
form of a power series around ɛ = 0viz.
(6.9)
y(t,ε)=y0(t)+εy1(t)+ε2y2(t)+…
Substituting this expansion in Equation (12.23) and collecting terms with the same power of ɛ
yields,
(6.10)
[dy0dt+ky0]+ε[dy1dt+ky1+y03]+ε2[dy2dt+ky2+3y02y1]+…=0.
(6.11)
y0(0)+εy1(0)+ε2y2(0)+…=2.
Since ɛ is a parameter, terms with diffe rent powers of ɛ in Equation (12.26) must each equal to
zero. Using Equation (12.27) we obtain
y0(0)=2,yi(0)=0,i≠0.
Thus we obtain the following system of diffe rential equations
(6.12)
dy0dt+ky0=0,y0(0)=2,
(6.13)
dy1dt+ky1+y03=0,y1(0)=0,
 (6.14)
dy2dt+ky2+3y02y1,y2(0)=0,
and so on. Observe that although the original Equation (12.23) is nonlinear the new system of
equations is actually linear and can be solved recursively. In fact the solution for y 0 is given by
Equation (12.24). Substituting this solution in Equation (9.21) we obtain
y1(t)=4k(e-2kt-1)e-kt
Similarly we can substitute the solutions for y 0 and y 1 in Equation (9.22) and solve for y 2 etc.
Hence to order ɛ the (approximate) solution of Equation(12.23) is
(6.15)
ya(t)=2e-kt+4εk(e-2kt-1)e-kt.
It is interesting to note that although, Equation (12.23) is nonlinear, it has an exact closed form
solution
(6.16)
ye(t)=2k-4ε+e2kt(4ε+k).
In Fig 6.1 we plotted the diffe rence y a - y e on the time interval [0, 5 ] for k = 1 and ɛ = 0.1.
Exercises
1. Use regular perturbations to find an approximate solution to order ɛ of the following equation
dydx+εy+3y2=0,y(0)=1.
2. The equation for the orbit of the planet Mercury around the Sun (in general relativity) can be
reduced to the following equation
d2udθ2+u=a(1+εu2)
where ɛ ≪ 1. Use first order perturbations to examine the impact of the term aɛu 2 on the period of the
orbit.
Hint: First solve this equation with ɛ = 0.
3. Use regular perturbations to find an approximate solution to order ɛ of the following initial value
problem
d2ydx2+4y-εy2=0,

6.4 Singular Perturbations

y(0)=1+ε,
In general, when the perturbation term is small, its impact on the solution of the corresponding
equation without the perturbation term will be small (as we showed in the previous section).
However, there are cases where the perturbation yields terms in the solution which are unbounded in
time. These are referred to as “Singular Perturbations” and we demonstrate their treatment through the
following example.
Example 6.4.1
Find an approximate solution to the following perturbed spring‐mass problem (with no friction)
(6.17)
d2ydt2+k2y+εy3=0,y(O)=1,dydt(0)=0
Solution 6.4.1
Using the regular perturbation expansion in Equation (12.25) and substituting it in Equation
(12.30) we obtain the following system of equations
(6.18)
d2y0dt2+k2y0=0,y0(0)=1,dy0dt(0)=0,
(6.19)
d2y1dt2+k2y1+y03=0,y1(0)=0,dy1dt(0)=0,
d2y2dt2+k2y2+3y02y1=0,
and so on. The solution for y 0 is y 0(0) = 0,
(6.20)
dy2dt(0)=0,
(6.21)
y 0 = cos ( k t )
and the equation for y 1 becomes
(6.22)
d 2 y 1 d t 2 + k 2 y 1 + cos 3 ( k t ) = 0 , y 1 ( 0 ) = 0 , d y 1 d t ( 0 ) = 0 .
However
cos 3 ( k t ) = cos ( k t ) ( 1 - sin 2 ( k t ) )
which implies that Equation (12.35) has a resonance term since cos ( k t ) is one of the solutions of
the homogeneous equation. In fact the solution of Equation (12.35) is
(6.23)
y 1 = 1 4 k 2 [ cos ( k t ) + 1 8 ( cos ( 3 k t ) - 9 cos ( k t ) - 12 k t sin ( k t ) ) ] .
Following the steps in the previous section, the approximate solution to order ɛ of Equation
(12.30) should be,
(6.24)
y ( t ) = cos ( k t ) +
ε 1 4 k 2 [ cos ( k t ) + 1 8 ( cos ( 3 k t ) - 9 cos ( k t ) - 12 k t sin ( k t ) ) ]
However even if ɛ ≪ 1 the last term in this equation will dominate as t increases. This term is refe
rred to as secular term. Furthermore such unbounded terms exist also in the solution for y 2 etc.
To overcome this problem (i.e. remove the secular term from the solution) it is customary to use a
perturbation expansion for both the dependent and independent variables, viz.
(6.25)
t=s(1+εa1+ε2a2+…)
(6.26)
y(s)=y0(s)+εy1(s)+ε2y2(s)+…)
where a i , i = 1, 2, . . . are constants whose value will be determined so that the secular term (s) in
the expansion disappear. (In more general expansions the constants a i are replaced by functions of
s.)
Using these formulas we have
dydt=dydsdsdt=dyds(1+εa1+ε2a2+…)-1
The Taylor expansion of
f(ε)=(1+εa1+ε2a2+…)-1
around ɛ = 0 is
f(ε)=(1-εa1+…).
Therefore,
dydt=dyds(1-εa1+…).
Similarly for the second order derivative we obtain
(6.27)
d2ydt2=ddt(dydsdsdt)=ddt[dyds(1-εa1+…)]=
dds[dyds(1-εa1+…)]dsdt=
d2yds2(1-εa1+…)2=d2yds2(1-2εa1+…).
Substituting this result in Equation (12.30) and using Equation (9.70) we have
(6.28)
(d2y0(s)ds2+εd2y1(s)ds2+…)(1-2εa1+…)+
k2(y0(s)+εy1(s)+…)+ε(y0(s)+εy1(s)+…)3=0
and for the initial conditions (since t = 0 corresponds to s = 0)
(6.29)
y0(0)+εy1(0)+…=1
(6.30)
dydt(0)=dyds(0)(1-εa1+…)=
(dy0ds(0)+εdy1ds(0)+…)(1-εa1+…)=0.
Hence to order ɛ we obtain the following two equations:
(6.31)
d2y0(s)ds2+k2y0(s)=0,y0(0)=1,dy0ds(0)=0.
(6.32)
d2y1(s)ds2+k2y1(s)+y0(s)3-2a1d2y0(s)ds2=0,y1(0)=0,dy1ds(0
)=0.
The solutions of these equations are
(6.33)
y 0 ( s ) = cos ( k s )
(6.34)
y 1 ( s ) = s 32 k 2 ( - 32 a 1 k 3 - 12 k ) sin ( k s ) + 1 32 k 2 ( cos ( 3 k s ) - cos ( k s ) ) .
We see then that the secular term will disappear if we let
a1=-38k2
and the approximate solution of Equation (12.30) to first order in ɛ is
y ( s ) = cos ( k s ) + ε 32 k 2 ( cos ( 3 k s ) - cos ( k s ) )
t=s(1-3ε8k2)
Exercises
Use singular perturbation expansion to first order in ɛ to find an approximate solution of the
following equations. Compare the solution with the numerical solution.
1.
d2ydt2+ε(dydt)2+k2y=0,y(O)=1,dydt(0)=0.

6.5 BOUNDARY LAYERS

When a viscous fluid flows over a plate (e.g. airplane wing), a thin layer is formed over the plate
where the flow velocity undergoes a very rapid change. This thin layer is referred to as a “Boundary
Layer”. Mathematically these phenomena are usually modeled by differential equations where
0 < ɛ ≪ 1 multiplies the highest derivative in the equations. To see how an approximate solution for
these problems can be obtained using perturbation theory we consider the following prototype
problem.
Example 6.5.1
Consider the following boundary value problem on the interval [0, 1 ];
(6.35)
εd2ydx2+dydx+ky=0,y(O)=0,y(1)=1,k>0,0<ε<<1.
Solution 6.5.1
The exact solution of this linear problem is
y ( x ) = A exp ( - x 2 ε ) sinh ( α x )
where
α = 1 - 4 k ε 2 ε , A = e x p ( 1 2 ε ) sinh ( α ) .
A plot of this solution with k = 1 and ɛ = 0.1 is shown in Fig 6.2. From this figure we see that the
solution rises rapidly from 0 to 2.3 over the “boundary layer” [0, 0.15 ] and then decays smoothly
over the rest of the interval.

Figure 6.2 Exact solution of (6.35)

To see the challenge that these types of problems present (from perturbation theory perspective) we
attempt, at first, to obtain an approximate solution via regular perturbations. Thus, if we write the
perturbation expansion
(6.36)
y(x)=y0(x)+εy1(x)+ε2y2(x)+…
and substitute in (Equation 6.35) we obtain the following system of equations:
(6.37)
dy0dx+ky0=0,y0(0)=0,y0(1)=1
(6.38)
dy1dx+ky1=-d2y0dx2,y1(0)=y1(1)=0
(6.39)
dy2dx+ky2=-d2y1dx2,y2(0)=y2(1)=0
etc. Thus although (Equation 6.35) is a second order equation with two boundary conditions, the
perturbation expansion yields a system of first order equations with two boundary conditions for
which no solution exists in general. In fact the general solution of (Equation 6.37) is
y0=Ce-kx,
but since y 0(0) = 0, it follows that C = 0 and, therefore, y 0(1) = 1 cannot be satisfied.
To see the origin of this peculiar behavior and how it can be ameliorated we note that the second
order derivative of the exact solution of (Equation 6.35) has a factor of
exp(-x2ε)ε
This observation shows that the term ε d 2 y d x 2 is small compared to d y d x and y for large x (e.g.,
x > 2 ε b . However, in a small interval around 0 it is of the same order as these two terms.
In view of this observation it is natural to divide the solution of our problem into two.
1. A solution in an inner region 0 ≥ x ≤ x 0 where x 0 > 0 is small.
2. A solution in an exterior region x 0 < x.
Observe that, at the present, the exact value of x 0 remains unspecified.
Denoting the exterior solution by v we note that in this region v, and its first and second order
derivatives have the same order of magnitude, and hence ε d 2 v d x 2 is negligible as compared to
the other terms in (Equation 6.35). Therefore, in the exterior region (Equation 6.35) can be
approximated correctly by a regular perturbation expansion,
(6.40)
dv0dx+kv0=0,v0(1)=1,
(6.41)
dv1dx+kv1=-d2v0dx2,v1(1)=0,
etc.
Observe that the point x = 0 is outside the “exterior region.” Therefore, we do not have to take into
account the boundary condition on the solution at this point.
From (Equations 6.40) and (6.41) we obtain
(6.42)
v 0 = exp [ - k ( x - 1 ) ]
(6.43)
v 1 = - k 2 ( x - 1 ) exp [ - k ( x - 1 ) ] ,
i.e., to first order in ɛ the outer solution is given by
 (6.44)
v = [ 1 - ε k 2 ( x - 1 ) ] exp [ - k ( x - 1 ) ] .
To obtain a similar perturbation expansion in the inner region we shall perform a stretching
transformation on this region so that the derivatives of the solution in this region have the “same
magnitude.” This can be accomplished by a transformation of the form
(6.45)
z=εaX
where a is a parameter to be determined. Substituting this transformation in (Equation 6.35) and
denoting the inner solution by u it follows that
(6.46)
ε1+2ad2udz2+εadudz+ku=0,u(O)=0
From this equation it follows that the derivatives of u will have the same power of ɛ if 1 + 2a = a i.e.,
a = - 1. (Observe that due to this stretching transformation x = ɛ will correspond to z = 1.) Hence in
the inner region (Equation 6.35) takes the form
(6.47)
d2udz2+dudz+εku=0,u(O)=0.
Applying regular perturbation expansion to (Equation 6.47) yields
(6.48)
d2u0dz2+du0dz=0,u0(0)=0
(6.49)
d2u1dz2+du1dz=-ku0,u1(0)=0
and so on.
The solution of these equations to first order in ɛ is given by
(6.50)
u0=c1(1-e-z).
(6.51)
u1=k[c2-c1z-(c2+c1z)e-z].
Now that we have perturbation solutions for the inner and outer regions a “matching principal” must
be formulated so that the two solutions blend smoothly at the “edges” of the two regions. This will
help to determine the constants c 1, c 2 which appear in (Equations 6.50) and (6.51). To accomplish
this task several “principles” were formulated in the past. The first due to Prandtl stipulates that
(6.52)
lim x → 0 v ( x , ε ) = lim z → ∞ u ( z , ε ) .
Applying this principle to the zeroth order perturbation solutions (6.42) and (6.51) of our problem we
find that we must satisfy
(6.53)
lim x → 0 A e k ( 1 - x ) = lim z → ∞ c 1 ( 1 - e - z ) .
Hence c 1 = Ae k . However, it is easy to verify that this principle fails to determine the constant c 2 in
the first order perturbation expansion.
To overcome this problem M. Van Dyke formulated the following generalized matching principle
for higher order perturbation expansions. To apply this matching principle to the perturbation
expansions in the inner and outer regions up to order ɛ m ,
v(x)=v0(x)+εv1(x)+…+εmvm(x)
uz=u0z+εu1z+…+εmumz,
one must express v(x) in terms of z and u(z) in terms of x, then expand the resulting expressions in
powers of ɛ up to ɛ m and finally, match the coefficients of ɛ k , k = 0, . . ., m in the two expressions to
determine the redundant constants.
As an example we implement this principle to determine the constants c 1, c 2 in (Equations 6.50),
(6.51) by matching the inner and outer solutions to first order in ɛ.
Rewriting the outer solution (Equation 6.44) in terms of z and expanding in ɛ we obtain
(6.54)
v(z)=Aek[e-εkz+εk2(1-εz)e-εkz]
=Aek[1-εkz+εk2+O(ε2)].
Similarly for the inner solution we have
(6.55)
u(x)=c1(1-e-x/ε)+εk[c2-c1x/ε-
(c2+c1x/ε)e-x/ε]=c1(1-kx)+εkc2+O(ε2).
(Note that the expression e -x/ɛ for any fixed x > 0 converge to 0 faster than any powers of ɛ and,
therefore, can be neglected). It follows then,
(6.56)
c1(1-kx)+εkc2=Aek[(1-kx)+εk2].
This yields
(6.57)
c1=Aek,c2=Akek.
Exercises
For the following two differential equations use first order perturbations to find inner, outer, and
matched solutions. Solve these equations numerically.
Compare the numerical and analytic solutions on the same plot.
εd2ydx2+(1+ax)dydx+k2y=0,y(O)=0y(1)=1.
εd2ydx2+(1+ε)dydx+y2=0,y(O)=0y(1)=1.
 CHAPTER 7

Modeling with Partial Differential Equations

CONTENTS
7.1 The Heat (or Diffusion) Equation
7.1.1 Burger’s Equation
7.1.2 Similarity Solutions
7.1.3 Stephan Problem(s)
7.2 Modeling Wave Phenomena
7.2.1 Nonlinear Wave Equations
7.2.2 Riemann Invariants
7.3 Shallow Water Waves
7.3.1 Tsunamis
7.4 Uniform Transmission Line
7.5 The Potential (Or Laplace) Equation
7.5.1 Kirchoff Transformation
7.6 The Continuity Equation
7.7 Electromagnetism
7.7.1 Maxwell Equations
7.7.2 Electrostatic Fields
7.7.3 Multipole Expansion
7.7.4 Magnetostatic
7.7.5 Electromagnetic Waves
7.7.6 Electromagnetic Energy and Momentum
7.7.7 Electromagnetic Potential
In this chapter, we illustrate this modeling process by considering various systems that are modeled in
terms of partial differential equations. In particular, we concentrate on the heat, wave, and potential
equations that are important in many scientific and engineering applications.

7.1 The Heat (or Diffusion) Equation

Objective
Build a model that describes the temperature distribution in a metal as a function of the position
and time.
Discussion
As stated, this problem does not specify if the metal composition of the body is homogeneous, nor
is any information given about its shape. Because such a general problem might require a complex
model, we will first attempt to build a prototype model and then compound it. To begin, we consider
the heat conduction problem in a rod of length L made of homogeneous metal with a constant cross
section A that is completely insulated along its lateral edges (see Fig. 7.1). (All these assumptions
are, naturally, mathematical idealizations.)

Figure 7.1 Perfectly insulated rod with constant cross section

Background
To build an acceptable model for this problem, an understanding of the concept of the flux and
basic laws of thermodynamics (and physics) is necessary. We present here a short review of these
pertinent ideas in one spatial dimension.
The Flux. Consider a flow of a certain physical quantity (such as mass, energy, heat, etc.). The flux
q(x,t) of this flow is defined as a vector in the direction of the flow [at (x, t)] whose length is given by
the amount of the quantity crossing a unit area (at x) normal to the flow in unit time; that is,
(7.1)
| q ( x , t ) | = △ S → 0 lim △ t → 0 Q u a n t i t y p a s s i n g t h r o u g h △ S i n t i m e [ t , t +
△ t]△ S△ t
where AS is a (small) surface area at x that is normal to the flow.
Thus, the approximate amount of the physical quantity passing through a
surface AS in time At is given by
(7.2)
Q(x,t,△ S,△ t)≅ |q(x,t)|△ S△ t
If AS is not normal to the flow, then it must be replaced by its projection in the direction normal to the
flow.

Figure 7.2 Flux-All the fluid in the tube will pass through S in time Δt

Example 7.1.1
Consider water flowing in a river with velocity v(x,t). To evaluate the flux of this flow at (x, t ) , we
consider a small surface element A S normal to v ( x,t ) . The amount of water flowing through A S
in time [ t,t + At] is given by the quantity present at t in a tube of base A S and height |v|At; that is
(7.3)
Q(x,t,△ S,△ t)≅ ρ|v|△ S△ t
where p is the mass density of the water. Hence,
(7.4)
q(x,t)=ρv
since the direction of the flow is given by v (see Fig. 7.2 ).
Basic Laws of Thermodynamics. A change AQ in the amount of heat in a body of mass m is
accompanied by a change A u in its (equilibrium) temperature. The relationship between these
changes is given by
(7.5)
△ Q=cm△ u
where c is the specific heat of the material of which the body is made, that is, the amount of heat
required to raise the temperature of a body of unit mass (made of the same material) by 1 degree.
In the following discussion, we assume that Q and u are normalized so that Q = cmu .
Remark About Units. In (Equation 7.3) (as in any equation that relates physical quantities), a
consistent set of units must be used. Thus, if the MKS system of units is used, then Q (energy) is
expressed in joules, mass in kilograms, u in degrees Kelvin (or Celsius), and c in joules/(kg.degK). In
this book we consistently use the MKS units unless otherwise noted.
Fourier Law of Heat Conduction . Heat is transported by diffusion in the direction opposite to the
temperature gradient and at a rate proportional to it. Thus, the heat flux q(x, t) is related to the
temperature gradient by
(7.6)
q(x,t)=-κgradu(x,t)=-κ(∂u(x,t)∂x,∂u(x,t)∂y,∂u(x.t)∂z)
where κ is the thermal conductivity of the material. [From (Equation 7.6) we infer that its units are
joules/(m.sec. degK).]
Remember that the gradient of a function gives the direction in which the function increases most
rapidly while in the direction opposite to it the function decreases most rapidly. Thus, a restatement
of the Fourier law is that heat flows in the direction in which the temperature decreases most rapidly
(and this is the reason for the minus sign in (Equation 7.6). Note that by convention any constant that
appears in a physical law is assumed to be positive).
Principle of Energy Conservation. Because heat is a form of energy, it must be conserved. Hence,
the rate of change in the amount of heat in a body must equal the rate at which heat is flowing in less
the rate at which it is flowing out (we assume that no heat is generated by the body).
Approximations and Idealizations
1. Since we assumed that the material of the rod that we are considering is homogeneous, it
follows that c, k, and ρ (the material density kg/m 3 ) are independent of the position x.
However, for the purpose of constructing a prototype model we further assume that they are
also independent of the temperature u .
2. We assume that the length of the rod remains constant in spite of the changes in its
temperature.
3. We assume that the rod is perfectly insulated along its lateral surface (idealization). Hence,
heat can flow only in the horizontal direction, since a vertical flow will lead to heat
accumulation along the edges, which is forbidden by the Fourier law of conduction.
Therefore, we infer that the temperature at every point on a vertical cross section of the rod
must be the same. Thus, the temperature u depends only on x and t; that is, u = u(x, t).
4. For definiteness, we assume that heat flows in the rod from left to right, which requires the
left side to be warmer than the right.
Modeling
A possible approach to model the system under consideration is to use the atomic and crystal
structure of the material of the rod and build a model for the heat conduction using these microscopic
variables. However, because this approach will lead to a very complex set of equations, it is not
useful in our context.
Thus, we present in the following two methods to derive the macroscopic heat equation. We use the
term macroscopic since we are using macroscopic variables such as u, c, k, and so on to model the
system.
Infinitesimal Approach
In this method, we consider an infinitesimal element of the rod between x and x + Δχ and write the
equation for the energy conservation in it.
Thus, since the volume of the element is ΑΔχ, its mass Δη is given by ρΑ Δ χ (see Fig. 7.3). The
amount of heat in this element at time t is or
(7.7)
Q(x,t,△ x)=c△ mu(x,t)
Figure 7.3 Heat flux through infinitesimal part of the rod.

or
(7.8)
Q(x,t,△ x)=cρAu(x,t)△ x.
The rate of change in Q is therefore given by
(7.9)
dQdt=cρA∂u∂t△ x.
By the principle of energy conservation, this rate of change must equal the rate at which heat is
flowing in less the rate at which it is flowing out. Hence,
(7.10)
dQdt=q(x,t)·A-q(x+△ x,t)·A.
Replacing d Q d t by
(7.11)
cρA∂u∂t△ x
we have
(7.12)
cρA∂u∂t=-Aq(x+△ x,t)-q(x,t)△ x.
Letting Δχ 0, we obtain
(7.13)
cρ∂u∂t=-∂q∂x.
From (7.6), the Fourier law of heat conduction in one dimension yields q =- κ( ∂ u/ ∂ x) (since u is a
function of x and t only!) and, therefore,
(7.14)
cρ∂u∂t=κ∂2u∂x2
or
(7.15)
1k∂u∂t=∂2u∂x2
where k−1 = cp/κ is called the thermal diffusivity. (Equation 7.15) is called the heat (or diffusion)
equation in one (space) dimension.
Integral Approach
In this method, we consider a finite section of the rod between a and b and use the principle of
energy conservation to write an equation for the heat balance in this segment.
Since the amount of heat in an infinitesimal section of the rod between x and x + Δχ is given by
(Equation 7.8), the total amount of heat in the section [a, b] is given by the integral of the expression
(7.16)
Q(t,a,b)=∫abcρAu(x,t)dx.
The rate of change in this quantity is therefore given by
(7.17)
dQdt=∫abcρA∂u∂tdx.
By the principle of energy conservation, dQ/dt must equal the rate at which heat enters the section
less the rate at which it leaves it. Thus,
(7.18)
dQdt=Aq(a,t)-Aq(b,t).
By the fundamental theorem of calculus, this equation can be rewritten as
(7.19)
dQdt=-∫abA∂q∂xdx
from which it follows that
(7.20)
∫abcρA∂u∂tdx=-∫abA∂q∂xdx.
By the Fourier law of heat conduction,
∫ab(cρA∂u∂t-κA∂2u∂x2)dx=0
(7.21)
∫ab(cρA∂u∂t-κA∂2u∂x2)dx=0.
But since a and b are arbitrary, Equation (17.21) implies that the integrand in this equation must also
be zero. Hence,
(7.22)
cρ∂u∂t-κ∂2u∂x2=0
which is the same equation we derived using the infinitesimal approach.
Remark 7.1.1
Although the infinitesimal and integral approaches must always (if applied correctly) yield the
same result, from a conceptual modeling point of view one might be superior to the other in a
given context.
To illustrate the process of model compounding, we now present the derivation of the heat equation
in two dimensions.
Example 7.1.2
Derive the heat equation for a thin homogeneous plate with constant cross section (height) h.
Assume that the plate is perfectly insulated on the top and bottom.
Solution 7.1.1 Since the plate is thin and perfectly insulated,
(7.23)
u=u(x,y,t).
To derive a model for u, we use the infinitesimal approach and consider a small rectangular element
that is located at a point (x, y) in the plate (see Fig. 7.4).

Figure 7.4 Flux in an infinitesimal section of the plate

The amount of heat Q in this element at time t is given (approximately) by

(7.24)
Q(x,y,t,△ x,△ y)=cρh△ x△ yu(x,y,t)
Hence, the rate of change in Q is
(7.25)
dQdt=cρh△ x△ y∂u∂t(x,y,t).
This rate of change must equal the rate at which heat flows into the element minus the rate at which it
flows out. To compute these rates we decompose q into
(7.26)
q=qi+q2j
and observe that qi is parallel to the boundary represented by the line between ( x , y ) and ( x + △ x ,
y ) . Hence, qi does not contribute to the flux through this boundary. Similar considerations apply to
other boundaries. It follows then that
dQdt=q2(x,y,t)△ xh+q1(x,y)△ yh-q2(x,y+△ y,t)
(7.27)
△ xh-q1(x+△ x,y,t)△ yh.
Equating (Equation 7.25) with (Equation 7.27) and dividing by ▵ x ▵ y yields
(7.28)
cρ∂u∂t=-q1(x+△ x,y,t)-q1(x,y,t)△ x-q2(x,y+△ y,t)-q2(x,y,t)△
y
Letting ▵ x, ▵ y → 0, we obtain
(7.29)
cρ∂u∂t=-(∂q1∂x+∂q2∂y)=-divq.
To obtain an equation containing u only, we apply (Equation 7.6) (in two dimensions). This leads to
(7.30)
1k∂u∂t=∇ 2u
where
(7.31)
∇ 2u=∂2u∂x2+∂2u∂y2.
∇2 is the Laplace operator and k = cρ/κ. (Equation 7.30) is the heat conduction equation in two
dimensions.

7.1.1 Burger’s Equation

Burger’s equation was introduced as a one‐dimensional model for turbulence and has since found
applications in the study of shock wave, wave transmission, traffic flow, etc. The equation is of the
form
∂u∂t+u∂u∂x=ν∂2u∂x2
where u is ((the amplitude of the disturbance” and ν is a constant. We now show that this equation
can be transformed into the heat equation.
To begin with, we remind the reader that a vector field F = ( f 1 ( x ) , f 2 ( x ) , f ( x ) ) is
conservative if curlF = 0. When this happens, there is a (potential) function φ(x) so that
gradφ(x) = F. This result follows from the fact that for any function g(x), curl (gradg(x)) = 0.
Similarly, a vector field F is called solenoidal if divF = 0. Since for any vector field V we have
div(curlV) = 0, it is possible to prove that if a vector field is solenoidal, then there exists V so that
curlV = F. This vector field V is called the ((vector potential” of F. In the two dimensional case
where F = ( f 1 ( x , y ) , f 2 ( x , y ) ) and divF = 0, i.e.,
∂f1∂x+∂f2∂y=0,
we conclude that there exists a function ψ so that
f1=∂ψ∂y,f2=-∂ψ∂x.
Going back to Burger’s equation, we first write it in divergence form as
∂u∂t+∂∂x(12u2-ν∂u∂x)=0.
We conclude, therefore, that there exists a function ψ so that
u=∂ψ∂x,(12u2-ν∂u∂x)=-∂ψ∂t.
Using the first part of this equation to substitute for u in the second part, we obtain the following
equation for ψ
(7.32)
∂ψ∂t=ν∂2ψ∂x2-12(∂ψ∂x)2.
We now introduce a new function η as
ψ = - 2 ν ln η .
(This is called the ((Hopf‐Cole transformation.”) We then have
(7.33)
u=∂ψ∂x=-2ν∂η∂xη,∂ψ∂t=-2ν∂η∂tη
and
∂2ψ∂x2=2ν[(∂η∂xη)2-∂2η∂x2η]
Substituting these relations in (Equation 7.32), we obtain that
∂η∂t=ν∂2η∂x2.
That is, η satisfies the heat equation. From this solution, we can recover u (that is the solution of the
Burger equation) from (Equation 7.33).

7.1.2 Similarity Solutions

Consider the heat equation in one dimension
1k∂u∂t=∂2u∂x2
It is easy to see that the equation remains invariant (unchanged) under the transformations
x¯=αx,t¯=α2t,u¯=βu+γ
where a, β and γ are constants. This three-parameter group of transformations contains the following
one-parameter subgroup
x¯=αx,t¯=α2t,u¯=α2nu
where m is a positive number. A similarity solution is a solution that is invariant under such a one-
parameter group. In fact, if we define
η=x2/tu=tnf(η)
then
∂u∂t=ntn-1f(η)+tnf′(η)(-x2t2),
∂u∂x=2tn-1xf′(η),
and
∂2u∂x2=2tn-1f′(η)+4x2tn-2f′′
Substituting these results in the heat equation, we obtain
4kηf′′+(η+2k)f′--nf=0
which is a hypergeometric equation (the standard form of this equation is obtained after the
substitution χ = - η).
The PDE has been reduced to an ODE. The same transformations and reductions are true in higher
dimensions if u = u ( r , t ) where r = |x|.
As another example for the application of similarity transformations, we consider a nonlinear heat
equation where the diffusion coefficient depends on u
(7.34)
∂u∂t=∂∂x(D(u)∂u∂x).
To find a similarity solution, we attempt to introduce a new variable η = x α t β where α and β are
constants to be determined later and attempt to find a solution of (7.34) which is dependent only on η,
viz. u = u(η) . For such a function, we have (primes denote differentiation with respect to η)
∂u∂t=βxαtβ-1u′,
∂u∂x=αxα-1tβu′,
and
∂∂x(D(u)∂u∂x)=α(α-1)xα-2tβD(u)u′ +α2x2α-2t2β[D(u)u′ ]′
Substituting these results in (Equation 7.34) and dividing by x α-2 t β , we obtain
βx2tu′ (η)=α(α-1)D(u)u′ +α2η[D(u)u′ ]′
We see that if we choose α = 1 and β = - 1/2, i.e.,
η=xt1/2
then the explicit dependence of this equation on x and t separately disappears. The equation
simplifies considerably and we have
(7.35)
[D(u)u′ ]′ +12ηu′ =0;
i.e., the partial differential equation has been reduced to an ODE. This trans‐ formation is due to
Boltzmann.

7.1.3 Stephan Problem(s)

“Stephan problems” refer to a class of interface models between two phases of material (e.g., solid
and liquid) where the boundary of the interface is moving in time.
For concreteness, we consider a solid and liquid in a slender pipe with a constant cross-section A
which is perfectly insulated. The solid phase is on the right side and the liquid phase on the left. We
denote the position of the interface between these phases by p(t). Both the solid and liquid phases are
assumed to satisfy the heat equation with proper coefficients
(7.36)
1kS∂uS∂t=∂2uS∂x2,1kL∂uL∂t=∂2uL∂x2(
where the subscripts S and L denote the solid and liquid phase respectively. At the interface, we also
have
uS(p(t),t)=uL(p(t),t)
i.e., the temperature varies continuously across the interface.
To model the movement of the interface between the two phases, we invoke the law of energy
conservation. The heat flux across the boundary p(t) is
AqS(p(t),t)-AqL(p(t),t)
This amount of heat is used to melt some of the solid in the tube. Denoting the latent heat of the solid
phase by L, we have
(7.37)
AqS(p(t),t)-AqL(p(t),t)=LρAp(t).
This is referred to as the “Stephan condition.” Since Fourier law of heat conduction holds for the two
phases, (7.37) becomes
(7.38)
-kS∂uS(p(t),t)∂x+kL∂uL(p(t),t)∂x=Lρp˙(t).
This equation is used to determine the position of the interface in time
Exercises
1. Generalize the prototype model to a situation in which heat is being generated in the rod at a
rate of r ( x,t ) per unit volume.
Hint . The rate at which heat is being generated in any infinitesimal slice [x,x + Ax] is given by
r(x,i)AAx. Use this to modify Equation (1.6) or (1.8).
1. Generalize the prototype model to the case in which c, k, and p are functions of x (non-
 homogeneous rod).
2. Repeat Exercise 2 when c, k, and p are functions of the temperature u .
3. Generalize the prototype model to the case A = A(x) where A(x) is a function that varies
slowly with x so we can still assume approximately that the temperature u is a function of x
and t .
4. Newton’s law of cooling states that for a non-insulated rod, the rate of the heat loss per unit
length is proportional to u — T 0 where u is the rod temperature and T 0 is the temperature of
the surroundings. Show that the heat equation in one dimension now takes the form
(7.39)
cρ∂u∂t=k∂2u∂x2-a(u-T0)
where a is constant.
Note. In this case we must assume that the rod is “thin” in order to be able to justify the
approximation that temperature u is a function of x and t.
1. A rod is made of a material that undergoes a chemical reaction as a result of which the
specific heat and conductivity change with respect to time [i.e., c = c(t) and k = k(t)]. Derive
the corresponding heat equation in one dimension.
2. Derive the heat conduction equation for a three-dimensional body by considering an
infinitesimal volume.
3. If the temperature inside a homogeneous and isotropic sphere is given to be a function of the
radial distance r only, show that the heat conduction equation is given by
1k∂u∂t=(∂2u∂r2+2r∂u∂r)
Hint. Use infinitesimal spherical shells.
1. Derive the heat conduction equation for an isotropic homogeneous and laterally insulated
circular plate in which u = u(r, t).
Hint. Consider infinitesimal rings.
1. Show that if u(x,t) is a solution of the heat equation, then du/dx and du/dt are also solutions
of this equation.
2. Show that Burger’s equation
∂u∂t+u∂u∂x=ν∂2u∂x2
can be transformed into the heat equation by introducing
w = - 2 ν ∂ ∂ x ( ln u )

7.2 Modeling Wave Phenomena

Objective
Construct a prototype mathematical model for the transverse vibrations of a string with fixed ends
(see Fig. 7.5).
Figure 7.5 Flexed string of length L

Background
In general, an in-depth treatment of wave phenomena requires considering the elastic properties of
matter and leads to a complicated set of equations. To overcome this difficulty, we make the
following simplifying approximations and idealizations so that a prototype model can be constructed
by applying only Newton’s second law F = ma (i.e., the force equals the mass multiplied by the
acceleration) to the system under study.
Approximations and Idealizations
1. The string is rigidly attached at its end points.
2. The string vibrates in one plane.
3. No external forces act on the string (prototype model).
4. The string does not suffer from damping forces (prototype model).
5. The string is homogeneous. In particular, this implies that the linear density ρ (i.e., the mass
per unit length) of the string is constant.
6. The deflection u of the string from equilibrium and its slope are always small.
Consequently, we are able to assume a point on the string moves only in the vertical
direction.
7. The tension force in the string is always tangential to it. This is usually expressed by saying
that the string is assumed to be perfectly flexible.
Modeling
Consider a small segment of the string between x and x + Δχ as shown in Fig. 7.6. Before we can
apply Newton’s Second Law to the motion of this segment, we must make the following observations:
1. By approximation 6, the segment is not moving in the horizontal direction. Therefore if we
denote by T(x,t) the tension in the string, the horizontal components of T(x,t) at x and x + Δχ
must be equal (see Fig 7.6).
Figure 7.6 Balance of forces on a small section of the string

(7.40)
T ( x ) cos α = T ( x + △ x ) cos β = R
1. The mass of the segment [ x , x + △ x ] is given by ρ ▵ s. Where ▵ s is the length of this
segment in the deformed state of the string. However, since we are considering only small
deflections, |u| 1, it follows that
△ s≈△ x.
1. The acceleration of the segment in the vertical direction is given by
(∂2u/∂t2).
1. The sum of the vertical forces acting on the segment (see Fig. 7.6)is
T ( x + △ x ) sin β - T ( x ) sin α .
Combining all these observations and approximations, we infer from Newton’s second law that
(7.41)
ρ △ x ∂ 2 u ∂ t 2 = T ( x + △ x ) sin β - T ( x ) sin α .
Using (7.40) to elliminate T form this equation yields
(7.42)
ρ△ x∂2u∂t2=R(tanβ-tanα)
However since T is always tangential to the string
(7.43)
tan α = ∂ u ∂ x ( x , t )
(7.44)
tan β = ∂ u ∂ x ( x + △ x , t ) .
Substituting (Equations 7.43) and (7.44) in (Equation 7.42) and dividing by ▵ x, we obtain
(7.45)
ρ∂2u∂t2=R∂u∂x(x+△ x,t)-∂u∂x(x,t)△ x.
Letting ▵ x → 0, we finally obtain
 (7.46)
1c2∂2u∂t2=∂2u∂x2
where c 2 = R/ρ.
(Equation 7.46) is called the wave equation in one dimension.
Remark 7.2.1 From Fig. 7.6, we can infer that the sum of the horizon‐ tal forces acting on the
string segment is T ( cos β - cos α ) ≠ 0 . Hence, the segment must have an acceleration in the
horizontal direction, which contradicts approximation 6b. However, since α and β are small, we
can argue that T ( cos β - cos α ) is negligible.
Compounding
Example 7.2.1
Derive a model equation for very small vibrations of a vertically suspended chain whose length is
L and whose mass density per unit length p is constant.
Approximations
1. Since the amplitude u of the vibrations is small, we assume that a point on the chain does
not change its x-coordinate (see Fig 7.7).
2. The tension T(x,t) in the chain cannot be assumed to be constant in the context of this
problem. In fact, in the equilibrium (vertical) position of the chain,

Figure 7.7 Vibrating chain

(7.47)
T(x)=ρg(L-x)
In the following model, we assume that (Equation 7.47) gives an accept‐ able approximation for the
tension in the vibrating chain when |u| ≪ 1 and | ∂ u ∂ t | ≪ 1 .
1. Other approximations and idealizations of the prototype model remain intact.
Modeling
To construct a mathematical model, we once again consider a small section of chain between [ x , x
+ △ x ] . Applying Newton’s second law in the horizontal direction to such a section (see Fig. 7.8),
we obtain
(7.48)
ρ △ x ∂ 2 u ∂ t 2 = T ( x + △ x ) sin β - T ( x ) sin α .

Figure 7.8 Balance of forces on a small section of the chain

But since α and β are small, we can once again use the approximations given by (Equations 7.43)
and (7.44). Hence,
(7.49)
ρ∂2u∂t2=1△ x[T(x+△ x)∂u∂x(x+△ x,t)-T(x)∂u∂x(x,t)].
Letting ▵ x → 0, it follows then that
(7.50)
ρ∂2u∂t2=∂∂x[T(x)∂u∂x(x,t)].
Substituting (Equation 7.47) for T(x), we finally obtain
(7.51)
∂2u∂t2=g∂∂x[(L-x)∂u∂x].
Other examples of physical systems describing wave phenomena will be discussed in Sections 4 and
5.

7.2.1 Nonlinear Wave Equations

Nonlinear wave equations have many important applications. We give here a brief overview of this
topic.
To begin with, we consider the nonlinear wave equation
utt+F′′(u)F′(u)ut2=λ[uxx+F′′(u)F′(u)ux2]
where F(u) is a smooth invertible function of u. This equation can be linearized by the transformation
ψ = F(u) . In fact,
ψt=F′(u)ut,ψtt=F′(u)[utt+F′′(u)F′(u)ut2]
with similar expressions for the derivatives with respect to x. We infer then that
ψtt=λψxx.
An important wave equation that models long water waves of low amplitude is the Kortewegde Vries
(KdV for short) equation
(7.52)
ut-6uux+uxxx=0.
This equation admits ((traveling wave solutions” of the the form u = f(x - ct) ( o r u = f ( x + c t ) ) in
which case this equation reduces to
(7.53)
-cf′-6ff′+f=0
where primes denote differentiation with respect to z = x - ct. This equation can be integrated once,
and we obtain
(7.54)
-(c+3f)f+f′ ′ =C1.
Multiplying this equation by f ’ and integrating again leads to
(7.55)
-(c+2f)f2+(f′ )2=2C1f+C2
where C 1, C 2 are integration constants. For the special case that C 1 = C 2 = 0 (which can be
justified by imposing proper boundary conditions on f and f ’), we obtain a simple first order equation
whose solution is
f=-c2sech2[c2(x-ct)]
Such a solution is called a soliton as it exhibits some special properties. One of these special
properties is that when two solitons traveling in opposite directions collide, they come out of the
collision with unchanged shape.
The KdV equation admits also n‐soliton solutions. A two soliton solution is given by
f = - 12 3 + 4 c o s h ( 2 x - 8 t ) + c o s h ( 4 x - 64 t ) [ 3 c o s h ( x - 28 t ) + c o s h ( 3 x - 36 t ) ] 2
.
The Sine‐Gordon (SG) equation is obtained from the linear wave equation by replacing the right hand
side of this equation by some elementary function, e.g.,
u x x - u t t = sin u .
This equation models superconducting transmission lines (“Josephson Junc‐ tions”) and the
propagation of crystal defects (and many other phenomena). The nonlinear term on the right hand side
of this equation models the quan‐ tum tunneling effect of electron‐pairs through the insulating material
of a superconducting transmission line.
 Looking again for traveling wave solutions, u = f(x - ct) reduces the SG equation to
( 1 - c 2 ) f ′ ′ = sin f .
We can integrate this equation with respect to z = x - ct by multiplying it by f ’ to obtain
( 1 - c 2 ) ( f ′ ) 2 = C 1 - 2 cos f .
For the special case C 1 = 2, we can solve this differential equation to obtain a soliton solution in the
form
u ( x , t ) = 4 arctan e x p ( ± x - c t ( 1 - c 2 ) ]
This solution with the plus (minus) sign is usually referred to as a((kink“ (anti‐kink) solution since u
increases (decreases) monotonically by 2π as z varies from - ∞ to ∞.
Another wave equation which models water waves in deep ocean and the propagation of nonlinear
pulses in fiber optics is the Nonlinear Schrodinger equation (NLS)
(7.56)
iut±uxx+2|u|2u=0.
The equation with the + ’sign is referred to as the focusing NSL equation while the one with “ sign is
called the defocusing NLS equation.
To derive this equation, we consider a one dimensional superposition of waves (a ((wave packet”)
in the form
(7.57)
f(x,t)=12π∫-∞∞F(k)ei(kx-ω(k)t)dk
where F(k) is the Fourier transform of f ( x , 0 ) . Observe that the different waves in this
superposition might have different speeds c(k) = ω(k)/k. The relation ω = ω(k) is referred to as the
((dispersion relation for the wave” and c(k) as the phase velocity, v p (k), of the individual waves in
the packet.
We now consider such a superposition under the assumption that F(k) is concentrated around a
certain value of k = k 0 (that is F(k) ≈ 0 outside a small interval around k 0). Such a superposition of
waves is called awave(( packet.” For such a packet, we can expand the dispersion relation around k 0
ω=ω0+a1(k-k0)+a2(k-k0)2+O[(k-k0)3].
Substituting this approximation in (7.57) we obtain
(7.58)
f(x,t)=ei(k0x-ω0t)ψ(x,t)
where
(7.59)
ψ(x,t)=12π∫-∞∞F(k)ei[(k-k0)x-a1(k-k0)t-a2(k-k0)2t]dk.
In (Equation 7.58), e i ( k 0 x - ω 0 t ) is called the ((carrier wave’’ while ψ ( x , t ) is the ((envelope
wave.”
Letting ν = k - k 0 and differentiating (Equation 7.59) with respect to t, we obtain
(7.60)
a2∂2ψ∂x2+i(∂ψ∂t+a1∂ψ∂x)=0.
By adding a “small nonlinear term a|ψ|2 ψ“ to this equation to account for the transmission medium
nonlinearities, we obtain the “Nonlinear Schrodinger equation”
a2∂2ψ∂x2+i(∂ψ∂t+a1∂ψ∂x)+a|ψ|2ψ=0.
This equation can be transferred to the standard form of the NLS (Equation 7.56) by a proper
transformation of the variables. The NLS equation admits solitons as solutions.

7.2.2 Riemann Invariants

Consider the nonlinear wave equation with c = c(ψ x ), i.e.,
(7.61)
c(ψx)2∂2ψ∂x2-∂2ψ∂t2=0.
We can rewrite this equation as a system of two equations by introducing ψ t = u and ψ x = v. In fact,
since u x = ψ xt = v t , we have
(7.62)
vt-ux=0.
Furthermore, from (Equation 7.61) we have
(7.63)
c(v)2vx-ut=0.
Multiplying (Equation 7.62) by c(v) and adding to (Equation 7.63) we have
(7.64)
c(v)[vt+c(v)vx]-[ut+c(v)ux]=0.
Similarly, by multiplying (Equation 7.62) by c(v) and subtracting from (Equation 7.63), we obtain
(7.65)
c(v)[vt-c(v)vx]-[ut-c(v)ux]=0.
On the line where c ( v ) = d x d t , (Equation 7.64) becomes
c(v)[vt+c(v)vx]-[ut+c(v)ux]=c(v)dvdt+dudt=0.
Letting dG = c(v)dv, we rewrite this equation as
ddt[G(v)+u]=0.
Similarly on the line c ( v ) = - d x d t , (Equation 7.65) becomes
c(v)[vt-c(v)vx]-[ut-c(v)ux]=c(v)dvdt-dudt=0
which leads to
ddt[-G(v)+u]=0.
It follows then that the quantities
r=G(v)+u,s=-G(v)+u
are conserved on the lines where c ( v ) = ± d x d t respectively. These quantities are called the
Riemann invariants of the (Equation 7.61). The lines where c ( v ) = ± d x d t are called the
“characteristic curves” of (Equation 7.61).
Exercises
1. Derive a model equation for the vibration of the string if a vertical external force F(x,t) per
unit length is acting on it. Especially consider the case of the gravitational force where F =
pg.
2. Derive a model equation for the vibration of the string when its motion is subject to an
elastic restraint and a damping force. (The restraint can be thought of as a force of ku per
unit of length acting to return the string to its equilibrium position. The damping force is
 given by b(du/dt) per unit length acting to oppose its motion.)
3. Generalize the model equation for the vibration of the string to the case where p = p(x).
4. Longitudinal elastic waves: Consider an elastic homogeneous rod with a constant cross
section placed along the x-axis. If we apply longitudinal stresses to the rod, then a particle
whose rest position is x will find itself at x + u(x,t), where we are assuming that the
displacement u = u(x,t). It is known that the local stress, T (force/unit area), in an elastic bar
satisfies T = Eu x where E is called the elastic modulus. Derive a model equation for u(x, t)
by considering a small section of the rod and the stress difference between its ends (see Fig.
7.9).

Figure 7.9 Acoustic vibrations in a metal rod

Hint. Let p be the mass density of the bar and A its cross section. The mass of the small section is
pAAx, and its acceleration is d 2 u/dt 2 . The stress difference across the section is EA[u x (x + Ax,t)
— ux(x,t)]. Note that c = y/E/p is called the speed of sound in the medium.
1. Derive a model equation for the vibrations of a membrane whose edges are fixed. Make
similar assumptions as for the vibrating string.
Hint. Consider a small rectangular area and apply the same analysis as for the string in the x and y
directions.
1. The pressure p and the mass flow rate u of a fluid flowing in a long pipe are related
approximately by the equations
∂p∂t=c∂u∂x
∂u∂p
∂t¯=∂x¯
where c is the compressibility of the fluid. Show that both p nd u satisfy the wave equation in one
dimension.
1. Show that
u(x,t)=f(x-ct)+g(x+ct)
is a solution of the wave equation in one dimension when f and g are any “smooth functions.”
1. Derive a model equation for the small vibrations of a string that is rotating around the x-axis
at a constant angular velocity ω. Assume that at each moment all the points of the string are
one plane.
2. A mass m is attached to the end of a suspended chain of length L and linear density p .
Derive a model equation for the small vibrations of this system.
 3. The results of Exercise 4 can be applied to the air vibrations in an organ pipe. However,
since it is not “easy” to follow the position of air molecules, it is more natural to derive an
equation for the pressure in the pipe. If it is known that the pressure is proportional to du/dx
, show that it must satisfy the wave equation.

7.3 Shallow Water Waves

Objective
Derive a prototype model equation that describes the phenomena of slow waves in shallow water
(i.e., waves in a pool or on the seashore). Background
An in-depth treatment of wave phenomena in fluids requires a knowledge of fluid mechanics. We
simplify the derivation of this problem by applying Newton’s second law and the following
elementary facts.
1. The hydrostatic pressure p (force/unit area) in a fluid at a point of depth D below its surface
is given by
(7.66)
p=ρgD
where p is the mass density of the fluid and g is the acceleration of gravity.
1. The principle of mass conservation states that the rate of change of mass in a given volume
equals the rate at which mass is entering the volume less the rate at which it is leaving it.
Approximations and Idealizations
1. Water is incompressible and hence has a constant density ρ independent of the pressure.
2. The motion of the fluid under consideration is two dimensional; that is, each fluid particle is
constrained to move in two dimensions, x and y (see Fig. 7.10).
3. Because we are considering only slow waves, water acceleration is small and we can
approximate the pressure at a point (x, y) beneath the water surface by the hydrostatic
pressure

Figure 7.10 Waves over topography

 (7.67)
p(x,y,t)=ρg[h(x,t)-y]
where h ( x,t ) is the wave height.
1. The force per unit volume in the fluid is given by the negative of the pressure gradient. In
particular, the force per unit volume in the x direction is
(7.68)
F=-px=-ρghx(x,t)
where p x = ∂ p/ ∂ x, and so on.
Since the force F is independent of y, it is reasonable to assume that the x component of a particle
velocity u is a function of x and t only; that is, u = u ( x , t ) . (We assume also that the y component of
the velocity is 0.)
Modeling
Consider a small volume ▵ V of water. Its mass is ρ ▵ , V and its acceleration is du/dt. The force
acting on it in the x direction is - ρ g h x ( x , t ) △ V . Hence, by Newton’s second law,
(7.69)
ρ△ Vdudt=-ρghx(x,t)△ V.
(This is Lagrange’s picture where the coordinate system is moving with the particle.) However, for an
inertial observer (Euler picture) u = u [ x ( t ) , t ] , there‐ fore,
(7.70)
dudt=∂u∂tdtdt+∂u∂xdxdt=∂u∂t+∂u∂xu.
Thus, we infer from (Equation 7.69) that
(7.71)
ut+uxu=-ghx.
Because (Equation 7.71) contains two unknown quantities, we need an‐ other independent equation
that relates h and u in order to be able to solve our model. Such an equation can be obtained by using
the principle of mass conservation.
To apply this principle, consider the column of water between x and x + ▵ x and the horizontal
length Δ z in the z direction (see Fig. 7.11).
Figure 7.11 Fluid flux in a small column

The difference between the amounts of water in this volume at t + Δ t and t is

(7.72)
△ z·△ x·ρ[h(x,t+△ t)-h(x,t)].
This quantity must equal the difference between the amount of water en‐ tering this volume in time ▵ t
less the amount leaving it. Hence,
(7.73)
△ z·△ t·ρ{[D(x)+h(x,t)]u(x,t)-[D(x+△ x)+h(x+△ x,t)]u(x+△ x,
t)}=△ z·△ x·ρ[h(x,t+△ t)-h(x,t)].
Dividing (Equation 7.73) by ▵ x and ▵ t and letting them approach zero, we obtain in the limit
(7.74)
ht=-[u(h+D)]x.
(Equations 7.71) and (7.74) constitute a system of coupled partial differential equations describing
shallow water waves.
We observe that (Equation 7.74) is in conservative form; i.e., if we define
f1=h,f2=u(h+D)
then
∂f1∂t+∂f2∂x=0.
However, (Equation 7.71) is not in this form. To re‐express this equation in conservative form, we
multiply (Equation 7.71) by (h + D) and (Equation 7.74) by u and sum. Combining the derivatives and
observing that D = D(x), we obtain
[(h+D)u]t+[(h+D)u2]x+g(h+D)hx=0.
Adding and subtracting gh(h + D) x in this equation we obtain
(7.75)
[(h+D)u]t+[(h+D)u2+g(h+D)h]x=gh(h+D)x.
Multiplying this equation by ρ, we can interpret it as a representation of the law of momentum
conservation where the right side of the equation represents the impulse due to gravity and non‐
constant bottom. The terms in the ex‐ pression [(h + D)u 2 + g(h + D)h] represent the kinetic and
potential energy of the fluid respectively.

7.3.1 Tsunamis
Waves on the oceans are generated by various causes. First, there are waves generated by winds
blowing on the ocean surface. Then, there are waves due to tidal forces exerted by the gravitational
interaction with the Sun and Moon systems. Finally, there are those that are generated as a result of
earthquakes or other natural catastrophes. This last category of waves is referred to as tsunamis.
Tsunamis are characterized by long wavelength λ and period ω. Typically, the wavelength can
range from few to hundreds of kilometers. Since the ocean depth D on Earth is less than llkm, it
follows that for these waves we always have λ / D ≫ 1 . Accordingly, tsunamis can be treated as
shallow water waves. As the wave approaches, land D decreases and λ / D becomes very large.
Thus, decreasing depth leads to a higher concentration of the energy.
In the deep ocean, we can approximate the horizontal and vertical dis‐ placements of a particle at (
x , y ) due to the wave by
X = A sin ( k x + ω ( k ) t ) , h = - B cos ( k x + ω ( k ) t )
where A is a constant and B = Aky. The factor ky has been added in this model of the tsunami wave
for several reasons. To begin with, k = 2 π / λ , and it follows that at the ocean surface the vertical
amplitude of this wave is 2 π A D / λ . However, D / λ ≪ 1 , which implies that (as expected) the
vertical amplitude of the wave is much smaller than the horizontal amplitude. In addition, the vertical
displacement is 0 at the ocean bottom. Furthermore, this form of the wave satisfies the continuity
equation div u = 0, u = ( u , v ) . In fact, the velocity components of the wave are
u = A ω cos ( k x + ω ( k ) t ) , v = A k y ω sin ( k x + ω ( k ) t )
and therefore
ux+vy=0.
Energy conservation implies then that ω 2 = (gD)k 2. (Such a relation that expresses the frequency as a
function of the wave number is called ((a dispersion relation.”) The group velocity is therefore
vg=|dωdk|=gD.
Observe that for this dispersion relation, the phase velocity v p = | ω k | is equal also to g D . From
this expression, we infer that in the deep ocean with D ≈ 4km, we have v g ≈ 200m/sec. Thus, a
tsunami can propagate very fast in the open ocean. However, as the water depth decreases, the
tsunami slows down. At the same time the tsunami’s energy flux, which is the sum of its kinetic and
potential energies, remains almost constant. Consequently, as the tsunami’s speed diminishes as it
travels into shallower water, its height grows. Because of this effect, a tsunami, imperceptible at sea,
may grow to be several meters or more in height near the coast.
Thus, assuming no dissipation due to friction or turbulence, the tsunami energy flux is proportional
to B 2 v g or B 2 D . It follows then that constant energy flow requires that the tsunami amplitude B is
proportional to D -1/4. Suppose now that at a depth of 4km the tsunami’s vertical amplitude is lm (a
“normal wave”), then at depth of 2m (near the shore) the wave height will be approximately 6.7m
Exercises
1. If D is a constant and u, h, and their derivatives are small (so that nonlinear terms in these
quantities can be neglected), show that h ( x , t ) must satisfy the wave equation
htt=(gD)hxx.
1. Under the same conditions as in the previous exercise, what is the equation that is satisfied
by u(x,t)?
2. Repeat Exercises 1 and 2 when D is a linear function of x; that is,
D=ax+ba>0,b<0.
1. Modify (Equation 7.71) to include the action of a damping force (in the x direction) that is
proportional to u.

7.4 Uniform Transmission Line

Objective
Derive a prototype model equation for the voltage and current in a long, uniform, two-wire
transmission line.
Background
(For a detailed introduction to electric circuits, see chapter 1.) The most common forms of
transmission lines are coaxial and two-wire types. The coaxial transmission line consists of two
concentric circular cylinders of metal. The two-wire types consist of two parallel wires, one of
which is used as ground. We consider here the two-wire line.
The passage of an electric current through a cable always involves a leakage, which leads to a loss
of electric energy. For short distances, this loss can usually be ignored. However, over long
distances, which are found in transmission lines, these losses must be taken into account.
Modeling
Since the transmission line is uniform, we assume that the resistance (R),
capacitance (C), inductance (L), and leakage (G) per unit length of the transmission line are
constant.
To derive the required model equations, we consider a small section of the line between x and x +
Δχ and apply Kirchoff’s laws to a circuit that is equivalent to it (see Fig. 7.12). Thus, ΕΔχ, ΤΔχ, CΔχ,
and GΔx are, respectively, the resistance, inductance, capacitance, and conductance of the section.
The quantity GΔx (where G is expressed in mhos or siemens) is a “virtual” resistance so that the
power lost through it to the ground is equal to that due to leakage.

Figure 7.12 Equivalent circuit for a small section of the transmission line
 Applying Kirchoff’s second law between A and D, (see Fig. 7.12) we obtain
(7.76)
e(x,t)-e(x+△ x,t)=R△ xi(x,t)+L△ x∂i(x,t)∂t.
Similarly, applying Kirchoff’s first law at the node B, we have
(7.77)
i(x,t)-i(x+△ x,t)=C△ x∂e(x+△ x,t)∂t+G△ xe(x+△ x,t).
Dividing these equations by Δχ and letting Δχ go to zero, we arrive at two differential equations:
(7.78)
∂e∂x(x,t)=-Ri(x,t)-L∂i(x,t)∂t
(7.79)
∂i∂x(x,t)=-Ge(x,t)-C∂e(x,t)∂t.
Finally, we can find an equation for e ( x , t ) only by differentiating (Equations 7.78) and (7.79) with
respect to x and t, respectively, to eliminate i ( x , t ) . We obtain
(7.80)
exx=LCett+(LG+RC)et+RGe.
Similarly, we can show that i ( x , t ) satisfies
(7.81)
ixx=LCitt+(LG+RC)it+RGi.
(Equations 7.80) and (7.81) are known as the telegraph equations.
Special Cases
High‐Frequency Limit. To qualitatively analyze this limit, consider the case where
(7.82)
e ( x , t ) = A ( x ) sin ( ω t + ϕ 1 ) , i ( x , t ) = B ( x ) sin ( ω t + ϕ 2 )
and ω ≫ 1. Under these assumptions, the second term in the right‐hand side of (Equation 7.78),
whose ((effective coefficient” is Lω (the impedance), is much larger than the first term, whose
effective coefficient is R. Hence, (Equation 7.78) can be approximated in this limit by
(7.83)
ex=-Lit.
Similarly, (Equation 7.79) reduces to
(7.84)
ix=Cet.
Combining (Equations 7.83) and (7.84) we obtain the wave equation
(7.85)
exx=LCett,ixx=LCitt.
Low‐Frequency Limit. In this case, i and e change very slowly with time. There‐ fore, from a similar
qualitative analysis as performed for the high frequency limit, with ω ≪ 1, we obtain that the
effective coefficient of the second term (Lω) is much smaller than R. Consequently, we can
approximate (Equations 7.78) and (7.79) by
(7.86)
ex=-Ri,ix=-Ge
and therefore,
 (7.87)
exx=RGe,ixx=RGi.
These are ordinary differential equations for i and e.
Submarine Cable. In the 19th century, telecommunication signals between the United States and
Europe were transmitted by cables that were laid down on the ocean floor. For these cables, G ≅ O
because of their extreme insulation. Moreover, the signal frequency ω is low. Under these
circumstances, we infer from (Equation 7.82) that the impedance Lω is much smaller than R. Hence,
(Equation 7.78) can be approximated by
(7.88)
ex=-Ri.
Furthermore, since we have assumed G = 0, (Equation 7.79) simplifies to
(7.89)
ix=-C∂e∂t.
Combining these two equations, we can approximate (Equations 7.78) and (7.79) by
(7.90)
exx=RCet,ixx=RCit
which show that e and i satisfy the one‐dimensional diffusion equation.
Exercises
1. Give explicit derivations of (Equations 7.80) and (7.81).
2. Compare (Equations 7.80) and (7.81) with the model equation for a spring mass system with
friction and identify the physical meaning of each term in these equations.
3. Explain in detail the approximations that lead to (Equations 7.83), (7.84), (7.86), (7.89),
and (7.90).
4. Derive a differential equation for the voltage e ( t ) in the circuit in Fig. 7.13 if the current i
( t ) is known.
Figure 7.13 LRC circuit

7.5 The Potential (Or Laplace) Equation

Objective
Derive model equations to compute the gravitational field of a material body.
Remark 7.5.1 Although we restrict our discussion to the gravitational field, the static electric
field would warrant similar treatment.
Background
Newton’s law of gravitation states that a point mass M attracts another point mass m by a force
(7.91)
F=-GMmr2er
where G is the gravitational constant, r is the distance between the two masses, and e r is a unit vector
along r (pointing outward from M, i.e., in the direction in which r increases) (see Fig. 7.14).
Figure 7.14 Gravitational force exerted by mass M on mass m

Figure 7.15 Spherical coordinate system

Since (Equation 7.91) can be rewritten as

 (7.92)
F=(-GMr2er)m,
we introduce the gravitational field generated by the mass M as
(7.93)
F=-GMr2er.
Thus, the gravitational field is the force that acts on a test particle of unit mass at a point due to the
presence of the mass M.
The gravitational field admits a potential; i.e., there exists a scalar function Φ so that
(7.94)
F=-∇ Φ.
The minus sign in this equation conforms to the convention that the force exerted by the field on a
point particle is always in the direction in which the potential (energy) is decreasing most rapidly.
For the gravitational field of a point mass M, the potential function Φ is
(7.95)
Φ(r)=-GMr.
Remark 7.5.2
Following the standard engineering convention, (see e.g. Stan‐ dard Mathematical Ta bles
published by CRC), spherical coordinates (see Fig. 7.15) are defined as
(7.96)
x = r sin ϕ cos θ
(7.97)
y = r sin ϕ sin θ
(7.98)
z = r cos ϕ
and the expression of the gradient operator is
(7.99)
∇ f = ∂ f ∂ r e r + 1 r ∂ f ∂ ϕ e ϕ + l r sin ϕ ∂ f ∂ θ e θ .
In our case, Φ is a function of r only.
The superposition principle states that the gravitational field at a point due to two point masses M 1
and M 2 is equal to the (vector) sum of their gravitational fields, which can be written as
(7.100)
Ftotal=F1+F2=-GM1r12er1-GM2r22er2
where r 1, r 2 are the distances from M 1 and M 2 to the point under consideration (see Fig. 7.16).
Figure 7.16 Gravitational force due to M1, M2 at x

As a corollary, we observe that if Φ 1 and Φ 2 are the potential functions for the gravitational field
of the masses M 1 and M 2, respectively, then
(7.101)
Ftotal=-∇ Φ1-∇ Φ2=-∇ (Φ1+Φ2)
where
(7.102)
Φi=-GMirii=1,2.
Thus, the potential function of the total gravitational field is given by the (scalar) sum of the
individual potential functions.
Similarly, if we are given a finite number of point masses, M i , i = 1, n with gravitational fields F
i and potential functions Φ i , then
(7.103)
Ftotal=∑i=1nFi=-∑i=1n∇ Φi=-∇ (∑i=1nΦi).
Remark 7.5.3
If the total potential function Φ of a gravitational field is known, then the gravitational field itself
is given as F = V$. This is one of the reasons for the introduction of the potential function, since Φ
is simply a (scalar) sum of the individual potentials and hence easier to compute than the (vector)
sum of the gravitational fields.
Idealizations
1. We assume that the concept of a point mass is valid. As a matter of fact, we note that due to
the discrete nature of matter, the notion of a (mathematical) point particle with mass m has
no physical meaning.
2. We assume that the field generated by a point particle does not act on itself; otherwise,
various contradictions will creep in.
Modeling
 To compute the gravitational field due to a continuous mass distribution with mass density ρ(x) in a
volume V, we divide the volume into small cells of volume ▵ V i . If we consider each of these cells
as a point particle of mass ρ ( x i ′ ) △ V i (where x i ′ ε △ V i ), then the gravitational field due to it
at a point
x = ( x , y , z ) is
(7.104)
△ Fi=-Gρ(xi′ )△ Viri2eri
where
(7.105)
ri=|x-xi′|=(x-xi′)2+(y-yi′)2+(z-zi′)2
(see Fig. 7.17).

Figure 7.17 Gravitational field at x due to a small volume of the body

The gravitational potential associated with this field is

(7.106)
△ Φi=-Gρ(xi′ )△ Viri.
Hence, the potential due to the whole mass is given approximately by
(7.107)
Φ=-∑i△ Φi=-∑iGρ(xi′ )△ Viri.
By letting ▵ V i → 0, the sum in (Equation 7.107) will be replaced by the volume integral over V
Φ(x,y,z)=-∫VGρ(x′ )dVr
(7.108)
=-∫VGρ(x′ ,y′ ,z′ )dx′ dy′ dz′ (x-x)2+(y-y′ )2+(z-z)2.
Notice that the integral is over the volume of the body whose coordinates are denoted by x ’, y ’, and
z ’, whereas the coordinates of the point where the potential is being computed are denoted by x, y,
and z.
 To find the differential equation satisfied by Φ ( x , y , z ) , we compute
(7.109)
∇ Φ=(∂Φ∂x,∂Φ∂y,∂Φ∂z)
and ∇ ⋅ (∇Φ) = ∇2 Φ to show that
(7.110)
∇ 2Φ=∂2Φ∂x2+∂2Φ∂y2+∂2Φ∂z2=0.
This equation is called the potential or Laplace equation (in three dimen‐ sions).
Compounding
In various applications, we must consider other equations that are closely related to the Laplace
equation in three dimensions. Three examples of such equations follow:
1. The Laplace equation in n dimensions is
(7.111)
∇ 2u=∑i=1n∂2u∂xi2=0.
1. The Poisson equation is given by
(7.112)
∇ 2u=fx
It can be shown that the gravitational field inside a body satisfies such an equation with f (x) =
4np(x). (See Exercise 6.)
1. The Helmholtz equation is
(7.113)
∇ 2u+k2u=0
where k is a constant.
Example 7.5.1
Compute the gravitational potential of a solid sphere of ra‐ dius a and a constant mass density ρ.
Solution 7.5.1 As noted earlier, in all such problems we first compute the gravitational potential
Φ and then evaluate the gravitational field F as ▵ Φ.
In this problem, because the sphere is isotropic, the gravitational potential must be the same for all
points whose distance from the center of the sphere is the same. It follows that if we let the center of
the sphere coincide with the origin of the coordinate system, then all points with r = constant must
have the same potential. It is enough, therefore, to evaluate Φ at ( 0 , 0 , z ) with 0 < a < z. To do so,
we infer from (Equation 7.108) that
(7.114)
Φ(0,0,z)=-Gρ∫Vdx′ dy′ dz′ (x′ )2+(y′ )2+(z-z′ )2.
To compute this integral, we introduce spherical coordinates on x ’, y ’ and z ’. This transformation is
given by
x ′ = r sin ϕ cos θ
y ′ = r sin ϕ sin θ
(7.115)
z ′ = r cos ϕ
and
 (7.116)
d V = r 2 sin ϕ d ϕ d r .
If we write the volume integral for Φ ( 0 , 0 , z ) as an iterated integral, we see that
(7.117)
Φ ( 0 , 0 , z ) = - G ρ ∫ 0 a ∫ 0 π ∫ 0 2 π r 2 sin ϕ d θ d ϕ d r r 2 + z 2 - 2 r z cos ϕ = - 2 π G ρ ∫ 0 a ∫ 0
π r 2 sin ϕ d ϕ d r r 2 + z 2 - 2 r z cos ϕ .
To compute the integral over φ, we introduce the substitution
(7.118)
w = r 2 + z 2 - 2 r z cos ϕ .
Remembering that z is a constant in this computation, we obtain
Φ(0,0,z)=-πGρz∫0ardr∫(r-z)2(r+z)2w-1/2dw
(7.119)
=-2πGρz∫0ardr·w1/2|(r-z)2(r+z)2.
But
(7.120)
w1/2|(r-z)2(r+z)2=(r+z)-|r-z|=2r
since r ≤ a ≤ z, and it follows that
(7.121)
Φ(0,0,z)=-4πa3ρG3z=-GMz
where is the total mass of the body. For a general point whose distance from the origin is R, we
obtain (due to symmetry)
(7.122)
Φ(R)=-GMR.
(Equation 7.122) implies that the potential of a solid sphere of mass M is equivalent to that of a point
particle with the same mass situated at its center.

7.5.1 Kirchoff Transformation

In many practical situations, one is called to consider a nonlinear version of Laplace equation in the
form
(7.123)
∇ ·(D(ϕ)∇ ϕ)=0.
We can linearize this equation by introducing
(7.124)
ψ=∫aϕD(t)dt
where a is an arbitrary constant. (This is called Kirchoff transformation.) We then have
(7.125)
dψdϕ=D(ϕ).
By the chain rule we obtain
(7.126)
∇ ψ=dψdϕ∇ ϕ=D(ϕ)∇ ϕ.
Hence, from (Equation 7.123) we infer that
(7.127)
∇ 2ψ=∇ ·(∇ ψ)=0;
i.e., (Equation 7.123) has been reduced to Laplace equation. We can recover ϕ from ψ using
Equation (7. 125).
Exercises
1. Compute the gravitational field of a spherical shell with constant density p , inner radius Rx,
and outer radius R 2 .
Hint. Consider two cases: one for a point outside the shell and another for a point inside the cavity.
1. Derive an expression for the gravitational potential of a thin metal wire bent to form a circle
of radius a if its mass density per unit length is p.
2. Derive a model equation for the gravitational field generated by a flat circular ring a < r <
b whose mass density per unit area p is constant. Consider points in the ring plane only.
3. Compute the expression for the two-dimensional Laplace equation in polar coordinates.
4. Compute the expression for the Laplace equation in three dimensions in cylindrical and
spherical coordinates.
5. Derive the Poisson equation as follows:
6. Show that the gravitational field for a point of mass m satisfies
∫SF·dS=4πm
where S is an (arbitrary) sphere around m.
Then deduce that the gravitational field of a continuous mass distribution satisfies
∫SF·dS=4π∫VρdV.
Use the fact that F = νφ and the divergence theorem to obtain the Poisson equation.
Show that ln ( x 2 + y 2 ) and 1 / x 2 + y 2 + z 2 are solutions of the Laplace equation in two
and three dimensions, respectively.
7.6 The Continuity Equation
Objective
Derive a model equation for the traffic flow on a highway without exits and with one entrance and
one lane (see Fig. 7.18).

Figure 7.18 Cars on a one lane highway

Discussion
One possible approach to modeling the traffic flow is to describe each car as a finite element on
the highway and then write a discrete model, which describes the motion of each such car. However,
if there are many cars on the highway, this approach is not practical, and it is better to construct a
continuous model, which treats these cars as “smeared out” quantities. We construct such a
continuous model in this section.
Approximation and Idealization
We assume that the highway is infinite.
1. We define the car density p ( x,t ) as
(7.128)
ρ(x,t)≅ Numberofcarsontheinterva1[x,x+△ x]attimet△ x
where Δχ must be large compared to a car length. [Otherwise, p(x, t) = 1 if there is a car at x in time
t, or p(x,t) = 0 if there is none.] In fact, the same approximation is made whenever we define the mass
density of a “continuous” body made of discrete atoms and molecules. Hence, the equations we
derive in this section apply also to fluid flow in a pipe.
1. We assume that there are no accidents on the highway (or that their number is negligible).
Hence, we can formulate the principle of “car conservation” (which is equivalent to that of
mass conservation) as follows:
The rate at which the number of cars on the segment [a,b] is changing equals the rate at which
they enter less the rate at which they are leaving.
1. We define the concept of car flux q ( x, t ) in the same way that we defined this concept in
Section 2. However, in this context it is natural to define the flux per lane rather than per
unit area. Equivalently, this can be considered as letting the unit length be equal to the width
of the lane. Moreover, note that
 (7.129)
q(x,t)=ρ(x,t)u(x,t)
where u ( x, t ) is the car’s speed at x at time t, and the dimension of q is
(7.130)
q=CarsTime.
Modeling
To derive a model equation for the traffic flow, consider a finite section of the road between a and
b . The number n of cars in this segment at time t is
(7.131)
n(t,a,b)=∫abρ(x,t)dx.
Hence, the rate of change in this quantity is
(7.132)
dndt=∫ab∂ρ(x,t)∂tdx.
This rate of change must equal the flux of cars entering at a less the flux of cars leaving at b .
(Remember that the flux was defined per lane!) Therefore,
(7.133)
dndt=q(a,t)-q(b,t)=-∫ab∂q∂x(x,t)dx.
Thus, we infer from (Equations 7.132) and (7.133) that
(7.134)
∫ab(∂ρ∂t+∂q∂x)dx=0.
(7.135)
∂ρ∂t+∂q∂x=0.
Using (Equation 7.129) to substitute for q ( x , t ) , we finally obtain
(7.136)
∂ρ∂t+∂(ρu)∂x=0.
This equation is the continuity equation in one dimension. Notice, how‐
ever, that this equation contains two unknown quantities, ρ and u. Therefore, to solve it we must
either be able to express u = u(ρ) or find an additional equation that relates these two quantities.
Compounding
To derive the version of the continuity equation in three dimensions, we consider a fluid flow with
mass density ρ ( x , t ) . Let V be a volume contained in the flow. The mass of the fluid in V at time t is
given by
(7.137)
m(t,V)=∫Vρ(x,t)dx.
Hence, the rate of change of mass in V is
(7.138)
dmdt=∫V∂ρ∂t(x,t)dx.
Now let S denote the boundary of V and n(x) the unit outward normal to S at x. The total mass flow
rate of the fluid across S in the outward direction is
(7.139)
∫Sq·ndS=∫S(ρu·n)dS.
The mass conservation principle implies, however, that the rate of change of mass in V must equal the
rate at which the mass is crossing S in the inward direction. Therefore,
(7.140)
∫V∂ρ∂t(x,t)dV=-∫Sρu·ndS.
To convert the right‐hand side of (Equation 7.140) into a volume integral, we now invoke the
divergence theorem, which states that for any smooth vector field F in V
(7.141)
∫SF·ndS=∫VdivFdV.
This yields
(7.142)
∫V[∂ρ∂t+div(ρu)]dV=0.
Since V is arbitrary, we infer that the integrand in (Equation 7.142) must be zero, or
(7.143)
∂ρ∂t+div(ρu)=0.
(Equation 7.143), which is a first‐order partial differential equation, is called the continuity equation
in three dimensions.
Exercises
1. From your experience, guess the general form of the relationship between u and p in a one-
lane highway.
2. Derive a model equation for an infinite one-lane highway where cars are entering and
leaving the highway at constant rates a and β per mile, respectively. Generalize to the case
where a = a ( x,t ) and β = β(x,t).
3. Compound the model in Exercise 2 to include accidents at a rate a ( x, t ) per mile.
4. Consider fluid flow in a long cylindrical pipe with constant cross section A whose axis is
along the x-axis. Let p ( x,t ) be the density of the fluid and q ( x, t ) be its flux. If the walls
of the pipe are made of porous material that allows the fluid to leak out at a rate L per unit
length, show that
∂ρ∂t+∂q∂x+LA=0.
1. Derive model equations for the car densities p1(x,t) and p2(x,t) in a two-lane infinite
highway with no entries or exits where cars are moving from lane 1 to lane 2 at a rate of a (
p 2 ) per mile and at a rate of b(p1) per mile from lane 2 to lane 1.
2. Compound the model of Exercise 5 to include entries and exits.

7.7 ELECTROMAGNETISM

7.7.1 Maxwell Equations

Maxwell equations form the core of all electromagnetic theory and applications. These equations (in
the rationalized unit system) are:
(7.144)
∇ ·D=ρ
 (7.145)
∇ ·B=0
(7.146)
∇ ×E=-∂B∂t
(7.147)
∇ ×H=J+∂D∂t
Here

E - electric field B - magnetic field D - electric displacement H - magnetic intensity J - current

density p - charge density.

To close these equations we need relations, D = D(E), H = H(B), and J = J(E) . In vacuum or in
homogeneous isotropic medium, we have the linear relations
(7.148)
D=εE,H=Bμ
where ɛ and μ are called the permittivity and permeability of the medium. Furthermore, in a
homogeneous conducting medium
(7.149)
J=σE
where σ is the medium conductivity. In the following, we assume the relations, (Equations 10.93) and
(10.94) implicitly.
By taking the divergence of (Equation 10.92) and using (Equation 10.89), we obtain the (electric
charge) continuity equation
(7.150)
∂ρ∂t+∇ ·J=0.
Remark 7.7.1
If a point charge ρ is moving with velocity y, then the induced current density is J = ρv.
Remark 7.7.2
The total force exerted by an electromagnetic field on charges and currents in a volume V is given
by the Lorentz force equation
(7.151)
F=∫V(ρE+J×B)dV.
Remark 7.7.3
To treat the special electromagnetic fields discussed in the rest of this section, we need the
following facts:
∇ × (∇ × A) = ∇(∇ ⋅ A) - ∇2 A
If ∇ × A = 0, then A is conservative and therefore there exists a scalar function ϕ so that
(7.153)
A=∇ ϕ.
If ∇ ⋅ A = 0, then A is solenoidal and, therefore, there exists a vector potential B so that
(7.154)
A=∇ ×B.
Furthermore, B can be chosen so that ∇ ⋅ B = 0

7.7.2 Electrostatic Fields

When p and J are time independent, it follows from (Equations 10.89)-(10.92) that the electric and
magnetic fields are also time independent. Furthermore, (Equations 10.89) and (10.91) decouple from
(Equations 10.90) and (10.92) and the equations for the electric and magnetic fields can be solved
separately.
For the electric field we have
(7.155)
∇ ×E=0
(7.156)
∇ ·E=ρε.
From (Equation 9.149), it follows that E is conservative and, hence, there exists a potential function φ
s.t.
(7.157)
E=-∇ ϕ.
From (Equation 9.150), we obtain
(7.158)
∇ 2ϕ=-ρε.
This is a Poisson equation whose solution can be obtained in integral form as (see Sec 3)
(7.159)
ϕ(r)=14πε∫Vρ(r′ )|r-r|dr′ .
The electric field can be obtained by taking the gradient of this equation
(7.160)
E(r)=14πε∫Vρ(r′ )(r-r′ )|r-r|3dr

7.7.3 Multipole Expansion

When |r - r ’| > > 1, we can obtain an approximation to the integral in (Equation 9.154) using the
expansion
(7.161)
1|r-r′|3=1[r2-2r·r′+(r′)2]3/2=1r3[1+3r·r′r2+⋯ ].
Substituting this expansion in (Equation 9.154), we obtain
(7.162)
E=Qr4πεr3+q4πεr5+⋯
where
(7.163)
Q=∫Vρ(r′ )dr′
(7.164)
qi=∑j(3xixj-r2δij)pj,r=(x1,x2,x3)
(7.165)
pj=∫Vxj′ ρ(r′ )dr′
p is called the dipole moment of the charge distribution. Observe that when Q = 0, the dipole term is
the leading term in the expansion (Equation 9.156).

7.7.4 Magnetostatic
The equations for the magnetic field in this case are
(7.166)
∇ ·B=0
(7.167)
∇ ×B=μJ.
From (Equation 7.166), we infer that B is a solenoidal vector field. Hence, there exists a vector
potential A so that
(7.168)
B=∇ ×A.
Hence, from Equation (7. 167)
(7.169)
∇ ×(∇ ×A)=μJ.
If we choose A so that ∇ ⋅ A = 0, we obtain
(7.170)
∇ 2A=-μJ,
i.e., A satisfies a vector Poisson equation. Hence,
(7.171)
A=μ4π∫VJ(r′ )dr′ |r-r|.
By taking the curl of this equation, we obtain Biot‐Savart law
(7.172)
B=μ4π∫VJ(r′ )×(r-r′ )|r-r|3dr.
Since the total magnetic charge is always zero, the multipole expansion for B yields
(7.173)
B=μM4πr5
where
(7.174)
Mi=∑j(3rirj-r2δij)mj
(7.175)
m=12∫Vr′ ×J(r′ )dr′ .

7.7.5 Electromagnetic Waves

In vacuum (and with no currents)
(7.176)
∇ ·E=0,∇ ·B=0
(7.177)
∇ ×E=-∂B∂t,∇ ×B=εμ∂E∂t.
By taking the curl of (Equation 7.177) and using (Equations 10.96) and (7.176), we obtain
(7.178)
∂2E∂t2=c2∇ 2E,∂2B∂t2=c2∇ 2B
where c 2 = 1 ε μ . Thus, both E and B satisfy the wave equation in three dimensions.

7.7.6 Electromagnetic Energy and Momentum

Consider a point charge ρ moving with velocity y in an E - M field. The force acting on the charge is
(from (Equation 10.97))
(7.179)
F=ρE+ρv×B.
The rate of work done by this force is
(7.180)
W=F·v=ρv·E=J·E.
We infer then that the rate of work done by an E - M field on a continuous distribution of currents is
(7.181)
W=∫J·EdV.
This power (which is converted into thermal or mechanical energy) must be accompanied by an equal
rate of decrease in the E - M energy in the volume V. To obtain an explicit expression for the
conservation of energy, we eliminate J in (Equation 7.180) using (Equation 10.92)
(7.182)
∫J·EdV=∫E·(∇ ×H-∂D∂t)dV.
Using the vector identity
(7.183)
∇ ·(E×H)=H·(∇ ×E)-E·(∇ ×H)
and (Equation 10.91), we have
(7.184)
E·(∇ ×H)=-H·∂B∂t-∇ ·(E×H).
Hence,
∫VJ·EdV=-∫V[H·∂B∂t+E·∂D∂t+∇ ·(E×H)]dV
(7.185)
=-∫V[12∂∂t(H·B+E·D)+∇ ·(E×H)]dV
where we have used the relations (10.93). Using the divergence theorem, we can rewrite (Equation
7.185) as
(7.186)
∫VJ·EdV=-∫12∂∂t(H·B+E·D)dV-∫S(E×H)·dS.
Since the volume in (Equation 7.185) is arbitrary, we can re-express this equation in a differential
form
(7.187)
∂U∂t+∇ ·P=-J·E
where
(7.188)
U=12(E·D+B·H),P=E×H.
U represents the E - M energy density and P, the Poynting vector, is the energy flux density. Observe
that only the divergence of P appears in (Equation 7.185). It follows then that we can change the
definition of the Poynting vector by adding the curl of an arbitrary vector field without changing the
physical contents of the theory.
To treat the conservation of linear momentum, we use (Equations 10.89) and (10.92) to eliminate ρ
and J from (Equation 10.97)
(7.189)
ρE+J×B=E(∇ ·D)+(∇ ×H-∂D∂t)×B=E(∇ ·D)+(∇ ×H)×B-[∂∂t(D×B
)-D×∂B∂t]=E(∇ ·D)-B×(∇ ×H)-D×(∇ ×E)â€​∂∂t(D×B).
By adding H(∇ ⋅ B) = 0 to the right hand side of this expression, we obtain:
(7.190)
∫V(ρE+J×B)dV+ddt∫V(D×B)=∫V[E(∇ ·D)-D×(∇ ×E)+H(∇ ·B)-B×
(∇ ×H)]dV=∫V∇ βTα,βdV.
Using the divergence theorem, we can rewrite this as
(7.191)
∫V(ρE+J×B)dV+ddt∫V(D×B)=∫S∑βTndS
where n is the normal to S and
(7.192)
Tαβ=EαDβ+HαBβ-12(E·D+B·H)δαβ
is Maxwell stress tensor.
1. The first and second terms on the left hand side of (Equation 7.191) rep‐ resent
(respectively) the rate of change in the linear moments of the particles and the E - M field in
the volume V. The right hand side of this equation is interpreted as the force per unit area
transmitted across S and acting on the fields and particles in V.

7.7.7 Electromagnetic Potential

Maxwell (Equations 10.89)-(10.92) with the relations (Equations 10.93) and (10.94) represent eight
equations for the six quantities E and B. We infer then that a more compact formulation may exist in
terms of only four unknowns. To see how this can be done, we first observe that (Equation 10.90) can
be satisfied identically if
(7.193)
B=∇ ×A.
Substituting this in (Equation 10.91), we obtain
 (7.194)
∇ ×(E+∂A∂t)=0,
i.e., E + ∂ A ∂ t is a conservative field. It follows then that there exists a scalar field φ so that
(7.195)
E=-∇ ϕ-∂A∂t.
We observe that the relations, (Equations 7.193) and (7.195) remain unchanged under the “gauge”
transformations
(7.196)
A→ A′ =A+∇ ∧
(7.197)
ϕ→ ϕ′=ϕ-∂∧ ∂t;
i.e., the physical contents of the theory (which is represented by E and B) remains unchanged under
these transformations.
Two particular choices of ∧ are of special interest.
1. Lorentz gauge
Under this gauge, ∧ is chosen so that
(7.198)
∇ 2∧ =∂2∧ ∂t2.
As a result, the quantity
(7.199)
I=∇ ·A+∂ϕ∂t=∇ ·A′ +∂ϕ′ ∂t
remains unchanged. Choosing the value of I to be zero and substituting (Equations 7.193) and (7.195)
in (Equations 10.89) and (10.92) leads to
(7.200)
(∇ 2-∂2∂t2)A=-J
(7.201)
(∇ 2-∂2∂t2)ϕ=-ρ/ε.
1. Coulomb gauge
Here we impose on ∧ the condition
(7.202)
∇ 2∧ =0.
As a result,
(7.203)
∇ ·A=∇ ·A′
and we can choose therefore ∇ ⋅ A = 0. With this choice, we substitute (Equations 7.193) and
(7.195) in (Equations 10.89) and (10.92) to obtain
(7.204)
(∇ 2-∂2∂t2)A=-J+∂∂t(∇ ϕ)
(7.205)
∇ 2ϕ=-ρ.
Remark 7.7.4
Sometimes the Coulomb gauge is referred to as the “Radiation gauge.”
Exercises
1. In free space ( i . e . , ρ = 0 , J = 0 ) , let the electric and magnetic field be given by
E=F(z-t),F=(f1,f2,f3)
B=G(z-t),G=(g1,g2,g3)
(plane waves). Show that to satisfy Maxwell equations we must require that
f3=g3=0,f1=g2,f2=-g1.
1. Show that for
f=f~(t)e-ik·r,F=F~(t)e-ik·r
∇ f=-ikf,∇ ·F=-ik·F,∇ ×F=-ik×F.
1. Use Ex. 2 to show that if E = E 0 e i(ωt-k⋅r), then B = B 0 e i(ωt-k⋅r), then B = (k × E)/ω, and
k ⋅ E = 0; i.e., the electric and magnetic fields are orthogonal to each other and k (the
propagation direction) is orthogonal to both.
2. Consider EM field in a homogeneous conducting medium where J = aE, and suppose ρ = 0
(change density). Show that
∂2E∂t2+σε∂E∂t=c2∇ 2E,c2=1εμ
 CHAPTER 8

Solutions of Partial Differential Equations

CONTENTS
8.1 Method of Separation of Variables
8.1.1 Method of Separation of Variables By Example
8.1.2 Non Cartesian Coordinate Systems
8.1.3 Boundary Value Problems with General Initial Conditions
8.1.4 Boundary Value Problems with Inhomogeneous Equations
8.2 Green’S Functions
8.3 Laplace Transform
8.3.1 Basic Properties of the Laplace Transform
8.3.2 Applications to the Heat Equation
8.4 Numerical Solutions Of PDES
8.4.1 Finite Difference Schemes
8.4.2 Numerical Solutions for the Poisson Equation
8.4.2.1Other Boundary Conditions
8.4.3 Irregular Regions
8.4.4 Numerical Solutions for the Heat and Wave Equations
8.1 Method of Separation of Variables
When it comes to solving boundary value problems involving partial differential equations, a number
of approaches are available. The method of separation of variables is very convenient because it
draws on many well‐known mathematical concepts and frequently works well. In this section, we
introduce this method and illustrate its application by considering various systems that were modeled
in the last chapter by partial differential equations. In particular, we concentrate on the heat, wave,
and potential equations that are important in many scientific and engineering applications.
The general objective of this method is to reduce the solution of a given partial differential
equation into the solution of a number of ordinary differential equations. Very often these ordinary
differential equations are well known, and their solutions are easily found. The whole process
follows a logical step‐by‐step development that terminates in the evaluation of the Fourier
coefficients of a Fourier‐type series. We suggest that you carefully catalog each step and its position
in the stairway to a successful conclusion.

8.1.1 Method of Separation of Variables By Example

Although the method of separation of variables contains many steps, it does follow a set pattern as we
move toward the solution. Probably the best way to understand the method is through observing a
number of examples. We start with a straightforward heat flow problem.
Example 8.1.1 For an introduction to this process, we will seek the solution of the following
boundary value problem:
∂u∂t=k∂2u∂x20<x<L,0<t
u(0,t)=00<t
u(L,t)=00<t
u(x,0)=x0<x<L.
This is a one‐dimensional heat flow problem in a rod of length L. We are to find the temperature u
( x , t ) if the temperature at the right- and left-hand ends is always zero and the initial
temperature in the rod is x.
Solution 8.1.1 : Step 1. We seek “elementary” solutions of the partial differential equation in the
special form ( x , t ) = X ( x ) T ( t ) . What we are suggesting is that the variable x occurs only in
the function X, whereas T is a function of t only. This device does not always work, but it does
solve many engineering problems and is usually a good method to use as a first approach.
Substituting u ( x , t ) = X ( x ) T ( t ) into the differential equation leads to
(8.1)
X(x)T′ (t)=kX′ ′ (x)T(t)
Notice that we are able to use the “prime” notation for the derivatives because each factor
depends only on one variable; that is,
T′(t)=dT(t)dt
X′ ′ (x)=d2Xdx2.
Step 2. We next see if it is possible to separate the variables. Is it possible to get all the functions
dependent on x‘s on one side of the equation and all those dependent on t‘s on the other side ’?
Since we are not looking for trivial solutions where either X or T is identically zero, we can divide
both sides of Equation (12.30) by X(x)T(t) provided X(x)T(t) ≠ 0. Thus,
1kXT′ XT=X′ ′ TXT
or
(8.2)
1kT′ (t)T(t)=X′ ′ (x)X(x)
Now pick a fixed value t = t 0. Then
1 k T ′ ( t 0 ) T ( t 0 ) = constant = X ′ ′ ( x ) X ( x ) .
But this is true for any x on the right‐hand side of the equation and , therefore (since the left
hand side is a constant),
X′ ′ (x)X(x)=c
In the same way we can show that
T′(t)T(t)=c
by fixing x = x 0. Observe that this constant is the same for both sides of Equation (12.31). In this
case (for convenience) let the constant be equal to - λ . Equation (12.31) becomes
(8.3)
1kT′ (t)T(t)=X′ ′ (x)X(x)=-λ
Step 3. From Equation (12.32) we see that
(8.4)
X′ ′ (x)+λX(x)=0
(8.5)
T′(t)+λkT(t)=0
which are two ordinary differential equations that are easily solved.
Comment. If X(x) or T(t) is zero at x = x 0 and t = t 0, respectively, the result in Step 3 is still
valid. Suppose X(x 0) = 0. Returning to Equation (12.30) (before we divide by XT) and substituting
X(x 0) = 0 into Equation (ref4.1), we have
0=kX′ ′ (x0)T(t)
Since this equation must hold for all t > 0, and T(t) is not identically zero, there must be some t = t
’’
0 so that T(t 0) ≠ 0. Since k > 0, it follows that X (x 0) = 0, and we can state that
X′ ′ (x0)+λX(x0)=0
is satisfied. The argument for showing that
T′(t0)+λkT(t0)=0
for T(t 0) = 0 is similar. □
Step 4. Before solving Equations ( 8.4 )‐(8.5), let us find the boundary conditions that go with
these ordinary differential equations. Now u ( 0 , t ) = X(0)T(t) = 0. If T(t) = 0 for all t, then u ( x , t
) = X ( x ) T ( t ) = X ( x ) · 0 = 0 . But if u ( x , t ) ≡ 0 , it is known as the trivial solution. Since we
are looking for nontrivial solutions, we must avoid setting T(t) = 0. Therefore,
X(0)=0
is one boundary condition.
 In a similar fashion we can show that
X(L)=0
is another boundary condition. In this problem there are no initial conditions for the first‐order
equation in t.
Warning. Do not attempt this type of argument on a nonhomogeneous boundary condition. For
example, the initial condition u ( x , 0 ) = X ( x ) T ( O ) = x cannot be solved uniquely for T(0).
Only when the product of two quantities equals zero can we conclude that one or both of the
factors is zero.
Step 5. Now we have the following ordinary differential equations plus boundary conditions
X′ ′ +λX=0X(0)=0X(L)=0
T′+λkT=0.
The first of these two equations with the two boundary conditions is known as an eigenvalue
problem and is solved first. Unfortunately, λ is unknown to us at the moment, and therefore we
must look at three problems because the character of the solution changes with each case. We shall
look at the eigenvalue problem for λ > 0 . λ = 0 , and λ < 0 .
Case 1. λ > 0 . Let α 2 = λ > 0 (again for convenience). Then the general solution to X ’’(x) + α 2
X(x) = 0 is
X ( x ) = A cos α x + B sin α x .
Now since X(0) = 0 we have
X(0)=0=A·1+B·0
which tells us that
A=0
and
X ( x ) = B sin α x .
Using the second boundary condition X(L) = 0, we have
X ( L ) = 0 = B sin α L .
Now if B = 0, then since A = 0, X(x) ≡ 0, which implies u ( x , t ) = X(x)T(t) ≡ 0, which is the trivial
solution we are trying to avoid. Therefore, sin α L = 0 . But recall from trigonometry that n π ( n = 0
, ± 1 , ± 2 , … ) are angles that make the sine zero, and therefore we write
αL=nπ
so that
α=nπL
and
λ = λ n = α n 2 = n 2 π 2 L 2 f or n = ± 1, ± 2,....
Note. The value n = 0 is dropped since this implies α = 0, and we are assuming α > 0. This case
is covered next.
The values of λ in Case 1 are called the eigenvalues of the boundary value problem.
Now the solutions of the ordinary differential equation and its boundary conditions
corresponding to these eigenvalues λ n are
X n = sin n π x L n = ± 1 , ± 2,...
These solutions are called eigenfunctions.
 Case 2. λ = 0 . The general solution of X ’’ = 0 is
X(x)=A+Bx.
Since
X(0)=0=A
and
X(L)=BL=0
implies B = 0, we see that the only solution to this problem is X(x) ≡ 0, which is the trivial
solution. λ = 0 is not an eigenvalue.
Note. Eigenfunctions corresponding to an eigenvalue λ must be nontrivial. Case 3. λ < 0 . Let λ =
- α 2 < 0 . The differential equation X’’ + α 2 X = 0 has the general solution
X(x)=C1eαx+C2e-αx
Although this solution may be used, it is more convenient (in this case) to write the solution in
terms of hyperbolic functions instead. Since
sinh α x = e α x - e - α x 2
and
cosh α x = e α x - e - α x 2 ,
which are linear independent functions, we can rewrite the general solution as
X ( x ) = A cosh α x + B sinh α x .
Now, since X(0) = 0 we have
X(0)=0=A·1+B·0
which shows that
A=0
and X ( x ) = B sinh α x .
Using X(L) = Bsinh αL = 0, we find either B = 0, which leads to the trivial solution or sinh α L =
0 . But sinh0 = 0 is the only zero of the hyperbolic sine function. Therefore, αL = 0. And since L ≠ 0,
it follows that α = 0. But this is a contradiction since we are assuming - α 2 < 0.
There are no nontrivial solutions under this case, which tells us that there are no negative
eigenvalues.
This long discourse completes the solution of the eigenvalue problem posed in Step 5.
Step 6. We can now shift to the differential equation
(8.6)
T′ +λkT=T′ +n2π2L2kT=0
Notice that the eigenvalues are now known to us from Step 5, and it is possible to solve Equation
(12.33) up to an arbitrary constant. The general solution to the first‐order equation is
T n ( t ) = C n e x p ( - k n 2 π 2 L 2 t ) n = ± 1 , ± 2,....
In solving boundary value problems by this approach, we find it convenient to
set all C n s = 1 because we will be multiplying each solution by an arbitrary constant later on.
Step 7. To continue with the solution, we multiply X n by T n because in Step 1 we assumed a
solution of the partial differential equation of the form ( x , t ) = X ( x ) T ( t ) . Therefo re, we can
write
u n ( x , t ) = e x p ( - k n 2 π 2 L 2 t ) sin n π x L n = ± 1 , ± 2, . . . .
 For = ± 1, ± 2, . . ., u n ( x , t ) satisfies the partial differential equation and the two boundary
conditions.
We still have to match the initial condition u ( x , 0 ) = x . To do this, we form an infinite linear
combination of all solutions u n ( x , t ) . This is made possible by the fact that the heat equation is
linear and the superposition principle applies. It follows then that the general solution to the
partial differential equation and boundary conditions can be written as
(8.7)
u ( x , t ) = ∑ n = 1 ∞ b n u n ( x , t ) = ∑ n = 1 ∞ b n e x p ( - k n 2 π 2 L 2 t ) sin n π x L
Note. When dealing with sines and cosines, one may question whether it is necessary to consider n
for negative integers. However, since sin ( - n π x / L ) = - sin ( n π x / L ) and cos ( - n π x / L ) =
cos ( n π x / L ) , each solution derived from a negative n can be combined with the corresponding
solution derived from a positive n. From here on we will assume this is done and let n = 1, 2,....
Step 8. Finally, using the condition ( x , 0 ) = x , 0 < x < a, we must satisfy
u ( x , 0 ) = x = ∑ n = 1 ∞ b n sin n π x L
since the time exponential is 1. But this is just a Fourier sine series when x on ( 0 , L ) is extended
into ( - L , 0 ) as an odd function, and it follows that
(8.8)
b n = 2 L ∫ 0 L x sin n π x L d x
= 2 L 2 L n 2 π 2 { sin n π x L - n π x L cos n π x L } | 0 L
=2Lnπ{-cosnπ}=-2Lnπ(-1)n
or
(8.9)
bn=(-1)n+12Lnπ
We have now determined all the constants b n . Substituting their value from Equation (12.35) in
Equation (12.34), we see that the final solution is
(8.10)
u ( x , t ) = 2 L π ∑ n = 1 ∞ ( - 1 ) n + 1 n e x p ( - k n 2 π 2 L 2 t ) sin n π x L
It is possible to garner some information about the solution u ( x , t ) without too much effort. True,
if we have access to a computer, we could numerically evaluate a large number of terms of the
series to construct a table of values or even plot a graph of an approximation to the solution. Of
course, how well this can be done depends on the convergence properties of the series which
represents the solution.
We observe that the solution we found for u ( x , t ) is a transient solution; that is, u ( x , t ) → 0
as t → + ∞. This follows easily from the fact that each term in the series is of the form e -αt where
α = kn 2 π 2/L 2. Carrying this idea a bit further, we notice that n enters into the exponent as n 2,
which indicates that e -αt decreases very rapidly as n increases. Therefore, most of the information
about the solution is carried in the first few terms of the series.
For example, if we let = 1, L = π, and t = 1, and compare the e -αt terms for n = 1, 2, 3, we shall
find that the magnitude of consecutive terms in the series is decreasing exponentially.
We now look at an example in which λ = 0 is an eigenvalue.
Example 8.1.2 : Consider a high‐frequency transmission line whose length is 40 meters.
Suppose the rate of change of voltage with respect to x at both ends is zero. If the initial voltage is
f(x) and the rate of change of voltage with respect to t is zero, find the solution as a function of x
and t. This problem can be written formally as
∂ e ∂ x ( 0 , t ) = 00 < t
∂ e ∂ x ( 40 , t ) = 00 < t
e ( x , 0 ) = f ( x ) 0 < x < 40
∂ e ∂ t ( x , 0 ) = 0 0 < x < 40 .
Solution 8.1.2 The differential equation is
∂ 2 e ∂ x 2 = L C ∂ 2 e ∂ t 2 0 < x < 40 , 0 < t .
Steps 1 and 2. Let ( x , t ) = X ( x ) T ( t ) . Substituting into the differential equation, we find
X′ ′ X=LCT′ ′ T=-λ.
Step 3. This result leads to the two ordinary differential equations
X′ ′ +λX=0
T′ ′ +λLCT=0.
Step 4. The three homogeneous boundary and initial conditions yield the following conditions for
the ordinary differential equations:
X ′ ( 0 ) = 0 , X ′ ( 40 ) = 0 , T ′ ( 0 ) = 0 .
Steps 5 and 6. When we solve the differential equation in x, we again consider three cases: λ > 0 ,
λ = 0 , and λ < 0 .
Case 1. λ > 0 . Let α 2 = λ > 0 . The general solution of x ’’ + α 2 X = 0 is then
X ( x ) = A cos α x + B sin α x
and
X ′ ( x ) = - α A sin α x + α B cos α x .
Applying the boundary conditions, we find
X′ (0)=-αA·0+αB·1=0
for which
αB=0.
However, since > 0, B = 0, and X ′ ( x ) = - α A sin α x , then X ’(40) = - α A sin 40 α = 0 . Neither
α nor A can equal zero, so we can write
sin 40 α = 0
from which it follows that
40 α = n π n = 1 ,
α = n π 40 .
The positive eigenvalues are
λ n = α 2 = [ n π 40 ] 2
and the eigenfunctions are
X n ( x ) = cos n π x 40 .
Solving the other differential equation, we have
T n ( t ) = D n cos n π t 40 L C + E n sin n π t 40 L C
and
T n ′ ( t ) = - D n n π 40 L C sin n π t 40 L C + E n n π 40 L C cos n π t 40 L C .
Using the condition
T n ′ ( 0 ) = E n π n 40 L C = 0
we observe that
En=0.
Therefore,
T n ( t ) = cos n π t 40 L C n = 1 ,
Combining X n and T n we find
e n ( x , t ) = cos n π x 40 cos n π t 40 L C n = 1 ,
Case 2. λ = 0 . In this case, X ’’ = 0 and X(x) = A + Bx. It follows that
X′ (x)=B
and
X′ (0)=B=0.
But now X ’(x) = 0, for all x and in particular for x = 40; that is, X ’(40) =
0. Therefore, both boundary conditions are satisfied: λ = 0 is an eigenvalue
and X(x) = A (a constant) is an eigenfunction. We find it convenient to let
A = 1 and use X(x) = 1 as our eigenfunction. To identify the eigenvalue and
eigenfunction associated with λ = 0 , we write λ 0 = 0 and X 0 = 1. Now solving
T ’’ = 0 we have T(t) = Dt + E and T ’(t) = D. Using the initial condition
T ’(0) = 0, we find D = 0 and T 0 = 1. Therefore, e 0 ( x , t ) = X 0 T 0 = 1 .
Case 3. λ < 0 . In a way similar to Case 3 of Example 1 we can show that
there are no negative eigenvalues.
Step 7. To finish the problem we form an infinite linear combination of all
the solutions e 0 and e n , which yields
(8.11)
e ( x , t ) = a 0 2 e 0 ( x , t ) + ∑ n = 1 ∞ a n e n ( x , t ) = a 0 2 + ∑ n = 1 ∞ a n cos n π x 40 cos n π t
40 L C .
Step 8. We use our final condition, which tells us that
e ( x , 0 ) = f ( x ) = a 0 2 + ∑ n = 1 ∞ a n cos n π x 40 on (0, 40).
Since this is just a Fourier cosine series, we know that
a 0 = 1 20 ∫ 0 40 f ( x ) d x
and
a n = 1 20 ∫ 0 40 f ( x ) cos n π x 40 d x .
Substituting these values back into Equation (12.36), we have the solution to our problem:
(8.12)
e ( x , t ) = 1 40 ∫ 0 40 f ( x ) d x +
1 20 ∑ n = 1 ∞ [ ∫ 0 40 f ( s ) cos n π s 40 d s ] cos n π x 40 cos n π t 40 L C .
We recognize immediately that the solution e ( x , t ) is an even function in both x and t. The
integral used for evaluating a 0 is the average value for f(x) over the half Fourier interval whose
length is 40. Half the value of this integral, that is, a 0/2, is equal to the DC component of the
series answer.
The frequency f n of the components of the voltage e ( x , t ) with respect to time is
f n = n π 2 π ( 40 L C ) = n 80 L C
Therefore, the fundamental frequency is 1 / 80 L C .
The previous two examples cover two of the three major classes of second‐order partial
differential equations: the parabolic type and the hyperbolic type. We now investigate the solution of
the elliptic type, which introduces us to some new techniques used in solving boundary value
problems.
Example 8.1.3 Recall that the temperature in a plate insulated above and below must satisfy the
differential equation
∂u∂t=k[∂2u∂x2+∂2u∂y2]
If the temperature is independent of t (i.e., steady tate), this differential equation becomes
∂2u∂x2+∂2u∂y2=0
which is Laplace’s equation in two variables.
Now suppose we have a rectangular plate whose dimensions are a by b and whose boundary
conditions are given by
u(x,0)=f1(x)0<x<a
u(a,y)=00<y<b
u(x,b)=f2(x)0<x<a
u(0,y)=00<y<b.
Find the temperature u as a function of x and y (see Figure 8.1).

Figure 8.1 Steady state heat conduction in a plate

Solution 8.1.3 In order to carry out specific operations, let us examine the special case where f
1(x) = 0 and f 2(x) = x. Steps 1,2, and 3. Assume ( x , y ) = X ( x ) Y ( y ) . Computing the necessary
derivatives and substituting into the diffe rential equation, we have
X′ ′ Y+XY′ ′ =0
or
X′ ′ X=-Y′ ′ Y=-λ
which yields the two ordinary differential equations
X′ ′ (x)+λX(x)=0
Y′ ′ (y)-λY(y)=0.
Steps 4 and 5. From our assumption and the given boundary conditions, we arrive at the following
boundary conditions associated with the ordinary differential equations:
.X(0)=0, X(a)=0
and
Y(0)=0.
Case 1. λ > 0 . Let λ = α 2 > 0 . Then we can show as before that the eigenvalues are
λn=αn2=n2π2a2n=1,
and the eigenfunctions are
X n ( x ) = sin n π x a .
Cases 2 and 3. λ = 0 and λ < 0 . It is easily shown that there are no non‐positive eigenvalues.
Step 6. One new idea introduced in this problem occurs when we wish to solve the second
differential equation,
Y′ ′ -n2π2a2Y=0.
Normally a beginning student would solve this equation using exponentials as follows
(8.13)
Yn(y)=Cnexpnπya+Dnexp(-nπya)
where e x p y = e y . Because neither exponential in Equation (12.37) vanishes, it is necessary to
solve a system of two equations to find D in terms of C. Therefore, it is much more convenient to
use hyperbolic functions, and we can write
Y n ( y ) = C n cosh n π y a + D n sinh n π y a .
Using the boundary condition Y(0) = 0, we see that
Yn(0)=0=Cn·1+Dn·0
or
Cn=0
and that
Y n ( y ) = sinh n π y a
is a solution to our ordinary differential equation in y and its one boundary condition. For
convenience we have again set the constant D = 1.
Step 7. Combining the two families of solutions X(x) and Y n (y), we construct solutions to the
partial differential equation and the three homogeneous boundary conditions that take the form
u n ( x , y ) = X n ( x ) Y n ( y ) = sin n π x a sinh n π y a .
Step 8. We still have to meet the final inhomogeneous boundary condition u ( x , b ) = x . In order
to do this we use the superposition principle and form the linear combination
(8.14)
u ( x , y ) = ∑ n = 1 ∞ b n sin n π x a sinh n π y a
expecting that we can find values for b n such that the inhomogeneous condition is met. Letting
y = b in Equation (9.69), we see that
 (8.15)
u ( x , b ) = x = ∑ n = 1 ∞ b n sinh n π b a sin n π x a
If we let
B n = b n sinh n π b a
Equation (9. 70) becomes
x = ∑ n = 1 ∞ B n sin n π x a
which is a Fourier sine series. Therefore,
(8.16)
B n = b n sinh n π b a = 2 a ∫ 0 a x sin n π x a d x
= 2 a a 2 n 2 π 2 [ sin n π x a - n π x a cos n π x a ] | 0 a
= 2 a n 2 π 2 [ - n π cos n π ] = ( - 1 ) n + 1 2 a n π .
The constants b n become
b n = ( - 1 ) n + 1 2 a n π sinh n π b a
and the final answer can be written as
u ( x , y ) = 2 a π ∑ n = 1 ∞ ( - 1 ) n + 1 sin ( n π x / a ) sin h ( n π y / a ) n sin h ( n π b / a ) .
Remark.
When solving the ordinary differential equation
y′′(x)-α2y(x)=0
it is useful to remember that there are three practical ways in which to express the solution. If one of
the boundary conditions is of the form
y(0) = 0 or y ’(0) = 0
then the general solution
(8.17)
y ( x ) = A cosh α x + B sinh α x
is most convenient to use because the value of A or B is quickly determined. On the other hand, if
y(L) = 0 or y ’(L) = 0
is given as a boundary condition, one of the following forms of the solution is useful:
y ( x ) = E sinh α ( x - d ) f o r y ( L ) = 0
y ( x ) = E cosh α ( x - d ) f o r y ( L ) = 0
where E and d are arbitrary constants. These forms of the solutions are not as well known as the form
in (Equation 9.71), but they can easily be shown to be equivalent for the particular boundary
condition given using the well‐known hyperbolic identities
sinh ( α - β ) = sinh α cosh β - cosh α sinh β
and
cosh ( α - β ) = cosh α cosh β - sinh α sinh β .
Finally, the classic solution
y(x)=Aeαx+Be-αx
finds its greatest use in solving boundary value problems over semi‐infinite intervals and the study of
Fourier integrals.
 Exercises
1. The temperature u ( x , t ) in a laterally insulated rod of length L satisfies the following boundary
value problem:
(8.18)
∂2u∂x2=1k∂u∂t0<x<L,0<t
u(0,t)=00<t
u(L,t)=00<t
u ( x , 0 ) = 100 0 < x < L .
Use the technique of separation of variables to find u ( x , t ) .
2. Use the technique of separation of variables to solve for the temperature u ( x , t ) if
(8.19)
∂ 2 u ∂ u 2 = 1 k ∂ u ∂ t 0 < x < 10 , 0 < t
∂u∂x(0,t)=00<t
∂ u ∂ x ( 10 , t ) = 0 0 < t
u ( x , 0 ) = 1 - x 0 < x < 10 .
3. The voltage e ( x , t ) along a submarine cable 3000 kilometers long satisfies the boundary value
problem
(8.20)
e x x = R C e t 0 < x < 3000 , 0 < t
e(0,t)=00<t
e ( 3000 , t ) = 0 0 < t
e ( x , 0 ) = sin x 100 0 < x < 3000 .
Use the technique of separation of variables to find e ( x , t ) .
4. The current i ( x , t ) along a submarine cable of length L satisfies
(8.21)
ixx=RCit0<x<L,0<t
∂i∂x(0,t)=00<t
∂i∂x(L,t)=00<t
i(x,0)=20<x<L.
Use the technique of separation of variables to find the current i ( x , t ) in the cable.
5. Find the temperature u ( x , t ) by the technique of separation of variables in a laterally insulated
rod of length 20 meters that has a heat source given by δ u ( x , t ) j o u l e s / m 3 . The other
conditions are
(8.22)
u(0,t)=0
(8.23)
u ( 20 , t ) = 0
(8.24)
u ( x , 0 ) = x ( 20 - x )
6. The voltage e ( x , t ) satisfies the differential equation e xx = RCe t + Ae x . Using separation of
variables find e ( x , t ) if
(8.25)
∂e∂x(0,t)=00<t
∂e∂x(L,t)=00<t
e ( x , 0 ) = 50 0 < x < L .
7. (a) A vibrating string is fastened to air bearings situated on vertical rods at x = 0 and x = 2. Find
the displacement y ( x , t ) if the conditions are
(8.26)
yx(0,t)=00<t
yx(2,t)=00<t
y(x,0)=x0<x<2
∂y∂t(x,0)=00<x<2
(b) Sketch y ( x , t ) over 0 < x < 2 for t = 0, t = 1, and t = 2.
Hint. Use only the first two terms of the series solution.
8. The pressure p ( x , t ) in an organ pipe satisfies the differential equation
pxx=1c2ptt.
If the pipe is L meters long and open at both ends, find the pressure p ( x , t ) if p ( x , t ) = 0 and ( ∂ p
/ ∂ t ) ( x , 0 ) = 40 .
9. (a) Given the telegraph equation for finding voltage e ( x , t ) on a transmission line of length a
along with the boundary and initial conditions, we can write
(8.27)
exx=LCett+(RC+GL)et+RGei
0<x<a,0<t
∂e∂x(0,t)=0,0<t
∂e∂x(a,t)=0,
e ( x , 0 ) = 100 , 0 < x < a
∂e∂t(x,0)=00<x<a.
Solve for e ( x , t ) using separation of variables.
(b) What is the frequency of the third harmonic?
10. The length of a guitar string is 65 centimeters. If the string is plucked 15 centimeters from the
bridge (i.e., at the end of the wire) by raising it 3 millimeters, find the displacement u ( x , t ) using
separation of variables.
11. The length of a piano string is 1 meter. When a pupil strikes a key, the following velocity is
imparted to the string:
∂ u ∂ t = velocity
(8.28)
= 0 0 < x < 49 c m 1 49 < x < 51 c m 0 51 < x < 100 c m .
Find the displacement u ( x , t ) using separation of variables. Do not evaluate the Fourier coefficients
of the final solution.
8.1.2 Non Cartesian Coordinate Systems
Examples 1, 3, and 4 were chosen especially so that you would see the method of separation of
variables applied to the three classes of second‐order differential equations: parabolic, hyperbolic,
and elliptic. All were set in terms of rectangular coordinates. In Example 5 we will examine a
boundary value problem that is more easily solved in polar coordinates.
Example 8.1.4 : Laplace Equation on a Disk
Let us consider a problem similar to Example 4 except that the shape of the plate is circular
rather than rectangular. We wish to find the steady‐state temperature u ( r , θ ) throughout a
circular plate of radius c that is insulated laterally. The temperature on the circumfe rence is 1000
C over one semicircle and 0 o over the other (see Figure 8.2).

Figure 8.2 Laplace equation on a disk of radius c

Expressed formally, u ( r , θ ) must satisfy

r2∂2u∂r2+r∂u∂r+∂2u∂θ2=0,-π<θ<π
(8.29)
u ( c , θ ) = 100 0 < θ < π 0 - π < θ < 0 .
Solution 8.1.4 To begin with we look for “elementary solutions” to this prob‐
lem in the form
u(r,θ)=R(r)Θ(θ)
Substituting in the diffe rential equation and using the method of separation of variables, we are
led to the following two ordinary diffe rential equations
(8.30)
Θ′′+λΘ=0
(8.31)
r2R′ ′ +rR′ -λR=0
When we attempt to solve Equation (9.72) (which is similar to the ones derived previously in
Examples 1, 3, and 4), we notice we have no boundary conditions since there are no exposed radial
edges. There are, however, implicit boundary conditions” due the periodic nature of the θ
coordinate. We require therefore that
(8.32)
u(r,-π)=u(r,π)
∂u∂θ(r,-π)=∂u∂θ(r,π)0<r<c.
Such boundary conditions are called periodic boundary conditions.
As before, we must consider the three cases λ > 0 , λ = 0 , and λ < 0 . If we let λ = α 2 > 0 , the
diffe rential equation in Equation (9.72) becomes
Θ′′+α2Θ=0
where the solution is
(8.33)
Θ ( θ ) = A cos α θ + B sin α θ
and
Θ ′ ( θ ) = - α A sin α θ + α B cos α θ .
From the periodic boundary conditions in Equation (9.74) we see that
(8.34)
Θ(-π)=Θ(π),Θ′ (-π)=Θ′ (π)
Substituting the solutions in the Equations ( 8.33 ) into the Equations ( 8.34 ), we have
(8.35)
A c o s α ( - π ) + B sin α ( - π ) = A cos α π + B sin α π
- α A sin α ( - π ) + α B cos α ( - π ) = - α A sin π + α B cos α π
which yields
B sin α π = 0 and A
(8.36)
sin α π = 0
Since A and B cannot both be zero, it follows that
sin α π = 0
or
α=nn=1,2,...
or
λn=n2
and
Θ n = A n cos n θ + B n sin n θ .
If λ = 0 , then Θ ’’ = 0 has the solution
Θ=Aθ+B
Substituting in the boundary conditions in Equation ( 8.34 ), we find
A(-π)+B=Aπ+B
or A = - A, which implies A = 0. TherefO re, λ = 0 is an eigenvalue, and the corresponding
eigenfunction is an arbitrary constant that we will choose as 1.
It can be shown that there are no negative eigenvalues.
We now move to the solution of the other ordinary diffe rential equation
(8.37)
r2R′ ′ +rR′ -n2R=0
We recognize that this is a Cauchy‐Euler differential. To obtain a solution to this diffe rential
equation, we attempt a solution of the form R(r) = r α , from which it follows that R ’(r) = αr α-1 and
R ’’(r) = α(α - 1)r α-2. Substituting these values in Equation ( 8.37 ), we find that
rα[α(α-1)+α-n2]=rα[α2-n2]=0.
This equation must hold for all r, 0 < r < c; therefore,
α 2 - n 2 = 0 or α = ± n.
Since r n and r -n are linearly independent solutions to Equation ( 8.37 ), the general solution of
the Cauchy‐Euler equation is
R(r) = Cr n + Dr -n n = 1, 2, 3, . . . .
Once again there is no explicit boundary condition, but we observe that since r = 0 is the center
of our circular plate, the
lim r → 0 + r - n = + ∞ .
Since we are only considering bounded solutions, we must set D = 0 and
R n (r) = r n , n = 1, 2, . . . .
For λ = 0 , the Cauchy‐Euler equation becomes
(8.38)
r2R′ ′ +rR′ =0
whose solution is
R(r)=C+Dlnr.
Since we are considering only bounded solutions, D must equal zero since
lim r → 0 l n r = - ∞
and therefO re we take R 0 = 1 as the solution of Equation ( 8.38 ).
The solution to the partial diffe rential equation is
u ( r , θ ) = A 0 2 + ∑ n = 1 ∞ r n ( A n cos n θ + B n sin n θ )
Using the inhomogeneous condition we can write
u ( c , θ ) = A 0 2 + ∑ n = 1 ∞ c n ( A n cos n θ + B n sin n θ )
where
(8.39)
A 0 = 1 π ∫ 0 π 100 d θ = 100
c n A n = 1 π ∫ 0 π 100 cos n θ d θ = 100 n π sin n θ | 0 π = 0
and
(8.40)
c n B n = 1 π ∫ 0 π 100 sin n θ d θ = - 100 n π cos n θ | 0 π
= 100 n π [ 1 - ( - 1 ) n ]
= 200 / n π n o d d 0 n e v e n
The solution to our problem is
u ( r , θ ) = 50 + 200 π ∑ n = 1 , 3 , 5 ∞ ( r c ) n sin n θ n .
If we set θ = 0 or π, we see immediately that the temperature u ( r , θ ) along the diameter θ = 0 or
π is 50 degrees. At points of symmetry with respect to this diameter, the temperature equals 50 plus
or minus the value of the series. Notice that only odd harmonics of the fundamental frequency sin θ
exist.
Example 8.1.5 : Flow Around A Cylinder
The steady‐state flow of a fluid in a direction transverse to a long metal cylinder can be
approximated at low velocities by Laplace’s equation in two dimensions. We will see that the
boundary conditions are of the Neumann
type. If ϕ is the velocity potential, then
∇ 2φ=r2∂2φ∂r2+r∂φ∂r+∂2φ∂θ2=0
where the velocity y = grad ϕ = ∇ϕ. For our purposes it is natural to locate the cylinder at the
origin. Its equation is then x 2 + y 2 = c 2. Far from the cylinder ( i . e . , a s r 2 = x 2 + y 2 → + ∞ ) ,
the velocity y of the fluid is constant and parallel to the x‐axis and can be written as v ( x , y ) = a
i as r 2 → + ∞ (see Figure 8.3).

Figure 8.3 Flow around a cylinder

Solution 8.1.5 To solve this problem we introduce polar coordinates and note that in these
coordinates the velocity is given by
(8.41)
v=vφ=φrer+1rφθeθ
where e r is the unit vector outward from the origin and e θ , a unit vector, is perpendicular to e r in
a counterclockwise direction.
Looking at Figure 8.4 we observe that the boundary conditions at r = + ∞
Figure 8.4 Radial coordinate system and the components of φ

are
(8.42)
φ r = a cos θ , 1 r φ θ = - a sin θ
Since the fluid cannot penetrate the cylinder, we must have φ r ( c , θ ) = 0 . Finally, since the
velocity at θ and θ + 2π is the same, we must have from Equation (8.41)
(8.43)
φr(r,θ+2π)=φr(r,θ)
φθ(r,θ+2π)=φθ(r,θ)
The method used in this problem is similar to that used in Example 5. We assume φ ( r , θ ) = R ( r )
Θ ( θ ) , which (aft er substituting into Laplace’s equation in polar form) yields the two ordinary
diffe rential equations and boundary conditions
(8.44)
Θ′ ′ +λΘ=0,Θ(θ+2π)=Θ(θ),Θ′ (θ+2π)=Θ′ (θ)
(8.45)
r2R′ ′ +R′ -λR=0,R′ (c)=0
If λ = α 2 > 0 , then the general solution of Θ ’’ + α 2 Θ = 0 is
Θ ( θ ) = A cos α θ + β sin α θ
and
Θ ′ ( θ ) = - α A sin α θ + α B cos α θ .
Substituting these two equations into the periodic conditions in Equation ( 8.42 ), we find (aft er
some effort) that
α = n, n = 1, 2, 3, . . .
or
λn=n2
and
Θ n ( θ ) = A n cos n θ + B n sin n θ .
With these eigenvalues the diffe rential equation in Equation ( 8.45 ) becomes
r 2 R ’’ + rR ’ - n 2 R = 0n = 1, 2, . . .
whose solution is
Rn(r)=Cnrn+Dnr-n
Using the boundary conditions in Equation ( 8.45 ), we have
R′ (c)=0=nCncn-1-nDnc-n-1
from which it follows that
Dn=c2nCn.
For λ = 0 , Θ ’’ = 0, and its solution is
Θ0=A0+B0θ.
The solution to the diffe rential equation in Equation ( 8.45 ) is
Now since ln r → + ∞ as r → + ∞, we must set C 0 = 0 and we can let
R0=D0=1.
Since there are no negative eigenvalues, we expect the solution to take the form
(8.46)
φ ( r , θ ) = A 0 + B 0 θ + ∑ n = 1 ∞ [ r n + c 2 n r - n ] [ A n cos n θ + B n sin n θ ]
We still have two conditions [Equation ( 8.42 ] that must be met. Differentiating Equation ( 8.46 )
with respect to r and θ, we have
(8.47)
φ r = ∑ n = 1 ∞ [ n r n - 1 - n c 2 n r - n - 1 ] [ A n cos n θ + B n sin n θ ]
and
(8.48)
φ θ = B 0 + ∑ n = 1 ∞ [ r n + c 2 n r - n ] n [ - A n sin n θ + B n cos n θ ]
Now the only term in Equation ( 8.47 ) that is bounded occurs when n = 1; therefO re, A n , B n = 0
for n = 2, 3, . . . and
lim r → ∞ φ r ( r , θ ) = a cos θ = A 1 cos θ + B 1 sin θ
or
A1=a
B1=0.
Under these conditions Equation ( 8.48 ) becomes
(8.49)
1 r φ θ = B 0 r + 1 r [ r + c 2 r - 1 ] [ - A 1 sin θ ]
= B 0 r + [ 1 + c 2 r 2 ] [ - a sin θ ] .
As r → + ∞,
1 r φ θ = - a sin θ
and the second condition in Equation ( 8.42 ) is satisfied.
Combining all these conditions in Equation ( 8.46 ), we write the solution as
φ ( r , θ ) = A 0 + B 0 θ + a [ r + c 2 r ] cos θ
where A 0 and B 0 are arbitrary constants.
 As we investigate this solution, we must recall that φ ( r , θ ) is the velocity potential, not the
velocity. The velocity field is given by
y = φ r e r + 1 r φ θ e θ = a [ 1 - c 2 r 2 ] cos θ e r - a [ 1 + c 2 r 2 ] sin θ e θ .
When we measure the velocity far from the axis of the cylinder,
v ≈ a cos θ e r —asin θe θ = ai.
In other words, the effect on the velocity due to the cylinder becomes less noticeable as we move
away from the axis. On the other hand, when we are near the cylindrical obstruction, that is r ≈ c,
v ≈ - 2 a sin θ e θ
which shows that the flow follows the shape of the cylinder when r ≈ c and approaches zero as we
get nearer to the x‐axis.
Example 8.1.6 :Vibrations in a Sphere
In all our examples up to this point it has been relatively easy to evaluate the eigenvalues and
eigenfunctions. We will now examine a problem in which this is not the case. We wish to study the
vibrations or pressure of air within a sphere of radius c which might be caused by an exploding
firecracker, for example.
The pressure equation in space is given by
(8.50)
( ρ 2 p ρ ) ρ + 1 sin 2 θ p φ φ + 1 sin θ ( p θ sin θ ) θ = ρ 2 a 2 p t t
Solution 8.1.6 If we assume the pressure is independent of θ and ϕ(i.e., in the radial direction
only), Equation ( 8.50 ) becomes
(8.51)
1ρ2(ρ2pρ)ρ=1a2ptt
Since the pressure under consideration is within the sphere, the directional derivative of p in the
direction ρ must be zero on the boundary, or
pρ(c,t)=0t>0.
Choosing a trial solution of the form p ( ρ , t ) = R ( ρ ) T ( t ) , we arrive at the two ordinary diffe
rential equations using the method of separation of variables on Equation ( 8.51 ):
(8.52)
ρ2R′ ′ +2ρR′ +λρ2R=0R′ (c)=0
(8.53)
T′′+a2λT=0
The general solution of Equation ( 8.52 ) for λ = α 2 > 0 is
R ( ρ ) = A cos α ρ ρ + B sin α ρ ρ .
This solution can be found by replacing R(ρ) by S(ρ)/ρ, which transfO rms the diffe rential
equation into one with constant coefficients (see Exercise 16). Since p ( ρ , t ) must be bounded in
the sphere, it is necessary that A = 0. trigonometric equation is difficult to solve in order to find
the eigenvalues. We can easily
see that there are an infinite number of eigenvalues by solving the equation graphically, as
shown in Figure 8.5.
Figure 8.5 Eigenvalues μ n are given by the intersections of y = μ with tanμ

The first nonzero root is 4.49, the second 7.73, and so on. Notice in particular that as
μ = αc → ∞, the intersection points approach (2n + 1)π/2, n = 1, 2, . . .. Ta ble 1 lists the first eight
nonzero solutions done numerically. Using these values for μ we see that the eigenvalues λ satisfy
λn=αn2=μn2c2n=1,2,...
and the eigenfunctions are
R n ( ρ ) = sin ( μ n ρ / c ) ρ .
To see if λ = 0 is an eigenvalue, we solve the diffe rential equation
R′ ′ +2ρR′ =0
whose general solution is
R(ρ)=Aρ+B.
Once again since R(ρ) must be bounded in the sphere, A = 0 and R(ρ) = B. Since R ’(ρ) = 0 for any
ρ, it certainly follows that R ’(c) = 0. Therefore, λ = 0 is an eigenvalue and its corresponding
eigenfunctions R 0(ρ) = 1. This ends our search for eigenvalues, for you can show there are no
negative eigenvalues.
We now proceed to the solution of Equation ( 8.53 ). For λ = μ n 2 / c 2 , n = 1, 2, . . ., its
solution is
T n ( t ) = C n cos a μ n t c + D n sin a μ n t c
and for λ = 0 ,
T0(t)=C0+D0t.
Combining the solutions R and T, we see that the answer to our boundary value problem takes the
form
(8.54)
p ( ρ , t ) = C 0 + D 0 t + ∑ n = 1 ∞ sin ( μ n ρ / c ) ρ
× ( C n cos a μ n t c + D n sin a μ n t c )
In order to solve for constants C n and D n , n = 0, 1, 2, . . ., we need two initial conditions:
p ( ρ , 0 ) = f ( ρ ) and ∂ p ( ρ , 0 ) ∂ t = g ( ρ ) 0 < ρ < c .
 When we attempt to solve for the coefficients, the series is not the Fourier series. But the family
of functions does possess an orthogonal property, and we can solve for the coefficients as follows:
(8.55)
C0=3c3∫0cρ2f(ρ)dρ,D0=3c3∫0cρ2g(ρ)dρ
C n = 2 ( 1 + μ n 2 ) c μ n 2 ∫ 0 c ρ f ( ρ ) sin μ n ρ c d ρ
D n = 2 ( 1 + μ n 2 ) a μ n 3 ∫ 0 c ρ g ( ρ ) sin μ n ρ c d ρ .
Exercises
1. You wish to find the temperature u ( ρ , θ ) in a laterally insulated pieshaped region of radius c.
The temperature satisfies the differential equation
ρ2uρρ+ρuρ+uθθ=0
and the boundary conditions
u ( ρ , 0 ) = 00 < ρ < c
u ( ρ , π 6 ) = 00 < ρ < c
u(c,θ)=θ0<θ<π6.
Use the technique of separation of variables to find the temperature
u(ρ,θ)
2. Find the steady‐state temperature u ( ρ , θ ) in a circular plate insulated laterally of radius 10 if the
temperature on the circumference is 3 - θ.
3. Find the electrostatic potential Φ ( ρ , θ ) on the half disk plate with radius 25cm with the
boundary conditions on the potential Φ as shown in Figure 8.6.

Figure 8.6 Laplace equation on the half disk plate

4. Show that (Equation 8.52) can be transformed into one with constant coefficients by using the
substitution R(ρ) = S(ρ)/ρ.
5. Evaluate and graph the velocity y of the fluid around the cylinder discussed in the previous
section for a = 1, r = 2c, θ = 0, π/4, π/2 radians.
6. Apply the method of separation of variables to Laplace’s equation in Example 5 to find the two
ordinary differential equations in (Equations 9.72) and (9.73).
7. Show that there are no negative eigenvalues to be found in the boundary value problem in
(Equation 8.44).
8. Prove that the solutions of (Equation 8.50) satisfy
∫ 0 c ρ sin μ n ρ c d ρ = 0
where tan μ n = μ n .
9. Show that
 (a) ∫ 0 c sin μ n ρ c sin μ m ρ c d ρ = 0 m ≠ n
(b) ∫ 0 c sin 2 μ n ρ c d ρ = c 2 ( μ m 2 1 + μ m 2 ) n = 1, 2, . . .
where tan μ n = μ n .
10. Using the series in (Equation 8.54) and the fact that p ( ρ , 0 ) = f ( ρ ) , justify C 0 in (Equation
8.55).
Hint. Multiply both sides of (Equation 8.53) by ρ 2 dρ and integrate from 0 to c.
11. Using the conditions in Exercise 22, justify C n in (Equation 8.55).
Hint. Multiply both sides of (Equation 8.54) by ρ sin ( μ m ρ / c ) d ρ and integrate from 0 to c.
12. The coefficients D n can be found easily from the equation for C n in (Equation 8.55). Using the
condition ∂ p ( ρ , 0 ) / ∂ t = g ( ρ ) , prove D n in (Equation 8.55).
13. Find p ( ρ , t ) if c = 1 and
pρ(1,t)=0t>0
p ( ρ , 0 ) = 00 < ρ < 1
pt(ρ,0)=ρ0<ρ<12012<ρ<1.
14. Write out the boundary value problem for finding the pressure p ( ρ , t ) in a spherical region if the
pressure is zero at 100 meters from the center and the initial pressure is zero, and the rate of change
of pressure with respect to t is h(p) .
15. Consider two concentric spheres of radius a and b, a < b, respectively. If the pressure on the
inner sphere is 1 and the rate of change of pressure with respect to ρ on the outer sphere is 0, and the
initial pressure is zero and the rate of change of pressure with respect to t is ρ, write the partial
differential equation and boundary and initial conditions satisfied.

8.1.3 Boundary Value Problems with General Initial Conditions

We will now examine the method of solution of boundary value problems where there are several
initial conditions and two or more are inhomogeneous. There are two well‐known methods for
solving such a problem.
The first approach consists of solving two boundary value problems, as we have done previously.
Suppose we wish to solve a vibrating string problem where both ends are fixed and initially the string
is not only displaced but a velocity is imparted to the string (see Figure 8.7).

Figure 8.7 Vibrating string, u is the displacement v = ∂ u ∂ t

 Formally, this problem is written as
(8.56)
∂2u∂t2=c2∂2u∂x2,0<x<L,0<t
u(0,t)=0,0<t,u(L,t)=0,0<t,
u(x,0)=f(x),0<x<L,∂u∂t(x,0)=g(x),0<x<L
We now assume that the solution to this problem can be written in the form u ( x , t ) = v ( x , t ) + w (
x , t ) where v and w satisfy the following boundary value problems, respectively:
∂2v∂t2=c2∂2v∂x2∂2w∂t2=c2∂2w∂x2
v(0,t)=0w(0,t)=00<t
v ( L , t ) = 0 w ( L , t ) = 00 < t
v(x,0)=f(x)w(x,0)=00<x<L
∂v∂t(x,0)=0∂w∂t(x,0)=g(x)0<x<L.
It follows from the method of superposition that if v and w satisfy their boundary value problem, u
satisfies the problem in (Equation 8.56).
The second way of solving might be called a straightforward attack. We solve (Equation 8.56)
directly as we have done in earlier examples (see Example 3). The eigenvalue problem will be the
same as before, but when we come to solve the other ordinary differential equation, we will have no
initial conditions. When we multiply X n by T n and then form the infinite linear combination to find u (
x , t ) , we will find it necessary to solve for two sets of arbitrary constants. Using one initial
condition at a time, we are led to solve Fourier series problems from which we evaluate our two sets
of constants.
Example 8.1.7 : Consider a vibrating string fastened to air bearings that move along two
parallel rods 4 meters apart. Find the displacement u ( x , t ) if the initial displacement is 1 meter
and the initial velocity is x meter, per second.
Solution 8.1.7 The formal statement of this problem is given as
∂2u∂x2=c2∂2u∂t20<x<4,0<t
∂u∂x(0,t)=00<t
∂u∂x(4,t)=00<t
u(x,0)=10<x<4
∂u∂t(x,0)=x0<x<4.
Using the method of separation of variables with u ( x , t ) = X ( x ) T ( t ) , we see easily that
(8.57)
X′ ′ +λX=0,X′ (0)=0,X′ (4)=0
(8.58)
T′′+λc2T=0
Solving the eigenvalue problem for Equation ( 8.57 ), we find the eigenvalues are
λ n = n 2 π 2 16
and the eigenfunctions are
X n = cos n π x 4 n = 0 , 1 , 2 , . . . .
Our next step is to solve the diffe rential equation in Equation ( 8.58 ). Since λ is known, the
solution of this equation is
T0(t)=A0+B0t
T n ( t ) = A n cos n π t 4 c + B n sin n π t 4 c n = 1 , 2 , . . . .
Unlike previous examples, no initial conditions are attached to the diffe rential equation in
Equation (8.58) and therefO re we cannot evaluate the A n s or B n s at this stage. We continue by
assuming a solution of the form
(8.59)
u ( x , t ) = 1 2 ( A 0 + B 0 t ) + ∑ n = 1 ∞ [ A n cos n π t 4 c + B n sin n π t 4 c ] cos n π x 4
from which it follows that
∂ u ∂ t ( x , t ) = B 0 2 + ∑ n = 1 ∞ ( n π 4 c ) ( - A n sin n π t 4 c + B n cos n π t 4 c ) cos n π x 4 .
Using the initial condition u ( x , 0 ) = 1 , we have
1 = A 0 2 + ∑ n = 1 ∞ A n cos n π x 4
which is a Fourier cosine series whose coefficients are A 0 = 2 and A n = 0. In the same way,
knowing that ( ∂ u / ∂ t ) ( x , 0 ) = x , we can write
x = B 0 2 + ∑ n = 1 ∞ n π 4 c B n cos n π x 4
on the interval (0, 4) . This is a Fourier cosine series representing an even function by extending x
into ( - 4, 0) as - x. Our calculations yield
B0=4
and
B n = - 64 c n 3 π 3 n o d d 0 n e v e n .
Substituting these coefficients into Equation ( 8.59 ), the solution to our problem is
(8.60)
u ( x , t ) = 1 + 2 t - 64 c π 3 ∑ 1 , 3 , 5 ∞ 1 n 3 sin n π t 4 c cos n π x 4
We see from the solution to Equation ( 8.60 ) that the center of the string, x = 2, has the
displacement u ( 2 , t ) = 1 + 2 t . This fact tells us that once the string is put in motion, the center
moves away from u(2, 0) = 1 at a constant velocity of 2 meters per second. When t = 4cq, where
q = 0, 1, 2, . . ., the string becomes parallel to the x‐axis.
The factor sin ( n π t / 4 c ) in Equation ( 8.60 ) allows us to determine the frequency of a
vibrating string. The period λ n of the individual terms in the series (8.60) is given by
λ n = period of oscillations = 2 π n π 4 a = 8 c n
from which we can find the frequency f n which is related to the period by
fn=1λn=n8c.
Now the smallest n allowed in the series (8.60), in this case n = 1, determines the fundamental
frequency, that is
f1=18c.
Notice that all higher frequencies are integer multiples of the fundamental frequency and are
called harmonics. The value n = 2 is the second harmonic, n = 3 the third harmonic,and so on.
8.1.4 Boundary Value Problems with Inhomogeneous Equations
In some boundary value problems the differential equation and the boundary conditions may be
inhomogeneous. In general, there is no straightforward way to solve these problems. However, if the
inhomogeneous part of the differential equation is a function of x and the boundary conditions are
constants, there is a step‐by‐step way to solve such a problem. The method is best shown by
example.
Example 8.1.8 Suppose we wish to find the temperature u ( x , t ) in a laterally insulated rod of
length π whose initial temperature is f(x) . For t > 0, the left end of the rod is fixed at 500 o and the
right end is fixed at a temperature of 100o. Furthermore, for t > 0, an electric current is made to
pass through the rod heating it, which introduces the sin x term in the partial diffe rential
equation.
Solution 8.1.8 Formally stated, this boundary value problem looks like
∂ u ∂ t = k ∂ 2 u ∂ x 2 + sin x
u ( 0 , t ) = 500 0 < t
u ( π , t ) = 100 0 < t
u(x,0)=f(x)0<x<π.
We assume the solution can be broken into two parts; that is,
(8.61)
u(x,t)=v(x,t)+h(x)
Our plan of attack is to choose constants of iteration so that eventually the v ( x , t ) term will be a
solution of a homogeneous boundary value problem.
Substituting the right‐hand side of Equation (8. 61) into the diffe rential equation, we have
∂ v ∂ t = k ∂ 2 v ∂ x 2 + k h ′ ′ ( x ) + sin x .
Therefore, to follow our plan, k h ′ ′ ( x ) = - sin x . Solving this simple differential equation, we find
(8.62)
h ( x ) = C 1 x + C 2 + sin x k
Next we look at the boundary conditions, which can be written as
v ( 0 , t ) + h ( 0 ) = 500
v ( π , t ) + h ( π ) = 100 .
In order that v ( 0 , t ) and v ( π , t ) equal zero,
h(0) = 500 and h(π) = 100.
We use these conditions to find the specific values of C 1 and C 2 in Equation ( 8.62 ). Therefore,
h ( 0 ) = 500 = C 1 ( 0 ) + C 2 + sin 0 k
which implies that C 2 = 500. Then,
h ( π ) = 100 = C 1 π + 500 + sin π k
or
C 1 = - 400 π .
Substituting these constants in the solution [Equation ( 8.62 it follows that
h ( x ) = - 400 π x + 500 + sin x k .
We have completely determined h(x) . Our next task is to find v ( x , t ) , which now satisfies the
boundary value problem
∂v∂t=k∂2v∂x2
v ( 0 , t ) = 00 < t
v ( π , t ) = 00 < t
v ( x , 0 ) = u ( x , 0 ) - h ( x ) = f ( x ) - sin x k + 400 π x - 5000 < x < π .
If f ( x ) = ( sin x / k ) + 500 , then v ( x , 0 ) = ( 400 / π ) x . Using the method of separation of
variables and setting v ( x , t ) = X ( x ) T ( t ) , we are led to the two diffe rential equations and
boundary conditions
X′ ′ (x)+λX(x)=0X(0)=X(π)=0
T′′(t)+λkT(t)=0.
The eigenvalues are λ = n 2 , n = 1, 2, . . ., and the eigenfunctions are X n (x) = sin nx. Solving the
other equation, find T n ( t ) = e - k n 2 t .
Combining this information, we expect our solution to be of the form
v(x,t)=∑n=1∞bne-kn2tsinnx.
And since
v ( x , 0 ) = 400 π x = ∑ n = 1 ∞ b n sin n x
we see that
(8.63)
b n = 2 π ∫ 0 π 400 x π sin n x
= ( - 1 ) n + 1 800 n π .
The solution to the homogeneous boundary value problem is
v ( x , t ) = 800 π ∑ n = 1 ∞ ( - 1 ) n n e - k n 2 t sin n x .
The solution to the inhomogeneous problem using the principle of superposition is
u ( x , t ) = 800 π ∑ n = 1 ∞ ( - 1 ) n n e - k n 2 t s i n n x - 400 π x + 500 + sin x k .
This problem is one that commonly occurs in heat flow problems. We observe that the answer for u (
x , t ) consists of two parts: the infinite series, which depends on x and t, and the remaining part h(x),
which depends only on x. Because of the exponential term in v ( x , t ) we see that as t → + ∞ the
value of the series approaches zero. This part of the solution is called the transient solution because it
passes away quickly. The other part of the solution h(x) does not vary with time and is called the
steady‐state solution.
Once again we notice that because the factor n 2 appears in the exponential term in the series, the
terms decrease quite rapidly in size for even small values of t. Therefore, we can often get a good
approximation of the transient term by taking only one or two terms.
Exercises
1. Given the wave equation y tt = c 2 y xx and the conditions
y(0,t)=00<t
y(L,t)=00<t
y ( x , 0 ) = 10 0 < x < L
yt(x,0)=-50<x<L
use the technique of separation of variables to find y ( x , t ) .
2. Find the voltage e ( x , t ) on a high‐frequency line of length 20 centimeters if both ends of the
line are shorted. The initial voltage e ( x , 0 ) is 20(1 - x); e t ( x , 0 ) is 20x.
3. Find the longitudinal displacement of a rod of length L whose ends are free if the initial
displacement is x(L - x) and u t ( x , 0 ) is 2.
4. Given the wave equation and the conditions
utt=c2uxx+e-x
u ( 0 , t ) = 100 0 < t
u ( L , t ) = 50 0 < t
u(x,0)=00<x<L
ut(x,0)=x0<x<L
and letting u ( x , t ) = v ( x , t ) + Φ ( x ) , solve for Φ completely and write (but do not solve) the
boundary value problem for v ( x , t ) .
5. The voltage e ( x , t ) satisfies the differential equation e xx = RCe t + Ax. Using separation of
variables, find e if
∂e∂x(0,t)=00<t
∂e∂x(L,t)=00<t
e ( x , 0 ) = 50 0 < x < L .
6. The current in a submarine cable of length 1000 meters is given by
ixx=RCit.
If the conditions are
i(0,t)=2a0<t
i ( 1000 , t ) = 0.1 a 0 < t
i ( x , O ) = 0 a 0 < x < 1000
find the current i ( x , t ) .
7. Solve the heat equation with decomposition
uxx-kut+A=0
with conditions
u(O,t)=00<t
u(L,t)=00<t
u(x,O)=00<x<L
8. The equation for a vibrating string with external force is given by y tt = c 2 y xx + F/δ, where F = δx
and where the left‐hand end x = 0 is fixed and the slope of the right‐hand end x = L is zero. The
initial displacement is f(x) while initial velocity is zero. Let y ( x , t ) = z ( x , t ) + Φ ( x ) . Solve for
Φ(x) completely and write (but do not solve) the boundary value problem for z ( x , t ) .
9. Poisson’s equation is given by u xx + u yy = y. If the conditions are
u(0,y)=56+y360<y<1
u(1,y)=y36-160<y<1
u(x,0)=-160<x<1
u ( x , 1 ) = 00 < x < 1
solve for u ( x , y ) .
Hint. Let u ( x , y ) = w ( x , y ) + g ( y ) .
10. The equation of a vibrating string with an external force is given by u t t = c 2 u x x + sin x .
The boundary and initial conditions are u ( 0 , t ) = 0 , u x ( L , t ) = - 1 , u ( x , 0 ) = 0 , and u t ( x , 0 )
= f ( x ) . Let u ( x , t ) = w ( x , t ) + Φ(x) . Find Φ(x) so that w ( x , t ) satisfies a homogeneous
boundary value problem. Write out the differential equation and conditions satisfied by w ( x , t ) , but
do not solve the equation for w.

8.2 GREEN’S FUNCTIONS

When boundary value problems are solved using separation of variables, the solution is represented
by an infinite series. However, a series solution is neither the only possible approach to the solution
of boundary value problems nor always the best. In this appendix we introduce a method that
represents the solution in terms of an integral. This method is usually referred to as “ Green’s function
method.”
Example 8.2.1 Consider Poisson equation
(8.64)
∇ 2u(x)=f(x)
on a domain D of R 2 having a piecewise smooth boundary with the boundary conditions
(8.65)
u(x)|∂D=g(x)
where ∂ D is the boundary of D. We want to find a function G ( x , η ) , η ∊ R 2, which is called the
Green’s function for the problem and depends only on the diffe rential operator and the domain D
so that the solution of Equations ( 8.64 ) and ( 8.65 ) can be represented as
(8.66)
u(x)=∫DG(x,η)f(η)dη+∫∂Dg(η)∂G∂ηds
where
∂G∂η=gradG·n=[∂G∂η1,∂G∂η2]·n
and n is the outward normal to the boundary ∂ D.
To see some of the advantages for using such an integral representation of the solution, we note that
(Equation 8.66) holds for all possible f(x) and g(x), thus providing us with the solution to (Equations
8.64) and (8.65) even when the method of separation of variables is not applicable. Furthermore, the
dependence of the solution on this function is explicit and in closed form. Therefore, it enables us to
investigate the dependence of the solution on these elements of the problem directly.
From an applied point of view, Green’s functions have a natural physical interpretation as the
influence of a unit source located at η on the point x. Accordingly, we can refer to G ( x , η ) as the “
influence function.” We emphasize this interpretation of the Green’s functions throughout this
presentation and use it to infer some of their properties.
The major undertaking in this solution methodology is the computation of the Green’s function,
which appears as a kernel in the integral representation of the solution.
Green’s Function for the Laplace Operator
Definition: The solution of
(8.67)
∇ 2G(x,η)=δ(x-η)
on a domain D in R n satisfying the boundary condition
(8.68)
G(x,η)|∂D=0
is called the Green’s function for the Laplace operator in D. Here δ(x) denotes the Dirac delta
function.
Theorem 1 G is symmetric in x and η; that is
G(x,η)=G(η,x)
Proof. To prove this theorem we use Green’s lemma, which states that under proper restrictions on φ,
ψ, and D, the following formula holds:
(8.69)
∫D(ψ∇ 2ϕ-ϕ∇ 2ψ)dA=∫∂D(ψ∂ϕ∂n-ϕ∂ψ∂n)ds
Substituting ψ = G ( x , η ) and ϕ = G ( x , η ∗ ) , where η, η * are arbitrary but fixed points in D, in
(Equation 8.69) we obtain
∫D[G(x,η)∇ 2G(x,η∗ )-G(x,η∗ )∇ 2G(x,η)]dA=
(8.70)
∫∂D[G(x,η)∂G∂n(x,η∗ )-G(x,η∗ )∂G∂n(x,η)]ds
Using (Equations 8.67),(8.68) this yields
∫D[G(x,η)δ(x-η∗ )-G(x,η∗ )δ(x-η)]dA=0.
Hence
G(x,η)=G(η,x)
Since η, η * are arbitrary points in D this proves the theorem. Theorem 2.
∂ G ∂ n ( x , η ) has a discontinuity at x = η. Moreover,
(8.71)
lim ε → 0 ∫ C ε ∂ G ∂ n d s = 1
where C ɛ is the circle of radius ɛ around x = ( x 1 , x 2 ) ; that is
(x1-η1)2+(x2-η2)2=∈ 2
Proof. Let D ɛ be the domain bounded by the circle C ɛ . From (8.67) we infer that
∫Dε∇ 2GdA=∫Dεδ(x-η)=1.
Hence, using the divergence theorem we infer that
1=∫Dε∇ 2GdA=∫Cε∇ G.nds=∫Cε∂G∂nds.
This shows that ∂ G/ ∂ n has a discontinuity at x = η since otherwise (i.e. if ∂ G/ ∂ n was continuous)
the integral in (Equation 8.71) would have to be equal to 0.
Theorem 3. The solution of the Dirichlet problem in (Equations 8.64), (8.65) is given by (Equation
8.66) where G is the Green’s function for the Laplace operator.
Proof. We prove that if u(x) satisfies (Equations 8.64), (8.65) then it must satisfy (Equation 8.66).
To prove this statement we use Green’s lemma with
ψ(η)=G(η,x),ϕ(η)=u(η)
We obtain
∫D[G(x,η)∇ 2u(η)-u(η)∇ 2G(η,x)]dη=
(8.72)
∫∂D[G(η,x)∂u(η)∂n-u(η)∂G∂n(x,η)]ds
Using (Equations 8.64),(8.65),(8.67),(8.68) this reduces to
(8.73)
∫D[G(η,x)f(η)-u(η)δ(η-x)]dη=-∫∂Dg(η)∂G∂nds
The desired result follows now from the symmetry of G ( η , x ) (Theorem 1) and the definition of the
Dirac δ function.
At this juncture it is natural to inquire into the intuitive (or physical) meaning of G and (Equation
8.66). To answer this question we note that the solution of Poisson (Equation 8.64) can be interpreted
as the gravitational potential due to a source distribution f(x) . It follows then that Green’s function
which satisfies (Equation 8.67) represents the gravitational potential at x due to a unit source at η.
(Equation 8.66) can be interpreted simply as a restatement of the superposition principle: that is the
total gravitational potential due to a volume distribution x and surface distribution g is equal to the
sum (represented by the integral) of the pointwise contributions of these sources.
Computation of Green’s Function
Different techniques exist for the computation of Green’s function for a given differential operator
on a domain D. In this section we give two examples for the computation of this function for the
Laplace operator.
Example 8.2.2 To begin with we compute Green’s function for the Laplace operator when D = R
2. In this case the domain has no boundaries and Green’s function has to satisfy only Equation (

8.67 ). This function is referred to as the “infinite space Green’s Function”

Solution 8.2.1 To compute G in this case we have to solve
(8.74)
∇ 2G=δ(x-η)
Since the right‐hand side of Equation ( 8.74 ) depends only on x - η we attempt to find G in the
form
G(x,y,η1,η2)=G(r)
where r 2 = (x - η 1)2 + (y - η 2)2. However, since δ(x - η) is zero for x ≠ η it follows that G must
satisfy
∇2 G = 0, for x ≠ η.
Using the expression of the Laplace operator in polar coordinates this equation reduces to
(8.75)
1r∂∂r(r∂G∂r)=0,r≠0
whose solution is
G = C 1 + C 2 ln r
where C 1, C 2 are constants. To determine C 2 we use Equation (8. 71). We
 obtain
1 = lim ε → 0 ∫ C ε ∂ G ∂ n d s = lim ε → 0 ∫ 0 2 π C 2 r r d θ = 2 π C 2 .
Hence
G = C 1 + 1 2 π ln r .
Although C 1 remains arbitrary, we set it to zero for convenience.
Example 8.2.3 Compute the Green’s function for the Laplace operator on the unit disk in R 2.
Solution 8.2.2 The usual technique to compute Green’s function on a domain with boundary is to
rewrite it as a sum
G=G1+v
where G 1 is the infinite space Green’s function. For the Laplace operator we have
∇ 2G=∇ 2G1+∇ 2v=δ(x-η)
However G 1 satisfies Equation ( 8.74 ); hence, v must satisfy
∇ 2v=0.
Moreover, from (8.68) we infer that on the boundary
v|∂D=-G1|∂D.
It follows then that v is a regular solution of Laplace equation whose values on ∂ D are equal to
those of - G 1 on this boundary.
To solve the boundary value problem
(8.76)
∂2v∂x2+∂2v∂y2=0
v | x 2 + y 2 = - 1 2 π ln r | x 2 + y 2
where r 2 = (x - η 1)2 + (y - η 2)2 (note that η is considered as a fixed but arbitrary point on the
disk). We now introduce polar coordinates for x, y and η 1, η 2:
x = ρ cos θ , y = ρ sin θ
η 1 = σ cos ϕ , η 2 = σ sin ϕ .
Solving Equation ( 8.76 ) by the method of separation of variables we have
(8.77)
v = a 0 2 + ∑ n = 1 ∞ ρ n ( a n cos n θ + b n sin n θ )
To satisfy the boundary condition we must have (using the cosine rule)
(8.78)
v | ρ = 1 = - 1 4 π ln [ 1 + σ 2 - 2 σ cos ( θ - ϕ ) ] = 1 2 π ∑ n = 1 ∞ σ n cos ( θ - ϕ ) n
From Equations (8.77), ( 8.78 ) it follows that:
(8.79)
a 0 = 0 , a n = σ n cos n ϕ 2 π n , b n = σ n sin n ϕ 2 π n
Combining Equation (8.77) and Equation ( 8.79 ) we have
(8.80)
v ( ρ , θ , σ , ϕ ) = 1 2 π ∑ n = 1 ∞ ( ρ σ ) n n cos n ( θ - ϕ )
= - 1 4 π ln [ 1 + ( ρ σ ) 2 - 2 ρ σ cos ( θ - ϕ ) ] .
Hence
 (8.81)
G = 1 4 π ln [ ρ 2 + σ 2 - 2 ρ σ cos ( θ - ϕ ) ] - 1 4 π ln [ 1 + ( ρ σ ) 2 - 2 ρ σ cos ( θ - ϕ ) ] .
Corollary. The solution of Laplace equation
(8.82)
∇ 2u(ρ,θ)=0
on the unit disk subject to the boundary condition
(8.83)
u|ρ=1=g(θ)
is
(8.84)
u(ρ,θ)=∫02πP(ρ,θ-ϕ)g(ϕ)dϕ
where P is the Poisson kernel
P ( ρ , θ - ϕ ) = 1 - ρ 2 1 + ρ 2 - 2 ρ cos ( θ - ϕ ) .
Proof. From (Equation 8.66) we know that the solution of (Equations 8.82), (8.83) can be expressed
as
(8.85)
u(ρ,θ)=∫∂Dg(η)∂G∂nds
In this case, however,
∂ G ∂ n | ∂ D = ( ∂ G ∂ σ ) | σ = 1 = 1 - ρ 2 1 + ρ 2 - 2 ρ cos ( θ - ϕ )
which yields the desired result. Note that in (Equation 8.85) ρ, θ are fixed and η 1, η 2 are variables.
Therefore we must compute ∂ G ∂ n with respect to these variables.
Remark: The formula that has been used in (Equation 8.78) follows from the identity
(8.86)
1 + ∑ n = 1 ∞ r n c o s ( n x ) = 1 - r cos x 1 + r 2 - 2 cos x , 0 ≤ r ≤ 1
To prove this identity we use the fact that
cos θ = e i θ + e - i θ 2 .
This enables us to rewrite the left hand side of (8.86) as
12
The desired identity follows by simplifying the right hand side of this equation. To apply this identity
in (Equation 8.78) we differentiate the log function in this equation with respect to σ, use the identity
above, and then integrate the result.

8.3 LAPLACE TRANSFORM

In some instances the Laplace transform method provides another technique for the solution of
boundary value problems by converting the partial differential equation into an ordinary differential
equation.
We start with a review of the Laplace transform and its basic properties.

8.3.1 Basic Properties of the Laplace Transform

The Laplace transform L of a function f(t) is defined as
L(f)(s)=∫0∞e-stf(t)dt.
Here we assumed implicitly that the improper integral in this definition converges. Obviously the
Laplace transform is a linear operator that is:
L(af+bg)=aL(f)+bL(g)
where a, b are constants.
Example 8.3.1 Compute the Laplace transform of f(x) = δ(x - a) .
Solution 8.3.1 By definition,
L(f)(s)=∫0∞e-stδ(t-a)dt=e-sa
Example 8.3.2 Compute the Laplace transform of f(t) = t k for k ≥ - 1.
Solution 8.3.2
L(tk)(s)=∫0∞te-stdt=1sk+1∫0∞rke-rdr=Γ(k+1)sk+1,
where we made the substitution r = st.
Example 8.3.3 Compute the Laplace transform of the translated Heaviside
function
(8.87)
H(t-a)=1,t≥a0,t<a.
Solution 8.3.3
L[(H(t-a)])(s)=∫0∞H(t-a)e-stdt=∫a∞e-stdt=e-ass.
The most important property of the Laplace transform is the relation between L ( f ) and the Laplace
transform of the derivatives of f, i.e. L ( f ′ ) , L ( f ′ ′ ) , and so on.
Theorem 1:
(8.88)
L(f′)=sL(f)-f(0)
(8.89)
L(f′ ′ )=s2L(f)-sf(0)-f′ (O)
and so on.
This theorem can be proved by repeated use of the formula for integration by parts, e.g.
L(f′)=∫0∞e-stf′(t)dt=f(t)e-st|0∞+s∫0∞e-stf(t)dt=sL(f)-f(0)
The final step in the application of the Laplace transform to ordinary differential equations requires
the inversion of the transform. This is usually done with the aid of a table of Laplace transforms and
some “factor theorems.”
We quote some of these:
Theorem 2: If L ( f ) ( s ) = g ( s ) , then
(8.90)
L[eatf(t)](s)=g(s-a)
(8.91)
L[H(t-a)f(t-a)](s)=e-asg(s)
(8.92)
L[tf(t)](s)=-dgds(s)
Example 8.3.4 Compute
L-1[s+a(s+a)2+k2]
Solution 8.3.4 From a table we have,
L ( cos k t ) = s s 2 + k 2 = g ( s )
Therefore,
s+a(s+a)2+k2=g(s+a)
Hence from Theorem 2 we infer that
L - 1 [ s + a ( s + a ) 2 + k 2 ] = e - a t cos
Another important property of the Laplace transform is related to the convolution of two functions.
Definition: The convolution of two functions f, g is defined as
f∗ g=∫0tf(τ)g(t-τ)dτ.
Theorem 3: Let F(s),G(s) be the Laplace transforms of f and g respectively then
L(f∗ g)=F(s)G(s)
Thus the transform of the convolution equals the multiplication of the two transforms.
Example 8.3.5 Use the Laplace transfO rm to solve the following initial value problem:
(8.93)
y′′+4y′+3y=0,y(0)=2,y′(0)=-4
Solution 8.3.5 Applying the Laplace transfO rm to Equation (8.93) and using Equations ( 8.88 ), (
8.89 ), we obtain for g ( s ) = L ( y ) ( s ) .
g(s)=2s+4s2+4s+3=1s+1+1s+3.
From a table of Laplace transforms we have, however,
L(eat)(s)=1s-a.
Hence, applying the inverse Laplace transform to g(s) we have
y(t)=L-1[g(s)]=L-1(1s+1)+L-1(1s+3)=e-t+e-3t
Note that this is the solution of the diffe rential equation and the initial conditions.

8.3.2 Applications to the Heat Equation

In this section we present the solution of two boundary value problems related to the heat equation in
one space dimension using the Laplace transform
Example 8.3.6 We consider here the heat conduction in a rod which cooled down by its
surroundings due to the fa ct that it is not insulated laterally one dimensional radiator”)According
to Newton’s law of cooling the equation which governs the heat conduction in this rod is
(8.94)
ρc∂u∂t=κ∂2u∂x2-a(u-T0)
where ρ, c, κ, a, T 0 are all constants (T 0 is the ambient temperature). Making the transformation u
¯ = u - T 0 , dividing by ρc and then dropping the bar on u ¯ , Equation (8. 94) becomes
(8.95)
∂u∂t=k∂2u∂x2-bu
We now want to solve this equation for a semi‐infinite rod, 0 ≤ x < ∞, subject to the following
initial and boundary conditions:
(8.96)
u(x,0)=0,u(0,t)=C=constant
Furthermore we assume (on physical grounds and the modeling assumptions that lead to the heat
equation) t h a t u ( x , t ) remains bounded as x → ∞ for all t.
Solution 8.3.6 Let g ( x , s ) be the Laplace transform of u ( x , t ) with respect to t:
g(x,s)=L[u(x,t)]=∫0∞e-stu(x,t)dt.
It follows then that:
L[∂2u∂x2]=∂2∂x2L(u)=∂2g∂x2.
From Equation (8.88) and the initial condition in Equation (8.96) we obtain
(8.97)
L[∂u∂t]=sg(x,s)-u(x,0)=sg(x,s)
Hence the application of Laplace transform to Equation ( 8.95 ) yields
(8.98)
Kgxx(x,s)-(s+b)g(x,s)=0
Moreover the boundary conditions on g ( x , s ) are
g ( 0 , s ) = C s , lim x → ∞ g ( x , s ) = 0 .
Equation (8.98) is an ordinary diffe rential equation with constant coefficients in x (s is
considered to be a constant) whose solution subject to the boundary conditions above is
g(x,s)=Cse-αx,
where
α=(s+bk)1/2
Thus
∂u∂t=L-1[sg(x,s)]=L-1(Ce-αx)
From a table of Laplace transfO rms and Equation ( 8.90 ) we infer that
L-1(e-αx)=x2πkt3exp[-bt-x24kt]
The required solution of our boundary value problem is
u(x,t)=Cx2πk∫0texp(-bτ-x24kτ)τ3/2dτ.
Example 8.3.7 The heat equation for a thin homogeneous and insulated rod in which heat is being
generated at the rate of r(t) per unit volume is
(8.99)
∂u∂t=k∂2u∂x2+r(t)
Solution 8.3.7 We now solve this equation for a semi‐infinite rod x ≥ 0 subject to the following
boundary conditions
u(0,t)=0,u(x,0)=0
and assuming that u ( x , t ) is bounded for all x and t.(These boundary conditions imply that heat
will flow out at x = 0.)
Applying the Laplace transform to Equation (8.99) and using the notations of the previous
example we obtain
(8.100)
Kgxx(x,s)-sg(x,s)=R(s)
where R ( s ) = L [ r ( t ) ] . The boundary conditions on g are
g ( 0 , s ) = 0 , lim x → ∞ g ( x , s ) = 0 .
The solution of Equation ( 8.10 0) subject to this boundary conditions is
g(x,s)=R(s)s(1-e-xs/k)
Hence
∂u∂t=L-1[sg(x,s)]=L-1[R(s)]-L-1{R(s)exp[-Xsk]}
Using the convolution theorem we obtain
∂u∂t=r(t)-r(t)∗ [x2πkt3exp(-x24kt)]
Hence
u(x,t)=∫0t{r(τ)-r(τ)∗ [x2πkτ3exp(-x24kτ)]}dτ.

8.4 Numerical Solutions Of PDES

8.4.1 Finite Difference Schemes

In this section we demonstrate the application of the finite difference approximation that was
introduced in a previous chapter to the solution of various partial differential equations.

8.4.2 Numerical Solutions for the Poisson Equation

To solve numerically the Dirichlet problem for the Poisson or Laplace equations on a rectangular
D ⊂ R 2, that is
∇2 u = f(x) on D
and
(8.101)
u|∂D=g(x)
we introduce on D an equispaced grid with step size h. Using the finite difference formula in
(Equation 8.101) for ∂ 2 u/ ∂ x 2 and its equivalent for ∂ 2 y/ ∂ y 2, we approximate the differential
equation at each interior grid point ( x i , y j ) by
(8.102)
ui+1,j+ui-1,j-2uijh2+ui,j+1+ui,j-1-2uijh2=f(xi,yj)
that is,
(8.103)
ui+1,j+ui-1,j-4uij+ui,j+1+ui,j-1=hf(xi,yj)
Thus, for each interior point we obtain a linear equation in the unknowns u ij . Furthermore, since the
values of u on the boundary are known, the total number of equations equals the number of the
unknowns; that is, the boundary value problem has been reduced to a system of linear equations.
Example 8.4.1 Derive a system of linear equations for the solution of
∇ 2u=0
on
D={(x,y)0≤x≤1,0≤y≤1}
with the boundary conditions
u(x,0)=1,u(1,y)=1,u(0,y)=u(x,1)=0
and h = 1 3 (see Figure 8.8).

Figure 8.8 Computational Grid for example 8.4.1

Solution 8.4.1 From the boundary conditions we have

u1,2=u1,3=u2,4=u3,4=1
and
u2,1=u3,1=u4,2=u4,3=0.
Hence, to solve this problem we have to compute only u 22, u 23, u 32, and u 33.
From Equation ( 8.103 ) we obtain the following equations for the four inner grid points:
(8.104)
u2,3+u2,1-4u2,2+u3,2+u1,2=0
u2,2+u2,4-4u2,3+u3,3+u1,3=0
u3,1+u3,3-4u3,2+u4,2+u2,2=0
u3,2+u3,4-4u3,3+u4,3+u2,3=0.
Using the boundary conditions this yields
(8.105)
u2,3-4u2,2+u3,2=-1
u2,2-4u2,3+u3,3=-2
u3,3-4u3,2+u2,2=0
u3,2-4u3,3+u2,3=-1
which is the required system of equations.
8.4.2.1 Other Boundary Conditions
When the boundary value problem is of Neumann or mixed type, the number of equations derived
using (Equation 8.103) at the inner grid points will not be equal to the number of the unknowns since
the values of u on the boundary are not known. To overcome this difficulty we add additional grid
points outside the region whenever the boundary conditions are given in terms of the derivatives and
then use a finite difference formula to equalize the number of unknown to the number of equations.
The details of this “trick” are demonstrated in the following example.
Example 8.4.2 Solve
∇ 2u=0
on
D={(x,y)0≤x≤1,0≤y≤1}
with the boundary conditions
∂u∂y(x,0)=1,u(1,y)=1,u(0,y)=u(x,1)=0
and h = 1 3 .
Solution 8.4.2 : Since the boundary conditions along y = 0 are given in terms of ∂ u/ ∂ y, we add
the grid points u 0,2 and u 0,3 and use Equation ( 8.103 ) at the four inner grid points and the two
grid points along x = 0[i.e., ( 1 3 , 0 ) and ( 3 2 , 0 ) ] . We obtain six equations in eight unknowns.
These equations consist of the four given by the system in Equation (8.104) and
(8.106)
u1,1+u1,3-4u1,2+u2,2+u0,2=0
u1,2+u1,4-4u1,3+u2,3+u0,3=0
(See Figure 8.9).

Figure 8.9 Computational Grid for example 8.4.3

But
u3,1=u2,1=u1,1=u4,2=u4,3=0
and
u1,4=u2,4=u3,4=1.
Hence,
u2,3-4u2,2+u3,2+u1,2=0
u2,2-4u2,3+u3,3+u1,3=-1
u3,3-4u3,2+u2,2=0
u3,2-4u3,3+u2,3=-1
u1,3-4u1,2+u2,2+u0,2=0
u1,2-4u1,3+u2,3+u0,3=-1.
To solve the system we need two additional equations. These equations are obtained from the
boundary condition at ( 1 3 , 0 ) and ( 2 3 , 0 ) using the approximation
(∂u∂y)ij=ui,j+1-ui,j-12h.
This yields
u2,2-u0,2=23,
u2,3-u0,3=23.
The system now consists of eight equations in eight unknowns.

8.4.3 Irregular Regions

When the domain D has an irregular shape, the finite difference forumlas for the derivatives will not
be applicable to grid points whose distance from the boundary is less than h. Under these
circumstances we must introduce appropriate approximation formulas for the derivatives (see
Chapter 3). Thus, using the notation shown in Figure 8.10, we obtain the following finite difference
approximations at the grid point a:
(8.107)
(∂u∂x)a=u2-u1h(1+α1),(∂u∂y)a=u4-u3h(α2+α3)

Figure 8.10 Computational Grid near the boundary of an irregular domain

(∂2u∂x2)a=2h2u2+α1u1-(1+α1)uaα1(1+α1)
(∂2u∂y2)a=2h2α2u4+α3u3-(α2+α3)uaα2α3(α2+α3)
where u a = u(a) . The finite difference approximation scheme for ∇2 u = f at a is given by
u2+α1u1-(1+α1)uaα1(1+α1)+α2u4+α3u3-(α2+α3)uaα2α3(α2+α3
)=h22f(a)
Symmetry Considerations
In practical applications of the finite difference method, the number of equations to be solved can
be very large; for example, a two‐dimensional grid with 50 divisions along the x-and y-axes will
yield 2500 equations. The same grid in three dimensions will lead to 625,000 equations! This number
can be reduced, however, when the region D, the boundary conditions, and the differential equation
are invariant under certain asymmetry operations.
Example 8.4.3 Since the region and the boundary conditions as shown in Figure 8.11 are
invariant under reflections with respect to the x and y axes, it follows that the solutions of ∇2
u = 0, which is also invariant with respect to these operations, will satisfy

Figure 8.11 Using symmetry to reduce the number of unknowns

u1=u2=u3=u4
and so on.
Exercises
1. Derive a finite difference approximation scheme for the Poisson equation in three dimensions.
2. For a boundary value problem with circular symmetry in R 2, it is natural to use polar rather than
Cartesian coordinates. Derive a finite difference approximation to ∇2 u = f in these coordinates.
Hint: use a grid with constant ▵ θ and ▵ r. Note also that
∇ 2u=∂2u∂r2+1r∂u∂r+1r2∂2u∂θ2
3. Solve
∇ 2 u = r cos 2 θ u ( 1 , θ ) = 0
on the unit disk using ▵ θ = π/6 and ▵ r = 0.2. Compare with the exact solution.
4. Solve
∇ 2u=x2-y2
on
D={(x,y)0≤x≤1,0≤y≤1}
with the boundary conditions
∂u∂x(0,y)=1,u(x,O)=1,u(1,y)=u(x,1)=0.
Use different step sizes and compare the solutions obtained.
5. Complete the solutions of Examples 3,4, and 5. Use several different h.
6. Derive a finite difference scheme to solve
∇ 2u+(x2+y2)∂u∂x=f(x)
7. Solve numerically the differential equations that appear in the previous exercise if
f(x)=x2+y2
and the boundary conditions are the same as in Example 3.

8.4.4 Numerical Solutions for the Heat and Wave Equations

To obtain numerical solutions for the heat equation in one dimension,
(8.108)
∂2u∂x2=1k∂u∂t
with the boundary conditions
u(x,0)=f(x),u(0,t)=g1(t),u(a,t)=g(t)
on 0 ≤ x ≤ a and 0 ≤ t ≤ T. We introduce a grid with step size ▵ x, ▵ t in x and t, respectively, and use
finite difference formulas to approximate (Equation 8.108). Denoting u ( x i , t j ) by u ij leads to the
following finite difference equation at the grid point (ij):
(8.109)
ui+1,j-2uij+ui-1,j(△ x)2=1kui,j+1-ui,j△ t
or
(8.110)
ui,j+1=k△ t(△ x)2(ui+1,j+ui-1,j)+(1-2k△ t(△ x)2)ui,j
We observe that in (Equation 8.109) the forward difference formula is used to approximate ∂ u/ ∂ t
since otherwise (for the central difference formula) the values of u at t ± ▵ t are needed to evaluate u
at t + ▵ t, and these data are not available.
The formula in (Equation 8.109) expresses u at t j+1 in terms of its values at time t j ; hence, since u
at time t = 0 is given, we can compute u on [ 0 , T ] consecutively, that is, at ▵ t, 2 ▵ t, and so on
(there is no need to solve systems of equations). Furthermore, (Equation 8.110) can be simplified by
choosing r = k ▵ t/( ▵ x)2 to be equal to 1 2 , which leads to
(8.111)
ui,j+1=12(ui+1,j+ui-1,j)
The meaning of this formula is that on such a grid the value of u ( x i , t j + 1 ) is the average of u ( x i
- 1 , t j ) and u(x i+1, t j ).
The algorithm defined by (Equation 8.110) or (Equation 8.111) is called the explicit method for the
numerical solution of the heat equation. Its drawback is that in order to ensure its stability, that is, to
prevent the accumulated numerical error from becoming too large and therefore rendering the
numerical solution meaningless, we must choose r ≤ 1 2 . This condition places a restriction on ▵ t
that must satisfy ▵ t ≤ ( ▵ x 2)/2k and, therefore, increases the computational effort needed to obtain
the required solution.
To overcome this restriction, Crank and Nicholson observed that the forward difference formula
(∂u∂t)ij=ui,j+1-ui,j△ t
can be interpreted as a central difference formula at time t j + 1 2 △ t and, hence, a consistent
approximation scheme must evaluate ∂ 2 u/ ∂ x 2 at this time. To accomplish this we average the
approximations for ∂ 2 u/ ∂ x 2 at this time. To accomplish this we average the approximations for ∂ 2
u/ ∂ x 2 at t j and t j+1 and, thus, obtain the following finite difference formula for (Equation 8.108) at
(ij):
(8.112)
ui+1,j-2ui,j+ui-1,j2(△ x)2+
ui+1,j+1-2ui,j+1+ui-1,j+12(△ x)2=1kui,j+1-ui,j△ t
that is
(8.113)
-rui+1,j+1+(2+2r)ui,j+1-rui-1,j+1
=rui+1,j+(2-2r)ui,j+rui-1,j.
The algorithm represented by (Equation 8.113) is called the Crank-Nicholson algorithm. It requires
the solution of a system of linear equations at each time step, but it is stable for all values of r.
As for the wave equation,
(8.114)
∂2u∂x2=1c2∂2u∂t2
with the boundary conditions
u(x,0)=f(x),∂u∂t(x,0)=g(x)
u(0,t)=u(a,t)=0.
We use (Equation 8.101) to derive a finite difference approximation for (Equation 8.114). We obtain
ui+1,j-2ui,j+ui-1,j(△ x)2=1c2ui,j+1-2ui,j+ui,j-1(△ t)2.
Hence,
(8.115)
ui,j+1=w2(ui+1,j+ui-1,j)+2(1-w2)ui,j-ui,j-1
where w = (c ▵ t)/ ▵ x. Furthermore, if we set w = 1, the algorithm formulas will be simplified and we
obtain
(8.116)
ui,j+1=ui+1,j+ui-1,j-ui,j-1
We infer from (Equation 8.115) or (Equation 8.116) that we can calculate the values of u at t j+1 if
these values are known at t j and t j-1. It follows then that in order to begin the computation of u at ▵ t,
its values at - ▵ t must be known. We can easily overcome this difficulty, however, if we observe that
( ∂ u / ∂ t ) ( x , 0 ) = g ( x ) implies
(8.117)
ui,1-ui,-12△ t=g(xi)
Thus, using (Equations 8.116) and (8.117) we obtain for u i,1,
(8.118)
ui,1=12(ui+1,0+ui-1,0)+g(xi)△ t
=12[f(xi+1)+f(xi-1)]+g(xi)△ t.
The algorithm given by (Equation 8.115) is stable for w ≤ 1. However, unexpectedly the numerical
results obtained by setting w = 1 are better than those for w < 1; that is, decreasing the time step will
not bring the numerical solution closer to the exact solution.
Exercises
1. Solve the heat equation in one dimension with the following boundary conditions using the
explicit and Crank‐Nicholson algorithms. Compare the accuracy of the solutions and the
computational effort.
(a)u(x,0)=x,u(0,t)=T,u(a,t)=0
( b ) u ( x , 0 ) = sin x , u ( 0 , t ) = T , u ( a , t ) = 0
( c ) u ( x , 0 ) = cos x , ∂ u ∂ x ( 0 , t ) = c , u ( a , t ) = 0 .
2. Solve the wave (Equation 8.114) in one dimension with w = 1 and w = 1 2
using the following initial conditions:
( a ) u ( x , 0 ) = sin x , ∂ u ∂ t ( x , 0 ) = 0
( b ) u ( x , 0 ) = 0 , ∂ u ∂ t = sin x
( c ) u ( x , 0 ) = cos 2 x , ∂ u ∂ t ( x , 0 ) = sin x .
Always assume that u ( 0 , t ) = u ( a , t ) = 0 . Compare the numerical and exact solutions.
3. Derive a finite difference algorithm to solve
∂2u∂x2+b(x)∂u∂x=1k∂u∂t
subject to the appropriate boundary conditions.
4. Solve numerically the differential equation that appears in Exercise 3 if the boundary conditions
on u are u ( x , 0 ) = sin 2 x , u ( 0 , t ) = T , u ( a , t ) = 0, and b(x) = x.
5. Derive a finite difference scheme for
∂2u∂x2+b(x)∂2u∂x∂t=1c2∂2u∂t2
if u ( x , 0 ) = sin 5 x , ( ∂ u / ∂ t ) ( x , 0 ) = 0 , and u ( 0 , t ) = u ( a , t ) = 0 .
 CHAPTER 9

Variational Principles

CONTENTS
9.1 Extrema of Functions
9.2 Constraints and Lagrange Multipliers
9.3 Calculus of Variations
9.3.1 Natural Boundary Conditions
9.3.2 Variational Notation
9.4 Extensions
9.5 Applications
9.6 Variation with Constraints
9.7 Airplane Control, Minimum Flight Time
9.8 Applications In Elasticity
9.9 Rayleigh‐ritz Method
9.10 The Finite Element Method in 2-D
9.10.1Geometrical Triangulations
9.10.2Linear Interpolation in 2D
9.10.3Galerkin Formulation of FEM
9.11 Appendix
9.1 Extrema of Functions
It is well known from elementary calculus that the local extrema of a smooth function f = f(x) in one
variable coincides with the points x i at which f ’(x i ) = 0. Furthermore, an extremal point x i is a local
maximum or minimum if f ’’(x i ) < 0 or f ’’(x i ) > 0 respectively.
Similar criteria exist naturally for multivariable functions. Thus, for functions in two variables F =
F ( x , y ) we have the following:
Definition: Let F ( x , y ) be C 2( = twice differentiable with continuous derivatives) in some region
R. The Hessian of F at ( x , y ) ε R is defined as
(9.1)
H(x,y)=Fxx(x,y)Fxy(x,y)Fyx(x,y)Fyy(x,y)
Remark: This definition can be extended naturally to functions with n variables. Thus if F = F ( x ) =
F ( x 1 , … , x n ) , then
(9.2)
H(x)=|∂2F(x)∂xi∂xj|
Definition: a = ( a 1 , … , a n ) is said to be a critical point of F(x) if
(9.3)
∂F∂xi(a)=0;i=1
, . . ., n
We quote the following theorem without proof:
Theorem 9.1.1 Let F ( x , y ) be C 2 in a domain D and let ( a , b ) be a critical point of F in the
interior of D. If
1. H ( a , b ) > 0 and F x x ( a , b ) < 0 then F has a maximum at ( a , b ) .
2. H ( a , b ) > 0 and F x x ( a , b ) > 0 then F has a minimum at ( a , b ) .
3. H ( a , b ) < 0 then F has neither a maximum nor a minimum at ( a , b ) .
Example: Let F ( x , y ) = x 2 - y 2 Then F x = 2x, F y = - 2y, and hence at (0, 0)F x (0, 0) = F y
(0, 0) = 0 ( i . e . ( 0 , 0 ) is a critical point of F) . However, H(0, 0) = - 4 < 0 and, therefore, (0, 0) is
not a maximum or a minimum point of F. To see what it is, we observe that when we approach (0, 0)
along the x‐axis (y = 0), F ( x , 0 ) = x 2 and hence (0, 0)“looks like” a minimum. On the other hand,
if we approach (0, 0) along the y‐axis then F ( O , y ) = - y 2 and looks like a maximum. Thus, (0, 0)
is a saddle point.

9.2 Constraints and Lagrange Multipliers

In many practical applications one wants to find the extremum points of F(x) subject to a finite
number of constraints on the value of
(9.4)
x,gkbfx=ck,k=1,....,m
where c k are constants.
To solve such problems one uses Lagrange multipliers. The algorithm is described by the
following:
Theorem 9.2.1 : The extremum points of F(x) subject to the constraints Equation (9.4) are the
solutions x 1,.....,x n , λ 1 ,...., λ m of the m + n equations
(9.5)
∂f∂xi=0,i=1,...,n
(9.6)
gk=ck,k=1,....,m
where
(9.7)
f(x,λ)=F-λ1g1-…-λkgk
Remark: λ 1 ,..., λ m are called the Lagrange multipliers and f the auxiliary function of the problem.
Example: Find the points on the ellipse 2x 2 + 2xy + y 2 = 1 that are closest to the origin.
Solution: we want to minimize
(9.8)
F(x,y)=x2+y2
(which is equivalent to minimizing (x 2 + y 2)1/2) subject to the constraint
(9.9)
g1(x,y)=2x2+2xy+y2=1
The auxiliary function f ( x , y , λ ) is therefore
(9.10)
f(x,y,λ)=x2+y2-λ(2x2+2xy+y2)
Hence the extremum points have to satisfy the equations
(9.11)
2x-4λx-2λy=0

(9.12)
2y-2λx-2λy=0
(9.13)
2x2+2xy+y2=1
Solving (Equation 9.11) for y yields
(9.14)
y=-x(-1+2λ)λ
Substituting this expression for y in (Equation 12.9) and solving for λ we obtain
λ=3±52.
Using this expression of λ in (Equation 9.14) and then substituting for y in (Equation 12.10) we obtain
the following solutions for ( x , y ) :
( x , y ) = ( ± 0.528 , ± 0.325 ) , ( x , y ) = ( ∓ 0.850 , ± 1.376 ) .
The first two pairs represent the points with minimum distance from the origin while the last two
pairs have a maximum distance from the origin.
Remark: It should be observed from this example that the method of Lagrange multipliers is
“easiest to apply” when F is at most a quadratic polynomial and g k are linear functions since in this
case the solution of the resulting system of equations for the extremum points is straightforward.
Exercises
1. Plot the ellipse 2x 2 + 2xy + y 2 = 1 and identify the points closest to the origin.
2. Find the point on the ellipse x 2 - 2xy + 2y 2 = 1 with maximum distance from the origin.
3. Find the dimensions of the right circular cylinder of fixed total surface area A (including top and
bottom) with maximum volume.
4. Find the relative extrema of
F(x,y,z)=xz+yz
which lie on the intersection of the surfaces
x2+y2=2,yz=2
(two constraints).
5. Show that
(a) Among all triangles with the same perimeter the equilateral triangle has the greatest area.
(b) Among all rectangles with the same perimeter the square encloses the greatest area.

9.3 Calculus of Variations

The basic problem of the calculus of variations can be stated as follows: Find the function y(x)
defined on [ x 1 , x 2 ] and satisfying the boundary conditions
(9.15)
y(x1)=y1,y(x2)=y2
so that the value of the integral
(9.16)
I(y)=∫x1x2f(x,y,y′)dx
is at extremum.
The following examples show how such problems arise in actual applications.
Example 1: Formulate a variational principle to find the arc of minimum length that passes through
given end points ( x 1 , y 1 ) , ( x 2 , y 2 ) .
Solution: The differential of the distance in two dimensions is
(ds)2= (dx)2+ (dy)2,
hence our problem is to find y(x) which minimizes the integral
(9.17)
I(y)=∫ds=∫(dx)2+(dy)2=∫x1x21+(y′)2dx
with
(9.18)
y(x1)=y1,y(x2)=y2
Example 2: Find the plane curve between ( x 1 , y 1 ) , ( x 2 , y 2 ) ,y 1, y 2 > 0 which
generates the smallest surface area when revolved around the x‐axis.
Solution: As is well known, the surface area generated by revolving a curve y = y(x) around the x‐
axis is
(9.19)
S(y)=∫x1x22πy1+(y′ )2dx
We have to find the function y(x) which minimizes the value of this integral and satisfies the boundary
conditions (Equation 12.26).
To solve the basic variational problem (or at least recast it in terms of differential equations) we
proceed as follows: let y(x) be the desired optimal solution (unknown as yet). Consider the following
set of functions:
(9.20)
Y(x,ε)=y(x)+εη(x)
where η(x) is an arbitrary but fixed function which satisfies the boundary conditions
(9.21)
η(x1)=η(x2)=0
For a function Y ( x , ε ) in this family the variational integral ((Equation 3.2)) takes the form
(9.22)
I(ε)=∫x1x2f(x,Y,Y′ )dx
Thus the variational integral “degenerates” into a one variable function in ɛ. Moreover, since we
know that at ɛ = 0I(ɛ) has an extremum (according to our assumption on y(x)), it follows that I
’(0) = 0, i.e.

(9.23)
I′ (0)=∫x1x2[∂f∂YdYdε+dfdYdY′ dε]|ε=0dx=0
But from (Equation 12.28)
dYdε=η,dY′ dε=η′
and Y ( x , 0 ) = y ( x ) , Y ′ ( x , 0 ) = y ′ ( x ) ; therefore, (Equation 10.35) reduces to
(9.24)
∫x1x2[∂f∂yη+∂f∂y·η′]dx=0
Integrating the last term in (Equation 10.36) by parts we obtain
(9.25)
[∂f∂y·η]|x1x2+∫x1x2[∂f∂y-ddx[∂f∂y]]η(x)dx=0
Figure 9.1 Optimal trajectory and another one close by

and hence by (Equation 9.21)

(9.26)
∫x1x2[∂f∂y-ddx[∂f∂y]]η(x)dx=0
(Equation 9.26) must hold true for any choice of η(x) which satisfies (Equation 10.46); hence we
infer that the extremum function y(x) must satisfy the partial differential equation
(9.27)
∂f∂y-ddx∂f∂y=0
This equation is called the Euler‐Lagrange equation.
We observe that when f has no explicit dependence on x, i.e. f = f ( y , y ′ ) , then (Equation 10.39)
can be written as
(9.28)
ddx[y′∂f∂y-f]=0
In fact
ddx[y′∂f∂y-f]=y′′∂f∂y+y′ddx[∂f∂y]-y′∂f∂y-y′′∂f∂y
(9.29)
=-y′[∂f∂y-ddx[∂f∂y]]=0
Example 1 (cont.): For this problem
(9.30)
f(x,y,y′)=1+(y′)2
Hence from (Equation 10.39) we obtain
(9.31)
ddx[y′1+(y)2]=0
i.e.
 (9.32)
y′1+(y′)2=const=c
or
(9.33)
y′ =c21-c2=A(const.)
Hence
(9.34)
y=Ax+B
as we expected!
Example 3: Find the plane curve along which a particle of mass m starting from rest will slide
without friction from ( x 1 , y 1 ) to ( x 2 , y 2 ) , y 2 < y 1, in the shortest time.
Remarks:
1. From a historical perspective this was the original problem that led Euler to “invent” the
calculus of variations. The solution curve to this problem is called the “brachistochrone.”
2. To simplify the algebra we can assume without loss of generality that
( x1,y1) =( 0,0) .
Solution: Since the speed of the particle along the curve is
(9.35)
v=dsdt
the total time of descent from 0 to ( x 2 , y 2 ) is
(9.36)
I(y)=∫0x2dt=∫0x2dsv=∫0x21+(y′)2vdx
However by conservation of energy we have
(9.37)
12mv2-12mv12=mg(y-0)
Since v 1 = 0 we infer that
(9.38)
I(y)=12g∫0x21+(y′)2y
i.e.
(9.39)
f(y,y′)=12g1+(y′)2y
From (Equation 10.40) we infer that
(9.40)
(y′)2y1+(y′)21+y′)2y=c¯
Solving this for y ’ and integrating we obtain
(9.41)
x=∫y2c-ydy
where c ¯ = 1 2 c . Substituting
(9.42)
y = 2 c sin 2 θ 2
we finally obtain a parametric representation of the required curve
 (9.43)
x = c ( θ - sin θ )
(9.44)
y = c ( 1 - cos θ )
which is a cycloid.
Example 4: Find the shortest path between two points on a sphere of radius R.
Remarks: A curve on a surface that provides the shortest path between two points is called a
“geodesic.”
Solution: A curve on a sphere is given by
(9.45)
x = R sin ϕ cos θ , y = R sin ϕ sin θ , z = R cos ϕ
where θ = θ(t), φ = φ(t) . Hence
(9.46)
( d s ) 2 = ( d x ) 2 + ( d y ) 2 + ( d z ) 2 = R 2 ( d ϕ ) 2 + R 2 sin 2 ϕ ( d θ ) 2
The variational problem is therefore to minimize
(9.47)
I = ∫ t 1 t 2 d s = R ∫ θ 1 θ 2 ( ϕ ′ ) 2 + sin 2 ϕ d θ
where ϕ ′ = d ϕ d θ . Hence
(9.48)
f ( θ , θ ′ ) = R ( ϕ ′ ) 2 + sin 2 ϕ
Since f does not have explicit dependence on θ, we can use (Equation 10.40) to obtain
(9.49)
R ( ϕ ′ ) 2 ( ϕ ) 2 + sin 2 ϕ - R ( ϕ ′ ) 2 + sin 2 ϕ = c 1
Solving for (φ ’) and integrating we obtain
(9.50)
ϕ = c 1 ∫ d ϕ R 2 sin 4 ϕ - c 1 2 sin 2 ϕ = - arcsin { cot ϕ b } - c 2
where c 1 c 2 are integration constants and b = R / c 1 2 - 1 . Hence
(9.51)
( sin c 2 ) R sin ϕ cos θ + ( cos c 2 ) R sin ϕ sin θ - R cos ϕ b = 0
or in Cartesian coordinates
(9.52)
x sin c 2 + y cos c 2 - z b = 0
Thus the shortest arc that connects ( θ 1 , ϕ 1 ) and ( θ 2 , ϕ 2 ) on the sphere is the intersection
between the sphere and a plane that passes through its center. Thus the arc is part of the great circle
connecting the points on the sphere.

9.3.1 Natural Boundary Conditions

In some variational problems the boundary values of y at one or both end points may not be specified.
Since y(x i ) remains arbitrary, it is impossible to choose η to be zero at such an end point ((Equation
9.21)). However, it is still possible to derive the Euler‐Lagrange equation for this problem if we note
that η(x i ) = 0 was used only to “discard” ∂ f ∂ y η | x 1 x 2 in (Equation 9.25). This can be also be
done, however, by choosing
(9.53)
∂f∂y=0atx=xi,i=1,2
(if y(x i ) is not specified). This condition is then called the natural boundary condition of the problem.

9.3.2 Variational Notation

Many books on variational problems introduce the concept of variation δy(x) to replace ɛη(x) that we
used above. This implies
(9.54)
δy′(x)=εη′(x)
and
(9.55)
ddx(δy)=εη′=δy′
Moreover, given a function F ( x , y , y ′ ) one defines the variation of F as
δF=△ F=F(x,y+εη,y′+εη′)-F(x,y,y′)
(9.56)
≅ ∂F∂y(εη)+∂F∂y(εη′)=∂F∂yδy+∂F∂yδy′
It is easy to verify the following:
(9.57)
δ(Fn)=nFn-1δF
(9.58)
δaF1+bF2=aδF1+bδF2,a,b,constants
(9.59)
δ(F1·F2)=F1δF2+F2δF1
The variational principle (Equation 12.24) can now be rewritten as
(9.60)
δI=△ I=I(y+εη)-I(y)=∫x1x2δf(x,y,y′)dx
Exercises
1. Fermat principle for the motion of a light ray in a medium states that “light will travel between
two points along the path which minimizes the transmission time.” Formulate the corresponding
variational principle for the motion of light in a medium in which the speed of light at any point in any
direction is a function of the position only.
2. Show the Euler‐Lagrange equation for a variational principle which contains second order
derivatives of y, viz.
I=∫x1x2f(x,y,y′,y
is
∂f∂y-ddx[∂f∂y′]+d2dx2[∂f∂y′′]=0.
Hint: Note that the boundary conditions on y(x) are
y(x1)=a1y(x2)=a2
y′(x1)=b1y′(x2)=b2
where a i , b i are constant.

9.4 EXTENSIONS
When the function f in the variational integral is a function of more than one dependent variable, e.g.
f=f( x,y,x˙ ,y˙ ,t) ,
the extremum of the variational integral
(9.61)
I=∫t1t2f(x,y,x˙,y˙,t)dt
is achieved by the functions x(t),y(t) satisfying the following Euler‐Lagrange equations.
(9.62)
∂f∂x-ddt[∂f∂x˙]=0

(9.63)
∂f∂y-ddt[∂f∂y˙]=0
Similarly for the n‐dimensional case (i.e. n‐dependent and one independent variables)
(9.64)
I=∫t1t2f(x,x˙,t)dt,xεRn
the extremum functions have to satisfy
(9.65)
∂f∂xi-ddt[∂f∂x˙i]=0,i=1,...,n
Another extension of the Euler‐Lagrange equations is required when f is a function of more than one
independent variable but one dependent variable, i.e.
(9.66)
I=∫∫Df(x,y,w,wx,wy)dA,DεR2
To prove the analog of (Equation 12.32) in this case let
(9.67)
W(x,y)=w(x,y)+εη(x,y)
where w ( x , y ) is the extremal function we are looking for.
(9.68)
0=dIdεε=0=∫∫D[∂f∂wη+∂f∂wxηx+∂f∂wyηy]dA
By applying Green’s theorem in two dimensions (see Appendix) we obtain
(9.69)
∫∫D[∂f∂wxηx+∂f∂wyηy]dA=
-∫∫Dη[∂∂x[∂f∂wx]+∂∂y[∂f∂wy]]dA
+∫∂Dη[∂f∂wxdy-∂f∂wydx]
Assuming that w is specified on the boundary of the domain ∂ D we deduce that η|∂D = 0. Hence the
extremum function w ( x , y ) has to satisfy the following partial differential equation:
 (9.70)
∂f∂w-∂∂x[∂f∂wx]-∂∂y[∂f∂wy]=0
Exercises
1. Use variations on (Equation 12.33) to derive (Equation 12.34).
2. Find the functions x(t),y(t),z(t) which minimize the integral
(9.71)
I=∫t1t2x˙2+y˙2+z˙2dt
and satisfy the boundary conditions
(9.72)
x(ti)=xiy(ti)=yiz(ti)=zii=1,2
3. On a Riemanian manifold the infinitesimal distance is given by
(9.73)
ds2=gijxdxidxj
(implicit simulation o n i , j ) . To find the equation of the curve x(t) of minimum length that connects
two points on the manifold ( = geodesic curve) it is enough to minimize the square of the distance, i.e.
minimize
(9.74)
I(x)=∫t1t2gij(x)dxidtdxjdtdt
Derive the explicit form of Euler‐Lagrange equation for x(t) .

9.5 APPLICATIONS
A key point in the application of variational principles to mechanical (and other) systems is
Hamilton’s principle.
Definition: If V:R n → R the gradient of V is defined as
gradV = ∇ V = [ ∂ V ∂ x 1 , ...., ∂ V ∂ x n ]
Definition: A system is called conservative if the forces acting on the system can be expressed as
the gradient of some scalar function V which is called the potential of the force.
Hamilton Principle: For a conservative system the actual motion (or trajectory) will minimize the
variational integral
(9.75)
I=∫t1t2(T-V)dt
where T is the kinetic energy of the system. (T - V) is also called the Lagrangian of the system.
Remark: The kinetic energy of a point particle of mass m is 1 2 m v 2 where y is the velocity of the
particle.
Example 1: Derive the equations of motion for a system of two masses and three springs as shown
in the diagram:
Figure 9.2 System of two masses and three springs

Solution: Let x, y be the displacement of the masses from equilibrium at time t. The kinetic energy
is then given by
(9.76)
T=12mx˙2+12my˙2
The potential energy of the springs in this position is given by
(9.77)
V=12kx2+12k(y-x)2+12.ky2
Therefore the variational integral for this system is
(9.78)
I=∫t1t2[12m(x˙2+y˙2)-k(x2-xy+y2)]dt
This is a variational integral with one independent and two dependent variables. Euler‐Lagrange
(Equations 12.31)‐(12.32) then lead to
(9.79)
mx¨+2kx-ky=0
(9.80)
my¨+2ky-kx=0
Example 2: Vibrating Membrane (Drum head)
We consider a “flexible membrane” which is stretched over some region D in the x - y plane and
bounded by a curve C. Denote the deviation of the membrane from the x - y plane by w ( x , y , t )
(where we assume that on the boundary w ( x , y , t ) | C = 0 ). The membrane is set in motion by an
initial disturbance from the equilibrium position, and we want to derive an equation of motion for w (
x,y,t) .
Solution: The kinetic energy of the membrane is obviously given by
(9.81)
T=12∫∫Dρ(x,y)wt2dxdy
To evaluate the expression for the potential energy we first note that for conservative systems
V(b) - V(a) is the amount of work needed to take the system from state a to state b.
For our system we can assume that the amount of work needed to take the membrane from its
equilibrium state (x - y plane) to another is proportional to the difference in the surface area (Hooke’s
law in two dimensions). Hence
(9.82)
V(w)=V(w)-V(w=0)=k∫∫D{1+wx2+wy2-1}dxdy.
For small deflections we can use the approximation
(9.83)
1+u≅ 1+12u,|u|<<1
to obtain
(9.84)
V(w)=k2∫∫D(wx2+wy2)dxdy
The variational integral which we have to consider is therefore
(9.85)
I=12∫t1t2{∫∫D[ρwt2-k(wx2+wy2)]dxdy}dt
Euler‐Lagrange equations for this case yield (one dependent variable and three independent ones)
(9.86)
k(wxx+wyy)-ρ(x,y)wtt=0
or equivalently
(9.87)
k∇ 2w=ρwtt
which is the wave equation in two dimensions.
Exercises
1. Modify the discussion above so that gravity is included.
2. Derive the equations of motion for the double pendulum.

Figure 9.3 Double pendulum

Hint: From the drawing it is obvious that

(9.88)
x 1 = ℓ 1 sin θ , y 1 = ℓ 1 cos θ
(9.89)
x 2 = ℓ 1 sin θ + ℓ 2 sin ϕ , y 2 = ℓ 1 cos θ + ℓ 2 cos ϕ
Hence
 (9.90)
x ˙ 1 = ℓ 1 cos θ d θ d t , y ˙ 1 = - ℓ 1 sin θ d θ d t

(9.91)
x ˙ 2 = ℓ 1 cos θ d θ d t + ℓ 2 cos ϕ d ϕ d t
(9.92)
y ˙ 2 = - ℓ 1 sin θ d θ d t - ℓ 2 sin ϕ d ϕ d t
The expressions for the kinetic and potential energy are
(9.93)
T=m12(x˙12+y˙12)+m22(x˙22+y˙22)
(9.94)
V = m 1 g ℓ 1 ( 1 - cos θ ) + m 2 g ( ℓ 1 + ℓ 2 - ℓ 1 cos θ - ℓ 2 cos ϕ )
3. Derive the equations of motion of the system shown in Figure 9 where m 1 is constrained to move
only in the vertical direction.

Figure 9.4 Pendulum suspended on a spring mass system

Hint: x 2 = ℓ 2 sin θ
(9.95)
y 2 = y 1 + ℓ 2 cos θ

(9.96)
T=12m1y˙12+m22(x˙22+y˙22)
(9.97)
V = 1 2 k ( y - ℓ 0 ) 2 + m 1 g [ - y + ℓ 0 ] + m 2 g [ ℓ o + ℓ 2 - ( y 1 + ℓ 2 cos θ ) ]
where ℓ0 is the natural length of the spring.
4. Solve the system described by (Equations 12.41)–(12.42) with m 1 ≠ m 2 and k 1 ≠ k 2 ≠ k 3.

9.6 Variation with Constraints

In many applications of variational principles we have to find the extremum of a variational integral
subject to constraints. The method used to solve these problems is an extension of the Lagrange
multiplier approach for functions. We shall illustrate it with few examples in this section.
Example 1: Find ψ(x) so that
1. (Variational Integral)
(9.98)
I(ψ)=∫x1x2(ψx+ψx2)dx
is extremum
2. (Constraint)
(9.99)
J(ψ)=∫x1x2ψ2(x)dx=1
3. (Boundary Conditions)
(9.100)
ψ(x1)=ψ(x2)=0
Solution: Let ψ(x) be the solution. For a small variation near ψ(x) we can write
(9.101)
ψ¯(x)=ψ(x)+εη(x)
Hence
(9.102)
I(ε)=∫x1x2[ψx+εηx(x)+(ψx+εηx)2]dx
(9.103)
J(ε)=∫x1x2(ψ+εη)2dx=1
and η(x 1) = η(x 2) = 0.
Applying the Lagrange multiplier approach to the function I(ɛ) and the constraint J(ɛ) we now form
(9.104)
K(ε,λ)=I(ε)+λJ(ε)
Since by assumption ψ is the solution of our problem we must have
(9.105)
∂K(ε,λ)∂ε|ε=0=0
and
(9.106)
J(ε=0)=1
But
(9.107)
∂K∂ε|ε=0=∫x1x2[ηx(x)+2ψxηx+2ηψλ]dx
However,
(9.108)
∫x1x2(ηx+2ψxηx)dx=(η+2ψxη)|x1x2-∫x1x22ψxxηdx=-∫x1x2ψxx
ηdx.
Thus we infer that for all η
(9.109)
∫x1x2(-2ψxx+2λψ)ηdx=0;
i.e. ψ must satisfy
(9.110)
ψxx-λψ=0
with the constraint
(9.111)
∫x1x2ψ2dx=1
A direct method to obtain the same result is to construct the Lagrange multiplier functional
(9.112)
K(ψ,ψx)=I(ψ,ψx)+λJ(ψ,ψx)
and compute the optimal solution as the simultaneous solution to a Euler‐Lagrange equation for the
variational principle K and the constraint J. In fact,
(9.113)
∂K∂ψ-ddx[∂K∂ψx]=2λψ-ddx[1+2ψx]=0
i.e.
ψxx-λψ=0.
Example 2: Find the curve of length 3 which passes through (0, 0) and ( 1 , 0 ) for which the area
between the curve and the x‐axis is a maximum.
Solution: If the equation of the curve is ψ = ψ(x), then we want to find the extremum of
I(ψ)=∫01ψ(x)dx,ψ(0)=0,ψ(1)=0
subject to the constraint
J(ψ)=∫011+ψ′2dx=3.
Following the Lagrange multiplier approach we obtain the following Euler‐Lagrange equation:
(9.114)
ddx[∂∂ψ[ψ+λ1+ψ′2]]-∂∂ψ[ψ+λ1+ψ′2]=0
Hence
(9.115)
λddxψ′1+ψ2-1=0
Integrating (Equation 9.115) we find (after some algebra) that ψ must satisfy
(x-c1)2+(ψ-c2)2=λ2;
i.e. ψ is a circle. The three constants c 1, c 2, and λ are to be determined so that ψ(0) = ψ(1) = 0 and
the requirement that the curve length is 2.
Exercises
1. Find the explicit expression(s) for ψ in Example 1 if x 1 = 0 and x 2 = 2.
 2. Find the extremum of the variational integral
I(ψ)=∫x1x2[ψ2ψx+ψx2]dx
subject to the constraint
J(ψ)=∫x1x2ψ2dx=1.
3. Show that the extremum of
I(ψ)=∫x1x2(p(x)(ψ′(x))2-q(x)ψ2)dx
subject to the constraint
J(ψ)=∫x1x2r(x)ψ2(x)dx=1
ψ(x1)=aψ(x2)=b
are solutions of the differential equation
(pψ′)′+(q+λr)ψ=0.

9.7 Airplane Control, Minimum Flight Time

As is well known, some (sea) fighter planes take off vertically. In this section we consider a
simplified model (with no air drag) for the vertical take off of such a plane in order to find out the
optimal thrust distribution so that the plane achieves a specific height in minimum time. We assume
that the total amount of fuel available is constant.
To formulate a mathematical model for this problem we introduce the following data and
approximations.
1. The simplified equation of motion for the plane (when air drag is neglected) is given by
(9.116)
x¨=-g+u(t)
where x(t) is the plane height at time t from the ground and u is the thrust.
2. The fuel consumption of the plane engine is proportional to the square of the thrust. Since the
amount of fuel available is constant, it follows that the total thrust of the engine over the flight time T
must satisfy the constraint
(9.117)
∫0Tu(t)2dt=constants=c
3. The boundary conditions on the flight of the plane are
(a)
(9.118)
x(O)=x˙(0)=0;
i.e. the plane starts from the ground with velocity 0.
(b)
(9.119)
x(T)=h;
i.e. the plane must achieve a predetermined height.
4. We want to minimize the vertical flight time T, i.e. minimize
(9.120)
I(T)=∫0T1.dt=T
Model: In this problem we must minimize (Equation 9.120) subject to the constraints given by
(Equations 9.116), (10.103) and the boundary conditions (Equation 9.118), and (Equation 9.119). As
a first step, however, we must convert the constraint given by the differential equation (Equation
9.116) into an “algebraic” relationship so that the problem becomes amenable for treatment by the
method of Lagrange multipliers as described in the previous section. Integrating (Equation 9.116)
over the interval [ 0 , τ ] and using the condition ẍ ( 0 ) = 0 we obtain
(9.121)
x˙(τ)=-gτ+∫0τu(s)ds
Integrating this equation over [ 0 , t ] and using the condition x(O) = 0 yields
(9.122)
x(t)=-gt22+∫0t[∫0τu(s)ds]dτ
We now invert the order of integration of the double integral in (Equation 9.122)
∫0t{∫0τu(s)ds]dτ=∫0t∫stu(s)dτds
(9.123)
=∫0t(t-s)u(s)ds
Hence,
(9.124)
x(t)=-gt22+∫0t(t-s)u(s)ds

Figure 9.5 Region of integration in (Equation 9.122)

Thus the variational problem under consideration is to minimize

(9.125)
I=∫0T1dt
subject to the constraints
(9.126)
J1(T,u)=-∫0Tgtdt+∫0T(T-t)u(t)dt=h
(9.127)
J2(T,u)=∫0Tu2(t)dt=c
We see then that the variational problem under consideration has two dependent variables T, u, one
independent variable t, and two constraints.
To compute the optimal solution we now follow the Lagrange multiplier approach and construct the
functional
(9.128)
K(T,u)=I-λ1J1-λ2J2
The optimal solution must then satisfy Euler‐Lagrange equations for T, u and the constraints given by
(Equations 9.126)‐(9.127). Thus
(9.129)
∂K∂T-ddt[∂K∂T]=1+λ1gT-λ2u2(T)-λ1∫0Tu(t)dt=0

(9.130)
∂K∂u-ddt[∂K∂u˙]=-λ1(T-t)-2λ2u(t)=0
Hence
(9.131)
u(t)=-λ1(T-t)2λ2
(and therefore, as expected u(T) = 0).
Substituting (Equation 9.131) in (Equations 9.126), (9.127) and (10.115) leads to the following
three equations in the unknowns T, λ 1 , λ 2 .
(9.132)
T2(λ1T+3λ2g)=-6λ2h
(9.133)
λ 1 2 T 3 = 12 λ 2 2 c 2
(9.134)
λ1T(λ1T+4λ2g)=-4λ2
From Equation (9. 133) we infer
(9.135)
λ1λ2=-23cT3/2;
hence for the optimal flight time (Equation 9.132) yields the equation
(9.136)
3gT2+6h=23cT3/2
Finally, using (Equations 9.131) and (9.135), we obtain an explicit expression for the thrust as a
function of time:
(9.137)
u(t)=3Tc[1-tT]
Exercises
1. Explain in detail the inversion of the integral in (Equation 9.123).
2. Reconsider the minimum lift off time problem when air drag has to be taken into account; i.e. the
equation of motion for the plane is given by
x¨=-g-αx˙+u(t)
where α is a positive constant.
 3. Reconsider the problem of this section when one wants to minimize the fuel consumption (e.g.,
in lift off from the moon surface).
9.8 APPLICATIONS IN ELASTICITy
In this section we discuss the modeling of transverse vibrations in elastic bars and thin plates using
variational principles.
A. Transverse vibrations in an elastic bar.
Consider an elastic bar of constant cross section, density ρ per unit length, and length L. To model
the transverse vibrations of such a bar we ignore the possible distortion of the cross sections and
assume that any small section of the bar moves as a rigid entity.
If u ( x , t ) is the displacement from equilibrium, the total kinetic energy of the bar is given by
(9.138)
T=ρ2∫0Lu˙2dx
The strain potential energy due to the vibrations is given by
(9.139)
J=∫∫Dz2dydz
By Hamilton’s principle the motion of the bar will be an extremum of the variational integral
(9.140)
I(u)=12∫t1t2∫0L(ρu˙2-EJuxx2)dxdt
with the boundary condition
(9.141)
u=ux=0atx=0,L
(clamped rod).
To derive the appropriate Euler‐Lagrange equation for I(u) we introduce
(9.142)
u=u+εη
in (Equation 10.91) where u is the (sought for) extremal solution. Hence
(9.143)
I′(0)=∫t1t2∫0L[∂f∂u˙η˙+∂f∂uxx]ηxxdxdt=0
where
(9.144)
f=12[ρu˙-EJuxx2]
Using integration by parts and the boundary conditions we obtain
(9.145)
∫t1t2∫0L∂f∂u˙η˙dxdt=-∫t1t2∫0L∂∂t[∂f∂u˙]ηdxdt
(9.146)
∫0L∂f∂uxxηxxdx=∫0L∂2∂x2[∂f∂uxx]ηdx
Hence
(9.147)
∂∂t[∂f∂u˙]-∂2∂x2[∂f∂uxx]=0
which yields
 (9.148)
ρ∂2u∂t2+EJ∂4u∂x4=0
To solve this equation we need the boundary conditions at 0, L and the initial position and velocity of
the bar.
B Transverse vibrations of thin plate.
Consider the transverse vibrations of a uniform thin plate over a domain
R with the boundary conditions
(9.149)
u=∂u∂n|∂R=0
The kinetic energy of the plate is
(9.150)
T=12ρ∫Ru˙2dA
and the total strain potential energy is
(9.151)
V=12D∫∫R[(∇ 2u)2-2(1-σ)(uxxyyy-uxy2)]dA
Here D is the “flexural rigidity” of the plate and σ is the “Poisson ratio” of the material (σ is related
to the relationship between the strain‐stress tensors in the material of the plate).
The variational integral for this problem is
(9.152)
I(u)=
12∫t1t2∫∫R{ρu˙-D[(∇ 2u)2-2(1-σ)(uxxuyy-uxy2)]}dAdt=
12∫t1t1∫∫RfdAdt
Letting
(9.153)
u=u+εη
we obtain
(9.154)
I′ (0)=∫t1t2∫∫R[∂f∂u˙η˙+∂f∂uxxηxx+∂f∂uyyηyy+∂f∂uxyηxy]dAd
t=0
Using integration by parts, Green’s theorem and the boundary conditions yields:
(9.155)
∫t1t2∫∫R∂f∂u˙η˙dAdt=-∫t1t2∫∫Rη∂∂t[∂f∂u˙]dAdt
(9.156)
∫∫R∂f∂uxxηxx=∫∫Rη∂2∂x2[∂f∂uxx]dA+∫C[ηx∂f∂uxx-η∂∂x[∂f∂ux
x]]dy
etc.
Combining all these results together we finally obtain
(9.157)
ρ∂2u∂t2+D∇ 2(∇ 2u)=0
This equation is referred to as the biharmonic equation.
9.9 RAYLEIGH‐RITZ METHOD
Variational formulation of differential equations can be used to obtain approximate solutions for these
equations. The basic technique is due to Rayleigh‐Ritz, and it was the “precursor” to the current
Finite Element Methods that are used for the numerical solution of partial differential equations in
various applications.
We present this method through two examples:
Example 1: The bending of an elastic bar under uniform loading.
In the previous section we derived a variational principle for the vibrations of an elastic bar. If we
apply on such a bar a uniform static loading, p, the appropriate variational principle for the static
shape of this bar will be
(9.158)
I(u)=∫0L[12EJ(u′′)2-pu]dx
Euler‐Lagrange equations then yield
(9.159)
EJuxxxx-p=0
If the rod is clamped at x = 0, the corresponding boundary conditions are
(9.160)
u(0)=u′(0)=0
We can solve (Equation 9.159) directly (as an ordinary differential equation) or use the variational
principle given by (Equation 9.158) to obtain at least an approximation to the solution. To this end we
consider (as an example) the function space S = { 1 , x , x 2 , x 3 , x 4 } and attempt to find the best
approximation to the solution in this space; i.e., we seek to find the values of a, b, c, d, e which yield
the best approximation to the solution u in the form
(9.161)
u=ax4+bx3+cx2+dx+e
These values of a, b, c, d, e will minimize I(u) in the function space S subject to the boundary
conditions.
As a first step toward the solution we apply the boundary conditions (Equation 9.160) to (Equation
9.161). This yields d = e = 0, i.e.
(9.162)
u(x)=ax4+bx3+cx2
Substituting (Equation 9.162) in (Equation 9.158) we obtain after integration:
(9.163)
I(u)=-pL3[15aL2+14bL+13c]+
E J [ 72 5 a 2 L 5 + 18 a b L 4 + 1 3 ( 24 a c + 18 b 2 ) ) L 3 + 6 b c L 2 + 2 c 2 L ]
This expression attains its minimum at a point where
(9.164)
∂I∂a=∂I∂b=∂I∂c=0
This leads to a system of three linear equations for a, b, c whose solution is
(9.165)
a = p 24 E J , b = - p L 6 E J , c = p L 2 4 E J
Example 2: Solve
(9.166)
∇ 2u=c
on the square Ω = [ - a , a ] × [ - a , a ] subject to the Dirichlet boundary condition
(9.167)
u|∂Ω=0
Solution: The solution of (Equation 9.166) satisfies the variational principle
(9.168)
I(u)=∫Ω[12(ux2+uy2)+cu]dxdy
Let {φ i }i = 1…N be a set of functions which satisfy the boundary condition (9.167). We seek an
approximate solution of (Equations 9.166), (9.167) in the form;
(9.169)
u(x,y)=∑aiϕi(x,y)
which minimizes the functional in (Equation 9.168), i.e.
(9.170)
∂I(∑aiϕi(x,y))∂ak=0k=1,...,N
This yields (after substitution in (Equation 9.168))
(9.171)
∑iai∫Ω[(∂ϕi∂x)(∂ϕk∂x)+(∂ϕi∂y)(∂ϕk∂y)]dxdy+c∫Ωϕkdxdy=0.
This is a system of N linear equations for the coefficients a i whose solution yields (through
Equation(9.169)) an approximate solution to (Equations 9.166), (9.167).

9.10 The Finite Element Method in 2-D

We shall discuss this method within the context of the following boundary value problem
(9.172)
∇ 2u=0inΩ
u=uDonΓD
∂ u ∂ n = ( ∇ u ) · n = u N on Γ N
where Γ D is the part of the boundary of Ω where Dirichlet boundary conditions are being specified
and Γ N is the part where Neumann boundary conditions are specified (n is the normal to the
boundary). Thus, the boundary of the domain satisfies
∂Ω=ΓD∪ ΓN=Γ.

9.10.1 Geometrical Triangulations

The first step towards the application of the Finite Element Method (FEM) is a geometrical one. We
have to triangulate the region, i.e. divide the region into triangles (not necessarily of equal size).
Remark: In the following we consider only “triangular elements;” however, in many applications
other shapes are used.
In this triangulation process one must be careful to leave no gaps between the elements and there
should NOT be any “free nodes i.e. nodes located on an element’s edges are not allowed (in this
context “nodes” is another name for the elements vertices or “grid points”). Also triangles should not
overlap with each other.
Guidelines for “good” triangulations:
1. Avoid extremely irregular mesh. Do not use obtuse triangles; i.e. the “sides“ of the elements
should be “comparable in size.”
2. If the physical problem has some sort of symmetry, subdivide the domain according to the
symmetry.
3. Use fine mesh in the vicinity of “abrupt changes” (in geometry, boundary conditions, etc.).
The motivation for these “guidelines” is intended to ensure that one obtains a well conditioned
system of linear equations for the solution at the vertices of the elements.
Topological Data: We have to number the elements and nodes (without gaps) and prepare where
1. The node number and its coordinates ( x , y ) are given.
2. The element number and its associated nodes (in the counterclockwise order) are given.
This numbering can be done using either row or column schemes (see Figure 9.6 and Figure 9.7).
However more general schemes are possible. These numbering schemes attempt to reduce the band of
the coefficient matrix for the linear system of equations one has to solve eventually.

Figure 9.6 Row numbering of the nodes

9.10.2 Linear Interpolation in 2D

Consider an element where the value of the solution u at the nodes ( x i , y i ) , i = 1, 2, 3 is given by u
1, u 2, u 3. We want to find a linear interpolating function
Figure 9.7 Column numbering of the nodes

over the element

ue(x,y)=α+βx+γy
using the values of the solution at the nodes, viz.
u1=α+βx1+γy1
u2=α+βx2+γy2
u3=α+βx3+γy3.
This is a system of linear equations which can be solved by Cramer’s method for α, β, γ. The solution
can be rewritten in the form
ue(x,y)=Σϕi(x,y)ui
where
ϕi(x,y)=12△ e(ai+bix+ciy)
and
△ e=12(b2C3-b3c2),
(9.173)
a1=x2y3-x3y2b1=y2-y3c1=x3-x2
(9.174)
a2=x3y1-x1y3b2=y3-y1c2=x1-x3
(9.175)
a3=x1y2-x2y1b3=y1-y2c3=x2-x1
▵ e is the element area up to a sign.
Remark: Observe that ϕ i ( x , y ) have the nice property that ϕ i ( x i , y i ) = 1 and ϕ i ( x j , y j ) =
0 j ≠ i. So this formulation preserves the fact that u i are the actual values of the solution at the nodes.
It follows then from this discussion that the values of the solution at the nodes, u i , can be used to
determine (by interpolation) the solution over the whole domain Ω.
9.10.3 Galerkin Formulation of FEM
We begin by reminding the reader of the following theorem:
Gauss‐Green’s Theorem: (Integration by parts in 2 - D)
(9.176)
∫Ω(∇ 2u)wdA=∫Γ∂u∂nwdΓ-∫Ω(∇ u·∇ w)dA
In particular if u is a solution of the Laplace equation then
(9.177)
∫∫Ω(∂u∂x∂w∂x+∂u∂y∂w∂y)dA=∫Γ∂u∂nwdΓ
In the Galerkin formulation we attempt to write the solution on each element as
(9.178)
ue(x,y)=∑i=13ϕie(x,y)ui
and let
Re(x,y)=∇ 2ue(x,y)
be the residual from the exact solution on each element. We then determine the u i ‘s by imposing the
orthogonality constraint
(9.179)
∑e,i∫∫Ωeϕie(x,y)Re(x,y)dA=0
If the residuals R e ( x , y ) are negligible we can assume that approximately
VariationaI Principles blacksquare387
∇ 2 u e ( x , y ) ≈ 0 and use (Equation 9.177) to rewrite (Equation 9.179) on each element as
(9.180)
∑j=13{∫∫Ωe(∂ϕi∂x∂ϕj∂x+∂ϕi∂y∂ϕj∂y)dA}uj=∫ΓNeuNϕidΓ,i=1,2,
3
where Γ N e denote the intersection, if any, of Γ N and the triangle e.
Introducing
Ki,je=∫∫Ωe(∂ϕi∂x∂ϕj∂x+∂ϕi∂y∂ϕj∂y)dA,
and
Fie=∫ΓNeuNϕidΓ,
(Equation 9.180) can be rewritten as
(9.181)
∑jKi,jeuj=Fie
Using the expression for the the functions ϕ i ( x , y ) , which was derived earlier, we find the
following matrix representation for K i , j e
(9.182)
Ke=14(△ e)2b12+c12b1b2+c1c2b1b3+c1c3b1b2+c1c2b22+c22b2
b3+c2c3b1b3+c1c3b2b3+c2c3b32+c32
Observe that the matrix K e is symmetric. These equations have to be “assembled” for all the elements
(since u i may appear in several elements) and solved for the nodal values u i .
Example 3: A Four Element Model.
Consider the following FEM model made of four equal size right angle, isosceles triangular
elements where the sides are of length 2 (see Figure 9.8)).
Column numbering is used for the nodes. Observe also the counterclockwise numbering of the
nodes in each element. The area of each triangle is 1 and the coordinates of the nodes in order are
(0, 0), ( 0 , - 2 ) , ( 2 , 0 ) , ( 2 , - 2 ) , ( 2 2 , 0 ) , ( 2 2 , - 2 ) (the element number is circled). Using
(Equations 9.173), (9.182) we obtain the following element equations:

Figure 9.8 Four element FEM model

1. 1 4 4 - 2 - 2 - 2 2 0 - 2 0 2 u 1 u 2 u 3 = F 1 1 F 2 1 F 3 1
2. 1 4 2 - 2 0 - 2 4 - 2 0 - 2 2 u 2 u 4 u 3 = F 2 2 F 4 2 F 3 2
3. 1 4 4 - 2 - 2 - 2 2 0 - 2 0 2 u 3 u 4 u 5 = F 3 3 F 4 3 F 5 3
4. 1 4 - 2 - 2 0 - 2 4 - 2 0 - 2 2 u 4 u 6 u 5 = F 4 4 F 6 4 F 3 4
Combining these equations yields:
144-2-2000-240-200-208-4-200-2-480-200-204-2000-2-24u1u2
u3u4u5u6=F11F21+F22F31+F32+F33F42+F43+F44F53+F54F64

9.11 Appendix
Green’s Theorems in Two and Three Dimensions
Green’s theorems are the analog of integration by parts” in higher dimensions. However, some
variations of these theorems are useful in other contexts.
Theorem (Green’s Theorem in two dimensions) Let D be a region in two dimensions and let ∂ D be
its boundary (traced in the positive direction), then for any two (smooth) functions P, Q (and proper
assumptions on D), ∫ ∫ D [ ∂ P ∂ x + ∂ Q ∂ y ] d A = ∫ ∂ D (Pdy—Qdx).
Corollary: (Integration by parts)
If P = Gη. A = Fη then we obtain
∫∫D[G∂η∂x+F∂η∂y]dA=-∫∫Dη[∂G∂x+∂F∂y]dA+∫∂Dη(Gdy-Fdx)
Corollary: (Second order integration by parts)
If we let Q = 0 and P = [ G ∂ η ∂ x - η ∂ G ∂ x ] in Green’s theorem we obtain
∫∫DG∂2η∂x2dA=∫∫Dη∂2G∂x2dA+∫∂D(G∂η∂x-η∂G∂x)dy.
Similar equation holds for differentiation with respect to y.
Theorem: (Green’s theorem in three dimensions)
∫ V F · g r a d η d V = - ∫ V η divFdV + ∫ ∂ V η F · n d S .
Note that the divergence theorem is obtained as a special case of this theorem with η = 1.
 CHAPTER 10

Modeling Fluid Flow

CONTENTS
10.1 Strain and Stress
10.2 Equations of Motion for Ideal Fluid
10.2.1Continuity equation
10.2.2Eulers’ equations
10.3 Navierstokes Equations
10.4 Similarity and Reynolds’ Number
10.5 Different Formulations of Navierstokes Equations
10.6 Convection and Boussinesq Approximation
10.7 Complex Variables in 2‐D Hydrodynamics
10.8 Blasius Boundary Layer Equation
10.9 Introduction to Turbulence Modeling
10.9.1Incompressible Turbulent Flow
10.9.2Modeling Eddy Viscosity
10.9.3k - ɛModel
10.9.4The Turbulent Energy Spectrum
10.10Stability of Fluid Flow
10.11Astrophysical Applications
10.11.1Derivation of the Model Equations
10.11.2Steady State Model Equations
10.11.3Physical Meaning of the Functions H(ρ),S(ρ)
10.11.4Radial Solutions for the Steady State Model
10.12Appendix A-Gauss Theorem And Its Variants
10.13Appendix B ‐Poincare Inequality And Burger’S Equation
10.14Appendix C ‐Gronwell Inequality
10.15Appendix D‐The Spectrum
10.1 Strain and Stress
By definition the relative positions of the points in a rigid body cannot change over time. For elastic
bodies, on the other hand, these relative positions can change. Strain measures the relative
deformation of the points in an elastic body when these deformations are “small.”
To analyze this concept, consider two nearby points whose coordinates in the undeformed position
are x = ( x 1 , x 2 , x 3 ) and y = ( y 1 , y 2 , y 3 ) . Let these points be displaced now to x+ d(x),
y+ d(y) (that is, the deformation is a function of the position).
The distance between the points before and after the deformation respectively is
(10.1)
r02=∑i=13(yi-xi)2
(10.2)
r12=∑i=13(yi-xi+di(y)-di(x))2
Since we are assuming that ∥ x‐y|| < < 1 ||d|| < < 1 we can write to a first order approximation
(10.3)
di(y)-di(x)≅ ∑j=13∂di∂xj(x)(yj-xj)
Expanding Equation (10.2) using Equation (10.3) and neglecting higher order terms we obtain
r12=r02+2∑i,j=13∂di∂xj(x)(yi-xi)(yj-xj)
We now define the extension of the element ||y—x|| as
e = lim y → x r 1 - r 0 r 0 = lim y → x ( r 1 - r 0 r 0 ) ( r 1 + r 0 2 r 0 ) =
= 1 2 lim y → x r 1 2 - r 0 2 r 0 2 = ∑ i , j = 1 3 ∂ d i ∂ x j ( x ) cos θ i cos θ j .
where we used the fact that
lim y → x ( r 1 + r 0 2 r 0 ) = 1 ,
and defined
cos θ i = lim y → x y i - x i r 0
are the cosine direction of the element. We therefore showed that the extension e is a quadratic form
in the direction cosines which can be rewritten as
(10.4)
e = ∑ i , j e i j cos θ i cos θ j
where
(10.5)
eij=12(∂di∂xj+∂dj∂xi).
Since e ij = e ji , this quadratic form is symmetric and e ij are called the components of the strain [the
symmetric strain matrix (e ij ) forms a second rank tensor].
Strains in elastic bodies are the result of forces acting on the boundary of each volume element.
The force per unit area is called the stress P. Observe however that in general P is not normal to the
surface element. The normal component of the stress is called the pressure (or tension), while the
tangential component is called the ((shearing stress To analyze the stress P acting on a surface
element dS with normal n we consider a tetrahedron of which three faces dS x , dS y , dS z lie in the
coordinate plane (see Figure 10.1).

Figure 10.1 Wolume element with stress P

Let τ i , i = 1, 2, 3, be the stresses on each face of the tetrahedron. By Newton’s third law the
condition of equilibrium for the volume enclosed by the tetrahedron is
(10.6)
PdS-τ1dSx-τ2dSy-τ3dSz=0
But
dS=ndS=dSxi+dSyj+dSzk,
i.e.
dSx=n1dS,dSy=n2dS,dSz=n3dS.
Therefore
(P-τ1n1-τ2n2-τ3n3)dS=0.
In component form this becomes
P 1 = τ 11 n 1 + τ 21 n 2 + τ 31 n 3
, etc.
Hence we can rewrite (Equation 10.6) as
(10.7)
Pi=∑j=13τjinj
It can be shown that (τ ij ) is a symmetric ( i . e . τ i j = τ j i ) second rank tensor which is called the
“stress tensor.”
Using the symmetry of (τ ij ) (Equation 10.7) can be written as
(10.8)
Pi=∑j=13τijnj
The basic assumption of linear elasticity is that for small deformations the strain and stress tensors
are linearly related. Thus
(10.9)
τij=∑m,nαijmnemn
where α ijmn depends on the elastic material.
Since e mn , τ ij are symmetric, it is easy to see that
(10.10)
αijmn=αijnm,αijmn=αjimn
Furthermore, it can be shown (using a thermodynamic argument) that
αijmn=αmnij
These constraints reduce the number of independent components in this tensor to 21. When the
material is isotropic (that is there is no preferred direction) the most general form of α ijmn is
(10.11)
αijmn=λδijδmn+μ(δimδjn+δinδjm)
where λ is the Lame’s constant and μ is the rigidity. Using (Equations 10.5), (10. 11) (Equation 10.9)
now becomes:
(10.12)
τij=λδij∑k∂dk∂xk+μ(∂di∂xj+∂dj∂xi)

10.2 Equations of Motion for Ideal Fluid

The equations which govern the motion of “ideal” fluid with zero viscosity (and constant
temperature) consist of one equation which is an expression of the law of “mass conservation” (the
“continuity equation”) and three equations which express the law of momentum conservation. These
are called Euler’s equations.

10.2.1 Continuity equation

To derive this equation consider a fluid flow with mass density ρ ( x , t ) . Let V be a volume
contained in the flow. The mass of the fluid in this volume V at time t is given by
(10.13)
m(t,v)=∫Vρ(x,t)dx
Hence the rate of change of the mass in V is
(10.14)
dmdt=∫V∂ρ∂t(x,t)dx
Now let ∂ V denote the boundary of V and n the unit outward normal to ∂ V. The total mass flow rate
of the fluid across ∂ V in the outward direction is
(10.15)
∫∂Vq·ndS=∫∂V(ρu)·ndS
Mass conservation principle implies, however, that a (positive) rate of change of the mass in V must
equal the rate at which the mass is crossing ∂ V in the inward direction. Therefore
 (10.16)
∫V∂ρ∂t(x,t)dV=-∫∂Vρu·ndS
To convert the right hand side of (Equation 10.16) into a volume integral we now invoke the
divergence theorem which states that for any “smooth” vector field F in V
∫∂ VF · nd S = ∫Vd i v F d V. .
Then (Equation 10.16) yields
(10.17)
∫V{∂ρ∂t+div(ρu)}dV=0
Since V is arbitrary we infer that the integrand in (Equation 10.17) is zero, i.e.
(10.18)
∂ρ∂t+div(ρu)=0
(Equation 10.18) is called the continuity equation.

10.2.2 Eulers’ equations

To derive these equations we consider a volume of fluid V and a coordinate system that is moving
with it (this is referred to as the “Lagrangian picture” for the fluid flow).
At time t the momentum of the fluid in V is
∫V(ρu)dV.
Hence by Newton’s second law and (Equation 10.7)
ddt∫V(ρui)dV=-∫∂V∑jτijnjdS.
The minus sign is due to the fact that dS = ndS where n is the outward normal to ∂ V. Applying the
divergence theorem yields
(10.19)
ddt∫V(ρui)dV=-∫V∑j∂τij∂xjdV
Since we are considering a fluid with zero viscosity, there are no shearing stresses and the stress
tensor τ ij has only a normal components, viz. the pressure p. Thus
τij=pδij.
Substituting this expression of τ in (Equation 10.19) yields (in vector form)
ddt∫V(ρu)dV=-∫V∇ pdV.
Since the volume V is arbitrary we infer therefore that
ddt(ρu)=-∇ p.
To convert this equation to one in a fixed (inertial) coordinate system (Eurelian Picture) we use the
fact that in a fixed coordinate system x = x(t), y = y(t), z = z(t) and hence for a general vector field
F(x) we have
dFdtLagrange=∂F∂t+∂F∂xdxdt+∂F∂ydydt+∂F∂zdzdt=∂F∂t+(u·∇
)F.
Therefore
(10.20)
∂(ρu)∂t+(u·∇ )(ρu)=-∇ p
When ρ = constant (i.e. the fluid is homogeneous and incompressible) (Equations 10.18), (10.20)
reduce to
(10.21)
∇ ·u=0
(10.22)
∂u∂t+(u·∇ u)=-1ρ∇ p
or in component form (where u = ( u , v , w ) )
(10.23)
∂u∂x+∂v∂y+∂w∂z=0
(10.24)
∂u∂t+u∂u∂x+v∂u∂y+w∂u∂z=-1ρ∂p∂x
(10.25)
∂v∂t+u∂v∂x+v∂v∂y+w∂v∂z=-1ρ∂p∂y
(10.26)
∂w∂t+u∂w∂x+v∂w∂y+w∂w∂z=-1ρ∂p∂z

10.3 NAVIERSTOKES EQUATIONS

The equations that govern the motion of viscous fluid consist of the continuity equation and a set of
three equations which express the law of momentum conservation. These equations are called
Navier-Stokes equations.
To derive these equations we consider again a fluid volume V. The resultant force in the ith‐
direction due to the stresses acting on this volume (see (Equation 10.8)) is
(10.27)
∫∂VPidS=∫∂V∑jτijnjdS
Hence momentum conservation yields
(10.28)
ddt∫V(ρui)dV=∫∂V(∑jτijnj)dS+∫VρFidV
where F represents (body force(s)” per unit mass (such as gravity) acting on the fluid. Using the
divergence theorem this converts to
(10.29)
∫Vd(ρui)dt-∑j∂τij∂xj-ρFi]dV
But since V is arbitrary, it follows that
(10.30)
d(ρui)dt=∑j∂τij∂xj+ρFi
In the following discussion we neglect the body forces.
To motivate the choice of the stress tensor which is appropriate for fluid motion we shall use both
formal and physical arguments. From a formal point of view we shall assume that there exists in the
fluid a linear relationship between the stress and strain
(10.31)
τij=sij+∑m,nαijmnemn
where s ij is the “static stress tensor.” This motivates us to write e ij in analogy with (10.5) as
(10.32)
eij=12(∂ui∂xj+∂uj∂xi)
Note, however, that in these equations the velocities u i replaced the displacements in Equation
(10.5). This is justified by the observation that the relation given by Equation (12.27) implicitly
assumes that the strains e ij are due to the fluid motion. We observe that in the literature there exist
several other models for the relationship between stress and strain especially in the context of
fluid turbulence.
For an isotropic fluid the tensors s ij , α ijmn are also isotropic and we must have
(10.33)
sij=-p0δij
(10.34)
αijmn=λδijδmn+
μ(δimδjn+δinδjm)
where p 0 is independent of e mn and is called the static pressure, λ is called the ((bulk viscosity” of
the fluid, while μ is called the “dynamic viscosity.” Hence ∑ m , n α i j m n e m n = ∑ m , n [ λ δ i j δ
m n + μ ( δ i m δ j n + δ i n δ j m ) ] e m n = λ δ i j ∑ m e m m + 2 μ e i j . Therefore
(10.35)
τij=-p0δij+λδij∑memm+2μeij
We now define the “dynamic pressure” p as the negative mean of the principal stresses
(10.36)
p=-13∑iτii=p0-(λ+23μ)∇ ·u
Substituting (Equation 10.36) in (Equation 10.35) we obtain
(10.37)
τij=-[p+23μ∇ ·u]δij+2μeij
Substituting this expression of τ ij in (12.26) we obtain after some algebra
(10.38)
d(ρu)dt=-∇ p+μ3∇ (∇ ·u)+μ∇ 2u
where we dropped the body force term ρF. If the fluid is incompressible, then ρ = constant and
∇ ⋅ u = 0 (continuity equation) which leads to
(10.39)
dudt=-1ρ∇ p+ν∇ 2u
where ν = μ/ρ is called the “kinematic viscosity” of the fluid. Replacing
( d u d t ) by its Eurelian equivalent, (Equation 10.39) takes the form
(10.40)
∂u∂t+(u·∇ )u=-1ρ∇ p+ν∇ 2u
Thus the motion of incompressibe viscous fluids is governed by (Equations 10.21), (10.40). They are
referred to as “Navier-Stokes equations (NSE).” In component form these equations are:
(10.41)
∂u∂x+∂v∂y+∂w∂z=0
(10.42)
∂u∂t+u∂u∂x+v∂u∂y+w∂u∂z=-1ρ∂p∂x+ν∇ 2u
(10.43)
∂v∂t+u∂v∂x+v∂u∂y+w∂v∂z=-1ρ∂p∂y+ν∇ 2v
(10.44)
∂w∂t+u∂w∂x+v∂w∂y+w∂w∂z=-1ρ∂p∂z+ν∇ 2w
In the following we consider only incompressible flows.
EXAMPLE-Poiseuille Flow:
This is a “uniform” steady state flow between two infinite parallel plates located at z = - h and
z = h which satisfies
(10.45)
u=u(z),v=0,w=0,dpdx=const
Substituting these assumptions NSE we obtain for the steady state
(10.46)
ν∂2u∂z2=1ρ∂p∂x
Integrating with respect to z leads to
(10.47)
u=z22μdpdx+Az+B,μ=νρ
Applying the no slip boundary conditions
(10.48)
u = 0 at z = ± h
yields
(10.49)
u=-12μdpdx(h2-z2)
To find p we use
(10.50)
u¯=12h∫-hhudz=-2h33μdpdx
Integrating with respect to x we have
(10.51)
p=3μu¯h2x+p0
Exercises
1. Prove Archimedes law: “When a body is immersed in a (static) fluid a force acts on it in
direction opposite to the gravity force. The magnitude of this “buoyancy force,” is equal to the weight
of the displaced fluid by the immersed body.”
Hint: The hydrostatic pressure at depth z is - ρgzk where ρ is the fluid density.
2. Show that if a flow u can be expressed in the form
u=∇ ϕ×∇ ψ
then it admits a vector potetial A = φ∇ψ so that
u=∇ ×A.
3. Fluid flow is irrotational if ∇ × u = 0. Show that under this assumption there exists a “potential
function φ so that u = ∇φ.
 4. Show that for incompressible flow in two dimensions one can define a Stream function ψ which
is defined as
u=∂ψ∂y,v=-∂ψ∂x.
5. Show that for irrotational and incompressible flow in two dimensions the function
f(z)=ψ(x,y)+iϕ(x,y),z=x+iy
is analytic. Hint: Show that f(z) satisfies Cauchy‐Reimann equations.
6. Sphere moving uniformly in a stationary viscous fluid.
Since motion is relative, we can consider the equivalent system where the sphere is fixed and the
fluid moving (in the reverse direction). Assuming that the fluid velocity is “small”, we can neglect the
convective term u ⋅ ∇u in NSE. Furthermore we neglect the gravitational force. In the steady state of
the flow NSE reduce then to
∇ ·u=0,∇ p=ν∇ 2u.
The boundary conditions are that u |R = 0 where R is the radius of the sphere.
Show that the solution for u = ( u , v , w ) and p of these equations is
u=-3w0Rr3xz(1-R2r2)
v=-3w0Rr3yz(1-R2r2)
w=-3w0Rr3z2(1-R2r2)+w0(1-R34r3-3R4r3)
p=-3νw0Rz2r2
where w 0 represents the uniform velocity of the sphere in the fluid (in the z-direction).

10.4 Similarity and Reynolds’ Number

Suppose that we wish to solve Navier-Stokes equations for a flow which involves a “representative
length” L and a “representative velocity” U. Furthermore, assume that the initial and boundary
conditions can be expressed in non-dimensional form using
(10.52)
u′ =uU,t′ =tUL,x′ =xL
Using these new variables and
(10.53)
p′ =p-p0ρU2
where p 0 is some “representative pressure” we can rewrite Navier-Stokes equations as
(10.54)
∑∂ui′∂xi=0
(10.55)
∂ui′ ∂t′ +∑juj′ ∂ui′ ∂xj=-∂p′ ∂xi+1Re∑j∂2ui′ ∂xj∂xj
where
(10.56)
Re=ρLUμ=LUν
is called “ Reynold’s number.”
Navier‐Stokes (Equations 12.32),(12.33) contain only the Reynold’s number as a parameter. The
solution u ’, p ’ depends only on x ’, t ’, Re and the dimensionless ratios which are needed to specify
the initial and boundary conditions.
This shows that all flows that satisfy the same initial and boundary conditions (when expressed in
non‐dimensional form) and have the same Reynold’s number (i.e. the ratio ρLU/μ is the same
although ρ, L, u, μ might be different) are described by the same non-dimensional solution. Such
flows are called dynamically similar. In the following we shall always use the non-dimensional form
of Navier-Stokes equations viz.
(10.57)
∇ ·u=0
(10.58)
∂u∂t+(u·∇ )u=-∇ p+1Re∇ 2u
This similarity principle is behind the use of a “wind tunnel” where appropriately scaled down
models of airplanes (and other systems) can be tested in flows which have the same Reynold’s
number to which an actual airplane will be subjected.

10.5 Different Formulations of Navierstokes Equations

The variables u, p which characterize the fluid flow are usually called “primitive variables.”
(Equations 12.35), (12.36), which are formulated in terms of these variables, constitute a system of
coupled nonlinear partial differential equations. The first of these equations, the continuity equation,
is a first order PDE which does not include the pressure. This fact can lead to instabilities when one
attempts to solve Navier-Stokes equations numerically, and several other formulations are used to
alleviate this problem.
We present here three such formulations for incompressible flow in two dimensions.
A. Navier-Stokes equations in conservative form.
The explicit form of Navier-Stokes equations in two dimensions is
(10.59)
∂u∂x+∂u∂y=0
(10.60)
∂u∂t+u∂u∂x+v∂u∂y=-∂p∂x+1Re∇ 2u
(10.61)
∂v∂t+u∂v∂x+v∂v∂y=-∂p∂y+1Re∇ 2v
Using the continuity equation we can write
u∂u∂x+v∂u∂y=∂∂x(u2)+∂∂y(uv)
u∂v∂x+v∂v∂y=∂∂x(uv)+∂∂y(v2)
Hence (Equations 12.39)–(12.40) can be rewritten in “conservative form” as:
(10.62)
∂u∂t+∂F(u)∂x+∂G(u)∂y=1Re(∂R(u)∂x+∂S(u)∂y)
where
u=1uv,F(u)=uu2+puv,G(u)=vuvv2+p
R(u)=0τxxτxy,
and
S(u)=0τxyτyy
τxx=(∂u∂x-∂v∂y),τxy=(∂u∂y+∂v∂x),τyy=(∂v∂y-∂u∂x)
B. Elliptic Pressure Equation
For steady state flows (in two dimensions) the continuity equation can be replaced by an elliptic
equation for the pressure. To derive this equation we differentiate (Equations 12.39), (12.40) with
respect to x, y respectively and add. Using the continuity equation this leads to
(10.63)
∇ 2p=2(∂u∂x∂v∂y-∂u∂y∂v∂x)
C. Vorticity‐Stream Function Formulations
The vorticity of the flow u is defined as
(10.64)
ω=∇ ×u
In two dimensions, ω = ( 0 , 0 , ω ) where
(10.65)
ω=∂v∂x-∂u∂y
Introducing the stream function ψ, which is defined as
(10.66)
u=∂ψ∂y,v=-∂ψ∂x
we find that the continuity equation is satisfied automatically and
(10.67)
∇ 2ψ=-ω
Furthermore, if we differentiate (12.39),(12.40) with respect to y, x respectively and subtract we
obtain
(10.68)
∂ω∂t+∂ψ∂y∂ω∂x-∂ψ∂x∂ω∂y=1Re∇ 2ω
(Equations 12.45)‐(12.46) give a complete description of the flow in two dimensions) in terms of the
non‐primitive variables ψ, ω.
Although this form of Navier-Stokes equations is robust numerically, it is usually more difficult to
find and apply boundary and initial conditions on the flow in this formulation than in the primitive
one.
D. Beltrami Flow
Definition: A flow for which ω × u = 0 is called a Beltrami flow. A flow for
which curl (ω × u) = 0 is called a generalized Beltrami flow.
To see the implications of these definitions we consider the vorticity equation in three dimensions
(10.69)
∂ω∂t+(u·∇ )ω-(ω·∇ )u=ν∇ 2ω
This equation is obtained by applying the curl operator to Navier-Stokes equations and using the
continuity equation to simplify the result. (Equation 12.47) can be rewritten as
(10.70)
∂ω∂t+curl(ω×u)=ν∇ 2ω
Thus for a Beltrami (or generalized Beltrami) flow the vorticity equation reduces to the heat equation.
We note that for steady two dimensional motion the condition curl (ω × u) = 0 is equivalent to
(10.71)
∂(ω,ψ)∂(x,y)=0
Therefore any functional relation in the form ω = f(ψ) is a solution of this equation.
E. Universal Solutions
There is a class of solutions to the NSE for which the viscous term vanishes i.e. the solutions are
independent of ν. These are called universal solutions. To derive these solutions we reconsider
(Equation 12.45) and observe that
∇ 2ω=-∇ ×(∇ ×ω)
(Since ∇ ⋅ ω = 0). Hence (Equation 12.47) can be rewritten as
∇ ×(∂u∂t+ω×u+ν∇ ×ω)=0.
It is obvious therefore that the condition
∇ ×(∇ ×ω)=0,
which is equivalent to
∂ω∂t+∇ ×(ω×u)=0,
leads to a universal solution. It follows then that steady Beltrami flows are universal solutions of
NSE.

10.6 Convection and Boussinesq Approximation

In many fluid problems one must consider the effects of temperature gradients on the flow. The motion
of a fluid subject to such temperature gradients is called convection.
To derive a model for these flows we shall use the “Boussinesq Approximation.”
In this approximation all the physical properties of the fluid are assumed to be independent of the
temperature except for the density. These density variations lead to a “rbBuoyancy term” in the
Navier Stokes equations.
To obtain an expression for this buoyancy term we consider a parcel of fluid which has a
temperature and density ρ respectively while the ambient temperature and density is T 0, ρ 0. Under
these conditions a buoyancy force will act on this parcel (Archimedes law)
(10.72)
F=-(ρ0-ρ)ρ0g
(F is force per unit mass).
If the density variations are small, then from the equation of state ρ = ρ(T) we obtain
(10.73)
ρρ0≅ 1-β(T-T0)
where β > 0 is the thermal expansion coefficient. Hence
(10.74)
F=-β(T-T0)g
Navier‐Stokes equations become then
(10.75)
∇ ·u=0
(10.76)
∂u∂t+(u·∇ )u=-∇ p+ν∇ 2u-β(T-T0)g
(observe that we are still using the incompressibility assumptions in these equations).
Since the temperature is part of (Equations 10.75)‐(10.76) we need an additional equation to close
this set of equations.
To do so we consider again a parcel of fluid (Lagrangian picture) and apply to it the law of energy
conservation.
Let Q(t) be the amount of heat in the parcel at time t.
(10.77)
Q=∫VcρT(x,t)dV
where c is the specific heat of the fluid. Then
(10.78)
dQdt=∫VcρdTdtdV=-∫∂Vq(x,t)·dS=-∫VdivqdV
where q is the heat flux. Using Fourier law of heat conduction q = - κ∇T this yields
(10.79)
dTdt=k∇ 2T
where k = κ c ρ is the thermodynamic conductivity. Rewriting (Equation 10.79) in the Eulerian
picture we therefore have
(10.80)
∂T∂t+(u·∇ )T=k∇ 2T
(Equations 10.75), (10.76), (10.80) are referred to as Boussinesq equations.
Boussinesq equations can be applied to situations where there is no “reference” velocity unit, e.g.
cavity flow. These flows are referred to as “natural” or “free convection.” In these cases we use ν/L
to represent the “reference velocity” when we perform the transformation to the nondimensional form
of the equations.
On the other hand, when a characteristic external velocity U exists, we refer to the flow as “forced
convection.” The nondimensional form of the equations is obtained by the transformation
(10.81)
t¯=tUL,x¯=xL,u¯=uU,T¯=T-T0△ T,p=p-p0ρU2
where T 0 is the characteristic temperature and ▵ T is “the characteristic temperature difference.”
With these scalings, Boussinesq equations for forced convection take the following form:
(10.82)
∇ ·u=0
(10.83)
∂u∂t+(u·∇ )u=-∇ p+1Re∇ 2u-1FrTk
(10.84)
∂T∂t+(u·∇ )T=1RePr∇ 2T
(where we dropped the bars over the scaled quantities).
In these equations we introduced the Prandtl and Froude numbers Pr, Fr which are defined as
(10.85)
1Fr=gβ△ TLU2,Pr=νk.
Other “nondimensional numbers” that appear frequently in the literature are the Grashof number
(10.86)
Gr=βg△ TL3ν2
and the Rayleigh number
(10.87)
Ra=βg△ TL3νk=GrPr
Observe also that for forced convection we have the relation
(10.88)
Gr=Re2·Fr
10.7 Complex Variables in 2‐D Hydrodynamics
If a 2-D flow is irrotational, i.e. ∇ × u = 0, then there exists a potential function φ so that
(10.89)
u=∂ϕ∂x,v=∂ϕ∂y
(or equivalently u = - ∂ ϕ ∂ x , v = - ∂ ϕ ∂ y ) . The continuity equation in the Navier-Stokes equation
will be satisfied if
(10.90)
∇ 2ϕ=0
We observe, however, that the continuity equation ∇ ⋅ u = 0 implies that there exists a vector
potential A so that
(10.91)
u=∇ ×A
In 2‐D A has only one component
(10.92)
A=ψk
which implies that
(10.93)
u=∂ψ∂y,v=-∂ψ∂x
Now the continuity equation is satisfied automatically, and if the flow is irrotational
(10.94)
∇ ×u=(∂v∂x-∂u∂y)k=0
then ψ must satisfy
(10.95)
∇ 2ψ=0
We stress that a stream function for incompressible flow in 2‐D always exists. A potential function on
the other hand exists only for irrotational flow.
We conclude that when the flow is two dimensional and irrotational both φ, ψ exists and each must
satisfy Laplace (Equations 10.90), (10.95). They satisfy also Cauchy‐Riemann equations, viz.
(10.96)
∂ϕ∂x=∂ψ∂y,∂ϕ∂y=-∂ψ∂x
Therefore we can define a complex analytic function W
(10.97)
W=ϕ+iψ
This allows for the use of complex variables methods for the solution of fluid flow problems under
these assumptions.
We also have
(10.98)
w=dWdz=dW/dxdz/dx=dWdx=∂ϕ∂x+i∂ψ∂x=u-iv
i.e. w represents the “conjugate velocity” and w ¯ = u + i v .
Example: Laminar parallel flow around a sphere of radius 1.
At x = ± ∞ the flow remains parallel; therefore, we shall assume that u ∞ = 1, v ∞ = 0. Furthermore,
no fluid can penetrate the sphere and hence we must have in polar coordinates that u r ( 1 , θ ) = 0 .
Since the relationship between the velocities ( u , v ) in Cartesian and ( u r , u θ ) in polar
coordinates is given by
u = u r cos θ - u θ sin θ , v = u r sin θ + u θ cos θ ,
we can satisfy these boundary conditions by a complex potential of the form
(10.99)
W=z+1z
Hence
(10.100)
dWdz=1-1z2=u-iv
which yields
(10.101)
u=1-x2-y2r4,v=-2xyr4

10.8 Blasius Boundary Layer Equation

To discuss the effects of viscosity on fluid motion near walls we consider in this section the steady
state solution of NSE over infinite plate in the x - z plane, 0 < x < ∞, - ∞ < z < ∞ subject to a uniform
flow
(10.102)
u = U i as x → - ∞ for all y , z
(see Figure 10.2). Due to the symmetry of the problem with respect to z, the governing equations for
this flow are the 2-D NSE, i.e., u = ( u ( x , y ) , v ( x , y ) , 0 ) , and
(10.103)
∂u∂x+∂v∂y=0
(10.104)
u∂u∂x+v∂u∂y=-1ρ∂p∂x+ν∇ 2u
(10.105)
v∂v∂x+v∂v∂y=-1ρ∂p∂y+ν∇ 2v
Figure 10.2 Unidirectional flow over a plate in the x - z plane

Figure 10.3 Spectrum of a turbulent flow demonstrating Kolmogorov 5 / 3 l a w

subject to the upstream boundary conditions Equation(10.102) and

(10.106)
u = v = 0 for 0 < x < ∞ at y = 0
We assume also that
lim y → ∞ p = c o n s tan t , lim y → ∞ v = 0 .
We now seek (following Prandtl) to find a solution to this system near y = 0 as the viscosity ν → 0.
From (Equation 10.103) we deduce the existence of a stream function ψ ( x , y ) so that
(10.107)
u=∂ψ∂yv=-∂ψ∂x
However, as we expect the solution to develop “a singularity” near y = 0 (i.e. we expect that there
will exist a region near y = 0 where the solution ( u , v ) changes very rapidly although ψ is very
small). To study this behavior we introduce a stretching transformation
(10.108)
x′=x,ζ=α-1y,ψ′(x′,ζ)=α-1ψ(x,y)
(where α will be determined later).
Then
(10.109)
u′(x′,ζ)=∂ψ′∂ζ=∂ψ∂y=u(x,y)
(10.110)
v′(x′,ζ)=-∂ψ′∂x′=-α-1∂ψ∂x=α-1v(x,y)
Dropping the primes from x, u (Equations 10.104), (10.105) transform into
(10.111)
u∂u∂x+v′∂u∂ζ=-1ρ∂p∂x+ν(∂2u∂x2+1α2∂2u∂ζ2)
(10.112)
α2(u∂v′∂x+v′∂v′∂ζ)=-1ρ∂p∂ζ+ν(α2∂2v′∂x2+∂2v′∂ζ2)
with the boundary conditions
(10.113)
lim ζ → ∞ lim ν → 0 u ( v , ζ ) = U , lim ζ → ∞ lim ν → 0 v ( x , ζ ) = 0
(10.114)
lim ζ → ∞ lim ν → 0 p ( x , ζ ) = p 0 = c o n s t .
In the limit α → 0, ν → 0 we infer from (Equation 10.112) that ∂ p ∂ ζ = 0 and hence from (Equation
10.114) that p = p 0 = constant. Letting α 2 = ν we now see that as ν → 0, the equations governing the
flow reduce to
(10.115)
u∂u∂x+v′∂u∂ζ=∂2u∂ζ2
(10.116)
u=∂ψ′∂ζ,v′=-∂ψ′∂x
with the boundary conditions
u(x,0)=v′(x,0)=00<x<∞
lim ζ → ∞ u ( x , ζ ) = U
0<x<∞.
Prandtl now observed that (Equation 10.115) is invariant under the transformation
(10.117)
(x,ζ,ψ′)→ (β2x,βζ,βψ′),β>0
( i . e . i f u ( x , ζ ) is a solution, then u ( β 2 x , β ζ ) is also a solution of the equation). This suggests
that we introduce a similarity variable (which remains unchanged by this transformation)
(10.118)
w=ζ(Ux)1/2=y(Uνx)1/2
and look for a solution ψ ’ in the form
(10.119)
ψ′ (x,ζ)=(Ux)1/2f(w)
i.e.
(10.120)
u=Uf′ (η),v′ =12(Ux)1/2(wf′ -f)
Substituting these expressions in (Equation 10.115) we find that f has to satisfy Blasius equation
(10.121)
f′ ′ ′ (w)+12f(w)f′ ′ (w)=1
with the boundary conditions f(0) = f ’(0) = 0 and
lim w → ∞ f ′ ( w ) = 1 .
Finally we observe that throughout our discussion we neglected the edge effects near x = 0. Thus the
solution derived above is valid only for x > > 0.

10.9 Introduction to Turbulence Modeling

Although turbulence is a rather familiar phenomenon, there is no formal definition which covers all of
its aspects. It is generally accepted however that “turbulence” refers to irregular fluid motion which
seems to have “random variations” in space-time so that a statistical treatment of the flow is justified.
This is in spite of the fact that the equations which govern the motion of the fluid are well known to be
deterministic in nature.
From a less formal point of view we note that in general “turbulence” is generated by the viscous
effects at fixed walls or by the flow of layers of fluids with different velocities over one another.
Furthermore, turbulence is said to consist of a (nonlinear) superposition of eddies. However, the size
of these eddies cannot go to zero since the smaller the eddy, the greater is the associated velocity
gradient. These eddies have kinetic energy which is determined by their vorticity. Thus a prominent
role in the description of turbulent motion is given to the “vortical structure” and the energy spectrum
of the flow.
In view of the “definition” given above it is natural to decompose the (Eulerian) velocity of the
fluid u as
(10.122)
u=u¯+u′
where u ¯ is the “ average value of the velocity” (or the “large scale component of the flow”) and u ’
are the fluctuations (or “turbulent residuals”) around this mean, i.e. u ¯ ′ = 0 . However, from a
mathematical point of view there are (at least) three ways to obtain u ¯ :
1. Space average (homogeneous turbulence)
(10.123)
u ¯ s ( t ) = lim V → ∞ 1 V ∫ V u ( x , t ) d x
2. Time average (stationary turbulence)
(10.124)
u ¯ t ( x ) = lim T → ∞ ∫ - T T u ( x , t ) d t
3. Ensemble average of N identical systems
(10.125)
u¯E(x,t)=1N∑n=1Nun(x,t).
For stationary and homogeneous turbulence the “ergodic hypothesis” asserts that these three averages
are identical:
(10.126)
u¯t=u¯s=u¯E.
Obviously actual turbulence does not satisfy the conditions needed to satisfy the ergodic hypothesis.
Moreover, in actual experiments the averaging must be carried over finite time intervals or spatial
extensions.
Whatever averaging procedure is used, we shall assume that the averaging process is linear i.e. if
f, g are two independent “flow variables” then
(10.127)
f+g¯=f¯+g¯,a¯=a(aconstant)
(10.128)
af¯=af¯.
It follows then that if
(10.129)
f=f¯+f′,g=g¯+g′
then
(10.130)
f¯¯=f¯
(10.131)
f¯=f¯+f′¯=f¯¯=+f¯=f¯+fi.e.f¯′=0
(10.132)
fg¯=(f¯+f′)(g¯+g′)¯=f¯g¯+f¯g′¯+f′g¯¯+f′g′¯=f¯g¯+f′g′¯.
Also observe that
(10.133)
∂f¯∂s=∂f∂s¯.

10.9.1 Incompressible Turbulent Flow

Substituting the decomposition
(10.134)
u=u¯+u′,p=p¯+p′
in Navier-Stokes equations and averaging with respect to time we obtain (using summation over
repeated indices)
(10.135)
∂u¯∂t+uj∂u¯∂xj=-∂p¯∂xi+1Re∇ 2u¯-∂∂xj(ui′ uj′ ¯)
(10.136)
∂u¯∂xi=0
where we have used the continuity equation to simplify some of the resulting expressions. It follows
then that in order to solve Navier‐Stokes equations for the mean flow we need to model the residual
term ∂ ∂ x j ( u i ′ u j ′ ¯ ) . In general this is referred to as the “closure problem.”
Rewriting the right hand side of (Equation 10.135) as
(10.137)
∂p¯∂xi+∂∂xj(1Re∂u¯∂xj-ui′ uj′ ¯)
we see that a natural interpretation of the term u i ′ u j ′ is as a “new” “turbulent stress tensor” in
addition to the viscous stresses. This is referred to as “Reynold’s stress tensor.” In view of the equal
footings in which the viscous and the Reynold’s tensors appear in Equation(10.135) it is natural to
assume that the effect of the new tensor is also proportional to the velocity gradients. This is called
the Boussinesq approximation.
(10.138)
ui′uj′¯=23kδij-2νT(∂u¯i∂xj+∂u¯j∂xi)
where
(10.139)
k=12ui′ui′¯
is the turbulent kinetic energy and ν T the “turbulent eddy viscosity.” In tensor notation (Equation
10.138) can be written as
(10.140)
τ¯=u′ u′ ¯=23kI-νT[gradu¯+(gradu¯)T]
From an intuitive point of view this treatment of the tensor τ ¯ i j corresponds to a physical picture
where turbulence eddies are considered as “lumps” of fluids. These lumps collide and exchange
momentum like molecules in Maxwell’s kinetic theory of gases.
In the literature there have been many attempts to generalize the relation in Equation(10.140). For
example, attempts were made to replace (Equation 10.140) by
τ¯ij=ui′ uj′ ¯=νT(0)δij+νT(1)Dij-+νT(2)Dik-D¯kj
where
D¯ij=∂u¯i∂xj+∂u¯j∂xi
and ν T ( 0 ) , ν T ( 1 ) , ν T ( 2 ) are scalars which might depend only on the invariants that can be
formed by the tensor D ¯ i j .
In principle, however, there is no reason why ν T ( i ) should be scalars rather than tensors. In fact,
models of the form
τ¯ij=(νT)ikD¯jk
or
τ¯ij=(νT)ijklD¯kl
(i.e. where ν T is a second or fourth order tensor) were also suggested. When aquation (10.138) is
used to model τ ij , the resulting Navier‐Stokes equations for the mean flow contain a new parameter ν
T which is a function of space-time. It is the objective of all (statistical) turbulent models to calculate
the correct value of the eddy viscosity ν T .
A completely different approach to model the Reynold’s stress tensor is offered by the observation
that the Navier-Stokes equations describe the behavior of a Newtonian fluid (i.e. where strain and
stress are linearly related). However, it may be argued that in turbulent flow the fluid behavior is not
Newtonian. As a result, a nonlinear stress‐strain relationship should be assumed. Another possibility
is to consider turbulent fluid as “viscoelastic” with memory effects (i.e. the state of the fluid at t
depends on its history in the time interval [ t - τ ~ , t ] , where τ ~ > 0 ).

10.9.2 Modeling Eddy Viscosity

Even though the eddy viscosity hypothesis (Equation 10.138), has some conceptual difficulties, it
proved successful in many flow simulations. Models for the computation ν T are usually classified as
zero, one, and two equations models.
2 . Zero equation models.
Models in this category stipulate an algebraic relationship to compute ν T . We mention two of
these.
‐Prandtl mixing length
This model was proposed for two dimensional boundary layers (near “walls”)
(10.141)
νT=ℓ2|∂u¯∂y|
where y is in the direction normal to the boundary and ℓ is determined experimentally.
‐Smagorinski model
(10.142)
νT=C(2D¯ijD¯ij)1/2.
This model usually is used in a “large eddy simulation” of atmospheric flow where the grid step is
large. Under these circumstances this model gives a computationally reasonable representation of the
subgrid‐scale‐turbulence.
2 . One equation model.
Based on conjectures by Prandtl and Kolmogorov, ν T is assumed to be a function of the turbulent
kinetic energy k
(10.143)
νT=cℓk1/2
where c is a constant and ℓ is the “turbulence scale” (eddy size). As we shall see in the next section
one can derive a partial differential equation for k. However, as the relation (10.143) requires the
specification of ℓ (for which no equations are given), the model is not complete. Due to this
deficiency most current applications today use two equation models. This is discussed in the next
section.

10.9.3 k - ɛModel
As we saw in the previous section, zero and one equation models require the use of an empirical
“length scale.” As this quantity depends on the geometry and the boundary conditions, it is apparent
that a second equation either for this quantity or its “equivalent” is needed for a complete
specification of ν T . Two equation models, in spite of their deficiencies, are utilized today for most
turbulence research and applications. They can be used to compute properties of turbulent flow with
no apriori assumptions on the structure of the flow.
The most prominent among these two equation models is the k - ɛ model and its variants. Here k, ɛ
represent respectively the turbulent kinetic energy and dissipation per unit mass.
(10.144)
k=12[(u1′)2+(u2′)2+(u31)3]

(10.145)
ε=νT∂ui′∂xk∂ui′∂xk¯
(summation over i , k ) and ν T is given by
(10.146)
ν T = μ k 2 / ε , c = 0.09
The derivation of the equations for the evolution of these quantities is rather involved algebraically
and requires the modeling of terms containing various double and triple averages of ∂ u i ′ ∂ x j , e.g.
∂ui′∂xk∂ui′∂xm∂uk′∂xm¯.
The resulting evolution equations contain several constants which have to be determined
experimentally (and adjusted for different applications). Overall, the resulting model should be
considered “phenomenological.” Yet the model has been successful in many applications and is
currently the “industry standard” (we do not present this derivation here).
We should mention at this juncture that there exists an ongoing research effort to model the
evolution of turbulent flow using the Reynold stress tensor (Equation 10.140). This approach lessens
to some extent the need for the various modeling approximations which are made in the k - ɛ model.
However, in this approach six coupled partial differential equations have to be solved in addition to
the Navier‐Stokes equations. Such a scheme requires heavy computational efforts. Some attempts for
reductions based on some algebraic relations were suggested in the literature.

10.9.4 The Turbulent Energy Spectrum

One of the major characteristics of turbulence is the existence of eddies in the flow. These eddies can
be viewed as a tangle of vortex elements. These vortex elements undergo “vortex stretching” which
leads to the breakup of the large eddies into smaller ones. Thus in turbulence we speak about “energy
cascade” from the large scale eddies to the smaller ones across a continuous spectrum of scales. At
the smallest scales eddies lose energy due to viscous stresses which convert energy to heat. This
suggests that for high Reynolds numbers (where we expect strong turbulence effects) the flow can be
decomposed into three “levels.” These are the mean flow, the large scale motion, or eddies, and the
small scale motion.
The large eddies determine the rate at which energy is fed to the turbulent motion. Due to their
dimension these large eddies depend strongly on the geometry and the boundary conditions of the
problem. On the other hand at the small scales (≈ large wave numbers) the character of the turbulent
motion is determined by the energy flux, and the rate of dissipation must equal to the energy supply in
this range. This led Kolmogorov to make the following conjecture:
At sufficiently high Reynolds numbers there exists a range of (high) wave numbers where
turbulence is statistically in equilibrium and uniquely determined by the dissipation ɛ and the
viscosity ν. This state of equilibrium is universal.”
To give this conjecture a more quantitative representation we note (following Kolmogorov) the
following dimensional analysis.
[E] = energy = [L 3 T -2]
[ν] = viscosity = [L 2 T -1]
[ɛ] = dissipation = [L 2 T -3]
[k] = wave number = [L -1].
It is then “easy” to see that proper dimensional relationship between E and ɛ, k can be obtained by
a formula of the form
(10.147)
E(k,t)=Aε2/3k-5/3
(where ν has been eliminated in favor of the wave number k). This is the famous “Kolmogorov 5/3
rule”, and the range of wave numbers for which it is assumed to be true is called “the inertial range.”
Since its inception, the Kolmogorov hypothesis has been the subject of spirited debate,
experimentation (which yielded conflicting data in some cases), and refinements. As (Equation
10.147) was derived by dimensional analysis, there have been many attempts to derive it from first
principles. So far, however, this remains as one of the fundamental open problems in fluid dynamics.
A typical energy spectrum of turbulence motion is presented in Fig. 8.3.

10.10 Stability of Fluid Flow

Notation: In this section indices which appear twice are summed over.
Navier‐Stokes equations for incompressible flow are
(10.148)
∂uk∂xk=0
(10.149)
∂uj∂t+uk∂uj∂xk=-∂p∂xj+1Re∇ 2uj
Let ( U j , P ) be the steady state solution of these equation on a domain Ω
(10.150)
∂Uk∂xk=0
(10.151)
Uk∂Uj∂Xk=-∂P∂xj+1Re∇ 2Uj
which satisfies the following boundary conditions on ∂ Ω,
(10.152)
Uj|∂Ω=Fj,P|∂Ω=g
(these are “generic” boundary conditions. They might be more general).
To investigate whether this steady state solution is stable we consider a small perturbation to the
flow ( u j ′ , p ′ )
(10.153)
uj′ =Uj+uj′ ,p=P+p′
Substituting in NSE and using (Equations 11.22), (10.150) we obtain
(10.154)
∂uk′∂xk=0
(10.155)
∂uj′ ∂t+Uk∂uj′ ∂xk+uk′ ∂Uj∂Xk=-∂p′ ∂xj+1Re∇ 2uj′
The boundary conditions on ( u j ′ , p ′ ) are homogeneous (viz. zero) since ( U j , P ) already satisfy
the boundary conditions (Equation 11.19).
(Equations 10.154), (10.155) form a system of linear partial differential equations for ( u j ′ , p ′ ) .
Therefore we can use the superposition principle to solve these equations in order to determine the
impact of these perturbations on the steady state solution. To this end we use separation of variables
to find “elementary solutions” of these equations in the form
 (10.156)
uj′=e-iωtϕj(x,y,z)
(10.157)
p′=e-iωtχ(x,y,z)
These yield
(10.158)
∂ϕk∂xk=0
(10.159)
-iωϕj+Uk∂ϕj∂xk+ϕk∂Uj∂xk=-∂χ∂xj+1Re∇ 2ϕj
We now specialize and consider the case of a flow in a channel [ - ∞ , ∞ ] × [ - h , h ] where (e.g.
Poiseuille flow)
(10.160)
U=(U(y),0,0)
and a perturbation in the form of (oblique) plane waves ( ≡ Tollmien −Schlicthing waves)
(10.161)
uj′=ϕj(y)ei(αx+βz-ωt)

(10.162)
pi′ =χ(y)ei(αx+βz-ωt),α,β∈ R,ω∈ C
By a proper transformation in the x - z plane we can set β = 0 (i.e. let the x-axis coincide with the
wave front). Substituting in (Equations 10.154), (10. 155) we obtain
(10.163)
iαϕ1+dϕ2dy=0
(10.164)
-iωϕ1+iUαϕ1+ϕ2dUdy=-iαχ+1Re(d2dy2-α2)ϕ1
(10.165)
-iωϕ2+iαUϕ2=-dχdy+1Re(d2dy2-α2)ϕ2
(10.166)
-iωϕ3+iαUϕ3=-1Re(d2dy2-α2)ϕ3
(Equation 10.163) can be satisfied by introducing φ so that
ϕ2=iαϕ,ϕ1=dϕdy.
Also note that (Equation 10.166) is an “independent equation.”
Substituting φ in Equations(10.164), (10.165) we obtain
(10.167)
iϕ′ (-ω+αU)-iαϕdUdy=-iαχ+1Re(d2dy2-α2)ϕ′
(10.168)
-αϕ(-ω+U)=-χ′ +1Re(d2dy2-α2)(-iαϕ)
Differentiating (Equation 10.167) with respect to y and substituting in(Equation 10.168) for χ ’ we
obtain
(10.169)
i(U-c)(ϕ′ ′ -α2ϕ)-ϕU′ ′ =1αRe(ϕ(4)-2α2ϕ′ ′ +α2ϕ)
(where c = ω/α is the “phase velocity”).
(Equation 10.169) is called “Orr-Somerfeld equation.” It is a 4th order differential equation for φ
with the boundary conditions
ϕ(-h)=ϕ(h)=ϕ′(h)=ϕ′(-h)=0.
The stability of the solution to Tollmien-Schlichting wave is controlled by the
value of ω. When Imω > 0, the perturbation “explodes” and the solution is unstable. Neutral
stability is obtained when Imω = 0.
Im ω = ψ ( α , R e )
The curve ψ ( α , R e ) = 0 in the α - Re plane separates the domain of stability from the unstable
domain. It follows that there is a critical value of the Reynold’s number where the solution U
becomes unstable. At this point waves
start to appear in the flow. THEOREM (Rayleigh):
At Re = ∞, U(y) must have a point of inflection for flow instability.
Proof When Re = ∞, the right hand side Orr‐Somerfeld equation is zero, and we have
(10.170)
ϕ′ ′ -α2ϕ=U′ ′ U-cϕ
Taking the complex‐conjugate of this equation yields
(10.171)
ϕ′ ′ -α2ϕ∗ =U′ ′ U-c∗ ϕ∗
Multiplying (Equation 10.170) by φ * and (Equation 10.171) by φ and subtracting leads to
ddy(ϕ′ ϕ∗ -ϕϕ∗ ′ )=2iIm(c)U′ ′ |U-c|2|ϕ|2
Integrating this equation on [ - h , h ] and using the boundary conditions on φ we obtain
2iIm(c)∫-hhU′ ′ |ϕ|2|U-c|2dy=0.
Since all the terms in the integrand except U ’’ are positive, it follows that U ’’ must be 0 for some y ∈
[ - h,h] .

10.11 Astrophysical Applications

Recent astronomical discoveries lead us to believe that our solar system is not unique. In fact the
reverse is true, viz. a large number of stars have planets orbiting around them and the number of
known exoplanets at the present time is over four thousand. This data leads to the hypothesis that there
is a fundamental physical process (which we do not understand fully as of yet) that leads to the
formation of planetary systems.
Many theories were put forward in the past about the origin of the solar system. Originally it was
Laplace in 1796 who put forward the hypothesis that planetary systems evolve from a family of
isolated rings that were formed from a primitive interstellar gas cloud. Actually such a system of
rings around a protostar was observed in 2014 by the Atacama Large Millimeter/submillimeter Array
in the constellation Taurus.
Currently the leading theory about the formation of planetary systems is the “nebula theory”
whereby a cloud of interstellar gas accreted under it own gravitation, to form in stages the protostar
and the planets. Many of the results related to this theory were obtained through elaborate modeling
and large scale numerical simulations. These involve, in general, thermodynamic considerations,
magnetohydrodynamics modeling, and turbulence. However, some questions about planet formation
still persist.
In this section, we present an idealized steady state hydrodynamic model which captures the
formation of ring structure in a self gravitating disk of stratified gas. Using this model we show that
matter density within the nebular cloud can exhibit oscillations whose peaks are separated by an
almost “empty space”, viz. a ring structure as hypothesized by Laplace.
The basic assumptions of this model are that the interstellar cloud can be treated as a two‐
dimensional self gravitating, incompressible, and stratified (viz. non‐constant density) gas in which
the particle velocities |u| are non-relativistic (i.e | u | ≪ c ¯ where c ¯ is the velocity of sound). The
justification for the reduction from three to two dimensions was discussed by many authors.

10.11.1 Derivation of the Model Equations

To model the time dependent non‐relativistic flow of a rotating incompressible fluid in two
dimensions ( x , y ) we use the hydrodynamic equations of inviscid and incompressible stratified fluid
(10.172)
ux+vy=0
(10.173)
ρt+uρx+vρy=0
(10.174)
ρut+ρ(uux+vuy)=-px-ρϕx+ρω2x
(10.175)
ρvt+ρ(uvx+vvy)=-py-ρϕy+ρω2y
(10.176)
∇ 2ϕ=4πGρ
where subscripts indicate differentiation with respect to the indicated variable, u = ( u , v ) is the
fluid velocity, ρ is its density, p is the pressure, φ is the gravitational field and G is the gravitational
constant. The terms ρω 2 x, ρω 2 y represent the components of the apparent centrifugal force due to
the rotation of the gas cloud with angular velocity ω.
We can nondimensionalize these equations by introducing the following scalings
(10.177)
t=Lt~U0,x=Lx~,y=Ly~,u=U0u~,v=U0v~
ρ=ρ0ρ~,p=ρ0U02p,ϕ=U02ϕ~,ω=U0Lω.
where L, U 0, ρ 0 are some characteristic length, velocity and mass density respectively that
characterize the problem at hand. Substituting these scalings in (Equations 10.172)‐(10.176) and
dropping the tildes, these equations remain unchanged (but the quantities that appear in these
equations become nondimensional) while G is replaced by G ~ = G ρ 0 L 2 U 0 2 (once again we
drop the tilde).
In view of eq. (10.172) we can introduce a stream function ψ so that
(10.178)
u=ψy,v=-ψx
Using this stream function we can rewrite (Equation 10.173) as
(10.179)
ρt+J{ρ,ψ}=0
where for any two (smooth) functions f, g
(10.180)
J{f,g}=∂f∂x∂g∂y-∂f∂y∂g∂x.
Using ψ the momentum (Equations 10.174),(10.175) become
(10.181)
ρ(ψyt+ψyψyx-ψxψyy)=-px-ρϕx+ρω2x

(10.182)
ρ(-ψxt-ψyψxx+ψxψxy)=-py-ρϕy+ρω2y
To eliminate p from these equations we differentiate (Equations 10.181), (10.182) with respect to y,
x respectively and subtract. This leads to
(10.183)
ρy(ψyt+ψyψyx-ψxψyy)+
ρ(ψyyt+ψyψyyx-ψxψyyy)-ρx(-ψxt-ψyψxx+ψxψxy)-ρ(-ψxxt-ψy
ψxxx+ψxψxxy)=-J{ϕ,ρ}+J{12ω2r2,ρ}
where r 2 = x 2 + y 2. The sum of the second and fourth terms in this equation can be rewritten as
(10.184)
ρ(∇ 2ψ)t+ρJ{∇ 2ψ,ψ}
To reduce the first and third terms in (Equation 10.183) we use (Equation 10.179). It follows that
(10.185)
ρy(ψyt+ψyψyx-ψxψyy)-
ρx(-ψxt-ψyψxx+ψxψxy)=ρyψyt+ρyψyψyx-(ρt+ρxψy)ψyy+ρxψx
t+(ψxρy-ρt)ψxx-ρxψxψxy=ρyψyt+ρxψxt-ρt∇ 2ψ+12J{(ψx)2+(
ψy)2,ρ}.
Combining the results of (Equations 10.184), (10.185), (Equation 10.183) becomes
(10.186)
ρyψyt+ρxψxt-ρt∇ 2ψ+ρ(∇ 2ψ)t+ρJ{∇ 2ψ,ψ}+12J{(ψx)2+(ψy)2
,ρ}=-J{ϕ,ρ}+J{12ω2r2,ρ}.
Thus we have reduced the original five (Equations 10.172)–(10.176) to three (Equations 10.176),
(10.179), and (10.186). Although (Equation 10.186) is rather cumbersome in general, it can be
simplified further when we consider only the steady state of the gas (a simplification for the time
dependent flow is also possible under some constraints but will not be presented here).

10.11.2 Steady State Model Equations

When we consider only steady states of the flow (Equation 10.179) implies that ψ = ψ(ρ) and
(Equation 10.186) can be rewritten as
(10.187)
ρJ{∇ 2ψ,ψ}+J{12(ψx2+ψy2),ρ}=-J{ϕ,ρ}+J{12ω2r2,ρ}
However, in view of eq. (10.179), ψ = ψ(ρ), and this fact can be used to eliminate ψ from (Equation
10.187). To this end we observe that
(10.188)
ψx=ψρρx,ψy=ψρρy,∇ 2ψ=ψρρ[ρx2+ρy2]+ψρ∇ 2ρ
Note also that for any function of F(ρ) we have J { F ( ρ ) , ρ } = 0 . This leads after some algebra to
the following relation
(10.189)
J{(ρψρ2)∇ 2ρ+12(ρψρ2)ρ[ρx2+ρy2]+ϕ-12ω2r2,ρ}=0
Hence we infer that
(10.190)
H(ρ)∇ 2ρ+12H′ (ρ)[ρx2+ρy2]+ϕ-12ω2r2=S(ρ),H′ =dH(ρ)dρ
where
(10.191)
H(ρ)=ρψρ2
and S(ρ) is some function of ρ. (Equation 10.190) can be rewritten as
(10.192)
H(ρ)1/2∇ ·(H(ρ)1/2∇ ρ)+ϕ-12ω2r2=S(ρ)
Thus the equations governing the steady state are (Equations 10.192), (10.176). H(ρ) and S(ρ) are
“parameter functions” which determine the nature of the steady state.

10.11.3 Physical Meaning of the Functions H(ρ),S(ρ)

The function H(ρ) is a parameter function which is determined by the momentum (and angular
momentum) distribution in the fluid. From a practical point of view the choice of this function
determines the structure of the steady state density distribution. The corresponding flow field can be
computed then aposteriori (that is, after solving f o r ρ ) from the following relations.
(10.193)
u=H(ρ)ρ∂ρ∂y,v=-H(ρ)ρ∂ρ∂x
The function S(ρ) that appears in (Equation 10.192) can be determined from the asymptotic values of
ρ and φ on the boundaries of the domain on which (Equations 10.176),(10.192) are solved. When
these asymptotic values are imposed or known, one can evaluate the left hand side of (Equation
10.192) on the domain boundaries and re-express it in terms of ρ only to determine S(ρ) on the
boundary of the domain. However, the resulting functional relationship of S on ρ must then hold also
within the domain itself since S does not depend on x, y directly.
For example, if we assume that on an infinite domain h(ρ) = 1, ω = 0, and the asymptotic
behaviour of ρ and φ is given by
(10.194)
lim r → ∞ ρ ( r ) = e - α r 2 , lim r → ∞ ϕ ( r ) = 4 α 2 r 2 e - α r 2
then (asymptotically), (Equation 10.192) evaluates to
(10.195)
S(ρ)=-4αe-αr2=-4αρ

10.11.4 Radial Solutions for the Steady State Model

It is natural to consider this special case using polar coordinates. Then ρ = ρ(r) and φ = φ(r). The
system consisting of (Equations 10.192)–(10.176) with H(ρ) = 1 reduces to
(10.196)
ρ′′=-ρ′r+S(ρ)-ϕ+12ω2r2
(10.197)
ϕ′ ′ =-1rϕ′ +cρ,c=4πG
To solve this system of equations we let S(ρ) = αρ, solve (Equation 10.196) for φ, and substitute the
result in (Equation 10.197). This leads to the following fourth order equation for ρ
(10.198)
ρ′′′′+2rρ′′′-(α+1r2)ρ′′+(1r3-1r)ρ′+cρ=2ω2
The general solution of this equation is
(10.199)
ρ=2ω2c+C1J0(a1r)+C2J0(b1r)+C3Y0(a1r)+C4Y0(b1r)
where J 0 and Y 0 are Bessel functions of the first and second kind of order 0 and
a1=12-2α+2α2-4c+α2,b1=12-2α-2α2-4c+α2.
Assuming no singularity at the origin we set C 3 = C 4 = 0. To assess the impact of the rotation term on
the steady state we solved this system for C 1, C 2 on a circular disk using the boundary conditions
ρ(0) = 1 and ρ(8) = 0 with c = 1, α = - 19.4. The results of these computations for different values of
ω are plotted in Fig. 10.4. In this figure we see that the separation between the density peaks becomes
more pronounced as ω increases. This might be interpreted as leading to the creation of protoplanets
around the central core.
Figure 10.4 Steady state of the interstellar gas with α = - 19.4, c = 1, and boundary conditions
ρ(0) = 1, ρ(8) = 0.1 with different values of ω

A strong dependence on ω is shown in Fig. 10.5 which has the same parameters as Fig. 10.4 except
that the boundary conditions on ρ are: ρ(0) = 0.35 and ρ(8) = 0.25. This figure clearly illustrates the
effect that rotation can have on the pattern of density fluctuations within the cloud. Furthermore, in this
figure the magnitude of the density fluctuations reverses itself as ω becomes larger viz. the higher
density peaks are placed at larger values of r (which is reminiscent of the situation in the solar
system).

Figure 10.5 Steady state of the interstellar gas with α = - 19.4, c = 1 and boundary conditions
ρ(0) =0.35, ρ(8) = 20.25 with different values of ω
10.12 Appendix A-Gauss Theorem And Its Variants
Notation: Let
x=(x1,x2,x3)=(x,y,z)
V = a three dimensional domain
∂ V = the boundary of V
n = ( n 1 , n 2 , n 3 ) the outward normal to V.
1. Basic Gauss Theorem: If φ(x) is a scalar field then
(A.1)
∫∂VϕnidS=∫V∂ϕ∂xidV
(this is basically the fundamental theorem of the calculus in 3-D).
2. Gradient form: Use the basic theorem for i = 1, 2, 3
∫∂Vϕn1dS=∫V∂ϕ∂x1dV
∫∂Vϕn2dS=∫∂ϕ∂x2dV
∫∂Vϕn3dS=∫∂ϕ∂x3dV.
Hence
(∫∂Vϕn1dS)i+(∫∂Vϕn2dS)j+(∫∂Vϕn3dS)k
=(∫V∂ϕ∂x1idV)+(∫V∂ϕ∂x2jdV)+(∫V∂ϕ∂x3kdV)
or in vector notation
(A.2)
∫∂VϕndS=∫VgradϕdV
3. Divergence form:
Let F = ( f 1 , f 2 , f 3 ) then from the basic theorem we have
∫∂Vf1n1dS=∫V∂f1∂x1dV
∫∂Vf2n2dS=∫V∂f2∂x2dV
∫∂Vf3n3dS=∫V∂f3∂x3dV.
Summing these three equations we have
(A.3)
∫∂VF·ndS=∫VdivFdV.
4. Tensor form:
Let T be a second rank tensor with vector components ( τ 1 , τ 2 , τ 3 ) then from (A.3) we have
∫∂Vτi·ndS=∫VdivτidV=
or in component form
∫∂Vτi·ndS=∫V∑j∂τij∂xjdV
or in tensor form
(A.4)
∫∂VT·ndS=∫VdivTdV.
5. Curl form:
∫∂Vf1n2dS=∫V∂f1∂x2dV
∫∂Vf2n1dS=∫V∂f2∂x1dV.
Hence
∫∂V(n1f2-n2f1)dS=∫V(∂f2∂x1-∂f1∂x2)dV.
This is the k component of
(A.5)
∫∂Vn×FdS=∫VcurlFdV.

10.13 Appendix B ‐Poincare Inequality And Burger’S Equation

Theorem (Poincare): Let u(x) be a bounded differentiable function on [0,1] with u(1) = 0. Then
(B.1)
∫01[u′ (x)]2dx≥π24∫01u2(x)dx
Proof: To prove this inequality we introduce an “auxiliary function”
(B.2)
h ( x ) = π 2 tan ( π x 2 )
This function satisfies h(0) = 0 and the differential equation
(B.3)
h′ -h2=π24
We now consider the integral
(B.4)
∫01[uh+u′]2dx≥0
(B.5)
∫01[uh+u′]2dx=∫01u2h2dx+∫01(u′)2dx+2∫01uhu′dx≥0
but
(B.6)
∫01uu′hdx=u2h201-12∫01u2h′dx=12[u2(1)h(1)-u2(0)h(0)]-12∫0
1u2h′dx=-12∫01u2h′dx
Hence from (B.5)
(B.7)
∫01(u′ )2dx≥∫01u2h′ dx-∫01u2h2dx=∫01u2(h′ -h2)dx=π24∫01u2dx
.
Remark: To obtain the same result on [ 0 , a ] we let z = ax
∫01[u′(x)]2dx=∫0aa2(dudz)21adz=a∫0a(u′(z))2dz
∫01u2(x)dx=∫0au2(z)1adz.
Hence
a∫0a(u′ (z))2dz≥π241a∫0au2dz,
i.e.
∫0a(u′ (z))2dz≥π24a2∫0au2(z)dz.
Burger’s Equation.
 Burger’s equation is:
(B.8)
ut+uux=νuxx
Observe that this is a nonlinear partial differential equation which contains a convective term uu x . It
was derived originally as a prototype, in one dimension, that can provide analytic insight about the
nature of turbulence and its modeling. Such equations with convective terms appear in applied
mathematics and theoretical physics (e.g. gas dynamics and traffic flow).
Lemma Let u = u ( x , t ) be a solution of Burger’s equation on the interval [ 0 , a ] , for t > 0. Show
that if u ( 0 , t ) = u ( a , t ) = 0 then
(B.9)
∫0au2dx≤Ce-νπ22a2t,Cisaconstant
Solution: Multiply (Equation B.8) by u and integrate over [ 0 , a ]
∫0auutdx+∫0au2uxdx=ν∫0auuxxdx.
But
∫0autudx=12ddt∫0au2dx
∫0au2uxdx=u330a
∫0auuxxdx=uux0a-∫0aux2dx.
Hence
12ddt∫02u2dx+u33|0a=νuux|0a-ν∫0aux2dx
(B.10)
12ddt∫0au2dx+ν∫0aux2dx=u(νux-u23)0a
with the boundary conditions u ( 0 , t ) = u ( a , t ) = 0 this leads to
12ddt∫0au2dx=ν∫0aux2dx≥νπ24a2∫0au2dx.
If we let
F(t)=∫0au2(x,t)dx≥0
then
-ddtF(t)≥νπ22a2F(t)
(multiply by - 1)
F′ (t)F(t)≤-νπ22a2.
Integrate with respect to t
ln F ( t ) | 0 t ≤ - ν π 2 2 a 2 t
ln F ( t ) F ( 0 ) ≤ - ν π 2 2 a 2 t
F(t)≤F(O)exp(-νπ22a2t)
Observe that if u ( O , t ) or u ( a , t ) are not equal to 0, one can still use Poincare inequality to obtain
a similar result.

10.14 Appendix C ‐Gronwell Inequality

Theorem (Gronwell): Let u: [ 0 , a ] → R be a continuous non‐negative function and let K, C be non‐
negative constants so that
(C.1)
u(t)≤C+∫0tKu(s)ds
for all t ∈ [ 0 , a ] . Then
(C.2)
u(t)≤CeKt
Proof: Define
(C.3)
U(t)=C+∫0tKu(s)ds
then it is obvious that u(t) ≤ U(t) for all t. However
(C.4)
dU(t)dt=U′ (t)=Ku(t)
Hence
(C.5)
U′ (t)U(t)=Ku(t)U(t)≤K
Integrating both sides of this inequality with respect to t we obtain
(C.6)
ln U ( t ) ≤ ln U ( 0 ) + K t = ln C + K t
Exponentiating this inequality we finally obtain
(C.7)
u(t)≤U(t)≤CeKt
As an application of this theorem we now prove the following
Theorem: Let f(x) be Lipschitz, i.e.
(C.8)
|f(x)-f(y)|≤K|x-y|,forallx,y
Then for any two solutions v(t), w(t) of
(C.9)
du(t)dt=f(u)
with initial conditions v(t 0) = v 0, w(t 0) = w 0 we have
(C.10)
|v(t)-w(t)|≤|v0-w0|eK(t-t0)
The solutions v(t), w(t) can be expressed as
(C.11)
v(t)=v0+∫t0tf(v(s))ds,w(t)=w0+∫t0tf(w(s))ds
Therefore
(C.12)
v(t)-w(t)=v0-w0+∫t0t[f(v(s))-f(w(s)]ds
Taking the absolute values on both sides of this equation we infer that
(C.13)
|v(t)-w(t)|≤|v0-w0|+∫t0t|[f(v(s))-f(w(s)]|ds
Using the fact that f(x) is Lipschitz, it follows that the function g(t) = |v(t) - w(t)| satisfies the
inequality
(C.14)
g(t)≤C+∫t0tKg(s)ds
where C = |v 0 - w 0|. All the conditions of Gronwell lemma are satisfied and therefore
g(t)≤CeK(t-t0),
i.e.
(C.15)
|v(t)-w(t)|≤|v0-w0|eK(t-t0)
which is the desired result.

10.15 Appendix D‐The Spectrum

Given a physical quantity, such as Energy, we can always rewrite it as a Fourier integral
(D.1)
E(τ)=12∫-∞∞Φ(ω)e-iωτdω
where Φ(ω) is the Fourier transform of E(τ) which is given explicitly by the “inverse transform”
(D.2)
Φ(w)=1π∫-∞∞E(τ)eiωτdτ
From (D. 1) we see that Φ(ω)dω is the contribution to the energy from harmonic oscillations in the
frequency interval [ ω , ω + d ω ] . By analogy with the spectra of light we therefore call Φ(ω) the
“spectral density function.”
Bibliography
[1] A. J. Chorin, J.E. Marsden, (1993) A Mathematical Introduction to Fluid Mechnics Springer Verlag NY
[2] M. Humi, (2006) Steady States of self gravitating incompressible fluid. J. Math. Phys. 47, 093101 (10 pages).
[3] H. Lamb, (1945) Hydrodynamics, Dover Publications NY
[4] L.D. Landau and E.M. Lifshitz, (1987) Fluid Mechanics Pergamon Press, NY
[5] Lissauer J.J, (1993) Planet formation, Ann. Rev. Astron. 31, pp. 129-174
[6] M. Ya Marov and A.V. Kolesnichenko (2013) Turbulence and Self-Organization, Modeling Astrophysical Objects, Springer, NY.
[7] L.M. Milne-Thompson, (1996) Theoretical Hydrodynamics, Dover Publications, NY
[8] C. Pozrikidis (1997) Introduction to Theoretical and Computational Fluid Dynamics, Oxford Univ. Press
 CHAPTER 11

Modeling Geophysical Phenomena

CONTENTS
11.1 Atmospheric Structure
11.2 Thermodynamics and Compressibility
11.2.1 Thermodynamic Modeling
11.2.2 Compressibility
11.3 General Circulation
11.4 Climate
11.1 ATMOSPHERIC STRUCTURE
On the large scale (in height), the atmosphere is divided into three sections. These are
1. Homosphere (up to a height of 100km),
2. Heterosphere (100 to 500km),
3. Exosphere (above 500km)
In the exosphere, the air density is very low (and the mean free path is large). As a result,
molecules in this region have a “fair chance” to escape into space. In the heterosphere, the strong
ultraviolet radiation from the sun dissociates the H 2 O and O2 molecules. By this process, part of this
harmful radiation is filtered out and does not reach the lower levels.
In the homosphere, the molecular mean free path is small. As a result, bulk transport by turbulent
air motion dominates the diffusive processes. This turbulent mixing “homogenizes” the passive
constituents of the atmosphere;
i.e., their densities decrease exponentially with altitude at the same rate which gives air a
homogeneous composition of 78% N 2 and 21% O2.
Due to its importance, the homosphere is further divided into
1. Troposphere (between 0-10km in height),
2. Stratosphere (between 10-50km),
3. Mesosphere (above 50km).
Remark: The boundary layers between these are referred to as tropopause and stratopause. In the
troposphere (which is also referred to as the biosphere), the temperature decreases at a rate of 6.5°
Kelvin/km. In the stratosphere, on the other hand, the temperature increases with height due to ozone
heating (as a result, the stratosphere is “stably stratified; “ i.e. “lighter air” is on top of the “denser”
air). In the mesosphere, the temperature again decreases with height.

11.2 Thermodynamics and Compressibility

In the previous section, we gave a qualitative overview of the atmospheric structure. In this section,
we concentrate on the troposphere and its quantitative modeling.
1. Thermodynamic Modeling

11.2.1 Thermodynamic Modeling

The “ideal gas law” states that for a parcel of gas
(11.1)
PV=mMRT
where m is the mass of the parcel, M is the molecular weight of the gas in the parcel, P is the
pressure, V is the volume, T is the temperature in Kelvins (K), and R is the (universal) “gas constant”
R = 8.314 J o u l e m o l K .
The gas density in the parcel is
 (11.2)
ρ=mV=MPRT.
Let U be the (internal) energy of a mole ideal gas. We define the heat capacity of a gas held at
constant volume as
CV=dUdT.
In the following, we assume that C is constant.
V

For a mole of an adiabatic parcel of gas (where no heat (Q) is exchanged between the parcel and
its surroundings), the first law of thermodynamics implies that
(11.3)
dQ=CVdT+PdV=0;
i.e. the sum of the change in the internal energy and the work done by the pressure is zero due to the
fact that there is no heat exchange with the environment. Using (Equation 12.1) leads to
(11.4)
CVdT+RTVdV=0.
Hence,
(11.5)
dTT=-RCVVdV.
Integrating this equation, we obtain
(11.6)
TVk-1=constant
where k = 1 + R C V . Using (Equation 12.1) we can rewrite (Equation 12.8) as
(11.7)
TP-k-1k=constant.
Taking the logarithmic derivative of (Equation 12.9) yields
(11.8)
dTT=k-1kdPP.
We assume now that the troposphere is composed of a gas which obeys the ((ideal gas law.” For an
adiabatic parcel of this gas betw TV^{{k - 1}} = een heights h and h + dh, the pressure difference
between the top and bottom is
(11.9)
dP=-ρgdh
where g is the acceleration due to gravity g = 9.8m/sec 2. Substituting for ρ from (Equation 12.2)
yields
(11.10)
dP=-gMPRTdh.
Using (Equation 12.10) to substitute for d P P in (Equation 12.4) leads to
(11.11)
dTdh=-k-1kgMR.
In the troposphere where the air is composed mostly of diatomic molecules k = 1.4, M = 28.88 which
yield
d T d h ≈ - 9.8 K / k m .
This value is greater (in absolute value) than the observed value quoted in the previous section of
- 6.5K/km. The difference is due to the fact that this ((simplistic” model neglects air moisture.
Observe that by integrating (Equation 12.11), one can compute P as a function of height using
(Equation 12.4).

11.2.2 Compressibility
In the previous chapter, we considered several aspects of Navier-Stokes equations (NSE) under the
assumption of incompressibility. However, in many applications (e.g., gas dynamics or atmospheric
applications), compressibility has to be taken into account. Boussinesq Approximation, in this
context, enables us to take into account some of these compressibility effects while retaining the
“flavor” of the incompressible equations (thereby reducing the nonlinearities in the equations). To this
end, we neglect the density variations in the continuity equation and the inertia term in the momentum
equations. However, these density variations do give rise to buoyancy forces in the momentum
equations.
The basic equations of the flow u = (u,v,w) are
∂ρ∂t+∇ ·(ρu)=0
∂u∂t+(u·∇ )u=-1ρ∇ p-gk
where p is the pressure, ρ is the density, k is a unit vector in the z direction, and g is the acceleration
due to gravity.
As a first step in this approximation, the mass continuity equation
∂ρ∂t+∇ ·(ρu)=0
is split into two equations
∇ ·u=0,∂ρ∂t+(u·∇ )ρ=0.
That is we consider the flow to be incompressible and ρ as a scalar that is carried over by the flow.
Furthermore, we assume that the density and pressure of the atmospheric fluid can be written as
ρ(x,t)=ρ0(z)+ρ′(x,t),p(x,t)=p0(z)+p′(x,t)
and
dp0(z)dz=-ρ0g
where ρ ′ ( x , t ) and p ′ ( x , t ) are small perturbations from ρ 0(z) and p(z) respectively. Introducing
the following definitions
P=pρ¯,N=-gρ-ρ0ρ¯,N02=-gρ¯dρ0(z)dz
where ρ ¯ is a constant ((reference density”, this leads to the following approximate system of
equations for the flow:
∇ ·u=0
∂u∂t+(u·∇ )u=-∇ P+Nk
∂N∂t+u·∇ N+N02w=0
where N 0 2 is called the Brunt‐Vaisala frequency. This system of equations yields a reasonable
approximation to the exact equations when the fluid is “almost incompressible.”
11.3 GENERAL CIRCULATION
The gross features of the atmospheric circulation are driven by convective currents due to differential
heating, Earth rotation, and the asymmetric distribution of land and sea. Due to these factors, the
atmosphere is divided into meridional and longitudinal convection cells.
To see how the meridional cells are formed, we observe that the equator receives much more heat
than the poles. Accordingly, we expect hot air to rise at the equator and travel in the upper
troposphere towards the poles, sink there, and then return along the surface to the equator. This global
picture was formulated by G. Hadley in the early 18th century and is named after him as “Hadley
circulation” (as a matter of fact such one-cell atmosphere exists on Venus).
On earth, the hot air rising at the equator sinks at about 30° (north and south) and thus forms the
tropical Hadley cell. At the descending branch of the Hadley cell, the air is dry. As a result, it creates
the great subtropical deserts on earth such as the Sahara and Gobi deserts.
A second meridional cell forms in the subtropics (between 30° and 60°). It is called the Ferrel
cell. Finally, there exists the polar cell in the arctic region.
The junction between the polar and Ferrel cells is called the polar front. Along this front in the
upper troposphere there is a strong band of westerly winds (i.e., winds blowing from the west) called
the jet stream. When the jet stream “meanders” south over the northern US, that region experiences
very cold weather.
We should note that due to the earth rotation, the winds in the upper troposphere of the Hadley cell
have a westerly direction. However, the returning surface wind is easterly (i.e., blowing from the
east). To understand this, observe that the tangential velocity of the earth is maximum at the equator.
For the Ferrel cell, this situation is reversed, and therefore we have “surface westerlies” in the mid-
latitudes. These are referred to as the “trade winds.”
Nonuniform heating due to the uneven distribution of land and sea drives the “zone overturning” or
Walker Circulation. In these cells, air rises at longitudes of heating (e.g., Indonesia) and sinks at
longitudes of cooling (west of South America). This circulation in normal years reinforces the
easterly trade wind across the equator. However, when this circulation reverses itself, it causes the
“El-Nino” current in the Pacific Ocean.

11.4 Climate
There exist various models for climate predictions in general and the computation of the mean
temperature of the earth. These models are usually classified by their degree of sophistication as 0-
dim, 1-dim, etc. The most sophisticated current models are the “Community Atmospheric Models”
(CAM) which were written at the National Center for Atmospheric Research (NCAR) [1].
In the 0-dimensional models, only the time dependence of the mean temperature is modeled and an
average is taken over the spatial dependence. For the one dimensional models, the dependence of the
temperature on the latitude and time is taken into consideration and so on.
In the following we provide a short narrative for Earth ‘‘climate modeling”
A key parameter in all these models is the albedo, which is the fraction of solar energy in the short
wave band which is reflected from Earth back into space. The value of the albedo depends on the
nature of the surface (ocean and different types of land, e.g., forests, deserts, etc.) and time (extent of
ice and snow cover). In the past, this “parameter” was modeled by various means, however, due to
recent advances in satellite imagery, it is now possible to compute the albedo accurately for each
location of the Earth.
In the following, we consider models with zero and one dimensions.
For the zero dimensional model, we take into consideration only the balance between the total
incoming and outgoing radiations which we denote by R and R o respectively.
(11.12)
CdTmdt=Ri-Ro
where T m is the global mean temperature, t is the time, and C is the heat capacity of the Earth system
(more precisely land, air, and oceans). To model the incoming and outgoing radiations, we let
(11.13)
Ri=Q{1-A(Tm)}
(11.14)
Ro=σg(Tm)Tm4.
Here, Q is the flux of the solar radiation, C is the heat capacity of the Earth system (more precisely
land, air, and oceans). A(T m ) is the mean albedo and g(T m ) is a “grayness-factor,” which measures
the deviation of Earth emissions from black body radiation due to the greenhouse effect.
Neglecting the spatial heat distribution on Earth, we can formulate a prototype model for the global
mean temperature (= climate). This can be done using the basic facts about the radiative balance of
Earth which were described in the previous sections. Thus, from Equations. (12.39)-(12.40) we infer
that
(11.15)
CdTmdt=Q[1-A(Tm)]-σ.g(Tm)Tm4
The factor g(T m ) was modeled by Sellers [2] by the following formula
(11.16)
g ( T m ) = 1 k tan h [ T m / T 6 ] , T 0 6 = 0.53 · 10 15 K 6
where k is the portion of Earth covered by clouds (under present conditions k ~ 0.5). In this model,
g(T m ) decreases as T m increases as the greenhouse effect becomes more pronounced.
As to the albedo, the following linear interpolation function was formulated by Sellers
(11.17)
A(T)=αM,T<T1αM-T-T1T2-T1(αM-αm),T1<T<T2αmT2<T
where a M and a m are the albedo values assigned to ice-covered and ice-free surfaces respectively (a
M = 0.85, a m = 0.25, T = 210°K, and T 2 = 275°K).
In a steady state, = 0 and (Equations 11.15) and (11.17) reduce to an algebraic equation which can
be solved for T m . We find then that this model has three equilibrium points, two of which are stable
while the third (in between) is unstable. The two stable equilibria correspond to “glaciation period”
and “present day” conditions.
Refinements to the model given by (Equations 11.15), (11.17) are obtained when we let Q depend
on a parameter A(t)
(11.18)
Q=λ(t)Q0
to take into account possible variations in the Sun radiative output with time.
We see that even this “prototype” model for Earth climate depends on many parameters whose
exact value is not known (and subject to change). This renders the predictions of this and more
sophisticated models somewhat unreliable with a large margin of error.
Many more elaborate models for the albedo have appeared in the literature. One of the major
sticking points in these models is the clouds that cover the Earth and their actual impact on the albedo.
For example, Bhattacharaya, et al. suggested a model for A(T m ) which is described by Fig. 11.1. In
this figure, the peak in the albedo near T = 220°K is attributed to the increased cloudiness near the
“ice-margin.” The use of this albedo model and (Equation 11.15) yield five steady state points, three
of which are stable.

Figure 11.1 Bhattacharaya, et al. model for the albedo

A more detailed model for the albedo takes into account the different albedos of ocean, land, and
ice and the meridional extent of the ice cover of the earth. Thus,
(11.19)
A=aLβ+(1-β)aoc,aL=a1+a2M
where β is the land and ice percentage of the earth surface and a oc is the albedo of the ocean. The
albedo of the land (a L ) is composed of two parts: a 1 is the albedo of the ice free land for which M =
0, and the albedo of the ice sheet a 1 + a 2 M where M is the meridional extent of the ice sheet.
In a more refined model for the albedo of the land, M is a function of time and is governed by the
following differential equation
(11.20)
dMdt=λM-1/2[(1+ε(T))MT-M].
Here, MT is the meridional extent of the ice accumulation zone and e(T) is a ramp function. Thus, in
this formulation the earth climate is governed by two differential Equations, (11.20) and (11.20),
which depend on several parameters. As these parameters can vary, the climate can pass through
various bifurcations.
To introduce spatial dependence in these models, we have to change dT/dt into
(11.21)
DT/Dt=∂T∂t+(u·∇ )T
where u is the wind speed. However, as the climate time-scale is long (æ O(104yrs)), it is usual in
this context to eliminate u by applying the “eddy diffusivity approximation”
(11.22)
-(u·∇ )T≅ ∇ ·(νe∇ T)
where ve is the “eddy diffusivity coefficient.” With this approximation, (Equation 11.15) takes the
form
(11.23)
C(x)∂T∂t=QS(x){1-A(x,T)}-
σg(x,T)T4+∇ ·(νe∇ T)
where S(x) is the distribution of the solar flux on earth (if the earth axis had no tilt then S(x) = jcosø,
where φ is the latitude). To simplify, to some extent, one can assume that all quantities in (Equation
11.23) depend only on time and latitude. (Equation 11.23) then takes the following form:
(11.24)
C(ϕ)∂T∂t=QS(ϕ){1-A(ϕ,T)}-
σ g ( ϕ , T ) T 4 + 1 cos ϕ ∂ ∂ ϕ { ν e ( ϕ ) cos ϕ ∂ T ∂ ϕ }
We see that due to the extreme complexity of the climate system, there remains a lot of uncertainty
about our ability to predict the future climate of the earth. In particular, it is questionable whether
current climate models can reliably predict the impact of man-made inputs to this system.
Bibliography
[1] CCSM3.0 Community Atmosphere Model (CAM), http://www.ccsm.ucar.edu/models/atm-cam
[2] K. Bhattacharaya, et al.(1982) J. Atmos. Sci. 39 p. 1747-1773.
[3] J. Pedlosky - Geophysical Fluid Dynamics 2nd edition. Springer,NY.
[4] W.D. Sellers (1969) J. App. Met. 8 p. 392-400.
[5] J. Pedlosky - Geophysical Fluid Dynamics 2nd edition. Springer NY.
 CHAPTER 12

Stochastic Modeling

CONTENTS
12.1 Introduction
12.2 Pure birth process
12.3 Kermackand mckendrick model
12.4 Queuing models
12.5 Markov chains
12.1 Introduction
In previous chapters, we considered the modeling of deterministic systems. For these systems,
information about the state of the system at time t determines with certainty its state at any later time.
For stochastic systems, on the other hand, no such certainty can be achieved, viz. the knowledge of the
state of the system at time t, enables us to predict only the probability that the system be in any of
several possible states in the future.
Our objective in this chapter is, therefore, to describe the basic logical steps that lead to such
stochastic models. To achieve this objective, we describe several stochastic models for various
growth and decay processes and compare to some extent their predictions with the corresponding
deterministic ones.

12.2 Pure birth process

Problem: A biologist is experimenting with a new hybrid tree of which there are very few plants. We
are asked to build a model which describes the population size of the trees in the next few years.
Approximations and Simplifications:
1. Since the size of the population N(t) at time t is expected to remain small, we must treat N(t)
as a discrete variable.
2. In order to build a prototype model for the tree population, we shall ignore the death
process. This is justified since trees are expected to live more than just “few years.”
3. The basic assumption in the deterministic treatment of such population models (that is when
N(t) is large and can be treated as a continuous variable) was that the change in the
population size on the time interval [ t , t + △ t ] is proportional to the size of the population
and the time interval ▵ t, i.e.,
(12.1)
N(t+△t)-N(t)=aN(t)△t
Assumptions and Abstractions:
1. Reproduction is an individual process, viz. reproduction of one tree does not affect the
reproduction of another tree in the population.
2. The probability P(t) that a tree reproduces itself in the time interval [ t , t + △ t ] is given
by
(12.2)
P( t) =k△t+p1( △t)

where k is a constant and

(12.3)
lim △ t → 0 p 1 ( △ t ) △ t = 0 i . e . p 1 ( △ t ) = O ( ( △ t ) 2 )
3. The probability p 2(t) that a tree reproduces more than once on [ t , t + △ t ] becomes
negligible, as ▵ t → 0, i.e.,
 (12.4)
lim △ t → 0 p 2 ( △ t ) △ t = 0 o r p 2 ( t ) = O ( ( △ t ) 2 ) .
4. The probabilities of reproduction of a tree on two disjoint time intervals are independent of
each other.
Remark
A process that satisfies assumptions 1—4 or their equivalents is referred to as a “Poisson process.”
Mathematical Model:
Our basic objective here is to derive differential equations for the probability P N (t) that the tree
population at time t is equal to N.To begin with, we compute the probability that the tree population
will increase from N to N + 1 on the interval [ t , t + △ t ] . To do so, we observe that the population
will increase by one on [ t , t + △ t ] if one tree reproduces itself once and all the others do not
reproduce themselves on this time interval. Hence, since there are N possibilities to choose the tree
that reproduces itself, this probability is given by
(12.5)
P(N→ N+1)=
N(k△ t+p1(t))[1-(k△ t+p1(t))-p2(t)]n-1≈Nk△ t+p(△ t)
where p( ▵ t) = O(( ▵ t)2) . From this result, we infer that if the tree population at time t is N, then the
probability that the population remains unchanged at a later time is
(12.6)
PN(t+△ t)=PN(t)·(1-k△ tN)+O((△ t)2)
Similarly, P N (t + ▵ t), viz. the probability that the tree population at t + ▵ t is (exactly) N( ≥ N 0) is
the sum of:
1. The probability that at t the tree population is N and there was no reproduction on [ t , t + △
t].
2. The probability that at t the tree population is N - 1 and there was exactly one reproduction
on [ t , t + △ t ] (remember that the probability of more than one reproduction in ▵ t is
O(( ▵ t)2).
Hence,
(12.7)
PN(t+△ t)=PN(t)(1-kN△ t)+PN-1(t)k(N-1)△ t+O((△ t)2).
If the tree population at time t = 0 is N 0, (Equation 12.6) implies that it remains unchanged at time ▵ t
if
PN0(△ t)-PN0(0)=-PN0(0)k△ tN0.
Dividing this equation by ▵ t and letting ▵ t → 0, we obtain the following differential equation for the
tree population to be the same at time t
(12.8)
dPN0(t)dt=-kN0,PN0(0)=1
where we used the fact that P N 0 ( 0 ) = 1 .
Similarly, dividing (Equation 12.7) by ▵ t and taking the limit as ▵ t → 0, we obtain the following
differential equation
(12.9)
dPN(t)dt=k[(N-1)PN-1(t)-NPN(t)],PN(0)=0,N>N0.
(Equations 12.8) and (12.9) describe the stochastic process under consideration.
Analysis of the Model:
Since N is an integer, the system (Equations 12.8) and (12.9) can be solved recursively for N = N
0, N 0 + 1,... Thus, from (Equation 12.8) we conclude that
(12.10)
PN0(t)=e-kN0t.
Substituting this result in the differential equation for N 0 + 1, we obtain
(12.11)
dPN0+1dt+k(N0+1)PN0+1=kNoe-kN0t
whose solution is
(12.12)
PN0+1=N0e-kN0t(1-e-kt).
In general, however, it is possible (after a long algebra) to show that
(12.13)
PN(t)=(N-1)!(N-N0)!(N0-1)!e-kN0t(1-e-kt)N-N0.
Other related quantities which are important in the analysis of such models are the expected value of
the population size at time t
(12.14)
μ(t)=∑N=N0∞NPN(t)
and the variance
(12.15)
σ2(t)=∑N=N0∞(N-μ(t))2PN(t)
To evaluate μ(t), we differentiate (Equation 12.14) with respect to t and use (Equations 12.8), (12.9)
to obtain
(12.16)
dμdt=∑N=N0∞NdPNdt==-kN02PN0-k∑N>N0N2PN+k∑N>N0N(N-1
)PN-1
but
(12.17)
∑N>N0N(N-1)PN-1=∑N>N0(N-1)2PN-1+
∑N>N0(N-1)PN-1=∑N=N0∞N2PN+∑N=N0∞NPN.
Hence,
(12.18)
dμdt=kμ,μ(0)=N0
which leads to
(12.19)
μ=N0ekt.
Similarly, for σ 2 we obtain the differential equation
(12.20)
dσ2dt-2kσ2=kμ,σ2(0)=0
whose solution is
(12.21)
σ2(t)=N02e2kt(1-e-kt).
From Equation (12.18) we see that the deterministic version of this stochastic model deals only with
the expected value of the population size. This can be justified for large populations since then as t
becomes large
(12.22)
σ(t)μ(t)≈12N0≈0
viz. for large N 0 the probability distribution is sharply centered around μ(t).
Exercises
1. Build and solve a stochastic model for a pure death process (e.g., radioactive decay).
2. Build a model for a population in which both birth and death occur.
3. In our discussion of the pure birth process we assumed that k is constant. Discuss what
happens if k = k(N) or k = k(t) .
4. Derive and solve (Equation 12.20).

12.3 Kermackand mckendrick model

In 1927, Kermack and McKendrick constructed a general mathematical model for the spread of
epidemics, rumors, etc. within a given population of M + 1 individuals. We shall describe here a
version of this model for the spread of rumors.
The basic starting point of this model is that at any time t the population can be divided into three
groups.
1. m(t) consists of persons who have not heard the rumor.
2. n(t) people who heard the rumor and are actively spreading it.
3. s(t) people who heard the rumor but stopped spreading it.
Remarks:
1. The corresponding version of this model for the spread of epidemics will have m(t) as those
who are susceptible to the disease but are not infected as yet, n(t) are people who are
infected and are spreading the disease, s(t) are people who are not susceptible to the
disease (due to inoculation, etc.). However, sometimes to make the model more realistic a
fourth group is added as those people who are infected but are not capable of spreading the
disease (hospitalized).
2. Since the total population is M + 1 and
m+n+s=M+1=constant,
it is obvious that a knowledge of two of these three quantities will automatically give the third.
However, to simplify our treatment further, we shall assume in the following that s(t) = 0.
Assumptions and Abstractions:
1. At time t = 0, m(O) = M, and n(O) = 1.
2. An individual in the sub-population m hears the rumor upon contact with a member of the
sub-population n.
 3. Once an individual is in the sub-population n, he remains there for all future times.
4. If I 1 and I 2 are disjoint time intervals, then the number of contacts between the sub‐
populations n and m on the interval I 1 does not affect the number of contacts on the interval
I 2.
5. The probability of exactly one contact between the sub‐populations n and m on [ t , t + △ t ]
is given by kn(t)m(t) ▵ t + O(( ▵ t)2) where k is constant.
6. The probability of more than one contact between the sub-populations n and m on [ t , t + △
t ] is O(( ▵ t)2) .
Mathematical Model:
To begin our derivation, we observe that if m(t) = R (and therefore n(t) = M + 1 - R) then the
probability of no contacts between the sub‐populations n and m on the time interval [ t , t + △ t ] is
given by (using assumption 5. above)
(12.23)
1-kR(M+1-R)△ t+O((△ t)2)
Hence, if we denote by P M (t) the probability that m(t) = M at t, then the probability that
m(t + ▵ t) = M is
(12.24)
PM(t+△ t)=PM(t)[1-kM]△ t]+O((△ t)2).
Similarly, if m(t + ▵ t) = R, R ≠ M we have three possibilities:
1. m(t) = R and there was no contact between the sub‐populations n and m on [ t , t + △ t ] .
2. m(t) = R + 1 and there was only one contact between the sub‐populations n and m on [ t , t +
△t].
3. m ( t ) = R + 2 a n d there were two contacts between n and m on [ t , t + △ t ] , etc..
However, the probability of this to happen is O(( ▵ t)2) in view of assumption 6.
Thus, we infer that:
(12.25)
PR(t+△ t)=PR(t)[1-kR(M+1-R)△ t]+
PR+1(t)k(R+1)(M-R)△ t+O((△ t)2).
Dividing Equation (12.24) and (Equation 12.25) by ▵ t and letting ▵ t → 0, we obtain the following
differential equations for P M (t) (viz. m(t) equals its initial size M at time t) and P R (t), R ≠ M
(12.26)
dPMdt=-kM,PM(0)=1
(12.27)
dPRdt=k{(R+1)(M-R)PR+1-R(M+1-R)PR},PR(0)=0,R≠M
Equations (12.26) and (12.27) can be solved recursively as in the previous section, viz.
(12.28)
PM(t)=e-kMt
PM-1(t)=MM-2e-kMt[1-e-k(M-2)t].
etc.
Exercises
1. The expected size of the population m at t is defined as
 (12.29)
μ ( t) =∑k=0Mkp( t) .

Evaluate μ(t) .
2. Evaluate P M-2(t) . Hint: use a computer algebra package.
3. A population of certain species consists of males and females. In a small colony, any male is
likely to mate with any female in any time interval of length ▵ t with probability
k ▵ t + O(( ▵ t)2) . Each such mating produces immediately one offspring which is equally
likely to be male or female. If M(t) and F(t) denote the number of males and females in the
population at time t, derive differential equations for P M,F (t).

12.4 Queuing models

Problem: A small bank has one teller. Recently, however, the manager received several complaints
about the time that customers have to wait in line for service. To determine if a second teller is
needed (or whether the current teller is slow), we are asked to formulate a model for the queue length
and waiting time in the bank.
Assumptions and Abstractions:
Let ℓ(t) be the queue length at time t, viz. the number of customers in line including the customer
that is being served. Thus, ℓ(t) = 0 if nobody is in line or being served at time t. Let P(t) be the
probability that the queue length at t is ℓ. We assume the following regarding the arrival rate of new
customers:

A1. The probability that one customer arrives at the queue in [ t , t + △ t ] is k ▵ t + O(( ▵ t)2)
where k is a constant which is referred to as the mean arrival rate.

A2. The probability that more than one customer arrives to the queue in [ t , t + △ t ] is
O(( ▵ t)2) .

A3. If I 1 and I 2 are two disjoint time intervals, then the number of customers which arrive in I 1
does not affect the number of arrivals in I 2.

Similarly, regarding the service time, we make the following assumptions:

S1. If a customer is being serviced at time t, then the probability that the service is completed in
[ t , t + △ t ] is s ▵ t + O(( ▵ t)2) where s is a constant representing the mean service time.

S2. The probability that service to more than one customer is completed in [ t , t + △ t ] is
O(( ▵ t)2) .

S3. If I 1 and I 2 are two disjoint time intervals, then the number of customers whose service is
completed in I 1 does not affect the number of customers whose service is completed in I 2.

Remark
We note that by our assumption the probability of one arrival and one completion in [ t , t + △ t ] is
 (12.30)
[k△ t+O((△ t)2)][s△ t+O((△ t)2)]=O((△ t)2)
Mathematical Model: To derive differential equations for P ℓ(t), we consider now the conditions
under which the queue length at t + ▵ t is ℓ > 0. These are:
1. The queue length at t is ℓ and there were no arrivals or departures during
[t,t+△t].
2. The queue length at t is ℓ - 1 and there was one arrival and no departures during [ t , t + △ t
].
3. The queue length at t is ℓ + 1 and there was one departure on [ t , t + △ t ] . Other possible
events are O(( ▵ t)2) by the assumptions and as noted in (Equation 12.30) Hence,
(12.31)
Pℓ(t+△ t)=Pℓ(t)(1-k△ t)(1-s△ t)+
Pℓ-1(t)(k△ t)(1-s△ t)+Pℓ+1(t)(s△ t)(1-k△ t)+O((△ t)2).
Dividing by ▵ t and letting ▵ t → 0 we obtain
(12.32)
dPℓdt=kPℓ-1+sPℓ+1-(k+s)Pℓ,ℓ>0.
Similarly, if ℓ = 0 then
(12.33)
P0(t+△ t)=P0(t)(1-k△ t)+P1(t)(s△ t)(1-k△ t)+O((△ t)2).
Hence,
(12.34)
dP0dt=-kP0+sP1.
Although this system of differential equations given by (Equations 12.32) and (12.34) can be solved
“in principle,” it is not possible to solve it recursively as we did in the previous sections. Hence, we
consider only the steady state solution for the queue, viz. the solution for P(t) when dP ℓ/dt = 0. Under
these conditions, (Equations 12.32) and (12.34) reduce to an algebraic system of equations:
(12.35)
-kP0+sP1=0
(12.36)
kPℓ-1+sPℓ+1-(k+s)Pℓ=0.
The solution of these equations is
(12.37)
Pℓ=ksℓP0.
The ratio q = k s is called the “traffic intensity” or the “utilization factor” of the queue.
Remarks:
If q > 1, then obviously the solution of (Equation 12.37) is meaningless since ∑ P ℓ = ∞
unless P 0 = 0.
If q < 1, then we must set ∑ P ℓ = 1 which yields P 0 = 1 - q and hence
Pℓ=qℓ(1-q).
For the expected queue length in the steady state, we obtain
 (12.38)
E(ℓ)=∑ℓ=0∞ℓPℓ=∑ℓqℓ(1-q)=(1-q)q∑ℓqℓ-1=(1-q)qddq∑qℓ=(1-q)q(
1-q)2=q1-q.
Exercises
1. Build a modConsider a Markov chainel for N server queue viz. a bank with one line and N
tellers. Hint: Only S1 has to be modified as follows:
S1’: If customers are being served at time t, then the probability that service to at least one
of them is completed in [ t , t + △ t ] is
2. ℓs ▵ t + O(( ▵ t)2) if ℓ < N
3. 2s ▵ t + O(( ▵ t)2) if ℓ > N.
4. Compute the steady state solution and the expected queue length for the queue in exercise 1.
5. What happens if N = ∞ in exercise 1 (infinite server queue)?

12.5 Markov chains

We start with some general definitions and background.
Consider a system which can be in any of a finite or countable number of states, and let S denote
this set of states (S is called the state space of the system). Obviously, we can identify S with
S={1,2,⋯ ,n}
or
S={1,2,⋯ ,∞}.
Furthermore, assume that the system changes its state at discrete times t i , i = 1, 2, . . . due to some
process occurring in the system. For brevity, however, we refer to these times by 0, 1, 2, … .
Definition: If the probability that the system is in the state y at time k + 1 depends only on its state at
time k, we shall then say that the system has the “Markov property” and refer to the system (or the
process which governs the system) as a “Markov chain”.
Definition
Let a system S be a Markov chain. The conditional probabilities
P ( x , y ) that the system in a state x at time k will transition to the state y at time k + 1 are called the
transition probabilities of the system.
Example 1
Birth‐Death Process.
Consider a Markov chain with state space S = { 1 , . . . , n } so that if the state of the system at time k
is x, then its state at k + 1 will be at x + 1, x - 1, and x. The transition function for this chain is given
by
P(x,y)=pxy=x+1qxy=x-1rxy=x0otherwise
where p x , q x and r x are the probabilities of birth, death, or neither birth nor death during [ t k , t k +
1 ] . Also, note that p x + q x + r x = 1. With these transition probabilities we define
(12.39)
w(x)=Px(ta<tb)
viz. the probability that a system which starts in a state x(a ≤ x ≤ b) will arrive to the state a before
arriving to the state b. (The states a, b are fixed.) We observe that w(a) = 1 and w(b) = 0. Moreover,
since a system in the state y can transit to the states y + 1, y - 1, and y with probabilities p y , q y ,
andr y , we infer that
(12.40)
{0,...,n}
Example 2
A gambler starts with m dollars and makes a series of one dollar bets against the House until either he
has n dollars (n > m) or his money runs out. Obviously, this is a Markov chain with state space { 0 , .
. . , n } . Furthermore, if p and q are the probabilities of winning and losing at each bet (p + q = 1),
then the transition probabilities of the chain are
P(x,y)=py=x+1qy=x-10otherwise
(note that r x = 0).
Hence, (Equation 12.40) implies that
(12.41)
w(y+1)-w(y)=qypy[w(y)-w(y-1)].
Introducing
(12.42)
αy=q1...qyp1..py,
we can rewrite (Equation 12.41) as
(12.43)
w(y+1)-w(y)=αyαy-1[w(y)-w(y-1)]
and by recursion it follows that
(12.44)
w(y+1)-w(y)=αyαy-1[w(y)-w(y-1)]=
αyαy-2[w(y-1)-w(y-2)]=…=αyαa[w(a+1)-w(a)].
Summing Equation(12.44) over y = a,.....,b - 1 we conclude that
(12.45)
w(a)-w(a+1)=αaA
where A = ∑ y = a b - 1 α y . Substituting (Equation 12.45) in (Equation 12.44) yields
(12.46)
w(y)-w(y+1)=αyA.
Summing (Equation 12.46) for y = x,....,b - 1 and using w(b) = 0 finally leads to
(12.47)
w(x)=Px(ta<tb)=1A∑y=xb-1αy.
Example 2 (continued): Assume that m = 100, n = 200, p = 0.4, and q = 0.6. The probability that the
gambler is ruined before having $200 is
(12.48)
W ( 0 ) = P 100 ( t 0 < t 200 ) = ∑ y = 100 199 ( 6 / 4 ) y ∑ y = 0 199 ( 6 / 4 ) y = ( 3 / 2 ) 200 - ( 3 /
2 ) 100 ( 3 / 2 ) 200 - 1 ≈ 1 1
i.e., the gambler has very little chance of winning.
Exercises
1. Consider a Markov chain with state space S = {0, 1, 2 } and transition matrix ( i . e . , a
matrix whose entries a r e t h e probabilities P ( x , y ) ) in the form
P( x,y) =010p01- p010

(thus P ( O , 1 ) = 1 , P(1, 0) = pP(1, 1) = 0, P(1, 2) = 1 - p, etc. ) Compute P n for all n and

interpret your results.
Hint: Show that P 4 = P 2
2. What strategy should the gambler in Example 2 use to maximize the chances of winning?
 CHAPTER 13

Answers to Problems

ANSWERS TO CHAPTER 1 EXERCISES

Sec. 2
All the problems in this section do not have a unique answer and in some cases they are “ ill
stated.” We present some pointers.
Ex. 1. The problem statement does not specify the initial location of the “tour around the world.” Thus
if the this location is at one of the Earth’s poles, It will take zero time to “go around the world.” On
the other hand if the tour is around the Earth equator, the answer will depend on some assumptions.
Thus if we assume that one can walk 5km an hour for 8 hours a day then since the Earth radius is
6400km then
Time in d a y s = ( 2 π · 6400 ) ( 5 · 8 ) ≈ 1005 days
Assuming a variation of 10% in the daily covered distance (i.e. the person covers 40 ± 4kms a day)
then the answer becomes 1005 ± 111 days.
Ex. 2. Use the Internet to find the size of the opening of the Mississippi to the ocean and the average
speed of the water flow in the river.
Ex. 3. For this project assume that one is required to remove only the mountain “cone” from it
surroundings (not sea level). Then find the average size of a dump truck. Also take into account that
the loading of a dump truck is not precise.
Ex. 4. One has to estimate the average volume of the human body to find the dimension of the box.
However, there are different levels of sophistication for the answer (e.g. taking into account gender
size, size of people from different cultures, etc.)
Ex. 5. The distance of Earth from the Sun is approximately 150⋅ 106 km (however the Earth orbit is
somewhat elliptical. The angular velocity of Earth is (approximately) 2π/365 radians per day.
Ex. 6. In this problem the concept of a “drop” has no exact value. Assuming that the volume of a drop
is 1cm 3, find the average depth of the Pacific ocean and it approximate size (as a portion of the Earth
surface) to estimate the answer.
Ex. 7. Assume that the average size of a bookstore is 30’ x30’ x12’ and the walls are covered
completely with books whose average thickness is 1.5 inches and height is 15 inches. (This includes
spaces between the shelves)
Ex. 8. The radius of an atom is about 10−8 cm assuming that an atom is a sphere (simplification) and a
volume of a cell is 10−3 cm 3 we can estimate the number of atoms in a cell.
Ex. 9. Use the data from ex. 4 and 8 to carry out this estimate. A more sophisticated model will use
different cell sizes for different organs in the body.
Ex. 10. Estimate that each human on average uses two light bulbs and each bulb has a life time of
1000 ± 100 hours. Each day these light bulbs is used on average 8 hours daily.
Ex. 11. Use the fact that the ‘radius’ of electron protons, and nuclei is about 10−13 cm. Assume that the
number of atoms in this volume is around 1025 to carry out this estimate.
Ex. 12. Make sure to list the different uses of the car (e.g. driving to work, leisure, etc.) and the
features that are important to you, (e.g. low maintenance, miles per gallon, etc.). Then follow each of
the model building steps that were discussed in this section.
Ex. 13. Follow essentially the same steps as in Ex. 12.
Sec. 3
Ex. 1(a). In this problem p 1 = SW and p 2 = WW. Using the same notation as in the text we have
R1=P(S|p1)P(S|p2)=0,
R2=P(S|p1)P(W|p2)+P(W|p1)P(S|p2)=12,
R3=P(W|p1)P(W|p2)=12.
Ex. 1 (b). In this problem p 1 = SW and p 2 = SS, therefore,
R1=12,R2=12,R3=0.
Ex. 2. In this example there are nine possible geno types of beans:
(SS,CC),(SW,CC),(WW,CC),(SS,Cc),(SW,Cc),(WW,Cc),
(SS,cc),(SW,cc),(WW,cc).
Therefore the reproduction function R ( p 1 , p 2 ) is a vector with nine components
R(p1,p2)=(R1,…,R9)
where
R1=P(S|p1)P(S|p2)P(C|p1)P(C|p2)
R2=P(S|p1)P(W|p2)P(C|p1)P(C|p2)+
P(W|p1)P(S|p2)P(C|p1)P(C|p2)
R3=P(W|p1)P(W|p2)P(C|p1)P(C|p2)
R4=P(S|p1)P(S|p2)P(C|p1)P(c|p2)+
P(S|p1)P(S|p2)P(c|p1)P(C|p2)
R5=P(S|p1)P(W|p2)P(C|p1)P(c|p2)+
P(W|p1)P(S|p2)P(C|p1)P(c|p2)+
P(S|p1)P(W|p2)P(c|p1)P(C|p2)+
P(W|p1)P(S|p2)P(c|p1)P(C|p2)
R6=P(W|p1)P(W|p2)P(C|p1)P(c|p2)+
P(W|p1)P(W|p2)P(c|p1)P(C|p2)
R7=P(S|p1)P(S|p2)P(c|p1)P(c|p2)
R8=P(S|p1)P(W|p2)P(c|p1)P(c|p2)+
P(W|p1)P(S|p2)P(c|p1)P(c|p2)
R9=P(W|p1)P(W|p2)P(c|p1)P(c|p2)
Ex. 3. Substitute p 1 = ( S S , C C ) and p 2 = ( W W , c c ) in the formulas of the previous exercise.
Ex. 4. The blood has six genotypes. These are AA, AO, BB, BO, AB and OO.
 Therefore the reproduction function R ( p 1 , p 2 ) is a vector with six components,
R(p1,p2)=(R1,…,R6)
where
R1=P(A|p1)P(A|p2)
R2=P(A|p1)P(O|p2)+P(O|p1)P(A|p2)
R3=P(B|p1)P(B|p2)
R4=P(B|p1)P(O|p2)+P(O|p1)P(B|p2)
R5=P(A|p1)P(B|p2)+P(B|p1)P(A|p2)
R6=P(O|p1)P(O|p2)
ANSWERS TO CHAPTER 2 EXERCISES
Sec. 2.1
Ex. 1. Assuming that the projectile range is not large we can treat approximately the Earth as “flat”.
The gravitational force is then in the y-direction, and we have to modify only the second equation in
(2.1.2) which is replaced by
y¨=-g(y),g(y)=GM/(R+y)2
where G is the gravitational constant, M is the Earth mass, and R is the Earth radius.
Ex. 2. At the maximum height v y = 0. From Equation (2.1.7) we have
v y = y ˙ = - g t + v 0 sin θ
but t = x v 0 cos θ hence v y = 0 when
g x v 0 cos θ = v 0 sin θ ,
i.e x = v 0 2 sin 2 θ 2 g .
Ex. 3. Assuming x(O) = 0 then, Equation (2.1.16) implies that c 1 = m x ˙ . The value of c 2 can be
obtained then by substituting t = 0 in Equation (2.1.17). Similarly from Equation (2.1.18) we have C 3
= y ˙ ( 0 ) + g b . c 4 is obtained by substituting t = 0 in Equation (2.1.19).
Ex. 4. The Taylor expansion of e -bt around bt = 0 is
e-bt=1-bt+(bt)22-…
Substituting the values of the constants c i , i = 1,2,3,4 in Equations (2.1.17) and (2.1.19) and
approximating e -bt by 1 - bt yields the desired results.
Ex. 5. The initial conditions in Equation (2.1.5) are modified as follows
x ˙ ( 0 ) = v 0 cos θ + w 1 , y ˙ ( 0 ) = v 0 sin θ + w 2 .
Ex. 6. Two angles.
Ex. 7. π 4
Ex. 8. Since the pilot directs the plane always towards N and the position of the plane at time t is ( x (
t ) , y ( t ) ) , it follows that the angle of u with the x‐direction satisfies tan θ = - d - y ( t ) x ( t ) .
Hence the components of u at this position are
u x = | u | cos θ , u y = | u | sin θ .
The equations of motion of the plane are
x˙=ux-v,y˙=uy.
Sec. 2
Ex. 1. m ẍ = - k x where 1 k = 1 k 1 + 1 k 2 .
Ex. 2. If the displacement of the mass m from equilibrium is x. then m ẍ = - (k 1 + k 2)x.
Ex. 3. The equivalent stiffness k of the springs in Ex. 1 satisfies
1k=1k1+1k2
(parallel resistors). For Ex. 2, k = k 1 + k2 (resistors in series).
Ex. 8. From Equation (2.2.42) we have x ( 0 ) = A cos ϕ and x ˙ ( 0 ) = - A ω sin ϕ
Ex. 9. Multiplying Equation (2.2.34) by x ˙ we obtain (F ext = 0)
m2ddt(x˙2)+k2ddt(x2)=-bx˙2
Hence
dEdt=ddt[m2x˙2+k2x2]=-bx˙2
This implies that the energy in the system decreases monotonically in time.
Ex. 15. Multiply the equation of motion by x ˙ .
Sec. 3
Ex. 1. For a closed circuit in which R 1 and R 2 are in series the potential drop on each is respectively
V 1 = R 1 * i and V 2 = R 2 * i where i is the current in the circuit. Hence
Vext=V1+V2=(R1+R2)∗ i.
It follows then that the equivalent resistance in the circuit is R = R 1 + R 2.
Similarly in a circuit where two resistors are connected in parallel to an external potential V ext we
have V ext = R 1 * i 1 and V ext = R 2 * i 2. However, the equivalent resistance in the circuit satisfies V ext
= R * i and i = i 1 + i 2. Hence
1R=1R1+1R2
Ex. 2.
e=Ldi1dt+R3∗ i1+R2∗ i3,QC1+R1∗ i2+R2∗ i3=0
i1=i2+i3,dQdt=i2.
Ex. 4.
e=R1∗ i1+Q1C1+Ldi1dt+R2∗ i2+Q2C2,
Q2C2+R2∗ i2+L2di3dt+Q3C3=0,
i1=i2+i3,Qkdt=ik,k=1,2,3.
Sec. 4
Ex. 1. Assume that the rate of birth, death and self‐cannibalization are proportional to the
population size N(t) (with proportionality constants a , b , c ) . Also assume that there is no
competition for plan food. Therefore
N(t+△ t)-N(t)=aN(t)△ t-bN(t)△ t-cN(t)△ t.
If competition for plant food exists, then additional term proportional to N(t)2 has to be added to this
equation.
Ex. 2.
dXdt=α1X-β1XY-γ1XZ,
dYdt=α2Y+β2XY-γ2YZ
dZdt=α3Z+β2XZ+γ3YZ
Ex. 3.
dXdt=α1X-β1(X+Y)2+r,dydt=α2Y-β2(X+Y)2
Ex. 4. The same number of people taken out of the S pool is added to the I pool.
Ex. 6. Substitute Equation (2.4.86) into Equation (2.4.85).
Ex. 7. Note that ∫ d x x 2 d x = - 1 x + C .
Ex 10. Overproduction of Q depresses the production and price of fuel.
Sec. 5
Ex. 1. The following approximations were made: 1. Neglect the stretch in the rod, 2. Neglect friction
in the joints (to the wall and the mass to the rod), 3. Assume no air drag, 4. “Small” angle of
vibrations. 5. Neglect the mass of the rod and spring.
Ex. 3. Equation (2.5.106) has to be modified to include the force exerted by the spring
m L θ ˙ 2 = T - m g cos θ - k ( L - L 0 )
where L is a function of time and L 0 is the length of the rod and spring without the mass attached (T
will not be present in this equation if we assume that the whole rod is replaced by the spring).
Ex. 5. When f = - μ r 2 Equation (2.5.129) becomes
d2udθ2+u=μh2.
This is an inhomogeneous equation with constant coefficients which can be solved by standard
methods.
Ex. 6. Substitute the approximations given in the text in Equations (2.5.137),(2.5.138) and keep only
the linear terms in x i , i = 1, 2, 3, 4 (Remember the x i are assumed to be small).
Ex. 7. Use simple trigonometry to express the components of e r and e θ in terms of their Cartesian
components.
Ex 8. For a one stage rocket with air friction Equation (2.5.131) is modified to,
d d t( m( t) v ( t) ) = d md t( v ( t) - u) - m( t) g( h( t) ) - α m( t) v ( t) .
If we neglect gravity this equation becomes
m(t)dvdt=-udmdt-αm(t)v(t)
changing the independent variable from time to the height of the rocket above ground and using the
fact that d h d t = v yields
dvdh=-u1mdmdh-α.
Integrating this equation with respect to h leads to
v ( h ) = - u ln ( m ( h ) ) - α h + C
where C is a constant of integration which is determined by the initial conditions.
Ex. 9. Modify Equation (2.5.131) with u = u(t) .
Ex. 11. The energy needed to overcome gravity can not be neglected for such a distance.
Ex. 13. Two different versions of Equation (2.5.131) modified to include frictional forces (see ex. 8)
have to be used.
Ex. 14. 35, 786 km above sea level.
ANSWERS TO CHAPTER 3 EXERCISES
Sec. 1
Ex. 1. If λ ≠ 0 is an eigenvalue of A then it is also an eigenvalue of A T and 1 λ is an eigenvalue of A -1.
Ex. 2. Observe that if A v = λ v then
Amv=Am-1(
Ex. 3. (a) The eigenpairs are ( 5 , ( - 1 / 2 , 1 ) T ) and ( - 1 , ( 1 , 1 ) T ) (where T indicates a vector
transpose).
(b) ( 0 , ( 2 , 1 ) T and ( 5 , ( - 1 / 2 , 1 ) T )
(c) ( 2 , ( 1 , 0 , 1 ) T ) : ( 2 , ( 1 , - 2 2 + 2 , 1 ) T ) : ( - 2 , ( 1 , - 2 2 - 2 , 1 ) T )
(d) ( 0 , ( - 5 , 4 , 1 ) T ) : ( 2 , ( 1 , 0 , 1 ) T ) : ( 2 , ( 0 , 0 , 0 ) T ) )
Sec. 2
Ex. 1(a).
dxdt=x1,dx1dt=-y1+3xy+t
dydt=y1,dy1dt=-x1-2x+3y+e2t
Initial conditions
x(1)=1,x1(1)=0,y(1)=-2,y1=0.
Sec. 3
Ex. 1.
x = - 3 25 e 5 t - e - t - 2 t 5 + 3 25
y = - 6 25 e 5 t + e - t + t 5 + 6 25 .
Ex. 2.
x = 7 30 e - 4 t + 3 5 e t + 1 6 e 2 t
Ex. 3.
x = - 2 5 e t cos t - 4 5 e t sin t - 3 5 e - 2 t
y = ( 2 5 cos t - 6 5 sin t ) e t + 3 5 e - 2 t - t .
Ex. 4.
x=34e2t-14te2t-34-t4
y=54e2t-14te2t-54-t4.
Sec. 4
Ex. 1(a). Exact solution y(1) = 4.1548, Euler’s solution y(1) = 3.465.
Ex. 1(b). Exact solution y(1) = 3.7844 Euler’s solution y(1) = - 15.32. (Solution has a singularity
around x = 0.7).
Ex. 1(c). Exact solution y(2) = - 0.27634 Euler’s solution y(2) = - 0.1509.
Ex. 1(d). Exact solution y(1) = - 0.89711 Euler’s solution y(1) = - 0.5700.
Ex. 1(e). Exact solution y(π) = 7.3891 Euler’s solution y(π) = 4.319.
Ex. 2.
(a) 3.781 (b) - 108.6(c) - 0.2024(d) - .7240(e) 5.587
Ex. 3. (a)4.155 (7th order Taylor expansion) (b) - 25.60 (order 5) (c) - .2763 (order 11) (d) - .8969
(order 16) (e) 7.389 (order 56).
Ex. 4. (a) 4.339 (b) = 3.622.
Ex. 6. (a) 8.620.
Ex. 7. (a)analyticaly u(1) = - 3.389451301, v(1) = - 2.492765599, Euler u(1) = - 2.14016,
v(1) = - 1.56672,
(b) analytically u(1) = - 1., v(1) = 0, Euler same as above.
Ex. 8. Extended Euler: 7.626026496, Euler 7.9863.
Sec. 3.5
Ex. 1. Use Taylor expansion around x i
fi-1=fi-hf′ (xi)+O(h2)
hence
f′ (xi)=fi-fi-1h+O(h)
Ex. 2. Use Taylor expansion up to order 5 (i.e. with a tail of order O(h 6)) around x i to express f i+2
etc. and then sum with the appropriate weights (as indicated by the formula in the text).
Ex. 3. Use the same strategy as indicated for Ex 2.
Ex. 4.
f i ′ ′ = - f i - 2 + 16 f i - 1 - 30 f i + 16 f i + 1 - f i + 2 12 h 2 + O ( h 4 )
Ex. 6.
fi′ ′ =fi+2-2fi+1+2fi-1-fi-22h3+O(h2)
Ex. 7.
f′(x)=h22f(x+h1)+(h12-h22)f(x)-h12f(x-h2)h2h1(h2+h1).
Sec. 3.6
Ex. 1. y(1) = 0.17141.
Ex. 2. y(1) = 0.78767.
Ex. 3. y(1) = - 0.66741.
Ex. 4. y(2) = 1.66677.
Ex. 5. y(1) = 0.
Ex. 6. (1) 0.16612 (2) 0.78330 (3)- .6669(4) 1.6644 (5) 0.
Ex. 7. (1) 0.1644 (2) 0.7819 (3)- .6667(4) 1.664 (5) 0.
Ex. 8. u = 1, v = 0.
Ex 9. modified Euler:u = 2.46345, v = .55591
Analytic: u = 2.52800, v = .55314.
Ex 10. u = 1.573312596 ⋅ 10232, v = 3.966500468 ⋅ 10115
Sec 3.7
Ex. 2. u(0.2) = 1.945386550
u ( 0.5 ) = 2.707238530
u ( 0.8 ) = 2.978086205
ANSWERS TO CHAPTER 4 EXERCISES
Sec. 1
Ex. 1. ( a ) N ( t ) = a b + C a e - a t , where C is an integration constant.
(c) N ( t ) = a - b + C a e - a t , where C is an integration constant.
Ex 2. (a) The steady state N = a b is asymptotically stable. The steady state N = 0 is unstable.
(c) The steady state N = - a b is asymptotically stable. The steady state N = 0 is unstable.
Ex. 3.
(a) The steady states are N = 0 and
N=-β±β2-4αγ2γ
The steady state in between the other two states is stable.
Sec. 3
Ex. 1(a). A steady state of the system is x = 2, y = 1. To move this steady state to the origin we make
the transformation
u=x-2,v=y-1
the new system is
dudt=(u+2)u(v-1),dvdt=(u+4)v2
Ex. 2(b). Trajectories
x=C1e-2t-C2e-4t
y=C1e-2t+C2e-4t
Integral curves
y(x)=1-2C1x±1-8C1xC1
Ex. 2(c). To find the integral curves note that
xdxdt+ydxdt+zdxdt=0
and
mdxdt+ndxdt+kdxdt=0.
Hence along an integral curve
ddt(x2+y2+z2)=0,ddt(mx+ny+kz)=0.
This mean that on an integral curve of the system
x2+y2+z2=c1,mx+ny+kz=c2.
This implies that the integral curve are the intersections of a sphere and a plane.
Ex. 3(c). The trajectories of this equation can be expressed in terms of elliptic functions. To find
the integral curves we rewrite the Equation as
d θ d t = v , d v d t = - ν 2 sin θ .
Hence
d θ d v = v - ν 2 sin θ
Therefore
2 ν 2 cos θ - v 2 = C
where C is a constant.
Sec. 4
Ex. 1. Observe that a solution of the equation is of the for e λ t where λ is a root of the polynomial p (
λ ) . Ex. 3. The point (0, 0) is a critical point of this system.
f 1= y ( 1 + cos x ) , f 2= x ( 1 + sin y )
Hence the Jacobian at (0, 0) is
(13.1)
J(0,0)=0110.
The determinant of the Jacobian is - 1 and the system is linearizable around this point. Using Taylor
expansions around (0, 0) we find that the linearized system is
dxdt=2y,dydt=x
Exs. 4,5. Rewrite these equations as a system of two first order equation.
Sec. 5
Ex. 1. Inward spiral for b > 0. Outward spiral for b < 0.
Ex. 3. Unstable node.
Ex. 5. Saddle point (one direction is stable another is unstable).
Ex. 6. Saddle point.
Sec. 6.
Ex. 1. If x > 0 then the integral is positive and F ( x , y ) > 0 . If x < 0 then
∫0xf(t)dt=-∫x0f(t)dt
and since f(t) < 0 on [‐b,0] the integral is positive.
Ex. 2. Follow similar steps as in example 4.6.4.
Ex. 3. From Ex. 1 we know that F ( x , x ) is positive definite. Now apply Liapounov theorem.
Ex. 4. First rewrite the pendulum equation as a system
x ˙ = y , y ˙ = - ω 2 sin x .
To determine the stability of (0, 0) define on [ - π / 2 , π / 2 ]
F ( x , y ) = 1 2 y 2 + ∫ 0 x sin t d t = 1 2 y 2 + ( 1 - c o s ( x ) )
and apply Liapounov theorem. This yields ∇F = gradF ⋅ G = 0.
Ex. 5. The results follow directly from Liapounov theorem.
Ex. 6. Use F ( x , y ) = x 2 + y 2 .
Sec. 7
Ex. 1. When |u| < 1 the effective frictional term is positive and therefore u is decreasing. On the other
hand if |u| > 1 the effective frictional term is negative and u is increasing. Therefore the limit cycle
u = 1 is unstable.
Ex. 2. In polar coordinates, using Equations (4.7.87) and (4.7.88) we obtain,
drdt=rh(r)dθdt=1
This implies that d r d t = 0 whenever h(r) = 0.
Ex. 3.
(a) For the limit cycle r = 1: If 1 < r < 3 then d r d t > 0 . If r < 1, d r d t is also positive. Hence
this limit cycle is (one sided) unstable.
For the limit cycle r = 3: If 3 < r < 4 then d r d t < 0 on the other hand if < 1r < 3 then, d r d t > 0 .
hence this limit cycle is asymptotically stable. (b) The limit cycle r = 2 is unstable.
Ex. 4. In polar representation we have
drdt=-r(h(r)-2)
Therefore limit cycles exist at the points where h(r) = 2.
Ex. 5. (a) Limit cycle exist only for n = 2.
(b) Limit cycle exists for r = 2k where k is an integer.
Ex. 6. When b > 0 the effective friction coefficient bx 2 is positive (for all values of x(t), and
therefore the steady state (0, 0) is asymptotically stable. When b < 0 the effective friction coefficient
is negative and (0, 0) is unstable.
ANSWERS TO CHAPTER 5 EXERCISES
Sec. 2
Ex. 1. (a) There is competition for (plant) food within each species and each species cannibalize the
other.
(b) One steady state is at (0, 0) the others are a combination of the roots of the factors e.g. x = 0
and y = a.
The linearization of the system around (0, a) is
x˙=xa(1-λ),z˙=a(z-2x)
where z = y - a. The bifurcation point is λ = 1 .
Ex. 2. The system has two steady states x = 0 and x = a + λ . The linearization around x = 0 is
x˙=x(λ+a)
and the bifurcation point is λ = - a . This is a transcritical bifurcation. The linearization around x = a
+ λ is
z˙=z(a+λ)
and the bifurcation point is λ = - a . This is a transcritical bifurcation.
Ex. 3. The linearized system around (0, 0) is
x˙=λx+y,y˙=-x+λy
The eigenvalues of this system are λ = ± i .
Ex. 4. The steady state (0, 0) is asymptotically stable for λ < 0 and unstable when λ > 0 .
Ex. 6. The equation is equivalent to the system
x˙=y,y˙=-(λ+x2)y-2x+x3
The linearization around (0, 0) is
x˙=y,y˙=-(1+λ2)y-2x
The eigenvalues of this system are ± i.
ANSWERS TO CHAPTER 6 EXERCISES
Sec. 3
Ex. 1. Writing the desired solution in the form
y=y0+εy1+ε2y2+…
We have to the zeroth order of ɛ
dy0dx+3y02=0,y0(0)=1
and to order ɛ
dy1dx+6y0y1=-y0,y1(0)=0.
Hence
y0=13x+1,y1=-12x(3x+2)(3x+1)2.
Ex. 2. Writing the desired solution in the form
u=u0+εu1+ε2u2+…
We have to the zeroth order of ɛ
d2u0dt2+u0=a
and to order ɛ
d2u1dθ2+u1=2au02.
Hence
u 0 = a + C cos ( θ + ϕ ) ,
where C, φ are constants. This solution has a period of 2π. The solution for u 1 is
(13.2)
u 1 = C 1 cos ( θ + ϕ 1 ) - a 6 ( C 2 cos ( 2 θ + 2 ϕ ) -
3 C a cos ( θ + ϕ ) - 3 C 2 - 6 a 2 - 6 C a θ sin ( θ + ϕ ) ) .
 Assuming a = 1, C = 1, C 1 = 0 (this term can be absorbed in u 0), φ = φ 1 = 0 and ɛ = 0.01 plot u
0 + ɛu 1 on the interval [ 0 , 200 π ] to estimate the deviation of the period from 2π.
Remark: Observe that this solution contains a (secular)term θ sin ( θ + ϕ ) which is unbounded as θ
grows.
Ex. 3. The equations for y 0 and y 1 are
d2y0dx2+4y0=0,y0(0)=1,dy0dx(0)=0,
d2y1dx2+4y1=y02,y1(0)=1,dy1dx(0)=0.
To order ɛ the solution is
y 0 = cos 2 t , y 1 = 11 12 cos 2 t - 1 24 cos 4 t + 1 8 .
Sec. 4
Ex. 1. Using the expansions in Equations (6.4.25) and (6.4.26) we have
(13.3)
(d2y0(s)ds2+εd2y1(s)ds2+…)(1-2εa1+…)+
ε[dyds(1-εa1+…]2+k2((y0(s)+εy1(s)+…)
To order ɛ we obtain the following two equations
d2y0ds2+k2y0=0,y0(0)=1,dy0dt(0)=0
d2y1ds2+k2y1=(dy0ds)2+2a1d2y0ds2,y1(0)=0,dy1dt(0)=0
where a 1 has to be determined so that the secular terms disappear. Hence y 0 = cos ks and
y 1 = C 1 cos k s + C 2 sin k s - a 1 cos k s - a 1 k s sin ks— 1 6 cos 2 k s - 1 2 .
We see that in order to get rid of the secular term we must have a 1 = 0. Applying the initial
conditions we finally have
y 1 = 2 3 cos ks— 1 6 cos 2 k s - 1 2 .
Sec. 5
Ex. 1. In the outer region one has to satisfy to order ɛ
(1+ax)dy0dx+ky0=0,y(1)=1
(1+ax)dy1dx+ky1=-d2y0dz2,y1(0)=0
whose solution is
y0=D1(1+ax)-k/a,y1=(D2+D1k(a+k)(2a(1+ax)2+)(1+ax)-k/a
where
D1=(1+a)k/a,D2=-D1k(a+k)2a.
Performing the stretching transformation z = ɛ b x in the inner region the equation becomes
ε1+2bd2ydz2+εb(1+aε-bz)dydz+ky=0,y(0)=0.
Hence we must choose b = - 1. The equation for the inner region becomes
d2ydz2+(1+aεz)dydz+kεy=0,y(0)=0.
If we write y = y 0 + ɛy 1 we obtain to order ɛ the following system of equations
d2y0dz2+dy0dz=0,y0(0)=0
d2y1dz2+dy1dz+azdy0dz+ky0=0,y1(0)=0.
Hence
y0=C1z,y1=C2(e-z-1)-C1(a+k)z(z-2)
To determine the constants C 1, C 2 apply Equation (6.5.52).
ANSWERS TO CHAPTER 7 EXERCISES
Sec. 1
Ex. 1.
1k∂u∂t=∂2u∂x2+r ( x,t) .
Ex. 4.
1k∂u∂t=∂∂x(A(x)∂u∂x)
Ex. 7.
1k∂u∂t=∇ 2u=∂2u∂x2+∂2u∂y2+∂2u∂z2.
Sec. 2
Ex. 1.
1c2∂2u∂t2=∂2u∂x2+g.
Ex. 2.
∂2u∂t2=ρ∂2u∂x2-ku-b∂u∂t.
Ex. 5.
1c2∂2u∂t2=∇ 2u.
Ex. 7. Let w = x - ct, z = x + ct.
Sec. 3
Ex. 3.
htt=g(ax+b)hxx+aghx.
Sec. 4
Ex. 2.
Mass ≈ LC, spring constant ≈ RG, coefficient of friction ≈ LG + RC
Sec. 5
Ex. 1. Zero inside the cavity; M/r outside the cavity.
Ex. 4. 1 r ∂ ∂ r ( r ∂ u ∂ r ) + 1 r 2 ∂ 2 u ∂ r 2 = 0 .
Ex. 7. Substitute these functions in Laplace equation.
Sec. 6
Ex. 1. u ≈ 1 ρ .
Ex. 3.
∂u∂t+∂(ρu)∂x=-a.
Ex. 5.
∂ρ1∂t+∂(ρ1u1)∂x=-a1(ρ1)+a2(ρ2)
∂ρ2∂t+∂(ρ2u2)∂x=-a2(ρ2)+a1(ρ1)
ANSWERS TO CHAPTER 8 EXERCISES
Sec. 1.1.
Ex. 1.
400 π ∑ n = 1 , 3 , 5 ∞ , … 1 n e x p ( - k n 2 π 2 t L 2 ) sin n π x L
Ex. 3.
π sin 30 ∑ n = 1 ∞ ( - 1 ) n n 900 - n 2 π 2 e x p ( - n 2 π 2 t 9 · 10 6 R C ) sin ( n π x 3000 )
Ex. 6.
3200 π 3 ∑ n = 1 , 3 , 5 ∞ , … 1 n 3 e x p ( 1 - k n 2 π 2 400 ) t sin ( n π x 20 )
Ex. 8.
1 1250 π 2 ∑ n = 1 , 3 , 5 ∞ , … 1 n 2 s i n ( n π x 2 ) sin 5000 n π t
Ex. 9.
80 L π c ∑ n = 1 , 3 , 5 ∞ , … 1 n sin ( n π x 2 L ) sin ( n π c t 2 L )
Ex. 12.
4 a π 2 ∑ n = 1 , 3 , 5 ∞ , … ( - 1 ) ( n - 1 ) / 2 n 2 sin ( 0.01 ) n π sin n π x sin a n π t
Sec 1.2
Ex. 1. 1 3 ∑ n = 1 ∞ ( - 1 ) ( n + 1 ) n ( ρ c ) 6 n sin 6 n θ
Ex. 3. 40 π ∑ n = 1 , 3 , 5 ∞ … ( r 20 ) n sin n θ n
Ex. 12. 200 a 2 π 4 ∑ n = 1 ∞ 1 + n 2 π 2 n 4 [ ∫ 0 100 σ h ( σ ) sin n π σ 100 d σ ] sin ( n π ρ / 100 ) ρ
sin a n π t 100
Ex. 13. 1 ρ 2 ( ρ 2 p ρ ) ρ = 1 a 2 p t t , p ( a , t ) = | 1 , ∂ p ∂ ρ ( b , t ) = 0 , p ( ρ , 0 ) = 0 , ∂ p ∂ t ( ρ ,
0)=ρ
Sec 1.3
Ex. 1. ∑ n = 1 , 3 , 5 ∞ … sin n π x L [ 40 n π cos a n π t L - 20 L a n 2 π 2 sin a n π t L ]
Ex. 3.
L 2 6 + 2 t + 2 L 2 π 2 ∑ n = 1 ∞ [ ( - 1 ) n ( 1 L ) - 1 ] cos n π x L cos a n π t L
Ex. 6.
2 - 0.0019 x + 1 5 π ∑ n = 1 ∞ ( - 1 ) n - 20 n sin n π x 100 e x p ( - n 2 π 2 t 10 6 R C )
Ex. 8.
(13.4)
Φ(x)=x2a2[L2-x23],ztt=a2zxx,z(O,t)=0,
∂z∂x(L,t)=0,z(x,0)=f(x)-Φ(x),∂x∂t(x,0)=0
Ex. 10.
(13.5)
Φ ( x ) = sin x a 2 - x ( 1 + cos L a 2 ) , w t t = a 2 w x x ,
w(O,t)=0,wx(L,t)=0,w(x,O)=-Φ(x),wt(x,0)=f(x)
Sec. 4.2
Ex. 1. Use the analog of Equation (8.4.16).
Ex. 2. To remove the singularity at r = 0 multiply the equation by r 2. Avoid r = 0 by using a circle of
radius ɛ ≪ 1 around this point.
Ex. 3. A particular solution of the equation is
r 3 ( 1 18 + cos 2 θ 10 )
Sec. 4.3
Ex. 3. Use the same finite difference formulas used in this section.
Ex. 5. To obtain a finite difference formula for the mixed derivative at the grid point ( i , j ) use the
following strategy
(13.6)
(∂2u∂t∂x)i,j=∂∂t(∂u∂x)i,j=∂∂t[(ui+1,j-ui-1,j)2h]=
12h[ ∂∂tui +1,j - ∂∂tui - 1,j ] =14h2( ui +1,j +1- ui +1,j - 1+ui - 1,j +1- ui - 1
,j-1)
ANSWERS TO CHAPTER 9 EXERCISES
Sec. 2
Ex. 2. First observe that we can choose to maximize the square of the distance from the origin rather
that the distance itself since when the square of the distance is maximum the distance is also at a
maximum.
The problem is then to maximize z = x 2 + y 2 subject to the constraint w = x 2 - 2xy + 2y 2 = 1. To
this end we consider the function
f(x,y,λ)=x2+y2-λ(x2-2xy+2y2)
Hence for extremum we must have
2x-2λx++2λy=0,2y+2λx+4λy=0,x2-2xy+2y2=1
which can solved for x, y (and λ ).
Ex. 3. The surface area and volume of the cylinder are respectively
A=2πr2+2πrh,V=πr2h
where r is the radius of the bottom and h the height. since A is constant we can solve for h and
substitute in the expression for V. The problem is then to find the max of a one variable function.
Otherwise one can use Lagrange multipliers by forming the function
f(r,h,λ)=πr2h-λ(2πr2+2πrh)
Ex. 4. Form the function
f(x,y,z,λ,μ)=xz+yz-λ(x2+y2)-μyz.
Ex. 5(b). The perimeter and area of the rectangle are P = 2(a + b), A = ab. Hence b = P/2 - a and
therefore
A=a(P/2-a)
The max of this function is obtained when a = P/4 and therefore a = b.
Sec. 3
Ex. 1. According to Fermat principle we want to minimize the time it takes light between x 1 and x 2.
This time is given by the integral
T=∫t1t2dt.
However d t = d s v ( x ) where ds is the distance traveled during the time interval dt and v(x) is the
light speed at the point. Hence we want to minimize
T=∫t1t2dt=∫x1x2dsv(x).
If we define the index of refraction in the medium as n ( x ) = c v ( x ) (where c is the velocity of light
in the vacuum) then
T=1c∫x1x2n(x)ds.
Sec. 4
Ex. 2. Using Equation (9.4.5) it follows that (since f is independent of x = ( x , y , z ) that
X˙=C=(C1,C2,C3)
Hence
x = ( C 1 x + D 1 , C 2 y + D 2 , C 3 z + D 3 ) ,where C, D are constants of integration.
Ex. 3. The differential equation for the geodesics is
d2xλdt2+Γμνλdxμdtdxνdt=0,
where Γ μ ν λ are the Christoffel symbols of the metric tensor g ij . (For the definition of the
Christoffel symbols consult a book on differential geometry.)
Sec. 5
Ex 1. The potential V(w) has to be modified to include (in the integral) an additional term - ρgw.
Exs. 2 and 3. Solutions are given in the text.
Sec. 6
Ex. 2. Following the example in the text (Equations. (9.6.4) - (9.6.6) we consider small variations
near the optimal solution ψ
ψ¯(x)=ψ(x)+εη(x)
Substituting this in the expressions for I and J yields,
I(ε)=∫x1x2[(ψ+εη)2(ψx+εηx)+(ψx+εηx)2]dx,
J(ε)=∫x1x2(ψ+εη)2dx=1.
Using Lagrange multipliers we generate the function
K(ε)=I(ε)+λJ(∈ )
The conditions for extremum become
∂K(ε,λ)∂εε=0=0,J(ε=0)=1.
This leads to
∫x1x2((2ψψx+2λψ)η(x)+(ψ2+2ψx)ηx))dx=0,∫x1x2ψ2dx=1.
Now use integration by parts to convert the integral containing η x to one over η to obtain a differential
equation for ψ. The constraint J is a normalization for ψ.
Sec. 7
Ex. 2 With drag Equation (9.7.6) is replaced by
x˙(τ)=-gτ-αx(τ)+F(τ)
where
∫0τu(s)ds.
The solution of the equation for x is
x(t)=Ce-αt-gt22+e-αt∫0tF(τ)e-ατdτ.
Then follow the steps in the text.
Ex. 3 This requirement replaces the constraint given by Equation (9.7.2) by requiring that
J(T)=∫0Tf(t)dt
is minimum, where f(t) is the fuel consumed at time t.
ANSWERS TO CHAPTER 10 EXERCISES
Sec. 3.
Ex. 1. Let V be the volume occupied a body of arbitrary shape. The total force acting on the body due
to the pressure at at each point on its surface S is
F=-∫SPndS
where n is the normal to the surface (The minus sign is due to the fact that n is in the outward
direction from the surface). However using Gauss Theorem Using (Equation A.2)in the appendix of
this chapter we have
 F = ∫ V gradPdV = ρgVk, since gradP = ( 0 , 0 , - ρ g ) (remember that gradP is in the direction in
which the pressure is increasing).
Ex. 2. Use the vector identity
∇ ×(ϕ∇ ψ)=ϕ∇ ×(∇ ψ)+∇ ϕ×∇ ψ.
Ex. 3. A well known theorem from vector calculus states that if the curl of a vector field u is zero then
(subject to some restrictions on the geometry of the domain) the vector field is conservative and there
exists a function φ so that u = gradφ
Ex. 4. Show that ψ satisfies the continuity equation.
ANSWERS TO CHAPTER 12 EXERCISES
Sec. 2.
Ex. 1.
Modify Equation (12.1.5) to read
P(N→ N-1)=NkΔt+p(Δt)
Eq (12.1.7) becomes
PN(t+Δt)=PN(1-kNΔt)+PN+1k(N+1)Δt.
Observe however that N ≥ 0.
Ex. 2.
Assume that the processes of birth and death are independent. On the time interval Δt the
probability of birth is
P(N → N + 1) = Nk 1 ▵ t + p 1( ▵ t),
and the probability of death is
P(N→ N-1)=Nk2△ t+p2(△ t)
Eq (12.1.7) becomes then
PN(t+Δt)=PN(1-(k1+k2)NΔt)+PN-1k1(N-1)Δt+PN+1k2(N+1)Δt
Ex. 4.
Observe that σ 2 can be rewritten as
σ2=∑N>N0N2PN(t)-μ(t)2
Now differentiate this formula with respect to t and use Equations (12.1.8), (12.1.9) and (12.1.18).
Sec. 3.
Ex. 1. Differentiate Equation (12.3.7) and use Equations. (12.3.4), (12.3.5).
Ex. 3.
Let the number of males and females in the colony at time t be denoted by M(t) and F(t)
respectively. Assume that on the time interval [ t , t + △ t ] the probability for one mating in the
colony is kM(t)N(t)t and the probability of more than one mating is O((Deltat)2). The probability that
at time t + ▵ t the number of males in the population is M(t) + 1 is
PM+1(t+△ t)=PM(t)+12kM(t)F(t)△ t+O((D
Similarly for the female population we have
PF+1(t+△ t)=PF(t)+12kM(t)F(t)△ t+O((Deltat)2)
This leads to a system of two coupled equations for the probability densities which can be solved
subject to initial conditions on the male and female populations at the initial time.
Sec. 4
Ex. 1.
Equation (12.4.2) is modified as follows:
(13.7)
Pℓ(t+Δt)=Pℓt(1-kΔt)(1-NsΔt)+Pℓ-1t(kΔt)(1-NsΔt)+Pℓ+1t(NsΔt)
(1-kΔt)+O((Δt)2)|.,ℓ≥N.
Similarly for ℓ < N we have
Pℓ(t+Δt)=Pℓt(1-kΔt)+Pℓ+1t(NsΔt)(1-kΔt)+O((Δt)2)
Sec. 5
Ex. 1.
P2k+1=P,P2(k+1)=P2,k=0,1,....
Ex. 2. Make one single bet with all the money the gambler has.