with Applications
in Mechanics
This page intentionally left blank This page intentionally left blank
NE W J E RSE Y • L ONDON • SI NGAP ORE • BE I J I NG • SHANGHAI • HONG KONG • TAI P E I • CHE NNAI
World Scientifc
Leonid P. Lebedev
National University of Colombia
Southern Federal University, Russia
Michael J. Cloud
Lawrence Technological University, USA
Victor A. Eremeyev
MartinLutherUniversity HalleWittenberg, Germany
Southern Scientifc Center of Russian Academy of Science
Southern Federal University, Russia
Tensor Analysis
with Applications
in Mechanics
British Library CataloguinginPublication Data
A catalogue record for this book is available from the British Library.
For photocopying of material in this volume, please pay a copying fee through the Copyright
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to
photocopy is not required from the publisher.
Desk Editor: Tjan Kwang Wei
ISBN13 9789814313124
ISBN10 9814313122
All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means,
electronic or mechanical, including photocopying, recording or any information storage and retrieval
system now known or to be invented, without written permission from the Publisher.
Copyright © 2010 by World Scientific Publishing Co. Pte. Ltd.
Published by
World Scientific Publishing Co. Pte. Ltd.
5 Toh Tuck Link, Singapore 596224
USA office: 27 Warren Street, Suite 401402, Hackensack, NJ 07601
UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
Printed in Singapore.
TENSOR ANALYSIS WITH APPLICATIONS IN MECHANICS
Foreword
Every science elaborates tools for the description of its objects of study. In
classical mechanics we make extensive use of vectorial quantities: forces,
moments, positions, velocities, momenta. Conﬁning ourselves to a single
coordinate frame, we can regard a vector as a ﬁxed column matrix. The
deﬁnitive trait of a vector quantity, however, is its objectivity; a vector does
not depend on our choice of coordinate frame. This means that as soon
as the components of a force are speciﬁed in one frame, the components
of that force relative to any other frame can be found through the use of
appropriate transformation rules.
But vector quantities alone do not suﬃce for the description of con
tinuum media. The stress and strain at a point inside a body are also
objective quantities; however, the speciﬁcation of each of these relative to a
given frame requires a square matrix of elements. Under changes of frame,
these elements transform according to rules diﬀerent from the transforma
tion rules for vectors. Stress and strain tensors are examples of tensors of
the second order. We could go on to cite other objective quantities that oc
cur in the mechanics of continua. The set of elastic moduli associated with
Hooke’s law comprise a tensor of the fourth order; as such, these moduli
obey yet another set of transformation rules. Despite the diﬀerences that
exist between the transformation laws for the various types of objective
quantities, they all ﬁt into a uniﬁed scheme: the theory of tensors.
Tensor theory not only relieves our memory from a huge burden, but en
ables us to carry out diﬀerential operations with ease. This is the case even
in curvilinear coordinate systems. Through the unmatched simplicity and
brevity it aﬀords, tensor analysis has attained the status of a general lan
guage that can be spoken across the various areas of continuum physics. A
full comprehension of this language has become necessary for those working
v
vi Tensor Analysis with Applications in Mechanics
in electromagnetism, the theory of relativity, or virtually any other ﬁeld
theoretic discipline. More modern books on physical subjects invariably
contain appendices in which various vector and tensor identities are listed.
These may suﬃce when one wishes to verify the steps in a development, but
can leave one in doubt as to how the facts were established or, a fortiori,
how they could be adapted to other circumstances. On the other hand,
a comprehensive treatment of tensors (e.g., involving a full excursion into
multilinear algebra) is necessarily so large as to be ﬂatly inappropriate for
the student or practicing engineer.
Hence the need for a treatment of tensor theory that does justice to
the subject and is friendly to the practitioner. The authors of the present
book have met these objectives with a presentation that is simple, clear,
and suﬃciently detailed. The concise text explains practically all those
formulas needed to describe objects in threedimensional space. Occur
rences in physics are mentioned when helpful, but the discussion is kept
largely independent of application area in order to appeal to the widest
possible audience. A chapter on the properties of curves and surfaces has
been included; a brief introduction to the study of these properties can be
considered as an informative and natural extension of tensor theory.
I.I. Vorovich
Late Professor of Mechanics and Mathematics
Rostov State University, Russia
Fellow of Russian Academy of Sciences
(1920–2001)
Preface
The ﬁrst edition of this book was written for students, engineers, and physi
cists who must employ tensor techniques. We did not present the mate
rial in complete generality for the case of ndimensional space, but rather
presented a threedimensional version (which is easy to extend to n dimen
sions); hence we could assume a background consisting only of standard
calculus and linear algebra.
We have decided to extend the book in a natural direction, adding two
chapters on applications for which tensor analysis is the principal tool. One
chapter is on linear elasticity and the other is on the theory of shells and
plates. We present complete derivations of the equations in these theories,
formulate boundary value problems, and discuss the problem of uniqueness
of solutions, Lagrange’s variational principle, and some problems on vibra
tion. Space restrictions prohibited us from presenting an entire course on
mechanics; we had to select those questions in elasticity where the role of
tensor analysis is most crucial.
We should mention the essential nature of tensors in elasticity and shell
theory. Of course, to solve a certain engineering problem, one should write
things out in component form; sometimes this takes a few pages. The corre
sponding formulas in tensor notation are quite simple, allowing us to grasp
the underlying ideas and perform manipulations with relative ease. Be
cause tensor representation leads quickly and painlessly to componentwise
representation, this technique is ideal for presenting continuum theories to
students.
The ﬁrst ﬁve chapters are largely unmodiﬁed, aside from some new
problem sets and material on tensorial functions needed for the chapters
on elasticity. The endofchapter problems are supplementary, whereas the
integrated exercises are required for a proper understanding of the text.
vii
viii Tensor Analysis with Applications in Mechanics
In the ﬁrst edition we used the term rank instead of order. This was
common in the older literature. In the newer literature, the term “rank” is
often assigned a diﬀerent meaning.
Because the book is largely selfcontained, we make no attempt at a
comprehensive reference list. We merely list certain books that cover similar
material, that extend the treatment slightly, or that may be otherwise useful
to the reader.
We are deeply grateful to our World Scientiﬁc editor, Mr. Tjan Kwang
Wei, for his encouragement and support.
L.P. Lebedev
Department of Mathematics
National University of Colombia, Colombia
M.J. Cloud
Department of Electrical and Computer Engineering
Lawrence Technological University, USA
V.A. Eremeyev
South Scientiﬁc Center of RASci
&
Department of Mathematics, Mechanics
and Computer Sciences
South Federal University, Russia
Preface to the First Edition
Originally a vector was regarded as an arrow of a certain length that could
represent a force acting on a material point. Over a period of many years,
this naive viewpoint evolved into the modern interpretation of the notion
of vector and its extension to tensors. It was found that the use of vectors
and tensors led to a proper description of certain properties and behaviors
of real natural objects: those aspects that do not depend on the coordinate
systems we introduce in space. This independence means that if we deﬁne
such properties using one coordinate system, then in another system we can
recalculate these characteristics using valid transformation rules. The ease
Preface ix
with which a given problem can be solved often depends on the coordinate
system employed. So in applications we must apply various coordinate
systems, derive corresponding equations, and understand how to recalculate
results in other systems. This book provides the tools necessary for such
calculation.
Many physical laws are cumbersome when written in coordinate form
but become compact and attractive looking when written in tensorial form.
Such compact forms are easy to remember, and can be used to state complex
physical boundary value problems. It is conceivable that soon an ability to
merely formulate statements of boundary value problems will be regarded
as a fundamental skill for the practitioner. Indeed, computer software is
slowly advancing toward the point where the only necessary input data will
be a coordinatefree statement of a boundary value problem; presumably
the user will be able to initiate a solution process in a certain frame and
by a certain method (analytical, numerical, or mixed), or simply ask the
computer algorithm to choose the best frame and method. In this way,
vectors and tensors will become important elements of the macrolanguage
for the next generation of software in engineering and applied mathematics.
We would like to thank the editorial staﬀ at World Scientiﬁc — espe
cially Mr. Tjan Kwang Wei and Ms. SookCheng Lim — for their assistance
in the production of this book. Professor Byron C. Drachman of Michigan
State University commented on the manuscript in its initial stages. Lastly,
Natasha Lebedeva and Beth LannonCloud deserve thanks for their pa
tience and support.
L.P. Lebedev
Department of Mechanics and Mathematics
Rostov State University, Russia
&
Department of Mathematics
National University of Colombia, Colombia
M.J. Cloud
Department of Electrical and Computer Engineering
Lawrence Technological University, USA
This page intentionally left blank This page intentionally left blank
Contents
Foreword v
Preface vii
Tensor Analysis 1
1. Preliminaries 3
1.1 The Vector Concept Revisited . . . . . . . . . . . . . . . . 3
1.2 A First Look at Tensors . . . . . . . . . . . . . . . . . . . 4
1.3 Assumed Background . . . . . . . . . . . . . . . . . . . . 5
1.4 More on the Notion of a Vector . . . . . . . . . . . . . . . 7
1.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2. Transformations and Vectors 11
2.1 Change of Basis . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Dual Bases . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Transformation to the Reciprocal Frame . . . . . . . . . . 17
2.4 Transformation Between General Frames . . . . . . . . . . 18
2.5 Covariant and Contravariant Components . . . . . . . . . 21
2.6 The Cross Product in Index Notation . . . . . . . . . . . 22
2.7 Norms on the Space of Vectors . . . . . . . . . . . . . . . 24
2.8 Closing Remarks . . . . . . . . . . . . . . . . . . . . . . . 27
2.9 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3. Tensors 29
3.1 Dyadic Quantities and Tensors . . . . . . . . . . . . . . . 29
xi
xii Tensor Analysis with Applications in Mechanics
3.2 Tensors From an Operator Viewpoint . . . . . . . . . . . 30
3.3 Dyadic Components Under Transformation . . . . . . . . 34
3.4 More Dyadic Operations . . . . . . . . . . . . . . . . . . . 36
3.5 Properties of SecondOrder Tensors . . . . . . . . . . . . . 40
3.6 Eigenvalues and Eigenvectors of a SecondOrder Symmet
ric Tensor . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.7 The Cayley–Hamilton Theorem . . . . . . . . . . . . . . . 48
3.8 Other Properties of SecondOrder Tensors . . . . . . . . . 49
3.9 Extending the Dyad Idea . . . . . . . . . . . . . . . . . . 56
3.10 Tensors of the Fourth and Higher Orders . . . . . . . . . 58
3.11 Functions of Tensorial Arguments . . . . . . . . . . . . . . 60
3.12 Norms for Tensors, and Some Spaces . . . . . . . . . . . . 66
3.13 Diﬀerentiation of Tensorial Functions . . . . . . . . . . . . 70
3.14 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4. Tensor Fields 85
4.1 Vector Fields . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.2 Diﬀerentials and the Nabla Operator . . . . . . . . . . . . 94
4.3 Diﬀerentiation of a Vector Function . . . . . . . . . . . . 98
4.4 Derivatives of the Frame Vectors . . . . . . . . . . . . . . 99
4.5 Christoﬀel Coeﬃcients and their Properties . . . . . . . . 100
4.6 Covariant Diﬀerentiation . . . . . . . . . . . . . . . . . . . 105
4.7 Covariant Derivative of a SecondOrder Tensor . . . . . . 106
4.8 Diﬀerential Operations . . . . . . . . . . . . . . . . . . . . 108
4.9 Orthogonal Coordinate Systems . . . . . . . . . . . . . . . 113
4.10 Some Formulas of Integration . . . . . . . . . . . . . . . . 117
4.11 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
5. Elements of Diﬀerential Geometry 125
5.1 Elementary Facts from the Theory of Curves . . . . . . . 126
5.2 The Torsion of a Curve . . . . . . . . . . . . . . . . . . . 132
5.3 Frenet–Serret Equations . . . . . . . . . . . . . . . . . . . 135
5.4 Elements of the Theory of Surfaces . . . . . . . . . . . . . 137
5.5 The Second Fundamental Form of a Surface . . . . . . . . 148
5.6 Derivation Formulas . . . . . . . . . . . . . . . . . . . . . 153
5.7 Implicit Representation of a Curve; Contact of Curves . . 156
5.8 Osculating Paraboloid . . . . . . . . . . . . . . . . . . . . 162
5.9 The Principal Curvatures of a Surface . . . . . . . . . . . 164
Contents xiii
5.10 Surfaces of Revolution . . . . . . . . . . . . . . . . . . . . 168
5.11 Natural Equations of a Curve . . . . . . . . . . . . . . . . 170
5.12 A Word About Rigor . . . . . . . . . . . . . . . . . . . . . 173
5.13 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 175
5.14 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Applications in Mechanics 179
6. Linear Elasticity 181
6.1 Stress Tensor . . . . . . . . . . . . . . . . . . . . . . . . . 181
6.2 Strain Tensor . . . . . . . . . . . . . . . . . . . . . . . . . 190
6.3 Equation of Motion . . . . . . . . . . . . . . . . . . . . . . 193
6.4 Hooke’s Law . . . . . . . . . . . . . . . . . . . . . . . . . 194
6.5 Equilibrium Equations in Displacements . . . . . . . . . . 200
6.6 Boundary Conditions and Boundary Value Problems . . . 202
6.7 Equilibrium Equations in Stresses . . . . . . . . . . . . . . 203
6.8 Uniqueness of Solution for the Boundary Value Problems
of Elasticity . . . . . . . . . . . . . . . . . . . . . . . . . . 205
6.9 Betti’s Reciprocity Theorem . . . . . . . . . . . . . . . . . 206
6.10 Minimum Total Energy Principle . . . . . . . . . . . . . . 208
6.11 Ritz’s Method . . . . . . . . . . . . . . . . . . . . . . . . . 216
6.12 Rayleigh’s Variational Principle . . . . . . . . . . . . . . . 221
6.13 Plane Waves . . . . . . . . . . . . . . . . . . . . . . . . . 227
6.14 Plane Problems of Elasticity . . . . . . . . . . . . . . . . . 230
6.15 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
7. Linear Elastic Shells 237
7.1 Some Useful Formulas of Surface Theory . . . . . . . . . . 239
7.2 Kinematics in a Neighborhood of Σ . . . . . . . . . . . . . 242
7.3 Shell Equilibrium Equations . . . . . . . . . . . . . . . . . 244
7.4 Shell Deformation and Strains; Kirchhoﬀ’s Hypotheses . . 249
7.5 Shell Energy . . . . . . . . . . . . . . . . . . . . . . . . . 256
7.6 Boundary Conditions . . . . . . . . . . . . . . . . . . . . . 259
7.7 A Few Remarks on the Kirchhoﬀ–Love Theory . . . . . . 261
7.8 Plate Theory . . . . . . . . . . . . . . . . . . . . . . . . . 263
7.9 On NonClassical Theories of Plates and Shells . . . . . . 277
Appendix A Formulary 287
xiv Tensor Analysis with Applications in Mechanics
Appendix B Hints and Answers 315
Bibliography 355
Index 359
Chapter 1
Preliminaries
1.1 The Vector Concept Revisited
The concept of a vector has been one of the most fruitful ideas in all of
mathematics, and it is not surprising that we receive repeated exposure to
the idea throughout our education. Students in elementary mathematics
deal with vectors in component form — with quantities such as
x = (2, 1, 3)
for example. But let us examine this situation more closely. Do the compo
nents 2, 1, 3 determine the vector x? They surely do if we specify the basis
vectors of the coordinate frame. In elementary mathematics these are sup
posed to be mutually orthogonal and of unit length; even then they are not
fully characterized, however, because such a frame can be rotated. In the
description of many common phenomena we deal with vectorial quantities
like forces that have deﬁnite directions and magnitudes. An example is the
force your body exerts on a chair as you sit in front of the television set.
This force does not depend on the coordinate frame employed by someone
writing a textbook on vectors somewhere in Russia or China. Because the
vector f representing a particular force is something objective, we should
be able to write it in such a form that it ceases to depend on the details of
the coordinate frame. The simplest way is to incorporate the frame vectors
e
i
(i = 1, 2, 3) explicitly into the notation: if x is a vector we may write
x =
3
i=1
x
i
e
i
. (1.1)
Then if we wish to change the frame, we should do so in such a way that
x remains the same. This of course means that we cannot change only the
3
4 Tensor Analysis with Applications in Mechanics
frame vectors e
i
: we must change the components x
i
correspondingly. So
the components of a vector x in a new frame are not independent of those
in the old frame.
1.2 A First Look at Tensors
In what follows we shall discuss how to work with vectors using diﬀerent
coordinate frames. Let us note that in mechanics there are objects of
another nature. For example, there is a socalled tensor of inertia. This
is an objective characteristic of a solid body, determining how the body
rotates when torques act upon it. If the body is considered in a Cartesian
frame, the tensor of inertia is described by a 3×3 matrix. If we change the
frame, the matrix elements change according to certain rules. In textbooks
on mechanics the reader can ﬁnd lengthy discussions on how to change the
matrix elements to maintain the same objective characteristic of the body
when the new frame is also Cartesian. Although the tensor of inertia is
objective (i.e., frameindependent), it is not a vector: it belongs to another
class of mathematical objects. Many such tensors of the second order arise
in continuum mechanics: tensors of stress, strain, etc. They characterize
certain properties of a body at each point; again, their “components” should
transform in such a way that the tensors themselves do not depend on the
frame. The precise meaning of the term order will be explained later.
For both vectors and tensors we can introduce various operations. Of
course, the introduction of any new operation should be done in such a way
that the results agree with known special cases when such familiar cases
are met. If we introduce, say, dot multiplication of a tensor by a vector,
then in a Cartesian frame the operation should resemble the multiplication
of a matrix by a column vector. Similarly, the multiplication of two ten
sors should be deﬁned so that in a Cartesian frame the operation involves
matrix multiplication. To this end we consider dyads of vectors. These are
quantities of the form
e
i
e
j
.
A tensor may then be represented as
i,j
a
ij
e
i
e
j
where the a
ij
are the components of the tensor. We compare with equation
(1.1) and notice the similarity in notation.
Preliminaries 5
The quantity e
i
e
j
is also called the tensor product of the vectors e
i
and
e
j
, and is sometimes denoted e
i
⊗ e
j
. Our notation (without the symbol
⊗) emphasizes that, for example, e
1
e
2
is an elemental object belonging to
the set of secondorder tensors, in the same way that e
1
is an elemental ob
ject belonging to the set of vectors. Note that e
2
e
1
and e
1
e
2
are diﬀerent
objects, however. The term “tensor product” indicates that the opera
tion shares certain properties with the product we know from elementary
algebra. We will discuss this further in Chapter 3.
Natural objects can possess characteristics described by tensors of higher
order. For example, the elastic properties of a body are described by a
tensor of the fourth order (i.e., a tensor whose elemental parts are of the
form abcd, where a, b, c, d are vectors). This means that in general the
properties of a body are given by a “table” consisting of 3 ×3 ×3 ×3 = 81
elements. The elements change according to certain rules if we change the
frame.
Tensors also occur in electrodynamics, the general theory of relativity,
and other sciences that deal with objects situated or distributed in space.
1.3 Assumed Background
In what follows we suppose a familiarity with the dot and cross products
and their expression in Cartesian frames. Recall that if a and b are vectors,
then by deﬁnition
a · b = ab cos θ,
where a and b are the magnitudes of a and b and θ is the (smaller)
angle between a and b. From now on we reserve the symbol i for the basis
vectors of a Cartesian system. In a Cartesian frame with basis vectors
i
1
, i
2
, i
3
where a and b are expressed as
a = a
1
i
1
+ a
2
i
2
+ a
3
i
3
, b = b
1
i
1
+ b
2
i
2
+ b
3
i
3
,
we have
a · b = a
1
b
1
+ a
2
b
2
+ a
3
b
3
.
Also recall that
a ×b =
i
1
i
2
i
3
a
1
a
2
a
3
b
1
b
2
b
3
.
6 Tensor Analysis with Applications in Mechanics
The dot product will play a role in our discussion from the very beginning.
The cross product will be used as needed, and a fuller discussion will appear
in § 2.6.
b
a
c
Fig. 1.1 Geometrical meaning of the scalar triple product.
Given three vectors a, b, c we can form the scalar triple product
a · (b ×c).
This may be interpreted as the volume of the parallelepiped having a, b, c
as three of its coterminal edges (Fig. 1.1). In rectangular components we
have, according to the expressions above,
a · (b ×c) =
a
1
a
2
a
3
b
1
b
2
b
3
c
1
c
2
c
3
.
Permissible manipulations with the scalar triple product include cyclic in
terchange of the vectors:
a · (b ×c) = c · (a ×b) = b · (c ×a).
Exercise 1.1. What does the condition a · (b × c) = 0 say about a, b,
and c? (Hints for this and many other exercises appear in Appendix B
beginning on page 315.)
So far we have made reference to vectors in a threedimensional space,
and shall continue this practice throughout most of the book. It is also
possible (and sometimes useful) to introduce vectors in a more general
space of n > 3 dimensions, e.g.,
a = i
1
a
1
+i
2
a
2
+· · · +i
n
a
n
.
Preliminaries 7
It turns out that many (but not all) of the principles and techniques we
shall learn have direct extensions to such higherdimensional spaces. It
is also true that many threedimensional notions can be reconsidered in
two dimensions. The reader should take the time to reduce each three
dimensional formula to its twodimensional analogue to understand what
happens with the corresponding assertion.
1.4 More on the Notion of a Vector
Before closing out this chapter we should mention the notions of a vector
as a “directed line segment” or “quantity having magnitude and direction.”
We ﬁnd these in elementary math and physics textbooks. But it is easy to
point to a situation in which a quantity of interest has magnitude and di
rection but is not a vector. The total electric current ﬂowing in a thin wire
is one example: to describe the current we must specify the rate of ﬂow of
electrons, the orientation of the wire, and the sense of electron movement
along the wire. However, if two wires lie in a plane and carry equal cur
rents running perpendicular to each other, then we cannot duplicate their
physical eﬀects by replacing them with a third wire directed at a 45
◦
angle
with respect to the ﬁrst two. Total electric currents cannot be considered
as vector quantities since they do not combine according to the rule for
vector addition.
Another problem concerns the notion of an ndimensional Euclidean
space. We sometimes hear it deﬁned as the set each point of which is
uniquely determined by n parameters. However, it is not reasonable to
regard every list of n things as a possible vector. A tailor may take mea
surements for her client and compose an ordered list of lengths, widths, etc.,
and by the above “deﬁnition” the set of all such lists is an ndimensional
space. But in R
n
any two points are joined by a vector whose components
equal the diﬀerences of the corresponding coordinates of the points. In the
tailor’s case this notion is completely senseless. To make matters worse,
in R
n
one can multiply any vector by any real number and obtain another
vector in the space. A tailor’s list showing that someone is six meters tall
would be pretty hard to ﬁnd.
Such simplistic deﬁnitions can be dangerous. The notion of a vector
was eventually elaborated in physics, more precisely in mechanics where
the behavior of forces was used as the model for a general vector. However,
forces have some rather strange features: for example, if we shift the point
8 Tensor Analysis with Applications in Mechanics
of application of a force acting on a solid body, then we must introduce a
moment acting on the body or the motion of the body will change. So the
mechanical forces that gave birth to the notion of a vector possess features
not covered by the mathematical deﬁnition of a vector. In the classical
mechanics of a rigid body we are allowed to move a force along its line of
action but cannot simply shift it oﬀ that line. With a deformable body we
cannot move a force anywhere because in doing so we immediately change
the state of the body. We can vectorially add two forces acting on a rigid
body (not forgetting about the moments arising during shift of the forces).
On the other hand, if two forces act on two diﬀerent material points then
we can add the vectors that represent the forces, but will not necessarily
obtain a new vector that is relevant to the physics of the situation. So we
should understand that the idea of a vector in mathematics reﬂects only
certain features of the real objects it describes.
We would like to mention something else about vectors. In elementary
mathematics students use vectors and points quite interchangeably. How
ever, these are objects of diﬀerent natures: there are vectors in space and
there are the points of a space. We can, for instance, associate with a vec
tor in an ndimensional Euclidean vector space a point in an ndimensional
Euclidean point space. We can then consider a vector x as a shift of all
points of the point space by an amount speciﬁed by x. The result of this
map is the same space of points; each pair of points that correspond under
the mapping deﬁne a vector x under which a point shifts into its image.
This vector is what we ﬁnd when we subtract the Cartesian coordinates
of the initial point from those of the ﬁnal point. If we add the fact that
the composition of two such maps obeys the rules of vector addition, then
we get a strict deﬁnition of the space introduced intuitively in elementary
mathematics. Engineers might object to the imposition of such formality
on an apparently simple situation, but mathematical rigor has proved its
worth from many practical points of view. For example, a computer that
processes information in the complete absence of intuition can deal prop
erly with objects for which rigorous deﬁnitions and manipulation rules have
been formulated.
This brings us to our ﬁnal word, regarding the expression of vectors in
componentfree notation. The simple and compact notation x for a vector
leads to powerful ways of exhibiting relationships between the many vector
(and tensor) quantities that occur in mathematical physics. It permits
us to accomplish manipulations that would take many pages and become
quite confusing if done in component form. This is typical for problems of
Preliminaries 9
nonlinear physics, and for those where change of coordinate frame becomes
necessary. The resulting formal nature of the manipulations means that
computers can be expected to take on more and more of this sort of work,
even at the level of original research.
1.5 Problems
1.1 Find 2a
1
+ 3a
2
for the vectors a
1
and a
2
given by
(a) a
1
= (1, 2, −1), a
2
= (1, 1, 2);
(b) a
1
= (−1, 2, 0), a
2
= (3, 1, −2);
(c) a
1
= (1, 3, 4), a
2
= (5, 1, −4).
1.2 Find a
1
+ 3a
2
−2a
3
for the vectors a
1
, a
2
, a
3
given by
(a) a
1
= (−1, 2, −2), a
2
= (1, −1, 2), a
3
= (1, −1, 2);
(b) a
1
= (1, 3, 2), a
2
= (2, −3, 2), a
3
= (3, 2, 3);
(c) a
1
= (2, 1, 2), a
2
= (4, 3, 0), a
3
= (1, 1, −2).
1.3 Find x satisfying the equation
(a) a + 2x −4b = 0;
(b) 2(a + 2x) +b = 3(b +x);
(c) x + 2a + 16(x −b) +c = 2(c + 3x) +a −x.
1.4 Find the values of a · b and a ×b if
(a) a = (0, 1, 1), b = (1, 1, 1);
(b) a = (1, 2, 3), b = (2, 3, 1);
(c) a = (1, 1, 1), b = (1, 1, 2);
(d) a = (−1, 1, −1), b = (2, −1, −1);
(e) a = (0, −1, 1), b = (−1, 1, 0);
(f) a = (−1, 1, 0), b = (2, 3, 0);
(g) a = (1, 2, −1), b = (1, 1, 2);
(h) a = (−1, 2, 0), b = (3, 1, −2);
(i) a = (1, 3, 4), b = (5, 1, −4).
1.5 Show that (a) i
1
×i
2
= i
3
; (b) i
2
×i
3
= i
1
; (c) i
3
×i
1
= i
2
.
10 Tensor Analysis with Applications in Mechanics
1.6 Show that the equation a × i
1
= −a
2
i
3
+ a
3
i
2
holds for an arbitrary
vector a = a
1
i
1
+ a
2
i
2
+ a
3
i
3
.
1.7 Let a = a
1
i
1
+ a
2
i
2
, b = b
1
i
1
+ b
2
i
2
, and let i
1
, i
2
, i
3
be a Cartesian
basis. Show that a ×b = (a
1
b
2
−a
2
b
1
)i
3
.
1.8 Suppose a ×x = 0, a · x = 0, and a = 0. Demonstrate that x = 0.
1.9 Find: (a) i
1
· (i
2
×i
3
); (b) i
1
· (i
3
×i
2
); (c) i
1
· (i
3
×i
1
).
1.10 Find a
1
· (a
2
×a
3
) when
(a) a
1
= (−1, 2, −2), a
2
= (1, −1, 2), a
3
= (1, −1, 3);
(b) a
1
= (−1, 1, 0), a
2
= (1, 1, 1), a
3
= (1, −1, 2);
(c) a
1
= (1, 1, 1), a
2
= (1, 2, 2), a
3
= (1, −3, 2);
(d) a
1
= (−1, 2, −1), a
2
= (1, −2, 2), a
3
= (3, −1, 3);
(e) a
1
= (9, 8, 4), a
2
= (7, 1, 3), a
3
= (5, 3, 6);
(f) a
1
= (1, 2, 3), a
2
= (7, 2, 3), a
3
= (1, 4, 6);
(g) a
1
= (−1, 2, −2), a
2
= (1, −1, 2), a
3
= (1, −1, 2);
(h) a
1
= (1, 3, 2), a
2
= (2, −3, 2), a
3
= (3, 2, 3);
(i) a
1
= (2, 1, 2), a
2
= (4, 3, 0), a
3
= (1, 1, −2).
1.11 Find (a
1
×a
2
) · (a
3
× a
4
), where
(a) a
1
= (−1, 2, −2), a
2
= (1, −1, 2), a
3
= (1, −1, 3), a
4
= (1, 0, 0);
(b) a
1
= (−1, 1, 0), a
2
= (1, 1, 1), a
3
= (1, −1, 2), a
4
= (0, −1, 0);
(c) a
1
= (1, 1, 1), a
2
= (1, 2, 2), a
3
= (1, −3, 2), a
4
= (1, 1, 1);
(d) a
1
= (−1, 2, −1), a
2
= (1, −2, 2), a
3
= (3, −1, 3), a
4
= (1, 0, 2);
(e) a
1
= (9, 8, 4), a
2
= (7, 1, 3), a
3
= (5, 3, 6), a
4
= (2, 3, −1);
(f) a
1
= (1, 2, 3), a
2
= (7, 2, 3), a
3
= (1, 4, 6), a
4
= (0, 0, 1);
(g) a
1
= (−1, 2, −2), a
2
= (1, −1, 2), a
3
= (1, −1, 2), a
4
= (2, −2, 4);
(h) a
1
= (1, 3, 2), a
2
= (2, −3, 2), a
3
= (3, 2, 3), a
4
= (1, −1, 1);
(i) a
1
= (2, 1, 2), a
2
= (−6, −3, −6), a
3
= (1, 1, −2), a
4
= (1, 12, 3).
Chapter 2
Transformations and Vectors
2.1 Change of Basis
Let us reconsider the vector
x = (2, 1, 3).
Fully written out in a given Cartesian frame e
i
(i = 1, 2, 3), it is
x = 2e
1
+ e
2
+ 3e
3
.
(This is one of the few times we do not use i as the symbol for a Cartesian
frame vector.) Suppose we appoint a new frame ˜ e
i
(i = 1, 2, 3) such that
e
1
= ˜ e
1
+ 2˜ e
2
+ 3˜ e
3
,
e
2
= 4˜ e
1
+ 5˜ e
2
+ 6˜ e
3
,
e
3
= 7˜ e
1
+ 8˜ e
2
+ 9˜ e
3
.
From these expansions we could calculate the ˜ e
i
and verify that they are
noncoplanar. As x is an objective, frameindependent entity, we can write
x = 2(˜ e
1
+ 2˜ e
2
+ 3˜ e
3
) + (4˜ e
1
+ 5˜ e
2
+ 6˜ e
3
) + 3(7˜ e
1
+ 8˜ e
2
+ 9˜ e
3
)
= (2 + 4 + 21)˜ e
1
+ (4 + 5 + 24)˜ e
2
+ (6 + 6 + 27)˜ e
3
= 27˜ e
1
+ 33˜ e
2
+ 39˜ e
3
.
In these calculations it is unimportant whether the frames are Cartesian;
it is important only that we have the table of transformation
⎛
⎝
1 2 3
4 5 6
7 8 9
⎞
⎠
.
11
12 Tensor Analysis with Applications in Mechanics
It is clear that we can repeat the same operation in general form. Let
x be of the form
x =
3
¸
i=1
x
i
e
i
(2.1)
with the table of transformation of the frame given as
e
i
=
3
¸
j=1
A
j
i
˜ e
j
.
Then
x =
3
¸
i=1
x
i
3
¸
j=1
A
j
i
˜ e
j
=
3
¸
j=1
˜ e
j
3
¸
i=1
A
j
i
x
i
.
So in the new basis we have
x =
3
¸
j=1
˜ x
j
˜ e
j
where ˜ x
j
=
3
¸
i=1
A
j
i
x
i
.
Here we have introduced a new notation, placing some indices as subscripts
and some as superscripts. Although this practice may seem artiﬁcial, there
are fairly deep reasons for following it.
2.2 Dual Bases
To perform operations with a vector x, we must have a straightforward
method of calculating its components — ultimately, no matter how ad
vanced we are, we must be able to obtain the x
i
using simple arithmetic.
We prefer formulas that permit us to ﬁnd the components of vectors using
dot multiplication only; we shall need these when doing frame transfor
mations, etc. In a Cartesian frame the necessary operation is simple dot
multiplication by the corresponding basis vector of the frame: we have
x
k
= x i
k
(k = 1, 2, 3).
This procedure fails in a more general nonCartesian frame where we do not
necessarily have e
i
e
j
= 0 for all j = i. However, it may still be possible
to ﬁnd a vector e
i
such that
x
i
= x e
i
(i = 1, 2, 3)
Transformations and Vectors 13
in this more general situation. If we set
x
i
= x e
i
=
⎛
⎝
3
¸
j=1
x
j
e
j
⎞
⎠
e
i
=
3
¸
j=1
x
j
(e
j
e
i
)
and compare the left and righthand sides, we see that equality holds when
e
j
e
i
= δ
i
j
(2.2)
where
δ
i
j
=
1, j = i,
0, j = i,
is the Kronecker delta symbol. In a Cartesian frame we have
e
k
= e
k
= i
k
for each k.
Exercise 2.1. Show that e
i
is determined uniquely by the requirement
that x
i
= x e
i
for every x.
Now let us discuss the geometrical nature of the vectors e
i
. Consider,
for example, the equations for e
1
:
e
1
e
1
= 1, e
2
e
1
= 0, e
3
e
1
= 0.
We see that e
1
is orthogonal to both e
2
and e
3
, and its magnitude is such
that e
1
e
1
= 1. Similar properties hold for e
2
and e
3
.
Exercise 2.2. Show that the vectors e
i
are linearly independent.
By Exercise 2.2, the e
i
constitute a frame or basis. This basis is said to
be reciprocal or dual to the basis e
i
. We can therefore expand an arbitrary
vector x as
x =
3
¸
i=1
x
i
e
i
. (2.3)
Note that superscripts and subscripts continue to appear in our notation,
but in a way complementary to that used in equation (2.1). If we dot
multiply the representation (2.3) of x by e
j
and use (2.2) we get x
j
. This
explains why the frames e
i
and e
i
are dual: the formulas
x e
i
= x
i
, x e
i
= x
i
,
14 Tensor Analysis with Applications in Mechanics
look quite similar. So the introduction of a reciprocal basis gives many
potential advantages.
Let us discuss the reciprocal basis in more detail. The ﬁrst problem
is to ﬁnd suitable formulas to deﬁne it. We derive these formulas next,
but ﬁrst let us note the following. The use of reciprocal vectors may not
be practical in those situations where we are working with only two or
three vectors. The real advantages come when we are working intensively
with many vectors. This is reminiscent of the solution of a set of linear
simultaneous equations: it is ineﬃcient to ﬁnd the inverse matrix of the
system if we have only one forcing vector. But when we must solve such
a problem repeatedly for many forcing vectors, the calculation and use of
the inverse matrix is reasonable.
Writing out x in the e
i
and e
i
bases, we used a combination of indices
(i.e., subscripts and superscripts) and summation symbols. From now on
we shall omit the symbol of summation when we meet matching subscripts
and superscripts: we shall write, say,
x
i
a
i
for the sum
¸
i
x
i
a
i
.
That is, whenever we see i as a subscript and a superscript, we shall under
stand that a summation is to be carried out over i. This rule shall apply
to situations involving vectors as well: we shall understand, for example,
x
i
e
i
to mean the summation
¸
i
x
i
e
i
.
This rule is called the rule of summation over repeated indices.
1
Note that
a repeated index is a dummy index in the sense that it may be replaced by
any other index not already in use: we have
x
i
a
i
= x
1
a
1
+ x
2
a
2
+ x
3
a
3
= x
k
a
k
for instance. An index that occurs just once in an expression, for example
the index i in
A
k
i
x
k
,
is called a free index. In tensor discussions each free index is understood to
range independently over a set of values — presently this set is ¦1, 2, 3¦.
1
The rule of summation was ﬁrst introduced not by mathematicians but by Einstein,
and is sometimes referred to as the Einstein summation convention. In a paper where
he introduced this rule, Einstein used Cartesian frames and therefore did not distinguish
superscripts from subscripts. However, we shall continue to make the distinction so that
we can deal with nonCartesian frames.
Transformations and Vectors 15
Let us return to the task of deriving formulas for the reciprocal basis
vectors e
i
in terms of the original basis vectors e
i
. We construct e
1
ﬁrst.
Since the cross product of two vectors is perpendicular to both, we can
satisfy the conditions
e
2
e
1
= 0, e
3
e
1
= 0,
by setting
e
1
= c
1
(e
2
e
3
)
where c
1
is a constant. To determine c
1
we require
e
1
e
1
= 1.
We obtain
c
1
[e
1
(e
2
e
3
)] = 1.
The quantity e
1
(e
2
e
3
) is a scalar whose absolute value is the volume of
the parallelepiped described by the vectors e
i
. Denoting it by V , we have
e
1
=
1
V
(e
2
e
3
).
Similarly,
e
2
=
1
V
(e
3
e
1
), e
3
=
1
V
(e
1
e
2
).
The reader may verify that these expressions satisfy (2.2). Let us mention
that if we construct the reciprocal basis to the basis e
i
we obtain the initial
basis e
i
. Hence we immediately get the dual formulas
e
1
=
1
V
(e
2
e
3
), e
2
=
1
V
(e
3
e
1
), e
3
=
1
V
(e
1
e
2
),
where
V
= e
1
(e
2
e
3
).
Within an algebraic sign, V
is the volume of the parallelepiped described
by the vectors e
i
.
Exercise 2.3. Show that V
= 1/V .
Let us now consider the forms of the dot product between two vectors
a = a
i
e
i
= a
j
e
j
, b = b
p
e
p
= b
q
e
q
.
16 Tensor Analysis with Applications in Mechanics
We have
a b = a
i
e
i
b
p
e
p
= a
i
b
p
e
i
e
p
.
Introducing the notation
g
ip
= e
i
e
p
, (2.4)
we have
a b = a
i
b
p
g
ip
.
(As a short exercise the reader should write out this expression in full.)
Using the reciprocal component representations we get
a b = a
j
e
j
b
q
e
q
= a
j
b
q
g
jq
where
g
jq
= e
j
e
q
. (2.5)
Finally, using a mixed representation we get
a b = a
i
e
i
b
q
e
q
= a
i
b
q
δ
q
i
= a
i
b
i
and, similarly,
a b = a
j
b
j
.
Hence
a b = a
i
b
j
g
ij
= a
i
b
j
g
ij
= a
i
b
i
= a
i
b
i
.
We see that when we use mixed bases to represent a and b we get formulas
that resemble the equation
a b = a
1
b
1
+ a
2
b
2
+ a
3
b
3
from ' 1.3; otherwise we get more terms and additional multipliers. We will
encounter g
ij
and g
ij
often. They are the components of a unique tensor
known as the metric tensor. In Cartesian frames we obviously have
g
ij
= δ
j
i
, g
ij
= δ
i
j
.
Transformations and Vectors 17
2.3 Transformation to the Reciprocal Frame
How do the components of a vector x transform when we change to the
reciprocal frame? We simply set
x
i
e
i
= x
i
e
i
and dot both sides with e
j
to get
x
i
e
i
e
j
= x
i
e
i
e
j
or
x
j
= x
i
g
ij
. (2.6)
In the system of equations
⎛
⎝
x
1
x
2
x
3
⎞
⎠
=
⎛
⎝
g
11
g
21
g
31
g
12
g
22
g
32
g
13
g
23
g
33
⎞
⎠
⎛
⎝
x
1
x
2
x
3
⎞
⎠
the matrix of the components of the metric tensor g
ij
is also called the
Gram matrix. A theorem in linear algebra states that its determinant is
not zero if and only if the vectors e
i
are linearly independent.
Exercise 2.4. (a) Show that if the Gram determinant vanishes, then the
e
i
are linearly dependent. (b) Prove that the Gram determinant equals V
2
.
We called the basis e
i
dual to the basis e
i
. In e
i
the metric components
are given by g
ij
, so we can immediately write an expression dual to (2.6):
x
i
= x
j
g
ij
. (2.7)
We see from (2.6) and (2.7) that, using the components of the metric tensor,
we can always change subscripts to superscripts and vice versa. These
actions are known as the raising and lowering of indices. Finally, (2.6) and
(2.7) together imply
x
i
= g
ij
g
jk
x
k
,
hence
g
ij
g
jk
= δ
k
i
.
Of course, this means that the matrices of g
ij
and g
ij
are mutually inverse.
18 Tensor Analysis with Applications in Mechanics
Quick summary
Given a basis e
i
, the vectors e
i
given by the requirement that
e
j
e
i
= δ
i
j
are linearly independent and form a basis called the reciprocal or dual basis.
The deﬁnition of dual basis is motivated by the equation x
i
= x e
i
. The
e
i
can be written as
e
i
=
1
V
(e
j
e
k
)
where the ordered triple (i, j, k) equals (1, 2, 3) or one of the cyclic permu
tations (2, 3, 1) or (3, 1, 2), and where
V = e
1
(e
2
e
3
).
The dual of the basis e
k
(i.e., the dual of the dual) is the original basis e
k
.
A given vector x can be expressed as
x = x
i
e
i
= x
i
e
i
where the x
i
are the components of x with respect to the dual basis.
Exercise 2.5. (a) Let x = x
k
e
k
= x
k
e
k
. Write out the modulus of x in all
possible forms using the metric tensor. (b) Write out all forms of the dot
product x y.
2.4 Transformation Between General Frames
Having transformed the components x
i
of a vector x to the corresponding
components x
i
relative to the reciprocal basis, we are now ready to take on
the more general task of transforming the x
i
to the corresponding compo
nents ˜ x
i
relative to any other basis ˜ e
i
. Let the new basis ˜ e
i
be related to
the original basis e
i
by
e
i
= A
j
i
˜ e
j
. (2.8)
This is, of course, compact notation for the system of equations
⎛
⎝
e
1
e
2
e
3
⎞
⎠
=
⎛
⎝
A
1
1
A
2
1
A
3
1
A
1
2
A
2
2
A
3
2
A
1
3
A
2
3
A
3
3
⎞
⎠
. .. .
≡A, say
⎛
⎝
˜ e
1
˜ e
2
˜ e
3
⎞
⎠
.
Transformations and Vectors 19
Before proceeding, we note that in the symbol A
j
i
the subscript indexes
the row number in the matrix A, while the superscript indexes the column
number. Throughout our development we shall often take the time to write
various equations of interest in matrix notation. It follows from (2.8) that
A
j
i
= e
i
˜ e
j
.
Exercise 2.6. A Cartesian frame is rotated about its third axis to give a
new Cartesian frame. Find the matrix of transformation.
A vector x can be expressed in the two forms
x = x
k
e
k
, x = ˜ x
i
˜ e
i
.
Equating these two expressions for the same vector x, we have
˜ x
i
˜ e
i
= x
k
e
k
,
hence
˜ x
i
˜ e
i
= x
k
A
j
k
˜ e
j
. (2.9)
To ﬁnd ˜ x
i
in terms of x
i
, we may expand the notation and write (2.9) as
˜ x
1
˜ e
1
+ ˜ x
2
˜ e
2
+ ˜ x
3
˜ e
3
= x
1
A
j
1
˜ e
j
+ x
2
A
j
2
˜ e
j
+ x
3
A
j
3
˜ e
j
where, of course,
A
j
1
˜ e
j
= A
1
1
˜ e
1
+ A
2
1
˜ e
2
+ A
3
1
˜ e
3
,
A
j
2
˜ e
j
= A
1
2
˜ e
1
+ A
2
2
˜ e
2
+ A
3
2
˜ e
3
,
A
j
3
˜ e
j
= A
1
3
˜ e
1
+ A
2
3
˜ e
2
+ A
3
3
˜ e
3
.
Matching coeﬃcients of the ˜ e
i
we ﬁnd
˜ x
1
= x
1
A
1
1
+ x
2
A
1
2
+ x
3
A
1
3
= x
j
A
1
j
,
˜ x
2
= x
1
A
2
1
+ x
2
A
2
2
+ x
3
A
2
3
= x
j
A
2
j
,
˜ x
3
= x
1
A
3
1
+ x
2
A
3
2
+ x
3
A
3
3
= x
j
A
3
j
,
hence
˜ x
i
= x
j
A
i
j
. (2.10)
It is possible to obtain (2.10) from (2.9) in a succinct manner. On the right
hand side of (2.9) the index j is a dummy index which we can replace with
20 Tensor Analysis with Applications in Mechanics
i and thereby obtain (2.10) immediately. The matrix notation equivalent
of (2.10) is
⎛
⎝
˜ x
1
˜ x
2
˜ x
3
⎞
⎠
=
⎛
⎝
A
1
1
A
1
2
A
1
3
A
2
1
A
2
2
A
2
3
A
3
1
A
3
2
A
3
3
⎞
⎠
⎛
⎝
x
1
x
2
x
3
⎞
⎠
and thus involves multiplication by A
T
, the transpose of A.
We shall also need the equations of transformation from the frame ˜ e
i
back to the frame e
i
. Since the direct transformation is linear the inverse
must be linear as well, so we can write
˜ e
i
=
˜
A
j
i
e
j
(2.11)
where
˜
A
j
i
= ˜ e
i
e
j
.
Let us ﬁnd the relation between the matrices of transformation A and
˜
A.
By (2.11) and (2.8) we have
˜ e
i
=
˜
A
j
i
e
j
=
˜
A
j
i
A
k
j
˜ e
k
,
and since the ˜ e
i
form a basis we must have
˜
A
j
i
A
k
j
= δ
k
i
.
The relationship
A
j
i
˜
A
k
j
= δ
k
i
follows similarly. The product of the matrices (
˜
A
j
i
) and (A
k
j
) is the unit
matrix and thus these matrices are mutually inverse.
Exercise 2.7. Show that x
i
= ˜ x
k
˜
A
i
k
.
Formulas for the relations between reciprocal bases can be obtained as
follows. We begin with the obvious identities
e
j
(e
j
x) = x, ˜ e
j
(˜ e
j
x) = x.
Putting x = ˜ e
i
in the ﬁrst of these gives
˜ e
i
= A
i
j
e
j
,
while the second identity with x = e
i
yields
e
i
=
˜
A
i
j
˜ e
j
.
From these follow the transformation formulas
˜ x
i
= x
k
˜
A
k
i
, x
i
= ˜ x
k
A
k
i
.
Transformations and Vectors 21
2.5 Covariant and Contravariant Components
We have seen that if the basis vectors transform according to the relation
e
i
= A
j
i
˜ e
j
,
then the components x
i
of a vector x must transform according to
x
i
= A
j
i
˜ x
j
.
The similarity in form between these two relations results in the x
i
being
termed the covariant components of the vector x. On the other hand, the
transformation law
x
i
=
˜
A
i
j
˜ x
j
shows that the x
i
transform like the e
i
. For this reason the x
i
are termed
the contravariant components of x. We shall ﬁnd a further use for this
nomenclature in Chapter 3.
Quick summary
If frame transformations
e
i
= A
j
i
˜ e
j
,
˜ e
i
=
˜
A
j
i
e
j
,
e
i
=
˜
A
i
j
˜ e
j
,
˜ e
i
= A
i
j
e
j
,
are considered, then x has the various expressions
x = x
i
e
i
= x
i
e
i
= ˜ x
i
˜ e
i
= ˜ x
i
˜ e
i
and the transformation laws
x
i
= A
j
i
˜ x
j
,
˜ x
i
=
˜
A
j
i
x
j
,
x
i
=
˜
A
i
j
˜ x
j
,
˜ x
i
= A
i
j
x
j
,
apply. The x
i
are termed contravariant components of x, while the x
i
are
termed covariant components. The transformation laws are particularly
simple when the frame is changed to the dual frame. Then
x
i
= g
ji
x
j
, x
i
= g
ij
x
j
,
where
g
ij
= e
i
e
j
, g
ij
= e
i
e
j
,
are components of the metric tensor.
22 Tensor Analysis with Applications in Mechanics
2.6 The Cross Product in Index Notation
In mechanics a major role is played by the quantity called torque. This
quantity is introduced in elementary physics as the product of a force mag
nitude and a length (“force times moment arm”), along with some rules
for algebraic sign to account for the sense of rotation that the force would
encourage when applied to a physical body. In more advanced discussions
in which threedimensional problems are considered, torque is regarded as
a vectorial quantity. If a force f acts at a point which is located relative
to an origin O by position vector r, then the associated torque t about O
is normal to the plane of the vectors r and f . Of the two possible unit
normals, t is conventionally (but arbitrarily) associated with the vector ˆ n
given by the familiar righthand rule: if the foreﬁnger of the right hand is
directed along r and the middle ﬁnger is directed along f , then the thumb
indicates the direction of ˆ n and hence the direction of t. The magnitude
of t equals [f [[r[ sin θ, where θ is the smaller angle between f and r. These
rules are all encapsulated in the brief symbolism
t = r f .
The deﬁnition of torque can be taken as a model for a more general
operation between vectors: the cross product. If a and b are any two
vectors, we deﬁne
a b = ˆ n[a[[b[ sin θ
where ˆ n and θ are deﬁned as in the case of torque above. Like any other
vector, c = a b can be expanded in terms of a basis; we choose the
reciprocal basis e
i
and write
c = c
i
e
i
.
Because the magnitudes of a and b enter into ab in multiplicative fashion,
we are prompted to seek c
i
in the form
c
i
=
ijk
a
j
b
k
. (2.12)
Here the ’s are formal coeﬃcients. Let us ﬁnd them. We write
a = a
j
e
j
, b = b
k
e
k
,
and employ the wellknown distributive property
(u +v) w ≡ u w +v w
Transformations and Vectors 23
to obtain
c = a
j
e
j
b
k
e
k
= a
j
b
k
(e
j
e
k
).
Then
c e
i
= c
m
e
m
e
i
= c
i
= a
j
b
k
[(e
j
e
k
) e
i
]
and comparison with (2.12) shows that
ijk
= (e
j
e
k
) e
i
.
Now the value of (e
j
e
k
) e
i
depends on the values of the indices i, j, k.
Here it is convenient to introduce the idea of a permutation of the ordered
triple (1, 2, 3). A permutation of (1, 2, 3) is called even if it can be brought
about by performing any even number of interchanges of pairs of these
numbers; a permutation is odd if it results from performing any odd number
of interchanges. We saw before that (e
j
e
k
) e
i
equals the volume of the
frame parallelepiped if i, j, k are distinct and the ordered triple (i, j, k) is
an even permutation of (1, 2, 3). If i, j, k are distinct and the ordered triple
(i, j, k) is an odd permutation of (1, 2, 3), we obtain minus the volume of
the frame parallelepiped. If any two of the numbers i, j, k are equal we
obtain zero. Hence
ijk
=
⎧
⎪
⎪
⎨
⎪
⎪
⎩
+V, (i, j, k) an even permutation of (1, 2, 3),
−V, (i, j, k) an odd permutation of (1, 2, 3),
0, two or more indices equal.
Moreover, it can be shown (Exercise 2.4) that
V
2
= g
where g is the determinant of the matrix formed from the elements g
ij
=
e
i
e
j
of the metric tensor. Note that [V [ = 1 for a Cartesian frame.
The permutation symbol
ijk
is useful in writing formulas. For example,
the determinant of a matrix A = (a
ij
) can be expressed succinctly as
det A =
ijk
a
1i
a
2j
a
3k
.
Much more than a notational device however,
ijk
represents a tensor (the
socalled Levi–Civita tensor). We discuss this further in Chapter 3.
Exercise 2.8. The contravariant components of a vector c = ab can be
expressed as
c
i
=
ijk
a
j
b
k
24 Tensor Analysis with Applications in Mechanics
for suitable coeﬃcients
ijk
. Use the technique of this section to ﬁnd the
coeﬃcients. Then establish the identity
ijk
pqr
=
δ
p
i
δ
q
i
δ
r
i
δ
p
j
δ
q
j
δ
r
j
δ
p
k
δ
q
k
δ
r
k
and use it to show that
ijk
pqk
= δ
p
i
δ
q
j
− δ
q
i
δ
p
j
.
Use this in turn to prove that
a (b c) = b(a c) −c(a b) (2.13)
for any vectors a, b, c.
Exercise 2.9. Establish Lagrange’s identity
(a b) (c d) = (a c)(b d) − (a d)(b c).
2.7 Norms on the Space of Vectors
We often need to characterize the intensity of some vector ﬁeld locally or
globally. For this, the notion of a norm is appropriate. The wellknown
Euclidean norm of a vector a = a
k
i
k
written in a Cartesian frame is
a =
3
¸
k=1
a
2
k
1/2
.
This norm is related to the inner product of two vectors a = a
k
i
k
and
b = b
k
i
k
: we have a b = a
k
b
k
so that
a = (a a)
1/2
.
In a nonCartesian frame, the components of a vector depend on the
lengths of the frame vectors and the angles between them. Since the sum
of squared components of a vector depends on the frame, we cannot use it to
characterize the vector. But the formulas connected with the dot product
are invariant under change of frame, so we can use them to characterize
the intensity of the vector — its length. Thus for two vectors x = x
i
e
i
and
y = y
j
e
j
written in the arbitrary frame, we can introduce a scalar product
(i.e., a simple dot product)
x y = x
i
e
i
y
j
e
j
= x
i
y
j
g
ij
= x
i
y
j
g
ij
= x
i
y
i
.
Transformations and Vectors 25
Note that only in mixed coordinates does this resemble the scalar product
in a Cartesian frame. Similarly, the norm of a vector x is
x = (x x)
1/2
=
x
i
x
j
g
ij
1/2
=
x
i
x
j
g
ij
1/2
=
x
i
x
i
1/2
.
This dot product and associated norm have all the properties required from
objects of this nature in algebra or functional analysis. Indeed, it is neces
sary only to check whether all the axioms of the inner product are satisﬁed.
(i) x x ≥ 0, and x x = 0 if and only if x = 0. This property holds
because all the quantities involved can be written in a Cartesian
frame where it holds trivially. By the same reasoning, we conﬁrm
satisfaction of the property
(ii) x y = y x. The reader should check that this holds for any
representation of the vectors. Finally,
(iii) (αx +βy) z = α(x z) +β(y z) where α and β are arbitrary real
numbers and z is a vector.
By the general theory then, the expression
x = (x x)
1/2
(2.14)
satisﬁes all the axioms of a norm:
(i) x ≥ 0, with x = 0 if and only if x = 0.
(ii) αx = [α[ x for any real α.
(iii) x +y ≤ x + y.
In addition we have the Schwarz inequality
x y ≤ x y , (2.15)
where in the case of nonzero vectors the equality holds if and only if x = λy
for some real λ.
The set of all threedimensional vectors constitutes a threedimensional
linear space. A linear space equipped with the norm (2.14) becomes a
normed space. In this book, the principal space is R
3
. Note that we can
introduce more than one norm in any normed space, and in practice a
variety of norms turn out to be necessary. For example, 2 x is also a
norm in R
3
. We can introduce other norms, quite diﬀerent from the above.
One norm can be introduced as follows. Let e
k
be a basis of R
3
and let
26 Tensor Analysis with Applications in Mechanics
x = x
k
e
k
. For p ≥ 1, we introduce
x
p
=
3
¸
k=1
[x
k
[
p
1/p
.
Norm axioms (i) and (ii) obviously hold. Axiom (iii) is a consequence of the
classical Minkowski inequality for ﬁnite sums. The reader should be aware
that this norm is given in a certain basis. If we use it in another basis, the
value of the norm of a vector will change in general. An advantage of the
norm (2.14) is that it is independent of the basis of the space.
Later, when investigating the eigenvalues of a tensor, we will need a
space of vectors with complex components. It can be introduced similarly
to the space of complex numbers. We start with the space R
3
having basis
e
k
, and introduce multiplication of vectors in R
3
by complex numbers. This
also yields a linear space, but it is complex and denoted by C
3
. An arbitrary
vector x in C
3
takes the form
x = (a
k
+ ib
k
)e
k
,
where i is the imaginary unit (i
2
= −1). Analogous to the conjugate
number is the conjugate vector to x, deﬁned by
x = (a
k
− ib
k
)e
k
.
The real and imaginary parts of x are a
k
e
k
and b
k
e
k
, respectively. Clearly,
a basis in C
3
may contain vectors that are not in R
3
. As an exercise, the
reader should write out the form of the real and imaginary parts of x in
such a basis.
In C
3
, the dot product loses the property that x x ≥ 0. However, we
can introduce the inner product of two vectors x and y as
'x, y` = x y.
It is easy to see that this inner product has the following properties. Let
x, y, z be arbitrary vectors of C
3
. Then
(i) x x ≥ 0, and x x = 0 if and only if x = 0.
(ii) x y = y x.
(iii) (αx + βy) z = α(x z) + β(y z) where α and β are arbitrary
complex numbers.
The reader should verify these properties. Now we can introduce the norm
related to the inner product,
x = 'x, x`
1/2
,
Transformations and Vectors 27
and verify that it satisﬁes all the axioms of a norm in a complex linear
space. As a consequence of the general properties of the inner product,
Schwarz’s inequality (2.15) also holds in C
3
.
2.8 Closing Remarks
We close by repeating something we said in Chapter 1:
A vector is an objective entity.
In elementary mathematics we learn to think of a vector as an ordered triple
of components. There is, of course, no harm in this if we keep in mind a
certain Cartesian frame. But if we ﬁx those components then in any other
frame the vector is determined uniquely. Absolutely uniquely! So a vector
is something objective, but as soon as we specify its components in one
frame we can ﬁnd them in any other frame by the use of certain rules.
We emphasize this because the situation is exactly the same with ten
sors. A tensor is an objective entity, and ﬁxing its components relative to
one frame, we determine the tensor uniquely — even though its components
relative to other frames will in general be diﬀerent.
2.9 Problems
2.1 Find the dual basis to e
i
.
(a) e
1
= 2i
1
+i
2
−i
3
, e
2
= 2i
2
+ 3i
3
, e
3
= i
1
+i
3
;
(b) e
1
= i
1
+ 3i
2
+ 2i
3
, e
2
= 2i
1
− 3i
2
+ 2i
3
, e
3
= 3i
1
+ 2i
2
+ 3i
3
;
(c) e
1
= i
1
+i
2
, e
2
= i
1
−i
2
, e
3
= 3i
3
;
(d) e
1
= cos φi
1
+ sin φi
2
, e
2
= −sinφi
1
+ cos φi
2
, e
3
= i
3
.
2.2 Let
˜ e
1
= −2i
1
+ 3i
2
+ 2i
3
, e
1
= 2i
1
+i
2
−i
3
,
˜ e
2
= −2i
1
+ 2i
2
+i
3
, e
2
= 2i
2
+ 3i
3
,
˜ e
3
= −i
1
+i
2
+i
3
, e
3
= i
1
+i
3
.
Find the matrix A
j
i
of transformation from the basis ˜ e
i
to the basis e
j
.
28 Tensor Analysis with Applications in Mechanics
2.3 Let
˜ e
1
= i
1
+ 2i
2
, e
1
= i
1
− 6i
3
,
˜ e
2
= −i
2
−i
3
, e
2
= −3i
1
− 4i
2
+ 4i
3
,
˜ e
3
= −i
1
+ 2i
2
− 2i
3
, e
3
= i
1
+i
2
+i
3
.
Find the matrix of transformation of the basis ˜ e
i
to e
j
.
2.4 Find
(a) a
j
δ
jk
,
(b) a
i
a
j
δ
i
j
,
(c) δ
i
i
,
(d) δ
ij
δ
jk
,
(e) δ
ij
δ
ji
,
(f) δ
j
i
δ
k
j
δ
i
k
.
2.5 Show that
ijk
ijl
= 2δ
l
k
.
2.6 Show that
ijk
ijk
= 6.
2.7 Find
(a)
ijk
δ
jk
,
(b)
ijk
mkj
δ
i
m
,
(c)
ijk
δ
k
m
δ
j
n
,
(d)
ijk
a
i
a
j
,
(e)
ijk
[
ijk
[,
(f)
ijk
imn
δ
j
m
.
2.8 Find (a b) c.
2.9 Show that (a b) a = 0.
2.10 Show that a (b c)d = (a d)b c + (b d)c a + (c d)a b.
2.11 Show that (e a) e = a if [e[ = 1 and e a = 0.
2.12 Let e
k
be a basis of R
3
, let x = x
k
e
k
, and suppose h
1
, h
2
, h
3
are ﬁxed
positive numbers. Show that h
k
[x
k
[ is a norm in R
3
.
Chapter 3
Tensors
3.1 Dyadic Quantities and Tensors
We have met sets of quantities like g
ij
or g
ij
. Such a table of 3 × 3 = 9
coeﬃcients could be considered as a vector in a ninedimensional space, but
we must reject this idea for an important reason: if we change the frame
vectors and calculate the relations between the new and old components,
the results diﬀer in form from those that apply to vector components. The
components of the metric tensor transform according to certain rules, how
ever, and it is found that these transformation rules also apply to various
quantities encountered in physical science. We indicated in Chapter 1 that
these quantities, represented by 3×3 matrices, form a class of objects known
as secondorder tensors. Our plan is to present the relevant theory in a way
that parallels the vector presentation of Chapter 2.
We begin to realize this program with the introduction of the dyad (or
tensor product ) of two vectors a and b, denoted a⊗b. We assume that the
tensor product satisﬁes many usual properties of a product:
(λa) ⊗b = a ⊗(λb) = λ(a ⊗b),
(a +b) ⊗c = a ⊗c +b ⊗c,
a ⊗(b +c) = a ⊗b +a ⊗c, (3.1)
where λ is an arbitrary real number. However, the tensor product is not
symmetric: if a is not proportional to b then a ⊗b = b⊗a. From now on,
we shall write out the dyad without the ⊗ symbol: ab = a ⊗b.
Let us once again consider the space of threedimensional vectors with
the frame e
i
. Using the expansion of the vectors in the basis vectors and
the properties (3.1), we represent the dyad ab as
ab = a
i
e
i
b
j
e
j
= a
i
b
j
e
i
e
j
.
29
30 Tensor Analysis with Applications in Mechanics
This introduces exactly nine diﬀerent dyads e
i
e
j
. We now consider a linear
space whose basis is this set of nine dyads and call it the space of second
order tensors (or tensors of order two). The numerical coeﬃcients of the
dyads are called the components of the tensor. Thus an element of this
space, a tensor A, has the representation
A = a
ij
e
i
e
j
.
To maintain the property of objectivity of the elements of this space, we
require that upon transformation of the frame the components of A trans
form correspondingly. Note that we have introduced superscript indices for
the components of A. This was done in keeping with the development of
Chapter 2.
In preparation for the next section let us introduce the dot product of
a dyad ab by a vector c:
ab · c = (b · c)a. (3.2)
So the result is a vector cooriented with a. Analogously we can introduce
the dot product from the left:
c · ab = (c · a)b. (3.3)
Exercise 3.1. (a) A dyad of the form ee, where e is a unit vector, is
sometimes called a projection dyad. Explain. (b) Write down matrices for
the dyads i
1
i
1
, i
2
i
2
, and i
3
i
1
.
3.2 Tensors From an Operator Viewpoint
An alternative to viewing a secondorder tensor as a weighted sum of dyads
is to view the tensor as an operator. From this standpoint a tensor A is
considered to map a vector x into a vector y according to the equation
y = A· x.
Conversely, a given linear relation between x and y will deﬁne the operator
A uniquely. Thus if we have A· x = B· x for all x, then we have A = B.
Let us show that the components are really uniquely deﬁned in any basis
by the equality y = A· x. The tensor A is represented by the expression
a
ij
e
i
e
j
in some basis e
i
. It is clear that the operation A· x is linear in x, so
we deﬁne A uniquely if we specify its action on all three vectors of a basis.
Taking x = e
k
, the corresponding y is
A· x = a
ij
e
i
e
j
· e
k
= a
ik
e
i
.
Tensors 31
Dot multiplying this by e
l
we get
a
lk
= e
l
· A· e
k
.
In this way we can ﬁnd the components of a tensor A in any basis:
a
ij
= e
i
· A· e
j
, a
i
·j
= e
i
· A· e
j
,
a
ij
= e
i
· A· e
j
, a
·j
i
= e
i
· A· e
j
.
Note that in “mixed components” we position the indices in such a way
that their association with the various dyads remains clear.
Analyzing the above reasoning, we can ﬁnd that we have proved the
quotient law for tensors of order two. If y is a given vector and there is a
linear transformation from x to y for an arbitrary vector x, then the linear
transformation is a tensor and we can write y = A· x. This statement is
sometimes useful in establishing the tensorial character of a set of scalar
quantities (i.e., the components of A).
We may also deﬁne common algebraic operations from the operator
viewpoint. Given tensors A and B, the sum is the tensor A+B uniquely
deﬁned by the requirement that
(A+B) · x = A· x +B· x
for all x. If c is a scalar, cA is deﬁned by the requirement that
(cA) · x = c(A· x)
for all x. In particular, any product of the form 0A gives a zero tensor
denoted 0. The dot product A · B is regarded as the composition of the
operators B and A:
(A· B) · x ≡ A· (B· x).
The dot product y · A, called premultiplication of A by a vector y, is
deﬁned by the requirement that
(y · A) · x = y · (A· x)
for all vectors x.
A simple but important tensor is the unit tensor denoted by E and
deﬁned by the requirement that for any x
E · x = x · E = x. (3.4)
32 Tensor Analysis with Applications in Mechanics
It is evident that in any Cartesian frame i
i
we must have
E =
3
¸
i=1
i
i
i
i
. (3.5)
In any frame we have
E = e
i
e
i
= e
j
e
j
(3.6)
for the mixed components. Consequently, the raising and lowering of indices
gives
E = g
ij
e
i
e
j
= g
ij
e
i
e
j
(3.7)
in nonmixed components. We see that the role of the unit tensor belongs
to the metric tensor! Throughout our discussion of secondorder tensors we
shall emphasize the close analogy between tensor theory and matrix theory.
Equations (3.5) and (3.6) show that the matrix representation of E in either
Cartesian or mixed components is the 3 ×3 identity matrix
⎛
⎝
1 0 0
0 1 0
0 0 1
⎞
⎠
.
This does not hold for the nonmixed components of (3.7).
Exercise 3.2. Use (3.4) along with (2.4) and (2.5) to show that the various
components of E are given by
e
ij
= g
ij
, e
i
·j
= δ
i
j
,
e
ij
= g
ij
, e
·i
j
= δ
i
j
.
Hence establish (3.6).
Our consideration of A as an operator leads us to introduce the notion
of an inverse tensor: if
A· A
−1
= E,
then A
−1
is called the inverse of A. The inverse of a tensor is also a tensor.
An important special case occurs when the matrix of the tensor is a diagonal
matrix. If in a Cartesian frame i
i
we have
A =
3
¸
i=1
λ
i
i
i
i
i
Tensors 33
then the corresponding matrix representation is
⎛
⎝
λ
1
0 0
0 λ
2
0
0 0 λ
3
⎞
⎠
.
If we take
B =
3
¸
j=1
λ
−1
j
i
j
i
j
to which there corresponds the matrix
⎛
⎝
λ
−1
1
0 0
0 λ
−1
2
0
0 0 λ
−1
3
⎞
⎠
and form the dot product A· B, we get
A· B =
3
¸
i=1
λ
i
i
i
i
i
·
3
¸
j=1
λ
−1
j
i
j
i
j
=
3
¸
i=1
λ
i
i
i
3
¸
j=1
λ
−1
j
(i
i
· i
j
)i
j
=
3
¸
i=1
λ
i
i
i
λ
−1
i
i
i
=
3
¸
i=1
i
i
i
i
= E.
This means that B = A
−1
. Correspondingly,
⎛
⎝
λ
1
0 0
0 λ
2
0
0 0 λ
3
⎞
⎠
⎛
⎝
λ
−1
1
0 0
0 λ
−1
2
0
0 0 λ
−1
3
⎞
⎠
=
⎛
⎝
1 0 0
0 1 0
0 0 1
⎞
⎠
.
Exercise 3.3. Establish the formula
(A· B)
−1
= B
−1
· A
−1
for invertible tensors A, B.
34 Tensor Analysis with Applications in Mechanics
A secondorder tensor A is singular if A · x = 0 for some x = 0.
Hence A is nonsingular if A · x = 0 only when x = 0. Recall that a
matrix A is said to be nonsingular if and only if det A = 0. The connection
between the uses of this terminology in the two areas is as follows. If we
take a mixed representation of the tensor A, the equation A· x = 0 yields
a set of simultaneous equations in the components of x; these equations
have a nontrivial solution if and only if the determinant of the coeﬃcient
matrix (i.e., the matrix representing A) is zero. Moreover, taking any other
representation of A and a dual representation of the vector, we again arrive
the same conclusion regarding the determinant. This brings the use of the
term “singular” to the tensor A. By deﬁnition the determinant of a second
order tensor A, denoted det A, is the determinant of the matrix of its mixed
components:
det A = a
·j
i
 = a
k
· m
 =
1
g
a
st
 = ga
pq
.
The ﬁrst equality is the deﬁnition as stated above; the rest are left for the
reader to establish. Various other formulas such as
det A =
1
6
ijk
mnp
a
· i
m
a
· j
n
a
· k
p
can be established for the determinant.
We close this section with an important remark. We can derive all de
sired properties of a tensor, and perform actions with the tensor, in any co
ordinate frame. Convenience will often dictate the use of Cartesian frames.
But if we obtain an equation or expression through the use of a Cartesian
frame and can subsequently represent this result in noncoordinate form,
then we have provided rigorous justiﬁcation of the latter. As we have said
before, tensors are objective entities and ultimately all results pertaining
to them must be frame independent.
3.3 Dyadic Components Under Transformation
The standpoint for deriving the transformation rules is that in any basis a
tensor is the same element of some space, and only (3.1) and the rules we
derived for vectors can govern the rules for transforming the components
of a tensor. Let us begin with the transformation of the components when
we go to the reciprocal basis. We set
a
ij
e
i
e
j
= a
ij
e
i
e
j
Tensors 35
and take dot products as in (3.2) and (3.3):
e
k
· a
ij
e
i
e
j
· e
m
= e
k
· a
ij
e
i
e
j
· e
m
.
This gives
a
ij
(e
k
· e
i
)(e
j
· e
m
) = a
ij
(e
k
· e
i
)(e
j
· e
m
),
hence
a
km
= a
ij
g
ki
g
jm
.
We see that the components of the metric tensor are encountered in this
transformation.
Now we can construct the formulas for transforming the tensor compo
nents when the change of basis takes the general form
e
i
= A
j
i
˜ e
j
.
From
˜ a
ij
˜ e
i
˜ e
j
= a
km
e
k
e
m
= a
km
A
p
k
˜ e
p
A
q
m
˜ e
q
we obtain
˜ a
ij
= a
km
A
i
k
A
j
m
. (3.8)
Similarly, the inverse transformation
˜ e
i
=
˜
A
j
i
e
j
leads to
a
ij
= ˜ a
km
˜
A
i
k
˜
A
j
m
. (3.9)
Equations (3.8) and (3.9) together imply that
A
k
j
˜
A
i
k
= δ
i
j
.
Various expressions for A,
A = a
ij
e
i
e
j
= a
kl
e
k
e
l
= a
·j
i
e
i
e
j
= a
k
· l
e
k
e
l
= ˜ a
ij
˜ e
i
˜ e
j
= ˜ a
kl
˜ e
k
˜ e
l
= ˜ a
· j
i
˜ e
i
˜ e
j
= ˜ a
k
· l
˜ e
k
˜ e
l
,
lead to other transformation formulas such as
˜ a
ij
=
˜
A
k
i
˜
A
l
j
a
kl
, a
ij
= A
k
i
A
l
j
˜ a
kl
,
and
˜ a
· j
i
=
˜
A
k
i
A
j
l
a
· l
k
, ˜ a
i
· j
= A
i
k
˜
A
l
j
a
k
· l
.
Remembering the terminology of § 2.5, we see why the a
ij
are called the
covariant components of A while the a
ij
are called the contravariant com
ponents. The components a
i
·j
and a
·j
i
are called mixed components.
36 Tensor Analysis with Applications in Mechanics
Quick summary
We have
A = ˜ a
ij
˜ e
i
˜ e
j
= ˜ a
kl
˜ e
k
˜ e
l
= ˜ a
· j
i
˜ e
i
˜ e
j
= ˜ a
k
· l
˜ e
k
˜ e
l
= a
ij
e
i
e
j
= a
kl
e
k
e
l
= a
·j
i
e
i
e
j
= a
k
· l
e
k
e
l
where
˜ a
ij
= A
i
k
A
j
l
a
kl
, a
ij
=
˜
A
i
k
˜
A
j
l
˜ a
kl
,
˜ a
ij
=
˜
A
k
i
˜
A
l
j
a
kl
, a
ij
= A
k
i
A
l
j
˜ a
kl
,
˜ a
i
· j
= A
i
k
˜
A
l
j
a
k
· l
, a
i
·j
=
˜
A
i
k
A
l
j
˜ a
k
· l
,
˜ a
· j
i
=
˜
A
k
i
A
j
l
a
· l
k
, a
·j
i
= A
k
i
˜
A
j
l
˜ a
· l
k
.
Exercise 3.4. (a) Express the transformation law
˜
b
ij
= A
i
k
A
j
m
b
km
in matrix notation. (b) Repeat for a transformation law of the form
˜
b
ij
=
˜
A
k
i
˜
A
m
j
b
km
.
Exercise 3.5. Our A
j
i
values give the transformation from one basis to
another; they deﬁne a transformation of the space that is an operator, and
hence a secondorder tensor. Write out the tensor for which the A
j
i
are
components. Repeat for the inverse transformation.
3.4 More Dyadic Operations
The dot product of two dyads ab and cd is deﬁned by
ab · cd = (b · c)ad.
The result is again a dyad, with a coeﬃcient b · c. Extensions of this and
the formulas (3.2) and (3.3) to operations with sums of dyads and vectors
(using (3.1) and the vectorial rules) gives us a number of rules which the
dot product obeys. Let A and B be dyads, a and b be vectors, and λ and
µ be any real numbers. Then
A· (λa + µb) = λA· a + µA· b,
(λA+ µB) · a = λA· a + µB· a.
Tensors 37
Similar identities hold for dot products taken in the opposite orders. These
results show that linearity may be assumed in working with these opera
tions.
Now let us pursue the close analogy between the dot product and matrix
multiplication. We begin with the simple case of a Cartesian frame. We
take a dyad A and a vector b and express these relative to a basis i
k
:
A = a
km
i
k
i
m
, b = b
j
i
j
.
Denoting
c = A· b (3.10)
we have
c
k
i
k
= a
km
i
k
i
m
· b
j
i
j
so that
c
k
=
3
¸
j=1
a
kj
b
j
.
(We have inserted the summation symbol because j stands in the upper
position twice and the summation convention would not apply.) Written
out, this is the system of three equations
c
1
= a
11
b
1
+ a
12
b
2
+ a
13
b
3
,
c
2
= a
21
b
1
+ a
22
b
2
+ a
23
b
3
,
c
3
= a
31
b
1
+ a
32
b
2
+ a
33
b
3
,
or
⎛
⎝
c
1
c
2
c
3
⎞
⎠
=
⎛
⎝
a
11
a
12
a
13
a
21
a
22
a
23
a
31
a
32
a
33
⎞
⎠
⎛
⎝
b
1
b
2
b
3
⎞
⎠
.
Here we have a matrix equation of the form
c = Ab (3.11)
where c and b are column vectors and A is a 3 × 3 matrix. The analogy
between the dot product and matrix multiplication is evident from (3.10)
and (3.11). This analogy extends beyond the conﬁnes of Cartesian frames.
Let us write, for example,
A = a
km
e
k
e
m
, b = b
j
e
j
.
38 Tensor Analysis with Applications in Mechanics
This time c = A· b gives
c
k
e
k
= a
km
e
k
e
m
· b
j
e
j
= a
km
e
k
δ
j
m
b
j
,
hence
c
k
= a
kj
b
j
.
The corresponding matrix equation is, of course,
⎛
⎝
c
1
c
2
c
3
⎞
⎠
=
⎛
⎝
a
11
a
12
a
13
a
21
a
22
a
23
a
31
a
32
a
33
⎞
⎠
⎛
⎝
b
1
b
2
b
3
⎞
⎠
.
With suitable understanding we could still write this as (3.11). Note what
happens when we express both the dyad and the vector in terms of covariant
components:
A = a
km
e
k
e
m
, b = b
j
e
j
.
We obtain
c
k
= a
km
g
mj
b
j
and the metric tensor appears. The corresponding matrix form is
⎛
⎝
c
1
c
2
c
3
⎞
⎠
=
⎛
⎝
a
11
a
12
a
13
a
21
a
22
a
23
a
31
a
32
a
33
⎞
⎠
⎛
⎝
g
11
g
12
g
13
g
21
g
22
g
23
g
31
g
32
g
33
⎞
⎠
⎛
⎝
b
1
b
2
b
3
⎞
⎠
.
Because the metric tensor can raise an index on a vector component, we
may also write these equations in the forms
c
k
= a
km
b
m
and
⎛
⎝
c
1
c
2
c
3
⎞
⎠
=
⎛
⎝
a
11
a
12
a
13
a
21
a
22
a
23
a
31
a
32
a
33
⎞
⎠
⎛
⎝
b
1
b
2
b
3
⎞
⎠
.
Let us examine the dot product between two dyads. There are various
possibilities for the components of the dyad
C = A· B,
depending on how we choose to express Aand B. If we use all contravariant
components and write
A = a
km
e
k
e
m
, B = b
km
e
k
e
m
,
Tensors 39
then
C = c
kn
e
k
e
n
where
c
kn
= a
km
g
mj
b
jn
.
Similarly, the use of all covariant components as in
A = a
km
e
k
e
m
, B = b
km
e
k
e
m
,
leads to
C = c
kn
e
k
e
n
where
c
kn
= a
km
g
mj
b
jn
.
Mixed components appear when we express
A = a
km
e
k
e
m
, B = b
km
e
k
e
m
.
Then
C = A· B = a
km
e
k
e
m
· b
jn
e
j
e
n
= a
km
δ
j
m
b
jn
e
k
e
n
= a
kj
b
jn
e
k
e
n
.
Deﬁning
c
k
· n
= a
kj
b
jn
we have
C = c
k
· n
e
k
e
n
.
We leave other possibilities to the reader as
Exercise 3.6. (a) Discuss how the formulation c
· n
k
= a
kj
b
jn
arises.
(b) Show how all the forms above correspond to matrix multiplication.
(c) What happens if mixed components are used on the righthand sides to
express A and B?
Another useful operation that can be performed between tensors is dou
ble dot multiplication. If ab and cd are dyads, we deﬁne
ab ·· cd = (b · c)(a · d).
That is, we ﬁrst dot multiply the near standing vectors, then the remaining
vectors, and thereby obtain a scalar as the result.
40 Tensor Analysis with Applications in Mechanics
Exercise 3.7. (a) Calculate A · · E if A is a tensor of order two. How
does this relate to the trace of the matrix that represents A in mixed
components? (b) Let A and B be tensors of order two. Write down several
diﬀerent component forms for the quantity A·· B.
Yet another operation is the scalar product of two secondorder tensors
A and B, denoted by A • B. This represents a natural extension of the
operation
ab • cd = (a · c)(b · d)
between two dyads ab and cd.
3.5 Properties of SecondOrder Tensors
Now we would like to consider in more detail those tensors that occur most
frequently in applications: tensors of order two. First we recall that such a
tensor is represented in dyadic form as
A = a
ij
e
i
e
j
.
Those who work in the applied sciences are probably more accustomed to
the matrix representation
⎛
⎝
a
11
a
12
a
13
a
21
a
22
a
23
a
31
a
32
a
33
⎞
⎠
. (3.12)
When we use the matrix form (3.12) the dyadic basis of the tensor remains
implicit. Of course when we use a unique, say Cartesian, frame for the
space of vectors, then it does not matter whether we show the dyads. The
correspondence between the dyadic and matrix representations suggests
that we can introduce many familiar ideas from the theory of matrices.
The tensor transpose
Let us begin with the notion of transposition. For a matrix A = (a
ij
) the
transposed matrix A
T
is
A
T
= (a
ji
).
Similarly we introduce the transpose operation for the tensor A:
A
T
= a
ji
e
i
e
j
. (3.13)
Tensors 41
This operation yields a new tensor, in each representation of which the
corresponding indices appear in reverse order:
A
T
= a
ji
e
i
e
j
= a
ji
e
i
e
j
= a
j
·i
e
i
e
j
= a
·i
j
e
i
e
j
.
A useful relation for any secondorder tensor A and any vector x is
A· x = x · A
T
. (3.14)
This follows when we write x = x
k
e
k
and use (3.13) to see that
A
T
= a
ij
e
j
e
i
.
Equation (3.14) can be used to deﬁne the transpose. Also note that
(A
T
)
T
= A
for any secondorder tensor A.
Exercise 3.8. Let A and B be tensors of order two. Demonstrate that
A• B = A·· B
T
= A
T
·· B.
Exercise 3.9. Let A be a secondorder tensor. Find A·· A
T
. Demonstrate
that A·· A
T
= 0 if and only if A = 0.
Exercise 3.10. (a) Show that if A and B are tensors of order two, then
(A· B)
T
= B
T
· A
T
.
(b) Let a and b be vectors and C be a tensor of order two. Show that
a · C
T
· b = b · C· a.
(c) Show that if Ais a nonsingular tensor of order two, then the components
of the tensor B = A
−1
are given by the formulas
b
·j
i
=
1
2 det A
ikl
jmn
a
· k
m
a
· l
n
.
(d) Verify the following relations:
det A
−1
= (det A)
−1
, (A· B)
−1
= B
−1
· A
−1
,
(A
T
)
−1
= (A
−1
)
T
, (A
−1
)
−1
= A.
42 Tensor Analysis with Applications in Mechanics
Tensors raised to powers
By analogy with matrix algebra we may raise a tensor to a positive integer
power:
A
2
= A· A, A
3
= A· A
2
, A
4
= A· A
3
,
and so on. Note that A
k
still represents a linear operator. Negative integer
powers are deﬁned by raising A
−1
to positive integer powers:
A
−2
= A
−1
· A
−1
, A
−3
= A
−2
· A
−1
, A
−4
= A
−3
· A
−1
,
and so on. These operations can be used to construct functions of tensors
using Taylor expansions of elementary functions. For example,
e
x
= 1 +
x
1!
+
x
2
2!
+
x
3
3!
+· · · .
By this we can introduce the exponential of the tensor A:
e
A
= E +
A
1!
+
A
2
2!
+
A
3
3!
+· · · .
The issue of convergence of such series is approached in a manner similar
to the absolute convergence of usual series, but with use of a norm of the
tensor A (see § 3.12). Note that e
A
represents a linear operator. We can
introduce other functions similarly. This technique is used in the study of
nonlinear elasticity, for example.
Symmetric and antisymmetric tensors
Among the class of all secondorder tensors, an important role is played
by the symmetric tensors. These include the strain and stress tensors of
the theory of elasticity. The tensor of inertia is symmetric, as is the metric
tensor. All these satisfy the relation
A = A
T
.
It follows from (3.14) that
A· x = x · A (3.15)
and
(A· x) · y = x · (A· y) (3.16)
if A is symmetric. The reader will recall that the unit tensor E satisﬁes a
relation of the form (3.15). A tensor A is said to be antisymmetric if
A = −A
T
.
Tensors 43
Exercise 3.11. Give the matrix forms corresponding to the cases of sym
metric and antisymmetric tensors. How many components can be indepen
dently speciﬁed for a symmetric tensor? For an antisymmetric tensor?
Both symmetric and antisymmetric tensors arise naturally in the phys
ical sciences. Their signiﬁcance is also shown by the following
Theorem 3.1. Any secondorder tensor can be decomposed as a sum of
symmetric and antisymmetric tensors:
A = B+C
where B = B
T
and C = −C
T
.
Proof. Take
B =
1
2
A+A
T
, C =
1
2
A−A
T
,
and check all the statements.
The dyad ab can be decomposed into symmetric and antisymmetric
parts as
ab =
1
2
(ab +ba) +
1
2
(ab −ba)
for example.
Exercise 3.12. Show that if A is symmetric and B is antisymmetric then
A·· B = 0.
Exercise 3.13. Demonstrate that the quadratic form x · A · x does not
change if the secondorder tensor A is replaced by its symmetric part.
Given an antisymmetric tensor C = c
ij
i
i
i
j
in a Cartesian frame, we can
construct a vector
ω = ω
k
i
k
according to the formulas
ω
1
= c
32
, ω
2
= c
13
, ω
3
= c
21
.
It is easy to verify directly that
C· x = ω ×x, x · C = x ×ω,
44 Tensor Analysis with Applications in Mechanics
where x is an arbitrary vector. These formulas are written in non
coordinate form so they hold in any frame. The reader can derive the
formulas for ω, which is called the conjugate vector, for an arbitrary frame.
The crossproducts of a tensor A = a
ij
e
i
e
j
and a vector x are deﬁned
by the formulas
A×x = a
ij
e
i
(e
j
×x), x ×A = a
ij
(x ×e
i
)e
j
.
Exercise 3.14. Show that C = E ×ω = ω ×E.
3.6 Eigenvalues and Eigenvectors of a SecondOrder Sym
metric Tensor
We now consider the question of which basis yields a tensor of simplest
form. As the analogous question in matrix theory relates to eigenvalues
and eigenvectors, we extend these notions to tensors. The pair
(λ, x) (x = 0)
is called an eigenpair if the equality
A· x = λx (3.17)
holds. Hence x is an eigenvector of A if A operates on x to give a vector
proportional to x. Equation (3.17) may also be written in the form
(A−λE) · x = 0.
Exercise 3.15. Find the eigenpairs of the dyad ab. Now try to position
an eigenvector on the left: x · ab = λx. You should ﬁnd that this diﬀers
from the previous eigenvector, so it makes sense to introduce left and right
eigenvectors. In the case of a symmetric tensor they coincide.
The eigenvalues of a secondorder tensor A are found as solutions of the
characteristic equation for A, which is derived as follows. In components
(3.17) becomes
a
ij
e
i
e
j
· x
k
e
k
= λx
i
e
i
or
a
ij
g
jk
x
k
e
i
= λx
k
δ
i
k
e
i
.
Writing this as
(a
i
·k
−λδ
i
k
)x
k
= 0,
Tensors 45
we have a system of three simultaneous equations in the three variables x
k
.
A nontrivial solution exists if and only if the determinant of the coeﬃcient
matrix vanishes:
a
1
· 1
−λ a
1
· 2
a
1
· 3
a
2
· 1
a
2
· 2
−λ a
2
· 3
a
3
· 1
a
3
· 2
a
3
· 3
−λ
= 0.
This is the characteristic equation
1
for A. Writing it in the form
−λ
3
+ I
1
(A)λ
2
−I
2
(A)λ + I
3
(A) = 0 (3.18)
we note that it is cubic in λ, hence there are at most three distinct eigen
values λ
1
, λ
2
, λ
3
. The coeﬃcients I
1
(A), I
2
(A), and I
3
(A) are called the
ﬁrst, second, and third invariants of A, and are expressed in terms of the
eigenvalues by the Vi`ete formulas
I
1
(A) = λ
1
+ λ
2
+ λ
3
,
I
2
(A) = λ
1
λ
2
+ λ
1
λ
3
+ λ
2
λ
3
,
I
3
(A) = λ
1
λ
2
λ
3
.
After representing the tensor in diagonal form it will be easy to see that
I
1
(A) and I
3
(A) are, respectively, the trace and determinant of the tensor
A. In fact,
I
1
(A) = tr A, I
2
(A) =
1
2
[tr
2
A−tr A
2
], I
3
(A) = det A.
In nonlinear elasticity, the invariants and eigenvalues of several tensors
play important roles in the formulation of various constitutive laws. See,
for example,
[
Lurie (1990); Lurie (2005); Ogden (1997)
]
.
Exercise 3.16. A tensor A, when referred to a certain Cartesian basis,
has matrix
⎛
⎝
1 0 1
2 −1 0
0 1 2
⎞
⎠
.
Find the ﬁrst, second, and third principal invariants of A.
1
Note that it is expressed in terms of mixed components of the tensor. However, it char
acterizes the properties of the tensor and has invariant properties since the eigenvalues
of a tensor do not depend on the coordinate frame in which they are obtained.
46 Tensor Analysis with Applications in Mechanics
In applications, the most important secondorder tensors are the real
valued symmetric tensors. These have special properties. For a realvalued
tensor that is considered as an operator in the complex space C
3
, we have
a formula analogous to (3.16):
(A· x) · y = x · (A· y). (3.19)
This will be used below. We recall that a bar over an expression denotes
complex conjugation.
Theorem 3.2. The eigenvalues of a real symmetric tensor are real. More
over, eigenvectors corresponding to distinct eigenvalues are orthogonal.
Proof. Let A be a real symmetric tensor and λ an eigenvalue of A that
corresponds to the eigenvector x = 0, so
A· x = λx.
Dotmultiply both sides of this equality by x. It follows that
λ =
(A· x) · x
x · x
.
Now we prove that λ is real. Indeed, x · x = x
2
is positive. To see that
(A· x) · x takes a real value, we write out
(A· x) · x = x · (A· x)
= (A· x) · x
(we have used a property of the inner product in a complex linear space,
and then (3.19)). Because the eigenvalues λ are real, the components of the
eigenvectors satisfy a linear system of simultaneous equations having real
coeﬃcients; hence they are real as well.
To prove the second part of the theorem, suppose that
A· x
1
= λ
1
x
1
, A· x
2
= λ
2
x
2
,
where λ
2
= λ
1
. From these we obtain
λ
1
x
1
· x
2
= (A· x
1
) · x
2
, λ
2
x
2
· x
1
= (A· x
2
) · x
1
,
and subtraction gives
(λ
1
−λ
2
)x
1
· x
2
= x
2
· A· x
1
−x
1
· A· x
2
= 0.
It follows that
x
1
· x
2
= 0. (3.20)
Tensors 47
From (3.20) we may obtain another property of the eigenvectors x
1
, x
2
,
known as generalized orthogonality:
x
1
· A· x
2
= 0.
This holds for the eigenvectors corresponding to diﬀerent eigenvalues of a
real symmetric tensor A, and is useful in applications.
Exercise 3.17. Show that for any secondorder tensor A (not necessarily
symmetric), eigenvectors corresponding to distinct eigenvalues are linearly
independent.
In solving the characteristic equation for λ, we may ﬁnd that there are
three distinct solutions or fewer than three. Note that if x is an eigenvector
of A then so is αx for any α = 0. In other words, an eigenvector is
determined up to a constant multiple. So when the eigenvalues are distinct
we can compose a Cartesian frame from the orthonormal eigenvectors x
k
and then express the tensor A in terms of its components a
ij
as
A =
¸
a
ij
x
i
x
j
.
Since the frame x
k
is Cartesian the reciprocal basis is the same, and we
may calculate the components of A from
a
ij
= x
i
· A· x
j
.
Since the x
i
are also eigenvectors we can write
A· x
j
= λ
j
x
j
,
and dot multiplication by x
i
from the left gives
x
i
· A· x
j
= x
i
· λ
j
x
j
= λ
j
δ
j
i
.
Hence the coeﬃcients of the dyads x
i
x
j
in A are nonzero only for those
coeﬃcients that lie on the main diagonal of the matrix representation of A;
moreover, these diagonal entries are the eigenvalues of A. We can therefore
write
A =
3
¸
i=1
λ
i
x
i
x
i
. (3.21)
This is called the orthogonal representation of A. The eigenvectors com
posing the coordinate frame give us the principal axes of A, and the process
of referring the tensor to its principal axes is known as diagonalization.
48 Tensor Analysis with Applications in Mechanics
When the characteristic equation of a tensor has fewer than three dis
tinct solutions for λ, then the repeated eigenvalue is said to be degenerate.
If A is symmetric, we may represent A in a Cartesian basis and apply facts
from the theory of symmetric matrices. Corresponding to a multiple root of
the characteristic equation we have a subspace of eigenvectors. If λ
1
= λ
2
is a double root, then the subspace is twodimensional and we can select an
orthonormal pair that is orthogonal to the third eigenvector (since λ
1
and
λ
3
are distinct). Such an eigenvalue corresponding to two linearly indepen
dent eigenvectors is regarded as a multiple eigenvalue (multiplicity two). If
λ
1
= λ
2
= λ
3
then any vector is an eigenvector, and choosing a Cartesian
frame we would have an eigenvalue of multiplicity three. However, this case
arises only when the tensor under consideration is proportional to the unit
tensor E. Such a tensor is called a ball tensor.
Exercise 3.18. Show directly that a secondorder tensor (not necessarily
symmetric) having three distinct eigenvalues cannot have more than one
linearly independent eigenvector corresponding to each eigenvalue.
3.7 The Cayley–Hamilton Theorem
The Cayley–Hamilton theorem states that every square matrix satisﬁes its
own characteristic equation. For example the 2 ×2 matrix
A =
a b
c d
has characteristic equation
a −λ b
c d −λ
= λ
2
−(a + d)λ + (ad −bc) = 0,
and the Cayley–Hamilton theorem tells us that A itself satisﬁes
A
2
−(a + d)A + (ad −bc)I = 0,
where I is the 2 ×2 identity matrix and the zero on the right side denotes
the 2×2 zero matrix. Similarly, the Cayley–Hamilton theorem for a second
order tensor A whose characteristic equation is given by (3.18) states that
A satisﬁes the equation
−A
3
+ I
1
(A)A
2
−I
2
(A)A+ I
3
(A)E = 0. (3.22)
This permits us to represent A
3
in terms of lower powers of A. Further
more, we may dot multiply (3.22) by A and thereby represent A
4
in terms
Tensors 49
of lower powers of A. It is clear that we could continue in this fashion and
eventually express any desired power of A in terms of E, A, and A
2
. This
is useful in certain applications (e.g., nonlinear elasticity) where functions
of tensors are represented approximately by truncated Taylor series.
It is easy to establish the Cayley–Hamilton theorem for the case of a
symmetric tensor. Such a tensor A has the representation (3.21), where
we redenote i
i
= x
i
because the eigenvectors x
i
constitute an orthonormal
basis:
A =
3
¸
i=1
λ
i
i
i
i
i
.
We get
A
2
= A· A =
3
¸
i=1
λ
i
i
i
i
i
·
3
¸
j=1
λ
j
i
j
i
j
=
3
¸
i,j=1
λ
i
i
i
λ
j
i
j
δ
i
j
=
3
¸
i=1
λ
2
i
i
i
i
i
.
Similarly
A
3
= A· A· A =
3
¸
i=1
λ
3
i
i
i
i
i
. (3.23)
Let us return to equation (3.18) written for the ith eigenvalue and put it
into the expression (3.23):
A
3
=
3
¸
i=1
[I
1
(A)λ
2
i
−I
2
(A)λ
i
+ I
3
(A)]i
i
i
i
= I
1
(A)
3
¸
i=1
λ
2
i
i
i
i
i
−I
2
(A)
3
¸
i=1
λ
i
i
i
i
i
+ I
3
(A)
3
¸
i=1
i
i
i
i
= I
1
(A)A
2
−I
2
(A)A+ I
3
(A)E,
as desired.
Exercise 3.19. Use the Cayley–Hamilton theorem to express A
3
in terms
of A
2
, A, and E if A = i
1
i
1
+i
2
i
1
+i
2
i
2
+i
3
i
2
.
3.8 Other Properties of SecondOrder Tensors
Tensors of rotation
A tensor Q of order two is said to be orthogonal if it satisﬁes the equality
Q· Q
T
= Q
T
· Q = E.
50 Tensor Analysis with Applications in Mechanics
We see that
Q
T
= Q
−1
for an orthogonal tensor. Furthermore,
det Q = ±1.
Indeed, det Q is determined by the determinant of the matrix of mixed
components of Q. Because of the properties of the determinant of a matrix
and the correspondence between tensors and matrices, we have for two
tensors A and B of order two
det(A· B) = det Adet B
and
det A = det A
T
.
Thus
1 = det E = det Qdet Q
T
= (det Q)
2
as desired. We call Q a proper orthogonal tensor if det Q = +1; we call Q
an improper orthogonal tensor if det Q = −1.
Exercise 3.20. (a) Show that the tensor Q = −i
1
i
1
+i
2
i
2
+i
3
i
3
is orthog
onal. Is it proper or improper? (b) Show that
q
ij
q
kj
= δ
k
i
if Q is orthogonal. (c) Show that if Q is orthogonal then so is Q
n
for every
integer n.
We now consider the orthogonal tensor Q as an operator in the space
of all vectors.
Theorem 3.3. The operator deﬁned by the orthogonal tensor Q preserves
the magnitudes of vectors and the angles between them.
Proof. Consider the result of application of Q to both the multipliers of
the inner product x · y:
(Q· x) · (Q· y) = (x · Q
T
) · (Q· y) = x · Q
T
· Q· y = x · E· y = x · y.
First we put y = x to see that Q preserves vector magnitudes: Q· x
2
=
x
2
. Thus, by deﬁnition of the dot product in terms of the cosine, angles
are also preserved.
Tensors 51
The action of Q amounts to a rotation of all vectors of the space. (More
precisely, this is the case for a proper orthogonal tensor; an improper tensor
also causes an axis reﬂection that changes the “handedness” of the frame.)
The situation is analogous to the case of a solid body where the position of
a point is deﬁned by a vector beginning at a ﬁxed origin of some frame and
ending at the point. Any motion of the solid with a ﬁxed point is a rotation
with respect to some axis by some angle. The equations of physics should
often be introduced in such a way that they are invariant under rotation
of the coordinate frame whose position is not determined in space. Such
invariance under rotation should be veriﬁed, and in large part this can be
done by showing that the application of Q to all the vectors of the relation
does not change the form of the relation.
Let us note that some quantities are always invariant under rotation.
One of them is the ﬁrst invariant I
1
(A) of a tensor. This quantity, also
called the trace of the tensor and denoted tr(A), is the sum of the diagonal
mixed components of A:
tr(A) = a
·i
i
.
The trace can be equivalently determined in the noncoordinate form
tr(A) = E·· A = E• A
(the reader should check this). To show invariance we consider the tensor
Q· A· Q
T
= Q· (a
ij
e
i
e
j
) · Q
T
= a
ij
(Q· e
i
)(e
j
· Q
T
)
= a
ij
(Q· e
i
)(Q· e
j
).
We see that this is a representation of the “rotated” tensor, derived as the
result of applying Q to each vector of the dyadic components of A. The
trace of the rotated tensor is given by
E·· (Q· A· Q
T
) = E·· (Q· a
·j
i
e
i
e
j
· Q
T
)
= E·· a
·j
i
(Q· e
i
)(Q· e
j
).
Under the action of Q the frame e
i
transforms to the frame Q· e
i
, to which
the reciprocal basis is Q· e
i
. This means that the unit tensor (the metric
tensor!) can be represented, in particular, as
E = (Q· e
i
)(Q· e
i
) = (Q· e
j
)(Q· e
j
).
52 Tensor Analysis with Applications in Mechanics
It follows that
E·· (Q· A· Q
T
) =
(Q· e
k
)(Q· e
k
)
·· a
·j
i
(Q· e
i
)(Q· e
j
)
= a
·j
i
(Q· e
k
) · (Q· e
i
)
(Q· e
k
) · (Q· e
j
)
= a
·j
i
(e
k
· Q
T
) · (Q· e
i
)
(e
k
· Q
T
) · (Q· e
j
)
= a
·j
i
e
k
· Q
T
· Q· e
i
e
k
· Q
T
· Q· e
j
= a
·j
i
e
k
· E· e
i
e
k
· E· e
j
= a
·j
i
e
k
· e
i
e
k
· e
j
= a
·j
i
δ
i
k
δ
k
j
= a
· k
k
= tr(A).
Another example demonstrates that under the transformation Q the
eigenvalues of a tensor A remain the same but the eigenvectors x
i
rotate.
Indeed, let
A· x
i
= λ
i
x
i
.
We show that Q· x
i
is an eigenvector of Q· A· Q
T
:
(Q· A· Q
T
) · (Q· x
i
) = Q· A· Q
T
· Q· x
i
= Q· A· E· x
i
= Q· (A· x
i
)
= Q· (λ
i
x
i
)
= λ
i
(Q· x
i
).
Let e be an axis of rotation deﬁned by an orthogonal tensor Q, and
let ω be the angle of rotation about e. It can be shown that the proper
orthogonal tensor has the representation
Q = Ecos ω + (1 −cos ω)ee − e ×Esinω.
Polar decomposition
A secondorder tensor A is nonsingular if
det A = 0
where A is the matrix of mixed components of A. It is possible to express
such a tensor as a product of a symmetric tensor and another tensor. A
Tensors 53
statement of this result, known as the polar decomposition theorem, requires
that we introduce some additional terminology.
A symmetric tensor is said to be positive deﬁnite if its eigenvalues are
all positive. By orthogonal decomposition such a tensor S may be written
in the form
S = λ
1
i
1
i
1
+ λ
2
i
2
i
2
+ λ
3
i
3
i
3
,
where all the λ
i
> 0 and the eigenvectors i
k
constitute an orthonormal
basis.
If A is nonsingular then the tensor A · A
T
is symmetric and positive
deﬁnite. Symmetry follows from the equation
(A· A
T
)
T
= (A
T
)
T
· A
T
= A· A
T
.
To see positive deﬁniteness, we begin with the deﬁnition
(A· A
T
) · x = λx
of an eigenvalue λ and dot with x from the left to get
λ =
x · (A· A
T
) · x
x
2
.
The numerator is positive because
x · (A· A
T
) · x = (x · A) · (A
T
· x) = (x · A) · (x · A) = x · A
2
.
Hence all the eigenvalues λ of A· A
T
are positive. The diagonalization of
A· A
T
now shows that it is positive deﬁnite.
With these facts in hand we may turn to our main result.
Theorem 3.4. Any nonsingular tensor A of order two may be written as
a product of an orthogonal tensor and a positive deﬁnite symmetric tensor.
The decomposition may be done in two ways: as a left polar decomposi
tion
A = S · Q (3.24)
or as a right polar decomposition
A = Q· S
. (3.25)
Here Q is an orthogonal tensor of order two, and S and S
are positive
deﬁnite and symmetric.
54 Tensor Analysis with Applications in Mechanics
Proof. Because A· A
T
is positive deﬁnite and symmetric, we have
A· A
T
= λ
1
i
1
i
1
+ λ
2
i
2
i
2
+ λ
3
i
3
i
3
,
where the i
k
are orthonormal. We deﬁne
S ≡ (A· A
T
)
1/2
=
λ
1
i
1
i
1
+
λ
2
i
2
i
2
+
λ
3
i
3
i
3
since the λ
i
are positive. We see that S
−1
exists and is equal to
S
−1
=
1
√
λ
1
i
1
i
1
+
1
√
λ
2
i
2
i
2
+
1
√
λ
3
i
3
i
3
.
Now we set
Q ≡ S
−1
· A.
To see that Q is orthogonal, we write
Q· Q
T
= (S
−1
· A) · (S
−1
· A)
T
= (S
−1
· A) · (A
T
· (S
−1
)
T
)
= S
−1
· (A· A
T
) · (S
T
)
−1
= S
−1
· S
2
· S
−1
= E.
Thus we have expressed A = S · Q as in (3.24). The validity of (3.25)
follows from deﬁning S
≡ Q
T
· S · Q.
Exercise 3.21. We have called a tensor positive deﬁnite if its eigenvalues
are all positive. An alternative deﬁnition is that A is positive deﬁnite if
x · (A· x) > 0 for all x = 0. Explain.
As an application of polar decomposition, let us show that a nonsingular
tensor A operates on the position vectors of the points on the unit sphere
to produce position vectors deﬁning the points of an ellipsoid. If r locates
any point on the unit sphere then
r · r = 1. (3.26)
Let x be the image of r under A:
x = A· r.
Tensors 55
Then r = A
−1
· x, and substitution into (3.26) along with (3.24) gives
1 = [(S · Q)
−1
· x] · [(S · Q)
−1
· x]
= x · {[(S · Q)
−1
]
T
· (S · Q)
−1
} · x
= x · {S
−1
· S
−1
} · x
(the reader can supply the missing details). Because
S =
λ
1
i
1
i
1
+
λ
2
i
2
i
2
+
λ
3
i
3
i
3
,
expansion in the Cartesian frame i
k
with use of
x =
3
¸
i=1
x
i
i
i
, S
−1
=
3
¸
j=1
1
λ
j
i
j
i
j
reduces the above equation x · {S
−1
· S
−1
} · x = 1 to the form
3
¸
i=1
1
λ
i
x
2
i
= 1.
This is the equation of an ellipsoid.
Polar decomposition provides the background for introducing measures
of deformation in nonlinear continuum mechanics.
Deviator and ball tensor representation
For A we can introduce the representation
A =
1
3
I
1
(A)E + dev A.
Such a representation is found useful in the theory of elasticity. Moreover,
it is used to formulate constitutive equations in the theories of plasticity,
creep, and viscoelasticity.
The tensor dev A is deﬁned by the above equality. It has the same
eigenvectors as A, but eigenvalues that diﬀer from the eigenvalues of A by
(1/3)I
1
(A):
˜
λ
i
= λ
i
−
1
3
tr A.
56 Tensor Analysis with Applications in Mechanics
3.9 Extending the Dyad Idea
Thirdorder tensors can be introduced in a way that parallels the introduc
tion of dyads in § 3.1. Using the tensor product as before, we introduce
triad quantities of the type
R = abc
where a, b, and c are vectors. Expanding these vectors in terms of a basis
e
i
we obtain
R = a
i
b
j
c
k
e
i
e
j
e
k
.
We then consider a linear space whose basis is the set of 27 quantities
e
i
e
j
e
k
and call it the space of thirdorder tensors. We continue to refer to
the numerical values a
i
b
j
c
k
as the tensor’s components. A general element
of this space, a tensor R of order three, has the representation
R = r
ijk
e
i
e
j
e
k
. (3.27)
The property of objectivity of R remains paramount, leading to the re
quirement that the components transform appropriately when we change
the frame. The now familiar procedure of setting
˜ r
ijk
˜ e
i
˜ e
j
˜ e
k
= r
mnp
e
m
e
n
e
p
under the change of frame
e
i
= A
j
i
˜ e
j
gives
˜ r
ijk
= r
mnp
A
i
m
A
j
n
A
k
p
— a direct extension of (3.8). As an alternative to the representation (3.27)
in contravariant components, we could use the covarianttype representa
tion
R = r
ijk
e
i
e
j
e
k
or either of the mixed representations
R = r
ij
· ·k
e
i
e
j
e
k
, R = r
i
·jk
e
i
e
j
e
k
.
These necessitate the respective transformation laws
˜ r
ijk
= r
mnp
˜
A
m
i
˜
A
n
j
˜
A
p
k
Tensors 57
and
˜ r
ij
··k
= r
mn
· · p
A
i
m
A
j
n
˜
A
p
k
, ˜ r
i
·jk
= r
m
· np
A
i
m
˜
A
n
j
˜
A
p
k
,
as is easily veriﬁed.
Dot products involving triads follow familiar rules. The dot product of
a triad with a vector is given by the formula
abc · x = ab(c · x),
while the double dot product of a triad with a dyad is given by
abc ·· xy = a(c · x)(b · y).
One may also deﬁne a triple dot product of a triad with another triad:
abc ··· xyz = (c · x)(b · y)(a · z).
The scalar product of triads is deﬁned by the rule
abc • xyz = (a · x)(b · y)(c · z).
The reader has surmised by now that the order of a tensor is always equal
to the number of free indices needed to specify its components. A vector,
for instance, is a tensor of order one. In Chapter 2 we also met a quantity
whose components are speciﬁed by three indices:
ijk
. This quantity, which
arose naturally in our discussion of the vector cross product, is known as
the Levi–Civita tensor and is given by
E =
ijk
e
i
e
j
e
k
.
Exercise 3.22. Verify that E = −E×E.
Exercise 3.23. Verify the following formulas for operations involving the
Levi–Civita tensor:
(a) E ··· zyx = x · (y ×z),
(b) E ·· xy = y ×x,
(c) E · x = −x×.
Note: The notation of (c) may require some explanation. The result of
applying the thirdorder tensor E to a vector x is a secondorder tensor
E · x. When E · x is applied to another vector y, it becomes equivalent
to the cross product y × x. This is the meaning of the right side of (c).
Although the notation is awkward, it is rare that the action of some tensor
can be described as the action of two vectors, and the development of a
special notation is unwarranted.
58 Tensor Analysis with Applications in Mechanics
3.10 Tensors of the Fourth and Higher Orders
We can obviously extend the present treatment to tensors of any desired
order. Let us illustrate the essential points using tensors of order four.
A fourthorder tensor C can be represented by several types of compo
nents:
c
ijkl
, c
·jkl
i
, c
· ·kl
ij
, c
· · · l
ijk
, c
ijkl
.
The ﬁrst and last are purely contravariant and purely covariant, respec
tively, while the other three are mixed with indices in various positions.
As before these components represent C with respect to various bases; for
instance,
C = c
ijkl
e
i
e
j
e
k
e
l
.
Dot products with vectors can be taken as before: the rule is that we simply
dot multiply the basis vectors positioned nearest to the dot. Carrying out
such operations we may obtain results of various kinds. A dot product of
a fourthorder tensor with a vector gives, for example,
R = C· x
= c
ijkl
e
i
e
j
e
k
e
l
· x
m
e
m
= c
ijkl
e
i
e
j
e
k
δ
m
l
x
m
= c
ijkl
x
l
e
i
e
j
e
k
— a tensor of order three. Similarly, a dot product between a thirdorder
tensor and a vector gives a secondorder tensor. Wherever the dot product
is utilized, it continues to enjoy the linearity properties stated earlier.
Double dot products also appear in applications. For example, in elas
ticity one encounters double dot products between the tensor of elastic
constants and the strain tensor. In generalized form Hooke’s law becomes
σ = C·· ε
where σ is the stress tensor, C is the tensor of elastic constants, and ε is
the strain tensor. The density of the function of internal (elastic) energy in
linear elasticity is
1
2
σ ·· ε =
1
2
(C·· ε) ·· ε.
Tensors 59
Because σ and ε are symmetric, the last expression can be put in a more
symmetrical form
1
2
ε ·· C·· ε
in which the result does not depend on the order of operations.
Isotropic tensors
In engineering, isotropic materials play an important role. These are the
materials whose properties are the same in all directions. Air, for example,
is isotropic: it is equally transparent in all directions. It is impossible to tell
whether a ball made of isotropic material has been rotated through some
angle. In mechanics, material properties are expressed via constitutive re
lations. From a mathematical point of view, a material is isotropic when its
constitutive equations are invariant with respect to certain transformations:
the rotations and mirror reﬂections.
First we consider the question of when various tensorial quantities can
be isotropic. We say that a tensor is isotropic if its individual components
are invariant under all possible rotations and mirror reﬂections in R
3
.
Any scalar quantity is isotropic. Clearly, the only isotropic vector is 0.
Let A be a secondorder tensor so that in a basis e
1
, e
2
, e
3
we have
A = a
ij
e
i
e
j
.
Recall that any rotation or mirror reﬂection of R
3
is uniquely deﬁned by
an orthogonal tensor Q. Let us apply Q to each vector of the basis:
˜ e
k
= Q· e
k
.
In the new basis we have
A = ˜ a
ij
˜ e
i
˜ e
j
.
Let A be isotropic. By the above deﬁnition we must have a
ij
= ˜ a
ij
. So
A = ˜ a
ij
˜ e
i
˜ e
j
= a
ij
Q· e
i
Q· e
j
= Q· (a
ij
e
i
e
j
) · Q
T
.
This means that A is isotropic if and only if the equation
A = Q· A· Q
T
holds for any orthogonal tensor Q.
60 Tensor Analysis with Applications in Mechanics
Common sense tells us that the metric tensor E, which is the unit tensor
as well, should be isotropic. Let us demonstrate this. For any orthogonal
Q we have
Q· Q
T
= E.
This can be rewritten as
E = Q· E· Q
T
.
Hence E is isotropic. If λ is a scalar, then clearly the ball tensor λE is
isotropic as well.
The following is, unfortunately, not a trivial exercise. It asserts that
any isotropic secondorder tensor takes the form λE.
Exercise 3.24. Show that A is an isotropic tensor of order two if and only
if it is a ball tensor: that is, A = λE for some scalar λ.
Under an orthogonal transformation Q of R
3
, a fourthorder tensor
C = c
ijmn
e
i
e
j
e
m
e
n
takes the form
C = ˜ c
ijmn
(Q· e
i
)(Q· e
j
)(Q· e
m
)(Q· e
n
).
By the general deﬁnition, it is isotropic if ˜ c
ijmn
= c
ijmn
for all Q.
In a Cartesian frame, the general form of the fourthorder isotropic
tensor is
αEE+ βe
k
Ee
k
+ γI, (3.28)
where α, β, γ are arbitrary scalars
[
Jeﬀreys (1931)
]
. The proof is cumber
some and we omit it. The properties of the tensors in the representation
are exhibited in Exercises 3.27, 3.28, and 3.29. This fact can be applied to
the tensor of elastic constants for an isotropic material. The quantities C,
σ, and ε will be considered further in Chapter 6.
3.11 Functions of Tensorial Arguments
The reader is familiar with the notion of a function f(x
1
, . . . , x
n
) in n
variables. If we regard x
k
as a Cartesian component of a vector
x = (x
1
, . . . , x
n
),
Tensors 61
then we can regard f as a function of the vectorial argument x:
f = f(x).
But a logically good deﬁnition of such a function dictated by physics should
require that f be independent of the representation of x in a basis.
Similarly, we can consider a function of one or more tensorial arguments.
In a ﬁxed basis, such a function reduces to a function in many variables,
the components of the tensorial arguments. Again, however, a true function
of a tensorial argument cannot depend on the basis representations of its
arguments.
To extend this notion further, we can consider functions that take values
in the set of vectors — or even tensors. Such functions arise in applications.
For example, a force vector f = f (t) can be given as a function of time t.
Later, we will encounter other functions that take values in the set of tensors
of some order. Depending on this latter set, the function may be termed
scalarvalued, vectorvalued, or tensorvalued.
As for any function in many variables, we can apply the tools of calculus
to tensorvalued functions. These include the notion of continuity, the ﬁrst
diﬀerential, and derivatives. We will consider these topics later.
Linear functions
In linear elasticity and linear shell theory, linear relations and quadratic
functions (such as occur in strain energy expressions) play central roles.
We deﬁne a linear function of a tensorial variable as a function f which,
for any tensors A, B and scalars λ, µ, satisﬁes the relation
f(λA+ µB) = λf(A) + µf(B).
This mimics the deﬁnition of a linear matrix operator. From this point of
view, the equation y = kx + b represents a linear function only if b = 0.
Theorem 3.5. Let f be a scalarvalued function of a vectorial argument
x. There is a unique c such that for all x,
f(x) = c · x. (3.29)
Proof. We expand x = x
k
e
k
with respect to the basis e
k
. By linearity,
f(x) = x
k
f(e
k
). Equation (3.29) holds with c = f(e
k
)e
k
. Supposing the
existence of two vectors c
1
and c
2
such that c
1
· x = c
2
· x, and putting
x = c
2
−c
1
, we get c
1
= c
2
.
62 Tensor Analysis with Applications in Mechanics
Now let us consider a scalarvalued function whose argument is a second
order tensor A. Clearly, for a secondorder tensor B, the function
f(A) = tr(B· A) = B·· A
T
is linear.
Exercise 3.25. Show that B·· A
T
= A·· B
T
.
Theorem 3.6. Let f be a scalarvalued function of a secondorder tensor
A. There is a unique secondorder tensor B such that for all A,
f(A) = tr(B· A
T
). (3.30)
The proof is left to the reader.
The representation of any tensorvalued function of a tensorial argument
is similar. We use it to introduce the tensor of elastic constants in linear
elasticity.
Theorem 3.7. Let F = F(A) be a linear function from the set of second
order tensors A to the same set of secondorder tensors. There is a unique
fourthorder tensor C such that for all A,
F(A) = C·· A
T
. (3.31)
Proof. Write A = a
mn
e
m
e
n
and introduce C by the formula
C = F(e
m
e
n
)e
m
e
n
.
Then
C·· A
T
= F(e
m
e
n
)e
m
e
n
·· a
ij
e
j
e
i
= F(e
m
e
n
)a
mn
= F(A).
So the representation is valid. Proof of uniqueness is left to the reader.
Exercise 3.26. An operator on the set of vectors x is given by the formula
y = B· x where B is a secondorder tensor. This can be extended to the set
of secondorder tensors by the equation Y = B· X
T
. Show that by using
the fourthorder tensor C = B· e
n
Ee
n
, we get C·· X
T
= B· X
T
for all X.
Note that we cannot represent Hooke’s law using only this operation B· X.
Looking back, we note that the general form of a fourthorder isotropic
tensor contains three independent isotropic tensors. Their properties are
exhibited in the following exercises.
Exercise 3.27. Show that the identity operator from the representation
(3.31) is I = e
k
e
m
e
k
e
m
; that is, for all A we have I · ·A
T
= A.
Tensors 63
Exercise 3.28. A linear function is deﬁned by the equality F(A) = A
T
.
Show that the corresponding fourthorder tensor is e
k
Ee
k
, i.e., show that
e
k
Ee
k
·· A
T
= A
T
.
Exercise 3.29. A linear function is deﬁned by F(A) = (tr A)E. Show
that the corresponding fourthorder tensor is EE, i.e., show that
EE·· A
T
= (tr A)E.
Isotropic scalarvalued functions
A scalar function of a tensor is said to be isotropic if it retains its form
under any orthogonal transformation of the space (or equivalently, of its
basis).
A scalarvalued function f(A) of a secondorder argument A is isotropic
if and only if for any orthogonal tensor Q we have
f(A) = f
Q· A· Q
T
.
Let us demonstrate that any eigenvalue λ of a secondorder tensor A,
when considered as a scalarvalued function of A (i.e., λ = f(A)), is
isotropic. Indeed, an eigenvalue satisﬁes the characteristic equation
det(A−λE) = 0.
The eigenvalues of Q· A· Q
T
satisfy
det(Q· A· Q
T
−λE) = 0.
The equality
det(Q· A· Q
T
−λE) = det(Q· A· Q
T
−λQ· E· Q
T
)
= det
Q· (A−λE) · Q
T
= (det Q)
2
det(A−λE)
= det(A−λE)
shows that the eigenvalues of A and Q· A· Q
T
satisfy the same equation.
Hence they coincide.
Exercise 3.30. Show that the invariants I
1
(A), I
2
(A), I
3
(A) are isotropic
functions of A.
64 Tensor Analysis with Applications in Mechanics
It can be shown that any scalarvalued isotropic function of a second
order tensor is a function of its invariants
[
Lurie (1990); Ogden (1997);
Truesdell and Noll (2004)
]
. This is used in nonlinear elasticity to introduce
the constitutive equations for isotropic bodies.
By Theorem 3.6, a scalarvalued linear function has the representation
f(A) = B·· A
T
for some secondorder tensor B.
Theorem 3.8. The function f(A) = B·· A
T
is isotropic if and only if B
is an isotropic tensor of order two, and hence a ball tensor: B = λE.
Proof. Because f is isotropic, we have
B·· A
T
= B·· (Q· A· Q
T
)
T
for any orthogonal Q. Using
B·· A
T
= tr(B· A)
we get
B·· (Q· A· Q
T
)
T
= tr(B· Q· A· Q
T
)
= tr(Q
T
· B· Q· A)
= (Q
T
· B· Q) ·· A
T
.
Hence for any A we have
B·· A
T
= (Q
T
· B· Q) ·· A
T
.
But this occurs if and only if the relation
B = Q
T
· B· Q
holds for any Q. Therefore B is isotropic.
Exercise 3.31. Show that a scalarvalued, linear, isotropic function of a
secondorder tensor A is a linear function of tr A: that is, f(X) = λtr X.
If a scalar function maintains its form under some subgroup of orthog
onal transformations, then one can ﬁnd functions that are invariant under
these. The reader will ﬁnd applications of this idea in books on crystallog
raphy and elasticity.
Tensors 65
Isotropic tensorvalued functions
Now we consider a function F whose domain and range are the set of second
order tensors. We say that F(A) is isotropic if the components of its image
value do not change from one Cartesian basis to another. So F is isotropic
if
F
Q· A· Q
T
= Q· F(A) · Q
T
holds for any orthogonal tensor Q. An example of an isotropic tensorvalued
function is F = λA.
As with a scalarvalued function, it can be shown that a tensorvalued
linear function F represented in terms of a fourthorder tensor C,
F = C·· A
T
,
is isotropic if and only if C is isotropic and therefore is given by (3.28). So
F takes the form
F(A) = αEtr A+ βA
T
+ γA.
Mechanicists employ functions whose domains and ranges can be sets of
symmetric tensors. This imposes certain additional restrictions on the form
of C. Indeed, let A = A
T
and F(A) = F(A)
T
. Using the representation
F(A) = c
ijmn
a
mn
i
i
i
j
with a Cartesian basis i
i
, we see that
c
ijmn
= c
jimn
= c
ijnm
.
In the general case, C has 81 independent components. But, in view of
symmetry, C has only 36 independent components. An isotropic fourth
order tensor C satisfying the symmetry conditions takes the form
αEE+ β(e
k
Ee
k
+I).
So the general form of an isotropic linear function satisfying the symmetry
condition is
F(A) = αEtr A+ 2βA.
We will introduce the elements of calculus for functions of tensorial
arguments. First, we require a way to gauge the magnitude of a vector or
tensor. Suitable norms exist for this purpose.
66 Tensor Analysis with Applications in Mechanics
3.12 Norms for Tensors, and Some Spaces
In § 2.7 we introduced a norm in the space R
3
. It can be immediately
extended to R
k
for any k as
x = (x · x)
1/2
,
where the inner product of x and y in the space is given by x · y.
Similarly, we can introduce a norm and inner product in the set of
secondorder tensors. We denote the inner product by (A, B) and deﬁne it
using the dot product as
(A, B) = A·· B
T
= a
ij
e
i
e
j
·· b
ts
e
s
e
t
= a
ij
b
ts
g
js
g
it
= a
ij
b
ts
g
js
g
it
= a
ij
e
i
e
j
·· b
ts
e
s
e
t
= a
ij
b
ts
δ
s
j
δ
t
i
= a
ij
b
ij
.
It is clear that in a Cartesian frame (A, A) is the sum of all the squared
components of A, so this is quite similar to the scalar product of vectors.
Using the same reasoning as above, we can show that the axioms of the
scalar product hold here as well (note that using only a Cartesian frame,
we could regard the components of a tensor as those of a ninedimensional
vector, so this is another reason why the axioms hold). In this case the inner
product axioms are written for arbitrary secondorder tensors A, B, C as
(i) (A, A) ≥ 0, and (A, A) = 0 if and only if A = 0;
(ii) (A, B) = (B, A);
(iii) (αA+ βB, C) = α(A, C) + β(B, C) for any real α, β.
By linear algebra, the expression
A = (A, A)
1/2
is a norm in the set of secondorder tensors. It satisﬁes the following axioms.
(i) A ≥ 0, with A = 0 if and only if A = 0;
(ii) αA = α A for any real α;
(iii) A+B ≤ A +B.
Tensors 67
We still have the Schwarz inequality
(A, B) ≤ A B .
In linear algebra, many particular implementations of the vector and
tensor norms can be introduced. We can do the same in any ﬁxed frame, but
if we wish to change the frames under consideration we must remember that
these norms should change in accordance with the tensor transformation
rules.
Note that the representation of a linear function in the previous section
used A
T
as an argument. This reﬂects the form of the inner product on
the set of secondorder tensors.
It is worth noting two important properties of the norms we have intro
duced, as these are used in analysis. Let A and B be secondorder tensors
and x a vector. Then the relations
A· x ≤ A x
and
A· B ≤ A B
hold. The latter implies that
A
k
≤ A
k
. (3.32)
Using this property, we can justify the introduction of tensorvalued func
tions like e
A
.
Exercise 3.32. Using (3.32), prove convergence of the series
e
A
= E+
1
1!
A+· · · +
1
k!
A
k
+· · · .
It is easy to extend the notion of inner product to tensors of any order;
we introduce the inner product
(A, B) = a
i
1
i
2
···i
n
b
i
1
i
2
···i
n
.
The reader can represent this in all its particular forms and verify the inner
product axioms.
68 Tensor Analysis with Applications in Mechanics
Some elements of calculus
As in ordinary calculus, we can introduce the notion of a function in one
or many variables that takes values in a set of vectors or tensors. Such a
function will be given on a certain set. Consider, for example, a function
on the segment [a, b] with values in R
3
. This mapping pairs each point
of [a, b] with at most one vector from R
3
. We may similarly construct a
function from [a, b] to the set of tensors of some order. Furthermore, we
can introduce the notions of limit and continuity at a point t
0
in the same
way as in calculus:
The function f : [a, b] → R
3
has limit a at t = t
0
∈ [a, b] if for any
ε > 0 there is a δ > 0, dependent on ε, such that for any t = t
0
and
t − t
0
 ≤ δ we have f (t) −a < ε. When a = f (t
0
), we say the
function is continuous at t
0
.
Various norms can be used on R
3
. However, in linear algebra it is shown
that on a ﬁnitedimensional space all norms are equivalent. Equivalence of
two norms ·
1
and ·
2
means that there exist positive constants c
1
and
c
2
such that for any element x of the space we have
0 < c
1
≤
x
1
x
2
≤ c
2
< ∞,
where c
1
and c
2
do not depend on x. Consequently, either norm can be
used in the deﬁnition of limit: the limit will exist or not, independently of
the form taken by the norm.
As in calculus, we say that a function f (t) is continuous on [a, b] if it is
continuous at each point of [a, b]. When a function takes values in R
3
, the
ordinary deﬁnitions of limit, derivative, integral, etc., can be modiﬁed by
replacing the absolute value with a suitable norm. The derivative of f (t)
at t
0
is given by
f
(t
0
) =
df
dt
t=t
0
= lim
t→t
0
f (t) −f (t
0
)
t −t
0
.
The derivative of a vectorvalued function has many properties familiar from
ordinary calculus. For example, if f (t) and g(t) are both diﬀerentiable at
t, then the product rule holds in the form
(f (t) · g(t))
= f
(t) · g(t) +f (t) · g
(t).
If we expand f in a basis e
1
, e
2
, e
3
so that f (t) = f
k
(t)e
k
, then
df (t)
dt
=
df
k
(t)
dt
e
k
.
Tensors 69
Here we have assumed that the e
k
do not depend on t; otherwise an appli
cation of the product rule would have been required (we will encounter this
situation later). Deﬁnite integration can also be carried out in component
wise fashion:
b
a
f (t) dt =
b
a
f
k
(t) dt
e
k
.
The integral has all the properties familiar from calculus.
Clearly, all this can be extended to the case of a function in one scalar
variable taking values in the set of tensors of some order. Moreover, the
theory of functions in many variables is generalized in a similar way to the
tensorial functions. The reader should remember that formally, in all the
deﬁnitions of ordinary calculus, we should change the absolute value to the
norm. We record here only the formula for the derivative of a tensorial
function F(t) = f
mn
(t)e
m
e
n
:
dF(t)
dt
=
df
mn
(t)
dt
e
m
e
n
.
Some normed spaces
In textbooks on functional analysis, the norm of a vector function is usually
introduced in a Cartesian frame. For example, the norm on the space C(V )
of continuous vector functions given on a compact region V is
f (x)
C
= f
k
(x)i
k
C
= max
k
max
V
f
k
(x)
. (3.33)
This formula depends on the Cartesian frame i
k
. If we use a similar norm
involving vector components in a frame having singular points in V — as
may be the case with spherical coordinates — we obtain a norm that is
not equivalent to (3.33). This means that (3.33) is an improper way to
characterize the intensity of a vector ﬁeld.
However, the proper norm for a function given in curvilinear coordinates
is based on the above norm:
f (x) =
f
i
(x)r
i
= max
V
f
i
(x)f
i
(x)
1/2
. (3.34)
If we would like to use a norm of the type (3.33), we need to remember that
during transformation of the frame we must change the form of the norm
accordingly.
70 Tensor Analysis with Applications in Mechanics
On the set of secondorder tensor functions continuous on a compact
region V , we can introduce a norm similar to (3.34):
A(x) =
a
ij
(x)r
i
r
j
= max
V
a
ij
(x)a
ij
(x)
1/2
.
Finally, note that instead of the norms for continuous vector and tensor
functions we can introduce the norms and scalar products corresponding to
the space of scalar functions L
2
(V ). The inner product is then
(A, B) =
V
a
ij
b
ij
dV. (3.35)
The reader should verify all the axioms of the inner product for this, and
introduce all forms of the inner product and norm corresponding to (3.35).
Exercise 3.33. Let A be a tensor of order two with A = q < 1. Demon
strate that E−A has the inverse (E−A)
−1
that is equal to
E+A+A
2
+A
3
+· · · +A
n
+· · · .
Exercise 3.34. Let A be a tensor of order two with A = q < 1. What
is the inverse to E+A?
3.13 Diﬀerentiation of Tensorial Functions
In the linear theory of elasticity, there are functions that relate two ten
sors. The stress and strain tensors, for example, are related though the
generalized form of Hooke’s law. More complex relations occur in nonlin
ear elasticity or plasticity. Here one must diﬀerentiate tensorial functions.
In elasticity, for example, the stress tensor can be found as the derivative
of the strain energy with respect to the strain tensor.
We recall that for an ordinary function f(x), the derivative is
f
(x) = lim
∆x→0
f(x + ∆x) −f(x)
∆x
.
The ﬁrst diﬀerential is given by the formula
df = f
(x) dx. (3.36)
For a function in n variables we have
df(x
1
, . . . , x
n
) =
n
¸
k=1
∂f
∂x
k
dx
k
.
Tensors 71
Using a Cartesian basis i
1
, . . . , i
n
, we can formally represent this as
df(x
1
, . . . , x
n
) =
n
¸
k=1
∂f
∂x
k
i
k
·
n
¸
m=1
dx
m
i
m
.
Let us regard
x =
n
¸
m=1
x
m
i
m
as a vector, write
f(x
1
, . . . , x
m
) = f(x),
and consider this as a function in a vectorial variable. In the same way we
introduce
dx =
n
¸
m=1
dx
m
i
m
where the dx
m
are some quantities that are not necessarily inﬁnitesimal.
Let ε be a real variable. For ﬁxed x and dx, the function f(x + ε dx) is a
function in one variable ε. The chain rule formally applied to this function
gives us
df(x + ε dx)
dε
ε=0
=
n
¸
k=1
∂f
∂x
k
i
k
·
n
¸
m=1
dx
m
i
m
, (3.37)
hence
df(x
1
, . . . , x
n
) =
df(x + ε dx)
dε
ε=0
. (3.38)
The righthand side of this equality is termed the Gˆ ateaux derivative of
f at the point x in the direction dx. For the intermediate expression, we
introduce the notation
f
,x
=
n
¸
k=1
∂f
∂x
k
i
k
and call it the derivative of f with respect to x. Later we will refer to this
vector quantity as the gradient of f.
We began with a function in n variables, but could in fact consider a
function of a vectorial argument f(x) and present the same operations in
noncomponent form. From (3.38) and (3.37) it follows that
df(x
1
, . . . , x
n
) =
df(x + ε dx)
dε
ε=0
= f
,x
· dx.
72 Tensor Analysis with Applications in Mechanics
When we derive a relation in Cartesian coordinates but present it in non
component form, it becomes valid for any basis. The reader may wish to
verify this by direct calculation. Note that in any basis, the components
of f
,x
are partial derivatives of f with respect to the components of the
expansion of x in the basis. When we wish to have this in a form that does
not include the basis vectors, we must use the expansion of dx in the dual
basis.
These ideas can be extended to tensorial functions of tensorial argu
ments in a straightforward manner.
First we consider a scalarvalued function f(X) whose argument X be
longs to the secondorder tensors. In a ﬁxed basis, it can be considered as
a function in 3 × 3 = 9 variables, the components of X. As above, in a
Cartesian basis in R
3
we can introduce the ﬁrst diﬀerential df. Then we
introduce the Gˆateaux derivative:
∂
∂ε
f(X+ ε dX)
ε=0
≡ lim
ε→0
f(X+ ε dX) −f(X)
ε
= f
,X
·· dX
T
(3.39)
for any tensor dX, not necessarily inﬁnitesimal. Here, in the Cartesian
basis we have
f
,X
=
∂f
∂x
mn
i
m
i
n
,
where the x
mn
are the Cartesian components of X. Then
df = f
,X
·· dX
T
.
The expression f
,X
is called the derivative of f with respect to the tensor
argument X. Although the last formula was derived in Cartesian coordi
nates, it holds in any basis.
For a function that maps values X in the set of secondorder tensors
into the same set, the derivative F
,X
(X) is deﬁned as
F
,X
·· dX
T
=
∂
∂ε
F(X+ εdX)
ε=0
≡ lim
ε→0
F(X+ εdX) −F(X)
ε
. (3.40)
This is a particular case of the Gˆ ateaux derivative. Again, we can repeat
the method used above for the ﬁrst diﬀerential written in Cartesian com
ponents. The order of F
,X
(X) is four. In component form it is
F
,X
=
∂f
ij
∂x
mn
i
i
i
j
i
m
i
n
.
(We recall that in Cartesian bases i
k
= i
k
, and hence the summation con
vention applies in situations such as this.) For
f
,X
(X) and F
,X
(X),
Tensors 73
the notations
df(X)
dX
and
dF(X)
dX
are also used.
In a similar way, we can deﬁne the derivative of a tensorvalued function
for tensors of any order.
Clearly, when calculating dF we get a linear function in dX. This is
why we studied linear functions earlier.
Now we introduce a partial derivative for a scalarvalued function
f(X
1
, . . . , X
m
) in several tensorial arguments. Let the X
i
be secondorder
tensors. The partial derivative of f with respect to X
i
, denoted by ∂f/∂X
i
,
is deﬁned by the equality
∂f
∂X
i
·· Y
T
=
∂
∂ε
f(X
1
, . . . , X
i
+ εY, . . . , X
m
)
ε=0
for any secondorder tensor Y. When function F takes values in the set of
secondorder tensors, the partial derivative ∂F/∂X
i
is similarly deﬁned by
equality
∂F
∂X
i
·· Y
T
=
∂
∂ε
F(X
1
, . . . , X
i
+ εY, . . . , X
m
)
ε=0
.
As in case of a tensorial function of one variable, the components of the
above partial derivatives can be expressed as ordinary partial derivatives of
the components. We describe this representation in Cartesian coordinates.
Let
f(X
1
, . . . , X
m
) = f(x
(1)
ij
, . . . , x
(m)
ij
)
be a function in 9m variables, the components x
(k)
ij
of X
k
. Then
∂f
∂X
i
=
∂f
∂x
(i)
jk
i
j
i
k
.
Exercise 3.35. Find the derivative of f(X) = I
1
(X) ≡ tr X with respect
to X.
Exercise 3.36. Find the derivative of f(X) = tr X
2
.
Exercise 3.37. Using the method applied to the previous exercise, show
that the derivative of f(X) = tr X
3
with respect to X is 3(X
T
)
2
.
74 Tensor Analysis with Applications in Mechanics
Exercise 3.38. Show that the derivative of
f(X) = I
2
(X) ≡
1
2
[tr
2
X−tr X
2
]
with respect to X is
I
2
(X)
,X
= I
1
(X)E −X
T
.
Exercise 3.39. Show that the derivative of
f(X) = I
3
(X) ≡ det X
with respect to X is
[X
2
−I
1
(X)X+ I
2
(X)E]
T
.
The formulas for diﬀerentiating the invariants of the strain tensor are
used to write down the constitutive equation for a nonlinear elastic isotropic
material under ﬁnite deformation.
Exercise 3.40. The strain energy of an isotropic elastic medium is a func
tion of the invariants I
k
of the strain tensor, f = f(I
1
, I
2
, I
3
). Using the
results of Exercises 3.35, 3.36, and 3.39, demonstrate that its derivative is
f
,X
=
¸
∂f
∂I
1
+ I
1
∂f
∂I
2
+ I
2
∂f
∂I
3
E−
∂f
∂I
2
+ I
1
∂f
∂I
3
X
T
+
∂f
∂I
3
X
T
2
.
We have presented examples of the derivatives of scalarvalued func
tions. The derivatives of tensorvalued functions are more complicated.
Note that the derivative of a linear function
F(X) = C·· X
T
is C:
F
,X
= C.
Exercise 3.41. Using Exercises 3.27 and 3.28, verify that if F(X) = X
then F
,X
= I. If F(X) = X
T
, then F
,X
= e
k
Ee
k
.
By the exercise above,
X
,X
= I.
That is, the derivative of a secondorder tensor X with respect to X is I, a
tensor of order four.
Tensors 75
On symmetric tensor functions
The principal tensors of linear elasticity, the stress tensor σ and strain
tensor ε, are symmetric (cf., Chapter 6). So we consider the problem of
diﬀerentiating a tensorvalued function of a symmetric tensorial argument,
as this speciﬁc case has some peculiarities.
Let us consider the derivative of a scalarvalued function f(X) of a sym
metric secondorder tensor X. We modify the deﬁnition (3.39) as follows.
The derivative f
,X
is a secondorder symmetric tensor that satisﬁes the
condition
∂
∂ε
f(X+ ε dX)
ε=0
= f
,X
·· dX (3.41)
for any symmetric tensor dX. The component representation of f
,X
in a
Cartesian basis is
f
,X
=
∂f
∂x
mn
i
m
i
n
.
So we have brought the symmetry of f
,X
into the deﬁnition:
(f
,X
)
T
= f
,X
.
Why do we require f
,X
to be symmetric? The sets of symmetric and
antisymmetric secondorder tensors are subspaces of the space of all second
order tensors. Let A be a symmetric tensor and B an antisymmetric
secondorder tensor. It is easy to see that A·· B = 0. So the subspaces of
symmetric and antisymmetric tensors are mutually orthogonal. In (3.41),
dX is arbitrary but symmetric. If in this deﬁnition we do not require f
,X
to be symmetric, then (3.41), holding only for all the symmetric tensors
dX, deﬁnes f
,X
nonuniquely up to an additive term B that can be any
antisymmetric secondorder tensor.
Note that the deﬁnition of derivative for a function with respect to a
symmetric tensor argument is closely related to the problem of representing
a linear function acting on the subspace of symmetric tensors, which diﬀers
slightly from the general case (see Problems 3.52, 3.53, and 3.54).
As an example, we ﬁnd the derivative of the function
f(X) = C·· X,
where C is a secondorder tensor and X
T
= X. By deﬁnition,
f
,X
·· dX = C·· dX
76 Tensor Analysis with Applications in Mechanics
for all dX such that dX = dX
T
. At ﬁrst glance it seems that f
,X
= C, and
this does hold for an arbitrary argument X. But for a symmetric argument
X the answer changes. Indeed the derivative must be a symmetric tensor,
so f
,X
= C if and only if C = C
T
. Now we consider the case when C is not
symmetric. We said that f
,X
must be symmetric. A consequence is that
f
,X
becomes
f
,X
=
1
2
(C+C
T
).
This expression is uniquely deﬁned. Indeed, we can represent C as a sum
of symmetric and antisymmetric tensors:
C =
1
2
(C+C
T
) +
1
2
(C−C
T
)
But
1
2
(C−C
T
) ·· dX =
1
2
(C−C) ·· dX = 0
and thus
C·· dX =
1
2
(C+C
T
) ·· dX
as claimed.
In a similar fashion, we should modify the deﬁnition of the derivative of
a tensorvalued function of a secondorder symmetric argument. If F takes
values in the set of tensors of order n, the derivative with respect to X is
denoted by F
,X
; it takes values in the subset of tensors of order n+2 which
are symmetric in the two last indices and satisfy the equality
F
,X
·· dX =
∂
∂ε
F(X+ ε dX)
ε=0
(3.42)
for all symmetric secondorder tensors dX. The symmetry of F
,X
in the
last two indices means that for the components F
ij...pt
of F
,X
we have
F
ij...pt
= F
ij...tp
.
Similar changes permit the deﬁnition of partial derivative for tensorial
functions in many symmetric tensorial arguments.
Exercise 3.42. Let
F(X) = C·· X,
Tensors 77
where C = c
mnpt
i
m
i
n
i
p
i
t
is a fourthorder tensor and X is a secondorder
symmetric tensor. Demonstrate that
F
,X
=
1
2
(C+C
),
where C
is derived from C by transposing the last two indices: C
=
c
mnpt
i
m
i
n
i
t
i
p
.
Exercise 3.43. Let
f(X) =
1
2
X·· C·· X,
where C = c
mnpt
i
m
i
n
i
p
i
t
and X is a secondorder symmetric tensor.
Demonstrate that
f
,X
=
1
4
(C·· X+C
·· X+X·· C+X·· C
),
where C
is derived from C by transposing the last two indices of the
components of C, i.e., C
= c
mnpt
i
m
i
n
i
t
i
p
, and C
by transposing the ﬁrst
two indices of the components, i.e., C
= c
mnpt
i
n
i
m
i
p
i
t
.
Exercise 3.44. Let
f(X) =
1
2
X·· C·· X,
where C is a fourthorder tensor and X is a secondorder symmetric tensor.
Suppose C = C
= C
and C·· X = X·· C for any symmetric tensor X; in
terms of components, this means the equalities c
mnpt
= c
nmpt
and c
mnpt
=
c
ptmn
hold for any sets of indices. Using the solution of Exercise 3.43,
demonstrate that f
,X
= C·· X.
3.14 Problems
In this section, unless otherwise stated, we use A, B, C, Xto denote second
order tensors, λ a scalar, Ω an antisymmetric secondorder tensor, and Q
an orthogonal secondorder tensor.
3.1 Write out the components of the dyad i
1
i
2
in a Cartesian basis
(i
1
, i
2
, i
3
).
3.2 Find the components of the tensor i
1
i
2
−i
2
i
1
+ 2i
3
i
3
in the Cartesian
basis (i
1
, i
2
, i
3
).
78 Tensor Analysis with Applications in Mechanics
3.3 Write out the components of the tensor that is the dyad a
1
a
2
in the
Cartesian basis (i
1
, i
2
, i
3
), where a
1
= (−1, 2, −2) and a
2
= (1, −1, 2).
3.4 Show that a0 = 0b = 0.
3.5 Determine the symmetric and antisymmetric parts of the following
tensors:
(a) i
1
i
2
;
(b) i
1
i
2
−i
2
i
1
+ 2i
3
i
3
;
(c) i
1
i
2
−2i
2
i
1
+i
1
i
3
;
(d) i
1
i
2
+i
2
i
3
+i
1
i
3
;
(e) i
1
i
1
+ 2i
1
i
2
+ 2i
2
i
1
+i
3
i
1
+i
1
i
3
.
3.6 Determine the ball and deviator parts of the following tensors:
(a) i
1
i
2
;
(b) i
1
i
2
+i
2
i
1
;
(c) i
1
i
1
;
(d) aa;
(e) i
1
i
1
+ 2i
1
i
2
+ 2i
2
i
1
+i
3
i
1
+i
1
i
3
.
3.7 Show that
I
1
(X) = tr X, I
2
(X) =
1
2
[tr
2
X−tr X
2
], I
3
(X) = det X.
3.8 Find the invariants of the following tensors:
(a) aa;
(b) i
1
i
2
;
(c) i
1
i
1
+i
2
i
2
;
(d) λE;
(e) 2i
1
i
1
+ 3i
2
i
2
+ 4i
3
i
3
.
3.9 Let A be a symmetric, invertible tensor. Show that
I
1
(A
−1
) =
I
2
(A)
I
3
(A)
, I
2
(A
−1
) =
I
1
(A)
I
3
(A)
, I
3
(A
−1
) = I
−1
3
(A).
3.10 Find the left and right polar decompositions of the following tensors:
(a) λE;
(b) aa +bb +cc if a, b, c are mutually orthogonal;
Tensors 79
(c) λE+ ai
1
i
1
;
(d) λE+ ai
1
i
1
+ bi
2
i
2
;
(e) ai
1
i
1
+ bi
2
i
2
+ ci
3
i
3
.
3.11 Suppose tr A
2
= 0, a is an arbitrary scalar, and Y is an arbitrary
secondorder tensor. Demonstrate that
X = aY−a
tr(A· Y)
tr A
2
A
satisﬁes the equation tr(A· X) = 0.
3.12 Let a nonzero scalar a and secondorder tensors A, B be given. Find
a solution X of the equation
aX+ tr(A· X)E = B.
3.13 For the equations
(a) X+ tr(A· X)B = C,
(b) X
T
+ tr(A· X)B = C,
(c) X+ a(tr X)A = B,
ﬁnd a solution X and establish conditions for uniqueness.
3.14 Under what conditions will the equation
aX+Etr X = 0
have a nonzero solution X?
3.15 Under what conditions will the equation
aX+Atr X = 0
have a nontrivial solution?
3.16 Demonstrate that in a Cartesian basis
E = i
1
i
2
i
3
+i
2
i
3
i
1
+i
3
i
1
i
2
−i
1
i
3
i
2
−i
3
i
2
i
1
−i
2
i
1
i
3
.
3.17 Let A be a symmetric tensor. Show that I
2
(dev A) ≤ 0.
3.18 Let A be a nonsingular tensor and let a and b be arbitrary vectors.
Prove that A+ab is a nonsingular tensor if and only if 1 +b· A
−1
· a = 0,
and
(A+ab)
−1
= A
−1
−
1
1 +b · A
−1
· a
(A
−1
· a)(b · A
−1
).
80 Tensor Analysis with Applications in Mechanics
3.19 Show that (E×ω)
2
= ωω −Eω · ω.
3.20 Show that
(a) (E×ω)
2n
= (−1)
n−1
(ω · ω)
n−1
(ωω −Eω · ω),
(b) (E×ω)
2n+1
= (−1)
n
(ω · ω)
n
E×ω,
(c) a ×E×b = ba −a · bE.
3.21 For a secondorder tensor A = a
mn
e
m
e
n
, the vectorial invariant was
introduced by J.W. Gibbs (1839–1903) as
A
×
= a
mn
e
m
×e
n
.
For example, (ab)
×
= a ×b. Find the vectorial invariants of the following
tensors:
(a) aa;
(b) E×ω;
(c) ab −ba;
(d) E;
(e) A if A = A
T
.
3.22 Show that A
×
=
1
2
(A−A
T
)
×
.
3.23 Show that (a ×A)
×
= A· a −atr A.
3.24 Show that (A×B) ·· B
T
= 0 holds for arbitrary secondorder tensors
A and B.
3.25 Show that tr [b ×(a ×A)] = b · A· a −a · btr A.
3.26 Show that
A×ω = −(ω ×A
T
)
T
, ω ×A = −(A
T
×ω)
T
.
3.27 Two secondorder tensors A and B are called commutative if A· B =
B · A. Show that symmetrical tensors A and B are commutative if their
sets of eigenvectors coincide.
3.28 Let a · b = 0. Demonstrate that the dyads aa and bb are commuta
tive.
3.29 Let the symmetric tensors A and B be commutative. Demonstrate
that (A· B)
×
= 0.
Tensors 81
3.30 Show that the tensor Q = i
1
i
2
−i
2
i
1
+i
3
i
3
is orthogonal.
3.31 Show that the principal invariants of an orthogonal tensor satisfy the
relations I
1
I
3
= I
2
and I
2
3
= 1.
3.32 Let e · e = 1. Show that the tensor Q = E−2ee is orthogonal.
3.33 Show that if Ω is an antisymmetric tensor, then the tensor
Q = (E+Ω) · (E−Ω)
−1
is orthogonal.
3.34 Show that the tensor Q = e
Ω
is orthogonal if Ω is antisymmetric.
3.35 Demonstrate that the tensor
Q =
1
4 + θ
2
(4 −θ
2
)E+ 2θθ −4E× θ
,
where θ
2
= θ · θ, is orthogonal. We call θ a ﬁnite rotation vector.
3.36 Let Q be an orthogonal tensor. Establish the identity
(Q−E) ·· (Q
T
−E) = 6 −2 tr Q.
3.37 Let e
k
(k = 1, 2, 3) and d
m
(m = 1, 2, 3) be two orthonormal bases.
Verify that Q = e
k
d
k
is an orthogonal tensor.
3.38 Let Q be a proper orthogonal tensor. Establish the identity
Q
×
= 2 sinωe,
where ω and e are the rotation angle and axis, respectively.
3.39 Find the principal invariants of the tensor αE + βee.
3.40 Find the principal invariants of the tensor E ×ω.
3.41 Find the principal invariants of the tensor (E ×ω)
2
.
3.42 Show that
I
3
(A) =
1
6
tr
3
A−3 tr Atr A
2
+ 2 tr A
3
.
3.43 From the Cayley–Hamilton theorem for a nondegenerate tensor, it
follows that
A
−1
=
1
det A
A
2
−I
1
(A)A+ I
2
(A)E
.
Use this to ﬁnd the inverses of the following tensors:
82 Tensor Analysis with Applications in Mechanics
(a) i
1
i
1
+ 2i
2
i
2
+i
3
i
3
+i
1
i
2
+i
2
i
1
;
(b) ai
1
i
3
+ bi
2
i
2
+ ci
3
i
1
, a, b, c = 0;
(c) ai
1
i
1
+i
2
i
2
+i
3
i
3
+ bi
1
i
2
;
(d) aE+ bi
1
i
2
.
3.44 Find the derivative with respect to the tensor X of the following
scalarvalued functions:
(a) tr X
4
;
(b) a · X· a;
(c) a · X· b;
(d) tr(B· X).
3.45 Let A be a secondorder tensor. A tensor X is called the cofactor of
A and denoted by X = cof A if X satisﬁes the equation
A· X
T
= X
T
· A = (det A)E.
If A is nonsingular, then cof A = (det A)A
−T
. Check the following prop
erties of cofactor:
(a) cof(A
T
) = (cof A)
T
;
(b) cof(A· B) = cof A· cof B;
(c) cof λA = λ
2
cof A;
(d) I
2
(A) = tr cof A.
3.46 Let A = λ
1
i
1
i
1
+ λ
2
i
2
i
2
+ λ
3
i
3
i
3
. Demonstrate that
cof A = λ
2
λ
3
i
1
i
1
+ λ
1
λ
3
i
2
i
2
+ λ
1
λ
2
i
3
i
3
.
3.47 Let f(x) be an analytic function in an open ball x < r centered at
x = 0; that is, suppose the representation
f(x) =
∞
¸
k=0
f
(k)
(0)
k!
x
k
holds and the series converges for x < r. In the ball, the relation
f
(x) =
∞
¸
k=1
f
(k)
(0)
(k −1)!
x
k−1
holds. Let
F(X) =
∞
¸
k=0
f
(k)
(0)
k!
X
k
.
Tensors 83
Demonstrate that
[tr(F(X)]
,X
=
∞
¸
k=1
f
(k)
(0)
(k −1)!
X
T
k−1
= f
(X
T
)
for X < r.
3.48 Let f(X) be a scalarvalued function. Show that the derivative of
f(X) with respect to X
T
is equal to the transpose of its derivative with
respect to X, i.e.,
f(X)
,X
T = (f(X)
,X
)
T
.
3.49 Find the derivative with respect to X of the following tensorvalued
functions:
(a) tr(A· X
T
)E;
(b) tr(A· X)E;
(c) a · X· bcd.
3.50 For a nonsingular tensor X, demonstrate that
(det X)
,X
= (det X)X
−T
.
3.51 Let S be the set of all symmetric secondorder tensors. Show that S
is a linear subspace of the linear space of all the secondorder tensors. On
S the inner product A·· B
T
of the whole space takes the form
A·· B
T
= A·· B.
3.52 Let f be a scalarvalued function of a symmetric secondorder tensor
A. Show that there is a unique symmetric secondorder tensor B such that
f(A) = tr(B· A)
for any symmetric A, that is A = A
T
.
3.53 Let F = F(A) be a linear function from the set of secondorder
symmetric tensors A to the set of secondorder tensors. Show that there is
a unique fourthorder tensor C such that
F(A) = C·· A, C = c
mnpt
i
m
i
n
i
p
i
t
, c
mnpt
= c
mntp
for all A = A
T
.
84 Tensor Analysis with Applications in Mechanics
3.54 Let F = F(A) be a linear function from the set of secondorder
symmetric tensors A to the same set of symmetric secondorder tensors
(F(A) = F(A)
T
). Show that there is a unique fourthorder tensor C such
that
F(A) = C·· A, C = c
mnpt
i
m
i
n
i
p
i
t
, c
mnpt
= c
mntp
, c
mnpt
= c
nmpt
whenever A = A
T
.
3.55 Let Q be the set of all antisymmetric secondorder tensors. Show
that Q is a linear subspace of the linear space of all secondorder tensors.
On Q the inner product A·· B
T
of the whole space takes the form
A·· B
T
= −A·· B = −A
T
·· B
T
.
3.56 Let C be a tensor of order n that is antisymmetric in the two last
components: c
mn...pt
= −c
mn...tp
. Let A be an arbitrary symmetric second
order tensor. Demonstrate that C · · A = 0, where 0 is the zero tensor of
order n −2.
Chapter 4
Tensor Fields
4.1 Vector Fields
The state of a point (more precisely of an inﬁnitely small volume) within
a natural object is frequently characterized by a vector or a tensor. Hence
inside a spatial object there arises what we call a vector or tensor ﬁeld. As a
rule these ﬁelds are governed by simultaneous partial diﬀerential equations.
Such equations are usually derived using a Cartesian space frame. In this
frame, the operations of calculus closely parallel those of one dimensional
analysis: the diﬀerentiation of a vector function proceeds on a component
by component basis, for example. However, it is often convenient to intro
duce curvilinear coordinates in the body, in terms of which the problem
formulation is simpler. In this way we get a frame that changes from point
to point, and componentwise diﬀerentiation is not enough to characterize
the change of the vector function. Thus we need to develop the apparatus
of calculus for vector and tensor functions when the frame in the object is
changeable. Another reason for introducing these tools is the objectivity of
the laws of nature: we must be able to formulate frame independent state
ments of these laws. Finally, there is an aesthetic reason: in noncoordinate
form many statements of mathematical physics look much less cumbersome
than their counterparts stated in terms of coordinates. It is said that beauty
governs the world; although this is not absolutely true, most students would
prefer a short and beautiful statement to a nightmarish formula taking half
a page.
The position of a point in space is characterized by three numbers called
coordinates of the point. Coordinate systems in common use include the
Cartesian, cylindrical, and spherical systems. The ﬁrst of these diﬀers from
the latter two in an important respect: the frame vectors of a Cartesian
85
86 Tensor Analysis with Applications in Mechanics
system are unique, while those for the curvilinear systems change from point
to point. In a general coordinate system the position of a point in space is
determined uniquely by three numbers q
1
, q
2
, q
3
. These are referred to as
curvilinear coordinates if the frame is not Cartesian. See Fig. 4.1.
Fig. 4.1 Curvilinear coordinates.
If we ﬁx two of the coordinates and change the third, we get a line
in space called a coordinate line. The Cartesian coordinates x
1
, x
2
, x
3
of
a point can be determined through the general curvilinear coordinates by
relations of the form
x
i
= x
i
(q
1
, q
2
, q
3
) (i = 1, 2, 3). (4.1)
Except for some set of singular points in space, the correspondence (4.1) is
one to one. We suppose the functions x
i
= x
i
(q
1
, q
2
, q
3
) are smooth (con
tinuously diﬀerentiable). In this case the local onetoone correspondence
is provided by the requirement that the Jacobian
∂x
i
∂q
j
= 0. (4.2)
Fixing some origin O, we characterize the position of a point P(q
1
, q
2
, q
3
)
by a vector r connecting the points O and P:
r = r(q
1
, q
2
, q
3
).
When a point moves along a coordinate line, along the q
1
line for in
stance, the end of its position vector moves along this line and the diﬀerence
vector
∆r = r(q
1
+ ∆q
1
, q
2
, q
3
) −r(q
1
, q
2
, q
3
)
Tensor Fields 87
Fig. 4.2 Generation of a local tangent vector to a coordinate line.
is directed along the chord (Fig. 4.2). The smaller the value of ∆q
1
, the
closer ∆r is to being tangent to the q
1
line. The limit as ∆q
1
→ 0 of the
ratio ∆r/∆q
1
is a vector
∂r(q
1
, q
2
, q
3
)
∂q
1
= r
1
tangent to the q
1
coordinate line. At the same point (q
1
, q
2
, q
3
) we can
introduce two other vectors
r
2
=
∂r(q
1
, q
2
, q
3
)
∂q
2
, r
3
=
∂r(q
1
, q
2
, q
3
)
∂q
3
,
tangent to the q
2
and q
3
coordinate lines, respectively. If the coordinate
point (q
1
, q
2
, q
3
) is not singular then the vectors r
i
are noncoplanar and
therefore constitute a frame triad. The mixed product r
1
· (r
2
× r
3
) is
the volume of the frame parallelepiped; renumbering the coordinates if
necessary, we obtain from this the same expression for the Jacobian (4.2).
Let us denote it
√
g = r
1
· (r
2
×r
3
) =
∂x
i
∂q
j
.
Unlike the previous chapters in which all frame vectors were constant, we
now deal with frame vectors that change from point to point. However, at
each point we can repeat our prior reasoning. In particular, let us introduce
the reciprocal basis r
i
by the relation
r
i
· r
j
= δ
i
j
.
By the previous chapter we deﬁne the metric coeﬃcients
g
ij
= r
i
· r
j
, g
ij
= r
i
· r
j
, g
j
i
= r
i
· r
j
= δ
j
i
,
88 Tensor Analysis with Applications in Mechanics
which are the components of the metric tensor E at each point in space.
Above we diﬀerentiated a vector function r(q
1
, q
2
, q
3
). Note that the
rules for diﬀerentiating vector functions are quite similar to those for dif
ferentiating ordinary functions. For brevity we consider the case of ﬁelds
depending on one variable t. Let e
1
(t) and e
2
(t) be continuously diﬀeren
tiable at some ﬁnite t, which means that for i = 1, 2 there exist
de
i
(t)
dt
= lim
∆t→0
e
i
(t + ∆t) −e
i
(t)
∆t
.
It is easily seen that
d(e
1
(t) +e
2
(t))
dt
=
de
1
(t)
dt
+
de
2
(t)
dt
.
Indeed,
d(e
1
(t) +e
2
(t))
dt
= lim
∆t→0
e
1
(t + ∆t) −e
1
(t) +e
2
(t + ∆t) −e
2
(t)
∆t
= lim
∆t→0
e
1
(t + ∆t) −e
1
(t)
∆t
+ lim
∆t→0
e
2
(t + ∆t) −e
2
(t)
∆t
=
de
1
(t)
dt
+
de
2
(t)
dt
.
Similarly, for a constant c we have
d(ce
1
(t))
dt
= c
de
1
(t)
dt
.
Finally, the product of a scalar function f(t) and a vector function e(t) is
diﬀerentiated by a rule similar to the formula for diﬀerentiating a product
of scalar functions:
d(f(t)e(t))
dt
=
df(t)
dt
e(t) + f(t)
de(t)
dt
.
The reader can adapt the proof for the scalar case almost word for word.
The rules for partial diﬀerentiation of vector functions in several variables
look quite similar and we leave their formulation to the reader. Repre
senting vectors in Cartesian coordinates it is easy to see the validity of the
following formulas:
d
dt
(e
1
(t) · e
2
(t)) = e
1
(t) · e
2
(t) +e
1
(t) · e
2
(t),
d
dt
(e
1
(t) ×e
2
(t)) = e
1
(t) ×e
2
(t) +e
1
(t) ×e
2
(t).
Tensor Fields 89
Exercise 4.1. (a) Diﬀerentiate e
1
(t) · e
2
(t) and e
1
(t) ×e
2
(t) with respect
to t if
e
1
(t) = i
1
e
−t
+i
2
, e
2
(t) = −i
1
sin
2
t +i
2
e
−t
.
(b) Show that [e(t) × e
(t)]
= e(t) × e
(t) for any diﬀerentiable vector
function e(t).
Exercise 4.2. For the mixed product
[e
1
(t), e
2
(t), e
3
(t)] = (e
1
(t) ×e
2
(t)) · e
3
(t)
show that
d
dt
[e
1
(t), e
2
(t), e
3
(t)]
= [e
1
(t), e
2
(t), e
3
(t)] + [e
1
(t), e
2
(t), e
3
(t)] + [e
1
(t), e
2
(t), e
3
(t)].
Cylindrical coordinates
In the cylindrical coordinate system we have
(q
1
, q
2
, q
3
) = (ρ, φ, z),
where ρ is the radial distance from the zaxis and φ is the azimuthal angle
(Fig. 4.3).
Fig. 4.3 Cylindrical coordinate system.
Using the expression
r = ˆ xx + ˆ yy + ˆzz
90 Tensor Analysis with Applications in Mechanics
for the position vector in rectangular coordinates and the coordinate trans
formation formulas
x = ρ cos φ, y = ρ sinφ, z = z,
we have
r(ρ, φ, z) = ˆ xρ cos φ + ˆ yρ sin φ + ˆzz.
Then
r
1
=
∂r(ρ, φ, z)
∂ρ
= ˆ xcos φ + ˆ ysin φ.
Similarly we compute
r
2
=
∂r(ρ, φ, z)
∂φ
= −ˆ xρ sin φ + ˆ yρ cos φ, r
3
=
∂r(ρ, φ, z)
∂z
= ˆ z.
Then
√
g = r
1
· (r
2
×r
3
)
= (ˆ xcos φ + ˆ ysin φ) · [(−ˆ xρ sin φ + ˆ yρ cos φ) × ˆ z]
= (ˆ xcos φ + ˆ ysin φ) · (ˆ yρ sin φ + ˆ xρ cos φ)
= ρ(cos
2
φ + sin
2
φ)
= ρ.
It is easy to verify that the frame vectors r
1
, r
2
, r
3
are mutually perpendic
ular. For instance,
r
1
· r
2
= (ˆ xcos φ + ˆ ysin φ) · (−ˆ xρ sinφ + ˆ yρ cos φ)
= −ρ cos φsin φ + ρ cos φsin φ = 0.
By direct computation we also ﬁnd that
r
1
 = 1, r
2
 = ρ, r
3
 = 1.
These facts may be used to construct the reciprocal basis as
r
1
= r
1
, r
2
= r
2
/ρ
2
, r
3
= r
3
.
The various metric coeﬃcients for the cylindrical frame are easily computed,
and are
(g
ij
) =
⎛
⎝
1 0 0
0 ρ
2
0
0 0 1
⎞
⎠
, (g
ij
) =
⎛
⎝
1 0 0
0 1/ρ
2
0
0 0 1
⎞
⎠
.
Tensor Fields 91
Commonly the unit basis vectors ˆ ρ,
ˆ
φ, and ˆ z are introduced in cylin
drical coordinates. These are unit vectors along the directions of r
1
, r
2
,
and r
3
, respectively, at the point of interest. Given any vector v at a point
(ρ, φ, z) in space, it is conventional in applications to write
v = ˆ ρv
ρ
+
ˆ
φv
φ
+ ˆ zv
z
.
The quantities v
ρ
, v
φ
, and v
z
are known as the physical components of v.
(For a vector, the physical components are the projections of the vector on
the directions of the corresponding frame vectors.
1
) But since we can also
express v in the forms
v = v
1
r
1
+ v
2
r
2
+ v
3
r
3
= v
1
r
1
+ v
2
r
2
+ v
3
r
3
,
we can identify the contravariant and covariant components of v from our
previous expressions. They are
(v
1
, v
2
, v
3
) = (v
ρ
, v
φ
/ρ, v
z
), (v
1
, v
2
, v
3
) = (v
ρ
, v
φ
ρ, v
z
).
We see that expression of v in terms of the basis (ˆ ρ,
ˆ
φ, ˆz) yields components
(v
ρ
, v
φ
, v
z
) that are neither contravariant nor covariant.
Exercise 4.3. Show that the acceleration of a particle in plane polar co
ordinates (ρ, φ) is given by
d
2
r
dt
2
= ˆ ρ
¸
d
2
ρ
dt
2
−ρ
dφ
dt
2
¸
+
ˆ
φ
ρ
d
2
φ
dt
2
+ 2
dρ
dt
dφ
dt
.
Spherical coordinates
In the spherical coordinate system we have
(q
1
, q
2
, q
3
) = (r, θ, φ),
1
Mathematicians often prefer to deal with dimensionless quantities. But in applications
many quantities have physical dimensions (e.g., N/m
2
for a stress). When we introduce
coordinates in space some can also have dimensions; in spherical coordinates, for instance,
r has the dimension of length whereas θ and φ are dimensionless. Since r has the length
dimension, the frame vectors corresponding to θ and φ must have the length dimension,
whereas the frame vector for r is dimensionless. In this way some components of the
metric tensor may have dimensions and, when involved in transformation formulas, may
thus introduce not only numerical changes but also dimensional changes in the terms so
that it becomes possible to add such quantities. When one uses physical components of
vectors and tensors they take the dimension of the corresponding tensor in full, and the
frame vectors become dimensionless.
92 Tensor Analysis with Applications in Mechanics
Fig. 4.4 Spherical coordinate system.
where r is the radial distance from the origin, φ is the azimuthal angle, and
θ is the polar angle (Fig. 4.4). The coordinate transformation formulas
x = r sin θ cos φ, y = r sin θ sin φ, z = r cos θ
give
r(r, θ, φ) = ˆ xr sinθ cos φ + ˆ yr sinθ sinφ + ˆ zr cos θ.
Then
r
1
=
∂r(r, θ, φ)
∂r
= ˆ xsin θ cos φ + ˆ y sinθ sin φ + ˆ z cos θ,
r
2
=
∂r(r, θ, φ)
∂θ
= ˆ xr cos θ cos φ + ˆ yr cos θ sin φ − ˆ zr sin θ,
r
3
=
∂r(r, θ, φ)
∂φ
= −ˆ xr sin θ sin φ + ˆ yr sin θ cos φ.
In this case
√
g =
sin θ cos φ sin θ sin φ cos θ
r cos θ cos φ r cos θ sin φ −r sin θ
−r sinθ sin φ r sin θ cos φ 0
= r
2
sinθ.
In this system we ﬁnd that
r
1
 = 1, r
2
 = r, r
3
 = r sin θ,
and the frame vectors are again mutually orthogonal. Hence we have
r
1
= r
1
, r
2
= r
2
/r
2
, r
3
= r
3
/r
2
sin
2
θ
Tensor Fields 93
for the vectors of the reciprocal basis. The metric coeﬃcients are
(g
ij
) =
⎛
⎝
1 0 0
0 r
2
0
0 0 r
2
sin
2
θ
⎞
⎠
, (g
ij
) =
⎛
⎝
1 0 0
0 1/r
2
0
0 0 1/r
2
sin
2
θ
⎞
⎠
.
Exercise 4.4. The unit basis vectors of spherical coordinates are denoted
ˆr,
ˆ
θ, and
ˆ
φ. Give expressions relating the various sets of components
(v
1
, v
2
, v
3
), (v
1
, v
2
, v
3
), and (v
r
, v
θ
, v
φ
) of a vector v.
We see that both the spherical and cylindrical frames are orthogonal.
In this case the reciprocal basis vectors have the same directions as the
vectors of the main basis, but have reciprocal lengths so that r
i
r
i
 = 1
for each i.
In § 2.4 we obtained the transformation laws that apply to the compo
nents of a vector under a change of frame. These laws are a consequence of
the change of the old curvilinear coordinates q
k
to other coordinates ˜ q
j
, so
˜ q
j
= ˜ q
j
(q
1
, q
2
, q
3
), q
k
= q
k
(˜ q
1
, ˜ q
2
, ˜ q
3
) (j, k = 1, 2, 3).
In the coordinates ˜ q
j
, we denote all the quantities in the same manner as
for the coordinates q
k
, but with the tilde above: ˜r
k
, ˜r
k
, and so on. We
now extend the transformation laws to the case of a vector ﬁeld in general
coordinates. Because
r
i
=
∂r
∂q
i
=
∂r
∂˜ q
j
∂˜ q
j
∂q
i
= ˜r
j
∂˜ q
j
∂q
i
we can write
r
i
= A
j
i
˜r
j
, A
j
i
=
∂˜ q
j
∂q
i
, (4.3)
to describe the change of frame. For the inverse transformation
˜r
i
=
˜
A
j
i
r
j
,
˜
A
j
i
=
∂q
j
∂˜ q
i
. (4.4)
The results of § 2.4 now apply, and we can write immediately
˜
f
i
= A
i
j
f
j
,
˜
f
i
=
˜
A
j
i
f
j
, f
i
=
˜
A
i
j
˜
f
j
, f
i
= A
j
i
˜
f
j
,
for the transformation laws pertaining to the components of a vector ﬁeld
f . The f
i
are still termed contravariant components of f , while the f
i
are
still termed covariant components — the main diﬀerence is that now A
j
i
and
˜
A
j
i
can change from point to point in space. The transformation laws
for the components of higherorder tensor ﬁelds are written similarly.
94 Tensor Analysis with Applications in Mechanics
Exercise 4.5. A set of oblique rectilinear coordinates (u, v) in the plane
is related to a set of Cartesian coordinates (x, y) by the transformation
equations
x = u cos α + v cos β, y = u sin α + v sin β,
where α, β are constants. (a) Sketch a few u and v coordinate curves
superimposed on the xyplane. (b) Find the basis vectors r
i
(i = 1, 2) and
the reciprocal basis vectors r
i
in the (u, v) system. Use these to calculate
the metric coeﬃcients. (c) Let z be a given vector. In the oblique system
ﬁnd the covariant components of z in terms of the contravariant components
of z.
4.2 Diﬀerentials and the Nabla Operator
Let us consider the inﬁnitesimal vector extending from point (q
1
, q
2
, q
3
) to
point (q
1
+ dq
1
, q
2
+ dq
2
, q
3
+ dq
3
), denoted by dr:
dr =
∂r
∂q
i
dq
i
= r
i
dq
i
.
Hence we can deﬁne dq
i
as
dq
i
= r
i
· dr = dr · r
i
. (4.5)
(Here there is no summation over i. Note that we only introduce spatial co
ordinates having superscripts. However, for Cartesian frames it is common
to see subscripts used exclusively.) The length of this inﬁnitesimal vector
is deﬁned by
(ds)
2
= dr · dr = r
i
dq
i
· r
j
dq
j
= g
ij
dq
i
dq
j
. (4.6)
On the right we have a quadratic form with respect to the variables dq
i
;
the coeﬃcients of this quadratic form are the covariant components of the
metric tensor. In a Cartesian frame (4.6) takes the familiar form
(ds)
2
= (dx
1
)
2
+ (dx
2
)
2
+ (dx
3
)
2
.
In cylindrical and spherical frames we have
(ds)
2
= (dρ)
2
+ ρ
2
(dφ)
2
+ (dz)
2
,
(ds)
2
= (dr)
2
+ r
2
(dθ)
2
+ r
2
sin
2
θ(dφ)
2
,
respectively.
Exercise 4.6. Find (ds)
2
for the oblique system of Exercise 4.5.
Tensor Fields 95
Let f(q
1
, q
2
, q
3
) be a scalar diﬀerentiable function of the variables q
i
.
Its diﬀerential is
df(q
1
, q
2
, q
3
) =
∂f(q
1
, q
2
, q
3
)
∂q
i
dq
i
.
Here and in similar situations the i in the denominator stands in the lower
position and so we must sum over i. Using (4.5) we can write
df = r
i
∂f
∂q
i
· dr.
The ﬁrst multiplier on the right is a vector r
i
∂f/∂q
i
. Let us introduce a
symbolic vector
∇ = r
i
∂
∂q
i
called the nabla operator, whose action on a function f is as given above:
∇f = r
i
∂f
∂q
i
(we repeat that there is summation over i here). The nabla operator is
often referred to as the gradient operator.
Exercise 4.7. (a) Show that the unit normal to the surface ϕ(q
1
, q
2
, q
3
) =
c = constant is given by
n =
g
ij
∂ϕ
∂q
i
g
mn
∂ϕ
∂q
m
∂ϕ
∂q
n
r
j
.
(b) Show that the angle at a point of intersection of the surfaces
ϕ(q
1
, q
2
, q
3
) = c
1
, ψ(q
1
, q
2
, q
3
) = c
2
,
is given by
cos θ =
g
ij
∂ϕ
∂q
i
∂ψ
∂q
j
g
mn
∂ϕ
∂q
m
∂ϕ
∂q
n
g
rt
∂ψ
∂q
r
∂ψ
∂q
t
(c) Show that the angle between the coordinate surfaces q
1
= c
1
, q
2
= c
2
is
given by
cos θ
12
=
g
12
g
11
g
22
.
96 Tensor Analysis with Applications in Mechanics
(d) Derive the condition for orthogonality of surfaces
g
ij
∂ϕ
∂q
i
∂ψ
∂q
j
= 0.
Exercise 4.8. Show that the gradient operation in the cylindrical and
spherical frames is given by the formulas
∇f = ˆ ρ
∂f
∂ρ
+
ˆ
φ
1
ρ
∂f
∂φ
+ ˆz
∂f
∂z
,
∇f = ˆr
∂f
∂r
+
ˆ
θ
1
r
∂f
∂θ
+
ˆ
φ
1
r sinθ
∂f
∂φ
,
respectively. What are the expressions for these when using the components
connected with the triads r
i
and r
i
?
Let us ﬁnd a formula for the diﬀerential of a vector function f (q
1
, q
2
, q
3
):
df (q
1
, q
2
, q
3
) =
∂f (q
1
, q
2
, q
3
)
∂q
i
dq
i
.
We use (4.5) again. Then
df (q
1
, q
2
, q
3
) =
∂f (q
1
, q
2
, q
3
)
∂q
i
(r
i
· dr)
= (dr · r
i
)
∂f (q
1
, q
2
, q
3
)
∂q
i
= dr ·
r
i
∂f (q
1
, q
2
, q
3
)
∂q
i
.
Thus we can represent this as
df = dr · ∇f .
The quantity ∇f , known as the gradient of f , is clearly a tensor of order
two. With the aid of transposition we can present df in the form
df = ∇f
T
· dr.
Sometimes ∇f
T
is called the gradient of f ; it is also called the derivative of
f in the direction of r and denoted as
df
dr
= ∇f
T
.
An application of the gradient to a tensor brings a new tensor whose order
is one higher than that of the original tensor.
For a pair of vectors, we introduced two types of multiplication: the dot
and cross product operations. We can apply these to the pair consisting
Tensor Fields 97
of the nabla operator (which is regarded as a formal vector) and a vector
function f = f (q
1
, q
2
, q
3
). In this way we get two operations: the divergence
div f = ∇· f = r
i
·
∂f
∂q
i
,
and the rotation
rot f = ∇× f = r
i
×
∂f
∂q
i
. (4.7)
The vector
ω =
1
2
rot f
is called the curl of f . In terms of the curl of f we can introduce a tensor
Ω as
Ω = E×ω.
This tensor is antisymmetric (the reader should verify this), and is called
the tensor of spin. It can be shown that
Ω =
1
2
∇f
T
−∇f
.
Similarly we can introduce the divergence and rotation operations for a
tensor A of any order:
∇· A = r
i
·
∂
∂q
i
A, ∇×A = r
i
×
∂
∂q
i
A.
Let us see what happens when we apply the nabla operator to the radius
vector:
∇r = r
i
∂
∂q
i
r = r
i
r
i
= E,
∇· r = r
i
·
∂
∂q
i
r = r
i
· r
i
= 3,
and
∇× r = r
i
×
∂
∂q
i
r = r
i
×r
i
= 0.
Exercise 4.9. Calculate ∇E, ∇· E, ∇×E, ∇E, ∇· E, and ∇×E.
98 Tensor Analysis with Applications in Mechanics
Exercise 4.10. Let f and g be functions, let f be a vector, and let Q be
a tensor of order two. Verify the following identities:
∇(fg) = g∇f + f∇g,
∇(ff ) = (∇f)f + f∇f ,
∇(fQ) = (∇f)Q+ f∇Q.
Exercise 4.11. Show that the following identities hold (f and g are any
vectors):
∇(f · g) = (∇f ) · g +f · ∇g
T
= (∇f ) · g + (∇g) · f ,
∇(f ×g) = (∇f ) ×g −(∇g) ×f ,
∇×(f ×g) = g · ∇f −g∇· f −f · ∇g +f ∇· g,
∇· (fg) = (∇· f )g +f ∇· g,
∇· (f ×g) = g · (∇×f ) −f · (∇×g).
4.3 Diﬀerentiation of a Vector Function
We have diﬀerentiated vector functions with respect to q
i
, with the tacit
understanding that the formulas necessary to do this resemble those from
ordinary calculus. Moreover, the reader certainly knows that to diﬀerentiate
a vector function
f (x
1
, x
2
, x
3
) = f
k
(x
1
, x
2
, x
3
)i
k
with respect to a Cartesian variable x
i
it is enough to diﬀerentiate each
component of f with respect to this variable:
∂
∂x
i
f (x
1
, x
2
, x
3
) =
∂f
k
(x
1
, x
2
, x
3
)
∂x
i
i
k
. (4.8)
Let us consider how to diﬀerentiate a vector function f that is written out
in curvilinear coordinates:
f (q
1
, q
2
, q
3
) = f
i
(q
1
, q
2
, q
3
)r
i
.
In this case the frame vectors r
i
depend on q
k
as well, which means
that simple diﬀerentiation of the components of f does not result in the
needed formula. To understand this let us consider a constant function
f (q
1
, q
2
, q
3
) = c. By the general deﬁnition of the partial derivative we have
∂f (q
1
, q
2
, q
3
)/∂q
i
= 0. But the components of f are not constant since the
Tensor Fields 99
r
k
are variable, hence the derivatives of the components of this function are
not zero.
The derivative of the product of a simple function and a vector function
is taken by the product rule:
∂
∂q
k
f (q
1
, q
2
, q
3
) =
∂
∂q
k
f
i
(q
1
, q
2
, q
3
)r
i
=
∂f
i
(q
1
, q
2
, q
3
)
∂q
k
r
i
+ f
i
(q
1
, q
2
, q
3
)
∂r
i
∂q
k
.
So the derivative of a vector function written in component form consists
of two terms. The ﬁrst is the same as for the derivative in the Cartesian
frame (4.8), and the other contains the derivatives of the frame vectors.
Thus we need to ﬁnd the latter.
4.4 Derivatives of the Frame Vectors
We would like to ﬁnd the value of the derivative
∂
∂q
j
r
i
=
∂
∂q
j
∂
∂q
i
r
=
∂
2
r
∂q
j
∂q
i
=
∂
2
r
∂q
i
∂q
j
=
∂
∂q
i
r
j
. (4.9)
Of course if we have the expression for the Cartesian components of the
frame vector we can compute these derivatives (and will do so in the exer
cises); however in the general case we exploit only the fact that the partial
derivative of a frame vector is a vector as well, hence it can be expanded
in the same frame r
i
(i = 1, 2, 3). Let us denote the coeﬃcients of the
expansion by Γ
k
ij
:
∂
∂q
j
r
i
= Γ
k
ij
r
k
. (4.10)
The quantities Γ
k
ij
are called Christoﬀel coeﬃcients of the second kind and
are often denoted by
Γ
k
ij
=
k
ij
. (4.11)
By (4.9) there is symmetry in the subscripts of the Christoﬀel symbols:
Γ
k
ij
= Γ
k
ji
. (4.12)
100 Tensor Analysis with Applications in Mechanics
Now we can write out the formula for the derivative of a vector function
in full:
∂
∂q
k
f (q
1
, q
2
, q
3
) =
∂f
i
(q
1
, q
2
, q
3
)
∂q
k
r
i
+ Γ
j
ki
f
i
(q
1
, q
2
, q
3
)r
j
=
∂f
i
(q
1
, q
2
, q
3
)
∂q
k
r
i
+ Γ
i
kt
f
t
(q
1
, q
2
, q
3
)r
i
=
∂f
i
∂q
k
+ Γ
i
kt
f
t
r
i
. (4.13)
This is called covariant diﬀerentiation. The coeﬃcients of r
i
are called the
covariant derivatives of the contravariant components of f , and are denoted
by
∇
k
f
i
=
∂f
i
∂q
k
+ Γ
i
kt
f
t
.
Let us discuss some properties of the Christoﬀel coeﬃcients.
4.5 Christoﬀel Coeﬃcients and their Properties
In a Cartesian frame the Christoﬀel symbols are zero, so it is impossible to
obtain them for another frame by the usual transformation rules for tensor
components. This is easy to understand: the Christoﬀel symbols depend
not only on the frame vectors themselves but on their rates of change from
point to point, and these rates do not appear in the transformation rules
for tensors. So the Christoﬀel symbols are not the components of a tensor,
despite their notation. This is why many authors prefer the notation shown
on the right side of equation (4.11).
It is important to have formulas for computing the Christoﬀel symbols.
Let us introduce the notation
r
ij
=
∂
∂q
j
r
i
=
∂
2
r
∂q
i
∂q
j
.
Relation (4.10) can be written as
r
ij
= Γ
k
ij
r
k
(4.14)
from which it follows that
r
ij
· r
t
= Γ
k
ij
r
k
· r
t
= Γ
k
ij
g
kt
.
Tensor Fields 101
The lefthand side of this can be expressed in terms of the components of
the metric tensor. Indeed
∂
∂q
j
g
it
=
∂
∂q
j
r
i
· r
t
= r
ij
· r
t
+r
i
· r
tj
. (4.15)
Similarly
∂
∂q
t
g
ji
=
∂
∂q
t
r
j
· r
i
= r
jt
· r
i
+r
j
· r
it
, (4.16)
∂
∂q
i
g
tj
=
∂
∂q
i
r
t
· r
j
= r
ti
· r
j
+r
t
· r
ji
. (4.17)
Now we obtain r
ij
· r
t
by subtracting (4.16) from the sum of (4.15) and
(4.17) and dividing by 2:
r
ij
· r
t
=
1
2
∂g
it
∂q
j
+
∂g
tj
∂q
i
−
∂g
ji
∂q
t
= Γ
ijt
. (4.18)
The quantities Γ
ijk
are called Christoﬀel coeﬃcients of the ﬁrst kind. They
are denoted frequently as
[ij, k] = Γ
ijk
.
It is clear that they are symmetric in the ﬁrst two subscripts: Γ
ijk
= Γ
jik
.
Thus we have
Γ
k
ij
g
kt
= Γ
ijt
(4.19)
and it follows that
Γ
k
ij
= g
kt
Γ
ijt
. (4.20)
Using these we can obtain
∂r
j
∂q
i
= −Γ
j
it
r
t
. (4.21)
Indeed
0 =
∂
∂q
t
δ
i
j
=
∂
∂q
t
r
i
· r
j
=
∂r
i
∂q
t
· r
j
+r
i
· r
jt
=
∂r
i
∂q
t
· r
j
+r
i
· Γ
k
jt
r
k
,
so
∂r
i
∂q
t
· r
j
= −r
i
· Γ
k
jt
r
k
= −Γ
k
jt
δ
i
k
= −Γ
i
jt
and we have
∂r
i
∂q
t
= −Γ
i
jt
r
j
.
102 Tensor Analysis with Applications in Mechanics
It is also useful to have formulas for transformation of the Christoﬀel
symbols under change of coordinates. Suppose the new coordinates ˜ q
i
are
determined by the old coordinates q
i
through relations of the form ˜ q
i
=
˜ q
i
(q
1
, q
2
, q
3
), and refer back to equations (4.3)–(4.4). Now (4.14) applied
in the new system gives
˜
Γ
k
ij
˜r
k
= ˜r
ij
. (4.22)
But
˜r
ij
=
∂
∂˜ q
j
˜r
i
=
∂q
m
∂˜ q
j
∂
∂q
m
˜r
i
=
˜
A
m
j
∂
∂q
m
(
˜
A
n
i
r
n
),
and expansion by the product rule gives
˜r
ij
=
˜
A
m
j
∂
˜
A
n
i
∂q
m
r
n
+
˜
A
n
i
∂r
n
∂q
m
=
˜
A
m
j
∂
˜
A
n
i
∂q
m
r
n
+
˜
A
m
j
˜
A
n
i
Γ
p
nm
r
p
=
˜
A
m
j
∂
˜
A
n
i
∂q
m
A
k
n
˜r
k
+
˜
A
m
j
˜
A
n
i
Γ
p
nm
A
k
p
˜r
k
.
Comparison with (4.22) shows that
˜
Γ
k
ij
=
˜
A
m
j
∂
˜
A
n
i
∂q
m
A
k
n
+
˜
A
m
j
˜
A
n
i
A
k
p
Γ
p
nm
.
The presence of the ﬁrst term on the right means that the Christoﬀel co
eﬃcient is not a tensor of order three. This conﬁrms our earlier statement
of this fact, which was based on diﬀerent reasoning.
Exercise 4.12. Show that the only nonzero Christoﬀel coeﬃcients of the
ﬁrst kind for cylindrical coordinates are
Γ
221
= −ρ, Γ
122
= Γ
212
= ρ.
Show that the nonzero Christoﬀel coeﬃcients of the ﬁrst kind for spherical
coordinates are
Γ
221
= −r, Γ
122
= Γ
212
= r,
Γ
331
= −r sin
2
θ, Γ
332
= −r
2
sinθ cos θ,
Γ
313
= Γ
133
= r sin
2
θ, Γ
233
= Γ
323
= r
2
sin θ cos θ.
Tensor Fields 103
Exercise 4.13. Show that the only nonzero Christoﬀel coeﬃcients of the
second kind for cylindrical coordinates are
Γ
1
22
= −ρ, Γ
2
12
= Γ
2
21
= 1/ρ.
Show that the nonzero Christoﬀel coeﬃcients of the second kind for spher
ical coordinates are
Γ
1
22
= −r, Γ
2
12
= Γ
2
21
= 1/r,
Γ
1
33
= −r sin
2
θ, Γ
2
33
= −sinθ cos θ,
Γ
3
31
= Γ
3
13
= 1/r, Γ
3
23
= Γ
3
32
= cot θ.
Exercise 4.14. A system of plane elliptic coordinates (u, v) is introduced
according to the transformation formulas
x = c coshu cos v, y = c sinh u sinv,
where c is a constant. Find the Christoﬀel coeﬃcients.
Euclidean vs. nonEuclidean spaces
When we derive the length of the elementary vector dr we write
(ds)
2
= dr · dr = r
i
dq
i
· r
j
dq
j
= g
ij
dq
i
dq
j
. (4.23)
With respect to the variables dq
i
and dq
j
we have, by construction, a posi
tive deﬁnite quadratic form. This is one of the main properties of the metric
tensor for a real coordinate space.
If a space is Euclidean, that is, if it can be described by a set of Cartesian
coordinates x
t
(t = 1, 2, 3), then the line element (ds)
2
can be written as
(ds)
2
=
3
¸
t=1
(dx
t
)
2
. (4.24)
Given a set of admissible transformation equations x
t
= x
t
(q
n
), the same
line element can be expressed in terms of general coordinates q
n
. Since
dx
t
=
∂x
t
∂q
n
dq
n
we have
(ds)
2
=
3
¸
t=1
∂x
t
∂q
n
dq
n
2
,
104 Tensor Analysis with Applications in Mechanics
and comparison with (4.23) gives
g
ij
=
3
¸
t=1
∂x
t
∂q
i
∂x
t
∂q
j
(i, j = 1, 2, 3), (4.25)
for the metric coeﬃcients in the q
n
system. We now pose the following
question. Suppose a space is originally described in terms of a set of general
coordinates q
n
. Such a description must include a set of metric coeﬃcients
g
ij
having (4.23) positive deﬁnite at each point. Will it be possible to
introduce a set of Cartesian coordinates x
t
that also describe this space?
That is, are we guaranteed the existence of functions x
t
(q
n
) such that in the
resulting coordinates x
t
= x
t
(q
n
) the line element takes the form (4.24)?
Put still another way, is the space Euclidean? Not necessarily. Given the
g
ij
in the q
n
system, (4.25) provides us with six equations for the three
unknown functions x
t
(q
n
). This implies that some additional conditions
must be fulﬁlled by the given metric coeﬃcients. It is possible to formulate
these restrictions neatly in terms of a certain fourthorder tensor R
p
· ijk
as
the equality
R
p
· ijk
= 0. (4.26)
The tensor R
p
· ijk
is known as the Riemann–Christoﬀel tensor, and its as
sociated tensor
R
nijk
= g
np
R
p
· ijk
is called the curvature tensor for the space. It can be shown that
R
p
· ijk
=
∂Γ
p
ij
∂q
k
−
∂Γ
p
ik
∂q
j
−
Γ
p
mj
Γ
m
ik
−Γ
p
mk
Γ
m
ij
.
Moreover, R
p
· ijk
actually has only six independent components, so (4.26)
represents six conditions on the g
ij
that must hold for the space described
by the g
ij
to be Euclidean
[
Sokolnikoﬀ (1994)
]
.
We have included this information only for the reader’s background,
as such considerations are important in certain application areas. Further
details and more rigorous formulations can be found in some of the more
comprehensive references.
Tensor Fields 105
4.6 Covariant Diﬀerentiation
Let us return to the problem of diﬀerentiation of a vector function. We
obtained the formula for the derivative (4.13)
∂
∂q
k
f =
∂
∂q
k
(f
i
r
i
) =
∂f
i
∂q
k
+ Γ
i
kt
f
t
r
i
and observed that the coeﬃcients of this expansion denoted by
∇
k
f
i
=
∂f
i
∂q
k
+ Γ
i
kt
f
t
(4.27)
are called the covariant derivatives of contravariant components of vector
function f . Let us express the same derivatives in terms of the covariant
components:
∂
∂q
k
f =
∂
∂q
k
(f
i
r
i
) =
∂f
i
∂q
k
r
i
+ f
i
∂
∂q
k
r
i
.
Using (4.21) we get
∂
∂q
k
f =
∂f
i
∂q
k
r
i
−f
i
Γ
i
kt
r
t
=
∂f
i
∂q
k
−Γ
j
ki
f
j
r
i
.
The coeﬃcients of this expansion are called covariant derivatives of the
covariant components and are denoted by
∇
k
f
i
=
∂f
i
∂q
k
−Γ
j
ki
f
j
. (4.28)
In these notations we can write out the formula for diﬀerentiation in the
form
∂f
∂q
i
= r
k
∇
i
f
k
= r
j
∇
i
f
j
. (4.29)
Because
∇f = r
i
∂
∂q
i
(f
j
r
j
) = r
i
(∇
i
f
j
)r
j
= r
i
r
j
∇
i
f
j
,
we see that ∇
i
f
j
is a covariant component of ∇f . Similarly,
∇f = r
i
∂
∂q
i
(f
j
r
j
) = r
i
r
j
∇
i
f
j
shows that ∇
i
f
j
is a mixed component of ∇f .
Exercise 4.15. Write out the covariant derivatives of the contravariant
components of a vector f in plane polar coordinates.
106 Tensor Analysis with Applications in Mechanics
Quick summary
The formulas for diﬀerentiation of a vector are
∂f
∂q
i
= r
k
∇
i
f
k
= r
j
∇
i
f
j
where
∇
k
f
i
=
∂f
i
∂q
k
−Γ
j
ki
f
j
, ∇
k
f
i
=
∂f
i
∂q
k
+ Γ
i
kt
f
t
.
The formulas
∇f = r
i
r
j
∇
i
f
j
= r
i
r
j
∇
i
f
j
show that ∇
i
f
j
is a covariant component of ∇f , while ∇
i
f
j
is a mixed
component of ∇f .
4.7 Covariant Derivative of a SecondOrder Tensor
Let us ﬁnd a partial derivative of a tensor A of order two:
∂
∂q
k
A =
∂
∂q
k
(a
ij
r
i
r
j
)
=
∂a
ij
∂q
k
r
i
r
j
+ a
ij
(r
ik
r
j
+r
i
r
jk
)
=
∂a
ij
∂q
k
r
i
r
j
+ a
ij
(Γ
t
ik
r
t
r
j
+r
i
Γ
t
jk
r
t
).
Changing dummy indices we get
∂
∂q
k
A =
∂a
ij
∂q
k
+ Γ
i
ks
a
sj
+ Γ
j
ks
a
is
r
i
r
j
.
The parenthetical expression is designated as a covariant derivative:
∇
k
a
ij
=
∂a
ij
∂q
k
+ Γ
i
ks
a
sj
+ Γ
j
ks
a
is
.
Tensor Fields 107
Similarly
∂
∂q
k
A =
∂
∂q
k
(a
ij
r
i
r
j
)
=
∂a
ij
∂q
k
r
i
r
j
+ a
ij
∂r
i
∂q
k
r
j
+r
i
∂r
j
∂q
k
=
∂a
ij
∂q
k
r
i
r
j
−a
ij
(Γ
i
kt
r
t
r
j
+r
i
Γ
j
kt
r
t
)
=
∂a
ij
∂q
k
−Γ
s
ki
a
sj
−Γ
s
kj
a
is
r
i
r
j
.
As before we denote
∇
k
a
ij
=
∂a
ij
∂q
k
−Γ
s
ki
a
sj
−Γ
s
kj
a
is
.
Exercise 4.16. Show that
∂
∂q
k
A =
∂a
·j
i
∂q
k
−Γ
s
ki
a
·j
s
+ Γ
j
ks
a
·s
i
r
i
r
j
,
∂
∂q
k
A =
∂a
i
·j
∂q
k
+ Γ
i
ks
a
s
· j
−Γ
s
kj
a
i
·s
r
i
r
j
.
The expressions in parentheses are all denoted the same way:
∇
k
a
·j
i
=
∂a
·j
i
∂q
k
−Γ
s
ki
a
· j
s
+ Γ
j
ks
a
·s
i
,
∇
k
a
i
·j
=
∂a
i
·j
∂q
k
+ Γ
i
ks
a
s
·j
−Γ
s
kj
a
i
·s
.
Quick summary
For a tensor A of order two, we have
∂
∂q
k
A = ∇
k
a
ij
r
i
r
j
= ∇
k
a
ij
r
i
r
j
= ∇
k
a
·j
i
r
i
r
j
= ∇
k
a
i
·j
r
i
r
j
where
∇
k
a
ij
=
∂a
ij
∂q
k
+ Γ
i
ks
a
sj
+ Γ
j
ks
a
is
, ∇
k
a
ij
=
∂a
ij
∂q
k
−Γ
s
ki
a
sj
−Γ
s
kj
a
is
,
∇
k
a
·j
i
=
∂a
·j
i
∂q
k
−Γ
s
ki
a
· j
s
+ Γ
j
ks
a
·s
i
, ∇
k
a
i
·j
=
∂a
i
·j
∂q
k
+ Γ
i
ks
a
s
·j
−Γ
s
kj
a
i
·s
.
Exercise 4.17. Show that any of the covariant derivatives of any compo
nent of the metric tensor is equal to zero.
108 Tensor Analysis with Applications in Mechanics
Exercise 4.18. Demonstrate that the components of the metric tensor
behave as constants under covariant diﬀerentiation of components:
∇
k
g
st
a
t
= g
st
∇
k
a
t
, ∇
k
g
st
a
t
= g
st
∇
k
a
t
.
4.8 Diﬀerential Operations
Here we look further at various diﬀerential operations that may be per
formed on vector and tensor ﬁelds. We begin with the rotation of a vector.
By (4.7) and (4.29) we have
∇×f = r
i
×
∂f
∂q
i
= r
i
×r
j
∇
i
f
j
=
ijk
r
k
∇
i
f
j
.
The use of (4.28) allows us to write this as
∇×f =
ijk
r
k
∂f
j
∂q
i
−Γ
n
ij
f
n
. (4.30)
Considering the second term in parentheses we note that
ijk
Γ
n
ij
=
ijk
Γ
n
ji
=
jik
Γ
n
ij
by (4.12) and a subsequent renaming of dummy indices. But
jik
= −
ijk
,
hence
ijk
Γ
n
ij
= −
ijk
Γ
n
ij
so that
ijk
Γ
n
ij
= 0. Equation (4.30) is thereby
reduced to
∇×f = r
k
ijk
∂f
j
∂q
i
.
As an example, we recall that the vector ﬁeld E of electrostatics satisﬁes
∇×E = 0. Such a ﬁeld is said to be irrotational. So the condition for any
vector ﬁeld f to be irrotational can be written in generalized coordinates as
ijk
∂f
j
∂q
i
= 0.
Of course, the operation ∇×f is given in a Cartesian frame by the familiar
formula
∇×f =
ˆ x ˆ y ˆ z
∂
∂x
∂
∂y
∂
∂z
f
x
f
y
f
z
.
Let us turn to the divergence of f . We start by writing
∇· f = r
i
·
∂f
∂q
i
= r
i
· r
j
∇
i
f
j
= g
ij
∇
i
f
j
.
Tensor Fields 109
By the result of Exercise 4.18 and equation (4.27) this can be written as
∇· f = ∇
i
g
ij
f
j
= ∇
i
f
i
=
∂f
i
∂q
i
+ Γ
i
in
f
n
. (4.31)
As was the case with (4.30) this can be simpliﬁed; we must ﬁrst develop a
useful identity for Γ
i
in
. This is done as follows. We begin by writing
∂
√
g
∂q
n
=
∂
∂q
n
[r
1
· (r
2
×r
3
)] =
∂r
1
∂q
n
· (r
2
×r
3
) +r
1
·
∂
∂q
n
(r
2
×r
3
)
where
r
1
·
∂
∂q
n
(r
2
×r
3
) = r
1
· r
2
×
∂r
3
∂q
n
+r
1
·
∂r
2
∂q
n
×r
3
=
∂r
3
∂q
n
· (r
1
×r
2
) +
∂r
2
∂q
n
· (r
3
×r
1
)
so that
∂
√
g
∂q
n
=
∂r
1
∂q
n
· (r
2
×r
3
) +
∂r
2
∂q
n
· (r
3
×r
1
) +
∂r
3
∂q
n
· (r
1
×r
2
).
Continuing to rewrite this we have
∂
√
g
∂q
n
= r
1n
· (r
2
×r
3
) +r
2n
· (r
3
×r
1
) +r
3n
· (r
1
×r
2
)
= Γ
i
1n
r
i
· (r
2
×r
3
) + Γ
i
2n
r
i
· (r
3
×r
1
) + Γ
i
3n
r
i
· (r
1
×r
2
)
= Γ
1
1n
√
g + Γ
2
2n
√
g + Γ
3
3n
√
g
=
√
g Γ
i
in
.
Hence
Γ
i
in
=
1
√
g
∂
√
g
∂q
n
.
This is the needed identity, and with it (4.31) may be written as
∇· f =
1
√
g
∂
∂q
i
√
gf
i
. (4.32)
We know that the condition of incompressibility of a liquid in hydrome
chanics is expressed as
∇· v = 0
where v is the velocity of a material point, so in general coordinates we can
write it out as
1
√
g
∂
∂q
i
√
gv
i
= 0.
110 Tensor Analysis with Applications in Mechanics
In electromagnetic theory the magnetic source law states that ∇ · B = 0
where B is the vector ﬁeld known as magnetic ﬂux density. In general, a
vector ﬁeld f is called solenoidal if ∇ · f = 0. Of course, ∇ · f is given in
Cartesian frames by the familiar expression
∇· f =
∂f
x
∂x
+
∂f
y
∂y
+
∂f
z
∂z
.
Exercise 4.19. Use (4.32) to express ∇· f in the cylindrical and spherical
coordinate systems.
The curl of a tensor ﬁeld A may be computed as follows. We start with
∇×A = r
k
×
∂
∂q
k
A
= r
k
×r
i
r
j
∇
k
a
ij
=
kin
r
n
r
j
∇
k
a
ij
=
kin
r
n
r
j
∂a
ij
∂q
k
−Γ
s
ki
a
sj
−Γ
s
kj
a
is
.
We then use
kin
Γ
s
ki
= 0, −r
j
Γ
s
kj
=
∂r
s
∂q
k
,
to get
∇× A =
kin
r
n
r
j
∂a
ij
∂q
k
+
∂r
s
∂q
k
a
is
and hence
∇×A =
kin
r
n
∂
∂q
k
r
j
a
ij
.
For the divergence of a tensor ﬁeld A, we write
∇· A = r
k
·
∂
∂q
k
A
= r
k
· r
i
r
j
∇
k
a
ij
= r
j
∇
i
a
ij
= r
j
∂a
ij
∂q
i
+ Γ
i
is
a
sj
+ Γ
j
is
a
is
.
But
r
j
Γ
j
is
=
∂r
s
∂q
i
Tensor Fields 111
so we have
∇· A = r
j
∂a
ij
∂q
i
+r
j
1
√
g
∂
√
g
∂q
s
a
sj
+
∂r
s
∂q
i
a
is
=
∂
∂q
i
r
j
a
ij
+r
j
1
√
g
∂
√
g
∂q
i
a
ij
.
Finally then,
∇· A =
1
√
g
∂
∂q
i
√
ga
ij
r
j
. (4.33)
Many problems of mathematical physics reduce to Poisson’s equation
or Laplace’s equation. The unknown function in these equations can be a
scalar function or (as in electrodynamics) a vector function. The equations
of the linear theory of elasticity in displacements also contain the Laplacian
operator and another type of operation involving the nabla operator. In
Cartesian frames it is simple to write out corresponding expressions. Solv
ing corresponding problems with the use of curvilinear coordinates, we need
to ﬁnd the representation of these formulas. We begin with the formulas
that relate to the secondorder tensor ∇∇f. We have
∇∇f = r
i
∂
∂q
i
r
j
∂
∂q
j
f
= r
i
∂
∂q
i
∂f
∂q
j
−Γ
k
ij
∂f
∂q
k
r
j
,
hence
∇∇f = r
i
r
j
∂
2
f
∂q
i
∂q
j
−Γ
k
ij
∂f
∂q
k
. (4.34)
From this we see that
(∇∇f)
T
= ∇∇f (4.35)
which is obvious by symmetry (in i and j) of the Christoﬀel coeﬃcient and
the rest of the expression on the right side of (4.34). A formal insertion
2
of the dot product operation between the vectors of (4.34) allows us to
generate an expression for the Laplacian ∇
2
f ≡ ∇· ∇f:
∇
2
f = r
i
· r
j
∂
2
f
∂q
i
∂q
j
−Γ
k
ij
∂f
∂q
k
= g
ij
∂
2
f
∂q
i
∂q
j
−Γ
k
ij
∂f
∂q
k
. (4.36)
2
This sort of operation can be done with any dyad of vectors; we can generate a scalar
a· b from ab by inserting a dot. This operation can have additional meaning when done
with a tensor A: if we write the tensor in mixed form a
·j
i
r
i
r
j
we obtain a
·j
i
r
i
· r
j
= a
·i
i
.
This is the ﬁrst invariant of A, known as the trace of A.
112 Tensor Analysis with Applications in Mechanics
Hence Laplace’s equation ∇
2
f = 0 appears in generalized coordinates as
g
ij
∂
2
f
∂q
i
∂q
j
−Γ
k
ij
∂f
∂q
k
= 0.
In a Cartesian frame (4.36) gives us
∇
2
f =
∂
2
f
∂x
2
+
∂
2
f
∂y
2
+
∂
2
f
∂z
2
,
while in cylindrical and spherical frames
∇
2
f =
1
ρ
∂
∂ρ
ρ
∂f
∂ρ
+
1
ρ
2
∂
2
f
∂φ
2
+
∂
2
f
∂z
2
,
∇
2
f =
1
r
2
∂
∂r
r
2
∂f
∂r
+
1
r
2
sin θ
∂
∂θ
sin θ
∂f
∂θ
+
1
r
2
sin
2
θ
∂
2
f
∂φ
2
,
respectively.
Exercise 4.20. Show that a formal insertion of the cross product operation
between the vectors of (4.34) leads to the useful identity
∇×∇f = 0,
holding for any scalar ﬁeld f.
Since usual and covariant diﬀerentiation amount to the same thing for
a scalar f, we can write
∇∇f = r
i
∂
∂q
i
r
j
∂
∂q
j
f = r
i
∂
∂q
i
r
j
∇
j
f
= r
i
r
j
∇
i
∇
j
f = r
j
r
i
∇
i
∇
j
f
where in the last step we used (4.35). The corresponding result for the
Laplacian is
∇
2
f = r
i
· r
j
∇
i
∇
j
f = g
ij
∇
i
∇
j
f = ∇
j
∇
j
f
where
∇
j
≡ g
ij
∇
i
.
The Laplacian of a vector arises in physical applications such as elec
tromagnetic ﬁeld theory. For this we have
∇
2
f = ∇· ∇f = r
k
·
∂
∂q
k
r
i
r
j
∇
i
f
j
= r
k
· r
i
r
j
∇
k
∇
i
f
j
= g
ki
r
j
∇
k
∇
i
f
j
hence
∇
2
f = r
j
∇
i
∇
i
f
j
.
Tensor Fields 113
Also
∇∇· f = r
i
∂
∂q
i
¸
1
√
g
∂
∂q
j
√
gf
j
= r
i
∇
i
∇
j
f
j
.
It is possible to demonstrate that
∇×∇×f = ∇∇· f −∇
2
f .
4.9 Orthogonal Coordinate Systems
The most frequently used coordinate frames are Cartesian, cylindrical, and
spherical. All of these are orthogonal. There are many other orthogonal
coordinate frames in use as well. It is sensible to give a general treatment
of these systems because mutual orthogonality of the frame vectors leads
to simpliﬁcation in many formulae. Additional motivation is provided by
the fact that in applications it is important to know the magnitudes of ﬁeld
components. The general formulations given above are inconvenient for
this; the general frame vectors are not of unit length, hence the magnitude
of the projection of a vector onto an orthogonal frame direction is not
the corresponding component of the vector. The physical components of
a vector are conveniently displayed in orthogonal frames with use of the
Lam´e coeﬃcients.
In this section we consider frames where the coordinate vectors are
mutually orthogonal:
r
i
· r
j
= 0, i = j.
The Lam´e coeﬃcients H
i
are
(H
i
)
2
= g
ii
= r
i
· r
i
(i = 1, 2, 3).
Note that H
i
is the length of the frame vector r
i
. At each (q
1
, q
2
, q
3
) the
coordinate frame is orthogonal. In the orthogonal frame, by the construc
tion of the reciprocal basis, the vectors r
i
are codirected with r
i
and the
product of their lengths is 1. Hence
r
i
= r
i
/(H
i
)
2
(i = 1, 2, 3).
Let us introduce the frame whose vectors are codirected with the basis
vectors but have unit length:
ˆr
i
= r
i
/H
i
(i = 1, 2, 3).
114 Tensor Analysis with Applications in Mechanics
Thus at each point the vectors ˆr
i
form a Cartesian basis, a frame that
rotates when the origin of the frame moves from point to point. We shall
present all the main formulas of diﬀerentiation when vectors are given in
this basis. They are not convenient in theory since many useful properties
of symmetry are lost, but they are necessary when doing calculations in
corresponding coordinates. We begin by noting that
r
i
= H
i
ˆr
i
, r
i
= ˆr
i
/H
i
(i = 1, 2, 3).
Let us compute the ˆr
i
in the cylindrical and spherical systems. In a cylin
drical frame where
H
1
= 1, H
2
= ρ, H
3
= 1,
we have
ˆr
1
= ˆ xcos φ + ˆ y sin φ,
ˆr
2
= −ˆ xsin φ + ˆ ycos φ,
ˆr
3
= ˆ z.
In a spherical frame where
H
1
= 1, H
2
= r, H
3
= r sin θ,
we have
ˆr
1
= ˆ xsin θ cos φ + ˆ y sinθ sinφ + ˆ z cos θ,
ˆr
2
= ˆ xcos θ cos φ + ˆ y cos θ sin φ −ˆz sin θ,
ˆr
3
= −ˆ xsin φ + ˆ ycos φ.
Diﬀerentiation in the orthogonal basis
Using the deﬁnition we can represent the ∇operator in new terms as
∇ =
ˆr
i
H
i
∂
∂q
i
.
(In this formula there is summation over i, and this continues to hold in
all formulas of this section below. The convention on summation over sub
and superindices is modiﬁed here since in a Cartesian system, which the
frame ˆr
i
locally constitutes, the reciprocal and main bases coincide.) Now
let us ﬁnd the formulas of diﬀerentiation of the new frame vectors. We
Tensor Fields 115
begin with the formula (4.18):
r
ij
· r
t
=
1
2
∂g
it
∂q
j
+
∂g
tj
∂q
i
−
∂g
ji
∂q
t
= H
t
∂H
i
∂q
j
δ
it
+ H
t
∂H
j
∂q
i
δ
jt
−H
i
∂H
j
∂q
t
δ
ij
.
Here we used the fact that g
ij
= 0 if i = j. On the other hand
r
ij
· r
t
=
∂
∂q
j
(H
i
ˆr
i
) · H
t
ˆr
t
=
∂H
i
∂q
j
ˆr
i
· H
t
ˆr
t
+ H
i
∂ˆr
i
∂q
j
· H
t
ˆr
t
=
∂H
i
∂q
j
H
t
δ
it
+ H
i
H
t
∂ˆr
i
∂q
j
· ˆr
t
.
Thus
∂ˆr
i
∂q
j
· ˆr
t
=
1
H
i
∂H
t
∂q
i
δ
jt
−
1
H
t
∂H
i
∂q
t
δ
ij
.
Therefore
∂ˆr
i
∂q
j
=
3
¸
t=1
1
H
i
∂H
t
∂q
i
δ
jt
−
1
H
t
∂H
i
∂q
t
δ
ij
ˆr
t
.
(Note that the components of vectors are given in the frame of the same
ˆr
i
.) Using this we derive
∇f =
3
¸
i,j=1
ˆr
j
H
j
∂
∂q
j
(f
i
ˆr
i
) .
The gradient, divergence, and rotation of a vector ﬁeld f are given by
∇f = ˆr
i
ˆr
j
1
H
i
∂f
j
∂q
i
−
f
i
H
i
H
j
∂H
i
∂q
j
+ δ
ij
f
k
H
k
1
H
i
∂H
i
∂q
k
∇· f =
1
H
1
H
2
H
3
∂
∂q
1
(H
2
H
3
f
1
) +
∂
∂q
2
(H
3
H
1
f
2
) +
∂
∂q
3
(H
1
H
2
f
3
)
(4.37)
and
∇×f =
1
2
ˆr
i
×ˆr
j
H
i
H
j
∂
∂q
i
(H
j
f
j
) −
∂
∂q
j
(H
i
f
i
)
. (4.38)
116 Tensor Analysis with Applications in Mechanics
Using the gradient we can write out the tensor of small strains for a dis
placement vector u = u
i
ˆr
i
, which is the main object of linear elasticity:
ε =
1
2
∇u +∇u
T
=
1
2
ˆr
i
ˆr
j
1
H
i
∂u
i
∂q
i
+
1
H
j
∂u
j
∂q
j
−
u
i
H
i
H
j
∂H
i
∂q
j
(4.39)
−
u
j
H
i
H
j
∂H
j
∂q
i
+ 2δ
ij
u
t
H
i
H
j
∂H
j
∂q
t
.
The Laplacian of a scalar ﬁeld f is
∇
2
f =
1
H
1
H
2
H
3
¸
∂
∂q
1
H
2
H
3
H
1
∂f
∂q
1
+
∂
∂q
2
H
3
H
1
H
2
∂f
∂q
2
+
∂
∂q
3
H
1
H
2
H
3
∂f
∂q
3
. (4.40)
In the cylindrical and spherical coordinate systems, for example, (4.37)
yields
∇· f =
1
ρ
∂
∂ρ
(ρf
ρ
) +
1
ρ
∂f
φ
∂φ
+
∂f
z
∂z
and
∇· f =
1
r
2
∂
∂r
r
2
f
r
+
1
r sin θ
∂
∂θ
(sin θf
θ
) +
1
r sin θ
∂f
φ
∂φ
,
while (4.38) yields
∇×f = ˆ ρ
1
ρ
∂f
z
∂φ
−
∂f
φ
∂z
+
ˆ
φ
∂f
ρ
∂z
−
∂f
z
∂ρ
+ ˆz
1
ρ
∂(ρf
φ
)
∂ρ
−
∂f
ρ
∂φ
and
∇×f = ˆr
1
r sinθ
∂
∂θ
(sin θf
φ
) −
∂f
θ
∂φ
+
ˆ
θ
1
r sin θ
∂f
r
∂φ
−sin θ
∂(rf
φ
)
∂r
+
ˆ
φ
1
r
∂(rf
θ
)
∂r
−
∂f
r
∂θ
.
The specializations of (4.40) to cylindrical and spherical coordinates were
given in § 4.8.
Tensor Fields 117
4.10 Some Formulas of Integration
Let f(x
1
, x
2
, x
3
) be a continuous function of the Cartesian coordinates
x
1
, x
2
, x
3
in a compact volume V . Let q
1
, q
2
, q
3
be curvilinear coordinates in
the same volume, in onetoone continuously diﬀerentiable correspondence
with the Cartesian coordinates, so after transformation of the coordinates
we shall write out the same function as f(q
1
, q
2
, q
3
). The transformation is
V
f(x
1
, x
2
, x
3
) dx
1
dx
2
dx
3
=
V
f(q
1
, q
2
, q
3
)J dq
1
dq
2
dq
3
where
J =
√
g =
∂x
i
∂q
j
is the Jacobian.
Exercise 4.21. Show that the Jacobian determinants of the transforma
tions from Cartesian coordinates to cylindrical and spherical coordinates
are, respectively,
∂(x, y, z)
∂(ρ, φ, z)
= ρ,
∂(x, y, z)
∂(r, θ, φ)
= r
2
sin θ.
Exercise 4.22. Two successive coordinate transformations are given by
x
i
= x
i
(q
j
) and q
i
= q
i
(˜ q
j
). Show that the Jacobian determinant of the
composite transformation x
i
= x
i
(˜ q
j
) is given by
∂(x
1
, x
2
, x
3
)
∂(˜ q
1
, ˜ q
2
, ˜ q
3
)
=
∂(x
1
, x
2
, x
3
)
∂(q
1
, q
2
, q
3
)
∂(q
1
, q
2
, q
3
)
∂(˜ q
1
, ˜ q
2
, ˜ q
3
)
.
For functions f(x
1
, x
2
, x
3
) and g(x
1
, x
2
, x
3
) that are continuously diﬀer
entiable on a compact volume V with piecewise smooth boundary S, there
is the well known Gauss–Ostrogradsky formula for integration by parts:
V
∂f
∂x
k
g dx
1
dx
2
dx
3
= −
V
∂g
∂x
k
f dx
1
dx
2
dx
3
+
S
fgn
k
dS,
where dS is the diﬀerential element of area on S and n
k
is the projection of
the outward unit normal n from S onto the axis i
k
. In the particular case
g = 1 we have
V
∂f
∂x
k
dx
1
dx
2
dx
3
=
S
fn
k
dS. (4.41)
118 Tensor Analysis with Applications in Mechanics
Let us use (4.41) to derive some formulas involving the nabla operator that
are frequently used in applications. These formulas will be valid in curvilin
ear coordinates, despite the fact that the intermediate transformations will
be done in Cartesian coordinates, because the ﬁnal results will be written
in noncoordinate form. We begin with the integral of ∇f:
V
∇f dV =
V
i
k
∂f
∂x
k
dx
1
dx
2
dx
3
= i
k
S
fn
k
dS =
S
fndS.
Now consider an analogous formula for a vector function f :
V
∇f dV =
V
i
k
∂f
t
∂x
k
i
t
dx
1
dx
2
dx
3
= i
k
i
t
S
n
k
f
t
dS
=
S
n
k
i
k
f
t
i
t
dS
=
S
nf dS.
Since the left and righthand sides are written in noncoordinate form we
can use this formula with any coordinate frame with dV =
√
g dq
1
dq
2
dq
3
.
This is the formula for a tensor ∇f ; for its trace we have
V
∇· f dV =
V
i
k
∂f
t
∂x
k
· i
t
dx
1
dx
2
dx
3
= i
k
· i
t
S
n
k
f
t
dS
=
S
n
k
i
k
· f
t
i
t
dS
=
S
n · f dS.
In a similar fashion we can derive
V
∇×f dV =
S
n ×f dS.
It is easily seen that
V
∇f
T
dV =
S
(nf )
T
dS =
S
fndS.
In a similar fashion the reader can use (4.41) to derive the formulas
V
∇AdV =
S
nAdS,
Tensor Fields 119
V
∇· AdV =
S
n · AdS,
and
V
∇×AdV =
S
n ×AdS
for a tensor ﬁeld A. Finally let us write out Stokes’s formula in non
coordinate form. Recall that this relates a vector function f given on a
simplyconnected surface S to its circulation over the piecewise smooth
boundary contour Γ:
Γ
f · dr =
S
(n ×∇) · f dS.
For a tensor A of secondorder this formula extends to the two formulas
Γ
dr · A =
S
(n ×∇) · AdS
and
Γ
A· dr =
S
(n ×∇) · A
T
dS.
We leave these for the reader to prove. We should note that Stokes’s for
mulas hold only when S is simplyconnected; for a doubly or multiply
connected surface, the formulas must be amended by some cyclic constants.
4.11 Problems
In this problem set, we let u denote a vector ﬁeld; f, g, h smooth functions;
a, b, c arbitrary constants; A a secondorder tensor; r the position vector of
a body point; n the unit external normal to the body boundary.
4.1 Let f = f(r), where r
2
= r · r. Find ∇f and ∇
2
f.
4.2 Let a, b, c be arbitrary constants, f, g arbitrary functions, and u a
given vector ﬁeld. Find ∇u and ∇· u for the following u.
(a) u = ax
1
i
1
+ bx
2
i
2
+ cx
3
i
3
;
(b) u = ax
2
i
1
;
(c) u = ar;
(d) u = f(r)e
r
(assume polar coordinates);
(e) u = f(r)e
φ
(assume polar coordinates);
120 Tensor Analysis with Applications in Mechanics
(f) u = f(r)e
z
(assume cylindrical coordinates);
(g) u = f(r)e
r
(assume spherical coordinates);
(h) u = ω ×r, ω = const;
(i) u = f(φ)e
z
+ g(φ)e
φ
(assume cylindrical coordinates);
(j) u = f(z)e
z
+ g(φ)e
φ
(assume cylindrical coordinates);
(k) u = A· r, A = const.
4.3 Demonstrate:
(a) ∇· (A· f ) = (∇· A) · f +A
T
·· ∇f ;
(b) ∇· (A· B) = (∇· A) · B+A
T
·· ∇B;
(c) ∇· (A×r) = (∇· A) ×r, if A is a symmetric tensor: A = A
T
;
(d) ∇· [(E ×ω) ×r] = 2ω + (∇×ω) ×r;
(e) ∇×(∇×A) = ∇(∇· A) −∇· (∇A);
(f) ∇×(f × r) = r · ∇f −r(∇· f ) + 2f ;
(g) tr [∇×(E ×ω)] = −2∇· ω;
(h) ∇·
(∇f )
T
−(∇· f )E
= 0;
(i) ∇· [∇f ×∇g] = 0;
(j) ∇×(∇×A)
T
is a symmetric tensor if A is symmetric;
(k) ∇· (f f ) = f ∇· f + (∇f ) · f −f ×(∇×f );
(l) tr [∇×(A· B)] = B·· (∇×A) −A
T
·· (∇×B
T
);
(m) (∇f )
×
= ∇×f .
4.4 Find
(a) ∇· (Er);
(b) ∇×[∇×(r ×A×r)]
T
, if A = A
T
= const;
(c) ∇· (A· r) if A = const;
(d) ∇· (fE);
(e) ∇· (rE);
(f) ∇· (rr).
4.5 Let f, g, h be arbitrary smooth functions and r
2
= r · r. Find the
divergence of the tensors A given by the following formulas.
(a) A = f(r)e
r
e
r
+ g(r)e
φ
e
φ
+ h(z)e
z
e
z
;
(b) A = f(r)e
r
e
r
+ g(r)e
φ
e
φ
+ g(r)e
θ
e
θ
;
(c) A = f(r)e
r
e
r
+ g(r)e
φ
e
φ
+ h(r)e
z
e
z
;
Tensor Fields 121
(d) A = f(x
1
)i
1
i
1
+ g(x
2
)i
2
i
2
+ h(x
3
)i
3
i
3
;
(e) A = f(x
2
)i
1
i
1
+ g(x
3
)i
2
i
2
+ h(x
1
)i
3
i
3
;
(f) A = f(r)e
r
e
φ
+ g(r)e
φ
e
r
+ h(r)e
r
e
z
;
(g) A = f(z)e
r
e
z
+ g(z)e
φ
e
r
+ h(z)e
z
e
r
.
4.6 Let A be a symmetric secondorder tensor depending on the coordi
nates. Denote f = ∇· A. Find ∇· (A×r) in terms of f .
4.7 The elliptic cylindrical coordinates are related to Cartesian coordinates
by the formulas
x
1
= aστ,
x
2
= ±a
(σ
2
−1)(1 −τ
2
),
x
3
= z,
where σ ≥ 1, τ ≤ 1, and a is a positive parameter. So σ, τ, z are inter
nal coordinates for the cylinder. Show that the internal coordinates are
orthogonal and ﬁnd their Lam´e coeﬃcients.
4.8 Parabolic coordinates σ, τ, φ in space are related to the Cartesian co
ordinates x
1
, x
2
, x
3
by the formulas
x
1
= στ cos φ,
x
2
= στ sin φ,
x
3
=
1
2
(τ
2
−σ
2
),
where a is a positive parameter. Show that the parabolic system of coor
dinates is orthogonal. Find its Lam´e coeﬃcients.
4.9 The bipolar cylindrical coordinates σ, τ, z are related to the Cartesian
coordinates x
1
, x
2
, x
3
by the formulas
x
1
=
a sinhτ
coshτ −cos σ
,
x
2
=
a sin σ
coshτ −cos σ
,
x
3
= z,
where a is a positive parameter. Show that the bicylindrical coordinate
system is orthogonal. Find its Lam´e coeﬃcients.
122 Tensor Analysis with Applications in Mechanics
4.10 The bipolar coordinates σ, τ, φ are related to the Cartesian coordi
nates x
1
, x
2
, x
3
by the formulas
x
1
=
a sin σ
coshτ −cos σ
cos φ,
x
2
=
a sin σ
coshτ −cos σ
sinφ,
x
3
=
a sinh τ
coshτ −cos σ
,
where 0 ≤ σ < π, 0 ≤ φ < 2π, and a is a positive parameter. Show that
the bipolar coordinate system is orthogonal. Find its Lam´e coeﬃcients.
4.11 The toroidal coordinates σ, τ, φ are related to Cartesian coordinates
by the formulas
x
1
=
a sinh τ
coshτ −cos σ
cos φ,
x
2
=
a sinh τ
coshτ −cos σ
sinφ,
x
3
=
a sin σ
coshτ −cos σ
,
where −π ≤ σ ≤ π, 0 ≤ τ, 0 ≤ φ < 2π, and a is a positive parameter.
Show that the toroidal coordinate system is orthogonal. Find its Lam´e
coeﬃcients.
4.12 Use the Gauss–Ostrogradsky theorem to show that
S
nr dS = V E,
where V is the volume of the domain bounded by surface S.
4.13 Show that the volume V of the body bounded by surface S is given
by the following formulas.
(a)
V =
1
6
S
n · ∇r
2
dS, r
2
= r · r,
(b)
V =
1
3
S
n · r dS.
Tensor Fields 123
4.14 Let a be a vector ﬁeld satisfying the condition ∇· a = 0. Using the
Gauss–Ostrogradsky formula, demonstrate that
S
fnadS =
V
(∇f) · adV.
4.15 Let S be a closed surface. Demonstrate that
S
ndS = 0.
4.16 Let the secondorder tensor A be symmetric. Prove the following
identity.
S
r ×(n · A) dS =
V
r ×(∇· A) dV.
4.17 Prove the identity
V
∇
2
AdV =
S
n · ∇AdS.
4.18 Let A be a given secondorder tensor. Denote f = ∇· A in volume
V and n · A
S
= g over S, the boundary surface of V . Find
V
AdV.
4.19 Prove the identity (n ×∇) · A = n · rot A.
Chapter 5
Elements of Diﬀerential Geometry
The standard fare of high school geometry consists mostly of material col
lected two millennia ago when geometry stood at the center of natural
philosophy. The ancient Greeks, however, did not limit their investigations
to the circles and straight lines of Euclid’s Elements. Archimedes, using
methods and ideas that were later to underpin the analysis of inﬁnitesimal
quantities, could calculate the length of a spiral and the areas and volumes
of other complex ﬁgures. In elementary algebra we learn to graph simple
quadratic functions such as the parabola and hyperbola, and then in ana
lytic geometry we learn to handle space ﬁgures such as the ellipsoid. The
methods involved are essentially due to Descartes, who connected the ideas
of geometry with those of algebra, and their application is largely limited
to objects whose describing equations are of the second order. Finally,
in elementary calculus we study formulas that permit us to calculate the
length of a curve given in Cartesian coordinates, etc. These more power
ful methods are now incorporated into a branch of mathematics known as
diﬀerential geometry.
Diﬀerential geometry allows us to characterize curves and ﬁgures of a
very general nature. The practical importance of this is well illustrated by
the problem of optimal pursuit, wherein one object tries to catch another
moving object in the shortest possible time.
1
Of course, we shall often make
use of standard ﬁgures such as circles, parabolas, etc., as speciﬁc examples
since we are fully familiar with their properties; in this way the objects of
both elementary and analytic geometry enter into the more general subject
of diﬀerential geometry.
1
A rather humorous statement of one such problem has two old ladies traveling in
opposite directions around the base of a hemispherical mountain while a mathematically
minded ﬂy travels back and forth between their noses in the least possible time.
125
126 Tensor Analysis with Applications in Mechanics
5.1 Elementary Facts from the Theory of Curves
In Chapter 4 we introduced the idea of a coordinate curve in space, which is
described by the tip of the radius vector as it moves in such a way that one
of the coordinates q
1
, q
2
, q
3
changes while the other two remain ﬁxed. Now
we consider a general curve described in a similar manner by a radius vector
whose initial point is the origin and whose terminal point moves through
space along the curve. We can describe the position of the radius vector
using some parameter t (for a coordinate curve this was the coordinate
value q
i
). Thus a curve is described as
r = r(t) (5.1)
where t runs through some set along the real axis. Each value of t corre
sponds to a point of the curve. Unless otherwise stated we shall suppose
that the dependence of r on t is smooth enough that r
(t) is continuous in
t at each point and, where necessary, that the same holds for r
(t). The
notion of vector norm is required for this (§ 2.7). Here we denote this norm
using the ordinary notation for the magnitude of a vector, e.g., r. Recall
that in Chapter 4 we introduced frame vectors r
i
tangential to the coordi
nate lines. We now introduce the tangential vector to an arbitrary curve at
a point t, which is r
(t). The diﬀerential of the radius vector corresponding
to the curve (5.1) is
dr(t) = r
(t) dt.
The length of an elementary section of the curve is
ds = r
(t) dt,
and the length of the portion of the curve corresponding to t ∈ [a, b] is
s =
b
a
r
(t) dt.
Exercise 5.1. (a) Find the length of one turn of the helix
r(t) = i
1
cos t +i
2
sin t +i
3
t.
(b) Calculate the perimeter of the ellipse
x
2
A
2
+
y
2
B
2
= 1
that lies in the z = 0 plane.
Elements of Diﬀerential Geometry 127
Exercise 5.2. Show that the general formulas for arc length in the rect
angular, cylindrical, and spherical coordinate systems are
s =
b
a
¸
dx
dt
2
+
dy
dt
2
+
dz
dt
2
¸
1/2
dt,
s =
b
a
¸
dρ
dt
2
+ρ
2
dφ
dt
2
+
dz
dt
2
¸
1/2
dt,
s =
b
a
¸
dr
dt
2
+r
2
dθ
dt
2
+r
2
sin
2
θ
dφ
dt
2
¸
1/2
dt.
Most convenient theoretically is the natural parametrization of a curve
where the parameter represents the length of the curve calculated from the
endpoint:
r = r(s).
With this parametrization
ds = r
(s) ds so r
(s) = 1.
Hence when a curve is parametrized naturally τ(s) = r
(s) is the unit tan
gential vector at point s. In this section s shall denote the length parameter
of a curve.
Exercise 5.3. Reparametrize the helix of Exercise 5.1 in terms of its nat
ural length parameter s. Then calculate the unit tangent to the helix as a
function of s.
If there is another parametrization of the curve that relates with the
natural parametrization s = s(t), then the vector dr(s(t))/dt also is tangent
to the curve at the point s = s(t). For the derivative of a vector function
the chain rule
dr(s(t))
dt
=
dr(s)
ds
s=s(t)
ds(t)
dt
holds.
As is known from analytic geometry, when we know the position of a
point a of a straight line and its directional vector b, then the line can be
represented in the parametric form
r = a +λb.
128 Tensor Analysis with Applications in Mechanics
This gives the vector equation of the line tangent to a curve at point r(t
0
):
r = r(t
0
) +λr
(t
0
).
In Cartesian coordinates (x, y, z) this equation takes the form
x = x(t
0
) +λx
(t
0
),
y = y(t
0
) +λy
(t
0
),
z = z(t
0
) +λz
(t
0
), (5.2)
which in nonparametric form is
x −x(t
0
)
x
(t
0
)
=
y −y(t
0
)
y
(t
0
)
=
z −z(t
0
)
z
(t
0
)
. (5.3)
Exercise 5.4. Describe the plane curve
r(t) = e
t
(i
1
cos t +i
2
sint),
and ﬁnd the line tangent to this curve at the point t = π/4.
Exercise 5.5. (a) Write out the equations for the tangent line to a plane
curve corresponding to (5.2) and (5.3) in polar coordinates. (b) Find the
angle between the tangent at a point of the curve and the radius vector
(from the origin) at this point.
Exercise 5.6. Write out the equations for the tangent line to a curve
corresponding to (5.2) and (5.3) in cylindrical and spherical coordinates.
Note that construction of a tangent vector is possible if r
(t
0
) = 0. A
point t
0
where r
(t
0
) = 0 is called singular.
Exercise 5.7. For a suﬃciently smooth curve r = r(t) ﬁnd a tangent at a
singular point.
Exercise 5.8. Under the conditions of the previous exercise write out the
representation of the type (5.2).
Curvature
For an element of circumference of a circle corresponding to the length ∆s,
the radius R relates to the central angle ∆ϕ according to ∆s = R∆ϕ. Thus
the curvature k = 1/R is
k =
∆ϕ
∆s
.
Elements of Diﬀerential Geometry 129
j
j
Fig. 5.1 Calculation of curvature.
The curvature of an arbitrary plane curve at a point s is deﬁned as the
limit
k = lim
∆s→0
∆ϕ
∆s
where ∆ϕ is the change in angle of the tangent to the element of the
curve. For a spatial curve the role of the tangent is played by the tangent
unit vector τ(s). When a point moves through the element of the curve
corresponding to ∆s, the tangent turns through an angle ∆ϕ. Fig. 5.1
demonstrates that τ(s + ∆s) −τ(s) is equal to 2 sin(∆ϕ/2). Since
lim
∆ϕ→0
sin(∆ϕ/2)
(∆ϕ/2)
= 1,
the curvature of a spatial curve is given by
k = lim
∆s→0
∆ϕ
∆s
= lim
∆s→0
τ(s + ∆s) −τ(s)
∆s
= lim
∆s→0
r
(s + ∆s) −r
(s)
∆s
= r
(s) . (5.4)
We intentionally introduced the deﬁnition of curvature in such a way that
it remains nonnegative. For a plane curve the deﬁnition is normally given
130 Tensor Analysis with Applications in Mechanics
so that k can be positive or negative depending on the sense of convexity
(“concave up” or “concave down”).
Exercise 5.9. A helix is described by
r(t) = i
1
αcos t +i
2
αsin t +i
3
βt.
Study the curvature as a function of the parameters α and β, showing that
k → 0 as α → 0 and k → 1/α as β → 0. Explain.
Moving trihedron
Let us note that τ
2
= τ · τ = 1 for all s. This means that
d
ds
τ
2
(s) = 2τ(s) · τ
(s) = 0, (5.5)
hence τ
(s) is orthogonal to τ(s).
Remark 5.1. It is clear that τ(s) in (5.5) need not be a tangent vector.
Thus we have a general statement: any unit vector e(s) is orthogonal to its
derivative e
(s). That is, e(s) · e
(s) = 0.
We deﬁne the principal normal ν at point s by the equation
ν =
r
(s)
k
. (5.6)
If the curve r = r(s) lies in a plane, it is clear that ν lies in the same plane.
Any plane through the point s that contains a tangent to the curve at this
point is called a tangent plane. Among all the tangent planes at the same
point of a nonplanar curve, there is a unique one that plays the role of
the plane containing a plane curve. It is the plane that contains τ and ν
simultaneously. This plane is said to be osculating, and can be thought of
as “the most locally tangent” plane to the curve at the point s. A unit
normal to the osculating plane at the same point s is introduced using the
relation
β = τ ×ν.
These vectors τ, ν, and β constitute what is called the moving trihedron
of the curve. Associated with this frame at each point along the curve is a
set of three mutually perpendicular planes. As we stated above, the plane
of τ and ν is called the osculating plane. The plane of ν and β is called
the normal plane, and the plane of β and τ is called the rectifying plane.
Elements of Diﬀerential Geometry 131
The equation of the osculating plane at point s
0
is
(r −r(s
0
)) · β(s
0
) = 0
where r is the radius vector of a point of the osculating plane. For general
parametrization of the curve r = r(t) the equation of the osculating plane
in Cartesian coordinates is
x −x(t
0
) y −y(t
0
) z −z(t
0
)
x
(t
0
) y
(t
0
) z
(t
0
)
x
(t
0
) y
(t
0
) z
(t
0
)
= 0.
Exercise 5.10. Calculate ν and β for the helix of Exercise 5.9. Then ﬁnd
the equation of the rectifying plane at t = t
0
.
Using the chain rule we can present the expression for k for an arbitrary
parametrization of a curve:
k
2
=
(r
(t) ×r
(t))
2
(r
2
(t))
3
. (5.7)
We shall now denote the curvature by k
1
, because there is another quantity
that characterizes how a curve diﬀers from a straight line. In terms of k
1
we may deﬁne the radius of curvature as the number R = 1/k
1
.
Exercise 5.11. Derive expression (5.7).
The principal normal and the binormal to a space curve are not deﬁned
uniquely when τ
(s) = 0. Any point at which this condition holds is called
a point of inﬂection of the curve. We shall assume that our curves satisfy
τ
(s) = 0 for all s.
Curves in the plane
The equations of this section take special forms when the curve under con
sideration lies in a plane. In such a case it is expedient to work in a concrete
coordinate system such as rectangular coordinates or plane polar coordi
nates. In a rectangular coordinate frame (ˆ x, ˆ y) where
r = ˆ xx(t) + ˆ yy(t),
it is easily seen that the length of the part of the curve corresponding to
the interval [a, b] of the parameter t is
s =
b
a
[x
(t)]
2
+ [y
(t)]
2
dt.
132 Tensor Analysis with Applications in Mechanics
The curvature is given by
k
1
=
x
(t)y
(t) −x
(t)y
(t)
{[x
(t)]
2
+ [y
(t)]
2
}
3/2
.
These formulas correspond to the familiar formulas
s =
b
a
1 + [f
(x)]
2
dx, k
1
=
f
(x)
{1 + [f
(x)]
2
}
3/2
from elementary calculus, which are written for a curve expressed in the
nonparametric form y = f(x). In polar coordinates where the curve is
expressed in the form r = r(θ), we have
s =
b
a
[r(θ)]
2
+ [r
(θ)]
2
dθ, k
1
=
[r(θ)]
2
+ 2[r
(θ)]
2
−r(θ)r
(θ)
{[r(θ)]
2
+ [r
(θ)]
2
}
3/2
.
Note that for a plane curve the curvature k
1
possesses an algebraic sign.
Exercise 5.12. (a) Find the radius of curvature of the curve y = x
3
at the
point (1, 1). Repeat for the curve y = x
4
at the point (0, 0). (b) Locate the
point of maximum curvature of the parabola y = ax
2
+bx +c.
Exercise 5.13. Suppose that all the tangents to a smooth curve pass
through the same point. Demonstrate that the curve is a part of a straight
line or the whole line.
Exercise 5.14. Suppose that all the tangents to a smooth curve are parallel
to a plane. Show that the curve lies in a plane.
Exercise 5.15. Suppose that all the principal normals of a smooth curve
are parallel to a plane. Does this curve lie in a plane? Repeat when all the
binormals are parallel to a plane.
5.2 The Torsion of a Curve
When a curve lies in a plane, the binormal β is normal to this plane.
Moreover, β is normal to the osculating plane to the curve at a point, so
the rate of rotation of this plane, which is measured by the rate of turn of
the binormal, characterizes how the curve is “nonplanar.” By analogy to
the curvature of a curve we introduce this characteristic of “nonplaneness”
called the torsion or second curvature, and deﬁne it as the limit of the ratio
∆ϑ/∆s where ∆ϑ is the angle of turn of the binormal β. Since β is a unit
Elements of Diﬀerential Geometry 133
vector we can use the same reasoning as we used to derive the expression
for the curvature (5.4). Let us denote the torsion by k
2
and write
k
2
= lim
∆s→0
∆ϑ
∆s
.
This quantity has an algebraic sign as we explain further below.
2
Theorem 5.1. Let r = r(s) be a three times continuously diﬀerentiable
vector function of s. At any point where k
1
= 0, the torsion of the curve is
k
2
= −
(r
(s) ×r
(s)) · r
(s)
k
2
1
.
Proof. At a point s where k
1
= 0 the binormal is deﬁned uniquely as is
ν. Since β is a unit vector we can deﬁne the absolute value of the turn
of the binormal, when it moves along the element corresponding to ∆s, in
the same manner as we used to introduce the curvature of a curve. Now
β(s + ∆s) −β(s) = 2 sin ∆ϑ/2 . Thus we have
k
2
 = lim
∆s→0
∆ϑ
∆s
= lim
∆s→0
∆ϑ
2 sin(∆ϑ/2)
2 sin(∆ϑ/2)
∆s
= lim
∆s→0
β(s + ∆s) −β(s)
∆s
=
β
(s)
.
Let us demonstrate that β
(s) and ν are parallel. The derivative of any
unit vector x(s) is normal to the vector:
0 =
d
ds
(x(s) · x(s)) = 2x(s) · x
(s).
So β
(s) is orthogonal to β(s) and hence parallel to the osculating plane.
Next
β
(s) = (τ(s) ×ν(s))
= τ
(s) ×ν(s) +τ(s) ×ν
(s) = τ(s) ×ν
(s) (5.8)
where we used the fact that τ
(s) is parallel to ν(s). By (5.8) it follows
that β
(s) is orthogonal to τ(s), so we have established the needed property.
Since ν(s) = 1 it follows that
β
(s) · ν(s)
=
β
(s)
and thus
k
2
 =
β
(s) · ν(s)
.
2
Basically, a curve having positive torsion will twist in the manner of a righthand screw
thread as s increases.
134 Tensor Analysis with Applications in Mechanics
Let us use the fact that ν = r
/k
1
, and thus
ν
= r
/k
1
−r
k
1
/k
2
1
, β = (r
×r
)/k
1
.
We get
k
2
 = (τ ×ν
) · ν =
(r
×r
) · r

k
1
2
.
Here we used the properties of the mixed product. We now deﬁne
k
2
= −
(r
×r
) · r
k
1
2
.
The rule for the sign is introduced as follows: if the binormal β turns in the
direction from β to ν, then the sign is positive; otherwise, it is negative.
The sign of k
2
is taken in such a way that
β
= k
2
ν. (5.9)
This completes the proof.
Let us also mention that if we consider another parametrization of the
curve, then a simple calculation using the chain rule yields
k
2
= −
(r
(t) ×r
(t)) · r
(t)
(r
(t) ×r
(t))
2
. (5.10)
We have said that the value of the torsion indicates the rate at which
the curve distinguishes itself from a plane curve. Let us demonstrate this
more clearly. Let k
2
= 0 for all s. We show that the curve lies in a plane.
Indeed 0 = k
2
 =
β
(s) · ν(s)
, thus
β
(s)
= 0 (since ν is a unit vector)
and so β(s) = β
0
= const. The tangent vector τ is orthogonal to β
0
, so
0 = τ · β
0
= r
(s) · β
0
and thus, integrating, we get (r(s) −r(s
0
)) · β
0
= 0.
This means that the curve lies in a plane.
By formula (5.9), k
2
= dβ/ds. So k
2
= lim
∆s→0
∆β/∆s. But up to
small quantities of the second order of ∆s, the change ∆e of a unit vector
e is equal to the angle of rotation of e when moved through a distance ∆s.
Thus ∆β is approximately equal to the angle of rotation of the binormal
during the shift of the point through ∆s, and so k
2
measures the rate of
rotation of the binormal when a point moves along the curve.
Exercise 5.16. Demonstrate that a smooth curve lies in a plane only if
k
2
= 0.
Elements of Diﬀerential Geometry 135
Exercise 5.17. Demonstrate that in Cartesian coordinates
k
1
=
(y
z
−z
y
)
2
+ (z
x
−x
z
)
2
+ (x
y
−y
x
)
2
1/2
[(x
)
2
+ (y
)
2
+ (z
)
2
]
3/2
and
k
2
=
x
y
z
x
y
z
x
y
z
(y
z
−z
y
)
2
+ (z
x
−x
z
)
2
+ (x
y
−y
x
)
2
where the prime denotes d/dt.
Exercise 5.18. What happens to k
1
and k
2
(and the moving trihedron of
a curve) if the direction of change of the parameter is reversed (t → (−t))?
Exercise 5.19. Calculate k
2
for the helix of Exercise 5.9.
5.3 Frenet–Serret Equations
The natural triad τ, ν, β can serve as coordinate axes of the space; this is
used when studying local properties of a curve. Frenet established a system
of ordinary diﬀerential equations which governs the triad along the curve.
We have already derived two of these three equations: (5.6) and (5.9). The
former will be written as τ
= k
1
ν. Let us derive the third formula of the
Frenet–Serret system. We have ν = β ×τ. By this,
ν
= (β ×τ)
= β
×τ +β ×τ
= k
2
ν ×τ +β ×(k
1
ν)
= −k
1
τ −k
2
β.
Let us collect the Frenet–Serret equations together:
τ
= k
1
ν,
ν
= −k
1
τ −k
2
β,
β
= k
2
ν. (5.11)
We recall that these equations are written out when the curve has the
natural parametrization with the length parameter s.
Note that if the curvatures k
1
(s) and k
2
(s) are given functions of s, the
system (5.11) becomes a linear system of ordinary diﬀerential equations; in
136 Tensor Analysis with Applications in Mechanics
component form, it becomes a system of nine equations in nine unknowns.
Fixing some point of the curve in space and an orthonormal triad τ, ν, β
at the point, by this system we can deﬁne τ(s), ν(s), and β(s) uniquely;
then, by the equation r
(s) = τ(s), we deﬁne the curve r = r(s) uniquely
as well. Thus k
1
(s) and k
2
(s) deﬁne the curve up to a motion in space.
That is why the pair of equations for k
1
(s), k
2
(s) is called the set of natural
equations of the curve.
Let us demonstrate how to use the Frenet–Serret equations to charac
terize the curve locally. We use the Taylor expansion of the radius vector
of a curve at point s:
r(s + ∆s) = r(s) + ∆sr
(s) +
(∆s)
2
2
r
(s) +
(∆s)
3
6
r
(s) +o(∆s
3
).
By the deﬁnition ν = r
/k
1
and by the second of equations (5.11) we get
r
= (k
1
ν)
= k
1
ν +k
1
ν
= k
1
ν +k
1
(−k
1
τ −k
2
β).
Substituting these we get
r(s+∆s) = r(s)+∆sτ +
(∆s)
2
2
k
1
ν+
(∆s)
3
6
(k
1
ν−k
2
1
τ −k
1
k
2
β)+o(∆s
3
).
This shows that, to the order of (∆s)
2
, the curve lies in the osculating
plane at point s. At a point s the triad τ, ν, β is Cartesian. Let us ﬁx this
frame and place its origin at r(s) = 0. Deﬁning x, y, z as the components
of r(s+∆s) in this frame we obtain the approximate representation for the
curve
x = ∆s −
k
2
1
(∆s)
3
6
+o(∆s
3
),
y =
k
1
(∆s)
2
2
+
k
1
(∆s)
3
6
+o(∆s
3
),
z = −
k
1
k
2
(∆s)
3
6
+o(∆s
3
).
As another application of the Frenet–Serret equations, we ﬁnd the ve
locity and acceleration of a point moving in space. Let s be the length
parameter of the trajectory of the point, and let t denote time. We shall
use the triad (τ, ν, β) of the trajectory r = r(s). The position of the point
is given by the equation
r = r(s(t)).
The velocity of the point is
v =
d
dt
r(s(t)) =
dr
ds
ds
dt
=
ds
dt
τ.
Elements of Diﬀerential Geometry 137
Denoting v = ds/dt, the particle speed at each point along its path, we can
write v = vτ. Now let us ﬁnd the acceleration of the point in the same
frame, which is
a =
d
2
dt
2
r(s(t)) =
d
dt
v(s(t)) =
d
dt
ds
dt
τ
=
d
2
s
dt
2
τ +
ds
dt
dτ
ds
ds
dt
= s
(t)τ +v
2
dτ
ds
.
Using the Frenet–Serret equations we get
a = s
(t)τ +k
1
v
2
ν
where k
1
is the principal curvature of the trajectory. In mechanics this is
commonly written as
a = s
(t)τ + (v
2
/ρ)ν
where ρ = 1/k
1
is the radius of curvature of the curve. Thus the acceleration
vector lies in the osculating plane at each point along the trajectory. On
this fact several practical methods of ﬁnding the acceleration are based.
Exercise 5.20. Express d
3
r/ds
3
in terms of the moving trihedron.
Exercise 5.21. By deﬁning a vector δ = k
1
β−k
2
τ, show that the Frenet–
Serret equations can be written in the form
τ
= δ ×τ, ν
= δ ×ν, β
= δ ×β.
The vector δ is known as the Darboux vector.
Exercise 5.22. A particle moves through space in such a way that its
position vector is given by
r(t) = i
1
(1 −cos t) +i
2
t +i
3
sin t.
Find the tangential and normal components of the acceleration.
5.4 Elements of the Theory of Surfaces
The reader is familiar with many standard surfaces: the sphere, the cone,
the cylinder, etc. These are easy to visualize, and in Cartesian coordinates
are described by simple equations. For example, the equation of a sphere
reﬂects the deﬁnition of a sphere: all points (x, y, z) of a sphere have the
same distance R from the center (x
0
, y
0
, z
0
):
(x − x
0
)
2
+ (y −y
0
)
2
+ (z −z
0
)
2
= R
2
. (5.12)
138 Tensor Analysis with Applications in Mechanics
An inﬁnite circular cylinder with generator parallel to the zaxis is given
by the equation
(x −x
0
)
2
+ (y −y
0
)
2
= R
2
,
in which the variable z does not appear. We can obtain other types of
cylindrical surfaces by considering a set of spatial points satisfying the
equation
f(x, y) = 0.
This equation does not depend on z, hence by drawing a generator parallel
to the zaxis through the point (x, y, 0) we get a more general cylindrical
surface. A paraboloid is an example of a more complex surface:
z = ax
2
+ 2bxy +cy
2
+dx +ey +f, (5.13)
where a, b, c, d, e, f are numerical coeﬃcients. Depending on the values of
a, b, c, the paraboloid can be elliptic, hyperbolic or parabolic. These terms
from analytic geometry will be used to characterize the shape of a surface
at a point. The areas and volumes associated with such standard surfaces
(or portions thereof) can be calculated by integration or, sometimes, by the
use of elementary formulas.
From a naive viewpoint a surface is something that fully or partially
bounds a spatial body. However, a precise deﬁnition of the term “surface”
is not easy to give. An attempt could be based on a local description, re
garding an elementary portion of a surface as a continuous image in space
of a small disk in the plane. Unfortunately such a deﬁnition would place
under consideration surfaces of extremely complex structure. To use the or
dinary tools of calculus we must invoke the idea of smoothness of a surface,
even if we do not deﬁne this term precisely (it is common in the natural
sciences to use notions and study objects that are not explicitly introduced,
and about which we know only some things).
As a ﬁrst step we could view a surface as something in space that re
sembles the ﬁgures mentioned above in the sense that it has zero thickness.
To use calculus we must represent the coordinates of the points of a surface
using functions. Note that the paraboloid (5.13) can be described using a
single function of the two variables x and y, whereas this is not possible
with the sphere (5.12). But we can divide a sphere into hemispheres in such
a way that each can be described by a separate function of x and y. This
can be done with more general surfaces, although it may not be possible
Elements of Diﬀerential Geometry 139
to have z as the dependent variable for all portions of a surface. In gen
eral, the position of any point of a surface is determined by the values of a
pair (u, v) of independent parameters which may or may not be Cartesian
coordinates. It is convenient to use the position vector
r = r(u, v)
to locate this point with respect to the coordinate origin. For a Cartesian
frame this vector equation is equivalent to the three scalar equations
x = x(u, v), y = y(u, v), z = z(u, v),
but the use of other coordinate frames is common. The pair (u, v) can be
regarded as the coordinates of a point in the uvplane, and a portion of
the surface may be regarded as the image of a domain in this plane; in
this sense we can say that a surface is twodimensional. As was the case
in earlier sections, we shall consider only suﬃciently smooth functions. We
have seen that the use of tensor symbolism simpliﬁes many formulas and
renders them clear. So instead of the notation (u, v) we shall use (u
1
, u
2
)
to denote the intrinsic coordinates of a surface.
3
One goal of this book is to collect major results and explain how these
can be used by practitioners of many sciences. In particular, we now present
formulas used in the mechanics of elastic shells, which relies heavily on the
theory of surfaces. First we show how to calculate distances and angles on a
general surface. This is done by introducing the “ﬁrst fundamental form” of
the surface. In this way we introduce a metric form of the surface analogous
to the one used for threedimensional space. Other surface properties can be
described once we possess the notion of the surface normal. The structure
of a surface at a point may cause us to experience diﬀerences as we depart
from the point in various directions. To study this we introduce the “second
fundamental form” of the surface. The two fundamental forms provide a
full local description of the surface. We shall study other notions as well,
concluding our treatment by adapting tensor analysis to two dimensions
and applying the resulting tools to the study of surfaces.
First fundamental form
We have said that a coordinate line in space is a set of points generated
by ﬁxing two of the coordinates q
1
, q
2
, q
3
and changing the remaining one.
3
We use this term to indicate that the coordinates refer to the surface; they are taken
in the surface, not from the surrounding space somehow.
140 Tensor Analysis with Applications in Mechanics
A coordinate level surface is generated by ﬁxing one of the coordinates
q
1
, q
2
, q
3
and changing the remaining two. A similar idea is used to intro
duce an arbitrary surface in space. The position of a point on a surface
can be characterized by a radius vector initiated from some origin of the
space and with terminus that moves along the surface. A point of a surface
can be characterized by two parameters that we shall denote by u
1
, u
2
or
sometimes, when convenient, by u, v:
r = r(u
1
, u
2
). (5.14)
Thus we consider a surface as the spatial image of some domain in the
(u
1
, u
2
) coordinate plane. For simplicity we shall restrict ourselves to those
surfaces that are smooth everywhere except possibly for some simple curves
or points which lie thereon. So we shall consider the case when (5.14) is
as smooth as we like (it is normally necessary for (5.14) to have all of its
second (and sometimes third) partial derivatives continuous except for some
lines or poles, as is the case with the vertex of a cone). Similar to the case
of a threedimensional space, on a surface there are coordinate lines that
arise when in (5.14) we ﬁx one of coordinates. In similar fashion we can
introduce the tangent vectors
r
i
=
∂r(u
1
, u
2
)
∂u
i
(i = 1, 2).
We suppose that, except at some singular lines or poles in the (u
1
, u
2
)
plane, these vectors are continuously diﬀerentiable in the coordinates. We
also suppose that, except at the same points, these two vectors are not
collinear so the vector
N = r
1
×r
2
= 0.
By this we can introduce the unit normal to the surface
n =
r
1
×r
2
r
1
×r
2

,
which is simultaneously normal to the plane that is osculating to the surface
at point (u
1
, u
2
). Let us demonstrate this. Let the plane through point
(u
1
, u
2
) have normal n. The osculating plane (Fig. 5.2) is the plane for
which
lim
∆s→0
h
∆s
= 0
where
∆s =
r(u
1
+ ∆u
1
, u
2
+ ∆u
2
) −r(u
1
, u
2
)
Elements of Diﬀerential Geometry 141
and h is the distance from the point with coordinates (u
1
+∆u
1
, u
2
+∆u
2
)
to the plane. Note that this must hold for any curve on the surface through
the point. Let the plane be orthogonal to n. Then
h =
r(u
1
+ ∆u
1
, u
2
+ ∆u
2
) −r(u
1
, u
2
)
· n
.
Using the deﬁnition of the diﬀerential we have
h
∆s
=
(r(u
1
+ ∆u
1
, u
2
+ ∆u
2
) −r(u
1
, u
2
)) · n
r(u
1
+ ∆u
1
, u
2
+ ∆u
2
) − r(u
1
, u
2
)
=
(r
i
∆u
i
+o(
∆u
1
+
∆u
2
)) · n
r
i
∆u
i
+o(∆u
1
 +∆u
2
)
=
o(
∆u
1
+
∆u
2
)
r
i
∆u
i
+o(∆u
1
 +∆u
2
)
.
This means that the plane we chose is osculating.
Fig. 5.2 Osculating plane to a surface.
It is clear that much of what was said about r
i
in Chapter 4 should be
simply reformulated for the present case if we use the vectors r
1
, r
2
, r
3
= n
as a basis in R
3
. First, we can introduce the reciprocal basis in space whose
third vector is n again and whose two others are deﬁned by the equations
r
i
· r
j
= δ
i
j
(i = 1, 2), r
i
· n = 0.
The vectors r
i
(i = 1, 2) constitute the reciprocal frame to r
i
(i = 1, 2) in
the plane osculating to the surface at (u
1
, u
2
).
142 Tensor Analysis with Applications in Mechanics
Let us consider the diﬀerential
dr = r
i
du
i
. (5.15)
In this chapter, when summing, the indices take the values 1 and 2. The
diﬀerential is the main linear part in du
i
of the diﬀerence
∆r = r(u
1
+du
1
, u
2
+ du
2
) −r(u
1
, u
2
).
The main part of the distance between nearby points r(u
1
+du
1
, u
2
+du
2
)
and r(u
1
, u
2
) is given by
(ds)
2
= dr · dr = r
i
du
i
· r
j
du
j
= g
ij
du
i
du
j
where g
ij
= r
i
· r
j
is the metric tensor of the surface. The form
(ds)
2
= g
ij
du
i
du
j
is called the ﬁrst fundamental form of the surface. Beginning with Gauss
the metric coeﬃcients were denoted by
E = g
11
, F = g
12
= g
21
, G = g
22
,
so the ﬁrst fundamental form is
(ds)
2
= E(du
1
)
2
+ 2F du
1
du
2
+G(du
2
)
2
.
Exercise 5.23. Find the ﬁrst fundamental form of a sphere of radius a.
Besides the length of an elementary curve on the surface, formula (5.15)
allows us to ﬁnd an angle between two elementary curves dr = r
i
du
i
and
d˜r = r
i
d˜ u
i
. Indeed, by the deﬁnition of the dot product we have
cos ϕ =
dr · d˜r
dr d˜r
=
E du
1
d˜ u
1
+F(du
1
d˜ u
2
+du
2
d˜ u
1
) +Gdu
2
d˜ u
2
E(du
1
)
2
+ 2F du
1
du
2
+G(du
2
)
2
E(d˜ u
1
)
2
+ 2F d˜ u
1
d˜ u
2
+G(d˜ u
2
)
2
.
(5.16)
Exercise 5.24. (a) Find the expression for the angle between coordinate
lines of a surface. (b) Calculate the angle at which the curve θ = φ crosses
the equator of a sphere.
The area of the parallelogram that is based on the elementary vectors
r
1
du
1
and r
2
du
2
which are tangent to the coordinate lines is
dS =
r
1
du
1
×r
2
du
2
.
Elements of Diﬀerential Geometry 143
This approximates up to higher order terms the area of the corresponding
curvilinear ﬁgure bordered by the coordinate lines
u
1
= const, u
2
= const, u
1
+du
1
= const, u
2
+du
2
= const.
Let us ﬁnd dS in terms of the ﬁrst fundamental form. We need to calculate
r
1
×r
2
. Let us demonstrate that
r
1
×r
2
 =
EG−F
2
. (5.17)
Indeed, by deﬁnition of the dot and cross product we have
r
1
×r
2

2
+ (r
1
· r
2
)
2
= r
1

2
r
2

2
sin
2
ϕ +r
1

2
r
2

2
cos
2
ϕ = r
1

2
r
2

2
.
Thus
r
1
×r
2

2
= r
1

2
r
2

2
−(r
1
· r
2
)
2
= EG−F
2
and (5.17) follows.
Summing the elementary areas corresponding to a domain A in the
(u
1
, u
2
) plane and doing the limit passage we get the value
S =
A
EG−F
2
du
1
du
2
.
It can be shown that for a smooth surface this limit S does not depend on
the parametrization and hence is the area of the portion A of the surface.
Exercise 5.25. A cone is described by the position vector
r(u
1
, u
2
) = i
1
u
1
cos u
2
+i
2
u
1
sinu
2
+i
3
u
1
,
where 0 ≤ u
1
≤ a and 0 ≤ u
2
< 2π. Find the surface area of the cone.
Exercise 5.26. Find the above formulas for a ﬁgure of revolution described
by the Cartesian coordinate position vector
r = ˆ xρ cos φ + ˆ yρ sinφ + ˆ zf(ρ)
where f(ρ) is a suitable proﬁle function.
Exercise 5.27. Let two smooth surfaces be parametrized in such a way
that the coeﬃcients of their ﬁrst fundamental forms are proportional at any
point with the same coordinates (so this deﬁnes a map of one surface onto
the other). Show that the map preserves the angles between corresponding
directions on the surfaces.
144 Tensor Analysis with Applications in Mechanics
Geodesics
A fundamental problem in surface theory is to ﬁnd the curve of minimum
length between two given points on a surface S. To treat such a problem
it is natural to begin with the ﬁrst fundamental form
(ds)
2
= g
ij
du
i
du
j
.
We know that parametric equations
u
1
= u
1
(t), u
2
= u
2
(t),
can specify a curve C on S. Writing (ds)
2
as
(ds)
2
= g
ij
du
i
dt
du
j
dt
(dt)
2
,
the expression for arc length along C from t = a to t = b becomes
s =
b
a
g
ij
du
i
dt
du
j
dt
1/2
dt. (5.18)
We seek a curve of minimum length between a given pair of endpoints: i.e.,
we seek functions u
1
(t) and u
2
(t) that minimize s.
Since
g
ij
= g
ij
(u
1
(t), u
2
(t)),
we see that (5.18) has the form
s(u) =
b
a
f(t, u, ˙ u) dt (5.19)
where u = u(t) = (u
1
(t), u
2
(t)) and the overdot denotes diﬀerentiation with
respect to t. Because s(u) is a correspondence that assigns a real number
s to each function u in some class of functions, we call it a functional in
u. Here the class of functions includes all admissible routes u between the
given endpoints on the surface.
The minimization of functionals is treated in the calculus of variations.
Fortunately the main ideas of this extensive subject lend themselves to a
brief discussion aimed at the present task. (Additional coverage appears in
Chapter 6, where we consider the variational principles of elasticity.)
The approach to ﬁnding a minimizer u(t) for the functional (5.19)
hinges on replacing u(t) by a new function u(t) + εϕ(t), where ϕ(t) is
an admissible variation of the function u(t) and ε is a small real parame
ter. In this discussion we shall consider a variation ϕ(t) to be admissible
Elements of Diﬀerential Geometry 145
if ϕ(a) = ϕ(b) = 0; that is, if each curve of the form u(t) +εϕ(t) connects
the same endpoints as the curve u(t). Such a replacement gives
s(u +εϕ) =
b
a
f(t, u +εϕ, ˙ u +ε ˙ ϕ) dt. (5.20)
For ﬁxed u and ϕ this is a function of the real variable ε, and takes its
minimum at ε = 0 for any ϕ. Let us vary the components of u(t) one
at a time. If we take ϕ of the special form ϕ
1
(t) = (ϕ(t), 0), then (5.20)
becomes
s(u +εϕ
1
) =
b
a
f(t, u
1
(t) +εϕ(t), u
2
(t), ˙ u
1
(t) +ε ˙ ϕ(t), ˙ u
2
(t)) dt.
We now set
ds(u +εϕ
1
)
dε
ε=0
= 0 (5.21)
and handle the lefthand side using the chain rule:
d
dε
b
a
f(t, u
1
+εϕ, u
2
, ˙ u
1
+ε ˙ ϕ, ˙ u
2
) dt
ε=0
=
=
b
a
¸
∂
∂u
1
f(t, u
1
, u
2
, ˙ u
1
, ˙ u
2
)ϕ +
∂
∂ ˙ u
1
f(t, u
1
, u
2
, ˙ u
1
, ˙ u
2
) ˙ ϕ
dt.
This is the ﬁrst variation of the functional s(u) with respect to variations of
the form u+εϕ
1
; we equate it to zero as in (5.21), and employ integration
by parts in the second term of the integrand to get
b
a
¸
∂
∂u
1
f(t, u
1
, u
2
, ˙ u
1
, ˙ u
2
) −
d
dt
∂
∂ ˙ u
1
f(t, u
1
, u
2
, ˙ u
1
, ˙ u
2
)
ϕdt = 0.
According to the fundamental lemma of the calculus of variations (cf.,
[
Lebedev and Cloud (2003)
]
; one version of this lemma is presented in
Theorem 6.8), this equation can hold for all admissible variations ϕ only if
the bracketed quantity vanishes:
∂
∂u
1
f(t, u
1
, u
2
, ˙ u
1
, ˙ u
2
) −
d
dt
∂
∂ ˙ u
1
f(t, u
1
, u
2
, ˙ u
1
, ˙ u
2
) = 0.
Repeating this process for variations of the special form ϕ
2
(t) = (0, ϕ(t))
we obtain a similar result
∂
∂u
2
f(t, u
1
, u
2
, ˙ u
1
, ˙ u
2
) −
d
dt
∂
∂ ˙ u
2
f(t, u
1
, u
2
, ˙ u
1
, ˙ u
2
) = 0.
146 Tensor Analysis with Applications in Mechanics
This system of two equations, which we can write in more compact notation
as
∂f
∂u
i
−
d
dt
∂f
∂ ˙ u
i
= 0 (i = 1, 2), (5.22)
is known as the system of Euler equations for the functional (5.19). Satis
faction of this system is a necessary condition for u = (u
1
(t), u
2
(t)) to be a
minimizer of s. In the calculus of variations a solution to an Euler equation
(or, in this case, a system of such equations) is called an extremal. This is
analogous to a stationary point of an ordinary function: further testing is
required to ascertain whether the extremal actually yields a minimum of
the functional (it could yield a maximum, say). We refer to solutions of the
system (5.22) as geodesics on the surface S.
We now use the fact that
f
2
= g
jk
˙ u
j
˙ u
k
(5.23)
to compute the derivatives in (5.22). First we have
2f
∂f
∂u
i
=
∂g
jk
∂u
i
˙ u
j
˙ u
k
so that
∂f
∂u
i
=
1
2f
∂g
jk
∂u
i
˙ u
j
˙ u
k
. (5.24)
Expanding the notation in (5.23) we see that
f
2
= g
11
˙ u
1
˙ u
1
+g
12
˙ u
1
˙ u
2
+g
21
˙ u
2
˙ u
1
+g
22
˙ u
2
˙ u
2
,
hence
2f
∂f
∂ ˙ u
1
= 2g
11
˙ u
1
+g
12
˙ u
2
+g
21
˙ u
2
= 2(g
11
˙ u
1
+ g
12
˙ u
2
) = 2g
1j
˙ u
j
and ∂/∂ ˙ u
2
is taken similarly. So
f
∂f
∂ ˙ u
i
= g
ij
˙ u
j
,
which gives
d
dt
∂f
∂ ˙ u
i
=
d
dt
g
ij
˙ u
j
f
=
1
f
d
dt
(g
ij
˙ u
j
) +g
ij
˙ u
j
d
dt
1
f
=
1
f
g
ij
¨ u
j
+ ˙ u
j
d
dt
g
ij
+g
ij
˙ u
j
−
1
f
2
df
dt
.
Elements of Diﬀerential Geometry 147
Here
d
dt
g
ij
=
∂g
ij
∂u
k
˙ u
k
by the chain rule, so
d
dt
∂f
∂ ˙ u
i
=
1
f
g
ij
¨ u
j
+
∂g
ij
∂u
k
˙ u
j
˙ u
k
−
g
ij
˙ u
j
f
2
df
dt
.
It simpliﬁes things if we take t to be the arc length parameter; then f ≡ 1
and df/dt ≡ 0 so that
d
dt
∂f
∂ ˙ u
i
= g
ij
¨ u
j
+
∂g
ij
∂u
k
˙ u
j
˙ u
k
. (5.25)
Putting (5.24) and (5.25) into (5.22) we have
g
ij
¨ u
j
+
¸
∂g
ij
∂u
k
−
1
2
∂g
jk
∂u
i
˙ u
j
˙ u
k
= 0. (5.26)
The second term on the left can be rewritten:
¸
∂g
ij
∂u
k
−
1
2
∂g
jk
∂u
i
=
¸
1
2
∂g
ij
∂u
k
+
∂g
ik
∂u
j
−
1
2
∂g
jk
∂u
i
˙ u
j
˙ u
k
=
1
2
¸
∂g
ij
∂u
k
+
∂g
ik
∂u
j
−
∂g
jk
∂u
i
˙ u
j
˙ u
k
= Γ
jki
˙ u
j
˙ u
k
.
Putting this into (5.26) and raising the index i, we have ﬁnally
¨ u
n
+ Γ
n
jk
˙ u
j
˙ u
k
= 0 (n = 1, 2). (5.27)
This is a system of two nonlinear diﬀerential equations for the unknown
u
n
(t). Again, t is assumed to be the natural length parameter; this means
that the constraint g
ij
˙ u
j
˙ u
k
= 1 must be enforced at each value of t along
the curve.
As a simple example we may treat a cylinder ρ = a. It is clear in
advance that if we cut the cylinder along a generator and “unroll” it onto
a plane, then the shortest route between two points is the direct segment
connecting them. Rolling this plane back into a cylinder we get a curve of
the form
φ(t) = c
1
t +c
2
, z(t) = c
3
t +c
4
, (5.28)
i.e., a helix. Now let us obtain the same result using (5.27). For this
particular surface (u
1
, u
2
) = (φ, z) and
r = ˆ xa cos φ + ˆ ya sinφ + ˆ zz.
148 Tensor Analysis with Applications in Mechanics
We ﬁnd
∂r
∂φ
= −ˆ xa sinφ + ˆ ya cos φ,
∂r
∂z
= ˆz,
giving g
11
= a
2
, g
12
= g
21
= 0, g
22
= 1. Since these are all constants the
Christoﬀel coeﬃcients are all zero, and (5.27) reduces to the system
¨
φ = 0, ¨ z = 0.
Integration replicates the result (5.28).
Exercise 5.28. Find the geodesics of a sphere. Show that they are great
circles of the sphere.
This exercise shows that a geodesic is not always a shortest route, since
we can get either part of a great circle on the sphere connecting two given
points, and both of them are geodesics.
5.5 The Second Fundamental Form of a Surface
Using the Taylor expansion we can ﬁnd the change of r(u
1
, u
2
) to a higher
degree of approximation:
r(u
1
+ ∆u
1
, u
2
+ ∆u
2
) = r(u
1
, u
2
) +
∂r(u
1
, u
2
)
∂u
i
∆u
i
+
+
1
2
∂
2
r(u
1
, u
2
)
∂u
i
∂u
j
∆u
i
∆u
j
+o
(∆u
1
)
2
+ (∆u
2
)
2
.
Here we apply the onotation to vector quantities, indicating that the norm
of the remainder is of a higher order of smallness. We are interested in the
deviation of the surface from the osculating plane in a neighborhood of a
point (u
1
, u
2
). This value is given through the dot product by the normal
vector n:
(r(u
1
+ ∆u
1
, u
2
+ ∆u
2
) −r(u
1
, u
2
)) · n
=
∂r(u
1
, u
2
)
∂u
i
· n∆u
i
+
1
2
∂
2
r(u
1
, u
2
)
∂u
i
∂u
j
· n∆u
i
∆u
j
+o
(∆u
1
)
2
+ (∆u
2
)
2
=
1
2
∂
2
r(u
1
, u
2
)
∂u
i
∂u
j
· n∆u
i
∆u
j
+o
(∆u
1
)
2
+ (∆u
2
)
2
.
Elements of Diﬀerential Geometry 149
We have used the fact that r
i
· n = 0. The terms of the second order of
smallness in ∆u
i
constitute the second fundamental form. Denoting
∂
2
r(u
1
, u
2
)
(∂u
1
)
2
· n = L(u
1
, u
2
) = L,
∂
2
r(u
1
, u
2
)
∂u
1
∂u
2
· n = M(u
1
, u
2
) = M,
∂
2
r(u
1
, u
2
)
(∂u
2
)
2
· n = N(u
1
, u
2
) = N,
we introduce the second fundamental form by
d
2
r · n = L(du
1
)
2
+ 2Mdu
1
du
2
+N(du
2
)
2
. (5.29)
The deviation of the surface from the osculating plane in a small neighbor
hood of (u
1
, u
2
) is given by the following approximation:
z =
1
2
L(du
1
)
2
+ 2Mdu
1
du
2
+N(du
2
)
2
, (5.30)
and so (5.30) deﬁnes the nature of the behavior of the surface in this neigh
borhood. This formula deﬁnes a paraboloid
z =
1
2
L(v
1
)
2
+ 2Mv
1
v
2
+N(v
2
)
2
(5.31)
written in the triad
r
1
(u
1
, u
2
), r
2
(u
1
, u
2
), n(u
1
, u
2
)
when the origin is the
point of the surface with coordinates (u
1
, u
2
). Depending on the coeﬃcients
L, M, N, this paraboloid can be elliptic, hyperbolic or parabolic. If L =
M = N = 0 then the corresponding point is called a planar point.
For the second fundamental form of the surface there is a representation
diﬀerent from (5.29). Diﬀerentiating the identity r
i
· n = 0 we get
∂r
i
∂u
j
· n = −r
i
·
∂n
∂u
j
. (5.32)
Dot multiplying dr · dn we get
r
i
du
i
·
∂n
∂u
j
du
j
= −
∂r
i
∂u
j
du
i
du
j
· n = −d
2
r · n,
which means that
−dr · dn = L(du
1
)
2
+ 2M du
1
du
2
+N(du
2
)
2
.
Exercise 5.29. Let a surface be given in Cartesian components as
r = x(u, v)ˆ x +y(u, v)ˆ y +z(u, v)ˆz.
150 Tensor Analysis with Applications in Mechanics
Let the subscripts u and v denote corresponding partial derivatives with re
spect to u and v. Demonstrate that the coeﬃcients of the ﬁrst fundamental
form are
E = r
2
u
= x
2
u
+y
2
u
+z
2
u
,
F = r
u
· r
v
= x
u
x
v
+y
u
y
v
+z
u
z
v
,
G = r
2
v
= x
2
v
+y
2
v
+z
2
v
.
Then show that the coeﬃcients of the second fundamental form are
L =
x
uu
y
uu
z
uu
x
u
y
u
z
u
x
v
y
v
z
v
√
EG−F
2
, M =
x
uv
y
uv
z
uv
x
u
y
u
z
u
x
v
y
v
z
v
√
EG−F
2
, N =
x
vv
y
vv
z
vv
x
u
y
u
z
u
x
v
y
v
z
v
√
EG−F
2
.
Exercise 5.30. Find L, M, N for a surface of revolution.
Normal curvature of the surface
Let us consider a curve lying on the surface. The curve can be uniquely
deﬁned by a parameter s in such a way that u
1
= u
1
(s), u
2
= u
2
(s); for
simplicity we assume s to be the length parameter. Then the equation of
the curve is
r = r(u
1
(s), u
2
(s)) = r(s)
and a tangential vector to the curve at (u
1
, u
2
) is (r
i
du
i
/ds) ds. Consider
the curvature of this curve, which can be found using r
(s): the normal ν
to the curve has the same direction and k
1
equals r
(s). It follows that
r
(s) · n = k
1
cos ϑ (5.33)
where ϑ is the angle between ν and n. Let us rewrite (5.33) in the form
k
1
cos ϑ =
=
∂
2
r
(∂u
1
)
2
du
1
ds
2
(ds)
2
+ 2
∂
2
r
∂u
1
∂u
2
du
1
ds
du
2
ds
(ds)
2
+
∂
2
r
(∂u
2
)
2
du
2
ds
2
(ds)
2
(ds)
2
· n
=
L(du
1
(s))
2
+ 2M du
1
(s) du
2
(s) +N(du
2
(s))
2
E(du
1
(s))
2
+ 2F du
1
(s) du
2
(s) +G(du
2
(s))
2
. (5.34)
For all the curves with the same direction deﬁned by the constant ratio
du
1
(s) : du
2
(s), the righthand side of (5.34) is the same as the ratio of the
Elements of Diﬀerential Geometry 151
second to the ﬁrst fundamental forms. We deﬁne this value k
0
as the normal
curvature of the surface in the direction du
1
(s) : du
2
(s). Geometrically, this
is the curvature of the curve formed by intersecting the surface with the
plane through (u
1
(s), u
2
(s)) that is parallel to n and r
i
du
i
(s). This normal
curvature satisﬁes the equality
k
0
= k
1
cos ϑ
which is the main part of Meusnier’s theorem. From this formula for cur
vatures it follows that the normal curvature of the surface in the direction
du
1
: du
2
has minimal absolute value among all the curvatures of the curves
on the surface through (u
1
(s), u
2
(s)) having the same direction du
1
: du
2
.
Let us note that the curvature of the osculating paraboloid is character
ized by the same equation. The normal curvature depends on the direction
du
1
: du
2
= x : y only. Let us rewrite its expression as
k
0
=
Lx
2
+ 2Mxy +Ny
2
Ex
2
+ 2Fxy +Gy
2
. (5.35)
The righthand side of (5.35) is a homogeneous form with respect to x, y
of zero order; it is easy to see that the minimum and maximum of k
0
can
be found as the minimum and maximum of the quadratic form
Lx
2
+ 2Mxy +Ny
2
when
Ex
2
+ 2Fxy +Gy
2
= 1. (5.36)
This problem can be solved with use of the Lagrange multiplier. On the
curve (5.36) denote
min(Lx
2
+ 2Mxy +Ny
2
) = k
min
, max(Lx
2
+ 2Mxy + Ny
2
) = k
max
.
Let the corresponding values for x, y be (x
1
, y
1
) and (x
2
, y
2
), respectively.
If k
min
= k
max
these values deﬁne only two directions x
j
: y
j
(j = 1, 2) and
it can be shown that
Ex
1
x
2
+F(x
1
y
2
+x
2
y
1
) +Gy
1
y
2
= 0,
which, by (5.16), means that the directions corresponding to the extreme
values of the normal curvature are mutually orthogonal. In the case when
k
min
= k
max
the normal curvature is the same in all directions (the oscu
lating paraboloid corresponds to a paraboloid of revolution). These values
k
min
, k
max
are the extreme normal curvatures of the surface at a point; they
characterize the surface at a point and hence are invariant with respect to
152 Tensor Analysis with Applications in Mechanics
the change of coordinates of the surface. So are two other characteristics
of the surface at a point: the mean curvature
H =
1
2
(k
min
+k
max
),
and the Gaussian or complete curvature
K = k
min
k
max
.
We present without proof the equations in terms of the fundamental forms
of the surface:
H =
1
2
LG−2MF +NE
EG−F
2
, K =
LN −M
2
EG−F
2
.
Since EG − F
2
> 0, the sign of K is the sign of LN − M
2
. When K = 0
everywhere the surface is developable.
4
These and many other facts on the
properties of surfaces with given H and K can be found in any textbook
on diﬀerential geometry (e.g.,
[
Pogorelov (1957)
]
).
Exercise 5.31. Find the mean and Gaussian curvatures for a sphere of
radius a.
Exercise 5.32. Find the mean and Gaussian curvatures at x = y = 0 of
the following paraboloids: (a) z = axy; (b) z = a(x
2
+y
2
); (c) z = ax
2
+by
2
.
Exercise 5.33. Suppose a surface is given in Cartesian coordinates:
r = ˆ xx + ˆ yy + ˆ zf(x, y).
Demonstrate that (a)
E = r
2
x
= 1 +f
2
x
, F = r
x
· r
y
= f
x
f
y
, G = r
2
y
= 1 +f
2
y
;
(b)
EG−F
2
= 1 +f
2
x
+f
2
y
;
(c)
L =
f
xx
1 +f
2
x
+f
2
y
, M =
f
xy
1 +f
2
x
+f
2
y
, N =
f
yy
1 +f
2
x
+f
2
y
;
4
Roughly speaking, a developable surface is one that can be ﬂattened into a portion
of a plane without compressing or stretching any part of it. Examples include cones
and cylinders, but not spheres. A developable surface can be generated by sweeping a
straight line (generator) along a curve through space. See page 168 for a more precise
deﬁnition.
Elements of Diﬀerential Geometry 153
(d) the area of a portion of the surface is
S =
D
1 +f
2
x
+f
2
y
dxdy;
(e) the Gaussian curvature is
K =
LN −M
2
EG−F
2
=
f
xx
f
yy
−f
2
xy
1 +f
2
x
+f
2
y
2
.
5.6 Derivation Formulas
At each point of a surface there is deﬁned the frame triad
r
1
(u
1
, u
2
), r
2
(u
1
, u
2
), n(u
1
, u
2
).
In the theory of shells they introduce the curvilinear coordinates in a neigh
borhood of the surface in such a way that on the surface the triad (r
1
, r
2
, n)
is preserved. Since we seek diﬀerent characteristics of ﬁelds given on the
surface and outside of it, we must ﬁnd the derivatives of the triad with
respect to the coordinates. The goal of this section is to present these
in terms of the surface we have introduced: i.e., in terms of the ﬁrst and
second fundamental forms of the surface.
We start with the representation for the derivatives through the
Christoﬀel notation. This is valid because, by assumption, (r
1
, r
2
, n) is
a basis of the space of vectors. So
r
ij
=
∂
2
r
∂u
i
∂u
j
= Γ
t
ij
r
t
+λ
ij
n. (5.37)
We recall that we use indices i, j, t taking values from the set {1, 2}, which is
why we introduce the notation λ
ij
for the coeﬃcients of n. Similarly let us
introduce the expansion for the derivatives of n. To derive the coeﬃcients
λ
ij
we dot multiply (5.37) by n. Using the expressions for the second
fundamental form we get
λ
11
= L, λ
12
= λ
21
= M, λ
22
= N.
Next, dot multiplying (5.37) ﬁrst by r
1
and then by r
2
, we get six equations
154 Tensor Analysis with Applications in Mechanics
in the six unknown Christoﬀel symbols:
1
2
∂E
∂u
1
= Γ
1
11
E + Γ
2
11
F,
∂F
∂u
1
−
1
2
∂E
∂u
2
= Γ
1
11
F + Γ
2
11
G,
1
2
∂E
∂u
2
= Γ
1
12
E + Γ
2
12
F,
1
2
∂G
∂u
1
= Γ
1
12
F + Γ
2
12
G,
∂F
∂u
2
−
1
2
∂G
∂u
1
= Γ
1
22
E + Γ
2
22
F,
1
2
∂G
∂u
2
= Γ
1
22
F + Γ
2
22
G.
This system can be easily solved for the Christoﬀel coeﬃcients since the
system splits into pairs of equations with respect to the pairs of Christoﬀel
symbols. The ﬁrst pair yields, for instance,
Γ
1
11
=
1
EG−F
2
¸
G
2
∂E
∂u
1
−F
∂F
∂u
1
+
F
2
∂E
∂u
2
,
and a similar expression for Γ
2
11
. We leave the rest of this work to the reader,
mentioning that the expressions for the Christoﬀel symbols are composed
only of the coeﬃcients of the ﬁrst fundamental form of the surface. Next,
we consider the expressions for the derivatives of the normal n:
n
i
≡
∂n
∂u
i
= µ
·t
i
r
t
+µ
·3
i
n. (5.38)
Let us ﬁnd the coeﬃcients of these expansions through the coeﬃcients of
the fundamental forms of the surface. Because of the equality
0 =
∂
∂u
i
1 =
∂
∂u
i
(n · n) = 2n
i
· n
we have
µ
·3
i
= 0.
Let us dot multiply (5.38) when i = 1 by r
1
and r
2
successively. Using the
deﬁnitions for the coeﬃcients of the ﬁrst and second fundamental forms we
have the equations
−L = µ
· 1
1
E +µ
· 2
1
F, −M = µ
· 1
1
F +µ
· 2
1
G,
from which we get
µ
· 1
1
=
−LG+MF
EG−F
2
, µ
· 2
1
=
LF −ME
EG−F
2
.
Elements of Diﬀerential Geometry 155
Repeating this procedure for (5.38) when i = 2 we similarly obtain
µ
· 1
2
=
NF −MG
EG−F
2
, µ
· 2
2
=
−NE +MF
EG−F
2
.
Exercise 5.34. Derive the Christoﬀel symbols for the case when the ﬁrst
fundamental form is (du
1
)
2
+G(du
2
)
2
.
Exercise 5.35. For a surface of revolution with coordinate lines being
parallels and meridians, derive all coeﬃcients of the expansions (5.37) and
(5.38).
Some useful formulas
Certain identities are frequently employed. Their derivation is lengthy so
we omit it here. Let us redenote the coeﬃcients of the second fundamental
form using index notations:
b
11
= L, b
12
= b
21
= M, b
22
= N.
These are the covariant components of a corresponding symmetric tensor
of order two. The ﬁrst formula is due to Gauss:
b
11
b
22
−b
2
12
=
∂
2
g
12
∂u
1
∂u
2
−
1
2
∂
2
g
11
(∂u
2
)
2
+
∂
2
g
22
(∂u
1
)
2
+
Γ
i
12
Γ
j
12
−Γ
i
11
Γ
j
22
g
ij
,
from which (and the form of K, see page 152) it follows that the Gaus
sian curvature of the surface can be expressed purely in terms of the ﬁrst
fundamental form. The next formulas are due to Peterson and Codazzi:
∂b
i1
∂u
2
−
∂b
i2
∂u
1
= Γ
t
i2
b
t1
−Γ
t
i1
b
t2
(i = 1, 2).
Using the formulas of this paragraph, we can derive the formulas for diﬀer
entiating a spatial vector
f = f
i
r
i
+f
3
n = f
i
r
i
+f
3
n
given on the surface. Denoting r
3
= r
3
= n, we can derive the following
formulas that mimic the formulas of Chapter 4:
∂f
∂u
i
= ∇
i
f
t
r
t
= ∇
i
f
t
r
t
.
156 Tensor Analysis with Applications in Mechanics
In these, the covariant derivatives are
∇
i
f
j
=
∂f
j
∂u
i
+ Γ
j
it
f
t
−b
j
i
f
3
,
∇
i
f
j
=
∂f
j
∂u
i
+ Γ
t
ij
f
t
−b
ij
f
3
,
∇
i
f
3
=
∂f
3
∂u
i
+b
it
f
t
=
∂f
3
∂u
i
+b
t
i
f
t
.
Here i, j take the values 1, 2, while t takes the values 1, 2, 3. Summation is
carried out over i, j, and t in the respective ranges. When f = n, i.e., when
f
1
= f
2
= 0 and f
3
= 1, we get the following formula for the derivatives of
n:
∂n
∂u
i
= −b
t
i
r
t
= −b
it
r
t
. (5.39)
Introducing the surface gradient or surface nabla operator by the for
mula
¯
∇ = r
i
∂
∂u
i
(i = 1, 2),
we ﬁnd that
¯
∇f = ∇
i
f
t
r
i
r
t
= ∇
i
f
t
r
i
r
t
,
¯
∇n = −B,
where B is the surface curvature tensor given by
B = b
j
i
r
i
r
j
= b
ij
r
i
r
j
= b
ij
r
i
r
j
.
One may introduce the surface divergence and rotation operations
div f =
¯
∇· f = r
i
·
∂f
∂u
i
, rot f =
¯
∇×f = r
i
×
∂f
∂u
i
.
The applications of the diﬀerential calculus on a surface are presented in
Chapter 7, where shell theories are considered.
5.7 Implicit Representation of a Curve; Contact of Curves
Our work in Chapter 5 has stressed technical topics needed to derive equa
tions describing certain natural objects. We now turn to some less technical
but quite important questions that can further our understanding of diﬀer
ential geometry.
We have described a curve by the representation r = r(t). From this
point of view, a curve is the image of some onedimensional domain of
Elements of Diﬀerential Geometry 157
the parameter t. Similarly, a surface was an image of a twodimensional
domain, given by a formula
r = r(u
1
, u
2
). (5.40)
However, we know that in Cartesian coordinates a sphere is described by
the equation
(x − x
0
)
2
+ (y −y
0
)
2
+ (z −z
0
)
2
= R
2
,
and this cannot be represented in the form (5.40) for all points simultane
ously. This is an example of the implicit form of description of a surface.
We can describe any small part of a sphere in an explicit form of the type
(5.40). This is a point from which the theory of manifolds arose: a surface
such as a sphere can be divided into small portions, each of which can be
described in the needed way.
However, implicit form descriptions of surfaces and curves are common
in practice. Let us suppose that the equation of a surface is
F(r) = 0.
In Cartesian coordinates this would look like
F(x, y, z) = 0.
Earlier we described coordinate curves in space as sets given by the equation
r = r(q
1
, q
2
, q
3
) when two of the three coordinate parameters q
1
, q
2
, q
3
are
ﬁxed. Each such curve can be described alternatively as the intersection of
two coordinate surfaces. Consider for instance a q
1
coordinate curve — the
curve corresponding to q
2
= q
2
0
, q
3
= q
3
0
, is the intersection of the surfaces
r = r(q
1
, q
2
, q
3
0
) and r = r(q
1
, q
2
0
, q
3
). Similarly a curve in space can be
represented as the set of points described by a vector r that satisﬁes two
simultaneous equations
F
1
(r) = 0, F
2
(r) = 0,
or, in coordinate form,
F
1
(q
1
, q
2
, q
3
) = 0, F
2
(q
1
, q
2
, q
3
) = 0.
A tangent t to the curve at a point must be orthogonal to the two surface
normals at the point; if n
1
and n
2
are these normals, we can write
t = n
1
×n
2
.
158 Tensor Analysis with Applications in Mechanics
Contact of curves
A curve in the plane can be also shown implicitly, but using just one equa
tion:
F(r) = 0 or F(q
1
, q
2
) = 0.
We would like to study the problem of approximating a given plane curve,
at a given point, by another curve taken from a parametric family of curves.
Let us begin with the following problem:
Given a smooth curve A and point C on it, select from all straight lines in
the plane that which best approximates the behavior of A at C.
Of course, one solution is the line tangent to A at C. We see this by
considering a line through C that intersects A at a point D close to C,
and then producing a limit passage under which this line rotates about C
in such a way that D tends to C. The line approached in the limit is the
tangent line.
Let us extend this idea somewhat and seek an approximation that can
reﬂect both the direction of the curve A at C and its curvature at that
point. Since straight lines cannot reﬂect curvature behavior, we shall have
to employ some other family of approximating curves. We could use the
family of all circles in the plane, since any three points in the plane deter
mine a unique circle. Taking two points D
1
and D
2
near C and drawing the
circle through these three points, we could attempt a limit passage under
which both D
1
and D
2
tend to C. This should yield a circle capable of
reﬂecting both desired properties (tangent and curvature) of A at C. The
family of all circles in the plane is described in Cartesian coordinates by
the equation
(x −x
0
)
2
+ (y −y
0
)
2
= R
2
, (5.41)
where the three values x
0
, y
0
and R are free parameters that should be
properly chosen to approximate a given curve at a point. So we see that
a threeparameter family of plane curves will be needed to get the better
approximation we seek. We could employ a parabola from the family y =
ax
2
+bx +c, but this manner of approximation would be less informative.
In general we can seek to approximate a plane curve
r = r(t) (5.42)
at a point t = t
0
by introducing a parametric family Φ
n
of curves
F(r, a
1
, . . . , a
n
) = 0,
Elements of Diﬀerential Geometry 159
where a
1
, . . . , a
n
are free parameters.
5
Suppose we take n points
r(t
k
) (k = 0, . . . , n −1)
of the curve (5.42), where all the t
k
are close to t
0
. With n free parameters
at our disposal we should be able to ﬁnd a curve from Φ
n
that goes through
these points. This means that the system
F(r(t
0
), a
1
, . . . , a
n
) = 0,
F(r(t
1
), a
1
, . . . , a
n
) = 0,
.
.
.
F(r(t
n−1
), a
1
, . . . , a
n
) = 0, (5.43)
is satisﬁed by some set of parameters a
1
, . . . , a
n
. We assume that the a
i
depend on the t
k
in such a way that when all the t
k
tend to t
0
the a
i
tend continuously to some respective values b
i
. Thus we ﬁnd the curve
F(r, b
1
, . . . , b
n
) = 0 that best approximates the behavior of (5.42) at t
0
,
and we have
F(r(t
0
), b
1
, . . . , b
n
) = 0. (5.44)
As a practical matter it is not convenient to use such a limit passage
to obtain the needed curve. It would be better to have conditions in which
only those characteristics of the curves at t
0
are involved. Since we suppose
the above limit passage is welldeﬁned, we can take the ordered set of points
t
k
:
t
0
< t
1
< · · · < t
n−1
.
Let a
1
, . . . , a
n
be a solution to (5.43) in this case, and consider the function
f(t) = F(r(t), a
1
, . . . , a
n
)
of the single variable t. This function vanishes at t = t
k
for k = 0, . . . , n−1.
By Rolle’s theorem there exists t
k
∈ [t
k−1
, t
k
] (k = 1, . . . , n − 1) such that
f
(t
k
) = 0. During the limit passage all the t
k
also tend to t
0
, so we have
d
dt
F(r(t), b
1
, . . . , b
n
)
t=t
0
= 0.
Similarly, the function
f
(t) = F
t
(r(t), a
1
, . . . , a
n
)
5
We suppose that the curve (5.42) is suﬃciently smooth at t
0
, as is each curve taken
from the family Φ
n
.
160 Tensor Analysis with Applications in Mechanics
takes n − 1 zeroes at the points t
1
< · · · < t
n−1
. So by Rolle’s theorem
there is a point t
k
on each segment [t
k−1
, t
k
] such that f
(t
k
) = 0. After
the main limit passage we will have
d
2
dt
2
F(r(t), b
1
, . . . , b
n
)
t=t
0
= 0.
Repeating this procedure for each derivative up to order n − 1, we obtain
the conditions
d
k
dt
k
F(r(t), b
1
, . . . , b
n
)
t=t
0
= 0 (k ≤ n −1).
These, taken together with (5.44), constitute n conditions that should be
fulﬁlled by the needed curve from the family Φ
n
. We deﬁne the order of
contact between this curve and (5.42) as the number n −1.
Contact of a curve with a circle; evolutes
Let us return to our previous problem and apply these conditions to ap
proximate a curve
r = i
1
x(t) +i
2
y(t)
at a point t = t
0
by a circle of the family (5.41). Our three free parameters
x
0
, y
0
, R should satisfy the system
(x(t
0
) −x
0
)
2
+ (y(t
0
) −y
0
)
2
= R
2
,
2(x(t
0
) −x
0
) x
(t
0
) + 2(y(t
0
) −y
0
) y
(t
0
) = 0,
2x
(t
0
)
2
+ 2(x(t
0
) −x
0
) x
(t
0
) + 2y
(t
0
)
2
+ 2(y(t
0
) −y
0
) y
(t
0
) = 0.
In particular, solution of these yields
R
2
=
{[x
(t
0
)]
2
+ [y
(t
0
)]
2
}
3
[x
(t
0
)y
(t
0
) −x
(t
0
)y
(t
0
)]
2
.
We see that the curvature 1/R of the circle coincides with the curvature k
1
of the curve. As the formulas for x
0
and y
0
are cumbersome, it is worthwhile
to recast the problem in vector notation.
Let r
0
locate the center of the contact circle
(r −r
0
)
2
= R
2
.
Let the curve be given in natural parametrization as r = r(s). To solve this
problem of secondorder contact, we deﬁne
F(s) = (r(s) −r
0
)
2
−R
2
Elements of Diﬀerential Geometry 161
and have
F(s
0
) = 0, F
(s
0
) = 0, F
(s
0
) = 0,
or
(r(s
0
) −r
0
)
2
−R
2
= 0,
2(r(s
0
) −r
0
) · τ(s
0
) = 0,
2 + 2(r(s
0
) −r
0
) · ν(s
0
)k
1
= 0.
By the second equation the vector r(s
0
) − r
0
is orthogonal to the tangent
τ(s
0
), hence is directed along the normal ν(s
0
) and we have
(r(s
0
) −r
0
) · ν(s
0
) = −r(s
0
) −r
0
 = −R.
This and the third equation yield 1 − k
1
R = 0, hence R = 1/k
1
as stated
above.
The locus of centers of all contact circles for a given curve is called the
evolute of the curve. The equation of the evolute is
ρ(t) = r(t) +Rν(t), R = 1/k
1
or in parametric Cartesian form (ρ = (ξ, η))
ξ = x −y
x
2
+y
2
x
y
−x
y
, η = y +x
x
2
+y
2
x
y
−x
y
.
Contact of nth order between a curve and a surface
Quite similarly we can solve a problem of nthorder contact between a space
curve r = r(t) at t = t
0
and a surface from an n + 1 parameter family of
surfaces given implicitly by
F(r, a
1
, . . . , a
n+1
) = 0.
Everything from the previous pages should be repeated word for word. First
we choose a surface from the family in such a way that it coincides with
the curve in n + 1 points close to t
0
. This gives n + 1 equations:
F(r(t
0
), a
1
, . . . , a
n+1
) = 0,
F(r(t
1
), a
1
, . . . , a
n+1
) = 0,
.
.
.
F(r(t
n−1
), a
1
, . . . , a
n+1
) = 0.
162 Tensor Analysis with Applications in Mechanics
Our previous reasoning carries through and we obtain
d
k
dt
k
F(r(t), b
1
, b
2
, . . . , b
n+1
)
t=t
0
= 0 (k = 0, 1, . . . , n),
as the equations that should hold at the point of nthorder contact.
We have introduced the osculating plane to a curve at a point A as the
plane through A that contains τ and ν. Note that it could also be deﬁned
as a surface of secondorder contact. The result is the same.
The reader should consider how to apply these considerations to the
problem of nthorder contact between a given surface and a surface from a
manyparameter family of surfaces.
Exercise 5.36. Treat the problem of thirdorder contact between a space
curve and a sphere. Show that denoting R = 1/k
1
and ρ the radius of the
sphere we get (in natural parametrization)
ρ
2
= R
2
+R
2
/k
2
2
.
Also show that the center of the contact sphere lies on the straight line
through the center of principal curvature that is parallel to the binormal
at the point of the curve.
5.8 Osculating Paraboloid
When considering the structure of a surface at a point, it is often helpful
to approximate the surface using another surface whose behavior is more
easily visualized. A spherical surface would be insuﬃcient for this purpose
because it has the same normal curvature in all directions. We can, however,
use the osculating paraboloid introduced in (5.31). Let us reconsider this
paraboloid from the point of view of local approximation.
Let O be a ﬁxed point of a surface, and assume the surface is suﬃciently
smooth at O. At O we determine the osculating plane and introduce a
Cartesian frame (i
1
, i
2
, n), where (i
1
, i
2
) is a Cartesian frame with origin
O on the osculating plane and n is normal to both the surface and the
osculating plane. For a smooth surface, Cartesian coordinates (x, y) can
play the role of surface coordinates since they uniquely deﬁne any point
of the surface at O. The surface in the vicinity of O can be described by
the equation z = z(x, y). We suppose that z(x, y) is twice continuously
diﬀerentiable near O. In the neighborhood of (0, 0) we can use the Taylor
Elements of Diﬀerential Geometry 163
expansion of z(x, y):
z(x, y) = z(0, 0) +z
x
(0, 0)x +z
y
(0, 0)y
+
1
2
z
xx
x
2
+ 2z
xy
xy +z
yy
y
2
+o(x
2
+y
2
),
where the indices x, y indicate that we take partial derivatives with respect
to x, y, respectively (and in this section, evaluate at the point (0, 0)). Since
the coordinates (x, y) are in the osculating plane, the partial derivatives
z
x
(0, 0) = 0, z
y
(0, 0) = 0.
By choice of the origin, z(0, 0) = 0. Thus the Taylor expansion is
z(x, y) =
1
2
z
xx
x
2
+ 2z
xy
xy +z
yy
y
2
+o(x
2
+y
2
). (5.45)
Let us consider the paraboloid
z(x, y) =
1
2
z
xx
x
2
+ 2z
xy
xy +z
yy
y
2
. (5.46)
The diﬀerence between the surfaces described by (5.45) and (5.46) is small
(as indicated by the o term). This allows us to show that it is an osculating
paraboloid that approximates the behavior of the surface at point O. A
change of coordinate frames can show that the paraboloid (5.46) coincides
with (5.31). Its coeﬃcients are L, M, N at the point O.
Let us note that the second fundamental form of the initial surface at
a point coincides with that of the osculating paraboloid, and the situation
is the same with the normal curvatures. We denote
r = z
xx
(0, 0), s = z
xy
(0, 0), t = z
yy
(0, 0),
so that the equation of the osculating paraboloid is
z =
1
2
rx
2
+ 2sxy +ty
2
.
Let us drawthe projection onto the xycoordinate plane of the crosssections
of the osculating paraboloid by the two planes z = ±h, h > 0. This is the
curve deﬁned by
1
2
rx
2
+ 2sxy +ty
2
= h. (5.47)
The curve is an ellipse when the paraboloid is elliptical (rt − s
2
> 0) or
a hyperbola when it is hyperbolic (rt − s
2
< 0), or a family of straight
lines when it is parabolic (rt −s
2
= 0). It can be shown that the radius of
normal curvature of the surface in the direction x : y is proportional to the
164 Tensor Analysis with Applications in Mechanics
squared distance from the origin to the point of the curve (5.47) taken in
the same direction. This curve is called the Dupin indicatrix.
The Dupin indicatrix is a curve of second order on the plane, and thus
it has some special directions (axes) and special values characterizing the
curve. The special directions are known as principal directions. In the next
section we consider this from another vantage point.
5.9 The Principal Curvatures of a Surface
Let us discuss in more detail the properties of a surface connected with
its second fundamental form. We now denote the coordinates without in
dices, using u = u
1
, v = u
2
. We also use subscripts u, v to indicate partial
diﬀerentiation:
r
1
= r
u
=
∂r
∂u
, r
2
= r
v
=
∂r
∂v
, n
u
=
∂n
∂u
, n
v
=
∂n
∂v
.
The second fundamental form is
Ldu
2
+ 2M du dv +N dv
2
= −dr · dn,
where
L = r
uu
· n, M = r
uv
· n, N = r
vv
· n.
Consider the ﬁrst diﬀerential of n at a point P:
dn = n
u
du +n
v
dv.
Since n is a unit vector, its diﬀerential dn is orthogonal to n and thus lies
in the osculating plane to the surface at P. We know that the diﬀerential
dr = r
u
du + r
v
dv also lies in the plane osculating at P. Let us consider
the relation between the diﬀerentials dr and dn with respect to the vari
ables du, dv now considered as independent variables. It is clearly a linear
correspondence dr → dn. Thus it deﬁnes a tensor A in twodimensional
space such that
dn = A· dr.
This tensor is completely deﬁned by its values n
u
= A· r
u
and n
v
= A· r
v
.
Lemma 5.1. The tensor A is symmetric.
Proof. It is enough to establish the equality
x
1
· (A· x
2
) = x
2
· (A· x
1
)
Elements of Diﬀerential Geometry 165
for a pair of linearly independent vectors (x
1
, x
2
). To show symmetry of A
consider
r
u
· (A· r
v
) = r
u
· n
v
.
Similarly
r
v
· (A· r
u
) = r
v
· n
u
.
The symmetry of A follows from the identity
r
u
· n
v
= r
v
· n
u
;
this is derived by diﬀerentiating the identity n · r
u
= 0 with respect to v,
then diﬀerentiating n · r
v
= 0 with respect to u and eliminating the term
containing r
uv
.
We know that a symmetric tensor has real eigenvalues; in this case there
are no more than two eigenvalues λ
1
and λ
2
to which there correspond
mutually orthogonal eigenvectors x
1
, x
2
, respectively (if λ
1
= λ
2
).
Let us ﬁnd the equation for the eigenvalues. A vector in the osculating
plane can be represented as xr
u
+ yr
v
since (r
u
, r
v
) is a basis in it. By
deﬁnition of A we have
A· (xr
u
+yr
v
) = xn
u
+yn
v
.
On the other hand, xr
u
+ yr
v
= 0 is an eigenvector if there exists λ such
that
A· (xr
u
+yr
v
) = λ(xr
u
+yr
v
) .
Thus, for the same vector we get
xn
u
+yn
v
= λ(xr
u
+yr
v
) . (5.48)
We see that an eigenvector is deﬁned by the same condition as the principal
directions of the previous section, since this equation means that we ﬁnd
the direction in which dr and dn are parallel.
From the last vector equation, let us derive scalar equations. For this,
dot multiply (5.48) ﬁrst by r
u
, then by r
v
. By (5.32) we have
L = −n
u
· r
u
, M = −n
u
· r
v
, N = −n
v
· r
v
,
hence
−Lx −My = λ(Ex +Fy),
−Mx −Ny = λ(Fx +Gy).
166 Tensor Analysis with Applications in Mechanics
Rewriting this as
(L +λE)x + (M +λF)y = 0,
(M +λF)x + (N +λG)y = 0, (5.49)
we have a homogeneous linear system of algebraic equations. This system
has nontrivial solutions when its determinant vanishes:
(EG−F
2
)λ
2
−(2MF −EN −LG)λ + (LN −M
2
) = 0. (5.50)
In the general case, there are two roots λ
1
and λ
2
, to which there correspond
two directions previously called the principal directions. They can be found
by elimination of λ from (5.49):
−x
2
xy −y
2
E F G
L M N
= 0.
By this we deﬁne two mutually orthogonal directions x : y. If λ
1
= λ
2
,
then all the directions are principal so we can choose any two mutually
orthogonal directions and regard them as principal.
Now we would like to consider another question that will bring us to the
same equations. It is the question of ﬁnding the extreme normal curvatures
at a point of a surface. We have established that in the direction x : y the
normal curvature of a surface is given by the formula
k =
Lx
2
+ 2Mxy +Ny
2
Ex
2
+ 2Fxy +Gy
2
.
Because of the secondorder homogeneity in x, y of the numerator and de
nominator we can reformulate the problem of ﬁnding extreme curvatures
as the problem of ﬁnding extremal points of the function
Lx
2
+ 2Mxy +Ny
2
under the restriction
Ex
2
+ 2Fxy +Gy
2
= 1.
Applying the theory of Lagrange multipliers to this, we should ﬁnd the
extreme points of the function
Lx
2
+ 2Mxy +Ny
2
−k(Ex
2
+ 2Fxy +Gy
2
),
which leads to the equations
(L −kE)x + (M −kF)y = 0,
(M −kF)x + (N −kG)y = 0. (5.51)
Elements of Diﬀerential Geometry 167
The systems (5.49) and (5.51) coincide if we put k = −λ, which means
that the principal directions and the directions found here are the same.
Moreover, it is seen that λ
1
and λ
2
found as the eigenvalues of the tensor A
give the extreme curvatures of the surface at a point, which are k
1
= −λ
1
and k
2
= −λ
2
. Remembering that they are the roots of the polynomial
(5.50) and using the Vi`ete theorem, we get
k
1
+k
2
=
EN −2FM +GL
EG−F
2
, k
1
k
2
=
LN −M
2
EG−F
2
.
We have met these expressions in the Gaussian curvature K = k
1
k
2
and
the mean curvature H = (k
1
+k
2
)/2.
We see that both are expressed through the coeﬃcients of the ﬁrst and
second fundamental forms of the surface. There is a famous theorem due
to Gauss that K can be expressed only in terms of the coeﬃcients of the
ﬁrst fundamental form.
When the principal curvatures and their directions are known, the Euler
formula gives the normal curvature at any direction of the curve which
composes the angle φ with the ﬁrst principal direction at the same point,
corresponding to k
1
:
k
φ
= k
1
cos φ +k
2
sin φ.
The principal directions of a surface deﬁne the lines of curvature of the
surface. A line in the surface is called a line of curvature if at each point
its tangent is directed along one of the principal directions at this point of
the surface.
Two lines of curvature pass through each point. Denoting their direc
tions by du : dv and δu : δv we have two equations, the ﬁrst of which means
orthogonality of the directions and the second their conjugation:
E du δu +F(du δv +dv δu) +Gdv δv = 0,
Ldu δu +M(du δv +dv δu) +N dv δv = 0.
It is convenient to take the family of the lines of curvature of the surface
as the coordinate lines of the surface. For these curvilinear coordinates
F = M = 0.
Two surfaces are called isometric if there is a onetoone correspondence
between the points of the surfaces such that corresponding curves in the
surfaces have the same length.
A planar surface and a cylindrical surface provide an example of iso
metric surfaces since we can develop the cylindrical surface over the plane.
168 Tensor Analysis with Applications in Mechanics
The correspondence between points is deﬁned by coincidence of the points
in this developing.
We can refer to local isometry of a surface at point A
1
to another surface
at point B
2
if there are neighborhoods of points on the surfaces that are
isometric.
Two smooth surfaces are locally isometric if and only if there are
parametrizations of the surfaces at corresponding points such that the co
eﬃcients of the ﬁrst principal forms of the surfaces coincide:
E
1
= E
2
, F
1
= F
2
, G
1
= G
2
.
A surface is developable if it is locally isometric to the plane at each
point. It turns out that a surface is developable if and only if its Gaussian
curvature is zero at each point.
Surfaces with zero Gaussian curvature appear once more in the problem
of ﬁnding spatial surfaces of minimal area that have given boundaries. This
popular physics problem describes the shape assumed by a soap ﬁlm on a
wire frame. The energy of such a ﬁlm is proportional to its area, and thus
the form actually taken by such a ﬁlm is the surface of minimum area.
5.10 Surfaces of Revolution
Surfaces of revolution are quite frequent in practice. Suppose that the
surface is formed by rotation of the proﬁle curve
x = φ(u), z = ψ(u), (5.52)
in the xzplane about the zaxis (Fig. 5.3). When we ﬁx u we get a circle
having center on the z axis; it is called a parallel. To deﬁne a point on the
parallel we introduce the angle of rotation v from the xzplane. When we
ﬁx v we get a meridian, a curve congruent to the initial curve (5.52).
It is easy to see that the equations of the surface of revolution corre
sponding to (5.52) are
x = φ(u) cos v, y = φ(u) sin v, z = ψ(u).
Let us ﬁnd the coeﬃcients of the ﬁrst fundamental form of the surface:
E = x
2
u
+y
2
u
+ z
2
u
= (φ
cos v)
2
+ (φ
sin v)
2
+ψ
2
= φ
2
+ψ
2
,
F = x
u
x
v
+y
u
y
v
+z
u
z
v
= (φ
cos v)(−φsin v) + (φ
sinv)(φcos v) = 0,
G = x
2
v
+y
2
v
+z
2
v
= (−φsin v)
2
+ (φcos v)
2
= φ
2
.
Elements of Diﬀerential Geometry 169
Fig. 5.3 Surface of revolution about the zaxis.
Note that F = 0 means the orthogonality of the parametrization net. Thus
the ﬁrst fundamental form is
(ds)
2
=
φ
2
+ψ
2
du
2
+φ
2
dv
2
.
For the components of the second fundamental form we have
L =
r
uu
· (r
u
×r
v
)
r
u
×r
v

=
x
uu
y
uu
z
uu
x
u
y
u
z
u
x
v
y
v
z
v
√
EG−F
2
=
ψ
φ
−φ
ψ
φ
2
+ψ
2
,
M =
r
uv
· (r
u
×r
v
)
r
u
×r
v

=
x
uv
y
uv
z
uv
x
u
y
u
z
u
x
v
y
v
z
v
√
EG−F
2
= 0,
and
N =
r
vv
· (r
u
×r
v
)
r
u
×r
v

=
x
vv
y
vv
z
vv
x
u
y
u
z
u
x
v
y
v
z
v
√
EG−F
2
=
ψ
φ
φ
2
+ψ
2
.
170 Tensor Analysis with Applications in Mechanics
So the second fundamental form is
−dn · dr =
ψ
φ
−φ
ψ
φ
2
+ ψ
2
du
2
+
ψ
φ
φ
2
+ψ
2
dv
2
.
We see that M = 0, which means the coordinate lines are conjugate.
6
Thus
the coordinate lines of a surface of revolution are the lines of curvature.
Exercise 5.37. Find the principal curvatures of the surface of revolution.
Exercise 5.38. Find the ﬁrst and second fundamental forms for (a) the
plane, and (b) the sphere.
Exercise 5.39. Find the ﬁrst and second fundamental forms for each of
the following paraboloids: (a) z = x
2
+ y
2
, (b) z = x
2
− y
2
, (c) z = x
2
.
Find H and K for each of these.
5.11 Natural Equations of a Curve
Suppose we are given functions k
1
(s) and k
2
(s) of a natural parameter
s, continuous on a segment [s
0
, s
1
], and that k
1
(s) is a positive function.
We are interested in whether there exists a space curve that has principal
curvature k
1
(s) and torsion k
2
(s). Geometrical considerations show that if
such a curve exists then it is uniquely deﬁned up to rigid motions. This
means that if we take two such curves, place them in space, and then shift
and rotate one of them so that their initial points and moving trihedra at
these points coincide, then all points of the curves coincide.
Now we will show that such a curve exists. Thus the form of the curve
is deﬁned by k
1
(s) and k
2
(s) uniquely; this prompts us to call the two
functions k
1
= k
1
(s) and k
2
= k
2
(s) the natural equations of the curve.
Let us demonstrate that from k
1
(s) and k
2
(s) we can ﬁnd the needed
curve. For this, we consider a vector system of diﬀerential equations that
6
To any symmetric quadratic form there corresponds something like an inner product of
vectors, and hence something akin to orthogonality. Conjugate directions at a point are
deﬁned as the directions on the surface du : dv and δu : δv for which Lduδu+M(du δv+
dv δu) + N dv δv = 0. Conjugate coordinate lines are those which are conjugate at each
point.
Elements of Diﬀerential Geometry 171
mimics the formula for a tangent vector and the Frenet–Serret equations:
dr
ds
= x,
dx
ds
= k
1
(s)y,
dy
ds
= −k
1
(s)x −k
2
(s)z,
dz
ds
= k
2
(s)y. (5.53)
Written in Cartesian components, this is a system of 12 linear ordinary
diﬀerential equations. ODE theory states that if we deﬁne the initial values
for all the unknowns (the Cauchy problem), we will have a unique solution
to this system.
We can choose arbitrary initial conditions for r(s
0
). The initial values
x(s
0
), y(s
0
), z(s
0
) must constitute an arbitrary righthanded orthonormal
trihedron (x
0
, y
0
, z
0
).
Thus on [s
0
, s
1
] there exists a unique solution (r(s), x(s), y(s), z(s)) of
the equations (5.53), satisfying some ﬁxed initial conditions. It can be
shown that because of skew symmetry of the matrix of the three last equa
tions of (5.53),
0 k
1
(s) 0
−k
1
(s) 0 −k
2
(s)
0 k
2
(s) 0
= 0,
the vectors (x(s), y(s), z(s)) constitute an orthonormal frame of the same
orientation as the initial one on the whole segment [s
0
, s
1
].
We can treat r = r(s) as the equation of the needed curve in natural
parametrization. Then the ﬁrst of the equations (5.53) states that x(s) is
its unit tangent. Comparing the equation dx/ds = k
1
(s)y with the ﬁrst
of the Frenet–Serret equations for a curve r = r(s), we see that y(s) is its
principal normal and k
1
(s) is its principal curvature. Similarly we ﬁnd that
z(s) is the binormal of the curve and k
2
(s) is its torsion. This completes
the necessary reasoning.
For a plane curve the natural equations reduce to a single equation for
the curvature. In the plane we associate the curvature with an algebraic
sign. The condition of positivity of k(s) is not necessary.
172 Tensor Analysis with Applications in Mechanics
Natural equation of a curve in the plane
A plane curve has zero torsion. Let us consider how to reconstruct a curve
if its curvature k(s) is a given function of the length parameter s. We ﬁx
the initial point of the curve, corresponding to s = s
0
, by the equation
r(s
0
) = r
0
. (5.54)
We must also ﬁx the direction of the curve at this point. Here it is useful
to introduce the angle φ = φ(s), measured between the unit tangent to the
curve and the vector i where {i, j} is an orthonormal basis in the plane. In
this way the unit tangent is given by
τ = i cos φ +j sin φ.
So to deﬁne the initial direction of the curve we introduce
φ(s
0
) = φ
0
. (5.55)
The equation relating τ and ν, which is dτ/ds = k(s)ν, becomes
(−i sinφ +j cos φ)
dφ
ds
= k(s)ν.
The vector −i sinφ + j cos φ has unit magnitude and is orthogonal to τ.
Hence it is parallel to ν, and we conclude that
k(s) =
dφ
ds
.
Taking
k(s) =
dφ
ds
,
we deﬁne a sign of k(s) in the standard manner where k(s) is positive for
points at which the curve is concave upwards.
We can integrate the last equation and obtain
φ(s) −φ
0
=
s
s
0
k(t) dt.
Knowing φ(s) we can integrate the equation for the unit tangent, which is
dr/ds = τ, rewritten as
dr
ds
= i cos φ(s) +j sin φ(s).
Integration with respect to s gives us
r(s) −r
0
=
s
s
0
[i cos φ(t) +j sin φ(t)] dt. (5.56)
Elements of Diﬀerential Geometry 173
By (5.56), the function k(s) and the initial values (5.54) and (5.55) uniquely
deﬁned the needed curve r = r(s). Changing the initial values of the curve,
we get a family of curves with the same k(s) such that all curves have the
same shape but diﬀerent positions with respect to the coordinate axes.
Exercise 5.40. Given the curvature k(s) = (as)
−1
of a plane curve, ﬁnd
the curve.
5.12 A Word About Rigor
The main goal of this book is the presentation of those tools and formu
las of tensor analysis that are needed for applications. Our approach is
typical of engineering books; we seldom oﬀer clear statements of the as
sumptions that guarantee validity of a formula, supposing instead that in
applications all the functions, curves, surfaces, etc., should be suﬃciently
smooth for our purposes. This approach is probably best for any practi
tioner who must simply get his or her hands on a formula. However, in
physics we see surfaces that do not necessarily bound realworld bodies;
purely mathematical surfaces can occur, such as those describing the en
ergy of a twoparameter mechanical system. Such a surface can be quite
complex, and a physicist may be largely interested in its singular points
since these represent the states at which the system changes its behavior
crucially. A reader interested in such applications would do well to study a
more sophisticated treatment where each major result is stated as the theo
rem under the weakest possible hypotheses. Many such treatments employ
much more advanced (e.g., topological) tools.
However, the reader should be aware that even on the elementary level
of our treatment there are questions that require additional explanation.
How, for example, should we deﬁne the length of a curve or the area of a
surface?
Early in our education we learn how to measure the length of a segment
or of a more complex set on a straight line. We also learn how to deﬁne
the circumference of a circle. For this we inscribe an equilateral triangle
and calculate its perimeter. Doubling the number of sides of the inscribed
triangle, we get an inscribed polygon whose perimeter approximates the
circumference. Doubling the number of sides to inﬁnity and calculating
the limit of their perimeters, we deﬁne this limit as the needed length. Of
174 Tensor Analysis with Applications in Mechanics
course, a limit passage of this type is used when we derive the formula
b
a
r
(t) dt (5.57)
for the length of a curve. Here we divide the segment [a, b] into small pieces
by the points t
0
= a, t
1
, . . . , t
n
= b, and draw the vector
∆r(t
i
+ 1) = r(t
i+1
) −r(t
i
).
In this way we “inscribe” a polygon into the curve. For a continuously
diﬀerentiable vector function r(t), the length of a side of a polygon can be
found by the mean value theorem:
r(t
i+1
) −r(t
i
) = r
(ξ
i+1
)(t
i+1
−t
i
),
where ξ
i+1
is a point in [t
i+1
− t
i
]. Thus the perimeter of the inscribed
polygon that approximates the length of the curve is
n
¸
i=1
r
(ξ
i+1
)(t
i+1
−t
i
).
We obtain (5.57) from a limit passage in which the number of partition
points tends to inﬁnity while the maximum length of a partition segment
tends to zero.
Although this might seem ﬁne, underlying the process is the notion of
approximating a small piece of the curve by a manysided polygon. Our
“intuition” tells us that the smaller its sides, the closer is the polygon to
the curve, hence its perimeter should approximate the curve length more
and more closely. But in elementary geometry one may “demonstrate” that
the sum of the legs of any right triangle is equal to its hypotenuse. The
construction is as follows. Suppose we are given a right triangle whose legs
have lengths a and b. Let us form a curve that looks like the toothed edge
of a saw, by placing along the hypotenuse of the given triangle a sequence
of small triangles each similar to the given triangle (Fig. 5.4). It is clear
that the sum of all the legs of the sawteeth does not depend on the number
of teeth and equals a +b. But as the number of teeth tends to inﬁnity, the
saw edge “coincides” with the hypotenuse of the original triangle. Hence
the limiting length is equal to the length of the hypotenuse c, and we
have c = a + b. So we come to understand that we cannot approximate a
curve by a polygon in an arbitrary fashion. Similar remarks apply to the
approximation of surfaces to calculate area.
Elements of Diﬀerential Geometry 175
Fig. 5.4 Fallacious estimation of the length of a line.
These remarks should serve as a warning that to apply formulas cor
rectly it is often necessary to understand the restrictions under which they
were derived.
5.13 Conclusion
Diﬀerential geometry is a well developed subject with many results and
formulas. We have presented the main technical formulas that are used
in applications. The interested reader could go on to study volumes de
voted to the theory of curves, surfaces, manifolds, etc. Such an extensive
undertaking falls outside the scope of this book.
5.14 Problems
5.1 Find the parametrization and singular points of the plane curve
x
2
3
+y
2
3
= a
2
3
,
where a is a parameter.
5.2 Find the singular points of the plane curve given by the equations
x = a(t −sin t), y = a(1 −cos t),
where a is a parameter.
5.3 Find the equation of the tangent to the curve
x
2
+y
2
+z
2
= 1, x
2
+y
2
= x
at the point (0, 0, 1).
176 Tensor Analysis with Applications in Mechanics
5.4 Find the length of the astroid
x = a cos
3
t, y = a sin
3
t.
5.5 Find the length of the part of the cycloid
x = a(t −sin t), y = a(1 − cos t)
deﬁned by t ∈ [0, 2π].
5.6 Find the length of the cardioid given in polar coordinates by
ρ = 2a(1 −cos φ).
See Fig. 5.5.
2a
4a
y
x
0
Fig. 5.5 Cardioid.
5.7 Find the length of the portion of the curve
x = a cosht, y = a sinh t, z = at
between the points t = 0 and t = T.
5.8 Find the curvature of the curve
x = t − sint, y = 1 −cos t, z = 4 sin
t
2
.
5.9 Find the curvature and torsion of the curve
x = a cosht, y = a sinh t, z = at
at an arbitrary point.
Elements of Diﬀerential Geometry 177
5.10 Find the torsion of the curve
x = a cosht cos t, y = a cosht sin t, z = at
at an arbitrary point.
5.11 Let a triad of orthonormal vectors e
1
(s), e
2
(s), e
3
(s) (i.e., vectors
satisfying e
i
· e
j
= δ
ij
) be given along a curve. Show that
d
ds
e
i
(s) = d ×e
i
(s)
where
d = −
1
2
(e
i
×e
i
).
Note that these formulas are analogous to the Frenet–Serret equations with
the Darboux vector δ.
5.12 Let Q(s) be an orthogonal tensor given along a curve. Verify the
formulas
Q
= d ×Q, d = −
1
2
(Q
×Q
T
)
×
.
5.13 Find the second quadratic form of the surface given by the equations
x = u cos v, y = u sinv, z = v.
5.14 A surface is deﬁned by the equation z = f(x, y). Show that the
coeﬃcients of its second principal form are
b
11
=
f
xx
1 +f
2
x
+f
2
y
, b
12
= b
21
=
f
xy
1 +f
2
x
+f
2
y
, b
22
=
f
yy
1 +f
2
x
+ f
2
y
.
5.15 A surface is deﬁned by the equation z = f(x, y), where f satisﬁes
Laplace’s equation ∇
2
f = 0. Demonstrate that its Gaussian curvature
satisﬁes K ≤ 0.
5.16 Show that the mean curvature of the surface z = f(x, y) is given by
H = div
gradf
1 + gradf
2
,
where
grad = i
1
∂
∂x
+i
2
∂
∂y
.
178 Tensor Analysis with Applications in Mechanics
5.17 Show that the mean curvature of a surface is given by H = −
¯
∇· n/2.
5.18 Demonstrate that
¯
∇· A = 2Hn where A = E−nn.
5.19 Let X be a secondorder tensor. Show that n·(
¯
∇×X) = −
¯
∇·(n×X).
5.20 Find the Gaussian curvature of a surface given by the relation
z = f(x) +g(y).
5.21 The ﬁrst principal form of a surface is A
2
du
2
+ B
2
dv
2
. Determine
its Gaussian curvature.
5.22 Let X be a secondorder tensor. Prove the following analog of the
Gauss–Ostrogradsky theorem on a surface S having boundary contour Γ:
S
¯
∇· X+ 2Hn · X
dS =
Γ
ν · Xds, (5.58)
where ν is the outward unit normal to Γ lying in the tangent plane, i.e.,
ν · n = 0. Note that when S is a closed surface,
S
¯
∇· XdS = −
S
2Hn · XdS.
5.23 Prove that (5.58) holds for a tensor ﬁeld X of any order.
5.24 Use the solution of the previous problem to prove that
S
¯
∇X+ 2HnX
dS =
Γ
νXds,
S
¯
∇×X+ 2Hn ×X
dS =
Γ
ν ×Xds,
and
S
¯
∇×(nX) dS =
Γ
τXds,
where τ = ν ×n is the unit tangent vector to Γ.
Chapter 6
Linear Elasticity
In this chapter we apply tensor analysis to linear elasticity. Linear elasticity
is a powerful tool of engineering design; using general computer programs,
engineers can calculate the strains and stresses within a complex elastic
body under load. It is useful to learn the principles behind these calcu
lations. Linear elasticity is based on the ideas of classical mechanics, but
its technical tools are those of tensor analysis. It treats the small defor
mations of elastic bodies described by a linear constitutive equation that
relates stresses and strains, extending the elementary form of Hooke’s law
for a spring.
Linear elasticity is the ﬁrst step (an elementary but not easy step) to
ward nonlinear mechanics. The latter considers other eﬀects in solids, such
as heat propagation, deformation due to piezoelectric or magnetic eﬀects,
etc. It therefore incorporates thermodynamics and other areas of physics.
The plan of this chapter is to introduce the principal tools and laws of
linear elasticity, to formulate and consider some properties of the boundary
value problems of elasticity, and to study some variational principles in
applied elasticity.
We start with the idea of the stress tensor, which was introduced by
Augustin Louis Cauchy (1789–1857).
6.1 Stress Tensor
The notion of stress is decisive in continuum mechanics. It is an extension
of the notion of pressure, and is introduced as the ratio of the value of a
force distributed over an elementary surface element to the element area.
Consider, for example, a long thin cylindrical bar having crosssectional
area S and stretched by a force f which is uniformly distributed over the
181
182 Tensor Analysis with Applications in Mechanics
cylinder faces (Fig. 6.1). This force is equidistributed over any normal
crosssection of the bar. The stress σ is deﬁned as
σ = f/S.
In elementary physics, the pressure is the only force characteristic at a
given point in a liquid or gas. But pressure is not suﬃcient to describe the
action of forces inside a threedimensional solid body. At a particular point
in such a body, we ﬁnd that the direction of the contact force may not be
normal to a given area element. Furthermore, as we change the orientation
of the area element, the density of the force acting across the element may
change in magnitude and direction.
f
f
S
Fig. 6.1 The stress σ = f/S in a bar having crosssectional area S and stretched by a
force f.
Forces
Let us discuss the conditions for equilibrium of a deformable body. From
classical mechanics, we know that the equilibrium conditions for a rigid
body consist of two vector equations. First, the resultant force (i.e., the
sum of all forces) acting on the body must be zero. Second, the resultant
moment (the sum of the moments of all the forces with respect to some
point) must be zero. We write
¸
k
f
k
= 0,
¸
k
(r
k
−r
0
) f
k
= 0, (6.1)
where f
k
denotes a force applied to a point located by position vector r
k
.
The position vector r
0
locates an arbitrary but ﬁxed point with respect to
which the moments are taken.
To deﬁne the equilibrium of a deformable body, we apply the equilibrium
equations for a rigid body to any portion of the deformable body. We accept
this as an axiom, known as the
Solidiﬁcation principle. In equilibrium, any part of a deformable body
obeys the equilibrium equations as if it were a rigid body under the action
of (1) all the external forces, and (2) the force reaction imposed by the
remainder of the body on the part under consideration.
Linear Elasticity 183
This principle is not a direct consequence of classical mechanics. Rather,
it is a kind of axiom which allows us to apply the results of classical me
chanics — obtained for nondeformable objects — to deformable objects.
We will use the following terminology. A solid body occupies a certain
volume in threedimensional space. The mapping which takes its material
points into the spatial points is called a conﬁguration of the body; roughly
speaking, this describes the geometry of the body in space. If the body
is not under load, the conﬁguration is termed the initial or reference con
ﬁguration. For a deformed body in equilibrium under a given load, the
conﬁguration is termed the actual conﬁguration.
Let us apply the solidiﬁcation principle in the actual conﬁguration of a
deformable body in equilibrium. We take an arbitrary portion { as shown
in Fig. 6.2. Two types of forces act on {. First, there are body forces.
These act on the interior of { and do not depend on the conditions over
the boundary surface of {. An example is the gravitational force. Second,
there are contact forces. These act on the boundary of { and represent the
reaction forces imposed on { by the remainder of the body. They arise as
follows. Imagine that we isolate { from the body and replace the eﬀects
of the rest of the body by some forces. These reactions are applied only to
that part of the boundary of { that comes into contact with the rest of the
body (hence the term “contact forces”). In addition, surface forces may
act over the external boundary of the entire body. We will refer to these,
together with the body forces, as the external forces. By the solidiﬁcation
principle, { must be in equilibrium under the action of all forces as though
it were a rigid body of classic mechanics. (Continuum mechanics permits
another approach, in which the reactions of the remaining part arise in the
volume as well. However, this approach is not used in modern engineering
practice.)
So the total force acting on { is
f ({) = f
B
({) +f
C
({),
where the subscripts B and C denote the body and contact forces, respec
tively.
To characterize the spatial distribution of the forces, we introduce force
densities deﬁned by the equalities
f
B
({) =
V
P
ρf dV, f
C
({) =
Σ
P
t dΣ,
where V
P
is the space volume of {, Σ
P
= ∂V
P
is the boundary of {, and ρ
is the (speciﬁc) density of the material composing the body. The density t
184 Tensor Analysis with Applications in Mechanics
carries dimensions of force per unit area, whereas f carries those of force per
unit mass. We emphasize that these are deﬁned in the actual conﬁguration.
The quantity t is called the stress vector.
t
f
V
P
V
P
Σ
P
Fig. 6.2 Forces acting on a portion V
P
of the body.
Equilibrium equations of a continuum medium
By the solidiﬁcation principle, the equilibrium equations (6.1) for a rigid
body become the following two conditions for a deformable body.
1. The resultant force acting on any portion { is zero:
V
P
ρf dV +
Σ
P
t dΣ = 0. (6.2)
2. The resultant moment of all forces acting on { is zero:
V
P
¦(r −r
0
) ρf ¦ dV +
Σ
P
¦(r −r
0
) t¦ dΣ = 0, (6.3)
where r locates a material point and r
0
locates a ﬁxed reference point. The
reader is encouraged to show, using (6.2), that (6.3) does not depend on
the choice of r
0
.
Note that { is not completely arbitrary, but is such that the integration
operation over { makes sense.
Linear Elasticity 185
Stress tensor
In general, the stress vector depends on the position r of a particle and on
the normal n to the area element in the body. We will always take the
normal outward from the portion of the body under consideration.
n
r
n
n
t( n) 
t(n)
i
i
i
1
2
3
Fig. 6.3 Interaction of two parts of the body.
In continuum mechanics, Newton’s third law on actionreaction pairs is
called Cauchy’s lemma and is expressed by the following relation:
t(r, n) = −t(r, −n). (6.4)
Formula (6.4) describes the interaction of the contacting parts of the body
shown in Fig. 6.3. Cauchy’s lemma allows us to introduce the stress tensor,
which describes the dependence of t on n, the normal to the area element
at a point. We will see this in Cauchy’s theorem below.
First, however, we will use (6.2) to obtain the equilibrium equation in
diﬀerential form. We ﬁx an arbitrary point P in V , the volume of the body,
and use it as a vertex of an arbitrary small parallelepiped Π having faces
parallel to the Cartesian coordinate planes as shown in Fig. 6.4. Hence the
normals to the faces lie along the orthonormal basis vectors i
1
, i
2
, i
3
.
Let us expand the stress vector t(i
k
) on the face having normal i
k
:
t(i
k
) = t
ks
i
s
. (6.5)
That is, the t
ks
are the components of t(i
k
) in the basis i
s
.
186 Tensor Analysis with Applications in Mechanics
x
x
x
P
Fig. 6.4 Parallelepiped Π.
Theorem 6.1. The diﬀerential equation
ρf +
∂t
ks
∂x
k
i
s
= 0, (6.6)
known as the equilibrium equation, holds in V .
Proof. Equation (6.2) for the parallelepiped Π is
V
Π
ρf dV +
Σ
Π
t dΣ = 0.
The surface integral is taken over the faces of Π, each of which is perpen
dicular to one of the i
k
. On the face whose normal is n = i
k
, we have
t(i
k
) = n
k
t
1s
i
s
, where n
k
= 1 and the other two components of n vanish.
On the opposite face the normal is −i
k
, so n
k
= −1 and the remaining
components vanish. By Cauchy’s lemma (6.4) we have t(−i
k
) = −t(i
k
),
hence t(−i
k
) = n
k
t
1s
i
s
. Therefore the above equation takes the form
V
Π
ρf dV +
Σ
Π
n
k
t
ks
i
s
dΣ = 0.
Applying (4.41) to the surface integral, we have
Σ
Π
n
k
t
ks
i
s
dΣ =
V
Π
∂t
ks
∂x
k
i
s
dV (6.7)
and it follows that
V
Π
ρf +
∂t
ks
∂x
k
i
s
dV = 0. (6.8)
Suppose the integrand in (6.8) is a continuous function. As the vertex P
of Π is ﬁxed and Π is arbitrarily small, the diﬀerential equation (6.6) must
hold at P. Since P is arbitrary, (6.6) holds in V .
Linear Elasticity 187
Exercise 6.1. Prove (6.7) by direct integration of the righthand side.
Now we formulate
Theorem 6.2. (Cauchy). At any point of a body, the dependence of t on
n, the normal to an elementary area at a point, is linear:
t = n σ.
Here σ is a secondorder tensor depending on the point; it is the Cauchy
stress tensor.
Proof. We construct a tetrahedron T with vertex O at an arbitrary but
ﬁxed point of V as shown in Fig. 6.5. Applying (6.2) to T, we get
V
T
ρf dV +
Σ
T
n
k
t
ks
i
s
dΣ +
M
1
M
2
M
3
t(n) dΣ = 0, (6.9)
where V
T
is the tetrahedron volume, Σ
T
is the part of tetrahedron bound
ary consisting of the faces that are parallel to the coordinate planes, and
M
1
M
2
M
3
is the inclined face. On the face that is orthogonal to i
k
, we
will use the representation t(i
k
) = n
k
t
1s
i
s
from the proof of the previous
theorem.
x
x
x
Ì
Ì
Ì
o
Fig. 6.5 Tetrahedron T.
Using equation (6.6), we transform (6.9) to the following equality:
V
T
∂t
ks
∂x
k
i
s
dV =
Σ
T
n
k
t
ks
i
s
dΣ +
M
1
M
2
M
3
t(n) dΣ.
188 Tensor Analysis with Applications in Mechanics
A consequence of (4.41) for V = V
T
is
V
T
∂t
ks
∂x
k
i
s
dV =
Σ
T
n
k
t
ks
i
s
dΣ +
M
1
M
2
M
3
n
k
t
ks
i
s
dΣ.
Comparing the last two equalities, we see that
M
1
M
2
M
3
t(n) −n
k
t
ks
i
s
dΣ = 0.
As the tetrahedron is small and arbitrary at point O, we get the following
identity:
t(n) −n
k
t
ks
i
s
= 0.
Because M
1
M
2
M
3
can have an arbitrary orientation n, this equality holds
for any n and, moreover, at any point of V .
We have shown that the dependence of t on n is linear. As we know
(' 3.2), a linear dependence between two vectors is given by a secondorder
tensor that we denote by σ:
t(n) = n σ. (6.10)
The proof is complete.
By the proof of Cauchy’s theorem, we see that the components of the
matrix (t
sk
) are the components of the stress tensor in a Cartesian frame:
σ = t
sk
i
k
i
s
.
To maintain the correspondence between the notations for σ and its com
ponents, we change t
sk
to σ
sk
and write
σ = σ
sk
i
k
i
s
.
Each subscript of σ
sk
has a certain geometric meaning. The ﬁrst subscript
designates the area element with normal i
k
, while the second designates the
direction of the projection of the stress vector onto i
s
. For example, σ
31
is
the projection of t(i
3
) onto the axis x
1
, and the stress vector t(i
3
) acts on
the elementary surface element having normal parallel to i
3
.
Now we return to the equilibrium equation. It is easy to see that the
equation
∇ σ + ρf = 0 (6.11)
written in Cartesian coordinates is (6.6). This means we have found the
componentfree form of the equilibrium equation for the body. Note that
the equation does not depend on the properties of the material that makes
Linear Elasticity 189
up the body. We recall that (6.11) is the diﬀerential form of the condition
that the resultant force applied to an arbitrary part of the body is zero.
So far we have exploited only the force equation (6.2). Now we will
formulate the consequences of the moment equation (6.3).
Theorem 6.3. Let equation (6.3) hold for any part of the body. It follows
that σ is a symmetric tensor: σ = σ
T
.
Proof. Using (4.41), we change the surface integral in (6.3) to an integral
over V
P
:
Σ
P
(r −r
0
) t dΣ =
Σ
P
(r −r
0
) (n σ) dΣ
= −
Σ
P
n σ (r −r
0
) dΣ
= −
V
P
∇ [σ (r −r
0
)] dV. (6.12)
Let us transform the integrand of the last integral. Because ∂r/∂x
k
= i
k
and ∂r
0
/∂x
k
= 0, we have
∇ [σ (r −r
0
)] = ∇ σ (r −r
0
) +i
k
σ
∂
∂x
k
(r −r
0
)
= −(r −r
0
) ∇ σ +i
k
σ i
k
= −(r −r
0
) ∇ σ −σ
ks
i
k
i
s
.
Thus the condition that the resultant moment of all forces acting on V
P
is
zero brings us to the relation
V
P
(r −r
0
)
ρf +∇ σ
dV +
V
P
σ
ks
i
k
i
s
dV = 0. (6.13)
The ﬁrst integral in (6.13) is zero by (6.11). So the second integral in (6.13)
is zero for arbitrary V
P
, and it follows that
σ
ks
i
k
i
s
= 0 in V.
This holds if and only if σ is symmetric at each point, i.e.,
σ = σ
T
.
Indeed, let us consider the part of the sum σ
ks
i
k
i
s
when k, s are 1 or 2.
We have
σ
ks
i
k
i
s
= σ
11
i
1
i
1
+ σ
22
i
2
i
2
+ σ
12
i
1
i
2
+ σ
21
i
2
i
1
= (σ
12
−σ
21
)i
3
= 0,
190 Tensor Analysis with Applications in Mechanics
which implies that σ
12
= σ
21
. Similarly we may demonstrate that σ
23
= σ
32
and σ
13
= σ
31
. This completes the proof.
It is worth noting that in continuum mechanics other types of stresses,
such as couple stresses, can be introduced
[
Cosserat and Cosserat (1909);
Eringen (1999)
]
. For such models, the Cauchy stress tensor is not symmetric
in general.
Principal stresses and principal area elements
In a general basis e
k
(k = 1, 2, 3), the Cauchy stress tensor σ takes the
form
σ = σ
sk
e
s
e
k
,
where the matrix σ
sk
has only six independent components.
Because σ is symmetric, there exists the spectral expansion (3.21):
σ = σ
1
i
1
i
1
+ σ
2
i
2
i
2
+ σ
3
i
3
i
3
. (6.14)
Here the eigenvalues σ
k
of the matrix (σ
sk
) are the principal stresses, and
the normalized eigenvectors i
k
of σ are the principal axes of σ. On the
principal area element having normal i
k
, the tangential stresses are absent.
When the σ
k
are distinct, the frame i
1
, i
2
, i
3
is orthonormal. For the case
of repeated σ
k
, the frame of i
k
is not unique; even in this case, however, we
can select an orthonormal set i
1
, i
2
, i
3
.
6.2 Strain Tensor
Under load, a body changes shape. We will consider how to describe defor
mation using the strain tensor. We restrict our consideration to very small
deformations.
Let us illustrate the notion of strain using a stretched bar as an example.
An undeformed bar has length l
0
; under load, the length becomes l. The
strain is
ε =
l −l
0
l
0
≡
∆l
l
0
.
Generalization of this to three dimensions is not straightforward: we should
consider changes in shape of the body in all directions.
Let a body initially occupy a volume V in space. Under some external
load, it occupies the volume v. The position vectors of a particle in the
Linear Elasticity 191
initial and deformed states are denoted by r
0
and r, respectively. The
displacement vector
u = r −r
0
describes the displacement of a particle due to deformation (Fig. 6.6). In
this book we restrict ourselves to the case in which u <1 and all the ﬁrst
derivatives of u are small in comparison with 1. So we will omit all terms
of the second order of smallness in any expression containing ﬁrstorder
terms. Moreover, in the case of small deformations we will not distinguish
the initial and actual states of the body; that is, all quantities will be
considered to be given in the initial volume V .
r
r
0
u
i
1
i
i
2
3
Fig. 6.6 Deformation of a threedimensional body.
Let us consider the change of an inﬁnitesimal vectorsegment dr
0
due
to deformation. After deformation the segment is given by the vector dr.
We have
dr = dr
0
F, (6.15)
where F = E+∇u is the gradient of the deformation.
1
Exercise 6.2. Derive (6.15) from the relation r = r
0
+u.
Next, we consider the change in length of the segment during deforma
tion. Before deformation, the squared length is
dS
2
= dr
0
dr
0
.
1
Some books use F = E+ ∇u
T
, in which case (6.15) takes the form dr = F · dr
0
.
192 Tensor Analysis with Applications in Mechanics
After deformation it is
ds
2
= dr dr = dr
0
F F
T
dr
0
.
We have
ds
2
−dS
2
= dr
0
(F F
T
−E) dr
0
= dr
0
(∇u + (∇u)
T
+ (∇u) ∇u
T
)
dr
0
.
For small deformations, we may omit all the squared quantities and write
ds
2
−dS
2
= dr
0
(∇u + (∇u)
T
)
dr
0
= 2dr
0
ε dr
0
,
where
ε =
1
2
∇u + (∇u)
T
(6.16)
is the linear strain tensor. It is clear that ε is a symmetric tensor.
In a Cartesian frame, ε is given by
ε = ε
mn
i
m
i
n
where
ε
11
=
∂u
1
∂x
1
, ε
12
=
1
2
∂u
1
∂x
2
+
∂u
2
∂x
1
,
ε
22
=
∂u
2
∂x
2
, ε
13
=
1
2
∂u
1
∂x
3
+
∂u
3
∂x
1
,
ε
33
=
∂u
3
∂x
3
, ε
23
=
1
2
∂u
2
∂x
3
+
∂u
3
∂x
2
.
The diagonal components ε
11
, ε
22
, ε
33
describe the changes in the lengths of
elementary segments along the i
1
, i
2
, i
3
directions, respectively. The other
components ε
mn
(m = n) represent skewing of the body; they characterize
the deformational changes in the angles between elementary segments lying
initially along the axes.
In arbitrary curvilinear coordinates q
1
, q
2
, q
3
with basis r
k
(k = 1, 2, 3)
and dual basis r
k
(k = 1, 2, 3), the tensor ε is given by
ε = ε
st
r
s
r
t
, ε
st
=
1
2
∂u
s
∂q
t
+
∂u
t
∂q
s
−Γ
r
st
u
r
.
In arbitrary orthogonal curvilinear coordinates, ε was represented in (4.39).
See Appendix A for ε in the cylindrical and spherical systems.
Linear Elasticity 193
Equation (6.16) deﬁnes ε as a tensorial function of u. The inverse
problem, of ﬁnding u when ε is given, has a solution if and only if the
compatibility condition
∇(∇ε)
T
= 0 (6.17)
holds. In this case, u can be found using Ces` aro’s formula
u = u
0
+ω
0
(r −r
0
) +
M
M
0
¦ε(s) + [r(s) −r] ∇ε(s)¦ dr(s), (6.18)
where u
0
and ω
0
are arbitrary but ﬁxed vectors, and the integration path
M
0
M joins the points M
0
and M whose position vectors are r
0
and r,
respectively. Here r(s) locates an arbitrary point on M
0
M.
The derivations of these formulas can be found in any book on elasticity,
e.g.,
[
Green and Zerna (1954); Lurie (2005)
]
.
6.3 Equation of Motion
Using the equilibrium equations (6.11) and d’Alembert’s principle of me
chanics, we can immediately obtain the equation of motion for a body. The
technique is to formally add the inertia forces to the body forces:
f →f −ρ
∂
2
u
∂t
2
,
where t is the time variable and ρ is the material density. The equation of
motion is
∇ σ + ρf = ρ
∂
2
u
∂t
2
. (6.19)
The form of (6.19) is simplest in Cartesian coordinates. Putting
σ = σ
mn
i
m
i
n
, u = u
m
i
m
,
we get
∂σ
ij
∂x
i
+ ρf
j
= ρ
∂
2
u
j
∂t
2
(j = 1, 2, 3)
194 Tensor Analysis with Applications in Mechanics
or explicitly
∂σ
11
∂x
1
+
∂σ
21
∂x
2
+
∂σ
31
∂x
3
+ ρf
1
= ρ
∂
2
u
1
∂t
2
,
∂σ
12
∂x
1
+
∂σ
22
∂x
2
+
∂σ
32
∂x
3
+ ρf
2
= ρ
∂
2
u
2
∂t
2
,
∂σ
13
∂x
1
+
∂σ
23
∂x
2
+
∂σ
33
∂x
3
+ ρf
3
= ρ
∂
2
u
3
∂t
2
.
With regard for (4.33), in curvilinear coordinates equation (6.19) takes
the form
1
√
g
∂
∂q
i
√
gσ
ij
r
j
+ ρf = ρ
∂
2
u
∂t
2
. (6.20)
In components it is
∂
∂q
i
√
gσ
ij
+ Γ
j
mn
σ
mn
+ ρ
√
gf
j
= ρ
√
g
∂
2
u
j
∂t
2
(j = 1, 2, 3). (6.21)
The equations of motion can be simpliﬁed in orthogonal coordinates. See
Appendix A for expressions in cylindrical and spherical coordinates.
6.4 Hooke’s Law
Equations (6.16) and (6.19) apply to any small deformation of a continu
ous medium. However, they do not uniquely deﬁne the deformations and
stresses in a body. To study a body under load, we should relate the stresses
to the strains using the material properties of the body. These relations are
called constitutive equations. In this book we consider linearly elastic ma
terials for which the constitutive equation represents a linear dependence
between σ and ε. The simplest version is Hooke’s law
σ = Eε, (6.22)
which describes the elastic properties of a thin rod under tension or com
pression. Here E is the elastic modulus of the material from which the rod
is made; it is known as Young’s modulus.
Robert Hooke (1635–1703) was the ﬁrst to establish the linear depen
dence f ∼ ∆l between the applied force and the elongation of a bar similar
to that in Fig. 6.1. Thomas Young (1773–1829) introduced the elastic mod
ulus E as a quantity that does not depend on the crosssectional area of
the bar; rather, E characterizes the material itself. The linear dependence
(6.22) holds only in some range [ε[ < ε
0
, where ε
0
depends on the material,
Linear Elasticity 195
temperature, and other factors. However, the importance of Hooke’s law
in engineering cannot be overestimated.
In the general case, a linear dependence between the secondorder ten
sors σ and ε is presented in equation (3.31), known as the generalized
Hooke’s law:
σ = C ε. (6.23)
In a Cartesian basis it is
σ
ij
= c
ijmn
ε
mn
.
The fourthorder tensor of elastic moduli
C = c
ijmn
i
i
i
j
i
m
i
n
has 81 components; only 36 of these are independent, however, as the sym
metries of σ and ε lead to the conditions
c
ijmn
= c
jimn
= c
ijnm
.
In linear elasticity it is shown
[
Lurie (2005)
]
that we can introduce the
strain energy
1
2
V
W dV
stored in the elastic body by virtue of its deformation. The integrand
W = W(ε), the strain energy function, is a quadratic form in ε:
W =
1
2
ε C ε =
1
2
ε
ij
c
ijmn
ε
mn
.
The fact that W is uniquely deﬁned for any deformation requires C to
possess an additional symmetry property which, in terms of components,
is
c
ijmn
= c
mnij
for any indices i, j, m, n. Indeed, we can represent C as the sum of two
tensors
C = C
+C
,
where the components of C
satisfy
c
ijmn
= c
mnij
and those of C
satisfy
c
ijmn
= −c
mnij
.
196 Tensor Analysis with Applications in Mechanics
(This is similar to the representation of a secondorder tensor as the sum
of symmetric and antisymmetric tensors.) For any symmetric tensors ε
k
,
we have
ε
1
C
ε
2
= −ε
2
C
ε
1
,
hence for any strain tensor ε,
ε C
ε = −ε C
ε
and
ε C
ε = 0.
It follows that
ε C ε = ε C
ε.
Thus C
does not aﬀect the values of W, and the constitutive relations
should not include it. Setting C
to zero, we get C = C
. Thus the
components of C have the following symmetry properties:
c
ijmn
= c
jimn
= c
ijnm
, c
ijmn
= c
mnij
,
for any indices i, j, n, m. Elementary calculation shows that these are 60
equalities. So in Hooke’s law, of the 81 components of C there remain only
81 −60 = 21 independent elastic constants.
It follows from Exercise 3.44 that W is a potential for σ, i.e., that
σ = W
,ε
where W
,ε
is the derivative of W with respect to ε.
A dual relation
ε = W
,σ
holds when we express W in terms of σ, i.e., W = W(σ). We urge the
reader to show this as an exercise.
In engineering analysis, the vector and matrix notations are typically
used to describe an elastic body. Because of the symmetry properties of
the tensors, all the relations are written in terms of formal sixdimensional
“vectors” for the components of σ and ε, and 6 6 matrices for C. Voigt’s
rule shows one how to transform the tensor notation to the matrixvector
notation. The rule for changing c
ijmn
→ C
pq
is as follows. The pairs of
indices 11, 22, 33 change to 1, 2, 3, the pairs 23 and 32 to 4, the pairs 13
and 31 to 5, and the pairs 12 and 21 to 6. For example, c
1122
→ C
12
and
Linear Elasticity 197
c
1232
→ C
64
. The symmetry property C
pq
= C
qp
holds. The components
of the stress and strain tensors are transformed by the formulas
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
σ
11
σ
22
σ
33
σ
23
σ
31
σ
12
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
=
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
σ
1
σ
2
σ
3
σ
4
σ
5
σ
6
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
,
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
ε
11
ε
22
ε
33
2ε
23
2ε
31
2ε
12
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
=
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
ε
1
ε
2
ε
3
ε
4
ε
5
ε
6
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
. (6.24)
In this notation, Hooke’s law takes the form
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
σ
1
σ
2
σ
3
σ
4
σ
5
σ
6
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
=
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
C
11
C
12
C
13
C
14
C
15
C
16
C
12
C
22
C
23
C
24
C
25
C
26
C
13
C
23
C
33
C
34
C
35
C
36
C
14
C
24
C
34
C
44
C
45
C
46
C
15
C
25
C
35
C
45
C
55
C
56
C
16
C
26
C
36
C
46
C
56
C
66
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
ε
1
ε
2
ε
3
ε
4
ε
5
ε
6
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
. (6.25)
The general case in which all 21 elastic constants in (6.25) are inde
pendent is not common in applications. In engineering practice, the ma
terials usually possess symmetries that reduce the number of independent
constants. Most common are isotropic materials; such a material exhibits
complete symmetry of its properties in space so that at each point the ma
terial properties do not vary with direction. Mathematically, symmetry is
expressed in the language of group theory. Examples of isotropic materials
are steel, aluminium, many other metals, polymeric materials, etc.
For an isotropic material, the relation between σ and ε is given by a
linear isotropic function. This function, considered in Chapter 3, now takes
the form
σ = λEtr ε + 2µε, (6.26)
where λ and µ are Lam´e’s moduli. We may also refer to µ as the shear
modulus. For an isotropic material,
C = λEE+ µ(e
k
Ee
k
+I).
198 Tensor Analysis with Applications in Mechanics
In matrix notation, to C there corresponds the diagonal matrix
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
λ + 2µ 0 0 0 0 0
0 λ + 2µ 0 0 0 0
0 0 λ + 2µ 0 0 0
0 0 0 2µ 0 0
0 0 0 0 2µ 0
0 0 0 0 0 2µ
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
.
Other elastic constants were used historically in engineering work, and
in terms of these, Hooke’s law takes other forms. Common engineering
constants include Young’s modulus E, Poisson’s ratio ν, the bulk modulus k,
and the shear modulus G. For isotropic materials, these are used in pairs: E
and ν, and k and G. The modulus E originated in bar stretching problems;
it relates the tension applied to the bar with the resulting strain. The
modulus k is used to describe uniform volume deformation of a material,
say of a ball under uniform pressure. G describes the shear characteristics
of a material. Finally, ν describes the lateral shortening of a stretched
band. We will derive some relations between the constants.
Let us split σ into a ball tensor and deviator:
σ =
1
3
σE + dev σ.
Then (6.26) takes the form
σ =
1
3
(3λ + 2µ) tr ε, dev σ = 2µdev ε.
The bulk modulus k relates the mean stress σ with the bulk strain tr ε
through
σ = k tr ε,
and so
k = λ + 2µ/3.
Shear stresses are related to shear strains by
σ
ij
= 2Gε
ij
(i = j).
The above deviator equation in components is
σ
ij
= 2µε
ij
(i = j),
which yields G = µ.
Linear Elasticity 199
Now we turn to Young’s modulus E. Let us consider a uniaxial homo
geneous deformation of a bar; this occurs when the material is uniformly
stretched or compressed. Take the unit vector i
1
along the bar axis. Then
σ = σ
11
i
1
i
1
. Relation (6.26) reduces to the three nontrivial component
equations
σ
11
= λtr ε + 2µε
11
, 0 = λtr ε + 2µε
22
, 0 = λtr ε + 2µε
33
.
Eliminating ε
22
and ε
33
from these, we get
σ
11
= Eε
11
, where E =
µ(3λ + 2µ)
λ + µ
.
When a bar is uniformly stretched by a force and its strain is ε, experi
ment shows that its transverse dimensions decrease. The transverse strain
for this deformation is proportional to the longitudinal strain −νε, the co
eﬃcient of proportionality being Poisson ratio. Its relation to the other
constants is deﬁned by the following
Exercise 6.3. For the above uniaxial strain state of the bar, ﬁnd ε
22
/ε
11
=
−ν. Show that ν = λ/(2λ +2µ). Note that in the uniaxially stretched bar,
ν deﬁnes the dependence of the lateral strain ε
22
on the axial strain ε
11
.
The relations between various pairs of moduli used in the literature are
summarized in Appendix A.
Thermodynamic considerations show that the strain energy should be
positive:
W(ε) > 0 whenever ε = 0. (6.27)
This puts additional restrictions on the elastic moduli for an isotropic ma
terial. We have
W(ε) =
1
2
σ ε =
1
2
λ(tr ε)
2
+ µε ε ≥ 0.
Let us represent ε as the sum of its ball and deviator terms:
ε =
1
3
Etr ε + dev ε.
Because
ε ε =
1
3
(tr ε)
2
+ dev ε dev ε,
we obtain
W(ε) =
1
6
(3λ + 2µ)(tr ε)
2
+ µdev ε dev ε.
200 Tensor Analysis with Applications in Mechanics
But tr ε and dev ε are independent quantities, and from (6.27) it follows
that
3λ + 2µ > 0, µ > 0. (6.28)
Hence the bulk modulus k = λ + 2/3µ and shear modulus G = µ are
positive. A consequence of (6.28) is explored in
Exercise 6.4. Demonstrate that E > 0 and −1 < ν ≤ 1/2.
For most engineering materials, we have ν > 0. The values of ν for
rubbers are close to 1/2; materials like reentrant foams can have negative
values of ν. In technical reference books, the reader can ﬁnd data for various
materials. For regular steels, the values of E are about 2 10
11
Pa, and ν
lies in the range 0.25 −0.33.
Exercise 6.5. For an isotropic body, express ε and W in terms of σ.
6.5 Equilibrium Equations in Displacements
We derived the equation of motion (6.19) and the equilibrium equation
(6.11). Equation (6.11) is written in terms of stresses. In component form,
(6.11) contains three equations in six independent unknowns σ
ij
. Using
the deﬁnition of strain tensor and Hooke’s law, we can transform (6.19)
and (6.11) to a system of three simultaneous equations with respect to
the components of the displacement vector u. So the systems (6.19) and
(6.11) reduce to systems involving three equations in the three unknown
components of u. This corresponds to the common viewpoint that a well
posed problem should contain the same number of equations as unknowns.
For simplicity, we derive the equilibrium equations for an isotropic ho
mogeneous material deﬁned by (6.26). Suppose λ and µ are constants.
First we derive ∇ σ in terms of u:
∇ σ = ∇ (λEtr ε + 2µε) = λ∇tr ε + µ∇ ∇u + µ∇ (∇u)
T
.
Because
tr ε = ∇ u
and
∇ (∇u)
T
= ∇∇ u,
Linear Elasticity 201
we get
∇ σ = (λ + µ)∇∇ u + µ∇ ∇u.
So the equilibrium equation takes the form
(λ + µ)∇∇ u + µ∇ ∇u + ρf = 0. (6.29)
Exercise 6.6. Show that ∇ (∇u)
T
= ∇∇ u.
In Cartesian coordinates, equation (6.29) is
(λ + µ)
∂u
i
∂x
k
∂x
i
+ µ
∂
2
u
k
∂x
i
∂x
i
+ ρf
k
= 0 (k = 1, 2, 3). (6.30)
That is,
(λ + µ)
∂
∂x
1
∂u
1
∂x
1
+
∂u
2
∂x
2
+
∂u
3
∂x
3
+ µ∆u
1
+ ρf
1
= 0,
(λ + µ)
∂
∂x
2
∂u
1
∂x
1
+
∂u
2
∂x
2
+
∂u
3
∂x
3
+ µ∆u
2
+ ρf
2
= 0,
(λ + µ)
∂
∂x
1
∂u
1
∂x
1
+
∂u
2
∂x
2
+
∂u
3
∂x
3
+ µ∆u
3
+ ρf
3
= 0, (6.31)
where
∆ =
∂
2
∂x
2
1
+
∂
2
∂x
2
2
+
∂
2
∂x
2
3
.
In curvilinear coordinates, the component representation of (6.29) is
(λ + µ)
∂
∂q
k
ε + µg
mn
∇
m
∇
n
u
k
+ ρf
k
= 0 (k = 1, 2, 3) (6.32)
where
ε = tr ε =
1
√
g
∂(
√
gg
mn
u
n
)
∂q
m
.
Exercise 6.7. Derive (6.32).
See Appendix A for the corresponding relations in cylindrical and spher
ical coordinates.
Exercise 6.8. Write down the equations of motion in displacements.
202 Tensor Analysis with Applications in Mechanics
6.6 Boundary Conditions and Boundary Value Problems
We have derived the equilibrium equations. To ensure uniqueness of so
lution, we must supplement the equations with certain conditions on the
boundary of the domain where the equations hold. We ﬁrst pose the main
equilibrium problems for elastic bodies. Then we touch on the existence
uniqueness question.
Physical intuition tells us that at a boundary point we should appoint
either a displacement vector or a stress vector deﬁned by the contact load.
Appointment of both at the same point is impossible. This will be con
ﬁrmed later. In elasticity, there are three main boundary value problems.
Two were mentioned above: one with given displacements, one with given
contact forces. The third is a mixed problem with some combination of
displacements and contact loads given on the boundary.
In the ﬁrst boundary value problem of elasticity, we supplement the equi
librium equations in displacements (6.29) or other forms of the equations
(6.30)–(6.32) with a given displacement ﬁeld u
0
on the whole boundary Σ:
u
Σ
= u
0
. (6.33)
This is known as the kinematic boundary condition.
In the second boundary value problem, we supplement (6.29) with a
prescribed contact load t
0
on Σ:
n σ
Σ
= t
0
. (6.34)
This is known as a static boundary condition. Since no point of the body
is ﬁxed, the body can move freely in space. With this condition, the static
boundary value problem is wellposed only if the load acting on the body
is selfbalanced; this means that the resultant force and resultant moment
of all the external forces acting on the body must vanish.
The reader familiar with partial diﬀerential equations may recognize
that these types of boundary conditions are analogous to the Dirichlet and
Neumann conditions for Poisson’s equation.
The third boundary value problem requires us to supplement (6.29) with
mixed boundary conditions. That is, we must assign kinematic conditions on
some portion Σ
1
of the boundary, and static conditions t
0
on the remainder
Σ
2
:
u
Σ
1
= u
0
, n σ
Σ
2
= t
0
. (6.35)
There exist other “mixed problems” of elasticity in which, at each
boundary point, we assign three conditions that combine the given dis
Linear Elasticity 203
placements and loads in some fashion. Moreover, on the boundary we can
combine conditions involving contact with other elastic or inelastic bodies
or media. As a rule, such problems require consideration of wellposedness,
although engineering intuition commonly permits investigators to propose
meaningful boundary conditions of that type. An example of a meaning
less problem is the problem with simultaneously given normal displacements
and normal stresses on the boundary. These cannot be speciﬁed indepen
dently.
Henceforth we will assume V is a bounded volume whose boundary Σ is
suﬃciently regular that we can apply the technique of integration by parts.
6.7 Equilibrium Equations in Stresses
A boundary value problem in displacements involves ﬁnding a displacement
ﬁeld satisfying the equilibrium equations (6.29) and one of the sets of bound
ary conditions (6.33), (6.34), or (6.35). It is assumed that the stress tensor
in (6.34) or (6.35) is expressed in terms of u by (6.26) or (6.16). When
the stresses are prescribed over the whole boundary, we can try ﬁnding the
unknown σ via (6.11) and (6.34). However, the solution for σ is not unique.
We must bring in other equations that take into account Hooke’s law. We
can do this as follows. Suppose we have found a solution σ of equations
(6.11) supplemented by (6.34). By σ, through Hooke’s law (6.26), we deﬁne
ε. Calculating the trace of equation (6.26), we get
tr ε =
1
3λ + 2µ
tr σ.
Substituting this into (6.26), we have
ε =
1
2µ
¸
σ −
λ
3λ + 2µ
Etr σ
=
1
2µ
¸
σ −
ν
1 + ν
Etr σ
. (6.36)
But we know that the strain tensor makes sense, i.e., by ε we can ﬁnd the
displacement ﬁeld, if and only if the compatibility equation (6.17) holds:
∇(∇ε)
T
= 0.
Substituting (6.36) into the compatibility equation, we get the equation in
terms of σ:
∇(∇ε)
T
=
1
2µ
¸
∇(∇σ)
T
−
ν
1 + ν
∇(∇(Etr σ)
T
= 0.
204 Tensor Analysis with Applications in Mechanics
Omitting cumbersome calculations, we can reduce this to the Beltrami–
Michell equations
∇ ∇σ +
ν
1 + ν
∇∇tr σ + ρ∇f + ρ(∇f )
T
+E
ν
1 −ν
ρ∇ f = 0. (6.37)
These equations, being supplementary to (6.11) and (6.34), present us with
the complete setup of the equilibrium problem in stresses. Note that this
problem has a solution only if the set of external forces, consisting of the
body forces and those acting on the boundary, is selfbalanced.
When body forces are absent, (6.37) reduces to
∇ ∇σ +
ν
1 + ν
∇∇tr σ = 0. (6.38)
In Cartesian coordinates, (6.38) implies
∆σ
11
+
1
1 + ν
∂
2
σ
∂x
2
1
= 0, ∆σ
12
+
1
1 + ν
∂
2
σ
∂x
1
∂x
2
= 0,
∆σ
22
+
1
1 + ν
∂
2
σ
∂x
2
2
= 0, ∆σ
23
+
1
1 + ν
∂
2
σ
∂x
2
∂x
3
= 0,
∆σ
33
+
1
1 + ν
∂
2
σ
∂x
2
3
= 0, ∆σ
13
+
1
1 + ν
∂
2
σ
∂x
1
∂x
3
= 0, (6.39)
where
σ = tr σ = σ
11
+ σ
22
+ σ
33
.
To solve the equilibrium problem completely, we ﬁrst ﬁnd σ by (6.11),
(6.34), and the Beltrami–Michell equations. Then, using Hooke’s law, we
ﬁnd ε. Finally, by Ces` aro’s formula (6.18), we ﬁnd the displacement ﬁeld
u that is uniquely deﬁned up to a rigid motion of the body.
We should note that, from the standpoint of classical mathematical
physics, the structure of the boundary value problem in stresses is a bit
strange. The unknowns are the six components of σ, satisfying nine equa
tions — three of ﬁrst order, and six of second order with respect to the
components of σ. These are supplemented with three boundary condi
tions. By its derivation, this boundary value problem is equivalent to the
boundary value problem in displacements. The theory of such “strange”
boundary value problems is far from complete.
Linear Elasticity 205
6.8 Uniqueness of Solution for the Boundary Value Prob
lems of Elasticity
Uniqueness of solution to a boundary value problem is an important indi
cation of its wellposedness. A uniqueness theorem for elasticity problems
was established by Gustav R. Kirchhoﬀ (1824–1887).
We will prove uniqueness of solution for equations (6.29) supplemented
with boundary conditions (6.35), which include conditions (6.33) as a par
ticular case.
To the contrary, we suppose that there exist two solutions u
1
and u
2
that satisfy (6.29) and (6.35). We denote the corresponding stress tensors
by σ
1
and σ
2
, respectively. So the following two sets of equations hold:
∇ σ
1
+ ρf = 0 in V, u
1
Σ
1
= u
0
, n σ
1
Σ
2
= t
0
,
∇ σ
2
+ ρf = 0 in V, u
2
Σ
1
= u
0
, n σ
2
Σ
2
= t
0
.
Consider the diﬀerence u = u
1
− u
2
and its corresponding stress ﬁeld
σ = σ
1
−σ
2
. Subtracting the equations for u
1
and u
2
, we see that u is a
solution of the following equilibrium problem:
∇ σ = 0 in V, u
Σ
1
= 0, n σ
Σ
2
= 0.
Dotmultiplying the equilibrium equation by u and integrating over V , we
get
V
(∇ σ) udV = 0.
Using the Gauss–Ostrogradsky theorem, we transform this to the equation
−
V
σ (∇u)
T
dV +
Σ
n σ u = 0.
At each point of Σ, one of the conditions u = 0 or n σ = 0 holds. So the
last surface integral is zero. The tensor σ is symmetric. So
σ (∇u)
T
= σ (∇u).
This allows us to represent the volume integral as follows:
−
V
σ ε dV = −2
V
W(ε) dV = 0.
Recall that W(ε) is positive deﬁnite. Hence the equality of the integral to
zero implies that its integrand vanishes everywhere in V : W(ε) = 0. By
positiveness of W, it follows that
ε = 0. (6.40)
206 Tensor Analysis with Applications in Mechanics
By Ces`aro’s formula (6.18), when ε(u) = 0, the displacement vector takes
the form
u = u
0
+ω
0
(r −r
0
), (6.41)
where u
0
, ω
0
and r
0
are constant vectors. But it is a small displacement
ﬁeld for the rigid volume V . Because u[
Σ
1
= 0, we have u = 0 in V . Thus
u
1
= u
2
.
We have proved uniqueness of solution for the ﬁrst and third boundary
value problems of elasticity.
For the second problem, the initial steps of the proof remain valid. We
ﬁnd that ε = 0 in V , hence the two solutions to the problem diﬀer by a
rigidbody displacement:
u
1
= u
2
+u
0
+ω
0
(r − r
0
).
But this time we cannot conclude that u
1
= u
2
. From physics, the situation
is clear. The second problem of elasticity describes a body free of geometric
restrictions, so its solution must be deﬁned up to a rigidbody displacement.
This is what the last formula states.
6.9 Betti’s Reciprocity Theorem
The solutions to the equilibrium problems for an elastic body under two
diﬀerent loads obey Betti’s reciprocal work theorem.
Let us consider a body under the action of contact forces t
and body
forces f
, and the same body under the action of another pair of external
forces t
and f
. The displacements and other quantities for the two cor
responding problems will be denoted similarly, using primes and double
primes, respectively. Now we have two solutions of two diﬀerent second
boundary value problems for the elastic body.
Theorem 6.4. The solutions u
and u
of the second boundary value prob
lem of elasticity that correspond to the respective loading pairs t
, f
and
t
, f
satisfy the following relation:
V
ρf
u
dV +
Σ
t
u
dΣ =
V
ρf
u
dV +
Σ
t
u
dΣ. (6.42)
This equality is Betti’s theorem. First derived for beam theory and later
extended to many portions of physics, it is used to derive the equations of
Linear Elasticity 207
the boundary ﬁnite element methods. Each of the expressions on the left
and right represents the work of a load over a displacement. So we can give
a mechanical formulation of Betti’s theorem.
The work of the ﬁrst system of forces over the displacements of the body
due to the action of the second system of forces is equal to the work of the
second system of forces over the displacements due to the action of the ﬁrst
system of forces.
Proof. The vectors u
and u
respectively satisfy the following equilib
rium equations and boundary conditions:
∇ σ
+ ρf
= 0 in V, n σ
Σ
= t
;
∇ σ
+ ρf
= 0 in V, n σ
Σ
= t
.
Let us consider the lefthand side of (6.42). Writing the forces in terms of
the stresses, we get
V
ρf
u
dV +
Σ
t
u
dΣ = −
V
(∇ σ
) u
dV +
Σ
n σ
u
dΣ.
Applying the Gauss–Ostrogradsky theorem to the surface integral, we get
−
V
(∇ σ
) u
dV +
Σ
n σ
u
dΣ
=
V
[−(∇ σ
) u
+∇ (σ
u
)] dV
=
V
σ
(∇u
)
T
dV
=
V
σ
ε
dV.
Similarly,
V
ρf
u
dV +
Σ
t
u
dΣ =
V
σ
ε
dV.
So the proof reduces to veriﬁcation of the equality
V
σ
ε
dV =
V
σ
ε
dV.
Using Hooke’s law and recalling the symmetry properties of C, we have
σ
ε
= (C ε
) ε
= ε
C ε
= ε
C ε
= σ
ε
.
This completes the proof.
208 Tensor Analysis with Applications in Mechanics
Exercise 6.9. Let a portion Σ
1
of the boundary be ﬁxed: u[
Σ
1
= 0. Sup
pose two boundary value problems include this condition and, moreover,
that the body is under the action of one of the two systems of forces t
, f
or t
, f
, with t
, t
given on Σ
2
= Σ ` Σ
1
. Prove that in this case Betti’s
equality takes the form
V
ρf
u
dV +
Σ
2
t
u
dΣ =
V
ρf
u
dV +
Σ
2
t
u
dΣ.
Exercise 6.10. Consider two equilibrium problems for a body under two
systems of external forces, as in the previous exercise. This time, however,
two distinct displacement ﬁelds are prescribed over Σ
1
:
u
Σ
1
= a
, u
Σ
1
= a
. (6.43)
As above, denote the solutions of the corresponding equilibrium problems
by u
and u
. Denote the solutions of the equilibrium problems for a body
free of external forces, that satisfy the corresponding conditions (6.43), by
u
0
and u
0
, respectively. Finally, introduce
˜ u
= u
−u
0
, ˜ u
= u
−u
0
.
Prove that Betti’s equality takes the form
V
ρf
˜ u
dV +
Σ
2
t
˜ u
dΣ =
V
ρf
˜ u
dV +
Σ
2
t
˜ u
dΣ.
6.10 Minimum Total Energy Principle
On a curved surface, a ball in equilibrium takes the lowest position; we call
the equilibrium stable. In elementary physics, it is said that at such an equi
librium point, the potential energy of the ball takes a minimum value. A
stationary point of potential energy corresponds to an equilibrium position
of the ball as well, but such an equilibrium may be unstable. An equilib
rium that corresponds to a nonminimum point normally is unstable. The
stability result for the potential energy of a particle was ﬁrst extended to
classical mechanics by Lagrange. The minimum potential energy principle
is now one of the most important principles of physics, holding also for
distributed systems like elastic bodies. However, it is not always straight
forward to formulate, since we must ﬁnd an expression for the potential
energy and prove that it takes a minimum value in equilibrium.
Linear Elasticity 209
As the expression for the energy of an elastic body under load, we
propose the total energy
E =
V
W(ε) dV −
V
ρf udV −
Σ
2
t
0
udΣ.
The ﬁrst term is the strain energy associated with deformation. If we slowly
decrease the external forces acting on the body to zero, the strain energy
is the maximum work that the body can produce. The other terms, with
negative signs, represent the work of external forces on the displacement
ﬁeld of the body. These terms are analogous to the expression for the work
of the gravitational force acting on a moving particle, and are so similar to
the gravitational potential that they may be called potential energy terms.
Again, we must prove that E really takes its minimum value when the
body is in equilibrium. Clearly, we should explain what is meant by “the
minimum value of E” and introduce the tools necessary to establish the
minimum principle.
The reader is familiar with the deﬁnition of a local minimum point for
a function in n variables: there should exist a neighborhood of the point
in which all the values of the function are no less than its value at the
point. For a global minimum point, this value should be no greater than
the values of the function at any point. These deﬁnitions can be extended
to quantities like E, as they take values in the set of real numbers. We
could call E a function, but because it depends on the vector function u,
it is called a functional. A functional is a correspondence that takes its
argument, which can be a whole function as it is here, to at most one real
number.
To extend the idea of minimum to functionals, we should treat the
notions of independent variable, the domain of such a variable, and the
neighborhood of a “point” (which is now a vector function u in the domain).
For simplicity, we will assume the domain of E consists of vector functions
possessing all continuous derivatives in V up to the second order. Moreover,
we suppose that for any u, the kinematic boundary conditions hold. Such
vector functions will be called admissible. To deﬁne a neighborhood, we
must introduce a metric or norm on the set of admissible vector functions.
The reader interested in the formalities should consult standard books on
the calculus of variations. Our present approach will be essentially that of
the pioneers of the subject, which was to obtain results without worrying
too much about formal justiﬁcation. In fact, we can completely avoid the
question about neighborhoods, since for the problems under consideration,
210 Tensor Analysis with Applications in Mechanics
in the equilibrium state E takes its global minimum value.
Now we will describe the technical tools that permit us to seek the
minimum points of a functional. We wish to derive for E some analogue
of Fermat’s theorem for a diﬀerentiable function in one variable: the ﬁrst
derivative must vanish at a minimum point. Fermat’s theorem extends to
the theory of functions in many variables: all the ﬁrst partial derivatives of
the function must vanish. Let us review how this extension is accomplished.
Let f be an ordinary realvalued function of a vector variable, and sup
pose f takes its minimum at a point x ∈ R
n
. If a is any vector and τ is
a real parameter, then f(x + τa) can be regarded as a function of τ that
takes its minimum value at τ = 0. Therefore, at τ = 0 its derivative with
respect to τ must vanish:
df(x + τa)
dτ
τ=0
= 0. (6.44)
At a minimum point x, this must hold for any a. As the reader is aware,
the expression on the left is the directional derivative of f at x along the
direction of a. Moreover, if we formally set a = dx, then the left member
takes the form of the ﬁrst diﬀerential df at point x:
df =
df(x + τ dx)
dτ
τ=0
= ∇f
x
dx.
Putting a = i
k
in (6.44), we obtain Fermat’s theorem for a function in n
variables: the partial derivative of f with respect to x
k
at x must vanish
if ∇f[
x
exists and x is a minimum point of f. We emphasize that this
is merely a necessary condition. If at some point x equation (6.44) holds,
then x may not be a minimum (or a maximum) point. Any x satisfying
(6.44) is called a stationary point of f. It could be a minimum point, a
maximum point, or a saddle point.
This procedure extends to functionals in a straightforward manner. Sup
pose u is a minimum point of E. Let us ﬁx some δu such that u + τ δu is
an admissible displacement for all small τ. Substitute this into E. Then
E(u+τ δu) is a function of the real variable τ that takes its minimum value
at τ = 0. If the derivative of this function with respect to τ exists, then it
vanishes at τ = 0:
δE =
d
dτ
E(u + τ δu)
τ=0
= 0
for any δu such that u + τ δu is admissible for any small τ.
Linear Elasticity 211
We call δE the ﬁrst variation of the functional E. Its zeros are called
stationary points of E; they are analogous to the stationary points of an
ordinary function f. The equation we obtained is a necessary condition for
u to be a minimum point of E.
We assumed the displacements u + τ δu were admissible for all small
τ. Clearly, if all derivatives of u and δu up to order two are continuous in
V , the same is true of u + τ δu. We required u to satisfy the kinematic
restrictions on Σ
1
. Hence if δu satisﬁes
δu
Σ
1
= 0,
then u+τ δu also satisﬁes the kinematic restrictions for any τ. In this case
δu is called a virtual displacement.
We note that the deﬁnition of δE can be applied to any suﬃciently
smooth E. (We leave this on an intuitive level, as a formal deﬁnition
would require a background in functional analysis.) It is called the Gˆ ateaux
derivative of the functional. In the notation δE, we regard the symbol δ
as some action on E given by a formula similar to the one used to ﬁnd
the diﬀerential of a function. By δu we denote a virtual displacement; this
could be denoted by v as well, but we would have to keep clarifying that
it is a virtual displacement. That is, in this case the symbol δ is merely a
notation; it is not an operation and is used for historical reasons. When
we apply the operation δ to an ordinary function, it coincides with the
operation of taking its ﬁrst diﬀerential. The diﬀerentials of variables in this
case are written with the use of the symbol δ instead of d; for example, we
have δ(x
2
) = 2xδx.
Using the deﬁnition, let us ﬁnd δE for our problem:
δE =
V
δW(ε) dV −
V
ρf δudV −
Σ
2
t
0
δudΣ.
Calculating δW in the integrand and recalling that for intermediate steps
we can use the formulas for the ﬁrst diﬀerential, we get
δW(ε) =
1
2
δ(ε C ε) = ε C δε = σ δε,
where
δε = ε(δu) =
1
2
∇δu + (∇δu)
T
.
It follows that
δE =
V
σ δε dV −
V
ρf δudV −
Σ
2
t
0
δudΣ.
212 Tensor Analysis with Applications in Mechanics
Now we wish to ﬁnd a relationship between a solution to the equilibrium
problem for an elastic body and the minimum problem for E. We start with
the following theorem.
Theorem 6.5. A stationary point u of E on the set of admissible displace
ments satisﬁes the equilibrium equations of the elastic body in the volume
V and the boundary condition n σ[
Σ
2
= t
0
, and conversely.
Proof. First we prove that u, a solution of the equilibrium problem for
an elastic body under load, is a stationary point of E; that is,
δE = 0
when u is a solution and δu is an arbitrary virtual displacement. Assume
u satisﬁes
∇ σ + ρf = 0 in V, u
Σ
1
= u
0
, n σ
Σ
2
= t
0
. (6.45)
Dotmultiply the equilibrium equation by an admissible δu and integrate
over V . Then apply the Gauss–Ostrogradsky theorem:
0 =
V
[(∇ σ) δu + ρf δu] dV
=
V
−σ (∇δu)
T
+ ρf δu
dV +
Σ
n σ δudΣ
=
V
[−σ δε + ρf δu] dV +
Σ
2
t
0
δudΣ
= −δE.
We used the fact that δu = 0 on Σ
1
and n σ = t
0
on Σ
2
. Hence u is a
stationary point of E.
Now we prove the converse statement. Let u be a stationary point of E,
so that it satisﬁes the equation δE = 0 for any virtual δu being suﬃciently
smooth, and let u[
Σ
1
= u
0
. We will show that u satisﬁes the equilibrium
equations in V as well as the contact condition n σ = t
0
on Σ
2
. Indeed,
Linear Elasticity 213
doing the above calculations in reverse order, we get
δE =
V
σ δε dV −
V
ρf δudV −
Σ
2
t
0
δudΣ
=
V
σ (∇δu)
T
dV −
V
ρf δudV −
Σ
2
t
0
δudΣ
= −
V
[(∇ σ) δu + ρf δu] dV +
Σ
n σ δudΣ −
Σ
2
t
0
δudΣ
= −
V
[∇ σ + ρf ] δudV
+
Σ
1
n σ δudΣ −
Σ
2
n σ −t
0
δudΣ.
Because δu = 0 on Σ
1
, we ﬁnally obtain
δE = −
V
[∇ σ + ρf ] δudV −
Σ
2
n σ −t
0
δudΣ = 0. (6.46)
We assumed u to be smooth enough that the above integrands are contin
uous. As admissible δu is arbitrary, from this equality it follows that
∇ σ + ρf = 0 in V, n σ
Σ
2
= t
0
. (6.47)
The justiﬁcation of this last step requires standard material from the calcu
lus of variations. We postpone this discussion until the end of the present
section.
Now we would like to prove
Theorem 6.6. Let u be a solution of the boundary problem (6.45). Then
it is a point of global minimum of E.
Proof. The statement is a consequence of the fact that W is a positive
quadratic form with respect to ε. Indeed, let v be an arbitrary admissi
ble displacement so that it is smooth enough and satisﬁes the kinematic
condition v[
Σ
1
= u
0
. Consider the diﬀerence
∆E = E(v) − E(u).
214 Tensor Analysis with Applications in Mechanics
We have
∆E =
V
W(ε(v)) dV −
V
ρf v dV −
Σ
2
t
0
vdΣ
−
V
W(ε(u)) dV +
V
ρf udV +
Σ
2
t
0
udΣ
=
V
[W(ε(v)) −W(ε(u))] dV
−
V
ρf (v −u) dV −
Σ
2
t
0
(v −u) dΣ.
Let w = v −u. Because u and v coincide on Σ
1
, we have w[
Σ
1
= 0. Next,
2 [W(ε(v)) −W(ε(u))] = ε(v) C ε(v) −ε(u) C ε(u)
= ε(w) C ε(w) + 2ε(u) C ε(w).
Therefore
∆E =
1
2
V
ε(w) C ε(w) dV
+
V
ε(u) C ε(w) dV −
V
ρf wdV −
Σ
2
t
0
wdΣ.
The sum of the terms in the second line constitutes δE at u with δu = w,
which is a virtual displacement, so this sum is zero. Thus we have
∆E =
1
2
V
ε(w) C ε(w) dV =
V
W(ε(w)) dV. (6.48)
As W is a positive deﬁnite form, ∆E ≥ 0 for any admissible v. Hence u is
a global minimizer of E.
The two preceding theorems can be summarized as Lagrange’s varia
tional principle.
Theorem 6.7. The solution of the equilibrium problem for an elastic body
is equivalent to the problem of minimizing the total energy functional E
over the set of all kinematically admissible smooth displacement ﬁelds.
To complete the proof of Theorem 6.5, we must show how to obtain
(6.47) from (6.46). It will suﬃce to do this for one of the component
equations. Denoting the kth component of the expressions in the brackets
in (6.46) by F and f respectively, the kth component of δu by δu, and
Linear Elasticity 215
setting the remaining components of δu to zero, we reduce equation (6.46)
to
V
F δu dV +
Σ
2
f δu dΣ = 0. (6.49)
Equation (6.47) follows from the next theorem.
Theorem 6.8. Let F be continuous in V and let f be continuous on Σ
2
.
Suppose (6.49) holds for all suﬃciently smooth functions δu that vanish on
Σ
1
. Then
F = 0 in V, f = 0 on Σ
2
.
Proof. The proof is done in two steps. We begin by restricting the set
of all admissible δu to those that vanish on the whole boundary Σ. Then
(6.49) becomes
V
F δu dV = 0.
Suppose, contrary to the theorem statement, that F(x
∗
) = a = 0 at some
point x
∗
. Without loss of generality we take x
∗
to be an interior point of
V and a > 0. By continuity of F there is an open ball B
r
with center x
∗
and radius r > 0 that lies inside V such that F(x) > a/2 > 0 on B
r
. Now
if we take a smooth function δu
0
that is positive on B
r
and zero outside
B
r
, we get
V
F δu
0
dV > 0,
which contradicts the above equality that must hold for all δu that vanish
on the boundary of V . It is clear that such a function δu
0
exists; the reader
can ﬁnd examples of “bellshaped functions” in textbooks on the calculus
of variations, cf.
[
Lebedev and Cloud (2003)
]
. So F = 0 in V .
The second step is to prove that f = 0 on Σ
2
. For this we return to
(6.49). Because F = 0, the ﬁrst integral is zero for any admissible δu. So
we ﬁnd that
Σ
2
f δu dΣ = 0
holds for all admissible δu that take arbitrary values on Σ
2
. But now we
are in a position similar to the proof for F. We can repeat that proof, but
on a twodimensional domain Σ
2
, so f = 0 on Σ
2
.
216 Tensor Analysis with Applications in Mechanics
A reader familiar with the calculus of variations may recognize the above
procedure as leading to one version of the “main lemma” of that subject.
It is worth noting that in the theory of elasticity there are several varia
tional principles. Some are of minimax type; others are of stationary type.
We can mention principles that carry the names of Castigliano, Reissner,
Tonti, Hamilton, etc.
6.11 Ritz’s Method
The minimum total energy principle is of great physical importance. It is
also the basis for introducing the generalized setup of the equilibrium prob
lems, which in turn deﬁnes weak solutions in elasticity. These solutions have
ﬁnite strain energies. The principle is even more important from a practical
standpoint. The variational equation δE = 0 is the basis for various nu
merical methods, including the Ritz and Galerkin methods, the variational
ﬁnite diﬀerence methods, and the ﬁnite element methods. Minimality of
the total energy on the solution warrants stability of the algorithms of these
methods. For details, the reader should consult specialized literature. Here
we present Ritz’s method, which spawned the other methods mentioned
above.
The minimum total energy principle states that a solution of the equi
librium problem
∇ σ + ρf = 0 in V, u
Σ
1
= u
0
, n σ
Σ
2
= t
0
,
expressed in terms of u, is a point of global minimum of the total energy
functional
E =
V
W(ε) dV −
V
ρf udV −
Σ
2
t
0
udΣ
on the set of admissible displacements. The converse statement is also valid,
hence by the uniqueness theorem this point of minimum is unique. Walter
Ritz (1878–1909) thought that by ﬁnding a minimum point on some subset
of admissible displacements, we can get an approximation to the solution.
The decisive idea was how to select this subset in such a way that solving
the approximate minimum problem would be relatively easy. Ritz proposed
to minimize E on the set of displacements of the form
u
N
=
N
¸
k=1
u
k
ϕ
k
+u
∗
, u
∗
Σ
1
= u
0
, ϕ
n
Σ
1
= 0 (6.50)
Linear Elasticity 217
with some ﬁxed ϕ
k
and u
∗
, by determining the numerical coeﬃcients u
k
.
The ﬁrst step is to ﬁnd a suﬃciently smooth vector function u
∗
that satisﬁes
the condition u
∗
[
Σ
1
= u
0
. Then we choose some set of basis elements
ϕ
k
that vanish on Σ
1
. Clearly, any u
N
satisﬁes the kinematic condition
u
N
[
Σ
1
= u
0
.
In Ritz’s time, when calculations where done manually, the selection
of the basis was extremely important; an engineer needed to ﬁnd a few
coeﬃcients u
k
in a reasonable time. Now, with powerful computers, the
number of basis elements can be large. One requirement for the basis
elements is that they constitute a linearly independent set. Another is that
using this basis, we can actually approximate a solution with satisfactory
precision. We shall not pursue this issue, as it requires a serious excursion
into mathematics.
Let us formulate Ritz’s minimization problem for the Nth approxima
tion:
Minimize E over the set of all u
N
satisfying (6.50).
That is, ﬁnd
˜ u
N
=
N
¸
k=1
˜ u
k
ϕ
k
+u
∗
such that
E(˜ u
N
) ≤ E(u
N
) for any u
N
from the set (6.50),
where
E(u
N
) =
V
W(ε(u
N
)) dV −
V
ρf u
N
dV −
Σ
2
t
0
u
N
dΣ.
Because u
∗
is ﬁxed, we should minimize E(u
N
) on the Ndimensional linear
space spanned by the elements ϕ
1
, . . . , ϕ
N
. Hence we must ﬁnd the values
of the real coeﬃcients ˜ u
1
, . . . , ˜ u
N
. From the minimum theorem on the whole
set of admissible displacements, it follows that E(u
N
) has a minimum point
in the spanned space.
Now we need a practical way of ﬁnding the coeﬃcients of the minimizer
of E. Let us write out the expression for E(u
N
) in detail. We introduce
the following notation:
'u, v` =
V
ε(u) C ε(v) dV. (6.51)
218 Tensor Analysis with Applications in Mechanics
Note that 'u, v` = 'v, u`, and that 'u, u` is twice the strain energy for
the displacement ﬁeld u. This form 'u, v` can be considered as an inner
product on any linear space of suﬃciently smooth u such that u[
Σ
1
= 0,
for which the expression 'u, v` makes sense.
Next we introduce
f
m
=
V
ρf ϕ
m
dV +
Σ
2
t
0
ϕ
m
dΣ −'u
∗
, ϕ
m
`
and
F =
V
ρf u
∗
dV +
Σ
2
t
0
u
∗
dΣ −
1
2
'u
∗
, u
∗
`.
In this notation we have
E(u
N
) =
1
2
N
¸
m=1
N
¸
n=1
u
m
u
n
'ϕ
m
, ϕ
n
` −
N
¸
n=1
f
n
u
n
−F.
So E(u
N
) is a quadratic function in the N variables u
n
, and we can apply
standard tools from calculus.
At the point of minimum of E(u
N
), the simultaneous equations
∂E(u
N
)
∂u
m
= 0 (m = 1, . . . , N)
must hold. Writing these as
N
¸
n=1
u
n
'ϕ
n
, ϕ
m
` = f
m
(m = 1, . . . , N), (6.52)
we have a system of linear algebraic equations. The matrix of the system
is
A =
⎛
⎜
⎜
⎜
⎜
⎜
⎝
'ϕ
1
, ϕ
1
` 'ϕ
2
, ϕ
1
` 'ϕ
3
, ϕ
1
` . . . 'ϕ
N
, ϕ
1
`
'ϕ
1
, ϕ
2
` 'ϕ
2
, ϕ
2
` 'ϕ
3
, ϕ
2
` . . . 'ϕ
N
, ϕ
2
`
'ϕ
1
, ϕ
3
` 'ϕ
2
, ϕ
3
` 'ϕ
3
, ϕ
3
` . . . 'ϕ
N
, ϕ
3
`
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
'ϕ
1
, ϕ
N
` 'ϕ
2
, ϕ
N
` 'ϕ
3
, ϕ
N
` . . . 'ϕ
N
, ϕ
N
`
⎞
⎟
⎟
⎟
⎟
⎟
⎠
.
Its determinant det(A) is Gram’s determinant. In linear algebra, it is
shown that for a linearly independent system of elements the Gram deter
minant is nonzero and conversely. But the elements ϕ
1
, . . . , ϕ
N
are linearly
independent by assumption, hence (6.52) has a unique solution.
Now let us touch on the question of convergence for Ritz’s approxima
tions. Our use of the term “approximations” does not, by itself, answer the
question whether u
N
is really close to the solution.
Linear Elasticity 219
We shall reformulate the problem. Let u be a solution to the equilib
rium problem under consideration. Then the Nth approximation of Ritz’s
method minimizes the functional E(u
N
)−E(u) as well. But by the formula
(6.48) in the proof of Theorem 6.6, we have
E(u
N
) −E(u) =
V
W(ε(u
N
−u)) dV. (6.53)
Because W is a positive deﬁnite form, the minimization procedure seems
to be convergent if we can approximate any admissible displacement u
with u
N
. The sense in which we should approximate the displacements is
given by the above limit E(u
N
) − E(u) → 0 as N → ∞. So to get an
approximation, it suﬃces to have the set ϕ
m
be complete in the sense that
for any admissible v such that v[
Σ
1
= 0 and ε > 0, we can ﬁnd a ˜ u
n
such
that
V
W(ε(˜ u
n
− v)) dV < ε.
In practice we can ﬁnd complete systems, but proof of completeness is not
easy.
Having a complete basis set of ϕ
m
, it seems we could say the following.
Because the Nth Ritz approximation is the best approximation of the so
lution from all the possible approximations, by completeness of the ϕ
m
we
immediately obtain convergence of the Ritz approximations u
N
to u in the
sense that
V
W(ε(u
N
−u)) dV →0 as N →∞.
This argument is, unfortunately, merely plausible.
We can show that for Ritz’s approximations the following holds:
V
W(ε(u
N
) −ε(u
M
)) dV →0 whenever N, M →∞. (6.54)
So ¦u
N
¦ seems to be a Cauchy sequence in the integral energy sense. Un
fortunately, using only classical calculus we cannot conclude that a limit
element exists and determine its properties. These issues are discussed in
the theory of Sobolev spaces. Engineers who use Ritz’s method should
understand that they work with smooth approximations that are ordinary
functions, but that the limit functions and convergence questions are func
tional analytic issues; see, for example,
[
Lebedev and Cloud (2003)
]
.
Often engineers attempt to compensate for a lack of theoretical justiﬁ
cation by means of experiment. However, experiments are quite restrictive
220 Tensor Analysis with Applications in Mechanics
and cannot provide real justiﬁcation in all practical cases. They can only
support opinion regarding the applicability of a method.
Korn’s inequality
For various qualitative questions about solutions of elasticity problems,
Korn’s inequality plays a signiﬁcant role. It states that a displacement
with ﬁnite energy belongs to the space W
1,2
(V ); that is, all of its Cartesian
components have squareintegrable ﬁrst derivatives. It can also yield an
estimate for the displacement ﬁeld through the expression for the strain
energy.
One form of Korn’s inequality is
V
(u u +∇u ∇u
T
) dV ≤ c
V
W(ε(u)) dV (6.55)
for any suﬃciently smooth u with a constant c that depends only on V and
Σ
1
, the part of the boundary on which u[
Σ
1
= 0. Because W(ε(u)) is a
positive deﬁnite quadratic form in ε and we are not interested in the exact
value of c, we can change W(ε(u)) to tr(ε ε). A general proof for Σ
1
= Σ
is diﬃcult
[
Ciarlet (1988)
]
, so we establish the inequality when u[
Σ
= 0.
We will prove
V
tr(∇u ∇u
T
) dV ≤ 2
V
tr(ε ε) dV, (6.56)
from which (6.55) follows. In Cartesian coordinates this is
V
3
¸
i,j=1
∂u
i
∂u
j
2
dV ≤ 2
V
⎛
⎝
3
¸
i,j=1
ε
2
ij
⎞
⎠
dV.
Another part of the inequality for u
i
follows from Friedrich’s inequality
V
3
¸
i=1
[u
i
[
2
dV ≤ c
1
V
3
¸
i,j=1
∂u
i
∂u
j
2
dV,
which holds with a constant c
1
independent of u
i
for any smooth function
u
i
that vanishes on the boundary. See
[
Lebedev and Cloud (2003)
]
for a
proof.
Linear Elasticity 221
So we prove (6.56). We have
V
tr(ε ε) dV =
1
2
V
tr(∇u (∇u)
T
) + tr(∇u ∇u)
dV
=
1
2
∇u
2
+
1
2
V
tr(∇u ∇u) dV.
On the boundary, u = 0. Using the Gauss–Ostrogradsky theorem, let us
transform the last integral:
V
tr(∇u ∇u) dV =
V
∂u
j
∂x
i
∂u
i
∂x
j
dV = −
V
u
j
∂
2
u
i
∂x
i
∂x
j
dV
=
V
∂u
i
∂x
i
∂u
j
∂x
j
dV =
V
(∇ u)
2
dV.
So
2 ε
2
= ∇u
2
+
V
(∇ u)
2
dV.
Inequality (6.56) follows from the positivity of the second term.
6.12 Rayleigh’s Variational Principle
The minimum total energy principle, also known as Lagrange’s principle,
is formulated for equilibrium problems; it cannot be applied to dynamics
problems. But there is one important dynamics problem for which there
exists a variational principle based on a similar minimization idea. This is
the eigenoscillation problem for an elastic body. In this problem we seek
solutions to a dynamic homogeneous problem in displacements in the form
u = u(r, t) = w(r)e
iωt
. (6.57)
The equations of the dynamical problem are
∇ σ = ρ¨ u in V, u
Σ
1
= 0, n σ
Σ
2
= 0. (6.58)
Substituting (6.57) into (6.58) expressed in displacements, and canceling
the factor e
iωt
, we get
∇ σ = −ρω
2
w in V, w
Σ
1
= 0, n σ
Σ
2
= 0. (6.59)
Here σ is given by
σ = C ε, ε = ε(w) =
1
2
∇w+ (∇w)
T
.
222 Tensor Analysis with Applications in Mechanics
Equations (6.59) constitute an eigenvalue problem: we must ﬁnd pos
itive values ω, known as eigenfrequencies of the elastic body, for which
(6.59) has a nontrivial solution w, called an eigenoscillation.
It can be shown that the problem has only nonnegative eigenfrequencies,
for which whas only real components, and that the set of eigenfrequencies is
countable. To demonstrate that the set of eigenfrequencies is countable, and
to show that a complete set of linearly independent eigenmodes constitutes
a basis in the space of modes having ﬁnite energy, we need techniques
that fall outside the scope of this book. Note that these proofs require
V to be a compact volume with a suﬃciently smooth boundary. The term
“suﬃciently smooth” covers the needs of engineering practice: the boundary
cannot have cusps, but V can be a pyramid or a cone.
Now we show that any eigenfrequency ω is real and, moreover, nonnega
tive. Suppose to the contrary ω is a complex number so that its correspond
ing mode w is also complexvalued. Dotmultiplying the ﬁrst equation of
(6.59) by w and integrating over V , we get
V
(∇ σ) wdV = −ω
2
V
ρw wdV.
Applying the Gauss–Ostrogradsky theorem to the left side, we get
V
σ ε dV = ω
2
V
ρw wdV. (6.60)
Because σ ε ≥ 0 and ρw w ≥ 0, from (6.60) it follows that ω
2
is a
real nonnegative number, hence so is ω. As the eigenfrequency equation
is linear with realvalued coeﬃcients, the real and imaginary parts of an
eigensolution are eigensolutions as well. This means that we can consider
only real eigenmodes.
Engineers are mostly interested in some range of eigenfrequencies, say
the lowest few, or those belonging to some ﬁnite range.
Let us note that for the ﬁrst and third boundary value problems, the
minimum eigenfrequency is positive. Indeed, for ω = 0 the problem (6.59)
is described by the equilibrium equations, and by uniqueness we have only
the zero solution w = 0. For the second boundary value problem, where
Σ
2
= Σ, to the value ω = 0 there corresponds a nontrivial solution that
represents a displacement of the body as a rigid whole:
w = w
0
+ω
0
(r −r
0
)
with arbitrary but constant vectors w
0
, ω
0
, and r
0
. We restrict ourselves
to the case of positive ω.
Linear Elasticity 223
By linearity of the problem, if w is an eigensolution, then so is aw for
any scalar a. We typically choose a so that aw has unit L
2
norm. Such an
eigensolution is called an oscillation mode.
We prove the following theorem.
Theorem 6.9. For oscillation eigenmodes w
1
and w
2
corresponding to
distinct eigenfrequencies ω
1
and ω
2
respectively, the relation
V
ρw
1
w
2
dV = 0 (6.61)
holds. Moreover,
'w
1
, w
2
` = 0. (6.62)
Equality (6.61) is the orthogonality relation, and (6.62) is the general
ized orthogonality relation between w
1
and w
2
.
Proof. The proof mimics that of Theorem 3.2. Let w
1
and w
2
satisfy
∇ σ
1
= −ρω
2
1
w
1
in V, w
1
Σ
1
= 0, n σ
1
Σ
2
= 0,
and
∇ σ
2
= −ρω
2
2
w
2
in V, w
2
Σ
1
= 0, n σ
2
Σ
2
= 0,
where σ
k
= C ε
k
and ε
k
= ε(w
k
). Dotmultiply the ﬁrst equation in V by
w
2
and integrate this over V . Applying the Gauss–Ostrogradsky theorem,
we get
−
V
σ
1
ε
2
dV + ω
2
1
V
ρw
1
w
2
dV = 0. (6.63)
Similarly,
−
V
σ
2
ε
1
dV + ω
2
2
V
ρw
2
w
1
dV = 0.
Subtracting these two integral equalities and using σ
1
ε
2
= σ
2
ε
1
, we
get
(ω
2
1
−ω
2
2
)
V
ρw
1
w
2
dV = 0.
Relation (6.61) follows from the fact that ω
1
= ω
2
. Substituting this into
the above equality, we obtain (6.62).
224 Tensor Analysis with Applications in Mechanics
Mathematically, Theorem 6.9 states that the oscillation eigenmodes cor
responding to distinct eigenfrequencies are orthogonal in the space L
2
(V )
with weight ρ. From (6.63) it follows that they are also orthogonal with re
spect to the energy inner product ', `. If to some ω there correspond a few
linearly independent eigenmodes, we can always construct an orthonormal
set of corresponding eigenmodes using the Gram–Schmidt procedure. See
[
Yosida (1980)
]
.
This orthogonality is the basis for practical application of the following
variational Rayleigh principle.
Theorem 6.10. Eigenmodes are stationary points of the energy functional
E
0
(w) =
1
2
V
W(ε(w)) dV
on the set of displacements satisfying the boundary conditions w[
Σ
1
= 0
and subject to the constraint
1
2
V
ρw wdV = 1. (6.64)
Conversely, all the stationary points of E
0
(w) on the above set of displace
ments are eigenmodes of the body that correspond to its eigenfrequencies.
Proof. Let us write out the stationarity condition for E
0
(w). Using the
same reasoning as in the proof of Theorem 6.3, we get
δE
0
= −
V
(∇ σ) δwdV −
Σ
2
n σ δwdΣ = 0. (6.65)
This should be considered on the set of displacements with the restrictions
described in the theorem statement. So δw is not independent as it was in
Theorem 6.3, but must satisfy the condition
δ
1
2
V
ρw wdV −1
=
V
ρw δwdV = 0. (6.66)
In calculus, the minimization of a function with constraints is solved
using Lagrange multipliers. We will adapt this method to our problem for
a functional. To E
0
(w) let us add
−λ
1
2
V
ρw wdV −1
where λ is an indeﬁnite multiplier (we shall call it a Lagrange multiplier).
Linear Elasticity 225
Let us introduce a functional
˜
E
0
(w, λ) depending on the variables w
and λ, as
˜
E
0
(w, λ) = E
0
(w) −λ
1
2
V
ρw wdV −1
.
Now w is not subjected to the constraint. We will show that w from a
stationary point (w, λ) of
˜
E
0
is a stationary point of E
0
(w) under the
constraint (6.64). Indeed, δ
˜
E
0
= 0 is
δ
˜
E
0
= δE
0
−λ
V
ρw δwdV −(δλ)
1
2
V
ρw wdV −1
= 0,
where δE
0
is given by (6.65). Because w and λ are independent, δ
˜
E
0
= 0
implies two simultaneous equations:
δE
0
−λ
V
ρw δwdV = 0,
1
2
V
ρw wdV −1 = 0. (6.67)
This shows the required property possessed by stationary points of
˜
E
0
.
So we consider a stationary point
˜
E
0
(w, λ) without constraints on w.
From the ﬁrst equation in (6.67) we get
−
V
[(∇ σ) δw+ λρw δw] dV −
Σ
2
n σ δwdΣ = 0. (6.68)
As in the derivation of the minimum total energy principle, it follows from
this integral equality for arbitrary δw that
∇ σ = −ρλw, n σ
Σ
2
= 0.
If we change λ to ω
2
, this equation coincides with (6.59). This explains the
meaning of the Lagrange multiplier: it is equal to the squared eigenvalue ω
whose nonnegativeness was proved above. Thus the stationarity condition
for E
0
(w) is valid on eigensolutions of the problem (6.59).
Now we prove the converse. Let w be a solution of the equation
∇ σ = −ρω
2
w
in V for some ω. We dotmultiply it by v, integrate over V , and apply the
Gauss–Ostrogradsky theorem to get
0 =
V
(∇ σ) v + ω
2
ρw v
dV
=
V
−σ ε(v) + ω
2
ρw v
dV +
Σ
2
n σ vdΣ, (6.69)
226 Tensor Analysis with Applications in Mechanics
where
ε(v) =
1
2
(∇v + (∇v)
T
).
Let us select v such that
V
ρw vdV = 0.
Denoting v = δw, we transform the second line of (6.69) to −δE
0
from
(6.65). Thus δE
0
, which is calculated at the eigenmode, is equal to zero.
This completes the proof.
We will reformulate Rayleigh’s principle for ease of application. In the
new formulation, one need not stipulate separate integral restrictions on
the set of w.
Theorem 6.11. On the set of admissible vectorfunctions satisfying the
condition w[
Σ
1
= 0, the oscillation eigenmodes are stationary points of the
functional
R(w) =
E
0
K
, where K =
1
2
V
ρw wdV.
R(w) is called Rayleigh’s quotient. Conversely, a normalized stationary
point of R(w) is an eigenmode that corresponds to some eigenfrequency;
the value of R(w) on an eigenmode is the squared eigenfrequency.
Proof. The proof of the ﬁrst part follows from the previous proof. Indeed,
R does not change when we multiply w by a constant factor. Selecting such
a factor that
1
2
V
ρw wdV = 1,
we see that R = E
0
.
Now we show that for the eigenmode w corresponding to the eigenfre
quency ω, we have
R(w) = ω
2
. (6.70)
We dotmultiply (6.59) by w, integrate over V , and apply the Gauss–
Ostrogradsky theorem. We get
V
σ ε dV = ω
2
V
ρw wdV,
from which (6.70) follows.
Linear Elasticity 227
Equation (6.70) is widely used in mechanics for ﬁnding approximate
values of eigenfrequencies. After using eigenmode orthogonality to approx
imate an eigenmode by ˜ w, we can approximate the eigenfrequency through
the equality
˜ ω =
R( ˜ w).
A ﬁnitedimensional analogue of Rayleigh’s quotient gave birth to a class
of numerical methods for large sparse systems of linear algebraic equations.
The interested reader should consult specialized literature.
6.13 Plane Waves
It is harder to obtain an exact solution for a dynamics problem than for
an equilibrium problem. However, there is an important dynamics problem
that can be solved analytically. This is the problem of plane waves in an
unbounded homogeneous anisotropic space that is free of body forces.
Such solutions describe propagation of sound far from the source, say
for distant earthquakes or explosions, so they are of signiﬁcant interest in
applications. On the other hand, their theory is relatively simple, and the
investigation reduces to an algebraic eigenvalue problem.
A solution of the form
u = U(k r −ωt)a
is called a plane wave. Here k is a constant vector called the wave vector,
ω is the frequency, a is a constant vector, and U = U(x) is an unknown
function in one variable x = k r −ωt . The equation
k r −ωt = const
represents a plane propagating in the space R
3
. At any instant, the normal
to this plane is parallel to k. The propagation velocity of the plane is
c = ω/[k[; this is the wave velocity. A solution of this form is constant
on each propagating plane. The vector k deﬁnes the direction of wave
propagation, whereas a is the direction of the displacement. For u we have
∇u = U
(kr−ωt)ka, ε =
1
2
U
(kr−ωt)(ka+ak), ¨ u = ω
2
U
(kr−ωt)a,
where the prime denotes diﬀerentiation with respect to x, where x = k r−
ωt, and the overdot denotes diﬀerentiation with respect to t.
228 Tensor Analysis with Applications in Mechanics
Using the symmetry properties of C, we write out Hooke’s law for this
displacement ﬁeld:
σ = C (ka)U
.
Substituting ¨ u and σ into the equations of motion, we get
k C (ka)U
= ρω
2
U
a.
This equation has the uninteresting trivial solution U
= 0. The necessary
nontrivial solutions are deﬁned by the following algebraic problem:
Find a nontrivial solution of the equation
A(k) a = ρω
2
a, where A(k) = k C k. (6.71)
We call A the acoustic tensor.
We recall that C has some symmetry properties, and that ε C ε is
positive for any ε = 0. It is easy to verify that A(k) is a positive deﬁnite
symmetric tensor. Considering ρω
2
as an eigenvalue of A(k), we have an
eigenvalue problem (6.71) that possesses three positive eigenvalues. In other
words, from (6.71) it follows that for any k there exist three values of ω. A
plane wave corresponds to each of these.
Let us consider the plane wave problem for an isotropic medium in more
detail. Now
C = λEE+ µ(e
k
Ee
k
+I),
where E = e
k
e
k
is the unit tensor and I = e
k
e
m
e
k
e
m
. The acoustic tensor
becomes
A = k C k = (λ + µ)kk + µ(k k)E. (6.72)
Let us take a Cartesian frame with basis vectors i
1
, i
2
,
˜
k = k/[k[, so
[i
1
[ = [i
2
[ = [
˜
k[ = 1 and i
1
i
2
= i
1
˜
k = i
2
˜
k = 0. In this basis, the matrix
A(k) is diagonal:
A(k) =
⎛
⎝
µ[k[
2
0 0
0 µ[k[
2
0
0 0 (2µ + λ)[k[
2
⎞
⎠
.
Linear Elasticity 229
The eigenvalues and eigenvectors from (6.71) take the form
ω
1
=
µ
ρ
[k[ for a = i
1
,
ω
2
=
µ
ρ
[k[ for a = i
2
,
ω
3
=
λ + 2µ
ρ
[k[ for a =
˜
k.
The ﬁrst two solutions describe transverse waves; their directions of dis
placement are those a that are perpendicular to k, the direction of wave
propagation. This type of wave is also called a shear or S wave. The third
equation describes a longitudinal wave; its displacement is along the direc
tion of propagation (Fig. 6.7). This type of wave is also called a dilatational,
pressure, or P wave. Such solutions are used in acoustics and seismology.
i
i
1
2
k
a
i
i
1
2
k
a
Fig. 6.7 Plane waves in an isotropic medium. Longitudinal wave is on the left. Trans
verse wave is on the right.
Let us note that for an arbitrary anisotropic medium, the spectral prob
lem describes waves that may be neither transverse nor longitudinal.
Exercise 6.11. Verify (6.72).
Exercise 6.12. Prove that A(k) is a symmetric tensor.
Exercise 6.13. Show that A(k) is positive deﬁnite.
230 Tensor Analysis with Applications in Mechanics
6.14 Plane Problems of Elasticity
Two important classes of deformation for which elastic problems can be sig
niﬁcantly simpliﬁed are the plane strain and plane stress deformations. The
problems for these deformations are called the plane problems of elasticity.
First we consider the plane deformation problem. A deformation is
called plane if all the displacements of the body points are parallel to a
plane and depend only on the coordinates in this plane. With reasonable
precision, this is the case for a central portion of a long cylindrical or
prismatic body stretched by axial forces.
The displacement vector for plane deformation takes the form
u = u
1
(x
1
, x
2
)i
1
+ u
2
(x
1
, x
2
)i
2
.
The body and contact forces should be of similar form:
f = f
1
(x
1
, x
2
)i
1
+ f
2
(x
1
, x
2
)i
2
, t
0
= t
1
(x
1
, x
2
)i
1
+ t
2
(x
1
, x
2
)i
2
.
In a Cartesian frame, the strain tensor components with indices 3, corre
sponding to u, are zero: ε
13
= ε
23
= ε
33
= 0, so
ε = ε
11
i
1
i
1
+ ε
12
(i
1
i
2
+i
2
i
1
) + ε
22
i
2
i
2
.
Using Hooke’s law, we ﬁnd the components of the stress tensor:
σ
11
= λtr ε + 2µε
11
, σ
22
= λtr ε + 2µε
22
, σ
33
= λtr ε,
σ
12
= µε
12
, σ
23
= 0, σ
13
= 0,
where tr ε = ε
11
+ε
22
. Because ε
33
= 0, we can express σ
33
in terms of σ
11
and σ
22
:
σ
33
=
λ
2(λ + µ)
(σ
11
+ σ
22
) = ν(σ
11
+ σ
22
).
All the components of σ do not depend on x
3
; this allows us to simplify
the equilibrium equations:
∂σ
11
∂x
1
+
∂σ
21
∂x
2
+ ρf
1
= 0,
∂σ
12
∂x
1
+
∂σ
22
∂x
2
+ ρf
2
= 0.
Let us consider a particular case when body forces are absent. Intro
ducing the Airy stress function Φ via the relations
σ
11
=
∂
2
Φ
∂x
2
2
, σ
22
=
∂
2
Φ
∂x
2
1
, σ
12
= −
∂
2
Φ
∂x
1
∂x
2
, (6.73)
Linear Elasticity 231
we identically satisfy the equilibrium equations. From the set of Beltrami–
Michell equations, in the plane problem there remain only three nontrivial
equations:
∆σ
11
+
1
1 + ν
∂
2
σ
∂x
2
1
= 0,
∆σ
22
+
1
1 + ν
∂
2
σ
∂x
2
2
= 0,
∆σ
12
+
1
1 + ν
∂
2
σ
∂x
1
∂x
2
= 0, (6.74)
where
σ = tr σ = (1 + ν)(σ
11
+ σ
22
).
Substitute (6.73) into (6.74). Since
σ = (1 + ν)∆Φ,
we reduce the ﬁrst of the equations (6.74) to the biharmonic equation for
Φ:
0 = ∆σ
11
+
1
1 + ν
∂
2
σ
∂x
2
1
= ∆
∂
2
Φ
∂x
2
2
+
∂
2
∂x
2
1
∆Φ = ∆
2
Φ.
We can reduce the second equation from (6.74) to the same biharmonic
equation, whereas the third equation from (6.74) holds identically. Thus Φ
satisﬁes
∆
2
Φ = 0. (6.75)
In the plane problem, this plays the role of the compatibility equation; it
is the equation we have to solve.
Similar simpliﬁcations can be done for the plane stress problem. This
type of deformation occurs in a thin plate with planar faces free of load.
We appoint the x
3
direction along the normal to the plate faces. Now the
stress components σ
13
, σ
23
, and σ
33
are small, so we set them to zero:
σ
13
= σ
23
= σ
33
= 0. The equality σ
33
= 0 distinguishes the plane strain
deformation from the above plane deformation where σ
33
= 0. For the
plane stress deformation, the analysis of the equilibrium equations can be
performed as above.
Exercise 6.14. Find the components of ε for the plane stress deformation.
Exercise 6.15. Find the components of the strain tensor for a plane stress
state.
232 Tensor Analysis with Applications in Mechanics
The results of Exercises 6.14 and 6.15 show that the equations of the
plane strain and plain stress problems diﬀer only in the values of the cor
responding elastic moduli λ.
Special tools have been developed for use in plane elasticity theory.
Some are based on the theory of functions of a complex variable
[
Muskhe
lishvili (1966)
]
. Using the complex potential method, we can ﬁnd exact
solutions to rather complex problems. However, computer techniques and
the ﬁnite element method have signiﬁcantly reduced interest in these.
6.15 Problems
6.1 Write out the boundary conditions for the problem of plane elasticity
on the square ABCD shown in Fig. 6.8.
6.2 Write out the boundary conditions for the plane elasticity boundary
value problem on the triangle ABC shown in Fig. 6.9.
6.3 Write out the boundary conditions for the plane elasticity problem on
the triangle shown in Fig. 6.10. Assume the forces are normal to the sides
and equal to p
1
, p
2
, p
3
, respectively.
6.4 Write out the boundary conditions for the plane elasticity problem on
the portion of a ring depicted in Fig. 6.11.
6.5 Write out the boundary conditions for the elasticity problem on the
visible portions of the cubes shown in Fig. 6.12.
6.6 At a point of an elastic body the principal stresses are σ
1
= 50 MPa,
σ
2
= −50 MPa, and σ
3
= 75 MPa. Referring to Fig. 6.13, ﬁnd the stress
vector on ABC that is equiinclined to the principal axes.
6.7 Referring to Fig. 6.14, write out the boundary conditions for the elas
ticity problem on the visible portions of the cylindrical bodies.
6.8 Let γ, λ, k be given parameters. For the following displacement vectors,
ﬁnd the expressions for the strain tensors.
(a) u = γx
2
i
1
;
(b) u = λx
1
i
1
;
(c) u = λr;
(d) u = u(r)e
r
+ kze
z
(use cylindrical coordinates);
Linear Elasticity 233
i
i
1
2
A
B C
D A
B C
D
A
B C
D A
B C
D
2s
s
2s
s
a) b)
d) c)
t
t
t
t
p p
s
s
t
t
t t
p
p
Fig. 6.8 Problem 6.1.
(e) u = u(r)e
φ
+ kze
z
(use cylindrical coordinates);
(f) u = (u(r) + kz)e
z
(use cylindrical coordinates);
(g) u = u(r)e
r
(use spherical coordinates).
6.9 Let ϕ be an arbitrary biharmonic function. Verify that the expression
2µu = 2(1 −ν)∇
2
ϕ−∇∇ ϕ
is a solution of Lam´e’s equations (6.29) at f = 0. It is the Boussinesq–
Galerkin solution of the equilibrium problem.
6.10 Let ψ be an arbitrary harmonic vector function and ψ
0
an arbitrary
234 Tensor Analysis with Applications in Mechanics
i
i
1
2
A B
C
a
y
x
gy
n
Fig. 6.9 Problem 6.2.
i
i
1
2
A
B
C
a
y
x
n
b
n
p
1
p
p
2
3
Fig. 6.10 Problem 6.3.
harmonic function. Verify that the expression
2µu = 4(1 −ν)ψ −∇(r ψ + ψ
0
)
is a solution of the equilibrium equations (6.29) at f = 0. It is called the
Papkovich–Neuber solution.
6.11 Let ψ be an arbitrary harmonic vector function, and suppose the
function η satisﬁes the equation
∇
2
η = 2∇ ψ.
Verify that the expression
2µu = 4(1 −ν)ψ −∇η
Linear Elasticity 235
i
i
1
2
a) b)
t
t
p
1
p
2
A B
C
D
s
s
p
1
2
A B
C
D
p
Fig. 6.11 Problem 6.4.
A
B C
D
2s
s
a) b)
t
t
p
i
1
i
3
i
2 A
B C
D
E F
G
E F
G
t
s
t
p
p
1
2
3
Fig. 6.12 Problem 6.5.
is a solution to (6.29) at f = 0.
6.12 Let ψ be an arbitrary harmonic vector function, and suppose the
function ξ satisﬁes the equation
∂ξ
∂z
=
1
4ν −3
∇ ψ.
Verify that the expression
2µu = ∇ψ + z∇ξ
represents a solution to equations (6.29) at f = 0.
236 Tensor Analysis with Applications in Mechanics
i
1
i
3
i
2
n
A
B
C
Fig. 6.13 Problem 6.6.
i
1
i
3
i
2
a) b) c)
t
s
p
2
s
t t
1
3
t
1
t
3
t
1
t
3
A
B
C
D
E
F
G
H
A
B
C
D
E
F
G
H
2
t
Fig. 6.14 Problem 6.7.
Chapter 7
Linear Elastic Shells
Thinwalled structures are common in engineering practice. Examples in
clude ship hulls, tank beds, rocket and aircraft airframes, and various tubu
lar structures such as arteries, membranes, and vases. A shell resembles a
surface having thickness. Mathematical models for such objects were devel
oped along with the theory of threedimensional elasticity. However, shell
theory is still under development; it is an approximation to describe the be
havior of threedimensional objects using twodimensional models given on
a surface. Clearly, any shell theory can be accurate only in some restricted
number of situations. A few popular shell models exist. We will present
some linear models that are used in applications.
Historically, the ﬁrst shell model was based on the Kirchhoﬀ–Love as
sumptions. Models are also attributed to Timoshenko, Reissner, Cosserat,
Naghdi, and others. In shell theory, we encounter the names of Euler,
Lagrange, Poisson, Kelvin, etc. A particular case of shell theory is plate
theory, where the object under consideration is akin to the top of a desk.
In this chapter, some features of shell theory will be demonstrated only for
plate models: our goal is an introduction to the theory and a demonstration
of how tensor analysis is applied in its construction.
A mathematical model of a shell takes the form of a boundary value
problem for simultaneous partial diﬀerential equations given on a base sur
face, usually the midsurface of the shell. Since the quantities describing
shell behavior are functions of the surface coordinates, shell theory is closely
related to the theory of surfaces from diﬀerential geometry (Chapter 5). The
decisive point in any version of shell theory is the method of eliminating the
third spatial coordinate — along the normal to the shell surface — from
the threedimensional model.
Before the invention of the computer, the numerical solution of three
237
238 Tensor Analysis with Applications in Mechanics
dimensional elasticity problems was nearly impossible. Much interest was
directed toward shell theories, where problems could be solved manually to
required accuracy. Modern computer solution has not changed the situation
for thinwalled problems. This is because numerical calculations of strains
in shelllike bodies harbor diﬃculties related to huge mesh sizes in discrete
models of a shell as a threedimensional body and the illposedness of the
corresponding threedimensional problems. It is interesting to note that the
application of powerful computers has increased interest in more accurate
shell models which allow engineers to apply twodimensional ﬁnite elements
that incorporate a more realistic strain distribution along the shell thickness
coordinate.
There are two approaches to the development of twodimensional shell
models. The ﬁrst is based on direct approximation of a given three
dimensional boundary value problem for a shelllike body in terms of a
boundary value problem given on the base surface of the body. This is
done in a few ways. One is the hypothesis method, in which we assume
the form of the dependence of the shell displacement and stress ﬁelds on
the thickness coordinate. We will present an example of this type of model,
based on the classical Kirchhoﬀ–Love hypotheses. Another method involves
asymptotic expansion of the solution of a threedimensional problem with
respect to a natural small parameter ε, the ratio of the shell thickness to a
characteristic shell size such as the diameter of a circular plate. There are
other principles for expanding a solution in a series with respect to ε. These
models commonly use as a starting point the equilibrium equations for the
shell as a threedimensional body. Afterwards, the expansions reduce the
equations to equations given on the shell surface. The models typically
result in a system of equations on the base surface whose order exceeds the
order of the initial system for a threedimensional body.
Another approach to deriving shell models can be called the direct ap
proach. In this case a shell is considered as a material surface having ad
ditional mechanical properties. So it is a surface that resists deformation,
possesses strain energy, can have distributed mass, etc. The dynamical or
equilibrium equations for such a surface are formulated directly through
fundamental laws of mechanics such as the law of impulse conservation. In
the direct approach, the equations of threedimensional elasticity are not
used; hence constitutive laws similar to Hooke’s law should be formulated
independently.
In this chapter, we will demonstrate how tensor calculus can be used to
derive a few linear shell and plate models for the case of small strains. We
Linear Elastic Shells 239
will also discuss some properties of the boundary value problems of shell
theory. We start with the Kirchhoﬀ–Love model.
7.1 Some Useful Formulas of Surface Theory
In this section we collect some tensorial formulas needed to formulate shell
relations. Many are taken from Chapter 5, with certain changes in notation.
Let Σ be a suﬃciently smooth surface in R
3
(Fig. 7.1). The position
vector of a point on Σ is denoted by ρ. We introduce
1
coordinates q
1
and
q
2
over Σ, which is then described by the equation ρ = ρ(q
1
, q
2
).
n
r
S
q
q
r r
i
1
i
2
i
3
1
2
1
2
Fig. 7.1 Surface Σ with curvilinear coordinate lines (q
1
, q
2
).
The basis vectors on Σ are tangent to the coordinate lines at a point;
they are denoted by ρ
1
and ρ
2
. The vectors ρ
1
and ρ
2
constitute the dual
basis. We recall that
ρ
1
=
∂ρ
∂q
1
, ρ
2
=
∂ρ
∂q
2
, ρ
α
· ρ
β
= δ
β
α
(α, β = 1, 2).
From now on, Greek indices take the values 1, 2, whereas Roman indices
take the values 1, 2, 3. The normal n to Σ at a point is deﬁned by the
relation
n =
ρ
1
×ρ
2
ρ
1
×ρ
2

.
1
In Chapter 5, following the tradition of diﬀerential geometry, we denoted the coordi
nates by u
1
and u
2
. To avoid confusion with the components of the displacement vector
u in this chapter, we rename the coordinates as indicated.
240 Tensor Analysis with Applications in Mechanics
Let us introduce the metric tensor A on the surface:
A = ρ
α
ρ
α
= E−nn,
where E is the metric tensor in R
3
. It plays the role of the unit tensor on
Σ at a point: if a vector v lies in the tangent plane at the point so that
v · n = 0, then A· v = v.
Exercise 7.1. Let v · n = 0. Show that A· v = v.
Exercise 7.2. Show that A = a
αβ
ρ
α
ρ
β
= a
αβ
ρ
α
ρ
β
, where a
αβ
= ρ
α
· ρ
β
and a
αβ
= ρ
α
· ρ
β
.
Exercise 7.2 states that the components of A are the coeﬃcients of
the ﬁrst fundamental form of Σ. This is why A is also called the ﬁrst
fundamental tensor of Σ.
At a point of Σ, the vectors (ρ
1
, ρ
2
, n) constitute a basis of R
3
. The
dual basis is (ρ
1
, ρ
2
, n). This means that an arbitrary vector ﬁeld v given
on Σ can be represented as
v = v(q
1
, q
2
) = v
1
(q
1
, q
2
)ρ
1
+v
2
(q
1
, q
2
)ρ
2
+v
3
(q
1
, q
2
)n
= v
1
(q
1
, q
2
)ρ
1
+v
2
(q
1
, q
2
)ρ
2
+v
3
(q
1
, q
2
)n.
Note that v
3
= v
3
.
The gradient operator on Σ is given by the formula
¯
∇ = ρ
α
∂
∂q
α
.
To apply the diﬀerential operations, we need the derivatives of the basis
vectors ρ
1
, ρ
2
, n, ρ
1
, ρ
2
. We recall that the derivatives of the normal n
are (see equation (5.39))
∂n
∂q
α
= −b
αβ
ρ
β
.
We use the components b
αβ
to construct the tensor
B = b
αβ
ρ
α
ρ
β
.
Using the deﬁnition of
¯
∇, we see that
B = −
¯
∇n.
The symmetric tensor B is called the curvature tensor of Σ or its second
fundamental tensor; the components of B are the coeﬃcients of the second
fundamental form of Σ.
Linear Elastic Shells 241
Exercise 7.3. Show that B = b
αβ
ρ
α
ρ
β
= b
αβ
ρ
α
ρ
β
. Also show that the
expressions for b
αβ
coincide with the coeﬃcients of the second fundamental
form (§ 5.5).
The derivatives of ρ
α
and ρ
α
are given by the formulas
∂ρ
α
∂q
β
= Γ
γ
αβ
ρ
γ
+b
αβ
n,
∂ρ
α
∂q
β
= −Γ
α
βγ
ρ
γ
+b
α
β
n,
where Γ
γ
αβ
is the Christoﬀel symbol introduced in § 4.5.
Let us derive the expression for the gradient
¯
∇v of a vector ﬁeld v given
on Σ. We split v into two components; one, ˜ v, is tangent to Σ such that
˜ v · n = 0, and the other is normal to Σ:
v = ˜ v +wn, ˜ v = v
1
(q
1
, q
2
)ρ
1
+v
2
(q
1
, q
2
)ρ
2
, w = v
3
= v
3
= v · n.
Then
¯
∇v = (
¯
∇˜ v) · A− wB+ (
¯
∇w +B· ˜ v)n. (7.1)
We also will need the expression for the tensor divergence:
¯
∇· T = ρ
γ
∂
∂q
γ
·
T
αβ
ρ
α
ρ
β
+ T
3β
nρ
β
+ T
α3
ρ
α
n +T
33
nn
=
∂T
αβ
∂q
α
ρ
β
+T
αβ
ρ
γ
·
∂ρ
α
∂q
γ
ρ
β
+T
αβ
∂ρ
β
∂q
α
+T
3β
ρ
γ
·
∂n
∂q
γ
ρ
β
+ T
33
ρ
γ
·
∂n
∂q
γ
n +
∂T
α3
∂q
α
n +T
α3
ρ
γ
·
∂ρ
α
∂q
γ
n +T
α3
∂n
∂q
α
=
∂T
αβ
∂q
α
ρ
β
+T
αβ
Γ
γ
αγ
ρ
β
+T
αβ
Γ
γ
βα
ρ
γ
+T
αβ
b
αβ
n −T
3β
b
γ
γ
ρ
β
− T
33
b
γ
γ
n +
∂T
α3
∂q
α
n +T
α3
Γ
γ
αγ
n −T
α3
b
β
α
ρ
β
. (7.2)
Exercise 7.4. Establish (7.1).
Exercise 7.5. Use (7.1) to show that
¯
∇· v =
¯
∇· ˜ v −wtr B.
Exercise 7.6. Let v = v(φ, θ) be a vector given on a sphere with spherical
coordinates φ and θ. Find
¯
∇v.
Exercise 7.7. Let v = v(φ, z) be a vector given on a cylindrical surface
with cylindrical coordinates φ and z. Find
¯
∇v.
Exercise 7.8. Let T be a secondorder tensor given on Σ, such that T·n =
n · T = 0. Show that the equation
¯
∇· T = 0 implies the algebraic relation
T
αβ
b
αβ
= 0.
242 Tensor Analysis with Applications in Mechanics
Exercise 7.9. Let Tbe a secondorder tensor given on a sphere with spher
ical coordinates φ, θ. Find
¯
∇· T.
Exercise 7.10. Let T be a secondorder tensor given on a cylindrical sur
face with cylindrical coordinates φ, z. Find
¯
∇· T.
7.2 Kinematics in a Neighborhood of Σ
To describe a spatial domain that surrounds Σ, we will use the coordi
nates and quantities of the previous section. A shell is a threedimensional
body occupying some neighborhood of the base surface Σ, which is used to
describe shell kinematics as well as displacements and strains under load.
The coordinates of a point Q in a neighborhood of Σ are introduced as
follows. Let n be the normal to Σ through a point Q. Let the base point
of n be P, whose position vector is ρ(q
1
, q
2
) so that its coordinates on Σ
are q
1
, q
2
. We appoint q
1
, q
2
as the two ﬁrst coordinates of Q, and the
distance z from P to Q as the third coordinate (Fig. 7.2). Note that z is
taken positive when
−−→
PQ is codirected with n and negative when oppositely
directed.
z
n
r
n
S
q
q
P
Q
i
1
i
2
i
3
r
r
r
1
1
2
2
Fig. 7.2 Kinematics in a neighborhood of Σ.
The position vector r of Q is
r = r(q
1
, q
2
, z) = ρ(q
1
, q
2
) +zn.
Linear Elastic Shells 243
The volume occupied by the shell is deﬁned by the inequalities
−h/2 ≤ z ≤ h/2,
where h is the shell thickness. In general, h = h(q
1
, q
2
) can vary from point
to point. At a point Q near Σ, the basis vectors are
r
α
=
∂r
∂q
α
= ρ
α
+z
∂n
∂q
α
= (A−zB) · ρ
α
, r
3
= n. (7.3)
The dual basis is given by
r
α
= (A−zB)
−1
· ρ
α
, r
3
= n. (7.4)
Let us explain the notation (A−zB)
−1
. It is seen that (A−zB)· n = 0,
so A−zBdegenerates as a threedimensional tensor and its inverse does not
exist. Here we restrict A−zB to an operator acting in the twodimensional
subspace tangent to Σ. This subspace has basis ρ
1
, ρ
2
, and the tensor A
plays the role of the unit operator on it. In what follows, we suppose that
hB is very small. Now, using the Banach contraction principle
[
Lebedev
and Cloud (2003)
]
, we can prove the existence of a unique inverse to A−zB
for small z ≤ h/2 on the tangent space. So we have
(A−zB)
−1
· (A−zB) = (A−zB) · (A−zB)
−1
= A. (7.5)
We illustrate how to derive (A−zB)
−1
via two important examples.
(1) Let Σ be a cylinder of radius R. In cylindrical coordinates
n = e
r
,
A = E−nn = e
φ
e
φ
+e
z
e
z
,
B = −
¯
∇e
r
= −e
φ
e
φ
/R.
Then
A−zB =
1 +
z
R
e
φ
e
φ
+e
z
e
z
.
It is easy to check that
(A−zB)
−1
=
1 +
z
R
−1
e
φ
e
φ
+e
z
e
z
satisﬁes (7.5) and hence is the needed inverse tensor.
(2) Now we deﬁne (A − zB)
−1
for a sphere of radius R. In spherical
coordinates,
n = e
r
,
A = E −nn = e
φ
e
φ
+e
θ
e
θ
,
B = −
¯
∇e
r
= −(e
φ
e
φ
+e
θ
e
θ
)/R.
244 Tensor Analysis with Applications in Mechanics
Then
A−zB =
1 +
z
R
(e
φ
e
φ
+e
θ
e
θ
).
Again, a direct check of formula (7.5) for
(A−zB)
−1
=
1 +
z
R
−1
(e
φ
e
φ
+e
θ
e
θ
)
demonstrates that it is the needed inverse tensor.
These examples show that (A− zB)
−1
cannot be found for all values
of z. In both cases, the inverse tensor exists for all z ≤ h/2 if h/2R < 1.
This is the domain in which each point is uniquely deﬁned by the coordinate
system introduced above. We shall use the deﬁnition of (A−zB)
−1
in this
restricted sense.
Finally, using the dual basis and (7.4), we ﬁnd the following represen
tation of the spatial nabla operator:
∇ = r
α
∂
∂q
α
+n
∂
∂z
= (A−zB)
−1
· ρ
α
∂
∂q
α
+n
∂
∂z
= (A−zB)
−1
·
¯
∇+n
∂
∂z
.
7.3 Shell Equilibrium Equations
Let us derive the twodimensional equations for shell equilibrium. These
will be a direct consequence of the threedimensional equilibrium equations
for the shell as a spatial body occupying a domain V in R
3
(see Fig. 7.3).
V is bounded by two surfaces (faces) S
−
and S
+
, each at distance h/2 from
the midsurface Σ, and by the lateral ruled surface S
ν
. S
ν
is formed by the
motion of the normal n to Σ along the boundary of Σ. In other words, V
is the set of all spatial points given by the position vector
r = ρ(q
1
, q
2
) +zn,
where ρ is the position vector of a point on Σ and −h/2 ≤ z ≤ h/2. For
simplicity, we suppose h is constant.
On the faces S
±
, the boundary conditions are
n
+
· σ
z=h/2
= t
0
+
, n
−
· σ
z=−h/2
= t
0
−
, (7.6)
Linear Elastic Shells 245
where n
±
are the normals to S
±
(Fig. 7.3) and t
0
±
are the surface loads
on S
±
, respectively. In the general case, on S
ν
we will assign the mixed
boundary conditions of the third problem of elasticity.
r
n
n
+
S
+
S
n
w
n
_
i
1
i
2
i
3
r
Fig. 7.3 A shelllike body.
Equation (6.20) with ¨ u = 0 describes the equilibrium of a three
dimensional elastic body occupying volume V :
1
√
g
∂
∂q
i
√
gσ
ij
r
j
+ρf = 0. (7.7)
Using q
1
, q
2
, q
3
= z as the coordinates, in equation (7.7) we can isolate the
diﬀerentiation with respect to z. Let us denote
σ
α
= r
α
· σ = σ
αj
r
j
, σ
3
= n · σ = σ
3j
r
j
.
We get
1
√
g
∂(
√
gσ
α
)
∂q
α
+
∂(
√
gσ
3
)
∂z
+ρf = 0. (7.8)
We recall that g is the determinant of the matrix g
ij
= r
i
· r
j
. We can
represent g in terms of the parameters of Σ. Indeed,
g
αβ
= ρ
α
· (A−zB)
2
· ρ
β
, g
13
= 0, g
23
= 0, g
33
= 1.
So g takes the form
g = aG
2
, a = a
11
a
22
−a
2
12
, G = det(A−zB), (7.9)
246 Tensor Analysis with Applications in Mechanics
where a
αβ
are equal to the values of g
αβ
taken on Σ. Here det(A−zB) is
calculated as the determinant of a twodimensional tensor:
det X = X
1
1
X
2
2
−(X
2
1
)
2
,
where X
β
α
are the mixed components of X = A−zB. Also observe that in
the new notation (7.6) is
σ
3
z=±h/2
= t
0
±
.
Let us rewrite (7.8) in the form
1
√
a
∂(
√
aGσ
α
)
∂q
α
+
∂(Gσ
3
)
∂z
+ρGf = 0. (7.10)
Now we can derive twodimensional equilibrium equations for the shell. We
integrate (7.10) over the thickness, i.e., along the normal coordinate z from
−h/2 to h/2. Taking into account the boundary conditions (7.6), we get
1
√
a
∂
∂q
α
√
a [[σ
α
]]
+G
+
t
0
+
−G
−
t
0
−
+ [[ρf ]] = 0, (7.11)
where
G
±
= G
z=±h/2
.
Here we introduced the notation [[· · · ]]. For any quantity f, this denotes
the deﬁnite integral of Gf over thickness:
[[f]] =
h/2
−h/2
Gf dz.
Let us transform (7.11) to componentfree form. The quantity
T = ρ
α
[[σ
α
]]
can be represented as follows:
T = ρ
α
[[r
α
· σ]]
= ρ
α
[[ρ
α
· (A−zB)
−1
· σ]]
= ρ
α
ρ
α
· [[(A−zB)
−1
· σ]]
= A· [[(A−zB)
−1
· σ]]
= [[(A−zB)
−1
· σ]].
We call T the stress resultant tensor. By deﬁnition, it is clear that n·T = 0.
The components of T·Aare the stress resultants acting in the tangent plane
Linear Elastic Shells 247
of the shell. T·n is the vector of transverse shear stress resultants. So (7.11)
becomes
¯
∇· T+q = 0, (7.12)
where
q = G
+
t
0
+
−G
−
t
0
−
+ [[ρf ]].
The vector q is the distributed load on the shell. Its component q · A acts
in the tangent plane to Σ, and q · n is the transverse load. So we have
derived the ﬁrst shell equilibrium equation.
To derive the second equilibrium equation, we crossmultiply the terms
of (7.8) by zn from the left:
1
√
a
zn ×
∂(
√
aGσ
α
)
∂q
α
+zn ×
∂(Gσ
3
)
∂z
+ρGzn ×f = 0. (7.13)
Moving n through the derivative sign, we get
1
√
a
∂
∂q
α
(n ×
√
aGzσ
α
) −
∂n
∂q
α
×Gzσ
α
+z
∂
∂z
(Gn ×σ
3
) +ρGzn ×f = 0.
Because
r
α
= ρ
α
+z
∂n
∂q
α
so that
z
∂n
∂q
α
= r
α
−ρ
α
,
we obtain
∂n
∂q
α
×zσ
α
= (r
α
−ρ
α
) ×σ
α
= −n ×σ
3
−ρ
α
×σ
α
.
Since
r
i
×σ
i
= r
i
×(r
i
· σ)
= r
i
×(r
i
· σ
kj
r
k
r
j
)
= r
i
×r
j
σ
ij
= −r
j
×r
i
σ
ij
= −r
j
×r
i
σ
ji
= −r
i
×r
j
σ
ij
= −r
i
×σ
i
,
248 Tensor Analysis with Applications in Mechanics
it follows that
r
i
×σ
i
= 0.
From
r
i
×σ
i
= r
α
×σ
α
+n ×σ
3
= 0
it follows that
r
α
×σ
α
= −n ×σ
3
.
Equation (7.13) becomes
1
√
a
∂
∂q
α
(n ×
√
aGzσ
α
) +ρ
α
×Gσ
α
+
∂
∂z
(Gn ×zσ
3
) +ρGzn ×f = 0.
Integrating with respect to z from −h/2 to h/2 and using (7.6), we get the
second vector equilibrium equation for the shell:
1
√
a
∂
∂q
α
√
a[[n ×zσ
α
]]
+ρ
α
×[[σ
α
]]
+
hG
+
2
n ×t
0
+
+
hG
−
2
n ×t
0
−
+ [[ρzn ×f ]] = 0. (7.14)
Now we represent it in componentfree form. Let us introduce the stress
couple tensor
M = ρ
α
[[n ×zσ
α
]],
which we represent as
M = −ρ
α
[[r
α
· zσ ×n]]
= −ρ
α
[[ρ
α
· (A−zB)
−1
· zσ ×n]]
= −ρ
α
ρ
α
· [[(A−zB)
−1
· zσ ×n]]
= −A· [[(A−zB)
−1
· zσ ×n]]
= −[[(A−zB)
−1
· zσ ×n]].
By deﬁnition of M, we see that n · M = M· n = 0. Let us denote
m =
1
2
hG
+
n ×t
0
+
+
1
2
hG
−
n ×t
0
−
+ [[ρzn ×f ]],
which represents the external bending couples applied to the shell surface.
By the deﬁnition of m, we see that m belongs to the tangent plane of Σ;
that is, m· n = 0.
Finally, we introduce a vectorial invariant of T, which is
T
×
= ρ
α
×[[σ
α
]];
Linear Elastic Shells 249
this notation is called the Gibbsian cross. Now we can rewrite (7.14) in
componentfree form as follows:
¯
∇· M+T
×
+m = 0. (7.15)
Note that (7.12) and (7.15) are exact consequences of the equilibrium
equations for the threedimensional continuum of the shell. They express
the equality to zero of the sum of all forces and moments acting on a shell
element.
So we have derived two twodimensional equilibrium equations (7.12)
and (7.15) for a shell. The unknowns T and M are deﬁned on Σ. In what
follows, Σ will be called the shell surface, or the midsurface or base surface
of the shell.
7.4 Shell Deformation and Strains; Kirchhoﬀ’s Hypotheses
Kirchhoﬀ formulated his famous hypotheses on the laws of plate defor
mation that extend Bernoulli’s assumptions for bending of a beam. Since
Kirchhoﬀ’s hypotheses were extended by Love to shell theory, they are often
called the Kirchhoﬀ–Love hypotheses.
Kirchhoﬀ’s kinematic hypothesis. Any shell cross section that is per
pendicular to the midsurface before deformation remains perpendicular to
the deformed midsurface. The transverse shear strains are negligibly small.
Mathematically, this means that n · ε · τ = 0 for all τ such that τ · n = 0.
Kirchhoﬀ’s static hypothesis. The normal stress is negligible in com
parison with the other stresses, so σ
33
= n · σ · n = 0.
Experimentally, the behavior of the normal often reﬂects the real sit
uation quite well. In this theory, we ignore shear strains and transverse
compression in the shell. In the literature we can ﬁnd diﬀerent ver
sions of Kirchhoﬀ’s hypotheses. We can also ﬁnd discussions regarding
their interpretations, applications, etc. See
[
Ciarlet (1997); Ciarlet (2000);
Chr´ o´scielewski, Makowski, and Pietraszkiewicz (2004); Donnell (1976);
Goldenveizer (1976); Libai and Simmonds (1998); Novozhilov, Chernykh,
and Mikhailovskiy (1991); Timoshenko (1985); Vorovich (1999); Zhilin
(2006); Zubov (1982)
]
. These topics fall outside the scope of the present
book.
Kirchhoﬀ’s static hypothesis allows us to eliminate ε
33
≡ n · ε · n from
Hooke’s law. This is done in a manner similar to the elimination of ε
33
in
250 Tensor Analysis with Applications in Mechanics
plane elasticity (cf., Exercises 6.14 and 6.15). By Kirchhoﬀ’s hypothesis,
σ
33
≡ n · σ · n = 0.
Let the shell material be isotropic. Because
σ
33
= λtr ε + 2µε
33
= 0
it follows that
ε
33
= −
λ
λ + 2µ
(ε
11
+ε
22
).
So for the stress and strain tensors of the shell we get the representations
σ = ˜ σ +σ
α3
(r
α
n +nr
α
),
ε = ˜ ε +ε
α3
(r
α
n +nr
α
) −
λ
λ + 2µ
(ε
11
+ε
22
)nn,
where ˜ σ and ˜ ε satisfy the relations
n · ˜ σ = 0, ˜ σ · n = 0, n · ˜ ε = 0, ˜ ε · n = 0.
Hooke’s law takes the form
σ = 2µ
¸
ε +
λ
λ + 2µ
Etr ˜ ε
=
E
1 +ν
¸
ε +
ν
1 −ν
Etr ˜ ε
. (7.16)
In component representation, this is
σ
1
1
=
E
1 −ν
2
(ε
1
1
+νε
2
2
), σ
2
2
=
E
1 −ν
2
(ε
2
2
+νε
1
1
),
σ
2
1
=
E
1 +ν
ε
2
1
, σ
3
1
=
E
1 +ν
ε
3
1
,
σ
3
2
=
E
1 +ν
ε
3
2
, (7.17)
where
σ
β
α
= r
α
· σ · r
β
, σ
3
α
= r
α
· σ · n.
The components of ε are denoted similarly.
Exercise 7.11. Derive (7.16) and (7.17).
Now we turn to the kinematic hypothesis. By this hypothesis, any
crosssection perpendicular to the midsurface before deformation remains
perpendicular after deformation and, moreover, remains planar. Let us ﬁx
a point q
1
, q
2
on Σ. Any crosssection through the point is planar after
deformation if and only if the dependence of the displacement of the points
Linear Elastic Shells 251
on the normal at q
1
, q
2
is linear in z. So we conclude that the general form
of the displacement vector is
u(q
1
, q
2
, z) = ˜ v(q
1
, q
2
) +w(q
1
, q
2
)n −zϑ(q
1
, q
2
), (7.18)
where
ϑ(q
1
, q
2
) =
˜
ϑ +ϑ
n
n, n ·
˜
ϑ = 0.
Let us derive the expression for the strain tensor corresponding to (7.18).
We have
∇u =
¯
∇+n
∂
∂z
(˜ v + wn −zϑ)
=
¯
∇˜ v + (
¯
∇w)n −wB−z
¯
∇
˜
ϑ +zϑ
n
B−nϑ (7.19)
and
(∇u)
T
= (
¯
∇˜ v)
T
+n
¯
∇w −wB−z(
¯
∇
˜
ϑ)
T
+zϑ
n
B−ϑn.
So the strain tensor is
2ε = ∇u +∇u
T
=
¯
∇˜ v + (
¯
∇˜ v)
T
−2wB−z
¯
∇
˜
ϑ + (
¯
∇
˜
ϑ)
T
+ 2zϑ
n
B
+ (
¯
∇w)n −nϑ +n
¯
∇w −ϑn. (7.20)
We recall that by (7.1),
¯
∇˜ v = (
¯
∇˜ v) · A+ (B· ˜ v)n,
¯
∇
˜
ϑ = (
¯
∇
˜
ϑ) · A+ (B·
˜
ϑ)n.
Substituting these into (7.20), we get
2ε = (
¯
∇˜ v) · A+A· (
¯
∇˜ v)
T
−2wB−z
(
¯
∇
˜
ϑ) · A+A· (
¯
∇
˜
ϑ)
T
+ (B· ˜ v)n +n(B· ˜ v) −z(B·
˜
ϑ)n −zn(B·
˜
ϑ) + 2zϑ
n
B
+ (
¯
∇w)n −nϑ +n
¯
∇w −ϑn. (7.21)
Let us derive some other consequences of Kirchhoﬀ’s kinematic hypoth
esis. By the hypothesis we have
n · ε · τ = 0, (7.22)
where τ is an arbitrary vector orthogonal to n; in components this is ε
31
=
ε
32
= 0. Using (7.21) we ﬁnd that
2n · ε = B· ˜ v −
˜
ϑ +
¯
∇w −zB·
˜
ϑ −2ϑ
n
n.
252 Tensor Analysis with Applications in Mechanics
All terms on the righthand side except 2ϑ
n
n are orthogonal to n. It follows
from (7.22), which holds for any τ orthogonal to n, that
B· ˜ v −
˜
ϑ +
¯
∇w −zB·
˜
ϑ = 0.
From this we ﬁnd
˜
ϑ:
˜
ϑ = (A+zB)
−1
· (
¯
∇w +B· ˜ v).
In linear elasticity, the displacements and their derivatives should be
very small. For a suﬃciently thin and smooth shell, we also can suppose
that hB is small. We formalize this as
Assumption S. Suppose hB 1.
So we assume that the terms zB are negligibly small in comparison
with unity. Note that by (7.9), from Assumption S it follows that G =
det(A−zB) is very close to 1. Under Assumption S we ﬁnd that
˜
ϑ =
¯
∇w +B· ˜ v. (7.23)
With this, we have
2n · ε · τ = −zτ · B·
˜
ϑ.
On the left side of this equality we see n · ε · τ which, according to Kirch
hoﬀ’s kinematic hypothesis, is zero. On the right, by Assumption S, we
see terms of a higher order of smallness than
˜
ϑ. So the above choice of
˜
ϑ makes Kirchhoﬀ’s kinematic hypothesis suﬃciently accurate. Note that
the quantity ϑ
n
is still undeﬁned.
Substituting (7.23) into (7.21), we ﬁnd that
2ε = (
¯
∇˜ v) · A+A· (
¯
∇˜ v)
T
−2wB−z
(
¯
∇
˜
ϑ) · A+A· (
¯
∇
˜
ϑ)
T
−2ϑ
n
nn + 2zϑ
n
B−z(B·
˜
ϑ)n −zn(B·
˜
ϑ). (7.24)
By Assumption S, we set the underlined terms in (7.24) to zero. Finally,
we get the expression for the strain tensor in the form
2ε = (
¯
∇˜ v) · A+A· (
¯
∇˜ v)
T
−2wB
−z
(
¯
∇
˜
ϑ) · A+A· (
¯
∇
˜
ϑ)
T
−2ϑ
n
nn
= 2˜ ε −2ϑ
n
nn, (7.25)
where
2˜ ε = (
¯
∇˜ v) · A+A· (
¯
∇˜ v)
T
−2wB−z
(
¯
∇
˜
ϑ) · A+A· (
¯
∇
˜
ϑ)
T
.
Linear Elastic Shells 253
We rewrite the last equation as
˜ ε = +zæ, (7.26)
where
2 = (
¯
∇˜ v) · A+A· (
¯
∇˜ v)
T
−2wB,
æ = −
1
2
(
¯
∇
˜
ϑ) · A+A· (
¯
∇
˜
ϑ)
T
.
The tensor is the tangential strain measure or the midsurface strain ten
sor. It describes the shell deformation in the tangent plane. The tensor æ
describes the shell deformation due to bending, and is called the bending
strain measure or the tensor of changes of curvature. The simplest inter
pretation for these quantities will be formulated for a plate, a particular
case of the shell, later. Various simpliﬁcations of the strain measures ap
pear in the literature. They are based on the assumptions of smallness of
certain terms in and æ. We will not consider these particular cases here.
The above deformation measures are used, for example, by W.T. Koiter
[
Koiter (1970)
]
.
Let us note that in the framework of Kirchhoﬀ’s hypotheses, the quan
tity ϑ
n
does not appear in the expressions for and æ. It is only involved
in ε
33
. Thus, in the Kirchhoﬀ–Love theory, ϑ
n
does not aﬀect the shell
deformation and stress characteristics: indeed, if we set ϑ
n
= 0 we get the
above expressions for and æ. The quantity ϑ
n
describes the rotation of
the shell cross section, which is tangent to the midsurface, about the nor
mal. From a physical standpoint, at least for a homogeneous shell, these
rotations are very small in comparison with the other rotations. This is
related to the fact that the shell twisting stiﬀness is much bigger than its
bending stiﬀness. Thus, in the Kirchhoﬀ–Love theory, a consideration of
ϑ
n
and the rotations of the cross sections about the normal related to ϑ
n
is
impossible. In much of the literature, it is assumed that ϑ
n
is zero; alterna
tively some analysis of its smallness, similar to the above, is proposed.
See
[
Novozhilov, Chernykh, and Mikhailovskiy (1991); Vorovich (1999);
Zhilin (2006)
]
. The reader can ﬁnd a rigorous asymptotic proof of the small
ness of ϑ
n
, under certain assumptions on the loads, the type of boundary
conditions, and shell homogeneity, in
[
Goldenveizer (1976)
]
.
Let us turn to Hooke’s law. Substitute ε into equation (7.17). The
expression for ˜ σ becomes
˜ σ =
E
1 +ν
¸
˜ ε +
ν
1 −ν
Atr ˜ ε
. (7.27)
254 Tensor Analysis with Applications in Mechanics
In component form it is
σ
β
α
=
E
1 +ν
¸
ε
β
α
+
ν
1 −ν
δ
β
α
tr ˜ ε
(7.28)
or, in more detail,
σ
1
1
=
E
1 −ν
2
(
1
1
+ν
2
2
) +
Ez
1 −ν
2
(æ
1
1
+νæ
2
2
),
σ
2
2
=
E
1 −ν
2
(
2
2
+ν
1
1
) +
Ez
1 −ν
2
(æ
2
2
+νæ
1
1
),
σ
2
1
=
E
1 +ν
2
1
+
Ez
1 +ν
æ
2
1
.
Using these, we can ﬁnd the resultants T· A and M via the formulas
T = [[(A−zB)
−1
· σ]], M = −[[(A−zB)
−1
· zσ ×n]].
In the context of Kirchhoﬀ’s theory, knowing the deformation, we cannot
recover the transverse shear stress resultants T · n; indeed by Kirchhoﬀ’s
hypothesis, the shear strains ε
· 3
1
and ε
· 3
2
in the shell are zero. But the
vector T · n, which represents the shear stress components, is not zero.
Until now it has been left indeﬁnite, but T· n can be found via (7.12) and
(7.15).
Assumption S allows us to overcome the diﬃculties of integrating
(A− zB)
−1
over the thickness. This simpliﬁes the formulas for the stress
resultants and stress couples to
T· A = [[A· ˜ σ]], M = −[[A· z ˜ σ ×n]]. (7.29)
The simpliﬁcation (7.29) implies that the twodimensional equilibrium
equations (7.12) and (7.15) with these versions of T and M are not ex
act consequences of the threedimensional equilibrium equations as earlier.
Rather, they are only ﬁrst approximations to the equilibrium equations that
diﬀer in terms of the second order of smallness. Integrating with respect to
z over the thickness, we get
T· A =
Eh
1 +ν
¸
+
ν
1 −ν
Atr
(7.30)
and
M = −
Eh
3
12(1 +ν)
¸
æ+
ν
1 −ν
Atr æ
×n. (7.31)
In the basis ρ
α
, we have
T· A = T
αβ
ρ
α
ρ
β
Linear Elastic Shells 255
and
=
αβ
ρ
α
ρ
β
.
Then the component representation of the stress resultant tensor is
T
αβ
=
Eh
1 + ν
¸
αβ
+
ν
1 −ν
a
αβ
γ
γ
. (7.32)
Comparing this with (7.28), we see that
T
αβ
= [[σ
αβ
]] ≡
h/2
−h/2
σ
αβ
dz. (7.33)
Here we have used the fact that, under Assumption S, we can take G = 1;
hence we redeﬁned
[[f]] ≡
h/2
−h/2
f(z) dz.
Similarly, we ﬁnd the components of the stress couple tensor:
M
αβ
= [[zσ
αβ
]] ≡
h/2
−h/2
zσ
αβ
dz. (7.34)
These are related to the components of æ through the formulas
M
αβ
=
Eh
1 +ν
¸
æ
αβ
+
ν
1 −ν
a
αβ
æ
γ
γ
. (7.35)
Comparing (7.33) and (7.34) with (7.29), we see that
T· A = T
αβ
ρ
α
ρ
β
, M = −M
αβ
ρ
α
ρ
β
×n.
These expressions for the components of stress resultants and stress couple
tensors were used in
[
Koiter (1970)
]
,
[
Donnell (1976)
]
,
[
Libai and Simmonds
(1998)
]
, and many others.
Donnell analyzed the conditions under which a particular class of shal
low shells that obey Assumption S can be used. The theory of shallow shells
is usable for moderate deformations where the shell shape remains close to
planar. For example, this is the case when the midsurface is a spherical
cap which is a relatively small part of the complete sphere. This theory
is attributed to Donnell, Marguerre, and Vlasov. In writing out æ, they
neglect B · ˜ v in the expression for ϑ =
¯
∇w as a small quantity of higher
order. Thus, the strains in shallow shell theory are given by
2 = (
¯
∇˜ v) · A+A· (
¯
∇˜ v)
T
−2wB, æ = −
¯
∇
¯
∇w.
256 Tensor Analysis with Applications in Mechanics
This simpliﬁcation for æ was ﬁrst introduced by the Russian engineer and
mechanicist Vasilij Z. Vlasov
[
Vlasov (1949)
]
.
In shallow shell theory, in the expressions for the stress resultants and
stress couples, we use a “strange” plane geometry; we formally set the
curvature components B to zero in æ but retain them in . The diﬀerential
operations in this theory are produced in the curvilinear coordinate system.
Note that Vlasov’s initial equations were derived when the geometry of a
shell was taken to be plane, but some nonzero curvatures were artiﬁcially
included in . A particular case of shallow shell theory is the plate theory
considered in § 7.8.
It is worth mentioning some aspects of the application of tensor analy
sis to shell theory. In the literature, one can ﬁnd many versions of linear
shell theory using various assumptions on the smallness of certain quan
tities. Some of these were derived in special coordinate systems like the
ones constituted by the principal curvature lines. In a certain sense, for an
isotropic material, such equations may lose their isotropic properties when
we express them in arbitrary curvilinear coordinates. Such a change of the
isotropic property was noted by
[
Zhilin (2006)
]
for an early version of the
shell equations by Novozhilov. Tensor analysis allows us to avoid similar
accidents.
7.5 Shell Energy
We saw the importance of strain and total energy in elasticity. Using these,
we proved uniqueness of solution for boundary value problems of elasticity,
and formulated variational principles of elasticity. In shell theory, the en
ergy also plays an important role. Kirchhoﬀ’s hypotheses aﬀect the form of
the stresses and strains in a shell as a threedimensional body, so we should
derive the form of the strain energy density for the shell.
The strain energy density W of the shell as a threedimensional body
takes the form
W =
1
2
σ ·· ε =
1
2
σ
i
j
ε
j
i
=
1
2
σ
α
β
ε
β
α
.
This follows from Kirchhoﬀ’s hypotheses that ε
1
3
= 0, ε
2
3
= 0, and σ
3
3
= 0.
Linear Elastic Shells 257
Substituting (7.28) into W, we get
W =
E
2(1 +ν)
¸
˜ ε ·· ˜ ε +
ν
1 −ν
tr
2
˜ ε
=
E
2(1 +ν)
¸
ε
β
α
ε
α
β
+
ν
1 −ν
(ε
α
α
)
2
. (7.36)
Using (7.26), we transform this to
W =
E
2(1 + ν)
¸
·· +
ν
1 −ν
tr
2
+
Ez
2
2(1 +ν)
¸
æ·· æ+
ν
1 −ν
tr
2
æ
+
Ez
1 + ν
¸
·· æ+
ν
1 −ν
tr tr æ
. (7.37)
So, using Kirchhoﬀ’s hypotheses, we have derived the expression for W in
terms of , æ, and z.
Consider the strain energy functional
E =
V
W dV.
We recall that in curvilinear coordinates q
1
, q
2
, z the volume element is
dV =
√
g dq
1
dq
2
dz =
√
aGdq
1
dq
2
dz, G = det(A−zB).
Under our assumptions on smallness of the displacements, W is a quantity
of the second order of smallness because it is a quadratic form in the strains.
Under Assumption S, we can neglect zB in comparison with A, for which
det A = 1, and change G to 1. Integrating (7.37) over z explicitly, we
transform E to
E =
V
W
√
a dq
1
dq
2
dz =
Σ
U
√
a dq
1
dq
2
,
where U = [[W]] denotes the shell strain energy density per area:
U =
Eh
2(1 +ν)
¸
·· +
ν
1 −ν
tr
2
+
Eh
3
24(1 +ν)
¸
æ·· æ+
ν
1 −ν
tr
2
æ
.
(7.38)
U is a quadratic form in two tensors, and æ, deﬁned on Σ. We see that U
splits into two terms; the ﬁrst depends only on , while the second depends
only on æ. From a physical viewpoint, this decomposition means that the
strain energy density splits into two parts. The ﬁrst is due to the shell
tangential deformation, i.e., the deformations in the tangent plane. The
second is due to shell bending.
258 Tensor Analysis with Applications in Mechanics
From (7.38) it follows that U ≥ 0, and that U = 0 if and only if both
ε = 0 and æ = 0. Indeed, we recall that −1 < ν ≤ 1/2 for an elastic
material. Then
·· +
ν
1 −ν
tr
2
> 0 when = 0.
Similarly,
æ·· æ+
ν
1 −ν
tr
2
æ > 0 when æ = 0.
Exercise 7.12. Show that
·· +
ν
1 −ν
tr
2
> 0 when = 0.
Using the component notation for T and M, we can represent U in
another form:
U =
1
2
T
αβ
αβ
+
1
2
M
αβ
æ
αβ
. (7.39)
From (7.39), with use of (7.32) and (7.35), it follows that
T
αβ
=
∂U
∂
αβ
, M
αβ
=
∂U
∂æ
αβ
. (7.40)
In componentfree form these are
U =
1
2
T·· +
1
2
M·· (n ×æ) =
1
2
T·· +
1
2
(M×n) ·· æ (7.41)
and
T =
∂U
∂
, M×n =
∂U
∂æ
. (7.42)
U inherits the positive deﬁniteness of the energy density in three
dimensional linear elasticity, where the energy density W = 0 if and only if
ε = 0. We recall that ε = 0 if and only if the displacement ﬁeld is a small
displacement of a rigid body, i.e., it is the sum of a parallel translation and
a rotation of the body in space. We encounter a similar situation in shell
theory. In Kirchhoﬀ’s theory, the expressions for and æ are cumbersome,
so we will demonstrate that fact later only for a simpler case of the plate
theory.
Exercise 7.13. Prove (7.41).
Linear Elastic Shells 259
7.6 Boundary Conditions
To uniquely deﬁne a solution of the equilibrium problem for an elastic shell,
we should supplement the equilibrium equations with boundary conditions.
This is no elementary task: in the beginnings of shell theory, the question of
boundary conditions was debated for a long time. Even the great Poisson
proposed an erroneous number of boundary conditions.
The equilibrium equations are deﬁned on the midsurface Σ, so the
boundary conditions are given on its boundary contour ω.
First we will introduce the kinematic boundary conditions: on the
boundary we assign the displacement v that deﬁnes the position of the
boundary contour after deformation. In the Kirchhoﬀ–Love theory we
should also assign the rotation of the normal n about the tangential to the
boundary shell contour vector — that is, about the vector n×ν (Fig. 7.6).
n
n
w
n n x t=
Fig. 7.4 The trihedron n, ν, τ = n ×ν used to formulate the boundary conditions.
The rotation is deﬁned by the formula ϑ =
¯
∇w +B· ˜ v, so the rotation
about τ = n ×ν is
ϑ
ν
≡ ν · ϑ = ν ·
¯
∇w +ν · B· ˜ v =
∂w
∂ν
+ν · B· ˜ v,
where ν ·
¯
∇w = ∂w/∂ν, the derivative with respect to the normal. Thus,
in the Kirchhoﬀ–Love theory, the kinematic boundary conditions consist of
four scalar equations:
v
ω
= v
0
(s),
∂w
∂ν
+ν · B· ˜ v
ω
= ϑ
0
(s), (7.43)
where v
0
(s) and ϑ
0
(s) are given on ω.
Some diﬃculties are involved in formulating static boundary conditions
in the Kirchhoﬀ–Love theory. At ﬁrst glance, it seems we could assign
the quantities ν · T and ν · (M× n) on the boundary. But this provides
ﬁve, not four, scalar conditions. Earlier we mentioned Poisson’s error in
constructing plate theory: he proposed too many static conditions on the
boundary.
260 Tensor Analysis with Applications in Mechanics
The derivation of the static boundary conditions can be based on the
Lagrange variational principle for the shell, which states that in equilibrium
the total energy functional takes its minimum value on the set of kinemati
cally admissible displacements. The static boundary conditions become the
natural boundary conditions for the problem of minimizing the total energy;
they are dual to the kinematic conditions. For the shell, the proof of the
Lagrange principle and the derivation of the static boundary conditions
are cumbersome, so we refer to more specialized literature (e.g.,
[
Green
and Zerna (1954); Novozhilov, Chernykh, and Mikhailovskiy (1991)
]
) and
merely sketch the procedure for deriving the boundary conditions. Later,
this will be given in more detail for the technically simpler case of a plate.
A consequence of the Lagrange variational principle is the statement
that the variation of the shell strain energy is equal to the work of external
forces on the kinematically admissible displacements:
δE = δA.
It turns out that on the set of admissible displacements of the shell edge,
the quantities v = ˜ v +wn and ϑ
ν
, which constitute four scalar quantities,
can be assigned independently. Hence the work of external forces on the
shell edge takes the form
δA =
ω
ϕ
0
(s) · δv −
0
(s)δϑ
ν
ds,
where ϕ
0
(s) and
0
(s) are given functions on ω. It is seen that δA depends
on δv and δϑ
ν
linearly. With regard for (7.40) we get
δE =
Σ
δU
√
a dq
1
dq
2
=
Σ
(T
αβ
δ
αβ
+M
αβ
δæ
αβ
)
√
a dq
1
dq
2
.
The further transformations of δE are done using integration by parts and
the Gauss–Ostrogradsky theorem. We should represent δ
αβ
and δæ
αβ
through δv and δϑ and eliminate their derivatives from the surface inte
grals. Unlike the case with threedimensional elasticity, δæ
αβ
contains the
second derivatives of w and hence integration by parts should be done twice.
For the boundary integrals we should also apply further transformations in
such a way that they contain only the factors δv and δϑ
ν
; the techniques
will be presented in the section for the plate. The equation δE = δA holds
for any admissible δv and δϑ. Selecting only δv and δϑ satisfying the ho
mogeneous kinematic conditions, we ﬁnd that the minimum point v and ϑ
satisﬁes the equilibrium equations on Σ. Next, extending the set of δv and
Linear Elastic Shells 261
δϑ to all the admissible δv and δϑ, we derive the following set of static
boundary conditions for v and ϑ. They are deﬁned on ω:
ν · T· A
ω
= ϕ
0
(s) · A,
ν · T· n
ω
−
∂
∂s
(ν · (M×n) · τ)
ω
= ϕ
0
(s) · n, (7.44)
ν · (M×n) · ν
ω
=
0
(s).
As with the kinematic boundary conditions, these are also four scalar equa
tions. Here ϕ
0
(s) and
0
(s) are given functions on ω that represent the
stress resultants and the bending couples on the shell edge.
The principal boundary conditions used in engineering are as follows.
Clamped (ﬁxed) edge:
v
ω
= 0,
∂w
∂ν
ω
= 0.
Simple support edge:
v
ω
= 0, ν · (M×n) · ν
ω
= 0.
Free edge:
ν · T· A
ω
= 0, ν · T· n
ω
−
∂
∂s
(ν · (M×n) · τ)
ω
= 0,
ν · (M×n) · ν
ω
= 0.
Mixed boundary conditions are also possible. For example, on a part
of ω we might specify (7.43) and on the remainder (7.44). Other com
binations of boundary conditions, four at each point of ω, deﬁne various
boundary value problems. Admissible combinations of boundary conditions
are deﬁned by the variational setups of these problems.
7.7 A Few Remarks on the Kirchhoﬀ–Love Theory
There exist various approaches in the Kirchhoﬀ–Love theory. They use
diﬀerent simplifying assumptions that depend on the choice of constitu
tive relations for U, T, and M, as well as on the choice of deformation
measures and æ. In a linear theory, equation (7.38) for U is a common
form of dependence of U on the deformation measures (see, e.g.,
[
Donnell
(1976); Koiter (1970); Novozhilov, Chernykh, and Mikhailovskiy (1991);
Timoshenko (1985)
]
).
262 Tensor Analysis with Applications in Mechanics
The versions may diﬀer in their deﬁnitions of T. We introduced T as
T = [[(A−zB)
−1
· σ]]
and used Assumption S to recast it in the form
T = [[A· σ]].
In the Kirchhoﬀ–Love literature, one can ﬁnd both the initial and simpliﬁed
deﬁnitions. Let us consider this diﬀerence in more detail. We introduce a
special coordinate system on the shell surface, deﬁned by the principal
curvature curves on Σ. Now B takes the diagonal form
B = e
1
e
1
/R
1
+ e
2
e
2
/R
2
,
where R
1
and R
2
are principal radii of curvature of Σ. Then
A = e
1
e
1
+e
2
e
2
and
G = (1 −z/R
1
)(1 −z/R
2
).
The tensor G(A−zB)
−1
takes the form
G(A−zB)
−1
=
1 −
z
R
2
e
1
e
1
+
1 −
z
R
1
e
2
e
2
.
Taking
T = [[(A−zB)
−1
· σ]],
we get the following formulas for the components of T:
T
11
=
h/2
−h/2
1−
z
R
2
σ
11
dz, T
22
=
h/2
−h/2
1 −
z
R
1
σ
22
dz,
T
12
= T
21
=
h/2
−h/2
σ
12
dz.
The alternative deﬁnition T = [[A· σ]] yields the expressions
T
11
=
h/2
−h/2
σ
11
dz, T
22
=
h/2
−h/2
σ
22
dz, T
12
= T
21
=
h/2
−h/2
σ
12
dz.
So the two deﬁnitions of T yield equilibrium equations that diﬀer by quan
tities that are secondorder small. The simpliﬁed form for T was used
by a number of authors, e.g.,
[
Goldenveizer (1976); Novozhilov, Chernykh,
and Mikhailovskiy (1991); Timoshenko (1985)
]
. Koiter
[
Koiter (1970)
]
used
(7.38) and (7.40) to relate the stress resultants and the bending couples with
Linear Elastic Shells 263
U, but he presented more complex expressions to calculate them through
the stresses in a threedimensional body.
The components of M
αβ
are also commonly calculated using (7.34),
and are related to U through (7.40). As is the case for T, however, in the
literature one can ﬁnd other ways to introduce M. In
[
Goldenveizer (1976)
]
,
for instance, the expressions for M
αβ
include terms that are quadratic in z.
Some works on shell theory feature a “mixed” formulation of the bound
ary value problems, introducing the stress resultants via the simpliﬁed for
mula T = [[A· σ]]. The constitutive equations take the form (7.38), but the
equilibrium equations include additional terms arising from the complete
deﬁnition T = [[(A−zB)
−1
· σ]]. In this way, the couple stresses appear in
all the equilibrium equations.
In
[
Koiter (1970); Novozhilov, Chernykh, and Mikhailovskiy (1991)
]
and
certain other works, the above tensor is also used. The second deformation
measure æ, which was derived by the simplifying replacement of ϑ by
¯
∇w,
also has a few diﬀerent forms (see the discussion in
[
Koiter (1970)
]
).
In the linear Kirchhoﬀ–Love shell theory, the reader can encounter var
ious forms of the equilibrium equations in displacements, as well as dif
ferences in the stress resultants and bending couples. Clearly, in diﬀerent
versions of shell theory the forms of the static boundary conditions can
diﬀer, and this is also reﬂected in the literature.
It is worth noting that in diﬀerent references, the same equilibrium
equations in stresses can lead to diﬀerent equilibrium equations in displace
ments. For a thin shell with a suﬃciently smooth midsurface and smooth
loads, this diﬀerence, as a rule, is of the second order of smallness in com
parison with the main terms in the equations.
7.8 Plate Theory
Resultant forces and moments in a plate; equilibrium equa
tions
A particular case of a shell is a thin plate. The base surface Σ of a plate
is part of a plane, so B = 0. The relations for a plate follow from shell
theory. The equilibrium equations are
¯
∇· T+ q = 0,
¯
∇· M+T
×
+m = 0, (7.45)
264 Tensor Analysis with Applications in Mechanics
where
T = [[A· σ]], M = −[[A· zσ ×n]]. (7.46)
In a Cartesian basis i
1
, i
2
, i
3
= n, we have
σ = σ
αβ
i
α
i
β
+σ
α3
(i
α
n+ni
α
)+σ
33
nn, A = i
α
i
α
,
¯
∇ = i
α
∂
∂x
α
, G = 1,
where x
1
, x
2
are Cartesian coordinates in the plane. The stress resultants
are
T = T
αβ
i
α
i
β
+T
αn
i
α
n, (7.47)
T
αβ
= [[σ
αβ
]] ≡
h/2
−h/2
σ
αβ
dz, T
αn
= [[σ
α3
]] ≡
h/2
−h/2
σ
α3
dz.
Note that σ
33
does not appear in the deﬁnition of the stress resultants or
the transverse shear stress resultants. It follows that the matrix T
αβ
is
symmetric: T
αβ
= T
βα
. The quantity T
×
is calculated by the formula
T
×
= T
αβ
i
α
×i
β
+T
αn
i
α
×n = T
αn
i
α
×n = T
2n
i
1
−T
1n
i
2
.
Next, we introduce the moments matrix
M
αβ
= [[zσ
αβ
]] ≡
h/2
−h/2
zσ
αβ
dz. (7.48)
Here, M
αβ
is a symmetric matrix as well. With regard for the equality
σ ×n = σ
αβ
i
α
i
β
×n
= −σ
11
i
1
i
2
+σ
12
i
1
i
1
−σ
21
i
2
i
2
+σ
22
i
2
i
1
,
from (7.46) we obtain
M = [[zσ
11
]]i
1
i
2
−[[zσ
12
]]i
1
i
1
+ [[zσ
21
]]i
2
i
2
−[[zσ
22
]]i
2
i
1
.
By (7.48), M takes the form
M = M
11
i
1
i
2
− M
12
i
1
i
1
+M
21
i
2
i
2
−M
22
i
2
i
1
= −M
αβ
i
α
i
β
×n.
So the equilibrium equations in Cartesian coordinates are
∂T
αβ
∂x
α
i
β
+
∂T
αn
∂x
α
n +q = 0, −
∂M
αβ
∂x
α
i
β
×n +T
×
+m = 0. (7.49)
Linear Elastic Shells 265
In component form, these are
∂T
11
∂x
1
+
∂T
21
∂x
2
+q
1
= 0,
∂T
12
∂x
1
+
∂T
22
∂x
2
+q
2
= 0,
∂T
1n
∂x
1
+
∂T
2n
∂x
2
+q
n
= 0, (7.50)
and
∂M
11
∂x
1
+
∂M
21
∂x
2
−T
1n
+m
1
= 0,
∂M
12
∂x
1
+
∂M
22
∂x
2
−T
2n
+m
2
= 0, (7.51)
where
q
α
= i
α
· q, q
n
= n · q, m
α
= i
α
· (n ×m).
Note that T
12
= T
21
and M
12
= M
21
; we keep the diﬀerent notation for
symmetry.
In applications of shell theory, m is often negligible in comparison with
the external forces. Hence, for simplicity we put m = 0. Now we exclude
T
1n
and T
2n
from equations (7.51). For this, we diﬀerentiate the second
equation of (7.51) with respect to x
2
, the ﬁrst from (7.51) with respect to
x
1
, and add the results. Using the third equation from (7.50), we obtain
the wellknown equation of plate theory
∂
2
M
11
∂x
1
2
+ 2
∂
2
M
12
∂x
1
∂x
2
+
∂
2
M
22
∂x
2
2
+q
n
= 0. (7.52)
In index notation this reads
∂
2
M
αβ
∂x
α
∂x
β
+q
n
= 0,
whereas in componentfree notation it takes the form
¯
∇· (
¯
∇· (M×n)) +q · n = 0. (7.53)
Indeed,
M = −M
αβ
i
α
i
β
×n.
Using the identity
(i
β
×n) ×n = −i
β
,
266 Tensor Analysis with Applications in Mechanics
the proof of which is left to the reader, we get
M×n = −M
αβ
i
α
(i
β
×n) ×n = M
αβ
i
α
i
β
.
Next we obtain
¯
∇· (M×n) =
∂M
αβ
∂x
α
i
β
,
¯
∇· (
¯
∇· (M×n)) =
∂
2
M
αβ
∂x
α
∂x
β
,
from which (7.53) follows.
In plate theory, Kirchhoﬀ’s kinematic hypotheses reduce to an assump
tion on the form of the displacement ﬁeld, which is
u(x
1
, x
2
, z) = v(x
1
, x
2
) −z
¯
∇w(x
1
, x
2
)
= ˜ v +wn −z
¯
∇w, ˜ v · n = 0. (7.54)
Let us ﬁnd the expression for the strain tensor that corresponds to u.
We have
∇u =
¯
∇+n
∂
∂z
(˜ v +wn −z
¯
∇w)
=
¯
∇˜ v −z
¯
∇
¯
∇w + (
¯
∇w)n −n
¯
∇w,
and
(∇u)
T
= (
¯
∇˜ v)
T
−z
¯
∇
¯
∇w +n
¯
∇w −(
¯
∇w)n.
Thus,
2ε =
¯
∇˜ v + (
¯
∇˜ v)
T
−2z
¯
∇
¯
∇w.
It follows that ε · n = n · ε = 0.
Using the notation
2 =
¯
∇˜ v + (
¯
∇˜ v)
T
, æ = −
¯
∇
¯
∇w,
we ﬁnd that the plate strains split into two parts:
ε = +zæ. (7.55)
Here represents the deformation in the plate plane, and æ is related to
plate bending. Geometrically, æ describes an inﬁnitesimal change of the
plate curvatures due to bending.
Linear Elastic Shells 267
Now let us use Kirchhoﬀ’s static hypothesis σ
33
= 0. From Hooke’s law
it follows that
σ
11
=
E
1 −ν
2
(ε
11
+ νε
22
),
σ
22
=
E
1 −ν
2
(ε
22
+ νε
11
),
σ
12
=
E
1 +ν
ε
12
. (7.56)
Substituting (7.55) into these, we get
σ
11
=
E
1 −ν
2
(
11
+ν
22
) +
Ez
1 −ν
2
(æ
11
+νæ
22
),
σ
22
=
E
1 −ν
2
(
22
+ν
11
) +
Ez
1 −ν
2
(æ
22
+νæ
11
),
σ
12
=
E
1 +ν
12
+
Ez
1 +ν
æ
12
.
Now by (7.47) and (7.48), integrating the expressions for σ
αβ
with respect
to z over thickness, we derive the expressions for the stress resultants and
stress couples:
T
11
=
Eh
1 −ν
2
(
11
+ν
22
),
T
22
=
Eh
1 −ν
2
(
22
+ν
11
),
T
12
=
Eh
1 +ν
12
,
M
11
=
Eh
3
12(1 −ν
2
)
(æ
11
+νæ
22
),
M
22
=
Eh
3
12(1 −ν
2
)
(æ
22
+νæ
11
),
M
12
=
Eh
3
12(1 +ν)
æ
12
. (7.57)
The constant
D = Eh
3
/12(1 −ν
2
)
is the bending stiﬀness. Substituting the expressions for M
αβ
into (7.52)
and taking into account that
æ
11
= −
∂
2
w
∂x
2
1
, æ
22
= −
∂
2
w
∂x
2
2
, æ
12
= −
∂
2
w
∂x
1
∂x
2
,
268 Tensor Analysis with Applications in Mechanics
we obtain
D
∂
4
w
∂x
4
1
+ 2
∂
4
w
∂x
2
1
∂x
2
2
+
∂
4
w
∂x
4
2
= q
n
. (7.58)
In componentfree form this is
D
¯
∇
4
w = q
n
. (7.59)
Because
¯
∇
4
= (
¯
∇
2
)
2
= ∆
2
,
we call (7.58) (or (7.59)) the biharmonic equation.
This equation has a long history in plate theory. The bending equation
was published by MarieSophie Germain, who received a prize from the
Paris Academy of Sciences in 1816. Later some of Germain’s errors were
corrected by Lagrange, hence the equation is now called Germain’s equation
or the Germain–Lagrange equation. The plate theory we have considered
is attributed to Kirchhoﬀ, who formulated correct boundary conditions for
a plate. Moreover, Kirchhoﬀ clearly introduced physically meaningful hy
potheses regarding the distribution of displacements and stresses in the
plate.
Note that the T
αβ
do not depend on w. Thus, the plate equilibrium
equations split into two groups. The ﬁrst consists of the two ﬁrst equations
of (7.50) with respect to ˜ v, whereas the second consists of the equation
(7.59) for variable w.
Let us write the equations for T· A in terms of the tangential displace
ment vector ˜ v. In tensor form, equations (7.50) are
¯
∇· (T· A) +q · A = 0. (7.60)
The constitutive equation for T· A is
T· A =
Eh
1 +ν
¸
+
ν
1 −ν
Atr
.
So we obtain
¯
∇· (T· A) =
Eh
1 +ν
¸
¯
∇· +
ν
1 − ν
¯
∇tr
=
Eh
2(1 +ν)
¸
¯
∇·
¯
∇˜ v +
¯
∇
¯
∇· ˜ v +
2ν
1 −ν
¯
∇
¯
∇· ˜ v
=
Eh
2(1 +ν)
¸
¯
∇·
¯
∇˜ v +
1 +ν
1 −ν
¯
∇
¯
∇· ˜ v
.
Linear Elastic Shells 269
In terms of ˜ v, equation (7.60) takes the form
Eh
2(1 +ν)
¸
¯
∇·
¯
∇˜ v +
1 +ν
1 −ν
¯
∇
¯
∇· ˜ v
+q · A = 0. (7.61)
There may be some concern over equations (7.51), which contain the
transverse shear stress resultants T
1n
and T
2n
. It may seem that by the
kinematic hypothesis n · ε = 0 and the last equation from (7.47), both T
1n
and T
2n
should vanish. But Kirchhoﬀ’s kinematic hypothesis restricts only
the form of deformation. So in Kirchhoﬀ’s plate theory, T
1n
and T
2n
are not
deﬁned by the strains; they are the reactions due to the strain constraints.
Note that the equations (7.51) are used to determine T
1n
and T
2n
.
Boundary conditions
In plate theory as a particular case of shell theory, on the boundary contour
ω we can assign the kinematic conditions
˜ v = v
0
(s), w = w
0
(s),
∂w
∂ν
= ϑ
0
(s), (7.62)
or the static conditions
ν · T· A = ϕ
0
(s)·A, ν · T· n −
∂
∂s
(ν · (M×n) · τ) = ϕ
0
n
(s),
ν · (M×n) · ν =
0
(s). (7.63)
For a mixed boundary value problem, these sets of conditions can be given
on diﬀerent portions of ω. Alternatively, at a point we can include condi
tions from (7.62) and the dual conditions from (7.63). At each point, four
scalar conditions must be appointed.
We split the boundary load into tangential and normal parts:
ϕ
0
= ϕ
0
· A+ϕ
0
n
n, ϕ
0
n
= ϕ
0
· n.
From equilibrium equations (7.51), we ﬁnd the transverse shear stress re
sultants in terms of the stress couples. As in our earlier treatment of shell
theory, we set m
1
= m
2
= 0. We have
T
1n
=
∂M
11
∂x
1
+
∂M
21
∂x
2
, T
2n
=
∂M
12
∂x
1
+
∂M
22
∂x
2
.
These can be written as the single vector equation
T· n =
¯
∇· (M×n).
There are various types of mixed boundary conditions. We restrict
ourselves to the two deﬁned by (7.62) and (7.63). Note that T· A depends
270 Tensor Analysis with Applications in Mechanics
only on ˜ v, whereas M is deﬁned through w. Thus, like the equilibrium
equations, the boundary conditions in plate theory split into two sets. One
is for normal displacement w:
w
ω
1
= w
0
(s),
∂w
∂ν
ω
1
= ϑ
0
(s), (7.64)
ν · (
¯
∇· (M×n))
ω
2
−
∂
∂s
(ν · (M×n) · τ)
ω
2
= ϕ
0
n
(s),
ν · (M×n) · ν
ω
2
=
0
(s).
The other is for tangential displacements ˜ v:
˜ v
ω
3
= v
0
(s), ν · T· A
ω
4
= ϕ
0
(s) · A. (7.65)
We have supposed the plate contour to be partitioned in such a way that
ω
2
= ω \ ω
1
and ω
4
= ω \ ω
3
.
In Appendix A we list some common homogeneous boundary conditions
in terms of w for the plate.
Strain energy for the plate
Let us consider the strain energy density for a plate in more detail. It is
U =
Eh
2(1 +ν)
¸
·· +
ν
1 −ν
tr
2
+
Eh
3
24(1 +ν)
¸
æ·· æ+
ν
1 −ν
tr
2
æ
.
(7.66)
A plate is a particular case of a shell considered above, but with B = 0.
Hence we can use the expression for shell strain density
U =
1
2
T
αβ
αβ
+
1
2
M
αβ
æ
αβ
=
1
2
T·· ε +
1
2
M·· (n ×æ).
Let us rewrite U in terms of the displacements:
U =
Eh
2(1 +ν)
¸
∂u
α
∂x
β
∂u
α
∂x
β
+
ν
1 −ν
∂u
α
∂x
α
2
¸
+
Eh
3
24(1 +ν)
¸
∂
2
w
∂x
α
∂x
β
∂
2
w
∂x
α
∂x
β
+
ν
1 −ν
∂
2
w
∂x
α
∂x
α
2
¸
.
Rigid motions
Now we consider the consequences of the equation U = 0. Because U is
positive deﬁnite, it is zero if and only if both and æ are zero. But = 0
implies that the plate can move in its plane as a rigid body only. In other
Linear Elastic Shells 271
words, it can translate and rotate about a normal to its midplane. The
corresponding displacement ﬁeld takes the form
˜ v = v
0
+ω
0
n × (ρ −ρ
0
),
where v
0
is an arbitrary ﬁxed vector in the midplane, i.e., v
0
· n = 0, the
angle ω
0
is an arbitrary but ﬁxed rotation angle about n, the vector ρ is
the position vector of a point on the midplane, and ρ
0
is an arbitrary but
ﬁxed vector in the same plane. This formula is a particular case of (6.41)
for the plane deformation.
The equation
æ ≡ −
¯
∇
¯
∇w = 0
implies that w is a linear function in x
1
and x
2
:
w = w
0
+w
1
x
1
+w
2
x
2
,
where w
0
, w
1
, and w
2
are arbitrary ﬁxed scalars. Physically, w corresponds
to a translation of the whole plate in the normal direction and a rotation
with respect to the axes i
1
, i
2
. Thus, as for threedimensional elasticity,
U = 0 only for those displacements that represent small displacements of
the plate as a rigid body.
Lagrange variational principle in plate theory
In shell theory, the Lagrange variational principle of elasticity changes some
what because of Kirchhoﬀ’s hypotheses. We will formulate Lagrange’s vari
ational principle for the plate. Recall that in threedimensional elasticity,
the total energy functional takes the form
E(u) =
V
W(ε) dV −
V
ρf · udV −
S
t
0
· udS.
We found that for a thin shell, the integral over V is represented through
the integral over the midplane Σ. The same is true for the plate:
V
W(ε) dV =
Σ
U dΣ,
where U is deﬁned by (7.66). Let us consider the remaining terms in E,
which represent the work of external forces. In terms of the plate theory,
they take the form
Σ
q · v dΣ +
ω
4
ϕ
0
(s) · ˜ v dS +
ω
2
ϕ
0
n
w −
0
∂w
∂ν
dS.
272 Tensor Analysis with Applications in Mechanics
So the total energy functional for the plate splits into two parts:
E = E
˜ v
(˜ v) +E
w
(w),
where
E
˜ v
(˜ v) =
Eh
2(1 +ν)
Σ
¸
·· +
ν
1 −ν
tr
2
dΣ
−
Σ
q · ˜ vdΣ −
ω
4
ϕ
0
(s) · ˜ v dS
is due to tangential displacements, and
E
w
(w) =
Eh
3
24(1 +ν)
Σ
¸
æ·· æ+
ν
1 −ν
tr
2
æ
dΣ
−
Σ
wq · ndΣ −
ω
2
ϕ
0
n
wdS +
ω
2
0
∂w
∂ν
dS
accounts for bending.
The arguments ˜ v and w of E
˜ v
and E
w
are independent, so the prob
lem of minimizing the total energy naturally splits into two minimization
problems. Moreover, we can formulate two variational principles: one for
tangential deformations, and one for plate bending.
Lagrange’s variational principle for tangential deformation
Theorem 7.1. On the set of admissible tangential displacement ﬁelds, a
stationary point ˜ v of E
˜ v
satisﬁes the plate equilibrium equations (7.60) on
Σ, or, equivalently (7.61) and the boundary condition
ν · T· A
ω
4
= ϕ
0
(s) · A,
and conversely. The stationary point, if it exists, minimizes E
˜ v
.
The set of admissible tangential displacements ˜ v consists of vector func
tions twice diﬀerentiable in Σ and satisfying the kinematic boundary con
dition on ω
3
from (7.65). The proof mimics that of Lagrange’s principle in
threedimensional elasticity (Exercise 7.18).
Note that we represented the work of external forces using the gen
eral deﬁnition of work and common sense. The fact that the equilibrium
equations and the kinematic boundary conditions really deﬁne a stationary
point of the energy functional shows that the work functional was written
correctly.
Linear Elastic Shells 273
Lagrange’s variational principle for plate bending
Theorem 7.2. On the set of admissible deﬂection ﬁelds, a stationary point
w of E
w
satisﬁes the plate equilibrium equations (7.53) or (7.59) in Σ and
the static boundary conditions from (7.64) on ω
2
, and conversely. The
stationary point, if it exists, minimizes E
w
.
Note that the set of admissible deﬂections w consists of the functions
that possess continuous derivatives in Σ up to order four and satisfy two
kinematic boundary conditions on ω
1
from (7.64).
Proof. Let us deﬁne the variation of E
w
. We have
δ
Eh
3
24(1 +ν)
¸
æ·· æ+
ν
1 − ν
tr
2
æ
=
Eh
3
12(1 +ν)
¸
æ+
ν
1 −ν
tr æ
·· δæ
= M
αβ
δæ
αβ
= −M
αβ
∂
2
δw
∂x
α
x
β
.
Using this and the Gauss–Ostrogradsky theorem, we get
−
Σ
M
αβ
∂
2
δw
∂x
α
x
β
dΣ =
Σ
∂M
αβ
∂x
α
∂ δw
∂x
β
dΣ −
ω
ν
α
M
αβ
∂ δw
∂x
β
ds
= −
Σ
∂
2
M
αβ
∂x
α
∂x
β
δwdΣ +
ω
ν
β
∂M
αβ
∂x
α
δwds
−
ω
ν
α
M
αβ
∂ δw
∂x
β
ds
= −
Σ
¯
∇· (
¯
∇· (M×n)) δwdΣ
+
ω
ν · (
¯
∇· (M×n)) δwds
−
ω
ν · (M×n) ·
¯
∇δwds, (7.67)
where ν
α
= ν · i
α
and s is the length parameter over ω.
Let us consider the last integral in (7.67). On ω the expression
¯
∇δw
can be represented as
¯
∇δw = ν
∂ δw
∂ν
+τ
∂ δw
∂τ
,
where τ = n × ν is the vector tangent to ω. Using the contour length
parameter s over ω, we get
∂ δw
∂τ
=
∂ δw
∂s
.
274 Tensor Analysis with Applications in Mechanics
Note that on ω the expressions ∂ δw/∂ν and δw are independent. The
expressions ∂ δw/∂s and δw are dependent because the value ∂ δw/∂s is
uniquely deﬁned when δw is known on ω, whereas ∂ δw/∂ν is not deﬁned
by the values of δw on ω. To exclude the derivative ∂ δw/∂s, we integrate
by parts in the integral over the closed contour ω:
ω
f
∂g
∂s
ds = −
ω
g
∂f
∂s
ds,
where f and g are arbitrary functions of s. Applying this to the last integral
in (7.67), we get
−
ω
ν · (M×n) ·
¯
∇δwds = −
ω
¸
ν · (M×n) · ν
∂ δw
∂ν
−
∂
∂s
(ν · (M×n) · τ) δw
ds.
On ω
1
we have
δw = 0 =
∂ δw
∂ν
,
hence
δE
w
= −
Σ
¯
∇· (
¯
∇· (M×n)) +q · n
δwdΣ
+
ω
2
¸
ν ·
¯
∇· (M×n)
−
∂
∂s
ν · (M×n) · τ
−ϕ
0
n
δwds
+
ω
2
ν · (M×n) · ν −
0
∂ δw
∂ν
ds. (7.68)
Using a standard procedure in the calculus of variations (see, e.g.,
[
Lebe
dev and Cloud (2003)
]
), we can show that the equation δE
w
= 0, which
holds for all admissible δw, implies both the equilibrium equation for the
plate
¯
∇·
¯
∇· (M×n)
+q · n = 0 in Σ
and the static boundary conditions
ν ·
¯
∇· (M×n)
−
∂
∂s
ν · (M×n) · τ
−ϕ
0
n
= 0
and
ν · (M×n) · ν −
0
= 0 on ω
2
.
Linear Elastic Shells 275
This derivation also indicates that we have introduced the work for the
transverse load correctly: a stationary point of E
w
satisﬁes (7.53) on Σ and
the boundary conditions (7.64) on ω
2
.
Now we can prove the converse statement: a solution w of the boundary
value problem (7.53), (7.64) is a stationary point of E
w
. Multiplying (7.53)
by an admissible δw, integrating appropriately by parts and therefore doing
the above transformations in the reverse order, we arrive at the equation
δE
w
= 0. This completes the proof.
In a manner similar to the proof of minimality in the Lagrange principle
in elasticity, we can show minimality of a stationary point in the bending
problem as well. The reader is encouraged to produce a complete proof.
Note that in the proof for threedimensional elasticity, the only thing that
mattered was the structure of the total energy functional: it is a sum
of quadratic and linear functionals, and its quadratic portion is positive
deﬁnite.
Note that by considering a plate as a threedimensional body, we obtain
a system of simultaneous equations in displacements of the second order,
accompanied by three boundary conditions at each point of the boundary.
Transformation of the problem to a twodimensional plate problem brings
us again to three diﬀerential equations in displacements, but the equation
for w is of fourth order; it is supplemented with two conditions at each point
of the plate edge, and the other two equations are of second order. The
increase in order of the system is the cost of reducing a threedimensional
problem to a twodimensional one.
In shell theory there are other variational principles. Their importance
in this theory is even more than in elasticity. They are the basis for nu
merical methods and investigations of qualitative questions of the theory.
Moreover, variational principles are used to construct various versions of
shell theory. For example, we can introduce the type of distribution of
the stresses and displacements along the thickness, and derive the equi
librium equations, ﬁnd admissible sets of boundary conditions, and justify
wellposedness of the corresponding boundary value problems.
Uniqueness of solution
For an equilibrium problem in plate theory, a solution is unique in the same
meaning as for a problem in threedimensional elasticity. That is, if some
part of the boundary contour is clamped, then the solution is truly unique.
276 Tensor Analysis with Applications in Mechanics
When the plate can move as a rigid body, the solution is uniquely deﬁned
by the equations and boundary conditions up to a small rigid motion. We
will brieﬂy outline a proof.
Suppose to the contrary the existence of two solutions v
1
and v
2
to
the boundary value problem (7.60), (7.65), (7.53), (7.64). The diﬀerence
v = v
2
− v
1
is a solution of this homogeneous boundary value problem.
It is easy to understand that it should be a stationary point of the total
energy functional
Σ
U dΣ.
This implies that
δ
Σ
U dΣ = 0.
But U is a homogeneous quadratic form, so putting δv = v in δU we get
δU = 2U(v) and thus
2
Σ
U dΣ = 0.
But earlier we considered the question of when the strain energy functional
takes zero value: the equation
Σ
U dΣ = 0
implies that v is a small rigid displacement of the plate. So the unique
ness of a solution for an equilibrium problem is demonstrated: a solution
is uniquely deﬁned up to a rigid displacement if there are no kinematic
boundary constraints.
Exercise 7.14. Derive (7.50) from (7.49).
Exercise 7.15. Show that tr ε =
¯
∇· ˜ v −z
¯
∇·
¯
∇w.
Exercise 7.16. Let w and its ﬁrst and second derivatives be small. Demon
strate that æ is the curvature tensor of the bent surface of the plate. Use
the results of Exercise 5.33.
Exercise 7.17. Derive (7.58).
Exercise 7.18. Demonstrate the variational principle for tangential dis
placements in the theory of plates.
Exercise 7.19. Show that at a point of plate equilibrium, E
w
takes its
minimum value.
Linear Elastic Shells 277
Remarks
Kirchhoﬀ’s plate theory was extended to shell theory. In shell theory,
Kirchhoﬀ’s hypotheses are called the Kirchhoﬀ–Love hypotheses. In the
Kirchhoﬀ–Love shell theory, the equations are written on the midsurface.
The strain energy density for the shell also splits into two parts, for tangen
tial and bending deformations. However, unlike the splitting of the plate
energy, these terms are dependent. Moreover, all three equilibrium equa
tions in displacements contain all components of the displacement vector.
Technically, the Kirchhoﬀ–Love shell theory is more complicated than plate
theory, but qualitatively they are quite similar. In shell theory, we have
Lagrange’s variational principle and uniqueness of solution to a boundary
value problem. In both cases, we can also prove existence of a solution, but
in shell theory this requires more advanced mathematical tools.
7.9 On NonClassical Theories of Plates and Shells
Reissner’s approach to plate and shell theory
We recall that Kirchhoﬀ’s hypotheses yield only an approximation to the
real deformations of thinwalled bodies. In the mechanics of plates and
shells, there are other approaches to the representation of threedimensional
deformation, where shear deformation, normal extension, and other factors
are taken into account. In particular, the Reissner and Mindlin approaches
allow us to construct more precise twodimensional models for the solution
of a threedimensional elastic problem for a shell. The main features of the
theory will be demonstrated using Reissner’s plate equations.
Unlike (7.54), in Reissner’s approach we take the displacement ﬁeld to
be of the form
u(x
1
, x
2
, z) = v(x
1
, x
2
) −zϑ(x
1
, x
2
)
= ˜ v +wn −zϑ, ϑ · n = 0. (7.69)
The components of ϑ can be interpreted as the rotation angles of the shell
cross section. In the general case, unlike in the Kirchhoﬀ theory, ϑ =
¯
∇w
and ϑ is considered to be independent of v. Thus, in Reissner’s theory
there are ﬁve unknown variables: the components v
1
, v
2
, w of v and the
components ϑ
1
, ϑ
2
of ϑ. In Reissner’s theory, the transverse shear stress
resultants are on equal footing with the stress resultants and the bending
and twisting stress couples. Unlike what happens in Kirchhoﬀ’s model, they
278 Tensor Analysis with Applications in Mechanics
are independent and so it is necessary to formulate additional constitutive
equations along the lines of (7.47). The equilibrium equations for Reissner’s
plate take the same form as in Kirchhoﬀ’s theory, i.e., (7.45):
¯
∇· T+ q = 0,
¯
∇· M+T
×
+m = 0, (7.70)
where T and M are the stress resultant tensor and the stress couple tensor,
respectively, and q and m represent the forces and couples distributed over
Σ. Note that m· n = 0. In the component representation, they coincide
with (7.50) and (7.51).
The dynamic equations of the theory are
¯
∇· T+q = ρ¨ v +ρΘ
1
·
¨
ϑ,
¯
∇· M+T
×
+m = ρΘ
T
1
·· ¨ v +ρΘ
2
·
¨
ϑ, (7.71)
where ρ is the surface shell density, Θ
1
and Θ
2
are the inertia tensors, and
Θ
T
2
= Θ
2
.
The boundary conditions for the plate are
v
ω
1
= v
0
(s), ϑ
ω
3
= ϑ
0
(s), (7.72)
ν · T
ω
2
= ϕ(s), ν · M
ω
4
= (s), (7.73)
where v
0
(s) and ϑ
0
(s) are given vector functions of the length parameter
s such that ϑ
0
· n = 0. They deﬁne the displacements and the rotation on
some part of the boundary contour, respectively. A given ϕ(s) and (s)
would deﬁne the stress resultants and the stress couples acting on the rest
of the plate edge, · n = 0.
In Reissner’s plate theory, the constitutive equations are
T· A = C·· µ, T· n = Γ · γ, M = D·· κ. (7.74)
Here T · A is the inplane stress resultant tensor, T · n represents the
transverse shear stress resultants, and M is the stress couple tensor. The
strain measures are denoted as follows: µ is the inplane strain tensor, γ
is the vector of transverse shear strains, and κ is the tensor of outofplane
strains. Their deﬁnitions are
µ =
1
2
¯
∇˜ v + (
¯
∇˜ v)
T
, γ =
¯
∇w −n ×ϑ, κ =
¯
∇ϑ.
The remaining notation is for the fourthorder tensors C and D and the
secondorder tensor Γ that describe the eﬀective stiﬀness properties of the
Linear Elastic Shells 279
plate; they depend on the material properties of the plate and on the cross
section geometry. In the case of an isotropic plate having properties sym
metric about the midplane, the eﬀective stiﬀness tensors take the form
C = C
11
AA+C
22
(A
2
A
2
+A
4
A
4
),
D = D
22
(A
2
A
2
+A
4
A
4
) +D
33
A
3
A
3
,
Γ = ΓA,
where
A = i
1
i
1
+i
2
i
2
, A
2
= i
1
i
1
−i
2
i
2
,
A
3
= i
1
i
2
−i
2
i
1
, A
4
= i
1
i
2
+i
2
i
1
,
and i
1
and i
2
are the unit basis vectors with i
1
· i
2
= 0 on the midplane.
Denote A
1
= A. The following orthogonality relation can be obtained:
1
2
A
i
·· A
j
= δ
ij
(i, j = 1, 2, 3, 4).
For an isotropic homogeneous plate, the components of the stiﬀness
tensors are deﬁned as follows:
C
11
=
Eh
2(1 −ν)
, C
22
=
Eh
2(1 +ν)
= µh,
D
33
=
Eh
3
24(1 −ν)
, D
22
=
Eh
3
24(1 +ν)
=
µh
3
12
.
The classical bending stiﬀness for a plate is
D = D
33
+D
22
=
Eh
3
12(1 −ν
2
)
. (7.75)
The surface density and the inertia tensors are
ρ = ρ
0
h, Θ
1
= 0, Θ
2
= ΘA, Θ =
ρ
0
h
3
12
, (7.76)
where ρ
0
is the plate material density.
The transverse shear stiﬀness is given by
Γ = kµh, (7.77)
where k is a shear correction factor ﬁrst introduced by Timoshenko in the
theory of beams
2
. For the value of k, Mindlin
3
proposed k = π
2
/12 while
2
Timoshenko, S. P. On the correction for shear of the diﬀerential equation for transverse
vibrations of prismatic bars. Phil. Mag. Ser. 6, 41, 744–746, 1921.
3
Mindlin, R.D. Inﬂuence of rotatory inertia and shear on ﬂexural motions of isotropic,
elastic plates. Trans. ASME J. Appl. Mech., 18, 31–38, 1951.
280 Tensor Analysis with Applications in Mechanics
Reissner
4
proposed a similar value of k = 5/6. The literature mentions
other values of k; for example, for plates that are strongly nonhomogeneous
in thickness, k can signiﬁcantly diﬀer from the above values
5
.
The strain energy density for Reissner’s plate is given by
U =
1
2
µ ·· C·· µ +
1
2
κ ·· D·· κ +
1
2
γ · Γ · γ.
It is seen that
T· A =
∂U
∂µ
, T· n =
∂U
∂γ
, M =
∂U
∂κ
.
The ﬁve scalar equations of (7.50) and (7.51) contain ﬁve scalar un
knowns v
1
, v
2
, w, ϑ
1
, and ϑ
2
. In Reissner’s model, equation (7.59) for w
changes to the following:
D
¯
∇
4
w = q
n
−
D
Γ
∆q
n
. (7.78)
Because of the additional term (D/Γ)∆q
n
, for strongly nonhomogeneous
loads the results obtained from Reissner’s theory can diﬀer signiﬁcantly
from those obtained from Kirchhoﬀ’s model.
For advanced readers, we propose
Exercise 7.20. Derive (7.78).
Reissner’s theory is usually used for nonthin plates, for dynamic prob
lems, and for cases of anisotropic materials having small shear stiﬀness. A
more detailed presentation of the theory can be found in
[
Wang, Reddy, and
Lee (2000); Zhilin (2006)
]
; see also Altenbach
6
and Grigolyuk and Selezov
7
.
Plate and shell theories of higher order
The approximation of the displacement ﬁeld (7.69) is linear in z. In the
literature on shell theory, we encounter displacement approximations in z of
4
Reissner, E. On the theory of bending of elastic plates. J. Math. Physics, 23, 184–194,
1944.
5
Altenbach, H., and Eremeyev, V.A. Direct approach based analysis of plates composed
of functionally graded materials. Arch. Appl. Mech., 78, 775–794, 2008.
6
Altenbach, H. An alternative determination of transverse shear stiﬀnesses for sandwich
and laminated plates. Int. J. Solids Struct., 37, 3503–3520, 2000. Altenbach, H. On the
determination of transverse shear stiﬀnesses of orthotropic plates. ZAMP, 51, 629–649,
2000.
7
Grigolyuk, E.I., and Selezov, I.T. Nonclassical theories of vibration of beams, plates
and shells (in Russian). In: Itogi nauki i tekhniki, Mekhanika tverdogo deformiruemogo
tela, vol 5, VINITI, Moskva, 1973.
Linear Elastic Shells 281
higher order. The natural idea of expanding the displacement in a series of
powers in z belongs to Cauchy and Poisson. We could say that Germain’s
equation for plate deﬂection represents the zeroorder theory, as here the
deﬂection ﬁeld does not depend on z. Equation (7.69) can be considered as
a basis of the ﬁrstorder theory. Some versions of higherorder theory were
developed in a number of works; certain diﬃculties arise, however, such
as how one should deﬁne the boundary conditions and physically interpret
moments of higher order.
As an example, we will consider the thirdorder plate theory. In
[
Wang,
Reddy, and Lee (2000)
]
there is a review of the literature on this version of
the theory. The displacement ﬁeld is approximated by the expressions
u(x
1
, x
2
, z) = v(x
1
, x
2
) +zφ(x
1
, x
2
) +z
3
ψ(x
1
, x
2
)
= ˜ v +wn +zφ +z
3
ψ, (7.79)
where ˜ v·n = 0, and φ and ψ are functions such that φ·n = 0 and ψ·n = 0.
In particular, Reddy proposed the following expression for ψ:
ψ = −α
φ +
¯
∇w
,
where α = 4/(3h
2
) and h is the plate thickness. The strain tensor becomes
ε = ˜ ε +
1
2
(γn +nγ) , (7.80)
where
˜ ε = +
z
2
¯
∇φ + (
¯
∇φ)
T
+
z
3
2
¯
∇ψ + (
¯
∇ψ)
T
= +
z
2
¯
∇φ + (
¯
∇φ)
T
−α
z
3
2
¯
∇φ + (
¯
∇φ)
T
+ 2
¯
∇
¯
∇w
,
γ = φ + 3z
2
ψ = φ −3z
2
αφ −3z
2
α
¯
∇w,
2 =
¯
∇˜ v + (
¯
∇˜ v)
T
.
In the thirdorder theory, we introduce the stress resultants T
αβ
, the
transverse shear stress resultants T
αn
, and the stress couples M
αβ
. But
we introduce some additional quantities such as the higherorder stress
282 Tensor Analysis with Applications in Mechanics
resultants P
αβ
and R
α
:
T
αβ
= [[σ
αβ
]] ≡
h/2
−h/2
σ
αβ
dz, T
αn
= [[σ
αn
]] ≡
h/2
−h/2
σ
α3
dz,
M
αβ
= [[zσ
αβ
]] ≡
h/2
−h/2
zσ
αβ
dz, (7.81)
P
αβ
= [[z
3
σ
αβ
]] ≡
h/2
−h/2
z
3
σ
αβ
dz, R
α
= [[z
2
σ
αn
]] ≡
h/2
−h/2
z
2
σ
α3
dz.
Exercise 7.21. Using (7.80) and (7.56), express (7.81) in terms of the
components of φ and w.
The equilibrium equations split into two systems. The ﬁrst is for the
plane state; it is of the form we had in Kirchhoﬀ’s theory:
∂T
11
∂x
1
+
∂T
21
∂x
2
+q
1
= 0,
∂T
12
∂x
1
+
∂T
22
∂x
2
+q
2
= 0. (7.82)
Equations (7.82) are supplemented with the boundary conditions
˜ v
ω
1
= ˜ v
0
(s), ν · T· A
ω
2
= ϕ
0
(s), (7.83)
where
n · ˜ v
0
= 0, n · ϕ
0
= 0, ω
1
∪ ω
2
= ω, ω
1
∩ ω
2
= ∅.
The constitutive equations are
T
11
=
Eh
1 −ν
2
(
11
+ν
22
),
T
22
=
Eh
1 −ν
2
(
22
+ν
11
),
T
12
=
Eh
1 +ν
12
.
The boundary value problem for plate bending is presented by the equi
librium equations
∂T
1n
∂x
1
+
∂T
2n
∂x
2
+α
∂
2
P
11
∂x
2
1
+ 2
∂
2
P
12
∂x
1
∂x
2
+
∂
2
P
22
∂x
2
2
+q
n
= 0,
∂
´
M
11
∂x
1
+
∂
´
M
21
∂x
2
−
´
T
1n
= 0,
∂
´
M
12
∂x
1
+
∂
´
M
22
∂x
2
−
´
T
2n
= 0, (7.84)
Linear Elastic Shells 283
where
´
M
αβ
= M
αβ
−αP
αβ
,
´
T
αn
= T
αn
−3αR
α
.
The kinematic boundary conditions for the bending problem consist of four
scalar equations:
w
ω
= w
0
(s),
∂w
∂n
ω
= w
0
n
(s), φ
ω
= φ
0
(s),
where n· φ
0
= 0. The static boundary conditions are expressed in terms of
stress resultants, stress couples, and moments of higher order; the interested
reader should consult the original sources
[
Wang, Reddy, and Lee (2000)
]
and Kienzler
8
, Levinson
9
, and Reddy
10
.
Note that the third and higherorder shell and plate theories show why
variational formulations are needed. These are the only way to introduce
correct static boundary conditions.
Micropolar shells or 6
th
parametrical shell theory
As an example of a nonclassical shell theory, we will sketch the theory of
micropolar shells. Its roots originate in the work of the Cosserat brothers,
Eug`en and Fran¸cois
[
Cosserat and Cosserat (1909)
]
. The micropolar theory
of shells is presented in the works of Eremeyev and Zubov
[
Eremeyev and
Zubov (2008)
]
, Zhilin
[
Zhilin (2006)
]
, and others.
In this theory, the dynamical equations and shell kinematics coincide
with those of a 6parametric shell theory, a nonlinear version of which is
presented in the books by Libai and Simmonds
[
Libai and Simmonds (1998)
]
and Chr´ o´scielewski et al
[
Chr´ o´scielewski, Makowski, and Pietraszkiewicz
(2004)
]
. There exists a micropolar plate theory by Eringen
[
Eringen (1999)
]
.
It is based on integration over the thickness of the plate when its material
is considered as a Cosserat continuum. This diﬀers from 6parametrical
shellplate theory, as it contains more than 6 unknown scalar functions.
The kinematics of the shell surface are described by six scalar quantities,
three of which are the components of the displacement vector v and the
rest of which are the components of the microrotation vector ϑ. So a
shell particle has six degrees of freedom described by the components of v
8
Kienzler, R. On consistent plate theories. Arch. Appl. Mech., 72, 229–247, 2002.
9
Levinson, M. An accurate, simple theory of the statics and dynamics of elastic plates.
Mech. Res. Comm., 7, 343–350, 1980.
10
Reddy, J.N. A simple higherorder theory for laminated composite plates. Trans.
ASME J. Appl. Mech., 51, 745–752, 1984.
284 Tensor Analysis with Applications in Mechanics
and ϑ. The vector ϑ satisﬁes ϑ · n = 0; the vectors v and ϑ are mutually
independent. This is analogous to a twodimensional version of the Cosserat
medium with couple stresses and the rotation interaction of the particles.
For a micropolar shell, we can assign the couple load acting on the shell
surface. Here the order of the equilibrium equations is 12, so we should
supplement them with 6 conditions on the shell edge. On the part of the
edge that is free of geometrical constraints, we can assign the distribution
of forces and couples. This extends the range of applicability of the the
ory in comparison with the theories considered above: we can, say, assign
boundary conditions that describe a shell clamped to a rigid body.
The micropolar theory is used, in particular, to describe branching shells
and thinwalled bodies with complex intrinsic structure that include mul
tilayered or composite plates and shells, shells with inner partition, with
stringers and those similar to honeycomb or made of highly porous mate
rials.
We present the micropolar shell equations for small deformation. The
dynamical equations are
¯
∇· T+q = ρ¨ v +ρΘ
1
·
¨
ϑ,
¯
∇· M+T
×
+m = ρΘ
T
1
··· ¨ v +ρΘ
2
·
¨
ϑ, (7.85)
where T and M are surface stress and couple stress tensors analogous to
those in Kirchhoﬀ’s theory, ρ is the shell surface density, Θ
1
and Θ
2
are the
inertia tensors, Θ
2
is symmetric, Θ
T
2
= Θ
2
, and q and m are distributed
surface forces and couples, respectively. Besides, in this theory, we can
assign a twisting couple (or drilling moment) on the shell surface. The
equilibrium equations take the form
¯
∇· T+ q = 0,
¯
∇· M+T
×
+m = 0. (7.86)
The conditions at a boundary point can be kinematic
v
ω
1
= v
0
(s), ϑ
ω
3
= ϑ
0
(s), (7.87)
or static
ν · T
ω
2
= ϕ(s), ν · M
ω
4
= (s), (7.88)
where v
0
(s) and ϑ
0
(s) are given vector functions of the length parameter
s; they deﬁne the displacements and microrotation of the shell edge. The
functions ϕ(s) and (s) determine the surface stresses and stress couples
on the edge. Here
ω = ω
1
∪ ω
2
= ω
3
∪ ω
4
, ω
2
= ω \ ω
1
, ω
4
= ω \ ω
3
.
Linear Elastic Shells 285
The form of the equilibrium and dynamical equations for a micropolar shell
does not diﬀer from (7.70) and (7.71). In general, in this theory m· n = 0
and · n = 0. In particular, M· n = 0.
The strain measures take the form
=
¯
∇v +A×ϑ, κ =
¯
∇ϑ, (7.89)
where and κ are nonsymmetric strain and bending strain tensors respec
tively.
The constitutive equations for an elastic shell are represented though
the strain energy density U = U(, κ):
T =
∂U
∂
, M =
∂U
∂κ
. (7.90)
For an isotropic shell, U is a quadratic form in its variables:
2U = α
1
tr
2
+α
2
tr ˜
2
+α
3
tr
˜ · ˜
T
+α
4
n ·
T
· · n
+β
1
tr
2
κ +β
2
tr ˜ κ
2
+β
3
tr
˜ κ · ˜ κ
T
+β
4
n · κ
T
· κ · n,
˜ = · A, ˜ κ = κ · A, (7.91)
where α
k
and β
k
(k = 1, 2, 3, 4) are elastic moduli. Substituting (7.91) into
(7.90), we get
T = α
1
Atr +α
2
˜
T
+α
3
˜ +α
4
( · n)n,
M = β
1
Atr κ +β
2
˜ κ
T
+β
3
˜ κ +β
4
(κ · n)n. (7.92)
Equations (7.86)–(7.88) and (7.92) constitute a linear boundary value
problem with respect to the ﬁelds of displacement and microrotation; they
describe the equilibrium of a micropolar shell in the case of small defor
mation. For dynamic problems, equations (7.86) change to (7.85). Under
some additional assumptions, Reissner’s shell theory is a consequence of
the micropolar theory.
As in the other shell theories, we can formulate some variational princi
ples. Lagrange’s variational principle for an elastic micropolar shell starts
with formulation of the total energy functional
E(v, ϑ) =
Σ
U(, κ) dΣ −A(v, ϑ), (7.93)
where the potential of external loads A(v, ϑ) is
A(v, ϑ) =
Σ
(q · v +m· ϑ) dΣ +
ω
2
ϕ· vds +
ω
4
· ϑds.
286 Tensor Analysis with Applications in Mechanics
The functional E(v, ϑ) is considered on the set of twice continuously diﬀer
entiable ﬁelds of displacements and microrotations that satisfy (7.87). The
pair (v, ϑ) that satisﬁes (7.86) and (7.88) is a stationary point of E(v, ϑ).
Lagrange’s stationary principle is minimal: on the equilibrium solution, the
functional (7.93) attains its minimum.
Rayleigh’s variational principle now takes the following form.
On the set of functions with boundary conditions v
ω
1
= 0, ϑ
ω
3
= 0 that
obey the constraint
K(v
◦
, ϑ
◦
) ≡
Σ
ρ
1
2
v
◦
· v
◦
+v
◦
· Θ
1
· ϑ
◦
+
1
2
ϑ
◦
· Θ
2
· ϑ
◦
dΣ = 1,
the eigenoscillation modes of the shell are stationary points of the strain
energy functional
E(v
◦
, ϑ
◦
) =
Σ
U(
◦
, κ
◦
) dΣ, (7.94)
where
◦
=
¯
∇v
◦
+A×ϑ
◦
and κ
◦
=
¯
∇ϑ
◦
.
Rayleigh’s principle also includes the reverse statement: on the set of
functions satisfying the above restrictions, the stationary points of E are
the modes of eigenoscillation. Here v
◦
and ϑ
◦
are the amplitudes of the
oscillations of the displacements and microrotations as the solutions of the
eigenoscillation problem are sought in the form v = v
◦
e
iωt
, ϑ = ϑ
◦
e
iωt
.
Rayleigh’s quotient takes the form
R(v
◦
, ϑ
◦
) =
E(v
◦
, ϑ
◦
)
K(v
◦
, ϑ
◦
)
.
Now the lowest eigenfrequency of the shell, which is the minimal frequency,
is equal to the minimum value of the functional R.
Finally, we would like to repeat that, contrary to the situation with lin
ear elasticity, shell theory is still under development; investigators continue
to seek better models and numerical methods for the solution of practical
problems. This is a consequence of the evident fact that it is impossible to
ﬁnd an accurate approximation to the results of a threedimensional elastic
ity problem under all the circumstances when one solves the same problem
using one of the twodimensional models of shell theory.
Appendix A
Formulary
For convenience we list the main formulas obtained in each chapter. The
symbol ∀ denotes the universal quantiﬁer (read as “for all” or “for every”).
Chapter 1
Dot product
a · b = ab cos θ = a
1
b
1
+ a
2
b
2
+ a
3
b
3
Cross product
a ×b =
i
1
i
2
i
3
a
1
a
2
a
3
b
1
b
2
b
3
Scalar triple product
a · (b ×c) =
a
1
a
2
a
3
b
1
b
2
b
3
c
1
c
2
c
3
287
288 Tensor Analysis with Applications in Mechanics
Chapter 2
Reciprocal (dual) basis
Kronecker delta symbol
δ
i
j
=
1 j = i
0 j = i
Deﬁnition of reciprocal basis
e
j
· e
i
= δ
i
j
Components of a vector
x = x
i
e
i
= x
i
e
i
x
i
= x · e
i
x
i
= x · e
i
Relations between dual bases
e
i
=
1
V
(e
j
×e
k
) e
i
=
1
V
(e
j
×e
k
)
where
(i, j, k) = (1, 2, 3) or (2, 3, 1) or (3, 1, 2)
and
V = e
1
· (e
2
×e
3
)
V
= e
1
· (e
2
×e
3
)
V
= 1/V
Metric coeﬃcients
g
jq
= e
j
· e
q
g
ip
= e
i
· e
p
g
ij
g
jk
= δ
k
i
In Cartesian frames,
g
ij
= δ
j
i
g
ij
= δ
i
j
Dot products in mixed and unmixed bases
a · b = a
i
b
j
g
ij
= a
i
b
j
g
ij
= a
i
b
i
= a
i
b
i
Formulary 289
Raising and lowering of indices
x
j
= x
i
g
ij
x
i
= x
j
g
ij
Frame transformation
Equations of transformation
e
i
= A
j
i
˜ e
j
A
j
i
= e
i
· ˜ e
j
˜ e
i
=
˜
A
j
i
e
j
˜
A
j
i
= ˜ e
i
· e
j
where
˜
A
j
i
A
k
j
= A
j
i
˜
A
k
j
= δ
k
i
Vector components and transformation laws
x = x
i
e
i
= x
i
e
i
= ˜ x
i
˜ e
i
= ˜ x
i
˜ e
i
and
˜ x
i
= A
i
j
x
j
x
i
=
˜
A
i
j
˜ x
j
˜ x
i
=
˜
A
j
i
x
j
x
i
= A
j
i
˜ x
j
Miscellaneous
Permutation (Levi–Civita) symbol
ijk
= e
i
· (e
j
×e
k
) =
⎧
⎪
⎪
⎨
⎪
⎪
⎩
+V (i, j, k) an even permutation of (1, 2, 3)
−V (i, j, k) an odd permutation of (1, 2, 3)
0 two or more indices equal
ijk
= e
i
· (e
j
×e
k
) =
⎧
⎪
⎪
⎨
⎪
⎪
⎩
+V
(i, j, k) an even permutation of (1, 2, 3)
−V
(i, j, k) an odd permutation of (1, 2, 3)
0 two or more indices equal
290 Tensor Analysis with Applications in Mechanics
Useful identities
ijk
pqr
=
δ
p
i
δ
q
i
δ
r
i
δ
p
j
δ
q
j
δ
r
j
δ
p
k
δ
q
k
δ
r
k
ijk
pqk
= δ
p
i
δ
q
j
−δ
q
i
δ
p
j
Determinant of Gram matrix
V
2
= det[g
ij
]
Cross product
a ×b = e
i
ijk
a
j
b
k
= e
i
ijk
a
j
b
k
Chapter 3
Dyad product
Properties
(λa) ⊗b = a ⊗(λb) = λ(a ⊗b)
(a +b) ⊗c = a ⊗c +b ⊗c
a ⊗(b +c) = a ⊗b +a ⊗c
Dot products of dyad with vector
ab · c = (b · c)a
c · (ab) = (c · a)b
Tensors from operator viewpoint
Equality of tensors
A = B ⇐⇒ ∀x, A· x = B· x
Formulary 291
Components
a
ij
= e
i
· A· e
j
a
ij
= e
i
· A· e
j
a
i
·j
= e
i
· A· e
j
a
·j
i
= e
i
· A· e
j
Deﬁnition of sum A+B
∀x, (A+B) · x = A· x +B· x
Deﬁnition of scalar multiple cA
∀x, (cA) · x = c(A· x)
Deﬁnition of dot product A· B
(A· B) · x = A· (B· x)
Deﬁnition of premultiplication y · A
∀x, (y · A) · x = y · (A· x)
Deﬁnition of unit tensor E
∀x, E· x = x · E = x
Unit tensor components
E = e
i
e
i
= e
j
e
j
= g
ij
e
i
e
j
= g
ij
e
i
e
j
Inverse tensor
A· A
−1
= E
(A· B)
−1
= B
−1
· A
−1
292 Tensor Analysis with Applications in Mechanics
Nonsingular tensor A
A· x = 0 =⇒ x = 0
Determinant of a tensor
det A =
a
·j
i
=
a
k
· m
=
1
g
a
st
 = g a
pq

=
1
6
ijk
mnp
a
· i
m
a
· j
n
a
· k
p
Dyadic components under transformation
Transformation to reciprocal basis
a
km
= a
ij
g
ki
g
jm
More general transformation
e
i
= A
j
i
˜ e
j
=⇒ ˜ a
ij
= a
km
A
i
k
A
j
m
˜ e
i
=
˜
A
j
i
e
j
=⇒ a
ij
= ˜ a
km
˜
A
i
k
˜
A
j
m
A
k
j
˜
A
i
k
= δ
i
j
A = ˜ a
ij
˜ e
i
˜ e
j
= ˜ a
kl
˜ e
k
˜ e
l
= ˜ a
· j
i
˜ e
i
˜ e
j
= ˜ a
k
· l
˜ e
k
˜ e
l
= a
ij
e
i
e
j
= a
kl
e
k
e
l
= a
·j
i
e
i
e
j
= a
k
· l
e
k
e
l
where
˜ a
ij
= A
i
k
A
j
l
a
kl
a
ij
=
˜
A
i
k
˜
A
j
l
˜ a
kl
˜ a
ij
=
˜
A
k
i
˜
A
l
j
a
kl
a
ij
= A
k
i
A
l
j
˜ a
kl
˜ a
i
· j
= A
i
k
˜
A
l
j
a
k
· l
a
i
·j
=
˜
A
i
k
A
l
j
˜ a
k
· l
˜ a
· j
i
=
˜
A
k
i
A
j
l
a
· l
k
a
·j
i
= A
k
i
˜
A
j
l
˜ a
· l
k
Formulary 293
More dyadic operations
Dot product
ab · cd = (b · c)ad
A· (λa + µb) = λA· a + µA· b
(λA+ µB) · a = λA· a + µB· a
Double dot product
ab ·· cd = (b · c)(a · d)
Scalar product of secondorder tensors
ab • cd = (a · c)(b · d)
Secondorder tensor topics
Transpose
A
T
= a
ji
e
i
e
j
= a
ji
e
i
e
j
= a
j
·i
e
i
e
j
= a
·i
j
e
i
e
j
A· x = x · A
T
(A
T
)
T
= A
(A· B)
T
= B
T
· A
T
a · C
T
· b = b · C· a
det A
−1
= (det A)
−1
(A· B)
−1
= B
−1
· A
−1
(A
T
)
−1
= (A
−1
)
T
(A
−1
)
−1
= A
294 Tensor Analysis with Applications in Mechanics
Tensors raised to powers
A
0
= E A
n
= A· A
n−1
for n = 1, 2, 3, . . .
A
−n
= A
−n+1
· A
−1
for n = 2, 3, 4, . . .
e
A
= E+
A
1!
+
A
2
2!
+
A
3
3!
+· · ·
Symmetric and antisymmetric tensors
symmetric: A = A
T
; ∀x, A· x = x · A
antisymmetric: A = −A
T
A =
1
2
A+A
T
+
1
2
A−A
T
Eigenpair
A· x = λx
Vi`ete formulas for invariants
I
1
(A) = λ
1
+ λ
2
+ λ
3
= tr A
I
2
(A) = λ
1
λ
2
+ λ
1
λ
3
+ λ
2
λ
3
=
1
2
[tr
2
A−tr A
2
]
I
3
(A) = λ
1
λ
2
λ
3
= det A
Orthogonal tensor
Q· Q
T
= Q
T
· Q = E
Polar decompositions
A = S · Q Q orthogonal
A = Q· S
S, S
positive deﬁnite and symmetric
Formulary 295
Chapter 4
Vector ﬁelds
Some rules for diﬀerentiating vector functions
d(e
1
(t) +e
2
(t))
dt
=
de
1
(t)
dt
+
de
2
(t)
dt
d(c e(t))
dt
= c
de(t)
dt
d(f(t)e(t))
dt
=
df(t)
dt
e(t) + f(t)
de(t)
dt
d
dt
(e
1
(t) · e
2
(t)) = e
1
(t) · e
2
(t) +e
1
(t) · e
2
(t)
d
dt
(e
1
(t) ×e
2
(t)) = e
1
(t) ×e
2
(t) +e
1
(t) ×e
2
(t)
and
d
dt
[e
1
(t) , e
2
(t) , e
3
(t)] = [e
1
(t) , e
2
(t) , e
3
(t)]
+ [e
1
(t) , e
2
(t) , e
3
(t)]
+ [e
1
(t) , e
2
(t) , e
3
(t)]
where
[e
1
(t) , e
2
(t) , e
3
(t)] = (e
1
(t) ×e
2
(t)) · e
3
(t)
Tangent vectors to coordinate lines
r
i
=
∂r
∂q
i
(i = 1, 2, 3)
Jacobian
√
g = r
1
· (r
2
×r
3
) =
∂x
i
∂q
j
Pointwise deﬁnition of reciprocal basis
r
i
· r
j
= δ
i
j
296 Tensor Analysis with Applications in Mechanics
Deﬁnition of metric coeﬃcients
g
ij
= r
i
· r
j
g
ij
= r
i
· r
j
g
j
i
= r
i
· r
j
= δ
j
i
Transformation laws
r
i
= A
j
i
˜r
j
A
j
i
=
∂˜ q
j
∂q
i
˜r
i
=
˜
A
j
i
r
j
˜
A
j
i
=
∂q
j
∂˜ q
i
˜
f
i
= A
i
j
f
j
˜
f
i
=
˜
A
j
i
f
j
f
i
=
˜
A
i
j
˜
f
j
f
i
= A
j
i
˜
f
j
Diﬀerentials and the nabla operator
Metric forms
(ds)
2
= dr · dr = g
ij
dq
i
dq
j
Nabla operator
∇ = r
i
∂
∂q
i
Gradient of a vector function
df = dr · ∇f = ∇f
T
· dr
Divergence of vector
div f = ∇· f = r
i
·
∂f
∂q
i
Rotation and curl of vector
rot f = ∇×f = r
i
×
∂f
∂q
i
ω =
1
2
rot f
Formulary 297
Divergence and rotation of secondorder tensor
∇· A = r
i
·
∂
∂q
i
A ∇×A = r
i
×
∂
∂q
i
A
Diﬀerentiation of a vector function
Christoﬀel coeﬃcients of the second kind
∂r
i
∂q
j
= Γ
k
ij
r
k
∂r
j
∂q
i
= −Γ
j
it
r
t
Γ
k
ij
= Γ
k
ji
Christoﬀel coeﬃcients of the ﬁrst kind
1
2
∂g
it
∂q
j
+
∂g
tj
∂q
i
−
∂g
ji
∂q
t
= Γ
ijt
Γ
ijk
= Γ
jik
Covariant diﬀerentiation
∂f
∂q
i
= r
k
∇
i
f
k
= r
j
∇
i
f
j
∇
k
f
i
=
∂f
i
∂q
k
−Γ
j
ki
f
j
∇
k
f
i
=
∂f
i
∂q
k
+ Γ
i
kt
f
t
∇f = r
i
r
j
∇
i
f
j
= r
i
r
j
∇
i
f
j
Covariant diﬀerentiation of secondorder tensor
∂
∂q
k
A = ∇
k
a
ij
r
i
r
j
= ∇
k
a
ij
r
i
r
j
= ∇
k
a
·j
i
r
i
r
j
= ∇
k
a
i
·j
r
i
r
j
298 Tensor Analysis with Applications in Mechanics
∇
k
a
ij
=
∂a
ij
∂q
k
+ Γ
i
ks
a
sj
+ Γ
j
ks
a
is
∇
k
a
ij
=
∂a
ij
∂q
k
−Γ
s
ki
a
sj
−Γ
s
kj
a
is
∇
k
a
·j
i
=
∂a
·j
i
∂q
k
−Γ
s
ki
a
·j
s
+ Γ
j
ks
a
·s
i
∇
k
a
i
·j
=
∂a
i
·j
∂q
k
+ Γ
i
ks
a
s
· j
−Γ
s
kj
a
i
·s
Diﬀerential operations
∇×f = r
k
ijk
∂f
j
∂q
i
Γ
i
in
=
1
√
g
∂
√
g
∂q
n
∇· f =
1
√
g
∂
∂q
i
√
gf
i
∇×A =
kin
r
n
∂
∂q
k
r
j
a
ij
∇· A =
1
√
g
∂
∂q
i
√
g a
ij
r
j
∇
2
f = ∇· ∇f = g
ij
∂
2
f
∂q
i
∂q
j
−Γ
k
ij
∂f
∂q
k
∇
2
f = ∇
j
∇
j
f ∇
j
= g
ij
∇
i
∇
2
f = r
j
∇
i
∇
i
f
j
∇∇· f = r
i
∇
i
∇
j
f
j
∇×∇×f = ∇∇· f −∇
2
f
Formulary 299
Orthogonal coordinate systems
Lam´e coeﬃcients
(H
i
)
2
= g
ii
r
i
= r
i
/(H
i
)
2
(i = 1, 2, 3)
ˆr
i
= r
i
/H
i
Diﬀerentiation in the orthogonal basis
∇ =
ˆr
i
H
i
∂
∂q
i
∇f = ˆr
i
ˆr
j
1
H
i
∂f
j
∂q
i
−
f
i
H
i
H
j
∂H
i
∂q
j
+ δ
ij
f
k
H
k
1
H
i
∂H
i
∂q
k
∇· f =
1
H
1
H
2
H
3
∂
∂q
1
(H
2
H
3
f
1
) +
∂
∂q
2
(H
3
H
1
f
2
) +
∂
∂q
3
(H
1
H
2
f
3
)
∇×f =
1
2
ˆr
i
×ˆr
j
H
i
H
j
∂
∂q
i
(H
j
f
j
) −
∂
∂q
j
(H
i
f
i
)
∇
2
f =
1
H
1
H
2
H
3
¸
∂
∂q
1
H
2
H
3
H
1
∂f
∂q
1
+
∂
∂q
2
H
3
H
1
H
2
∂f
∂q
2
+
∂
∂q
3
H
1
H
2
H
3
∂f
∂q
3
Integration formulas
Transformation of multiple integral
V
f(x
1
, x
2
, x
3
) dx
1
dx
2
dx
3
=
V
f(q
1
, q
2
, q
3
)J dq
1
dq
2
dq
3
J =
√
g =
∂x
i
∂q
j
300 Tensor Analysis with Applications in Mechanics
Integration by parts
V
∂f
∂x
k
g dx
1
dx
2
dx
3
= −
V
∂g
∂x
k
f dx
1
dx
2
dx
3
+
S
fg n
k
dS
Miscellaneous results
V
∇f dV =
S
fndS
V
∇· f dV =
S
n · f dS
V
∇f dV =
S
nf dS
V
∇×f dV =
S
n ×f dS
V
∇AdV =
S
nAdS
V
∇· AdV =
S
n · AdS
V
∇×AdV =
S
n ×AdS
Γ
f · dr =
S
(n ×∇) · f dS
Γ
dr · A =
S
(n ×∇) · AdS
Γ
A· dr =
S
(n ×∇) · A
T
dS
Chapter 5
Elementary theory of curves
Parametrization
r = r(t) or r = r(s)
Length
s =
b
a
r
(t) dt
Formulary 301
Unit tangent
τ(s) = r
(s)
Equation of tangent line
r = r(t
0
) + λr
(t
0
)
x −x(t
0
)
x
(t
0
)
=
y −y(t
0
)
y
(t
0
)
=
z −z(t
0
)
z
(t
0
)
Curvature
k
1
= r
(s) k
2
1
=
(r
(t) ×r
(t))
2
(r
2
(t))
3
Radius of curvature
R = 1/k
1
Principal normal, binormal
ν =
r
(s)
k
1
β = τ ×ν
Osculating plane
[r −r(s
0
)] · β(s
0
) = 0
x −x(t
0
) y −y(t
0
) z −z(t
0
)
x
(t
0
) y
(t
0
) z
(t
0
)
x
(t
0
) y
(t
0
) z
(t
0
)
= 0
Torsion
k
2
= −
(r
(s) ×r
(s)) · r
(s)
k
2
1
k
2
= −
(r
(t) ×r
(t)) · r
(t)
(r
(t) ×r
(t))
2
302 Tensor Analysis with Applications in Mechanics
Serret–Frenet equations
τ
= k
1
ν
ν
= −k
1
τ −k
2
β
β
= k
2
ν
Theory of surfaces
Parametrization
r = r(u
1
, u
2
)
Tangent vectors, unit normal
r
i
=
∂r
∂u
i
(i = 1, 2)
n =
r
1
×r
2
r
1
×r
2

First fundamental form
(ds)
2
= g
ij
du
i
du
j
= E(du
1
)
2
+ 2F du
1
du
2
+ G(du
2
)
2
E = r
1
· r
1
F = r
1
· r
2
G = r
2
· r
2
Orthogonality of curves
E du
1
d˜ u
1
+ F(du
1
d˜ u
2
+ du
2
d˜ u
1
) + Gdu
2
d˜ u
2
= 0
Area
S =
A
EG−F
2
du
1
du
2
Formulary 303
Second fundamental form
d
2
r · n = L(du
1
)
2
+ 2M du
1
du
2
+ N(du
2
)
2
= −dr · dn
L =
∂
2
r
(∂u
1
)
2
· n M =
∂
2
r
∂u
1
∂u
2
· n N =
∂
2
r
(∂u
2
)
2
· n
Normal curvature, mean curvature, Gaussian curvature
k
0
= k
1
cos ϑ ϑ = angle between ν and n
H =
1
2
(k
min
+ k
max
) =
1
2
LG−2MF + NE
EG−F
2
K = k
min
k
max
=
LN −M
2
EG−F
2
Surface given by z = f(x, y) in Cartesian coordinates
Subscripts x, y denote partial derivatives with respect to x, y respectively.
E = 1 + f
x
2
F = f
x
f
y
G = 1 + f
y
2
EG−F
2
= 1 + f
x
2
+ f
y
2
S =
D
1 + f
x
2
+ f
y
2
dxdy
n =
−f
x
i
1
−f
y
i
2
+i
3
1 + f
x
2
+ f
y
2
L = r
xx
· n =
f
xx
1 + f
x
2
+ f
y
2
M = r
xy
· n =
f
xy
1 + f
x
2
+ f
y
2
N = r
yy
· n =
f
yy
1 + f
x
2
+ f
y
2
304 Tensor Analysis with Applications in Mechanics
K =
f
xx
f
yy
−f
xy
2
1 + f
x
2
+ f
y
2
2
Surface of revolution about zaxis
x = φ(u) z = ψ(u)
(ds)
2
=
φ
2
+ ψ
2
du
2
+ φ
2
dv
2
−dn · dr =
ψ
φ
−φ
ψ
φ
2
+ ψ
2
du
2
+
ψ
φ
φ
2
+ ψ
2
dv
2
Surface gradient operator and Gauss–Ostrogradsky theorems
Surface gradient operator
¯
∇ = r
i
∂
∂u
i
(i = 1, 2)
Surface analogues of the Gauss–Ostrogradsky (divergence) theorem
S
¯
∇· X+ 2Hn · X
dS =
Γ
ν · Xds
S
¯
∇X+ 2HnX
dS =
Γ
νXds
S
¯
∇×X+ 2Hn ×X
dS =
Γ
ν × Xds
S
¯
∇×(nX) dS =
Γ
τXds
Chapter 6
Stress tensor
Relation between stress tensor and stress vector
t = n · σ
Formulary 305
Equilibrium and motion equations
∇· σ + ρf = 0 ∇· σ + ρf = ρ
d
2
u
dt
2
Strain tensor
ε =
1
2
∇u + (∇u)
T
Compatibility deformation conditions
∇×(∇×ε)
T
= 0
Hooke’s law
σ = C·· ε
c
ijmn
= c
jimn
= c
ijnm
c
ijmn
= c
mnij
Isotropic material
σ = λEtr ε + 2µε
Recalculation of elastic moduli for isotropic body
Moduli λ, µ k, µ µ, ν E, ν E, µ
λ λ k −
2
3
µ
2µν
1−2ν
νE
(1+ν)(1−2ν)
(E−2µ)µ
3µ−E
µ = G µ µ µ
E
2(1+ν)
µ
k λ +
2
3
µ k
2µ(1+ν)
3(1−2ν)
E
3(1−2ν)
Eµ
3(3µ−E)
E
µ(3λ+2µ)
λ+µ
9kµ
3k+µ
2µ(1 + ν) E E
ν
λ
2(λ+µ)
3k−2µ
6k+2µ
ν ν
1
2
E
µ
−1
306 Tensor Analysis with Applications in Mechanics
Principal equations in Cartesian coordinates
Equations of motion
∂σ
ij
∂x
i
+ ρf
j
= ρ
∂
2
u
j
∂t
2
∂σ
11
∂x
1
+
∂σ
21
∂x
2
+
∂σ
31
∂x
3
+ ρf
1
= ρ
∂
2
u
1
∂t
2
∂σ
12
∂x
1
+
∂σ
22
∂x
2
+
∂σ
32
∂x
3
+ ρf
2
= ρ
∂
2
u
2
∂t
2
∂σ
13
∂x
1
+
∂σ
23
∂x
2
+
∂σ
33
∂x
3
+ ρf
3
= ρ
∂
2
u
3
∂t
2
Strains
ε
11
=
∂u
1
∂x
1
ε
12
=
1
2
∂u
1
∂x
2
+
∂u
2
∂x
1
ε
22
=
∂u
2
∂x
2
ε
13
=
1
2
∂u
1
∂x
3
+
∂u
3
∂x
1
ε
33
=
∂u
3
∂x
3
ε
23
=
1
2
∂u
2
∂x
3
+
∂u
3
∂x
2
Principal equations in a curvilinear coordinate system
Equations of motion
∂
∂q
i
√
gσ
ij
+ Γ
j
mn
σ
mn
+ ρ
√
gf
j
= ρ
√
g
∂
2
u
j
∂t
2
Strains
ε = ε
st
r
s
r
t
ε
st
=
1
2
∂u
s
∂q
t
+
∂u
t
∂q
s
−Γ
r
st
u
r
Formulary 307
Principal equations in cylindrical coordinates
Equations of motion
∂σ
rr
∂r
+
σ
rr
−σ
φφ
r
+
1
r
∂σ
rφ
∂φ
+
∂σ
zr
∂z
+ ρf
r
= ρ
∂
2
u
r
∂t
2
∂σ
rφ
∂r
+ 2
σ
rφ
r
+
1
r
∂σ
φφ
∂φ
+
∂σ
zφ
∂z
+ ρf
φ
= ρ
∂
2
u
φ
∂t
2
∂σ
rz
∂r
+
σ
rz
r
+
1
r
∂σ
zφ
∂φ
+
∂σ
zz
∂z
+ ρf
z
= ρ
∂
2
u
z
∂t
2
Strains
ε
rr
=
∂u
r
∂r
ε
rφ
=
1
2
∂u
φ
∂r
+
1
r
∂u
r
∂φ
−
u
φ
r
ε
φφ
=
1
r
∂u
φ
∂φ
+
u
r
r
ε
φz
=
1
2
1
r
∂u
z
∂φ
+
∂u
φ
∂z
ε
zz
=
∂u
z
∂z
ε
rz
=
1
2
∂u
r
∂z
+
∂u
z
∂r
Equilibrium equations in displacements
µ
∆u
r
−
u
r
r
2
−
2
r
2
∂u
φ
∂φ
+ (λ + µ)
∂
∂r
¸
1
r
∂
∂r
(ru
r
) +
1
r
∂u
φ
∂φ
+
∂u
z
∂z
+ ρf
r
= 0
µ
∆u
φ
−
u
φ
r
2
+
2
r
2
∂u
r
∂φ
+ (λ + µ)
1
r
∂
∂φ
¸
1
r
∂
∂r
(ru
r
) +
1
r
∂u
φ
∂φ
+
∂u
z
∂z
+ ρf
φ
= 0
µ∆u
z
+ (λ + µ)
∂
∂r
¸
1
r
∂
∂r
(ru
r
) +
1
r
∂u
φ
∂φ
+
∂u
z
∂z
+ ρf
z
= 0
where
∆ =
∂
2
∂r
2
+
1
r
∂
∂r
+
1
r
2
∂
2
∂φ
2
+
∂
2
∂z
2
308 Tensor Analysis with Applications in Mechanics
Principal equations in spherical coordinates
Equations of motion
∂σ
rr
∂r
+
1
r
∂σ
rθ
∂θ
+
1
r sinθ
∂σ
rφ
∂φ
+
1
r
(2σ
rr
−σ
θθ
−σ
φφ
+ σ
rθ
cot θ) + ρf
r
= ρ
∂
2
u
r
∂t
2
∂σ
rφ
∂r
+
1
r
∂σ
θφ
∂θ
+
1
r sin θ
∂σ
φφ
∂φ
+
1
r
(3σ
rφ
+ 2σ
θφ
cot θ) + ρf
φ
= ρ
∂
2
u
φ
∂t
2
∂σ
rθ
∂r
+
1
r
∂σ
θθ
∂θ
+
1
r sin θ
∂σ
θφ
∂φ
+
1
r
[(σ
θθ
−σ
φφ
) cot θ + 3σ
rθ
] + ρf
θ
= ρ
∂
2
u
θ
∂t
2
Strains
ε
rr
=
∂u
r
∂r
ε
θθ
=
1
r
∂u
θ
∂θ
+
u
r
r
ε
φφ
=
1
r sin θ
∂u
φ
∂φ
+
u
θ
r
cot θ +
u
r
r
ε
θφ
=
1
2r
∂u
φ
∂θ
+
1
sin θ
∂u
θ
∂φ
−u
φ
cot θ
ε
rφ
=
1
2
1
r sin θ
∂u
r
∂φ
+
∂u
φ
∂r
−
u
φ
r
ε
rθ
=
1
2
∂u
θ
∂r
+
1
r
∂u
r
∂θ
−
u
θ
r
Formulary 309
Equilibrium equations in displacements
µ
∆u
r
−
2
r
2
¸
u
r
+
1
sin θ
∂
∂θ
(u
θ
sin θ) +
1
sin θ
∂u
φ
∂φ
+ (λ + µ)
∂
∂r
¸
1
r
2
∂
∂r
(r
2
u
r
) +
1
r sin θ
∂
∂θ
(u
θ
sin θ) +
1
r sin θ
∂u
φ
∂φ
+ ρf
r
= 0
µ
∆u
θ
−
2
r
2
¸
∂u
r
∂θ
−
1
2 sin
2
θ
u
θ
−
cos θ
sin
2
θ
∂u
φ
∂φ
+
λ + µ
r
∂
∂θ
¸
1
r
2
∂
∂r
(r
2
u
r
) +
1
r sinθ
∂
∂θ
(u
θ
sin θ) +
1
r sin θ
∂u
φ
∂φ
+ ρf
θ
= 0
µ
∆u
φ
+
2
r
2
sin θ
¸
∂u
r
∂φ
+ cot
∂u
θ
∂φ
−
u
φ
2 sin θ
+
λ + µ
r sinθ
∂
∂φ
¸
1
r
2
∂
∂r
(r
2
u
r
) +
1
r sinθ
∂
∂θ
(u
θ
sin θ) +
1
r sin θ
∂u
φ
∂φ
+ ρf
φ
= 0
where
∆ =
1
r
2
∂
∂r
r
2
∂
∂r
+
1
r
2
sin θ
∂
∂θ
sin θ
∂
∂θ
+
1
r
2
sin θ
∂
2
∂φ
2
Chapter 7
Formulas of Surface Theory
Position vector of the middle surface and base vectors
ρ = ρ(q
1
, q
2
)
ρ
1
=
∂ρ
∂q
1
ρ
2
=
∂ρ
∂q
2
ρ
α
· ρ
β
= δ
β
α
(α, β = 1, 2)
Normal to midsurface
n =
ρ
1
×ρ
2
ρ
1
×ρ
2

310 Tensor Analysis with Applications in Mechanics
Metric tensor
A = ρ
α
ρ
α
= E−nn
Nabla operator
¯
∇ = ρ
α
∂
∂q
α
Curvature tensor
B = b
αβ
ρ
α
ρ
β
= −
¯
∇n
Derivatives of ρ
α
and ρ
α
∂ρ
α
∂q
β
= Γ
γ
αβ
ρ
γ
+ b
αβ
n
∂ρ
α
∂q
β
= −Γ
α
βγ
ρ
γ
+ b
α
β
n
Gradient of vector v
¯
∇v = (
¯
∇˜ v) · A−wB+ (
¯
∇w +B· ˜ v)n
v = ˜ v + wn
˜ v = v
1
(q
1
, q
2
)ρ
1
+ v
2
(q
1
, q
2
)ρ
2
w = v
3
= v
3
= v · n
Divergence of secondorder tensor T
¯
∇· T = ρ
γ
∂
∂q
γ
·
T
αβ
ρ
α
ρ
β
+ T
3β
nρ
β
+ T
α3
ρ
α
n + T
33
nn
=
∂T
αβ
∂q
α
ρ
β
+ T
αβ
ρ
γ
·
∂ρ
α
∂q
γ
ρ
β
+ T
αβ
∂ρ
β
∂q
α
+ T
3β
ρ
γ
·
∂n
∂q
γ
ρ
β
+ T
33
ρ
γ
·
∂n
∂q
γ
n +
∂T
α3
∂q
α
n + T
α3
ρ
γ
·
∂ρ
α
∂q
γ
n + T
α3
∂n
∂q
α
=
∂T
αβ
∂q
α
ρ
β
+ T
αβ
Γ
γ
αγ
ρ
β
+ T
αβ
Γ
γ
βα
ρ
γ
+ T
αβ
b
αβ
n −T
3β
b
γ
γ
ρ
β
− T
33
b
γ
γ
n +
∂T
α3
∂q
α
n + T
α3
Γ
γ
αγ
n −T
α3
b
β
α
ρ
β
Formulary 311
Kinematics in a Neighborhood of a Shell MidSurface
Position vector
r = r(q
1
, q
2
, z) = ρ(q
1
, q
2
) + zn
Basis vectors
r
α
=
∂r
∂q
α
= ρ
α
+ z
∂n
∂q
α
= (A−zB) · ρ
α
r
α
= (A−zB)
−1
· ρ
α
r
3
= r
3
= n
Spatial nabla operator
∇ = r
α
∂
∂q
α
+n
∂
∂z
= (A−zB)
−1
·
¯
∇+n
∂
∂z
Shell Equilibrium Equations
Stress resultant tensor and couple stress tensor
T = [[(A−zB)
−1
· σ]]
M = −[[(A−zB)
−1
· zσ ×n]]
where
[[f]] =
h/2
−h/2
Gf dz
Equilibrium equations
¯
∇· T+q = 0 T
×
= ρ
α
×[[σ
α
]]
¯
∇· M+T
×
+m = 0 σ
α
= r
α
· σ
Strain measures
=
1
2
(
¯
∇˜ v) · A+A· (
¯
∇˜ v)
T
−wB
æ = −
1
2
(
¯
∇
˜
ϑ) · A+A· (
¯
∇
˜
ϑ)
T
312 Tensor Analysis with Applications in Mechanics
Constitutive equations for Kirchhoﬀ–Love shells
T· A =
Eh
1 + ν
¸
+
ν
1 −ν
Atr
M = −
Eh
3
12(1 + ν)
¸
æ+
ν
1 −ν
Atr æ
×n
Component representation
T· A = T
αβ
ρ
α
ρ
β
=
αβ
ρ
α
ρ
β
M = −M
αβ
ρ
α
ρ
β
×n æ = æ
αβ
ρ
α
ρ
β
T
αβ
= [[σ
αβ
]] =
h/2
−h/2
σ
αβ
dz
M
αβ
= [[zσ
αβ
]] =
h/2
−h/2
zσ
αβ
dz
T
αβ
=
Eh
1 + ν
¸
αβ
+
ν
1 −ν
a
αβ
γ
γ
M
αβ
=
Eh
1 + ν
¸
æ
αβ
+
ν
1 −ν
a
αβ
æ
γ
γ
Shell surface energy density
U =
Eh
2(1 + ν)
¸
·· +
ν
1 −ν
tr
2
+
Eh
3
24(1 + ν)
¸
æ·· æ+
ν
1 −ν
tr
2
æ
Common principal boundary conditions
(1) Clamped (ﬁxed) edge
v
ω
= 0
∂w
∂ν
ω
= 0
(2) Simple support edge
v
ω
= 0 ν · (M×n) · ν
ω
= 0
Formulary 313
(3) Free edge
ν · T· A
ω
= 0
ν · T· n
ω
−
∂
∂s
(ν · (M×n) · τ)
ω
= 0
ν · (M×n) · ν
ω
= 0
Deﬂection of a plate
Resultant moments
M
11
=
Eh
3
12(1 −ν
2
)
(æ
11
+ νæ
22
) æ
11
= −
∂
2
w
∂x
2
1
M
22
=
Eh
3
12(1 −ν
2
)
(æ
22
+ νæ
11
) æ
22
= −
∂
2
w
∂x
2
2
M
12
=
Eh
3
12(1 + ν)
æ
12
æ
12
= −
∂
2
w
∂x
1
∂x
2
Bending stiﬀness
D =
Eh
3
12(1 −ν
2
)
Equilibrium equation
D
¯
∇
4
w = q
n
Equilibrium equation in Cartesian coordinates
D
∂
4
w
∂x
4
1
+ 2
∂
4
w
∂x
2
1
∂x
2
2
+
∂
4
w
∂x
4
2
= q
n
This page intentionally left blank This page intentionally left blank
Appendix B
Hints and Answers
Chapter 1 Exercises
Exercise 1.1 The given equation states that a, b, c are not coplanar;
because a has a nonzero component perpendicular to the plane of b and c,
it cannot be in this plane.
Chapter 1 Problems
Problem 1.1 (a) (5, 7, 4); (b) (7, 7, −2); (c) (17, 11, −4).
Problem 1.2 (a) (0, 1, 0); (b) (1, −10, 2); (c) (12, 8, 6).
Problem 1.3 (a) x = −
1
2
a+2b; (b) x = 2b−2a; (c) x =
1
12
(c−a) +
4
3
b.
Problem 1.4 (a) 2, (0, 1, −1); (b) 11, (−7, 5, −1); (c) 4, (1, −1, 0); (d) −2,
(−2, −3, −1); (e) −1, (−1, −1, −1); (f) 1, (0, 0, −5); (g) 1, (5, −3, −1); (h)
−1, (−4, −2, −7); (i) −8, (−16, 24, −14).
Problem 1.9 (a) 1; (b) −1; (c) 0.
Problem 1.10 (a) −1; (b) −4; (c) 5; (d) 5; (e) −179; (f) 0; (g) 0; (h) 13;
(i) −2.
Problem 1.11 (a) −1; (b) 4; (c) 3 ; (d) −7; (e) −826; (f) −18; (g) 0; (h)
105; (i) 0.
315
316 Tensor Analysis with Applications in Mechanics
Chapter 2 Exercises
Exercise 2.1 Suppose that e
i
and f
i
are two vectors such that x
i
= x· e
i
and x
i
= x · f
i
for all x. Then x · (e
i
− f
i
) = 0 for all x, from which it
follows that e
i
−f
i
= 0. So f
i
= e
i
.
Exercise 2.2 We must show that the equation α
1
e
1
+ α
2
e
2
+ α
3
e
3
= 0
implies α
1
= α
2
= α
3
= 0. To get α
1
= 0, for example, we simply dot
multiply the equation by e
1
and use e
i
· e
j
= δ
j
i
.
Exercise 2.3 We have
e
2
×e
3
=
1
V
2
(e
3
×e
1
) ×(e
1
×e
2
)
=
1
V
2
{[e
2
· (e
3
×e
1
)]e
1
−[e
1
· (e
3
×e
1
)]e
2
}
=
1
V
2
e
1
[e
2
· (e
3
× e
1
)]
by the vector triple product identity, hence
V
=
1
V
2
e
1
· e
1
[e
2
· (e
3
×e
1
)] =
1
V
2
[e
1
· (e
2
×e
3
)] =
1
V
.
Exercise 2.4 (b) First use common properties of determinants to establish
the identity
[a · (b ×c)][u · (v ×w)] =
a · u a · v a · w
b · u b · v b · w
c · u c · v c · w
. (*)
Write
[a · (b ×c)][u · (v ×w)] =
a
1
a
2
a
3
b
1
b
2
b
3
c
1
c
2
c
3
u
1
u
2
u
3
v
1
v
2
v
3
w
1
w
2
w
3
in a Cartesian frame. Then
[a · (b ×c)][u · (v ×w)] =
a
1
a
2
a
3
b
1
b
2
b
3
c
1
c
2
c
3
u
1
v
1
w
1
u
2
v
2
w
2
u
3
v
3
w
3
=
a
1
u
1
+a
2
u
2
+a
3
u
3
a
1
v
1
+a
2
v
2
+a
3
v
3
a
1
w
1
+ a
2
w
2
+a
3
w
3
b
1
u
1
+b
2
u
2
+b
3
u
3
b
1
v
1
+b
2
v
2
+b
3
v
3
b
1
w
1
+b
2
w
2
+b
3
w
3
c
1
u
1
+c
2
u
2
+c
3
u
3
c
1
v
1
+c
2
v
2
+c
3
v
3
c
1
w
1
+c
2
w
2
+c
3
w
3
.
Hints and Answers 317
(Note that our intermediate steps occurred in a Cartesian frame, but the
ﬁnal result is still frame independent.) Finally, write
V
2
= [e
1
· (e
2
×e
3
)]
2
=
e
1
· e
1
e
1
· e
2
e
1
· e
3
e
2
· e
1
e
2
· e
2
e
2
· e
3
e
3
· e
1
e
3
· e
2
e
3
· e
3
= g.
Exercise 2.5
(a) x
2
= x
k
x
k
= g
ij
x
i
x
j
= g
mn
x
m
x
n
.
(b) x · y = x
k
y
k
= x
k
y
k
= g
ij
x
i
y
j
= g
mn
x
m
y
n
.
Exercise 2.6 The transformation matrix is of the form
⎛
⎝
cos θ sin θ 0
−sinθ cos θ 0
0 0 1
⎞
⎠
,
where θ is the angle of rotation.
Exercise 2.7 Equate two diﬀerent expressions for x as is done in the text.
Exercise 2.8
ijk
=
⎧
⎪
⎪
⎨
⎪
⎪
⎩
+V
, (i, j, k) an even permutation of (1, 2, 3),
−V
, (i, j, k) an odd permutation of (1, 2, 3),
0, two or more indices equal.
(Recall that V
= 1/V .) Then use the determinantal identity established
in Exercise 2.4: put a = e
i
, b = e
j
, c = e
k
, u = e
p
, v = e
q
, w = e
r
. To
prove the vector triple product identity we write
a ×(b ×c) = e
i
ijk
a
j
kpq
b
p
c
q
= e
i
(δ
p
i
δ
q
j
−δ
q
i
δ
p
j
)a
j
b
p
c
q
= e
i
b
i
(a
q
c
q
) −e
i
c
i
(a
p
b
p
).
Exercise 2.9 Using the scalar triple product identity and (2.13) we have
(a ×b) · (c ×d) = a · [b ×(c ×d)]
= a · [(b · d)c −(b · c)d],
and the result follows.
318 Tensor Analysis with Applications in Mechanics
Chapter 2 Problems
Problem 2.1 (a) e
1
= (2i
1
+ 3i
2
− 2i
3
)/9, e
2
= (−3i
1
+ 3i
2
+ i
3
)/9,
e
3
= (5i
1
− 6i
2
+ 4i
3
)/9; (b) e
1
= −i
1
+ i
3
, e
2
= (−5i
1
− 3i
2
+ 7i
3
)/13,
e
3
= (12i
1
+2i
2
−9i
3
)/13; (c) e
1
= (i
1
+i
2
)/2, e
2
= (i
1
−i
2
)/2, e
3
= 1/3i
3
;
(d) e
1
= cos φi
1
+ sin φi
2
, e
2
= −sinφi
1
+ cos φi
2
, e
3
= i
3
.
Problem 2.2 Using the formulas ˜ e
1
= (˜ e
2
× ˜ e
3
)/V , etc., ﬁrst ﬁnd
˜ e
1
= (1, 1, 0), ˜ e
2
= (−1, 0, −1), ˜ e
3
= (−1, −2, 2).
Then, with regard for the formula A
j
i
= e
i
· ˜ e
j
, arrive at
⎛
⎝
3 −1 −6
2 −3 2
1 −2 1
⎞
⎠
.
Problem 2.3
1
3
⎛
⎝
5 14 2
−10 −10 −1
2 −1 −1
⎞
⎠
.
Problem 2.4 (a) a
k
; (b) a
i
a
i
; (c) 3; (d) δ
k
i
; (e) 3; (f) 3.
Problem 2.7 (a) 0; (b) −6; (c)
inm
; (d) 0; (e) 0; (f) 2δ
n
k
.
Problem 2.8 (a ×b) ×c = b(a · c) −a(b · c).
Chapter 3 Exercises
Exercise 3.1 (b)
⎛
⎝
1 0 0
0 0 0
0 0 0
⎞
⎠
,
⎛
⎝
0 0 0
0 1 0
0 0 0
⎞
⎠
,
⎛
⎝
0 0 0
0 0 0
1 0 0
⎞
⎠
.
Exercise 3.2 E can be written in any of the forms
E = e
ij
e
i
e
j
= e
ij
e
i
e
j
= e
i
·j
e
i
e
j
= e
·j
i
e
i
e
j
.
Let us choose the ﬁrst form as an example. Equation (3.4) with x = e
k
implies e
ij
e
i
e
j
· e
k
= e
k
. Predotting this with e
m
yields e
mk
= g
mk
.
Exercise 3.3 (A· B) · (B
−1
· A
−1
) = A· (B· B
−1
) · A
−1
= A· E· A
−1
=
A· A
−1
= E.
Exercise 3.4 (a)
˜
B = A
T
BA. (b)
˜
B = ABA
T
.
Hints and Answers 319
Exercise 3.5 We are given e
i
= A
j
i
˜ e
j
. Now it is necessary to understand
what the needed tensor does. We denote it by A. It is the map y = A· x
that must take x = e
i
into y = A
j
i
˜ e
j
for each i = 1, 2, 3. Let us write out
these equations:
A· e
i
= A
j
i
˜ e
j
.
The decisive step is to suppose that A should be written in a “mixed basis”
as A = a
m
· n
˜ e
m
e
n
. We have
a
m
· n
˜ e
m
e
n
· e
i
= A
j
i
˜ e
j
and it follows that
a
m
· i
˜ e
m
= A
j
i
˜ e
j
.
Thus (because the indices are dummy)
a
j
·i
= A
j
i
and so the tensor is A = a
m
· n
˜ e
m
e
n
where a
m
· n
= A
m
n
.
Exercise 3.7
(a) a
ij
e
i
e
j
·· e
k
e
k
= a
ij
δ
k
j
g
ik
= a
· k
k
.
(b) a
ij
e
i
e
j
·· b
kn
e
k
e
n
= a
ij
b
kn
δ
j
k
δ
i
n
= a
ik
b
ki
. Having obtained this we may
use the metric tensor to raise and lower indices and thereby obtain
other forms; for example, a
ik
b
ki
= a
p
· k
g
pi
b
ki
= a
p
· k
b
k
· p
.
Exercise 3.10
(a) Write
A = a
ij
e
i
e
j
, B = b
kn
e
k
e
n
,
A
T
= a
ji
e
i
e
j
, B
T
= b
nk
e
k
e
n
,
and show that both sides can be written as a
jp
b
· k
p
e
k
e
j
.
(b) b · C· a = b · (C· a) = b · (a · C
T
) = (a · C
T
) · b = a · C
T
· b.
(c) See
[
Lurie (2005)
]
.
(d) The relation (A
T
)
−1
= (A
−1
)
T
follows from the two relations
A
T
· (A
−1
)
T
= (A
−1
· A)
T
= E
T
= E,
(A
−1
)
T
· A
T
= (A· A
−1
)
T
= E
T
= E.
320 Tensor Analysis with Applications in Mechanics
Exercise 3.11
⎛
⎝
a
11
a
12
a
13
a
12
a
22
a
23
a
13
a
23
a
33
⎞
⎠
,
⎛
⎝
0 a
12
a
13
−a
12
0 a
23
−a
13
−a
23
0
⎞
⎠
.
Exercise 3.12 With A = a
ij
e
i
e
j
and B = b
kn
e
k
e
n
we have
A·· B = a
ij
b
kn
g
jk
g
in
. (*)
Since A is symmetric and B is antisymmetric, we may also write A·· B =
−a
ji
b
nk
g
jk
g
in
; let us swap i with j and k with n in this expression to get
A·· B = −a
ij
b
kn
g
in
g
jk
. (**)
Adding (*) and (**) we obtain 2(A·· B) = 0.
Exercise 3.13
x ·
¸
1
2
(A+A
T
)
· x =
1
2
x · A· x +x · A
T
· x
=
1
2
[x · A· x +x · A· x]
= x · A· x.
Exercise 3.15 They are (1) x = a with λ = b · a, and (2) any x that is
perpendicular to b, with λ = 0.
Exercise 3.16 The characteristic equation −λ
3
+2λ
2
+λ = 0 has solutions
λ
1
= 0, λ
2
= 1 +
√
2, λ
3
= 1 −
√
2. The Vi`ete formulas give
I
1
(A) = λ
1
+λ
2
+λ
3
= 2,
I
2
(A) = λ
1
λ
2
+λ
1
λ
3
+λ
2
λ
3
= −1,
I
3
(A) = λ
1
λ
2
λ
3
= 0.
Exercise 3.17 We prove this for two eigenvectors; the reader may readily
generalize to any number of eigenvectors. Let A· x
1
= λ
1
x
1
and A· x
2
=
λ
2
x
2
where λ
2
= λ
1
. Now suppose
α
1
x
1
+α
2
x
2
= 0. (*)
Hints and Answers 321
Let us operate on both sides of (*) with the tensor A−λ
2
I. After simpli
ﬁcation we obtain
α
1
(λ
1
−λ
2
)x
1
= 0.
Because x
1
= 0 by deﬁnition, we have α
1
= 0. Putting this back into (*),
we also have α
2
= 0.
Exercise 3.18 Assume x
1
, x
2
, x
3
are eigenvectors of A corresponding
to the distinct eigenvalues λ
1
, λ
2
, λ
3
, respectively. Let us show that any
eigenvector corresponding to λ
1
must be a scalar multiple of x
1
. So suppose
that
A· x = λ
1
x. (**)
By the linear independence of x
1
, x
2
, x
3
, these vectors form a basis for
threedimensional space and we can express
x = c
1
x
1
+c
2
x
2
+c
3
x
3
.
Applying A to both sides we have, by (**),
λ
1
x = λ
1
c
1
x
1
+λ
1
c
2
x
2
+λ
1
c
3
x
3
= c
1
λ
1
x
1
+c
2
λ
2
x
2
+c
3
λ
3
x
3
.
This simpliﬁes to
(λ
1
−λ
2
)c
2
x
2
+ (λ
1
−λ
3
)c
3
x
3
= 0.
By linear independence of x
2
, x
3
, we get c
2
= c
3
= 0. Hence x = c
1
x
1
.
Exercise 3.19 The characteristic equation of A is
1 −λ 0 0
1 1 −λ 0
0 1 −λ
= 0,
which is λ
3
−2λ
2
+λ = 0. Thus A
3
= 2A
2
−A.
Exercise 3.20
(a) Q · Q
T
= Q
T
· Q = i
1
i
1
+ i
2
i
2
+ i
3
i
3
= E. Since we can write Q =
−i
1
i
1
+i
2
i
2
+i
3
i
3
, the matrix of Q in mixed components is
⎛
⎝
−1 0 0
0 1 0
0 0 1
⎞
⎠
.
Hence det Q = −1.
322 Tensor Analysis with Applications in Mechanics
(b) The deﬁning equation Q· Q
T
= E implies
q
ij
e
i
e
j
· q
km
e
m
e
k
= q
ij
q
kj
e
i
e
k
= δ
k
i
e
i
e
k
.
(c) If n is a positive integer,
Q
n
· (Q
n
)
T
= Q
n
· (Q
T
)
n
= (Q· Q
T
)
n
= E
n
= E.
Exercise 3.21 Predot A· x = λx with x to get
λ =
x · (A· x)
x
2
.
Clearly λ > 0 under the stated condition.
Exercise 3.23
(a) E ··· zyx =
ijk
(e
k
· z)(e
j
· y)(e
i
· x) = x
i
(
ijk
y
j
z
k
).
(b) E ·· xy = e
i
ijk
(e
k
· x)(e
j
· y) = e
i
ijk
y
j
x
k
.
(c) E · x =
ijk
e
i
e
j
x
k
so that (E · x) · y =
ijk
e
i
y
j
x
k
= y ×x.
Exercise 3.28 Write E = e
m
e
m
.
Exercise 3.30 Use the representation of the invariants through the eigen
values of A.
Exercise 3.31 Use Theorem 3.8, the fact that B in the representation is
a ball tensor, and the equality E·· A
T
= tr A.
Exercise 3.32 Introduce the nth partial sum
S
n
= E +
1
1!
A+· · · +
1
n!
A
n
.
It follows that
S
n+m
−S
n
=
1
(n + 1)!
A
(n+1)
+· · · +
1
(n +m)!
A
(n+m)
≤
1
(n + 1)!
A
n+1
+· · · +
1
(n +m)!
A
n+m
.
Next, use the proof of convergence of the Taylor series for e
x
from any
textbook with x = A.
Exercise 3.33 Show that
(1) A
n
≤ q
n
→0 as n →∞;
(2) (E−A)(E +A+A
2
+A
3
+· · · +A
n
) = E−A
n+1
;
Hints and Answers 323
(3) limit passage may be justiﬁed in the last equality.
Exercise 3.34 (E +A)
−1
= E−A+A
2
−A
3
+· · · + (−1)
n
A
n
+· · · .
Exercise 3.35 Consider f(X+εB) = tr X+ε tr B. Using the deﬁnition,
we get
∂
∂ε
f(X+εdX)
ε=0
= tr(dX).
As tr B = E·· B
T
, we get f
,X
= E.
Exercise 3.36 Consider f(X+εB) = tr X
2
+ε tr(X· B) +ε tr(B· X) +
ε
2
tr B
2
. By the deﬁnition, we get
∂
∂ε
f(X+εB)
ε=0
= tr(X· B) + tr(B· X) = 2 tr(X· B)
With regard for tr(X· B) = X·· B = X
T
·· B
T
, we get f
,X
= 2X
T
.
Exercise 3.38 Use the results of Exercises 3.35 and 3.36.
Exercise 3.39 Use a consequence of the Cayley–Hamilton theorem, which
is
I
3
=
1
3
tr
X
3
−I
1
X
2
+I
2
X
,
and the results of Exercises 3.35, 3.36 and 3.37.
Exercise 3.43 Let us derive f(X+εdX). We have
f(X+εdX) =
1
4
X·· C·· X+
ε
2
dX·· C·· X
+
ε
2
X·· C·· dX+
ε
2
2
dX·· C·· dX.
So
∂
∂ε
f(X+ εdX)
ε=0
=
1
2
dX·· C·· X+
1
2
X·· C·· dX
=
1
2
(dx
mn
c
mnpt
x
pt
+x
mn
c
mnpt
dx
pt
)
=
1
2
(c
mnpt
x
pt
+x
ij
c
ijmn
)dx
mn
.
324 Tensor Analysis with Applications in Mechanics
Equating the components of this to the components of f
,X
·· dX, we get
∂f
∂x
11
=
1
2
(c
11pt
x
pt
+x
mn
c
mn11
),
∂f
∂x
12
=
1
4
(c
12pt
x
pt
+c
21pt
x
pt
+x
mn
c
mn12
+x
mn
c
mn21
),
∂f
∂x
22
=
1
2
(c
22pt
x
pt
+x
mn
c
mn22
),
∂f
∂x
13
=
1
4
(c
13pt
x
pt
+c
31pt
x
pt
+x
mn
c
mn13
+x
mn
c
mn31
),
.
.
.
In tensor notation these formulas are
f
,X
=
1
4
(C·· X+C
·· X+X·· C+X·· C
),
where C
= c
mnpt
i
m
i
n
i
t
i
p
and C
= c
mnpt
i
n
i
m
i
t
i
p
.
Exercise 3.44 Consider the tensor equality
C·· X =
1
4
(C·· X+C
·· X+X·· C+X·· C
)
in components.
Chapter 3 Problems
Problem 3.1 They are the components of the matrix
⎛
⎝
0 1 0
0 0 0
0 0 0
⎞
⎠
.
Problem 3.2
⎛
⎝
0 1 0
−1 0 0
0 0 2
⎞
⎠
Problem 3.3
⎛
⎝
−1 1 −2
2 −2 4
−2 2 −4
⎞
⎠
.
Problem 3.5
(a)
1
2
(i
1
i
2
+i
2
i
1
),
1
2
(i
1
i
2
−i
2
i
1
);
(b) 2i
3
i
3
, i
1
i
2
−i
2
i
1
;
(c)
1
2
(−i
1
i
2
−i
2
i
1
+i
1
i
3
+i
3
i
1
), 3(i
1
i
2
−i
2
i
1
) +
1
2
(i
1
i
3
−i
3
i
1
);
Hints and Answers 325
(d)
1
2
(i
1
i
2
+i
2
i
3
+i
1
i
3
+i
2
i
1
+i
3
i
2
+i
3
i
1
),
1
2
(i
1
i
2
+i
2
i
3
+i
1
i
3
−i
2
i
1
−i
3
i
2
−i
3
i
1
);
(e) i
1
i
1
+ 2i
1
i
2
+ 2i
2
i
1
+i
3
i
1
+i
1
i
3
, 0.
Problem 3.6
(a) 0, i
1
i
2
;
(b) 0, i
1
i
2
+i
2
i
1
;
(c)
1
3
E,
2
3
i
1
i
1
−
1
3
i
2
i
2
−
1
3
i
3
i
3
;
(d)
1
3
a · aE, aa −
1
3
a · aE;
(e)
1
3
E, 2i
1
i
2
+ 2i
2
i
1
+i
3
i
1
+i
1
i
3
+
2
3
i
1
i
1
−
1
3
i
2
i
2
−
1
3
i
3
i
3
.
Problem 3.8
(a) I
1
= a · a, I
2
= 0, I
3
= 0;
(b) I
1
= 0, I
2
= −
1
2
, I
3
= 0;
(c) I
1
= 2, I
2
= 1, I
3
= 0;
(d) I
1
= 3λ, I
2
= 3λ
2
, I
3
= λ
3
;
(e) I
1
= 9, I
2
= 26, I
3
= 24.
Problem 3.10
(a) S = S
= λE, Q = sgnλE;
(b) S = S
= aa +bb +cc, Q =
1
a·a
aa +
1
b·b
bb +
1
c·c
cc ≡ E;
(c) S = S
= λ+ai
1
i
1
+λi
2
i
2
+λi
3
i
3
, Q = sgn(λ+a)i
1
i
1
+sgnλi
2
i
2
+
sgnλi
3
i
3
;
(d) S = S
= λ+ai
1
i
1
+λ+bi
2
i
2
+λi
3
i
3
, Q = sgn(λ+a)i
1
i
1
+sgn(λ+
b)i
2
i
2
+ sgnλi
3
i
3
;
(e) S = S
= ai
1
i
1
+bi
2
i
2
+ci
3
i
3
, Q = sgnai
1
i
1
+ sgnbi
2
i
2
+ sgnci
3
i
3
.
Problem 3.12 Let X be a needed solution. Multiply the above equality
by A from the left and take the trace of both sides:
a tr(A· X) + tr(A· X) tr A = tr(A· B). (*)
If a+tr A = 0, then tr(A· X) = tr(A· B)/(a+tr A). Substituting this into
the above equality we get
X =
1
a
B−
tr(A· B)
a + tr A
E
.
326 Tensor Analysis with Applications in Mechanics
If a + tr A = 0 then (*) reduces to 0 = tr(A· B). So if tr(A· B) = 0 the
equation has no solution X. Let us see what happens when a + tr A = 0
and tr(A · B) = 0. Now tr(A · X) can take any value but the dev X is
uniquely deﬁned as follows from the initial equation: dev X =
1
a
dev B as
a = 0.
Answer. (1) If a+tr A = 0, then X =
1
a
B−
tr(A·B)
a+tr A
E
; (2) if a+tr A = 0
and tr(A · B) = 0 then a solution does not exist; (3) if a + tr A = 0 and
tr(A · B) = 0, then only dev X is uniquely deﬁned X =
1
a
dev B and so
X = λE+
1
a
B.
Problem 3.13 (a) X = C −
tr(A·C)
1+tr(A·B)
B is uniquely deﬁned when 1 +
tr(A · B) = 0; (b) X = C
T
−
tr(A·C
T
)
1+tr(A·B
T
)
B
T
is uniquely deﬁned when
1 + tr(A · B
T
) = 0; (c) X = B −
a tr B
1+a tr A
A is uniquely deﬁned when
1 +a tr A = 0.
Problem 3.14 Taking the devoperation on the above equality, we get
a dev X = 0. This means that dev X = 0 when a = 0. If a = 0 then dev X
can take any value.
Applying the trace operation to the equality, we get (a +tr E) tr X = 0.
It is valid when (1) a = −3 or (2) tr X = 0.
If a = −3, then X = λE with any scalar λ.
If tr X = 0, then the equation reduces to aX = 0 and so it can have a
nonzero solution X only if a = 0.
Answer. (1) Let a = −3. Then X = λE with any scalar λ; (2) Let a = 0.
Then X is an arbitrary tensor such that tr X = 0.
Problem 3.15 Let X be a solution. Calculate the trace of both sides
of the equation given in the problem. It follows that (a + tr A) tr X = 0.
There follow two possibilities: (1) a = −tr A, or (2) tr X = 0.
When a = −tr A then tr X can take any value; now X = λA is a
solution for any scalar λ.
Let tr X = 0. Then the equation reduces to aX = 0, which has a
nonzero solution only if a = 0.
Answer. (1) a = −tr A, X = λA; (2) If a = 0, then X is an arbitrary
tensor such that tr X = 0.
Hints and Answers 327
Problem 3.19
(E×ω)
2
= (E ×ω) · (E×ω)
= ω ×E· E×ω
= ω ×E× ω
= ω ×i
k
i
k
×ω
= ω
p
i
p
×i
k
i
k
×i
t
ω
t
= ω
p
ω
t
pkm
ktn
i
m
i
n
= ω
p
ω
t
mpk
tnk
i
m
i
n
= ω
p
ω
t
(δ
mt
δ
pn
−δ
mn
δ
pt
)i
m
i
n
= ω
m
ω
n
i
m
i
n
−ω
p
ω
p
i
m
i
m
= ωω −ω · ωE.
Problem 3.21 (a) 0; (b) −2ω; (c) 2a ×b; (d) 0; (e) 0.
Problem 3.39 I
1
= 3α+βe · e, I
2
= 3α
2
+2αβe · e, I
3
= α
2
(α +βe · e).
Problem 3.40 I
1
= 0, I
2
= ω · ω, I
3
= 0.
Problem 3.41 I
1
= −2ω · ω, I
2
= (ω · ω)
2
, I
3
= 0.
Problem 3.43
(a) 2i
1
i
1
+i
2
i
2
+i
3
i
3
−i
1
i
2
−i
2
i
1
;
(b) c
−1
i
1
i
3
+b
−1
i
2
i
2
+c
−1
i
3
i
1
;
(c) a
−1
i
1
i
1
+i
2
i
2
+i
3
i
3
−a
−1
bi
1
i
2
(d) a
−1
E−a
−2
bi
1
i
2
.
Problem 3.44 (a) 4X
T
3
; (b) aa; (c) ab; (d) B
T
.
Problem 3.49 (a) EA; (b) EA
T
; (c) cdab.
Chapter 4 Exercises
Exercise 4.1
(a) Consider the derivative of the dot product. Since we are given the
vectors explicitly, an easy way is to dot the vectors ﬁrst:
e
1
(t) · e
2
(t) = e
−t
(1 −sin
2
t) = e
−t
cos
2
t,
328 Tensor Analysis with Applications in Mechanics
hence
[e
1
(t) · e
2
(t)]
= −e
−t
cos t(2 sin t + cos t).
However, it is easily checked that the productrule identity given in the
text yields the same result.
(b) [e(t) ×e
(t)]
= e
(t) ×e
(t) +e(t) ×e
(t) = e(t) ×e
(t).
Exercise 4.3 Use the product rule to take the indicated second derivative
of r = ˆ ρρ, noting that
dˆ ρ
dt
=
d
dt
(ˆ xcos φ + ˆ ysin φ) = (−ˆ xsin φ + ˆ y cos φ)
dφ
dt
=
ˆ
φ
dφ
dt
and, similarly,
d
ˆ
φ
dt
= −ˆ ρ
dφ
dt
.
Exercise 4.4 v
1
= v
r
, v
2
= v
θ
/r, v
3
= v
φ
/r sin θ.
Exercise 4.5
(a) Fix v = v
0
and eliminate u from the transformation equations to get
y = xtan α +v
0
(sin β −cos β tan α);
thus the v = v
0
coordinate curve is a straight line in the xyplane
making an angle α with the xaxis. Similarly, the u = u
0
line makes
an angle β with the xaxis.
(b) r = ˆ xx + ˆ yy = ˆ x(u cos α +v cos β) + ˆ y(u sinα +v sin β) gives
r
1
=
∂r
∂u
= ˆ xcos α + ˆ y sin α, r
2
=
∂r
∂v
= ˆ xcos β + ˆ ysin β.
Note that these are both unit vectors. To ﬁnd r
1
we write r
1
= ˆ xa+ˆ yb
and solve the equations
r
1
· r
1
= 1, r
1
· r
2
= 0,
simultaneously to get a, b. By this method we ﬁnd that
r
1
= ˆ x
sin β
sin(β −α)
− ˆ y
cos β
sin(β −α)
,
r
2
= −ˆ x
sin α
sin(β −α)
+ ˆ y
cos α
sin(β −α)
.
Hints and Answers 329
These are not orthonormal vectors; moreover the coordinate system is
nonorthogonal except in such trivial cases as β −α = π/2. We have
g
11
= 1 = g
22
, g
12
= g
21
= cos(β −α),
and
g
11
=
1
sin
2
(β −α)
= g
22
, g
12
= g
21
= −
cos(β −α)
sin
2
(β −α)
.
(c) Write z = z
1
r
1
+ z
2
r
2
= z
1
r
1
+ z
2
r
2
. Substitute for the r
i
and r
i
in
terms of ˆ x and ˆ y and then equate coeﬃcients of ˆ x, ˆ y to get simultaneous
equations for the z
i
in terms of the z
i
. The answers are
z
1
= z
1
+z
2
cos(β −α), z
2
= z
1
cos(β −α) +z
2
.
Exercise 4.6 (ds)
2
= (du)
2
+ 2 du dv cos(β −α) + (dv)
2
.
Exercise 4.7
(a) The total diﬀerential of ϕ is
dϕ =
∂ϕ
∂q
i
dq
i
.
If we start at a point on the given surface and move in such a way
that we stay on the surface, then dϕ = 0. In the text it is shown that
dq
i
= r
i
· dr, so if we move along dr satisfying
∂ϕ
∂q
i
r
i
· dr = 0
then we will stay on the surface. It is clear that the vector
∇ϕ =
∂ϕ
∂q
i
r
i
is normal to the surface, since it is perpendicular to every tangent
direction. We have
∇ϕ =
∇ϕ · ∇ϕ
=
∂ϕ
∂q
m
r
m
·
∂ϕ
∂q
n
r
n
=
g
mn
∂ϕ
∂q
m
∂ϕ
∂q
n
,
330 Tensor Analysis with Applications in Mechanics
hence
n =
∇ϕ
∇ϕ
=
∂ϕ
∂q
i
r
i
g
mn
∂ϕ
∂q
m
∂ϕ
∂q
n
=
g
ij
∂ϕ
∂q
i
g
mn
∂ϕ
∂q
m
∂ϕ
∂q
n
r
j
.
(b) Use the result of part (a).
(c) Use the result of part (b).
(d) Use the result of part (b).
Exercise 4.8 For cylindrical coordinates we have
∇f = r
i
∂f
∂q
i
= r
1
∂f
∂ρ
+r
2
∂f
∂φ
+r
3
∂f
∂z
= r
1
∂f
∂ρ
+
r
2
ρ
2
∂f
∂φ
+r
3
∂f
∂z
= ˆ ρ
∂f
∂ρ
+
ˆ
φ
ρ
∂f
∂φ
+ ˆz
∂f
∂z
.
Exercise 4.9 The results are all zero by expansion in Cartesian frame.
Exercise 4.10 Use expansion in Cartesian frame.
Exercise 4.11 Use expansion in Cartesian frame.
Exercise 4.12 Equation (4.18) gives, for instance,
Γ
221
=
1
2
∂g
12
∂q
2
+
∂g
21
∂q
2
−
∂g
22
∂q
1
.
In the cylindrical system this is
Γ
221
=
1
2
∂0
∂φ
+
∂0
∂φ
−
∂ρ
2
∂ρ
= −ρ.
Exercise 4.13 Use the results of the previous exercise and equation (4.20).
Exercise 4.14 From r = ˆ xc coshu cos v + ˆ yc sinhu sinv we ﬁnd
r
u
=
∂r
∂u
= ˆ xc sinhu cos v + ˆ yc coshu sinv,
r
v
=
∂r
∂v
= −ˆ xc coshu sinv + ˆ yc sinhu cos v,
Hints and Answers 331
and consequently
g
uu
= g
vv
= c
2
(cosh
2
u −cos
2
v), g
uv
= g
vu
= 0
(the system is orthogonal). Then
r
u
=
r
u
c
2
(cosh
2
u −cos
2
v)
, r
v
=
r
v
c
2
(cosh
2
u −cos
2
v)
,
from which we get
g
uu
= g
vv
=
1
c
2
(cosh
2
u −cos
2
v)
, g
uv
= g
vu
= 0.
An application of (4.18) gives
Γ
uuu
= c
2
coshu sinhu,
Γ
uuv
= −c
2
cos v sinv,
Γ
uvu
= Γ
vuu
= c
2
cos v sinv,
Γ
vvu
= −c
2
coshu sinhu,
Γ
vuv
= Γ
uvv
= c
2
coshu sinhu,
Γ
vvv
= c
2
cos v sin v.
Finally, from (4.20) we obtain
Γ
u
uu
=
cosh u sinhu
cosh
2
u −cos
2
v
,
Γ
v
uu
= −
cos v sinv
cosh
2
u −cos
2
v
,
Γ
u
uv
= Γ
u
vu
=
cos v sin v
cosh
2
u −cos
2
v
,
Γ
u
vv
= −
coshu sinhu
cosh
2
u −cos
2
v
,
Γ
v
vu
= Γ
v
uv
=
coshu sinhu
cosh
2
u −cos
2
v
,
Γ
v
vv
=
cos v sin v
cosh
2
u −cos
2
v
.
Exercise 4.15 We switch to a notation in which the coordinate symbols
ρ and φ are used for the indices. The only nonzero Christoﬀel symbols are
332 Tensor Analysis with Applications in Mechanics
Γ
ρ
φφ
= −ρ, Γ
φ
ρφ
= 1/ρ. Then
∇
ρ
f
ρ
=
∂f
ρ
∂ρ
, ∇
φ
f
ρ
=
∂f
ρ
∂φ
−ρf
φ
,
∇
ρ
f
φ
=
∂f
φ
∂ρ
+
1
ρ
f
φ
, ∇
φ
f
φ
=
∂f
φ
∂φ
+
1
ρ
f
ρ
.
Exercise 4.16 Use the method shown in the text. For example,
∂
∂q
k
(a
·j
i
r
i
r
j
) =
∂a
·j
i
∂q
k
r
i
r
j
+a
·j
i
∂r
i
∂q
k
r
j
+r
i
∂r
j
∂q
k
=
∂a
·j
i
∂q
k
r
i
r
j
+a
·j
i
−Γ
i
kt
r
t
r
j
+r
i
Γ
t
jk
r
t
=
∂a
·j
i
∂q
k
r
i
r
j
+
−a
·j
s
Γ
s
ki
r
i
r
j
+r
i
a
·s
i
Γ
j
sk
r
j
=
∂a
·j
i
∂q
k
−Γ
s
ki
a
·j
s
+ Γ
j
sk
a
·s
i
r
i
r
j
.
Exercise 4.17 Consider, for example,
∇
k
g
ij
=
∂g
ij
∂q
k
−Γ
s
ki
g
sj
−Γ
s
kj
g
is
.
By (4.19), this is
∇
k
g
ij
=
∂g
ij
∂q
k
−Γ
kij
−Γ
kji
.
Elimination of Γ
kij
and Γ
kji
via (4.18) shows that ∇
k
g
ij
= 0.
Exercise 4.18 Equate components of the ﬁrst and third members of the
trivial identity ∇(E· a) = ∇a = E· ∇a.
Exercise 4.19 In cylindrical coordinates,
∇· f =
1
√
g
∂
∂q
i
(
√
gf
i
)
=
1
ρ
¸
∂
∂ρ
(ρf
1
) +
∂
∂φ
(ρf
2
) +
∂
∂z
(ρf
3
)
=
1
ρ
∂
∂ρ
(ρf
1
) +
1
ρ
∂
∂φ
(ρf
2
) +
∂
∂z
(f
3
)
=
1
ρ
∂
∂ρ
(ρf
ρ
) +
1
ρ
∂f
φ
∂φ
+
∂f
z
∂z
.
Hints and Answers 333
In spherical coordinates,
∇· f =
1
√
g
∂
∂q
i
(
√
gf
i
)
=
1
r
2
sinθ
¸
∂
∂r
(r
2
sin θf
1
) +
∂
∂θ
(r
2
sin θf
2
) +
∂
∂φ
(r
2
sin θf
3
)
=
1
r
2
sinθ
¸
∂
∂r
(r
2
sin θf
r
) +
∂
∂θ
(r sin θf
θ
) +
∂
∂φ
(rf
φ
)
=
1
r
2
∂
∂r
(r
2
f
r
) +
1
r sin θ
∂
∂θ
(sin θf
θ
) +
1
r sin θ
∂f
φ
∂φ
.
Exercise 4.20
∇×∇f =
ijk
r
k
∂
2
f
∂q
i
∂q
j
−Γ
n
ij
∂f
∂q
n
= 0.
(Swapping two adjacent indices on
ijk
causes a sign change; hence, when
ever the symbol multiplies an expression that is symmetric in two of its
subscripts, the result is zero.)
Exercise 4.22 We have
∂(x
1
, x
2
, x
3
)
∂(˜ q
1
, ˜ q
2
, ˜ q
3
)
=
∂x
1
∂˜ q
1
∂x
1
∂˜ q
2
∂x
1
∂˜ q
3
∂x
2
∂˜ q
1
∂x
2
∂˜ q
2
∂x
2
∂˜ q
3
∂x
3
∂˜ q
1
∂x
3
∂˜ q
2
∂x
3
∂˜ q
3
=
∂x
1
∂q
i
∂q
i
∂˜ q
1
∂x
1
∂q
i
∂q
i
∂˜ q
2
∂x
1
∂q
i
∂q
i
∂˜ q
3
∂x
2
∂q
i
∂q
i
∂˜ q
1
∂x
2
∂q
i
∂q
i
∂˜ q
2
∂x
2
∂q
i
∂q
i
∂˜ q
3
∂x
3
∂q
i
∂q
i
∂˜ q
1
∂x
3
∂q
i
∂q
i
∂˜ q
2
∂x
3
∂q
i
∂q
i
∂˜ q
3
=
⎛
⎜
⎜
⎜
⎜
⎜
⎜
⎝
∂x
1
∂q
1
∂x
1
∂q
2
∂x
1
∂q
3
∂x
2
∂q
1
∂x
2
∂q
2
∂x
2
∂q
3
∂x
3
∂q
1
∂x
3
∂q
2
∂x
3
∂q
3
⎞
⎟
⎟
⎟
⎟
⎟
⎟
⎠
⎛
⎜
⎜
⎜
⎜
⎜
⎜
⎝
∂q
1
∂˜ q
1
∂q
1
∂˜ q
2
∂q
1
∂˜ q
3
∂q
2
∂˜ q
1
∂q
2
∂˜ q
2
∂q
2
∂˜ q
3
∂q
3
∂˜ q
1
∂q
3
∂˜ q
2
∂q
3
∂˜ q
3
⎞
⎟
⎟
⎟
⎟
⎟
⎟
⎠
.
The result follows from the fact that the determinant of a product equals
the product of the determinants.
Chapter 4 Problems
Problem 4.1 ∇f = f
(r)/rr, ∇
2
f = f
(r) + 2f
(r)/r.
334 Tensor Analysis with Applications in Mechanics
Problem 4.2
(a) ai
1
i
1
+bi
2
i
2
+ci
3
i
3
, a +b +c;
(b) ai
2
i
1
, 0;
(c) aE, 3a;
(d) f
(r)e
r
e
r
+f(r)/r e
φ
e
φ
, f
(r) +f(r)/r;
(e) f
(r)e
r
e
φ
−f(r)/r e
φ
e
r
, 0;
(f) f
(r)e
r
e
z
, 0;
(g) f
(r)e
r
e
r
+f(r)/r e
φ
e
φ
+f(r)/r e
θ
e
θ
, f
(r) + 2f(r)/r;
(h) −E×ω, 0;
(i) f
(φ)/r e
φ
e
z
+g
(φ)/r e
φ
e
φ
−g(φ)/r e
φ
e
r
, g
(φ)/r;
(j) f
(z)e
z
e
z
+g
(φ)/r e
φ
e
φ
−g(φ)/r e
φ
e
r
, f
(z) +g
(φ)/r;
(k) A
T
, tr A.
Problem 4.4 (a) E; (b) −6A; (c) tr A; (d) ∇f; (e) 3E; (f) 4r.
Problem 4.5
(a) f
e
r
+
f−g
r
e
r
+h
e
z
;
(b) f
e
r
+ 2
f−g
r
e
r
;
(c) f
e
r
+
f−g
r
e
r
;
(d) f
i
1
+g
i
2
+h
i
3
;
(e) 0;
(f) f
e
φ
+
f+g
r
e
φ
+h
e
z
;
(g)
f
r
e
z
+
g
r
e
φ
+h
e
r
.
Problem 4.6 f ×r.
Problem 4.7 Find the coordinate vectors for the new coordinates and
demonstrate their orthogonality. The answers are
H
σ
= a
2
σ
2
− τ
2
σ
2
−1
, H
τ
= a
2
σ
2
−τ
2
1 −τ
2
, H
z
= 1.
Problem 4.8 H
σ
= H
τ
= σ
2
+τ
2
, H
φ
= σ
2
τ
2
.
Problem 4.9
H
σ
= H
τ
=
a
2
(coshτ −cos σ)
2
, H
z
= 1.
Hints and Answers 335
Problem 4.10
H
σ
= H
τ
=
a
2
(coshτ − cos σ)
2
, H
φ
=
a
2
sin
2
σ
(coshτ −cos σ)
2
.
Problem 4.11
H
σ
= H
τ
=
a
2
(coshτ − cos σ)
2
, H
φ
=
a
2
sinh
2
τ
(coshτ −cos σ)
2
.
Problem 4.18 Consider the integral
V
∇· (Ar) dV.
Using Gauss–Ostrogradsky theorem we get
V
∇· (Ar) dV =
S
n · Ar dS =
S
gr dS.
On the other hand, with regard for the identity
∇· (Ar) = (∇· A)r +A
T
· ∇r = (∇· A)r +A
T
we have
V
∇· (Ar) dV =
V
(fr +A
T
) dV.
Comparing these two expressions we ﬁnd
V
A
T
dV =
S
gr dS −
V
fr dV,
from which the answer follows. The answer is
S
rg dS −
V
rf dV.
336 Tensor Analysis with Applications in Mechanics
Chapter 5 Exercises
Exercise 5.1
(a) r
(t) = −i
1
sint +i
2
cos t +i
3
, r
(t) =
√
2, so the required length is
2π
0
√
2 dt = 2π
√
2.
(b) The ellipse can be described as the locus of the tip of the vector
r(t) = i
1
Acos t +i
2
Bsint (0 ≤ t < 2π).
Hence
s =
2π
0
(A
2
sin
2
t +B
2
cos
2
t)
1/2
dt.
The integral on the right is an elliptic integral and cannot be evaluated
in closed form.
Exercise 5.3 s(t) =
t
0
√
2 dt = t
√
2, so
r(s) = i
1
cos(s/
√
2) +i
2
sin(s/
√
2) +i
3
(s/
√
2).
The unit tangent is
r
(s) =
1
√
2
¸
−i
1
sin
s
√
2
+i
2
cos
s
√
2
+i
3
.
Exercise 5.4 The curve is an exponential spiral. We have
r
(t) = e
t
[i
1
(cos t −sint) +i
2
(cos t + sint)],
hence r
(π/4) = i
2
e
π/4
√
2. Also
r(π/4) =
√
2
2
e
π/4
(i
1
+i
2
),
so the tangent line is described by
x =
√
2
2
e
π/4
.
Exercise 5.5 Assuming the curve can be expressed in the form ρ = ρ(θ),
we can write the position vector as
r(θ) = i
1
ρ(θ) cos θ +i
2
ρ(θ) sin θ.
Hints and Answers 337
Diﬀerentiation gives
r
(θ) = i
1
[−ρ(θ) sin θ +ρ
(θ) cos θ] +i
2
[ρ(θ) cos θ +ρ
(θ) sin θ].
(a) The equation r = r(θ
0
) +λr
(θ
0
) gives upon substitution and matching
of coeﬃcients
ρ cos θ = ρ(θ
0
) cos θ
0
+λ[−ρ(θ
0
) sin θ
0
+ρ
(θ
0
) cos θ
0
],
ρ sin θ = ρ(θ
0
) sin θ
0
+λ[ρ(θ
0
) cos θ
0
+ρ
(θ
0
) sin θ
0
],
where (ρ, θ) locates a point on the tangent line. The desired analogues
of (5.2) can be obtained by (1) squaring and adding, whereby we elim
inate θ from the lefthand side, and (2) dividing the two equations,
whereby we eliminate ρ from the lefthand side. Elimination of λ in
stead gives
ρ cos θ −ρ(θ
0
) cos θ
0
−ρ(θ
0
) sin θ
0
+ρ
(θ
0
) cos θ
0
=
ρ sinθ −ρ(θ
0
) sin θ
0
ρ(θ
0
) cos θ
0
+ρ
(θ
0
) sin θ
0
for the analogue of (5.3).
(b) We compute r(θ) · r
(θ) to get
r(θ) · r
(θ) = ρ(θ)ρ
(θ).
We also ﬁnd
r(θ) = ρ(θ), r
(θ) = [ρ
2
(θ) +ρ
2
(θ)]
1/2
,
so that by deﬁnition of the dot product
cos φ =
r(θ) · r
(θ)
r(θ)r
(θ)
=
ρ
(θ)
[ρ
2
(θ) +ρ
2
(θ)]
1/2
.
Use of the identity
tan φ =
[1 −cos
2
φ]
1/2
cos φ
allows us to get
tan φ =
ρ(θ)
ρ
(θ)
.
Exercise 5.7 Let r
(t
0
) = 0. For determining a tangent vector at t
0
we
can use the Taylor expansion of r = r(t):
r(t
0
+ ∆t) = r(t
0
) +r
(t
0
)∆t
+
1
2!
r
(t
0
)(∆t)
2
+· · · +
1
n!
r
(n)
(t
0
)(∆t)
n
+o(∆t
n
).
338 Tensor Analysis with Applications in Mechanics
If the nth derivative is the ﬁrst that is not zero at t = t
0
, then
r(t
0
+ ∆t) = r(t
0
) +
1
n!
r
(n)
(t
0
)(∆t)
n
+o(∆t
n
).
When ∆t → 0, the direction of the vector r(t
0
+ ∆t) − r(t
0
) tends to the
tangential direction of the curve at t = t
0
. Thus a tangent vector t to the
curve at t = t
0
can be found as
t = lim
∆t→0
r(t
0
+ ∆t) −r(t
0
)
(∆t)
n
=
1
n!
r
(n)
(t
0
).
Exercise 5.8
x = x(t
0
) +λ
1
n!
x
(n)
(t
0
),
y = y(t
0
) +λ
1
n!
y
(n)
(t
0
),
z = z(t
0
) +λ
1
n!
z
(n)
(t
0
).
Exercise 5.9 This time s(t) = t
α
2
+β
2
, hence
r(s) = i
1
αcos
s
α
2
+β
2
+i
2
αsin
s
α
2
+β
2
+i
3
β
s
α
2
+β
2
.
This gives
r
(s) =
1
α
2
+β
2
¸
−i
1
αsin
s
α
2
+β
2
+i
2
αcos
s
α
2
+β
2
+i
3
β
,
r
(s) =
α
α
2
+β
2
¸
−i
1
cos
s
α
2
+β
2
−i
2
sin
s
α
2
+β
2
¸
,
and hence
k = r
(s) =
α
α
2
+β
2
.
Exercise 5.10 The principal normal is
ν = −i
1
cos
s
α
2
+β
2
−i
2
sin
s
α
2
+β
2
.
Hints and Answers 339
The binormal is obtained as
β =
i
1
i
2
i
3
−
α
√
α
2
+β
2
sin
s
√
α
2
+β
2
α
√
α
2
+β
2
cos
s
√
α
2
+β
2
β
√
α
2
+β
2
−cos
s
√
α
2
+β
2
−sin
s
√
α
2
+β
2
0
and is
β =
1
α
2
+β
2
¸
i
1
β sin
s
α
2
+β
2
−i
2
β cos
s
α
2
+β
2
+i
3
α
¸
.
We can determine a plane in space by specifying the normal to the
plane and a single point through which the plane passes. If the plane
passes through the point located by the position vector r
0
, and N is any
normal vector, then the plane is described by the vector equation
N· (r −r
0
) = 0
(that is, all points whose position vectors r satisfy the above equation will
lie in the plane). Hence the equations for the three fundamental planes are
β · (r −r
0
) = 0 osculating plane,
τ · (r −r
0
) = 0 normal plane,
ν · (r −r
0
) = 0 rectifying plane.
The rectifying plane for the helix is found to be xcos t
0
+y sin t
0
= α.
Exercise 5.11 Writing r = r(t), we diﬀerentiate with respect to t and use
the chain rule to obtain
dr
dt
= τ
ds
dt
,
d
2
r
dt
2
= τ
d
2
s
dt
2
+νk
ds
dt
2
.
These give
dr
dt
×
d
2
r
dt
2
= βk
ds
dt
3
or
r
(t) ×r
(t) = βkr
(t)
3
.
Dotting this equation to itself we obtain
k
2
=
[r
(t) ×r
(t)]
2
r
(t)
6
.
340 Tensor Analysis with Applications in Mechanics
The formula in the text was presented in such a way that the absolute value
function (with its inherent discontinuity at the origin) would be avoided.
Exercise 5.12
(a) R = 5
√
10/3. R = ∞.
(b) x = −b/2a, the vertex of the parabola.
Exercise 5.15 No; construct counterexamples.
Exercise 5.16 First show that β is a constant β
0
. Then use the deﬁnition
of a plane as a set of points described by a radius vector r such that r·n = c,
where c is a constant.
Exercise 5.17 Straightforward using equations (5.7) and (5.10).
Exercise 5.18 The vectors r
(t) and r
(t) reverse their directions, while
r
(t) does not. So τ reverses direction, ν maintains its direction, and
therefore β reverses direction. The moving trihedron changes its handed
ness. The formulas for k
1
and k
2
show that these quantities do not change
sign.
Exercise 5.19 We have
k
2
= −
β
α
2
+β
2
.
Note that the ratio of the curvature to the torsion is constant for the helix.
Exercise 5.20
d
3
r
ds
3
= −k
2
1
τ +
dk
1
ds
ν − k
1
k
2
β.
Exercise 5.22 We have
r
(t) = i
1
sin t +i
2
+i
3
cos t, r
(t) =
√
2,
and can ﬁnd the length parameter for the curve by setting
s(t) =
t
0
r
(x) dx =
t
0
√
2 dx =
√
2t.
Hints and Answers 341
So t = s/
√
2 and we have
r(s) = i
1
¸
1 −cos
s
√
2
+i
2
s
√
2
+i
3
sin
s
√
2
,
r
(s) = i
1
1
√
2
sin
s
√
2
+i
2
1
√
2
+i
3
1
√
2
cos
s
√
2
,
r
(s) = i
1
1
2
cos
s
√
2
−i
3
1
2
sin
s
√
2
.
The desired decomposition of the acceleration is
a = s
(t)τ +k
1
v
2
ν.
In the present case we have s
(t) = 0, while k
1
= r
(s) = 1/2 and
v
2
= (ds/dt)
2
= 2. Therefore a = ν.
Exercise 5.23 The sphere is described by
r = ˆ xa sinθ cos φ + ˆ ya sinθ sinφ + ˆ za cos θ.
We have
∂r
∂θ
= ˆ xa cos θ cos φ + ˆ ya cos θ sin φ − ˆ za sinθ,
∂r
∂φ
= −ˆ xa sinθ sinφ + ˆ ya sin θ cos φ,
hence
E = a
2
, F = 0, G = a
2
sin
2
θ,
so that (ds)
2
= a
2
(dθ)
2
+a
2
sin
2
θ(dφ)
2
.
Exercise 5.24
(a) The diﬀerentials of coordinates for elementary curves on the coordinate
lines are (du
1
, du
2
= 0) and (du
1
= 0, du
2
). The answer is given by
cos θ = F/
√
EG.
(b) The desired angle is ψ = π/4.
Exercise 5.25 We ﬁnd
r
1
= i
1
cos u
2
+i
2
sin u
2
+i
3
,
r
2
= −i
1
u
1
sin u
2
+ i
2
u
1
cos u
2
,
hence E = 2, F = 0, G = (u
1
)
2
. Then
S =
2π
0
a
0
√
2u
1
du
1
du
2
= π
√
2a
2
.
342 Tensor Analysis with Applications in Mechanics
Exercise 5.26 Computing dr/dρ and dr/dφ, we ﬁnd that
E = 1 + [f
(ρ)]
2
, F = 0, G = ρ
2
.
From these we get
(ds)
2
= {1 + [f
(ρ)]
2
}(dρ)
2
+ρ
2
(dφ)
2
and
S =
A
EG−F
2
dρ dφ =
1 + [f
(ρ)]
2
ρ dρ dφ.
Exercise 5.30 From
r = ˆ xρ cos φ + ˆ yρ sinφ + ˆ zf(ρ)
we ﬁnd that
n =
∂r
∂ρ
×
∂r
∂φ
∂r
∂ρ
×
∂r
∂φ
=
−ˆ xf
(ρ) cos φ − ˆ yf
(ρ) sin φ + ˆ z
1 + [f
(ρ)]
2
and
∂
2
r
∂ρ
2
= ˆ zf
(ρ),
∂
2
r
∂ρ∂φ
= −ˆ xsin φ + ˆ ycos φ,
∂
2
r
∂φ
2
= −ˆ xρ cos φ − ˆ yρ sin φ.
Hence
L =
f
(ρ)
1 + [f
(ρ)]
2
, M = 0, N =
ρf
(ρ)
1 + [f
(ρ)]
2
.
Exercise 5.31 In Exercise 5.23 we found that E = a
2
, F = 0, and
G = a
2
sin
2
θ for the sphere. The outward normal to the sphere is
n = ˆ xsin θ cos φ + ˆ y sinθ sin φ + ˆ z cos θ,
and this yields the values L = −a, M = 0, N = −a sin
2
θ. The mean
curvature is H = −1/a, and the Gaussian curvature is K = 1/a
2
.
Hints and Answers 343
Exercise 5.35 The answers include
µ
· 1
1
=
−f
(ρ)
{1 + [f
(ρ)]
2
}
3/2
, µ
· 2
1
= µ
· 1
2
= 0, µ
· 2
2
=
−f
(ρ)
ρ
1 + [f
(ρ)]
2
,
and
Γ
1
11
=
f
(ρ)f
(ρ)
1 + [f
(ρ)]
2
.
Exercise 5.40 A logarithmic spiral
r = ce
aφ
[i cos(φ +φ
0
) +j sin(φ +φ
0
)],
where the constant c is deﬁned by the initial values of the curve.
Chapter 5 Problems
Problem 5.1 The curve is called the astroid (Fig. B.1(a)). Its parametric
representation is x = a cos
3
t, y = a sin
3
t. The singular points are (0, a),
(0, −a), (a, 0), and (−a, 0).
a
a
a
x
y
2a
2p
 p 2
a) b)
x
y
Fig. B.1 (a) astroid curve, t ∈ [0, 2π]; (b) cycloid curve, t ∈ [−2π, 2π].
Problem 5.2 The curve is a cycloid (Fig. B.1(b)). The singular points
are (0, 2πk) for k = 0, ±1, ±2, . . ..
Problem 5.3
x
0
=
y
1
=
z −1
0
.
344 Tensor Analysis with Applications in Mechanics
Problem 5.4 s = 6a.
Problem 5.5 s = 8a.
Problem 5.6 s = 16a.
Problem 5.7 s = a
√
2 sinh t.
Problem 5.8
k
1
=
1
4
1 + sin
2
t
2
.
Problem 5.9
k
1
=
1
2a cosh
2
t
, k
2
=
1
2a cosh
2
t
.
Problem 5.10 k
2
= a cosht.
Problem 5.12 Diﬀerentiating the relation Q· Q
T
= E with respect to s,
we obtain
Q
· Q
T
+Q· Q
T
= 0.
Hence Q
· Q
T
is an antisymmetric tensor, so it takes the form Q
= d×Q
where d is the corresponding conjugate vector. Thus Q
· Q
T
= d×E. Now
refer to the result of Problem 3.21 (b).
Problem 5.13
−2du dv
√
u
2
+ 1
.
Problem 5.17 2H = tr B = −tr
¯
∇n = −
¯
∇· n.
Problem 5.18
0 =
¯
∇· E =
¯
∇· (A+nn) =
¯
∇· A+ (
¯
∇· n)n =
¯
∇· A−2Hn.
Problem 5.20
K =
f
xx
g
yy
1 +f
2
x
+g
2
y
2
.
Hints and Answers 345
Problem 5.21
K = −
1
AB
¸
A
v
B
v
+
B
u
A
u
,
where indices u and v denote partial derivatives with respect to u and v
respectively.
Problem 5.22 Let us consider Stokes’ formula from Chapter 4 for a tensor
ﬁeld X and a vector ﬁeld x:
Γ
τ · Xds =
S
n · (∇×X) dS,
Γ
τ · xds =
S
n · (∇×x) dS, (B.1)
where we have used the relations dr = τ ds and (n ×∇) · X = n · (∇×X).
Using the identities
A = −n ×(n ×A)
and
A = E · X = (A+nn) · X = −n ×(n ×A) · X+nn · X,
we represent the tensor X as
X = n ×X
1
+nx
2
(B.2)
where X
1
= −n × X and x
2
= n · X. Note that the surface and spatial
gradient operators
¯
∇ and ∇ are related by
∇ =
¯
∇+n
∂
∂z
,
where z is the distance coordinate along the normal to S. Then
n · (∇×X
1
) = n · (
¯
∇×X
1
).
Formula (B.1) for X
1
yields
S
n · (
¯
∇×X
1
) dS =
Γ
τ · X
1
ds. (B.3)
In view of the identities τ = ν × n and (ν × n) · X
1
= ν · (n × X
1
),
equation (B.3) reduces to
S
¯
∇· (n ×X
1
) dS =
Γ
ν · (n ×X
1
) ds. (B.4)
Using the relation
¯
∇· (nx
2
) = (
¯
∇· n)x
2
= −2Hx
2
,
346 Tensor Analysis with Applications in Mechanics
from (B.2) and (B.4) we derive the ﬁnal formula
S
¯
∇· X+ 2Hn · X
dS =
Γ
ν · Xds.
Problem 5.23 Similar to Problem 5.22.
Problem 5.24 Applying (5.58) to the tensor AX, we obtain the ﬁrst
formula
S
¯
∇X+ 2HnX
dS =
Γ
νXds.
The second one,
S
¯
∇×X+ 2Hn ×X
dS =
Γ
ν ×Xds,
follows from this. Substituting the tensor nX into the last formula, we
obtain the third formula
S
¯
∇×(nX) dS =
Γ
τXds.
Chapter 6 Exercises
Exercise 6.5 Start with Hooke’s law: σ = λEtr ε+2µε. Apply the trace
operation to it and ﬁnd tr ε:
tr ε =
1
3λ + 2µ
tr σ.
Substituting this into Hooke’s law, obtain ε:
ε =
1
2µ
¸
σ −
λ
3λ + 2µ
Etr σ
=
1
2µ
¸
σ −
ν
1 +ν
Etr σ
.
Now
W =
1
2
σ ·· ε =
1
4µ
¸
σ ·· σ −
ν
1 +ν
(tr σ)
2
.
Exercise 6.8 Use d’Alembert’s principle, which states that in dynamics
the system of all forces — including the inertia forces — are in equilibrium.
So in the equilibrium equations, change
f
i
→f
i
−ρ
d
2
u
i
dt
2
.
Hints and Answers 347
Exercise 6.12 Consider A(k) and C in a Cartesian frame. By deﬁnition,
A(k) = k · C· k
= k
m
i
m
· c
pqrt
i
p
i
q
i
r
i
t
· k
s
i
s
= k
m
k
s
c
mqrs
i
q
i
r
= k
m
k
s
c
mrqs
i
q
i
r
= k
m
k
s
c
mqrs
i
r
i
q
= A(k)
T
.
Exercise 6.13 It is suﬃcient to show that a · A(k) · a > 0 for a = 0. The
tensor C is such that ε · · C· · ε > 0 whenever ε = 0. Substitute into this
ε = ka. Because C is symmetric, the identity (ka) ·· C = (ak) ·· C holds.
Thus
ε ·· C·· ε = (ka) ·· C·· (ka) = a · (k · C· k)a = a · A(k) · a > 0.
Exercise 6.14 From Hooke’s law it follows that
tr ε = ε
11
+ε
22
=
σ
11
+σ
22
λ + 2µ
.
So we have
ε
11
=
1
2µ
σ
11
−
λ
2µ(λ + 2µ)
(σ
11
+σ
22
),
ε
22
=
1
2µ
σ
22
−
λ
2µ(λ + 2µ)
(σ
11
+σ
22
),
ε
12
=
σ
12
µ
.
Exercise 6.15 By Hooke’s law we get
σ
11
= λtr ε + 2µε
11
, σ
22
= λtr ε + 2µε
22
, 0 = λtr ε + 2µε
33
,
σ
12
= µε
12
, 0 = µε
13
, 0 = µε
23
.
It follows that
ε
13
= 0, ε
23
= 0, ε
12
=
σ
12
µ
, ε
33
= −
λ
λ + 2µ
(ε
11
+ε
22
).
348 Tensor Analysis with Applications in Mechanics
Substituting ε
33
into the ﬁrst two equations, we obtain
σ
11
=
2λµ
λ + 2µ
(ε
11
+ε
22
) + 2µε
11
, σ
22
=
2λµ
λ + 2µ
(ε
11
+ε
22
) + 2µε
22
.
These equations coincide with the equations of the plane deformation prob
lem when we change λ to λ
=
2λµ
λ+2µ
. Then we have
ε
11
+ε
22
=
σ
11
+σ
22
λ
+ 2µ
=
λ + 2µ
4µ(λ +µ)
(σ
11
+σ
22
),
ε
11
=
1
2µ
σ
11
−
λ
2µ(λ
+ 2µ)
(σ
11
+σ
22
),
ε
22
=
1
2µ
σ
22
−
λ
2µ(λ
+ 2µ)
(σ
11
+σ
22
).
Chapter 6 Problems
Problem 6.1
(a) On sides AB and CD: σ
11
= σ, σ
12
= 0; on sides AD and BC: σ
22
= 2σ,
σ
12
= 0.
(b) On sides AB and CD: σ
11
= 0, σ
12
= −τ; on sides AD and BC: σ
22
= 0,
σ
12
= −τ.
(c) On sides AB and CD: σ
11
= −p, σ
12
= 0; on sides AD and BC: σ
22
= σ,
σ
12
= 0.
(d) On sides AB and CD: σ
11
= 0, σ
12
= τ; on sides AD and BC: σ
22
= −p,
σ
12
= τ.
Problem 6.2 On AB we have n = −i
2
. The boundary condition is n· σ =
0. Using the representation
σ = σ
11
i
1
i
1
+σ
12
(i
1
i
2
+i
2
i
1
) +σ
22
i
2
i
2
,
we get
n · σ = −σ
22
i
2
−σ
12
i
1
= 0.
So the conditions on side AB are: σ
22
= 0, σ
12
= 0.
On BC we have n = i
1
. The conditions on side BC are
n · σ = σ
11
i
1
+σ
12
i
2
= 0.
Hints and Answers 349
On side AC the task is a bit more diﬃcult. On AC we have
n = −sinαi
1
+ cos αi
2
.
The boundary condition on AC is n · σ = 0. Using the representation
σ = σ
11
i
1
i
1
+σ
12
(i
1
i
2
+i
2
i
1
) +σ
22
i
2
i
2
,
we get
n · σ = (−sin αi
1
+ cos αi
2
) · (σ
11
i
1
i
1
+σ
12
(i
1
i
2
+i
2
i
1
) +σ
22
i
2
i
2
)
= i
1
(−σ
11
sinα +σ
12
cos α) +i
2
(−σ
12
sin α +σ
22
cos α)
= 0,
which yields the needed result.
The ﬁnal answers are as follows. Side AB:
σ
22
= 0, σ
12
= 0.
Side BC:
σ
11
= −γy, σ
12
= 0.
Side AC:
−σ
11
sin α +σ
12
cos α = 0, −σ
12
sin α +σ
22
cos α = 0.
Problem 6.3 On side BC, the normal is
n = sin βi
1
+ cos βi
2
.
Then
n · σ = (sin βi
1
+ cos βi
2
) · (σ
11
i
1
i
1
+σ
12
(i
1
i
2
+i
2
i
1
) +σ
22
i
2
i
2
)
= i
1
(σ
11
sin β +σ
12
cos β) + i
2
(σ
12
sin β +σ
22
cos β).
The force on BC is
−p
3
n = −p
3
(sin βi
1
+ cos βi
2
).
The conditions on side BC follow from the relation n · σ = −p
3
n.
The normal to AC is
n = −sinαi
1
+ cos αi
2
.
The rest is similar to the above solution for BC.
The ﬁnal answers are as follows. Side AB:
σ
22
= −p
1
, σ
12
= 0.
350 Tensor Analysis with Applications in Mechanics
Side BC:
σ
11
sin β +σ
12
cos β = −p
3
sin β,
σ
12
sin β +σ
22
cos β = −p
3
cos β.
Side AC:
−σ
11
sin α +σ
12
cos α = p
2
sin α,
−σ
12
sin α +σ
22
cos α = −p
2
cos α.
Problem 6.4
(a) On side AB: σ
φφ
= σ, σ
φr
= 0.
On a part of circle BC: σ
rr
= −p
2
, σ
rφ
= 0.
On a part of circle AC: σ
rr
= −p
1
, σ
rφ
= 0.
On side CD: σ
φφ
= σ, σ
φr
= 0.
(b) On side AB: σ
φφ
= −p, σ
φr
= 0.
On a part of circle BC: σ
rr
= 0, σ
rφ
= −τ
2
.
On a part of circle AC: σ
rr
= 0, σ
rφ
= −τ
1
.
On side CD: σ
φφ
= −p, σ
φr
= 0.
Problem 6.5
(a) On side ABCD: σ
11
= σ, σ
12
= 0, σ
13
= τ.
On a part of circle BEFC: σ
33
= 2σ, σ
32
= 0, σ
31
= τ.
On a part of circle DCFG: σ
22
= σ, σ
21
= 0, σ
23
= 0.
(b) On side ABCD: σ
11
= −p
1
, σ
12
= τ, σ
13
= 0.
On a part of circle BEFC: σ
33
= −p
3
, σ
32
= 0, σ
31
= 0.
On a part of circle DCFG: σ
22
= −p
2
, σ
21
= τ, σ
23
= 0.
Problem 6.6 Denote the principal axes by i
1
, i
2
, i
3
. In this Cartesian
system, the normal to triangle ABC is
n =
1
√
3
(i
1
+i
2
+i
3
).
The stress tensor is
σ = σ
1
i
1
i
1
+σ
2
i
2
i
2
+σ
3
i
3
i
3
.
Hints and Answers 351
The required stress vector is
t = n · σ
=
1
√
3
(σ
1
i
1
+σ
2
i
2
+σ
3
i
3
)
=
1
√
3
(50i
1
−50i
2
+ 75i
3
).
Problem 6.7
(a) On the upper cylinder face: σ
zz
= σ, σ
zφ
= τ
3
, σ
zr
= 0.
On the internal surface part: σ
rr
= 0, σ
rφ
= τ
2
, σ
rz
= 0.
On the external later surface: σ
rr
= 0, σ
rφ
= τ
1
, σ
rz
= 0.
(b) On side ABCD: σ
φφ
= 0, σ
φr
= 0, σ
φz
= τ
1
.
On a part of ring BCEH: σ
zz
= −p, σ
zφ
= 0, σ
zr
= 0.
On a part of cylinder CEFD: σ
rr
= 0, σ
rφ
= τ
3
, σ
rz
= 0.
On side FEHG: σ
φφ
= 0, σ
φr
= 0, σ
φz
= −τ
2
.
(c) On side ABCD: σ
φφ
= 0, σ
φr
= τ
1
, σ
φz
= 0.
On a part of ring BCEH: σ
zz
= σ, σ
zφ
= 0, σ
zr
= 0.
On a part of cylinder CEFD: σ
rr
= 0, σ
rφ
= 0, σ
rz
= −τ
3
.
On side FEHG: σ
φφ
= 0, σ
φr
= −τ
2
, σ
φz
= 0.
Problem 6.8
(a) ε =
1
2
γ(i
1
i
2
+i
2
i
1
);
(b) ε = λ
1
i
1
i
1
;
(c) ε = λE;
(d) ε = u
(r)e
r
e
r
+u(r)/r e
φ
e
φ
+ke
z
e
z
;
(e) ε = (u
(r) −u(r)/r)(e
r
e
φ
+e
φ
e
r
) +ke
z
e
z
;
(f) ε = u
(r)(e
r
e
z
+e
z
e
r
) +ke
z
e
z
;
(g) ε = u
(r)e
r
e
r
+u(r)/r e
φ
e
φ
+u(r)/r e
θ
e
θ
.
Chapter 7 Exercises
Exercise 7.11 Hooke’s law is given by the formula
σ = λEtr ε + 2µε.
352 Tensor Analysis with Applications in Mechanics
Dotmultiply this by n from the left and the right. From σ
33
= 0 it follows
that
λtr ε + 2µε
33
= 0.
Because tr ε = tr ˜ ε +ε
33
where tr ˜ ε = ε
1
1
+ε
2
2
, we get
ε
33
= −
λ
λ + 2µ
tr ˜ ε.
So
tr ε =
2µ
λ + 2µ
tr ˜ ε.
It follows that
σ = 2µ
¸
ε +
λ
λ + 2µ
Etr ˜ ε
≡
E
1 +ν
¸
ε +
ν
1 −ν
Etr ˜ ε
.
Exercise 7.12 In Cartesian coordinates we get
·· =
2
11
+
2
22
+ 2
2
12
, tr
2
= (
11
+
22
)
2
=
2
11
+
2
22
+ 2
11
22
.
So,
·· +
ν
1 −ν
tr
2
=
1 +
ν
1 −ν
2
11
+
1 +
ν
1 −ν
2
22
+
2ν
1 −ν
11
22
+ 2
2
12
=
1
1 −ν
2
11
+
2
22
+ 2ν
11
22
+ 2
2
12
.
If ν < 1 and
11
= 0,
22
= 0 then
2
11
+
2
22
+ 2ν
11
22
> 0. This proves
the required positivity.
Exercise 7.13 Doubly dotmultiplying T by , we get
T·· = (T· A+ (T· n)n) ·· = (T· A) ·· = T
αβ
αβ
.
Substituting (7.30) into this, we get
T·· =
Eh
1 +ν
¸
+
ν
1 −ν
Atr
·· =
Eh
1 +ν
¸
·· +
ν
1 −ν
tr
2
.
Using the equality
M = −M
αβ
ρ
α
ρ
β
×n,
Hints and Answers 353
we ﬁnd that
M
αβ
= −ρ
α
· M· (ρ
β
×n).
Then
M
αβ
æ
αβ
= −ρ
α
· M· (ρ
β
×n)æ
αβ
= −ρ
α
· M· (ρ
β
×n)ρ
α
· æ· ρ
β
= −M··
(ρ
β
×n)ρ
α
ρ
α
· æ· ρ
β
= M··
(n ×ρ
β
)æ· ρ
β
= M··
(n ×ρ
β
)ρ
β
· æ
= M·· (n ×æ)
= (M×n) ·· æ.
In addition, we obtain the equality
M·· (n ×æ) =
Eh
3
12(1 +ν)
¸
æ·· æ+
ν
1 −ν
tr
2
æ
.
Exercise 7.15 tr
¯
∇˜ v = tr(
¯
∇˜ v)
T
=
¯
∇· ˜ v and tr
¯
∇
¯
∇w =
¯
∇·
¯
∇w. It follows
that tr ε =
¯
∇· ˜ v −z
¯
∇·
¯
∇w.
Exercise 7.16 Let the equation of the bent plate surface be z = w(x
1
, x
2
).
In vector form, the surface is given by
r = x
1
i
1
+x
2
i
2
+nw(x
1
, y
1
).
With slightly diﬀerent notation, this was analyzed in Exercise 5.33; it was
shown that
L =
w
xx
1 +w
2
x
+w
2
y
, M =
w
xy
1 +w
2
x
+w
2
y
, N =
w
yy
1 +w
2
x
+w
2
y
.
We recall that L = b
11
, M = b
12
, N = b
22
are the components of the
curvature tensor B. If w and its derivatives are small, we get
B = w
xx
i
1
i
1
+ w
xy
(i
1
i
2
+i
2
i
1
) +w
yy
i
2
i
2
= −æ.
In other words, the second fundamental tensor of the bent plate surface
coincides with æ up to a diﬀerence in algebraic sign.
Exercise 7.18 Modify the proof in threedimensional elasticity, changing
σ to T· A, and u to ˜ v.
This page intentionally left blank This page intentionally left blank
Bibliography
Ciarlet, Ph. G. (1988). Mathematical Elasticity. Vol. I. ThreeDimensional Elas
ticity (NorthHolland, Amsterdam).
Ciarlet, Ph. G. (1997). Mathematical Elasticity. Vol. II. Theory of Plates (North
Holland, Amsterdam).
Ciarlet, Ph. G. (2000). Mathematical Elasticity. Vol. III. Theory of Shells (North
Holland, Amsterdam).
Ciarlet, Ph. G. (2005). An Introduction to Diﬀerential Geometry with Application
to Elasticity (Springer, Dordrecht).
Chr´ o´scielewski, J., Makowski, J. and Pietraszkiewicz, W. (2004). Statyka i
dynamika pow lok wieloplatowych. Nieliniowa teoria i metoda element`ow
sko´ nczonych (Wydawnictwo IPPT PAN, Warszawa).
Cosserat, E. and Cosserat, F. (1909). Th´eorie des corps d´eformables (Herman et
Fils, Paris). English translation: NASA TT F11, 561 (NASA, Washington,
DC).
Danielson, D. A. (1992). Vectors and Tensors in Engineering and Physics
(Addison–Wesley, New York).
Donnell, L. H. (1976). Beams, Plates and Shells (McGrawHill, New York).
Eremeyev, V. A. and Zubov L. M. (2008). Mechanics of Elastic Shells (in Russian)
(Nauka, Moscow).
Eringen, A. C. (1999) Microcontinuum Field Theory. I. Foundations and Solids
(SpringerVelag, New York).
Green, A. E. and Zerna, W. (1954). Theoretical Elasticity (Clarendon Press, Ox
ford).
Goldenveizer, A. L. (1976). Theory of Thin Elastic Shells (in Russian) (Moscow,
Nauka).
Goodbody, A. M. (1982). Cartesian Tensors, with Applications to Mechanics,
Fluid Mechanics, and Elasticity (Halsted Press, Ellis Horwood, New York).
Heinbockel, J. H. (2001). Introduction to Tensor Calculus and Continuum Me
chanics (Traﬀord Publishing, British Columbia, Canada).
Jeﬀreys, H. (1931). Cartesian Tensors (Cambridge University Press, Cambridge,
UK).
Kay, D. C. (1988). Tensor Calculus (Schaum’s Outline Series, McGrawHill, New
355
356 Tensor Analysis with Applications in Mechanics
York).
Knowles, J. (1997). Linear Vector Spaces and Cartesian Tensors (Oxford Univer
sity Press, Oxford).
Koiter, W. T. On the foundations of the linear theory of thin elastic shells. I.
Proc. Kon. Ned. Ak. Wet. B73, 169–195.
Lebedev, L. P. and Cloud, M. J. (2003). The Calculus of Variations and Functional
Analysis with Optimal Control and Applications in Mechanics (World Sci
entiﬁc, Singapore).
Libai, A. and Simmonds, J. G. (1998). The Nonlinear Theory of Elastic Shells,
2nd ed (Cambridge University Press, Cambridge, UK).
Lipschutz, M. M. (1969). Diﬀerential Geometry (Schaum’s Outline Series,
McGrawHill, New York).
Lurie, A. I. (1990). Nonlinear Theory of Elasticity, Series in Applied Mathematics
and Mechanics, 36 (NorthHolland, Amsterdam).
Lurie, A. I. (2005). Theory of Elasticity (Springer, Berlin).
McConnell, A. (1957). Application of Tensor Analysis (Dover, New York).
Muskhelishvili N. I., (1966). Some Basic Problems of the Mathematical Theory
of Elasticity; Fundamental Equations, Plane Theory of Elasticity, Torsion
and Bending, 5th ed. (in Russian), Nauka, Moscow. Transl. P. Noord
hoﬀ, Groningen, 1953 (Translation of the 3rd ed., Izd. Akad. Nauk SSSR,
MoscowLeningrad, 1949).
Naghdi, P. (1972). The Theory of Plates and Shells. In: Fl¨ ugge, S. (Ed.), Hand
buch der Physik, Vol. VIa/2, SpringerVerlag, Berlin. Pp. 425–640.
Novozhilov, V. V., Chernykh, K. Ph. and Mikhailovskiy E. M. (1991). Linear The
ory of Thin Shells (in Russian) (Politechnika, Leningrad).
Ogden, R. W. (1997). NonLinear Elastic Deformations (Dover, New York).
O’Neill, B. (1997). Elementary Diﬀerential Geometry (Academic Press, New
York).
Papastavridis, J. (1998). Tensor Calculus and Analytical Dynamics (CRC Press,
Boca Raton).
Pogorelov, A. V. (1957). Diﬀerential Geometry. Translated from the ﬁrst Russian
ed. by L. F. Boron (P. Noordhoﬀ, Groningen).
Rubin, M. B. (2000). Cosserat Theories: Shells, Rods and Points (Kluwer, Dor
drecht).
Simmonds, J. G. (1982). A Brief on Tensor Analysis, 2nd ed (Springer, New
York).
Schouten, J. A. (1951). Tensor Analysis for Physicists (Clarendon Press, Oxford).
Sokolnikoﬀ, I. S. (1994). Tensor Analysis: Theory and Applications to Geometry
and Mechanics of Continua (Wiley, New York).
Synge, J. and Schild, A. (1978). Tensor Calculus (Dover, New York).
Timoshenko, S. P., WoinowskyKrieger, S. (1985). Theory of Plates and Shells
(McGraw Hill, New York).
Truesdell, C. and Noll, W. (2004). The Nonlinear Field Theories of Mechanics.
3rd ed (Springer, Berlin).
Wang, C. M., Reddy, J. N. and Lee, K. H. (2000). Shear Deformable Beams and
Shells (Elsevier, Amsterdam).
Bibliography 357
Vlasov, V. Z. (1949). General Theory of Shells and its Applications in Technics
(in Russian) (Moscow, SaintPetersburg, Gostekhizdat).
Vorovich, I. I. (1999). Nonlinear Theory of Shallow Shells (Springer, New York).
Wang, C.C. and Truesdell, C. (1973). Introduction to Rational Elasticity (No
ordhoof Int. Publishing, Leyden).
Young, E. (1993). Vector and Tensor Analysis (Marcel Dekker, New York).
Yosida, K. (1980). Functional Analysis (Springer, Berlin).
Zhilin, P. A. (2006). Applied Mechanics. Foundations of the Theory of Shells (in
Russian) (St. Petersburg State Polytechnical University, St. Petersburg).
Zubov, L. M. and Karyakin M. I. (2006). Tensor Calculus (in Russian) (Vu
zovskaya Kniga, Moscow).
Zubov, L. M. (1982). Methods of Nonlinear Elasticity in the Theory of Shells (in
Russian) (Rostov State University, Rostov on Don).
This page intentionally left blank This page intentionally left blank
Index
6parametric shell theory, 283
acoustic tensor, 228
admissible function, 209
admissible variation, 144
Airy’s function, 230
antisymmetric tensor, 42
Archimedes, 125
astroid, 343
ball tensor, 48
Beltrami–Michell equations, 204
bending stiﬀness, 267
bending strain measure, 253
Betti’s reciprocity theorem, 206
biharmonic equation, 268
body forces, 183
boundary conditions
kinematic, 202, 259
mixed, 202
plate, 269
static, 202, 261
bulk modulus, 198
calculus of variations, 144
fundamental lemma, 145
Cauchy
lemma, 185
stress tensor, 187
theorem, 187
Cayley–Hamilton theorem, 48
Ces`aro formula, 193
Christoﬀel coeﬃcients, 99, 101
for a surface, 154
cofactor, 82
commutative tensors, 80
compatibility condition, 193
complete curvature, 152
conﬁguration, 183
actual, 183
initial, 183
reference, 183
conjugate vector, 44
constitutive equations, 194
contact forces, 183
continuity, 68
contravariant components, 21
coordinate line, 86
covariant components, 21
covariant derivatives, 100
covariant diﬀerentiation, 100
cross product, 22
curl, 97
curvature, 128
curvature tensor, 104, 156, 240
curve
curvature, 128
moving trihedron, 130
normal plane, 130
osculating plane, 130
principal normal, 130
rectifying plane, 130
second curvature, 132
torsion, 132
359
360 Tensor Analysis with Applications in Mechanics
curvilinear coordinates, 86
cycloid, 343
d’Alembert’s principle, 193
Darboux vector, 137
degenerate eigenvalue, 48
derivative, 71
determinant, 34
developable surface, 152, 168
diagonalization, 47
displacement vector, 191
divergence, 97, 156
dot product, 5
double dot product, 39
dual basis, 13
dummy index, 14
Dupin indicatrix, 164
dyad, 4, 29
dynamical equations, 193
eigenfrequencies, 222
eigenoscillation, 222
eigenpair, 44
eigenvalue problem, 222
equal tensors, 30
equilibrium equations
continuum medium, 184, 186, 188
elastic body, 201
micropolar shell, 284
plate, 263
rigid body, 182
shell, 247, 249
Euclid, 125
Euclidean space, 103
Euler equations, 146
even permutation, 23
evolute, 161
external forces, 183
extremal, 146
Fermat’s theorem, 210
ﬁnite rotation vector, 81
ﬁrst fundamental form, 142
ﬁrst fundamental tensor, 240
ﬁrst variation, 145, 211
force density, 183
force(s)
body, 183
contact, 183
external, 183
free index, 14
Friedrich inequality, 220
function(s)
isotropic, 63, 65
linear, 61
of tensorial argument, 60
symmetric tensor, 75
functional, 144, 209
Gˆateaux derivative, 71, 211
Gauss–Ostrogradsky theorem, 117,
205, 207, 212, 221–223, 225, 226,
260, 273
surface analog, 178
Gaussian curvature, 152
generalized Hooke’s law, 195
generalized orthogonality, 47
generator, 152
geodesic, 146
Germain–Lagrange equation, 268
Gibbsian cross, 249
gradient, 71, 96
Gram determinant, 17, 218
Gram matrix, 17
inner product, 66
integration by parts, 117
intrinsic coordinates, 139
inverse tensor, 32
irrotational ﬁeld, 108
isometric surfaces, 167
isotropic function, 63, 65
isotropic tensor, 59
kinematic boundary conditions, 202,
259
Kirchhoﬀ’s hypotheses, 249, 266
kinematic, 249
static, 249
Kirchhoﬀ–Love hypotheses, 249
Korn’s inequality, 220
Kronecker delta, 13
Index 361
Lagrange identity, 24
Lagrange’s principle, 271
Lagrange’s variational principle, 214,
271, 285
Lam´e coeﬃcients, 113
Lam´e’s moduli, 197
level surface, 140
Levi–Civita tensor, 57
limit, 68
lines of curvature, 167
lowering indices, 17
mean curvature, 152
meridian, 168
metric tensor, 16
Meusnier’s theorem, 151
micropolar shells, 283
minimizer, 144
mixed boundary conditions, 202
moving trihedron, 130
multiple eigenvalue, 48
nabla operator, 95
natural equations, 170
natural parametrization, 127
nonsingular tensor, 34, 52
norm, 66
normal curvature, 151
normal plane, 130
oblique coordinates, 94
odd permutation, 23
order of contact, 160
orthogonal representation, 47
orthogonal tensor, 49
oscillation mode, 223
osculating plane, 130
parallel, 168
permutation symbol, 23
physical components, 91
plane elasticity problems, 230
plane strain, 230
plane stress, 230
plane wave, 227
plate, 237, 263
bending stiﬀness, 267
boundary conditions, 269
equilibrium equations, 263
Germain’s equation, 268
higherorder theory, 280
nonclassical theory, 277
Reissner’s equations, 277
rigid motions, 270
strain energy, 270
thirdorder theory, 281
uniqueness of solution, 275
point of inﬂection, 131
Poisson’s ratio, 198
polar decomposition, 53
positive deﬁnite tensor, 53
premultiplication, 31
principal axes, 47, 190
principal directions, 164
principal normal, 130
principal stresses, 190
projection dyad, 30
quotient law, 31
radius of curvature, 131
raising indices, 17
Rayleigh’s quotient, 226, 286
Rayleigh’s variational principle, 221,
286
reciprocal basis, 13
rectifying plane, 130
Reissner’s plate theory, 277
resultant force, 182
resultant moment, 182
Riemann–Christoﬀel tensor, 104
righthand rule, 22
Ritz’s method, 216
rotation, 97, 156
scalar product, 40
scalar triple product, 6
Schwarz inequality, 67
second curvature, 132
second fundamental form, 149
second fundamental tensor, 240
shear correction factor, 279
362 Tensor Analysis with Applications in Mechanics
shear modulus, 197, 198
shell, 237
base surface, 249
boundary conditions, 259
deformation, 249
direct approach, 238
energy, 256
equilibrium equations, 244
higherorder theory, 280
hypothesis approach, 238
midsurface, 249
nonclassical theory, 277
strain, 249
thickness, 243
singular point, 128
singular tensor, 34
solenoidal ﬁeld, 110
solidiﬁcation principle, 182
static boundary conditions, 202, 261
stationary point, 210, 211
Stokes’s formula, 119
strain energy, 195
strain energy function, 195
strain tensor, 192
stress couple tensor, 248
stress resultant, 246
stress resultant tensor, 246
stress tensor, 187
stress vector, 184
sum of tensors, 31
summation convention, 14
surface
complete curvature, 152
Gaussian curvature, 152
gradient operator, 156, 240
mean curvature, 152
normal curvature, 151
surface gradient, 156
symmetric tensor, 42
tangential strain measure, 253
tensor ﬁeld, 85
tensor of changes of curvature, 253
tensor of elastic moduli, 195
tensor product, 5, 29
tensor(s)
acoustic, 228
antisymmetric, 42
commutative, 80
diagonalization of, 47
equal, 30
inverse, 32
isotropic, 59
Levi–Civita, 57
nonsingular, 34, 52
of fourth order, 5
of second order, 4
operator viewpoint, 30
orthogonal, 49
orthogonal representation, 47
positive deﬁnite, 53
principal axes, 47
resultant stress, 246
singular, 34
strain , 192
stress, 187
stress couple, 248
sum of, 31
symmetric, 42
trace of, 51
transpose, 40
unit, 31
thinwalled structure, 237
torque, 22
torsion, 132
total energy, 209
trace, 51
transpose, 40
transverse shear stress resultants, 247
triads, 56
unit basis vectors, 91
unit tensor, 31
variational principle
Lagrange’s, 214, 271, 285
Rayleigh’s, 221, 286
vector ﬁeld, 85
vector(s), 3
contravariant components, 21
covariant components, 21
cross product of, 22
Index 363
dot product of, 5
unit basis, 91
vectorial invariant, 80
Vi`ete formulas, 45
virtual displacement, 211
Voigt’s rule, 196
wave vector, 227
wave velocity, 227
waves
dilatational, 229
longitudinal, 229
plane, 227
shear, 229
transverse, 229
weak solution, 216
Young’s modulus, 194, 198
zero tensor, 31
Tensor Analysis with Applications in Mechanics
This page intentionally left blank
Tensor Analysis with Applications in Mechanics
Leonid P. Lebedev
National University of Colombia Southern Federal University, Russia
Michael J. Cloud
Lawrence Technological University, USA
Victor A. Eremeyev
MartinLutherUniversity HalleWittenberg, Germany Southern Scientific Center of Russian Academy of Science Southern Federal University, Russia
World Scientific
NEW JERSEY
•
LONDON
•
SINGAPORE
•
BEIJING
•
SHANGHAI
•
HONG KONG
•
TA I P E I
•
CHENNAI
Published by World Scientific Publishing Co. Pte. Ltd. 5 Toh Tuck Link, Singapore 596224 USA office: 27 Warren Street, Suite 401402, Hackensack, NJ 07601 UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
British Library CataloguinginPublication Data A catalogue record for this book is available from the British Library.
TENSOR ANALYSIS WITH APPLICATIONS IN MECHANICS Copyright © 2010 by World Scientific Publishing Co. Pte. Ltd. All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means, electronic or mechanical, including photocopying, recording or any information storage and retrieval system now known or to be invented, without written permission from the Publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from the publisher.
Desk Editor: Tjan Kwang Wei
ISBN13 9789814313124 ISBN10 9814313122
Printed in Singapore.
Foreword
Every science elaborates tools for the description of its objects of study. In classical mechanics we make extensive use of vectorial quantities: forces, moments, positions, velocities, momenta. Conﬁning ourselves to a single coordinate frame, we can regard a vector as a ﬁxed column matrix. The deﬁnitive trait of a vector quantity, however, is its objectivity; a vector does not depend on our choice of coordinate frame. This means that as soon as the components of a force are speciﬁed in one frame, the components of that force relative to any other frame can be found through the use of appropriate transformation rules. But vector quantities alone do not suﬃce for the description of continuum media. The stress and strain at a point inside a body are also objective quantities; however, the speciﬁcation of each of these relative to a given frame requires a square matrix of elements. Under changes of frame, these elements transform according to rules diﬀerent from the transformation rules for vectors. Stress and strain tensors are examples of tensors of the second order. We could go on to cite other objective quantities that occur in the mechanics of continua. The set of elastic moduli associated with Hooke’s law comprise a tensor of the fourth order; as such, these moduli obey yet another set of transformation rules. Despite the diﬀerences that exist between the transformation laws for the various types of objective quantities, they all ﬁt into a uniﬁed scheme: the theory of tensors. Tensor theory not only relieves our memory from a huge burden, but enables us to carry out diﬀerential operations with ease. This is the case even in curvilinear coordinate systems. Through the unmatched simplicity and brevity it aﬀords, tensor analysis has attained the status of a general language that can be spoken across the various areas of continuum physics. A full comprehension of this language has become necessary for those working
v
vi
Tensor Analysis with Applications in Mechanics
in electromagnetism, the theory of relativity, or virtually any other ﬁeldtheoretic discipline. More modern books on physical subjects invariably contain appendices in which various vector and tensor identities are listed. These may suﬃce when one wishes to verify the steps in a development, but can leave one in doubt as to how the facts were established or, a fortiori, how they could be adapted to other circumstances. On the other hand, a comprehensive treatment of tensors (e.g., involving a full excursion into multilinear algebra) is necessarily so large as to be ﬂatly inappropriate for the student or practicing engineer. Hence the need for a treatment of tensor theory that does justice to the subject and is friendly to the practitioner. The authors of the present book have met these objectives with a presentation that is simple, clear, and suﬃciently detailed. The concise text explains practically all those formulas needed to describe objects in threedimensional space. Occurrences in physics are mentioned when helpful, but the discussion is kept largely independent of application area in order to appeal to the widest possible audience. A chapter on the properties of curves and surfaces has been included; a brief introduction to the study of these properties can be considered as an informative and natural extension of tensor theory. I.I. Vorovich Late Professor of Mechanics and Mathematics Rostov State University, Russia Fellow of Russian Academy of Sciences (1920–2001)
formulate boundary value problems. hence we could assume a background consisting only of standard calculus and linear algebra. The corresponding formulas in tensor notation are quite simple. and discuss the problem of uniqueness of solutions. We should mention the essential nature of tensors in elasticity and shell theory. vii . The ﬁrst ﬁve chapters are largely unmodiﬁed. and physicists who must employ tensor techniques. whereas the integrated exercises are required for a proper understanding of the text. sometimes this takes a few pages. adding two chapters on applications for which tensor analysis is the principal tool. Of course. Lagrange’s variational principle. aside from some new problem sets and material on tensorial functions needed for the chapters on elasticity. we had to select those questions in elasticity where the role of tensor analysis is most crucial. one should write things out in component form.Preface The ﬁrst edition of this book was written for students. Space restrictions prohibited us from presenting an entire course on mechanics. allowing us to grasp the underlying ideas and perform manipulations with relative ease. and some problems on vibration. The endofchapter problems are supplementary. We present complete derivations of the equations in these theories. Because tensor representation leads quickly and painlessly to componentwise representation. but rather presented a threedimensional version (which is easy to extend to n dimensions). One chapter is on linear elasticity and the other is on the theory of shells and plates. this technique is ideal for presenting continuum theories to students. engineers. We have decided to extend the book in a natural direction. We did not present the material in complete generality for the case of ndimensional space. to solve a certain engineering problem.
we make no attempt at a comprehensive reference list. Mechanics and Computer Sciences South Federal University. Cloud Department of Electrical and Computer Engineering Lawrence Technological University. that extend the treatment slightly. Because the book is largely selfcontained. This was common in the older literature. We are deeply grateful to our World Scientiﬁc editor. We merely list certain books that cover similar material. This independence means that if we deﬁne such properties using one coordinate system. then in another system we can recalculate these characteristics using valid transformation rules. or that may be otherwise useful to the reader. Colombia M. Eremeyev South Scientiﬁc Center of RASci & Department of Mathematics. the term “rank” is often assigned a diﬀerent meaning. Over a period of many years.P.J. Mr. L. It was found that the use of vectors and tensors led to a proper description of certain properties and behaviors of real natural objects: those aspects that do not depend on the coordinate systems we introduce in space. In the newer literature. Lebedev Department of Mathematics National University of Colombia.A. this naive viewpoint evolved into the modern interpretation of the notion of vector and its extension to tensors. Tjan Kwang Wei. Russia Preface to the First Edition Originally a vector was regarded as an arrow of a certain length that could represent a force acting on a material point. The ease . for his encouragement and support. USA V.viii Tensor Analysis with Applications in Mechanics In the ﬁrst edition we used the term rank instead of order.
Tjan Kwang Wei and Ms. or simply ask the computer algorithm to choose the best frame and method. Such compact forms are easy to remember. Many physical laws are cumbersome when written in coordinate form but become compact and attractive looking when written in tensorial form. vectors and tensors will become important elements of the macrolanguage for the next generation of software in engineering and applied mathematics. This book provides the tools necessary for such calculation. Professor Byron C. or mixed). and understand how to recalculate results in other systems. Indeed.P. So in applications we must apply various coordinate systems. numerical. SookCheng Lim — for their assistance in the production of this book. We would like to thank the editorial staﬀ at World Scientiﬁc — especially Mr. presumably the user will be able to initiate a solution process in a certain frame and by a certain method (analytical. Natasha Lebedeva and Beth LannonCloud deserve thanks for their patience and support.J. USA . and can be used to state complex physical boundary value problems. Cloud Department of Electrical and Computer Engineering Lawrence Technological University. Drachman of Michigan State University commented on the manuscript in its initial stages. Lebedev Department of Mechanics and Mathematics Rostov State University. It is conceivable that soon an ability to merely formulate statements of boundary value problems will be regarded as a fundamental skill for the practitioner. derive corresponding equations. computer software is slowly advancing toward the point where the only necessary input data will be a coordinatefree statement of a boundary value problem. Lastly. In this way. Colombia M. L.Preface ix with which a given problem can be solved often depends on the coordinate system employed. Russia & Department of Mathematics National University of Colombia.
This page intentionally left blank .
. . . . . . . .2 1. . . . . . . . . . 1 3 3 4 5 7 9 11 . . . . . . . . . . . . . xi . . . . . . . .7 2. . . . .8 2. . . . .2 2. . . .5 2. . Tensors 3. . . . . . . Covariant and Contravariant Components The Cross Product in Index Notation . . . . . . . . . Transformation to the Reciprocal Frame . . . . . . .4 1. . Transformation Between General Frames . . . . .4 2. . . . . . . . . A First Look at Tensors . . . . . . . . . . . .1 Dyadic Quantities and Tensors . . . . . . . . . . . . . . .1 2. . . . . . . . . . . Closing Remarks . . . . . .Contents Foreword Preface v vii Tensor Analysis 1. . . . . . . . . . .6 2. . . . . . More on the Notion of a Vector Problems . . . . . . . . . . . . . . . . . . Preliminaries 1. . . . . . . .1 1. . . . .3 2. . . .5 The Vector Concept Revisited . . . . . . . 3. . . .3 1. . . . . Norms on the Space of Vectors . . . . . . . . . . Problems . . . . . 11 12 17 18 21 22 24 27 27 29 29 2. . Dual Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . Assumed Background . . . . . . . . . . .9 Change of Basis . . . . . . . . . . . . . . . Transformations and Vectors 2. . . . . . . . .
. . . . . .6 5. . . .4 5. . . . The Principal Curvatures of a Surface . . . . . . . . . . . . .11 3. . . . . . . . . . . .2 5. . . . . . . . . . . . . The Torsion of a Curve . . . . . . . . . . Derivation Formulas . . . . .6 4. . . . . . . . . . . . . . . . .3 5. . . . 126 132 135 137 148 153 156 162 164 . . Christoﬀel Coeﬃcients and their Properties . . . .2 4. . . . . . . .3 4. . . . . . .5 5. . 85 94 98 99 100 105 106 108 113 117 119 125 5. . . . . . . . .10 4. . . .1 5. . . . . . . . 30 34 36 40 44 48 49 56 58 60 66 70 77 85 4. . . . . . . . . . .3 3. . Problems . . . . . . . . Extending the Dyad Idea . .4 3. . . . . . . . .9 Elementary Facts from the Theory of Curves . . . . . . .9 4. . . . . . . .8 5. . . . Eigenvalues and Eigenvectors of a SecondOrder Symmetric Tensor . . Other Properties of SecondOrder Tensors . . . . .10 3. Dyadic Components Under Transformation . . . . . . . . . . . . . The Second Fundamental Form of a Surface . . . . . . . . . . . .13 3. More Dyadic Operations . . . . . . . . . . Diﬀerentiation of a Vector Function .11 Vector Fields . . . . . . . Covariant Diﬀerentiation . . . . . . . . . . . . . . Derivatives of the Frame Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8 4. . . . Implicit Representation of a Curve. . . Covariant Derivative of a SecondOrder Tensor Diﬀerential Operations . . . . . . Norms for Tensors. . Frenet–Serret Equations . . . . . . . . . . . . . . . . . . .2 3. . . . . . Elements of the Theory of Surfaces . . . .14 Tensors From an Operator Viewpoint . . .12 3. . . . Properties of SecondOrder Tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Some Formulas of Integration . . . . . . . . . .xii Tensor Analysis with Applications in Mechanics 3. .6 3. . . . . . . . . . . . Problems .5 3. . . . . . . . . . . . .7 5. . . . . .5 4.8 3. . . . . . . . . . . .4 4. . . . . . . . . . . . . . The Cayley–Hamilton Theorem . . . . . . Orthogonal Coordinate Systems . . Elements of Diﬀerential Geometry 5. . .7 3. . . Diﬀerentials and the Nabla Operator . . . . . . . . . . . . Tensors of the Fourth and Higher Orders . . . . . . . . .7 4. Contact of Curves Osculating Paraboloid . . . . Diﬀerentiation of Tensorial Functions . .9 3.1 4. . . . . . . . . . . Functions of Tensorial Arguments . Tensor Fields 4. . . . . . . . . . . . . and Some Spaces . . . . .
. . . . . . . .10 6. .7 6.2 6. . . . . . . . . . . . . . . . . . . . . . . 168 170 173 175 175 Applications in Mechanics 6. . . . . Linear Elastic Shells 7. . . . . . . . . . . . Hooke’s Law . . . . 239 242 244 249 256 259 261 263 277 287 7. . . . . . . . . . . . . . . Problems . Betti’s Reciprocity Theorem .15 Stress Tensor .8 7. Equilibrium Equations in Displacements . . . . . .6 7. . . . . . . . . . . . . . . . . . . . . . . . . . . .12 6. . . . . . . .13 5. . . . . . . . . . . . . . . . . . . . .9 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . On NonClassical Theories of Plates and Shells . . . . . . Plane Problems of Elasticity . . . . Linear Elasticity 6. . . . . . . . . . . . .13 6. . . Problems . . .1 6. .4 7. Appendix A Formulary . Uniqueness of Solution for the Boundary Value Problems of Elasticity . . . Kinematics in a Neighborhood of Σ . . A Few Remarks on the Kirchhoﬀ–Love Theory . .4 6. . . . . . . . . . . . . . . . . Strain Tensor . . . . . . . . . . .14 Surfaces of Revolution Natural Equations of a A Word About Rigor . . . .7 7. . . . . . . . . . . . Ritz’s Method . . . . . . . . . . . . . . . . . . .2 7. . . . . . . . . . . . .1 7. . . . Shell Deformation and Strains. . .11 5. . . . . . . Plate Theory . . . . . . . . Boundary Conditions . . . . . Equilibrium Equations in Stresses . . . . . . . . . . . . . . . . . . . . . . . . . Boundary Conditions and Boundary Value Problems . Shell Equilibrium Equations . . . Rayleigh’s Variational Principle . . .9 Some Useful Formulas of Surface Theory . . . . . . . . . . . . . . . . . . . . . . . . . Kirchhoﬀ’s Hypotheses Shell Energy . . . .10 5. . . . . . . . . .Contents xiii 5. . . . . . .5 6. . . . . . . . . . . . . . . . . . Plane Waves . . . .3 7. . . . . . . .5 7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Curve . . . . . . . . . . . . . . . . 179 181 181 190 193 194 200 202 203 205 206 208 216 221 227 230 232 237 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 6.14 6. . . . . . Minimum Total Energy Principle . . . . . Conclusion . . . . . . . . . . .12 5. . . . .11 6. . .8 6.6 6. . . . . . . . . . . . . . . Equation of Motion .
xiv Tensor Analysis with Applications in Mechanics Appendix B Hints and Answers Bibliography Index 315 355 359 .
3) for example. The simplest way is to incorporate the frame vectors ei (i = 1. This force does not depend on the coordinate frame employed by someone writing a textbook on vectors somewhere in Russia or China.Chapter 1 Preliminaries 1. 1. we should be able to write it in such a form that it ceases to depend on the details of the coordinate frame. In elementary mathematics these are supposed to be mutually orthogonal and of unit length. 3) explicitly into the notation: if x is a vector we may write 3 x= i=1 xi ei . because such a frame can be rotated. This of course means that we cannot change only the 3 . we should do so in such a way that x remains the same. 1.1) Then if we wish to change the frame. however. 2. (1. In the description of many common phenomena we deal with vectorial quantities like forces that have deﬁnite directions and magnitudes. even then they are not fully characterized.1 The Vector Concept Revisited The concept of a vector has been one of the most fruitful ideas in all of mathematics. But let us examine this situation more closely. 3 determine the vector x? They surely do if we specify the basis vectors of the coordinate frame. Do the components 2. Because the vector f representing a particular force is something objective. Students in elementary mathematics deal with vectors in component form — with quantities such as x = (2. and it is not surprising that we receive repeated exposure to the idea throughout our education. An example is the force your body exerts on a chair as you sit in front of the television set.
then in a Cartesian frame the operation should resemble the multiplication of a matrix by a column vector. In textbooks on mechanics the reader can ﬁnd lengthy discussions on how to change the matrix elements to maintain the same objective characteristic of the body when the new frame is also Cartesian. For example. . Let us note that in mechanics there are objects of another nature. determining how the body rotates when torques act upon it. the tensor of inertia is described by a 3 × 3 matrix. These are quantities of the form ei ej .2 A First Look at Tensors In what follows we shall discuss how to work with vectors using diﬀerent coordinate frames. etc.e. So the components of a vector x in a new frame are not independent of those in the old frame. strain. there is a socalled tensor of inertia. They characterize certain properties of a body at each point. The precise meaning of the term order will be explained later. This is an objective characteristic of a solid body. If we change the frame. say.1) and notice the similarity in notation. their “components” should transform in such a way that the tensors themselves do not depend on the frame. If we introduce. again. the matrix elements change according to certain rules. Similarly. To this end we consider dyads of vectors. Many such tensors of the second order arise in continuum mechanics: tensors of stress. For both vectors and tensors we can introduce various operations. We compare with equation (1..j where the aij are the components of the tensor. dot multiplication of a tensor by a vector. frameindependent). A tensor may then be represented as aij ei ej i. the introduction of any new operation should be done in such a way that the results agree with known special cases when such familiar cases are met. Although the tensor of inertia is objective (i. the multiplication of two tensors should be deﬁned so that in a Cartesian frame the operation involves matrix multiplication.4 Tensor Analysis with Applications in Mechanics frame vectors ei : we must change the components xi correspondingly. If the body is considered in a Cartesian frame. Of course. 1. it is not a vector: it belongs to another class of mathematical objects.
Note that e2 e1 and e1 e2 are diﬀerent objects. Also recall that i1 i2 i3 a × b = a1 a2 a3 . We will discuss this further in Chapter 3. we have a · b = a 1 b 1 + a2 b 2 + a3 b 3 . where a and b are the magnitudes of a and b and θ is the (smaller) angle between a and b. For example.Preliminaries 5 The quantity ei ej is also called the tensor product of the vectors ei and ej . i2 .3 Assumed Background In what follows we suppose a familiarity with the dot and cross products and their expression in Cartesian frames. b. The elements change according to certain rules if we change the frame. Tensors also occur in electrodynamics. a tensor whose elemental parts are of the form abcd. in the same way that e1 is an elemental object belonging to the set of vectors. i3 where a and b are expressed as a = a 1 i 1 + a2 i 2 + a3 i 3 . Our notation (without the symbol ⊗) emphasizes that. however. This means that in general the properties of a body are given by a “table” consisting of 3 × 3 × 3 × 3 = 81 elements. Natural objects can possess characteristics described by tensors of higher order. The term “tensor product” indicates that the operation shares certain properties with the product we know from elementary algebra. From now on we reserve the symbol i for the basis vectors of a Cartesian system.. Recall that if a and b are vectors. In a Cartesian frame with basis vectors i1 . the elastic properties of a body are described by a tensor of the fourth order (i. and is sometimes denoted ei ⊗ ej . b1 b2 b3 b = b1 i 1 + b2 i 2 + b3 i 3 . 1. where a.e. then by deﬁnition a · b = ab cos θ. c. and other sciences that deal with objects situated or distributed in space. for example. d are vectors). e1 e2 is an elemental object belonging to the set of secondorder tensors. the general theory of relativity. .
1). e. b. 1.g.. c as three of its coterminal edges (Fig. This may be interpreted as the volume of the parallelepiped having a. and a fuller discussion will appear in § 2.6 Tensor Analysis with Applications in Mechanics The dot product will play a role in our discussion from the very beginning. a = i 1 a1 + i 2 a2 + · · · + i n an . and shall continue this practice throughout most of the book. b.6. The cross product will be used as needed. . and c? (Hints for this and many other exercises appear in Appendix B beginning on page 315. c1 c2 c3 Permissible manipulations with the scalar triple product include cyclic interchange of the vectors: a · (b × c) = c · (a × b) = b · (c × a).1. It is also possible (and sometimes useful) to introduce vectors in a more general space of n > 3 dimensions.1 Geometrical meaning of the scalar triple product. What does the condition a · (b × c) = 0 say about a. 1. In rectangular components we have. Given three vectors a. c b a Fig. a 1 a2 a3 a · (b × c) = b1 b2 b3 . Exercise 1.) So far we have made reference to vectors in a threedimensional space. b. c we can form the scalar triple product a · (b × c). according to the expressions above.
in Rn one can multiply any vector by any real number and obtain another vector in the space. widths. etc. Total electric currents cannot be considered as vector quantities since they do not combine according to the rule for vector addition. forces have some rather strange features: for example. The notion of a vector was eventually elaborated in physics. 1.4 More on the Notion of a Vector Before closing out this chapter we should mention the notions of a vector as a “directed line segment” or “quantity having magnitude and direction. Such simplistic deﬁnitions can be dangerous.” We ﬁnd these in elementary math and physics textbooks. if we shift the point . However. In the tailor’s case this notion is completely senseless. the orientation of the wire. Another problem concerns the notion of an ndimensional Euclidean space. it is not reasonable to regard every list of n things as a possible vector. The total electric current ﬂowing in a thin wire is one example: to describe the current we must specify the rate of ﬂow of electrons. A tailor may take measurements for her client and compose an ordered list of lengths. A tailor’s list showing that someone is six meters tall would be pretty hard to ﬁnd. It is also true that many threedimensional notions can be reconsidered in two dimensions. However.Preliminaries 7 It turns out that many (but not all) of the principles and techniques we shall learn have direct extensions to such higherdimensional spaces. But in Rn any two points are joined by a vector whose components equal the diﬀerences of the corresponding coordinates of the points. if two wires lie in a plane and carry equal currents running perpendicular to each other. To make matters worse. But it is easy to point to a situation in which a quantity of interest has magnitude and direction but is not a vector.. The reader should take the time to reduce each threedimensional formula to its twodimensional analogue to understand what happens with the corresponding assertion. and by the above “deﬁnition” the set of all such lists is an ndimensional space. However. We sometimes hear it deﬁned as the set each point of which is uniquely determined by n parameters. and the sense of electron movement along the wire. more precisely in mechanics where the behavior of forces was used as the model for a general vector. then we cannot duplicate their physical eﬀects by replacing them with a third wire directed at a 45◦ angle with respect to the ﬁrst two.
each pair of points that correspond under the mapping deﬁne a vector x under which a point shifts into its image. The result of this map is the same space of points. In elementary mathematics students use vectors and points quite interchangeably. However. In the classical mechanics of a rigid body we are allowed to move a force along its line of action but cannot simply shift it oﬀ that line. This vector is what we ﬁnd when we subtract the Cartesian coordinates of the initial point from those of the ﬁnal point. With a deformable body we cannot move a force anywhere because in doing so we immediately change the state of the body. So the mechanical forces that gave birth to the notion of a vector possess features not covered by the mathematical deﬁnition of a vector. but will not necessarily obtain a new vector that is relevant to the physics of the situation. We can then consider a vector x as a shift of all points of the point space by an amount speciﬁed by x. For example. but mathematical rigor has proved its worth from many practical points of view. So we should understand that the idea of a vector in mathematics reﬂects only certain features of the real objects it describes. On the other hand. then we get a strict deﬁnition of the space introduced intuitively in elementary mathematics. We would like to mention something else about vectors. This brings us to our ﬁnal word. then we must introduce a moment acting on the body or the motion of the body will change. We can vectorially add two forces acting on a rigid body (not forgetting about the moments arising during shift of the forces). associate with a vector in an ndimensional Euclidean vector space a point in an ndimensional Euclidean point space. Engineers might object to the imposition of such formality on an apparently simple situation. a computer that processes information in the complete absence of intuition can deal properly with objects for which rigorous deﬁnitions and manipulation rules have been formulated. these are objects of diﬀerent natures: there are vectors in space and there are the points of a space. The simple and compact notation x for a vector leads to powerful ways of exhibiting relationships between the many vector (and tensor) quantities that occur in mathematical physics. regarding the expression of vectors in componentfree notation. We can. If we add the fact that the composition of two such maps obeys the rules of vector addition.8 Tensor Analysis with Applications in Mechanics of application of a force acting on a solid body. It permits us to accomplish manipulations that would take many pages and become quite confusing if done in component form. This is typical for problems of . if two forces act on two diﬀerent material points then we can add the vectors that represent the forces. for instance.
1. (d) a = (−1. 1. (h) a = (−1. (c) a = (1. 3. (c) a1 = (1.Preliminaries 9 nonlinear physics. 2). a2 = (3. a2 . (c) x + 2a + 16(x − b) + c = 2(c + 3x) + a − x. 2. 2. 3). 2). a3 = (1. −4). 2). 2). −2). b = (2. 1). 1). 1. 2. 2. b = (3. 1. 2. a2 = (4. (g) a = (1. 1). (b) i2 × i3 = i1 . 1). 4). b = (−1.3 Find x satisfying the equation (a) a + 2x − 4b = 0. −4). 1. −1). 0). 1. −2). 0). 1. (b) a = (1. (e) a = (0. 3. −1). 1. 3. −1). 2). . a2 = (5. −2).2 Find a1 + 3a2 − 2a3 for the vectors a1 . 1. 2). 1. 2. 1). b = (1. 1. (b) a1 = (1. 1. a2 = (1. (f) a = (−1. 1. a3 given by (a) a1 = (−1. (i) a = (1. 0).1 Find 2a1 + 3a2 for the vectors a1 and a2 given by (a) a1 = (1. (c) a1 = (2. −3. 0). a2 = (2. 1. b = (5. 4). The resulting formal nature of the manipulations means that computers can be expected to take on more and more of this sort of work. 3. a3 = (1. 1. −1. −1. and for those where change of coordinate frame becomes necessary. 3.5 Problems 1. −1. −2). 2). a2 = (1. b = (2. even at the level of original research. a3 = (3. 2. b = (1. (b) a1 = (−1. 0). b = (1.5 Show that (a) i1 × i2 = i3 . −1). 1. 1. −1.4 Find the values of a · b and a × b if (a) a = (0. 0). 1. 1. 3). 1. 2). b = (2. 3. (b) 2(a + 2x) + b = 3(b + x). (c) i3 × i1 = i2 .
3). 3. 1). −3. −2). 2.10 Tensor Analysis with Applications in Mechanics 1. a4 = (1. a3 = (1. −2). 1. 2). a2 = (1. 2). (e) a1 = (9. a2 = (1. 2). 6). 0. −3. −1. 2. 2. −2). a3 = (1. 2). 1). 1. i2 . a3 = (1. 3). 3). 6). b = b1 i1 + b2 i2 . 2). a4 = (2. 2). a2 = (1. a4 = (1. −2). a3 = (3. −1. 1. a3 = (1. 2. (d) a1 = (−1. a2 = (7. 2). 1. 8. (g) a1 = (−1. a3 = (1. 2. 2. a3 = (1. 0). a3 = (3. a3 = (1. a2 = (1. a2 = (7.10 Find a1 · (a2 × a3 ) when (a) a1 = (−1. a2 = (1. 0). a4 = (1. −1. 1. 4. −1. 2. a2 = (−6. a4 = (2. 4). 3. −1. 4). 3). (c) i1 · (i3 × i1 ). a4 = (1. 2. (g) a1 = (−1. −1. a2 = (2. 3). a2 = (2. −2).6 Show that the equation a × i1 = −a2 i3 + a3 i2 holds for an arbitrary vector a = a1 i1 + a2 i2 + a3 i3 . 3. −1. 0.7 Let a = a1 i1 + a2 i2 . 0. (b) a1 = (−1. 3. a4 = (1. 3). 1. a3 = (1. −1. −2. 2). a2 = (1. (c) a1 = (1. (i) a1 = (2. 0). (c) a1 = (1. a4 = (0. (i) a1 = (2. 3). 3). (h) a1 = (1. 3). (f) a1 = (1. a3 = (1. 1. 1). 1. 12. 0). and a = 0. −1. a2 = (1. 1. a3 = (5. 4). 2). 1.11 Find (a1 × a2 ) · (a3 × a4 ). 8. 1. −1. 1. a2 = (7. 2). (f) a1 = (1. 1. 2). (d) a1 = (−1. 1. −1. 2.8 Suppose a × x = 0. 3. 1). Demonstrate that x = 0. −2). −1. 2). a3 = (3. −1. 2). 1. −1). 0). −6). −1). 2). 1. 6). 2). 1). −3. 3). a2 = (1. a · x = 0. 2). 2). a2 = (1. −3. a2 = (1. 4. 2. a4 = (0. 2). and let i1 . Show that a × b = (a1 b2 − a2 b1 )i3 . where (a) a1 = (−1. −2. 1). a3 = (1. 3). 1. 3). 2. −1). a2 = (7. −3. a3 = (5. 2. 2. . i3 be a Cartesian basis. 1). 3).9 Find: (a) i1 · (i2 × i3 ). (b) i1 · (i3 × i2 ). 6). (h) a1 = (1. −2. 1. a2 = (4. −1. a3 = (3. 2). a3 = (1. 2). 2). (b) a1 = (−1. 3. (e) a1 = (9. a3 = (1. 2.
e e e e3 = 7˜1 + 8˜2 + 9˜3 . Fully written out in a given Cartesian frame ei (i = 1. 1. we can write e e e e e e e e x = 2(˜1 + 2˜2 + 3˜3 ) + (4˜1 + 5˜2 + 6˜3 ) + 3(7˜1 + 8˜2 + 9˜3 ) e e e = (2 + 4 + 21)˜1 + (4 + 5 + 24)˜2 + (6 + 6 + 27)˜3 e = 27˜1 + 33˜2 + 39˜3 . e e e In these calculations it is unimportant whether the frames are Cartesian. it is x = 2e1 + e2 + 3e3 . 7 8 9 11 . 3).) Suppose we appoint a new frame ei (i = 1. 3). it is important only that we have the table of transformation ⎛ ⎞ 1 2 3 ⎝ 4 5 6 ⎠. ˜ From these expansions we could calculate the ei and verify that they are noncoplanar. frameindependent entity. 2. 3) such that ˜ e e e1 = e1 + 2˜2 + 3˜3 .Chapter 2 Transformations and Vectors 2. e e e e2 = 4˜1 + 5˜2 + 6˜3 .1 Change of Basis Let us reconsider the vector x = (2. As x is an objective. (This is one of the few times we do not use i as the symbol for a Cartesian ˜ frame vector. 2.
In a Cartesian frame the necessary operation is simple dot multiplication by the corresponding basis vector of the frame: we have xk = x · ik (k = 1. This procedure fails in a more general nonCartesian frame where we do not necessarily have ei · ej = 0 for all j = i. placing some indices as subscripts and some as superscripts. 3) . i So in the new basis we have 3 3 x= j=1 xj ej ˜ ˜ where xj = ˜ i=1 Aj xi . Although this practice may seem artiﬁcial. no matter how advanced we are. However.2 Dual Bases To perform operations with a vector x. Let x be of the form 3 x= i=1 xi ei (2. we must have a straightforward method of calculating its components — ultimately. we must be able to obtain the xi using simple arithmetic. we shall need these when doing frame transformations. 2. there are fairly deep reasons for following it. 3).12 Tensor Analysis with Applications in Mechanics It is clear that we can repeat the same operation in general form. it may still be possible to ﬁnd a vector ei such that xi = x · ei (i = 1. 2. We prefer formulas that permit us to ﬁnd the components of vectors using dot multiplication only. i˜ Then 3 3 3 3 x= i=1 x i j=1 Aj ej i˜ = j=1 ˜ ej i=1 Aj xi . 2. etc.1) with the table of transformation of the frame given as 3 ei = j=1 Aj ej . i Here we have introduced a new notation.
If we set ⎛ ⎞ xi = x · ei = ⎝ 3 xj ej ⎠ · ei = 3 xj (ej · ei ) j=1 j=1 and compare the left.2) we get xj . In a Cartesian frame we have ek = e k = i k for each k. we see that equality holds when i e j · e i = δj (2. This explains why the frames ei and ei are dual: the formulas x · ei = xi . x · ei = xi . but in a way complementary to that used in equation (2. Exercise 2. (2. 0.Transformations and Vectors 13 in this more general situation. for example. . the equations for e1 : e1 · e1 = 1. and its magnitude is such that e1 · e1 = 1.3) of x by ej and use (2. Show that the vectors ei are linearly independent. We see that e1 is orthogonal to both e2 and e3 . Show that ei is determined uniquely by the requirement that xi = x · ei for every x. Consider. If we dotmultiply the representation (2. the ei constitute a frame or basis. Exercise 2. We can therefore expand an arbitrary vector x as 3 x= i=1 xi ei . By Exercise 2. Similar properties hold for e2 and e3 . This basis is said to be reciprocal or dual to the basis ei . j = i. Now let us discuss the geometrical nature of the vectors ei .and righthand sides.3) Note that superscripts and subscripts continue to appear in our notation.1).1. e2 · e1 = 0. is the Kronecker delta symbol.2.2) where i δj = 1. j = i.2. e3 · e1 = 0.
Let us discuss the reciprocal basis in more detail. This rule shall apply to situations involving vectors as well: we shall understand. xi ei to mean the summation i xi ei . the calculation and use of the inverse matrix is reasonable. we shall understand that a summation is to be carried out over i. whenever we see i as a subscript and a superscript. This is reminiscent of the solution of a set of linear simultaneous equations: it is ineﬃcient to ﬁnd the inverse matrix of the system if we have only one forcing vector.1 Note that a repeated index is a dummy index in the sense that it may be replaced by any other index not already in use: we have xi ai = x1 a1 + x2 a2 + x3 a3 = xk ak for instance. we used a combination of indices (i. The ﬁrst problem is to ﬁnd suitable formulas to deﬁne it. say. However. So the introduction of a reciprocal basis gives many potential advantages. for example the index i in Ak xk . for example. we shall continue to make the distinction so that we can deal with nonCartesian frames.. An index that occurs just once in an expression. We derive these formulas next. This rule is called the rule of summation over repeated indices. But when we must solve such a problem repeatedly for many forcing vectors. From now on we shall omit the symbol of summation when we meet matching subscripts and superscripts: we shall write. In tensor discussions each free index is understood to range independently over a set of values — presently this set is {1. In a paper where he introduced this rule. The real advantages come when we are working intensively with many vectors. 1 The rule of summation was ﬁrst introduced not by mathematicians but by Einstein.e. i is called a free index. xi ai for the sum i xi ai . Writing out x in the ei and ei bases. Einstein used Cartesian frames and therefore did not distinguish superscripts from subscripts. 3}. That is. but ﬁrst let us note the following.14 Tensor Analysis with Applications in Mechanics look quite similar. and is sometimes referred to as the Einstein summation convention. 2. . The use of reciprocal vectors may not be practical in those situations where we are working with only two or three vectors. subscripts and superscripts) and summation symbols.
Let us now consider the forms of the dot product between two vectors a = ai e i = a j e j . Denoting it by V .Transformations and Vectors 15 Let us return to the task of deriving formulas for the reciprocal basis vectors ei in terms of the original basis vectors ei . We obtain c1 [e1 · (e2 × e3 )] = 1. Since the cross product of two vectors is perpendicular to both. The quantity e1 · (e2 × e3 ) is a scalar whose absolute value is the volume of the parallelepiped described by the vectors ei . we can satisfy the conditions e2 · e1 = 0. Exercise 2. 1 2 (e × e3 ). e2 = 1 (e3 × e1 ). Within an algebraic sign. The reader may verify that these expressions satisfy (2. V is the volume of the parallelepiped described by the vectors ei . V e2 = 1 3 (e × e1 ). Hence we immediately get the dual formulas e1 = where V = e1 · (e2 × e3 ). V . b = b p ep = b q eq . V e3 = 1 (e1 × e2 ). by setting e1 = c1 (e2 × e3 ) where c1 is a constant. V e3 = 1 1 (e × e2 ). V e3 · e1 = 0. Let us mention that if we construct the reciprocal basis to the basis ei we obtain the initial basis ei . To determine c1 we require e1 · e1 = 1.3. Show that V = 1/V . V 1 (e2 × e3 ). we have e1 = Similarly.2). We construct e1 ﬁrst.
3. otherwise we get more terms and additional multipliers. similarly. Introducing the notation gip = ei · ep . Finally. (As a short exercise the reader should write out this expression in full.5) and. a · b = aj b j . They are the components of a unique tensor known as the metric tensor. .) Using the reciprocal component representations we get a · b = aj ej · bq eq = aj bq g jq where g jq = ej · eq . In Cartesian frames we obviously have j gij = δi . We will encounter gij and g ij often. Hence a · b = ai bj gij = ai bj g ij = ai bi = ai bi . i g ij = δj .16 Tensor Analysis with Applications in Mechanics We have a · b = ai e i · b p e p = a i b p e i · e p . using a mixed representation we get q a · b = ai e i · b q e q = ai b q δ i = ai b i (2.4) (2. We see that when we use mixed bases to represent a and b we get formulas that resemble the equation a · b = a 1 b 1 + a2 b 2 + a3 b 3 from § 1. we have a · b = ai bp gip .
Transformations and Vectors 17 2. (a) Show that if the Gram determinant vanishes. (2. this means that the matrices of gij and g ij are mutually inverse. In ei the metric components are given by g ij . Of course.6) and (2. In the system of equations ⎞ ⎛ ⎞⎛ 1 ⎞ ⎛ x g11 g21 g31 x1 ⎝ x2 ⎠ = ⎝ g12 g22 g32 ⎠ ⎝ x2 ⎠ x3 g13 g23 g33 x3 the matrix of the components of the metric tensor gij is also called the Gram matrix. using the components of the metric tensor. we can always change subscripts to superscripts and vice versa. A theorem in linear algebra states that its determinant is not zero if and only if the vectors ei are linearly independent.6): xi = xj g ij .6) and (2. Finally.7) (2.3 Transformation to the Reciprocal Frame How do the components of a vector x transform when we change to the reciprocal frame? We simply set xi ei = xi ei and dot both sides with ej to get xi ei · ej = xi ei · ej or xj = xi gij . Exercise 2.6) We see from (2. (b) Prove that the Gram determinant equals V 2 . (2. These actions are known as the raising and lowering of indices.4. so we can immediately write an expression dual to (2. then the ei are linearly dependent.7) together imply xi = gij g jk xk .7) that. We called the basis ei dual to the basis ei . . hence k gij g jk = δi .
of course. (b) Write out all forms of the dot product x · y. The dual of the basis ek (i. and where V = e1 · (e2 × e3 ). 2).4 Transformation Between General Frames Having transformed the components xi of a vector x to the corresponding components xi relative to the reciprocal basis. the vectors ei given by the requirement that i e j · e i = δj are linearly independent and form a basis called the reciprocal or dual basis. 1. we are now ready to take on the more general task of transforming the xi to the corresponding compo˜ ˜ nents xi relative to any other basis ei . the dual of the dual) is the original basis ek .18 Tensor Analysis with Applications in Mechanics Quick summary Given a basis ei .e. compact notation for the system of equations ⎞ ⎛ 1 2 3 ⎞⎛ ⎞ ⎛ ˜ A1 A1 A1 e1 e1 ⎝ e2 ⎠ = ⎝ A1 A2 A3 ⎠ ⎝ e2 ⎠ . k) equals (1. j. The deﬁnition of dual basis is motivated by the equation xi = x · ei . say (2.8) . 2. Exercise 2. ˜ 2 2 2 ˜ e3 e3 A1 A2 A3 3 3 3 ≡A.5. 3) or one of the cyclic permutations (2. Write out the modulus of x in all possible forms using the metric tensor. A given vector x can be expressed as x = xi ei = xi ei where the xi are the components of x with respect to the dual basis. The ei can be written as 1 ei = (ej × ek ) V where the ordered triple (i. 1) or (3. 3. i˜ This is.. (a) Let x = xk ek = xk ek . Let the new basis ei be related to ˜ the original basis ei by ei = Aj ej . 2.
x = xi ei . ˜˜ hence xi ei = xk Aj ej .Transformations and Vectors 19 Before proceeding. we note that in the symbol Aj the subscript indexes i the row number in the matrix A. we may expand the notation and write (2. ˜ j (2.9) the index j is a dummy index which we can replace with . 3˜ 3˜ 3˜ 3˜ ˜ Matching coeﬃcients of the ei we ﬁnd x1 = x1 A1 + x2 A1 + x3 A1 = xj A1 . Aj ej = A1 e1 + A2 e2 + A3 e3 .9) as ˜ ˜ ˜ ˜ ˜ x1 e1 + x2 e2 + x3 e3 = x1 Aj ej + x2 Aj ej + x3 Aj ej ˜ ˜ 1˜ 2˜ 3˜ where. i Exercise 2. It follows from (2.9) in a succinct manner. Find the matrix of transformation.8) that ˜ Aj = ei · ej . Throughout our development we shall often take the time to write various equations of interest in matrix notation. of course. 2˜ 2˜ 2˜ 2˜ Aj ej = A1 e1 + A2 e2 + A3 e3 . while the superscript indexes the column number. ˜ 1 2 3 j x3 = x1 A3 + x2 A3 + x3 A3 = xj A3 .6.9) To ﬁnd xi in terms of xi . A vector x can be expressed in the two forms x = xk ek . 1˜ 1˜ 1˜ 1˜ Aj ej = A1 e1 + A2 e2 + A3 e3 . we have xi ei = xk ek . ˜ 1 2 3 j x2 = x1 A2 + x2 A2 + x3 A2 = xj A2 .10) It is possible to obtain (2. ˜ 1 2 3 j hence xi = xj Ai .10) from (2. ˜˜ Equating these two expressions for the same vector x. On the righthand side of (2. A Cartesian frame is rotated about its third axis to give a new Cartesian frame. ˜˜ k˜ (2.
˜ i xi = xk Ak .8) we have ˜ ˜ j˜ ˜ ei = Aj ej = Aj Ak ek . From these follow the transformation formulas ˜ xi = xk Ak . ˜ Putting x = ei in the ﬁrst of these gives ˜ ei = Ai ej .7. The product of the matrices (Aj ) and (Ak ) is the unit j i matrix and thus these matrices are mutually inverse.10) immediately. i i ˜ and since the ei form a basis we must have k ˜ Aj Ak = δi . Since the direct transformation is linear the inverse must be linear as well. j while the second identity with x = ei yields ˜j ˜ ei = Ai ej .20 Tensor Analysis with Applications in Mechanics i and thereby obtain (2. the transpose of A. We begin with the obvious identities ej (ej · x) = x. so we can write ˜ ˜ ei = Aj ej i where ˜ ˜ Aj = ei · ej . By (2. ˜ e ej (˜j · x) = x. The matrix notation equivalent of (2. Exercise 2. ˜ i . i j (2.11) and (2. Show that xi = xk Ai .10) is ⎛ 1 ⎞ ⎛ 1 1 1 ⎞⎛ 1 ⎞ A1 A2 A3 x x ˜ ⎝ x2 ⎠ = ⎝ A2 A2 A2 ⎠ ⎝ x2 ⎠ ˜ 1 2 3 x3 ˜ A3 A3 A3 x3 1 2 3 and thus involves multiplication by AT . ˜ ˜k Formulas for the relations between reciprocal bases can be obtained as follows.11) The relationship k ˜ Aj Ak = δi i j ˜ follows similarly. i ˜ Let us ﬁnd the relation between the matrices of transformation A and A. ˜ We shall also need the equations of transformation from the frame ei back to the frame ei .
˜ j apply. then x has the various expressions x = xi ei = xi ei = xi ei = xi ei ˜˜ ˜˜ and the transformation laws xi = Aj xj . g ij = ei · ej . the transformation law ˜j ˜ xi = Ai xj shows that the xi transform like the ei . where gij = ei · ej . j xi = Ai xj . xi = g ij xj . The transformation laws are particularly simple when the frame is changed to the dual frame. We shall ﬁnd a further use for this nomenclature in Chapter 3. i˜ then the components xi of a vector x must transform according to xi = Aj xj .Transformations and Vectors 21 2. are components of the metric tensor. Quick summary If frame transformations ei = Aj ej . i˜ ˜j xj . xi = A ˜ i ˜ ˜ xi = Ai xj .5 Covariant and Contravariant Components We have seen that if the basis vectors transform according to the relation ei = Aj ej . j are considered. i˜ The similarity in form between these two relations results in the xi being termed the covariant components of the vector x. . i ˜ ˜ ei = Ai ej . Then xi = gji xj . i˜ ˜ ˜ ei = Aj ej . For this reason the xi are termed the contravariant components of x. while the xi are termed covariant components. j ˜ ei = Ai ej . On the other hand. The xi are termed contravariant components of x.
In more advanced discussions in which threedimensional problems are considered. Because the magnitudes of a and b enter into a×b in multiplicative fashion. This quantity is introduced in elementary physics as the product of a force magnitude and a length (“force times moment arm”). The magnitude of t equals f r sin θ. then the thumb ˆ indicates the direction of n and hence the direction of t. c = a × b can be expanded in terms of a basis. If a and b are any two vectors.12) Here the ’s are formal coeﬃcients. We write a = aj ej . we choose the reciprocal basis ei and write c = ci e i . Of the two possible unit ˆ normals. The deﬁnition of torque can be taken as a model for a more general operation between vectors: the cross product. and employ the wellknown distributive property (u + v) × w ≡ u × w + v × w .6 The Cross Product in Index Notation In mechanics a major role is played by the quantity called torque. If a force f acts at a point which is located relative to an origin O by position vector r. These rules are all encapsulated in the brief symbolism t = r × f. Like any other vector. we deﬁne ˆ a × b = nab sin θ ˆ where n and θ are deﬁned as in the case of torque above. we are prompted to seek ci in the form ci = ijk a j k b . Let us ﬁnd them.22 Tensor Analysis with Applications in Mechanics 2. b = b k ek . torque is regarded as a vectorial quantity. t is conventionally (but arbitrarily) associated with the vector n given by the familiar righthand rule: if the foreﬁnger of the right hand is directed along r and the middle ﬁnger is directed along f . (2. then the associated torque t about O is normal to the plane of the vectors r and f . along with some rules for algebraic sign to account for the sense of rotation that the force would encourage when applied to a physical body. where θ is the smaller angle between f and r.
k) is an odd permutation of (1. We saw before that (ej × ek ) · ei equals the volume of the frame parallelepiped if i. (i. j. the determinant of a matrix A = (aij ) can be expressed succinctly as det A = ijk a1i a2j a3k .8. Exercise 2. ⎪ ⎨ = −V. 2. j. The contravariant components of a vector c = a × b can be expressed as ci = ijk aj b k . For example. 3). j. ijk represents a tensor (the socalled Levi–Civita tensor ). 2. Here it is convenient to introduce the idea of a permutation of the ordered triple (1. 3). Moreover. k) an even permutation of (1. k. A permutation of (1. k) is an even permutation of (1. The permutation symbol ijk is useful in writing formulas. j. a permutation is odd if it results from performing any odd number of interchanges. two or more indices equal. 2. Hence ⎧ ⎪+V. Now the value of (ej × ek ) · ei depends on the values of the indices i. Then c · ei = cm em · ei = ci = aj bk [(ej × ek ) · ei ] and comparison with (2. Note that V  = 1 for a Cartesian frame. If any two of the numbers i. j. k) an odd permutation of (1. it can be shown (Exercise 2.12) shows that ijk = (ej × ek ) · ei .Transformations and Vectors 23 to obtain c = aj ej × bk ek = aj bk (ej × ek ). (i. k are distinct and the ordered triple (i. k are distinct and the ordered triple (i. 2. 2. We discuss this further in Chapter 3. we obtain minus the volume of the frame parallelepiped. 2. k are equal we obtain zero. If i. 3) is called even if it can be brought about by performing any even number of interchanges of pairs of these numbers. ijk ⎪ ⎪ ⎩0. 3). Much more than a notational device however. j. 3). 3).4) that V2 =g where g is the determinant of the matrix formed from the elements gij = ei · ej of the metric tensor. j. j.
the components of a vector depend on the lengths of the frame vectors and the angles between them. 2.24 Tensor Analysis with Applications in Mechanics for suitable coeﬃcients ijk . Thus for two vectors x = xi ei and y = y j ej written in the arbitrary frame.7 Norms on the Space of Vectors (2. Use this in turn to prove that a × (b × c) = b(a · c) − c(a · b) for any vectors a.e. b. But the formulas connected with the dot product are invariant under change of frame. a simple dot product) x · y = xi ei · y j ej = xi y j gij = xi yj g ij = xi yi . so we can use them to characterize the intensity of the vector — its length.9.13) We often need to characterize the intensity of some vector ﬁeld locally or globally. Exercise 2. For this.. . the notion of a norm is appropriate. In a nonCartesian frame. This norm is related to the inner product of two vectors a = ak ik and b = bk ik : we have a · b = ak bk so that a = (a · a)1/2 . The wellknown Euclidean norm of a vector a = ak ik written in a Cartesian frame is 3 1/2 a = k=1 a2 k . we cannot use it to characterize the vector. Establish Lagrange’s identity (a × b) · (c × d) = (a · c)(b · d) − (a · d)(b · c). Use the technique of this section to ﬁnd the coeﬃcients. Since the sum of squared components of a vector depends on the frame. Then establish the identity pqr p q r δi δi δi p q r = δj δj δj p q r δk δk δk ijk and use it to show that ijk pqk p q q p = δi δj − δi δj . c. we can introduce a scalar product (i.
with x = 0 if and only if x = 0. we conﬁrm satisfaction of the property (ii) x · y = y · x. One norm can be introduced as follows. By the general theory then. Similarly. This dot product and associated norm have all the properties required from objects of this nature in algebra or functional analysis. it is necessary only to check whether all the axioms of the inner product are satisﬁed. The reader should check that this holds for any representation of the vectors. Indeed. the principal space is R3 . We can introduce other norms. (ii) αx = α x for any real α. (2. By the same reasoning. For example. Finally. (iii) x + y ≤ x + y . (iii) (αx + βy) · z = α(x · z) + β(y · z) where α and β are arbitrary real numbers and z is a vector. the norm of a vector x is x = (x · x)1/2 = xi xj gij 1/2 = xi xj g ij 1/2 = xi xi 1/2 . the expression x = (x · x)1/2 satisﬁes all the axioms of a norm: (i) x ≥ 0. In this book. A linear space equipped with the norm (2. and in practice a variety of norms turn out to be necessary.14) where in the case of nonzero vectors the equality holds if and only if x = λy for some real λ.14) becomes a normed space.Transformations and Vectors 25 Note that only in mixed coordinates does this resemble the scalar product in a Cartesian frame. and x · x = 0 if and only if x = 0. The set of all threedimensional vectors constitutes a threedimensional linear space. Let ek be a basis of R3 and let . In addition we have the Schwarz inequality x·y ≤ x y .15) (2. (i) x · x ≥ 0. 2 x is also a norm in R3 . This property holds because all the quantities involved can be written in a Cartesian frame where it holds trivially. quite diﬀerent from the above. Note that we can introduce more than one norm in any normed space.
Norm axioms (i) and (ii) obviously hold. we will need a space of vectors with complex components. For p ≥ 1. y = x · y. In C3 . where i is the imaginary unit (i2 = −1). and introduce multiplication of vectors in R3 by complex numbers. but it is complex and denoted by C3 . we introduce 3 1/p x p = k=1 xk p . Later. This also yields a linear space. The reader should be aware that this norm is given in a certain basis. It can be introduced similarly to the space of complex numbers. the reader should write out the form of the real and imaginary parts of x in such a basis.26 Tensor Analysis with Applications in Mechanics x = xk ek .14) is that it is independent of the basis of the space. Axiom (iii) is a consequence of the classical Minkowski inequality for ﬁnite sums. when investigating the eigenvalues of a tensor. An arbitrary vector x in C3 takes the form x = (ak + ibk )ek . a basis in C3 may contain vectors that are not in R3 . However. and x · x = 0 if and only if x = 0. (iii) (αx + βy) · z = α(x · z) + β(y · z) where α and β are arbitrary complex numbers. we can introduce the inner product of two vectors x and y as x. (ii) x · y = y · x. As an exercise. Let x. We start with the space R3 having basis ek . x 1/2 . If we use it in another basis. respectively. y. deﬁned by x = (ak − ibk )ek . z be arbitrary vectors of C3 . It is easy to see that this inner product has the following properties. Clearly. Then (i) x · x ≥ 0. . The real and imaginary parts of x are ak ek and bk ek . Now we can introduce the norm related to the inner product. An advantage of the norm (2. The reader should verify these properties. the dot product loses the property that x · x ≥ 0. Analogous to the conjugate number is the conjugate vector to x. x = x. the value of the norm of a vector will change in general.
(b) e1 = i1 + 3i2 + 2i3 . A tensor is an objective entity. Schwarz’s inequality (2.1 Find the dual basis to ei . As a consequence of the general properties of the inner product. 2. ˜ e2 = −2i1 + 2i2 + i3 . 2. e3 = i1 + i3 . e3 = i 1 + i 3 . i . But if we ﬁx those components then in any other frame the vector is determined uniquely.2 Let ˜ e1 = −2i1 + 3i2 + 2i3 .15) also holds in C3 .8 Closing Remarks We close by repeating something we said in Chapter 1: A vector is an objective entity. (d) e1 = cos φi1 + sin φi2 .9 Problems 2. We emphasize this because the situation is exactly the same with tensors. ˜ e3 = −i1 + i2 + i3 . ˜ Find the matrix Aj of transformation from the basis ei to the basis ej . (a) e1 = 2i1 + i2 − i3 . e3 = 3i3 . In elementary mathematics we learn to think of a vector as an ordered triple of components. e1 = 2i1 + i2 − i3 .Transformations and Vectors 27 and verify that it satisﬁes all the axioms of a norm in a complex linear space. Absolutely uniquely! So a vector is something objective. e2 = 2i2 + 3i3 . e3 = i3 . no harm in this if we keep in mind a certain Cartesian frame. e2 = i1 − i2 . but as soon as we specify its components in one frame we can ﬁnd them in any other frame by the use of certain rules. 2. e2 = 2i1 − 3i2 + 2i3 . e3 = 3i1 + 2i2 + 3i3 . we determine the tensor uniquely — even though its components relative to other frames will in general be diﬀerent. There is. (c) e1 = i1 + i2 . and ﬁxing its components relative to one frame. e2 = 2i2 + 3i3 . e2 = − sin φi1 + cos φi2 . of course.
6 Show that 2. and suppose h1 . ˜ e2 = −i2 − i3 . j k i (f) δi δj δk . 2.28 Tensor Analysis with Applications in Mechanics 2. 2.11 Show that (e × a) × e = a if e = 1 and e · a = 0. i j ijk a a . mkj i δm . (d) δij δ jk . ijk k j ijk δm δn . ijk . i (b) ai aj δj . let x = xk ek .9 Show that (a × b) · a = 0. e1 = i1 − 6i3 . h2 . 2.4 Find (a) aj δ jk . ijk 2. ijk  imn j δm .12 Let ek be a basis of R3 . e3 = i 1 + i 2 + i 3 . jk ijk δ . 2. ˜ Find the matrix of transformation of the basis ei to ej . 2.3 Let ˜ e1 = i1 + 2i2 . .10 Show that a · (b × c)d = (a · d)b × c + (b · d)c × a + (c · d)a × b. ijl ijk l = 2δk .7 Find (a) (b) (c) (d) (e) (f) ijk ijk = 6.5 Show that 2. e2 = −3i1 − 4i2 + 4i3 .8 Find (a × b) × c. ˜ e3 = −i1 + 2i2 − 2i3 . Show that hk xk  is a norm in R3 . 2. (e) δij δ ji . h3 are ﬁxed positive numbers. i (c) δi .
The components of the metric tensor transform according to certain rules.1 Dyadic Quantities and Tensors We have met sets of quantities like g ij or gij . We indicated in Chapter 1 that these quantities. Such a table of 3 × 3 = 9 coeﬃcients could be considered as a vector in a ninedimensional space. however. we shall write out the dyad without the ⊗ symbol: ab = a ⊗ b.1). We assume that the tensor product satisﬁes many usual properties of a product: (λa) ⊗ b = a ⊗ (λb) = λ(a ⊗ b). 29 . From now on. represented by 3×3 matrices. form a class of objects known as secondorder tensors.1) where λ is an arbitrary real number. However.Chapter 3 Tensors 3. but we must reject this idea for an important reason: if we change the frame vectors and calculate the relations between the new and old components. Our plan is to present the relevant theory in a way that parallels the vector presentation of Chapter 2. and it is found that these transformation rules also apply to various quantities encountered in physical science. the results diﬀer in form from those that apply to vector components. Let us once again consider the space of threedimensional vectors with the frame ei . We begin to realize this program with the introduction of the dyad (or tensor product ) of two vectors a and b. (3. (a + b) ⊗ c = a ⊗ c + b ⊗ c. a ⊗ (b + c) = a ⊗ b + a ⊗ c. we represent the dyad ab as ab = ai ei bj ej = ai bj ei ej . the tensor product is not symmetric: if a is not proportional to b then a ⊗ b = b ⊗ a. denoted a ⊗ b. Using the expansion of the vectors in the basis vectors and the properties (3.
so we deﬁne A uniquely if we specify its action on all three vectors of a basis. Thus if we have A · x = B · x for all x. has the representation A = aij ei ej . we require that upon transformation of the frame the components of A transform correspondingly. Analogously we can introduce the dot product from the left: c · ab = (c · a)b. Conversely. (3.2 Tensors From an Operator Viewpoint An alternative to viewing a secondorder tensor as a weighted sum of dyads is to view the tensor as an operator. (b) Write down matrices for the dyads i1 i1 . To maintain the property of objectivity of the elements of this space. the corresponding y is A · x = aij ei ej · ek = aik ei . The numerical coeﬃcients of the dyads are called the components of the tensor. . and i3 i1 . (3. (a) A dyad of the form ee. Note that we have introduced superscript indices for the components of A. 3.30 Tensor Analysis with Applications in Mechanics This introduces exactly nine diﬀerent dyads ei ej . i2 i2 . is sometimes called a projection dyad. It is clear that the operation A · x is linear in x. a given linear relation between x and y will deﬁne the operator A uniquely. This was done in keeping with the development of Chapter 2. Let us show that the components are really uniquely deﬁned in any basis by the equality y = A · x. Thus an element of this space. a tensor A. where e is a unit vector.1. We now consider a linear space whose basis is this set of nine dyads and call it the space of secondorder tensors (or tensors of order two). Taking x = ek . In preparation for the next section let us introduce the dot product of a dyad ab by a vector c: ab · c = (b · c)a. then we have A = B. From this standpoint a tensor A is considered to map a vector x into a vector y according to the equation y = A · x. Explain. The tensor A is represented by the expression aij ei ej in some basis ei .3) Exercise 3.2) So the result is a vector cooriented with a.
ai = e i · A · e j . In particular.4) . A simple but important tensor is the unit tensor denoted by E and deﬁned by the requirement that for any x E · x = x · E = x. Analyzing the above reasoning. called premultiplication of A by a vector y. Given tensors A and B.. We may also deﬁne common algebraic operations from the operator viewpoint. (3. the sum is the tensor A + B uniquely deﬁned by the requirement that (A + B) · x = A · x + B · x for all x. If c is a scalar. the components of A). The dot product y · A. Note that in “mixed components” we position the indices in such a way that their association with the various dyads remains clear. cA is deﬁned by the requirement that (cA) · x = c(A · x) for all x. If y is a given vector and there is a linear transformation from x to y for an arbitrary vector x. The dot product A · B is regarded as the composition of the operators B and A: (A · B) · x ≡ A · (B · x). is deﬁned by the requirement that (y · A) · x = y · (A · x) for all vectors x.Tensors 31 Dot multiplying this by el we get alk = el · A · ek . then the linear transformation is a tensor and we can write y = A · x. any product of the form 0A gives a zero tensor denoted 0. we can ﬁnd that we have proved the quotient law for tensors of order two.e. This statement is sometimes useful in establishing the tensorial character of a set of scalar quantities (i. aij = ei · A · ej . ·j ·j ai = e i · A · e j . In this way we can ﬁnd the components of a tensor A in any basis: aij = ei · A · ej .
5) and (3. (3. eij = gij . The inverse of a tensor is also a tensor.7). We see that the role of the unit tensor belongs to the metric tensor! Throughout our discussion of secondorder tensors we shall emphasize the close analogy between tensor theory and matrix theory. the raising and lowering of indices gives E = gij ei ej = g ij ei ej (3.32 Tensor Analysis with Applications in Mechanics It is evident that in any Cartesian frame ii we must have 3 E= i=1 ii ii . An important special case occurs when the matrix of the tensor is a diagonal matrix. Exercise 3. Consequently. Equations (3. 0 0 1 This does not hold for the nonmixed components of (3.5) In any frame we have E = ei ei = ej ej (3.5) to show that the various components of E are given by eij = g ij .4) and (2. If in a Cartesian frame ii we have 3 i e i = δj . ·j ·i i e j = δj . Hence establish (3.6) show that the matrix representation of E in either Cartesian or mixed components is the 3 × 3 identity matrix ⎛ ⎞ 1 0 0 ⎝ 0 1 0 ⎠. then A−1 is called the inverse of A. Our consideration of A as an operator leads us to introduce the notion of an inverse tensor : if A · A−1 = E. A= i=1 λi ii ii .2.4) along with (2. Use (3.6).6) for the mixed components.7) in nonmixed components.
⎛ ⎞ ⎛ −1 ⎞ ⎛ ⎞ 0 0 λ1 0 0 λ1 1 0 0 ⎝ 0 λ2 0 ⎠ ⎝ 0 λ−1 0 ⎠ = ⎝ 0 1 0 ⎠ . we get 3 3 A·B= i=1 3 λi ii ii · j=1 3 λ−1 ij ij j = i=1 3 λi ii j=1 λ−1 (ii · ij )ij j = i=1 3 λi ii λ−1 ii i ii ii = i=1 = E. B.Tensors 33 then the corresponding matrix representation is ⎛ ⎞ λ1 0 0 ⎝ 0 λ2 0 ⎠ . This means that B = A−1 . 0 0 λ3 If we take 3 B= j=1 λ−1 ij ij j to which there corresponds the matrix ⎛ −1 ⎞ 0 0 λ1 ⎝ 0 λ−1 0 ⎠ 2 0 0 λ−1 3 and form the dot product A · B. 2 0 0 λ3 0 0 λ−1 0 0 1 3 Exercise 3. Correspondingly. Establish the formula (A · B)−1 = B−1 · A−1 for invertible tensors A.3. .
We close this section with an important remark. g The ﬁrst equality is the deﬁnition as stated above. Various other formulas such as 1 · · · det A = ijk mnp ami anj apk 6 can be established for the determinant. tensors are objective entities and ultimately all results pertaining to them must be frame independent. in any coordinate frame. As we have said before. We set aij ei ej = aij ei ej . Convenience will often dictate the use of Cartesian frames. 3. The connection between the uses of this terminology in the two areas is as follows.. these equations have a nontrivial solution if and only if the determinant of the coeﬃcient matrix (i.34 Tensor Analysis with Applications in Mechanics A secondorder tensor A is singular if A · x = 0 for some x = 0. the equation A · x = 0 yields a set of simultaneous equations in the components of x. then we have provided rigorous justiﬁcation of the latter. and only (3. Recall that a matrix A is said to be nonsingular if and only if det A = 0. Let us begin with the transformation of the components when we go to the reciprocal basis. Hence A is nonsingular if A · x = 0 only when x = 0. If we take a mixed representation of the tensor A. we again arrive the same conclusion regarding the determinant. denoted det A.3 Dyadic Components Under Transformation The standpoint for deriving the transformation rules is that in any basis a tensor is the same element of some space. the rest are left for the reader to establish. By deﬁnition the determinant of a secondorder tensor A. We can derive all desired properties of a tensor. the matrix representing A) is zero. is the determinant of the matrix of its mixed components: ·j det A = ai  = akm  = · 1 ast  = gapq . taking any other representation of A and a dual representation of the vector. and perform actions with the tensor.1) and the rules we derived for vectors can govern the rules for transforming the components of a tensor. But if we obtain an equation or expression through the use of a Cartesian frame and can subsequently represent this result in noncoordinate form. Moreover. This brings the use of the term “singular” to the tensor A.e.
Tensors 35 and take dot products as in (3. the inverse transformation ˜ ˜ ei = Aj ej i leads to aij = akm Ai Aj . ˜ ˜˜ ˜ ˜ ˜ ˜i ˜ ˜ ˜· ˜ ˜ lead to other transformation formulas such as ˜ ˜ aij = Ak Al akl . We see that the components of the metric tensor are encountered in this transformation.5.8) and (3. The components ai and ai are called mixed components. Now we can construct the formulas for transforming the tensor components when the change of basis takes the general form ei = Aj ej . j k Various expressions for A. This gives aij (ek · ei )(ej · em ) = aij (ek · ei )(ej · em ). i˜ From aij ei ej = akm ek em = akm Ap ep Aq eq ˜ ˜˜ m˜ k˜ we obtain aij = akm Ai Aj . ˜ aij = Ak Al akl . ˜· k j · Remembering the terminology of § 2. ˜i i l ˜ ai j = Ai Al akl . we see why the aij are called the covariant components of A while the aij are called the contravariant com·j ponents.9) = aij ei ej = akl ek el = a· j ei ej = akl ek el .3): ek · aij ei ej · em = ek · aij ei ej · em . i j i j˜ and · ˜ a· j = Ak Aj akl .2) and (3. hence akm = aij gki gjm . ˜ k m Similarly.9) together imply that i ˜ Ak Ai = δj . ·j .8) (3. ·j A = aij ei ej = akl ek el = ai ei ej = akl ek el · (3. ˜ ˜k ˜m Equations (3.
Extensions of this and the formulas (3. ˜i l ˜k ˜ ˜ aij = Ai Aj akl . l aij = Ak Al akl .4. ˜ ˜ ai j = Ai Al akl . ˜· k j · · ˜i a· j = Ak Aj akl . . (b) Repeat for a transformation law of the form ˜ij = Ak Am bkm .36 Tensor Analysis with Applications in Mechanics Quick summary We have ˜ ˜ ˜ ˜i ˜ ˜ ˜· ˜ ˜ A = aij ei ej = akl ek el = a· j ei ej = akl ek el ˜ ˜˜ ·j = aij ei ej = akl ek el = ai ei ej = akl ek el · where aij = Ai Aj akl . they deﬁne a transformation of the space that is an operator. (a) Express the transformation law ˜ij = Ai Aj bkm b k m in matrix notation. (λA + µB) · a = λA · a + µB · a. i l ˜k Exercise 3. 3. i j˜ ˜ ai = Ai Al akl . and λ and µ be any real numbers. The result is again a dyad. Let A and B be dyads. Repeat for the inverse transformation.2) and (3.1) and the vectorial rules) gives us a number of rules which the dot product obeys. ˜ k l ˜i ˜j aij = Ak Al akl . and hence a secondorder tensor. ˜i ˜j b Exercise 3. a and b be vectors.5. ·j k j ˜· ·j ˜ ai = Ak Aj a· l .4 More Dyadic Operations The dot product of two dyads ab and cd is deﬁned by ab · cd = (b · c)ad.3) to operations with sums of dyads and vectors (using (3. Then A · (λa + µb) = λA · a + µA · b. with a coeﬃcient b · c. Our Aj values give the transformation from one basis to i another. Write out the tensor for which the Aj are i components.
These results show that linearity may be assumed in working with these operations. c2 = a21 b1 + a22 b2 + a23 b3 . this is the system of three equations c1 = a11 b1 + a12 b2 + a13 b3 . Denoting c= A·b we have ck ik = akm ik im · bj ij so that 3 b = bj ij .11) where c and b are column vectors and A is a 3 × 3 matrix. (3. or ⎞ ⎛ 11 12 13 ⎞ ⎛ 1 ⎞ b a a a c1 ⎝ c2 ⎠ = ⎝ a21 a22 a23 ⎠ ⎝ b2 ⎠ . c3 a31 a32 a33 b3 ⎛ Here we have a matrix equation of the form c = Ab (3.11).Tensors 37 Similar identities hold for dot products taken in the opposite orders. .10) and (3. Now let us pursue the close analogy between the dot product and matrix multiplication.10) ck = j=1 akj bj . c3 = a31 b1 + a32 b2 + a33 b3 . We take a dyad A and a vector b and express these relative to a basis ik : A = akm ik im . We begin with the simple case of a Cartesian frame. (We have inserted the summation symbol because j stands in the upper position twice and the summation convention would not apply. b = b j ej . for example. The analogy between the dot product and matrix multiplication is evident from (3. A = akm ek em . Let us write.) Written out. This analogy extends beyond the conﬁnes of Cartesian frames.
of course. There are various possibilities for the components of the dyad C = A · B. B = bkm ek em . a33 g 31 g 32 g 33 b3 b = b j ej . The corresponding matrix equation ⎛ 1 ⎞ ⎛ 11 c a ⎝ c2 ⎠ = ⎝ a21 c3 a31 is. 32 33 a a b3 With suitable understanding we could still write this as (3. ⎛ ⎞ ⎛ c1 a11 a12 ⎝ c2 ⎠ = ⎝ a21 a22 c3 a31 a32 The corresponding matrix form is ⎞ ⎛ 11 12 13 ⎞ ⎛ ⎞ a13 g g g b1 a23 ⎠ ⎝ g 21 g 22 g 23 ⎠ ⎝ b2 ⎠ . c3 a31 a32 a33 b3 Let us examine the dot product between two dyads.38 Tensor Analysis with Applications in Mechanics This time c = A · b gives j ck ek = akm ek em · bj ej = akm ek δm bj . hence ck = akj bj . Note what happens when we express both the dyad and the vector in terms of covariant components: A = akm ek em . . Because the metric tensor can raise an index on a vector component. We obtain ck = akm g mj bj and the metric tensor appears. depending on how we choose to express A and B. If we use all contravariant components and write A = akm ek em .11). we may also write these equations in the forms ck = akm bm and ⎛ ⎞ ⎛ ⎞⎛ 1 ⎞ c1 b a11 a12 a13 ⎝ c2 ⎠ = ⎝ a21 a22 a23 ⎠ ⎝ b2 ⎠ . ⎞⎛ ⎞ a12 a13 b1 22 23 ⎠ ⎝ a a b2 ⎠ .
then the remaining vectors. That is. If ab and cd are dyads. B = bkm ek em . Deﬁning ckn = akj bjn · we have C = ck n e k e n . (c) What happens if mixed components are used on the righthand sides to express A and B? Another useful operation that can be performed between tensors is double dot multiplication. and thereby obtain a scalar as the result. (b) Show how all the forms above correspond to matrix multiplication. . leads to C = ckn ek en where ckn = akm g mj bjn . the use of all covariant components as in A = akm ek em . Mixed components appear when we express A = akm ek em . we ﬁrst dot multiply the near standing vectors.Tensors 39 then C = ckn ek en where ckn = akm gmj bjn . Then j C = A · B = akm ek em · bjn ej en = akm δm bjn ek en = akj bjn ek en . (a) Discuss how the formulation ckn = akj bjn arises. B = bkm ek em . · We leave other possibilities to the reader as · Exercise 3. Similarly.6. we deﬁne ab ·· cd = (b · c)(a · d).
First we recall that such a tensor is represented in dyadic form as A = aij ei ej . For a matrix A = (aij ) the transposed matrix AT is AT = (aji ). then it does not matter whether we show the dyads.12) a31 a32 a33 When we use the matrix form (3. say Cartesian. 3.40 Tensor Analysis with Applications in Mechanics Exercise 3. Similarly we introduce the transpose operation for the tensor A: AT = aji ei ej . Yet another operation is the scalar product of two secondorder tensors A and B. Those who work in the applied sciences are probably more accustomed to the matrix representation ⎛ 11 12 13 ⎞ a a a ⎝ a21 a22 a23 ⎠ . The correspondence between the dyadic and matrix representations suggests that we can introduce many familiar ideas from the theory of matrices. Of course when we use a unique. frame for the space of vectors.5 Properties of SecondOrder Tensors Now we would like to consider in more detail those tensors that occur most frequently in applications: tensors of order two.13) . denoted by A • B. (3.7.12) the dyadic basis of the tensor remains implicit. Write down several diﬀerent component forms for the quantity A ·· B. The tensor transpose Let us begin with the notion of transposition. (a) Calculate A · · E if A is a tensor of order two. (3. This represents a natural extension of the operation ab • cd = (a · c)(b · d) between two dyads ab and cd. How does this relate to the trace of the matrix that represents A in mixed components? (b) Let A and B be tensors of order two.
then the components of the tensor B = A−1 are given by the formulas ·j bi = (3. (A · B)−1 = B−1 · A−1 .10.8.14) 1 2 det A ikl jmn · · amk anl . (c) Show that if A is a nonsingular tensor of order two. Let A be a secondorder tensor. . Show that a · CT · b = b · C · a. Equation (3. Let A and B be tensors of order two. This follows when we write x = xk ek and use (3. (AT )−1 = (A−1 )T . Exercise 3. Also note that (AT )T = A for any secondorder tensor A.14) can be used to deﬁne the transpose.9. Exercise 3.13) to see that AT = aij ej ei . then (A · B)T = BT · AT . (d) Verify the following relations: det A−1 = (det A)−1 . (a) Show that if A and B are tensors of order two.Tensors 41 This operation yields a new tensor. Demonstrate that A ·· AT = 0 if and only if A = 0. ·i A useful relation for any secondorder tensor A and any vector x is A · x = x · AT . (A−1 )−1 = A. in each representation of which the corresponding indices appear in reverse order: ·i AT = aji ei ej = aji ei ej = aj ei ej = aj ei ej . (b) Let a and b be vectors and C be a tensor of order two. Exercise 3. Demonstrate that A • B = A ·· BT = AT ·· B. Find A·· AT .
A−4 = A−3 · A−1 . A4 = A · A3 . . A3 = A · A2 . This technique is used in the study of nonlinear elasticity. eA = E + Symmetric and antisymmetric tensors Among the class of all secondorder tensors. but with use of a norm of the tensor A (see § 3. 1! 2! 3! The issue of convergence of such series is approached in a manner similar to the absolute convergence of usual series. for example.42 Tensor Analysis with Applications in Mechanics Tensors raised to powers By analogy with matrix algebra we may raise a tensor to a positive integer power: A2 = A · A.15) if A is symmetric. A tensor A is said to be antisymmetric if A = −AT . Note that Ak still represents a linear operator. It follows from (3. The reader will recall that the unit tensor E satisﬁes a relation of the form (3.15).12). These include the strain and stress tensors of the theory of elasticity. These operations can be used to construct functions of tensors using Taylor expansions of elementary functions. We can introduce other functions similarly. For example. A−3 = A−2 · A−1 . x x2 x3 + + + ··· .16) (3. All these satisfy the relation A = AT . and so on. 1! 2! 3! By this we can introduce the exponential of the tensor A: ex = 1 + A3 A A2 + + + ··· . Note that eA represents a linear operator.14) that A·x =x·A and (A · x) · y = x · (A · y) (3. an important role is played by the symmetric tensors. Negative integer powers are deﬁned by raising A−1 to positive integer powers: A−2 = A−1 · A−1 . and so on. The tensor of inertia is symmetric. as is the metric tensor.
Any secondorder tensor can be decomposed as a sum of symmetric and antisymmetric tensors: A=B+C where B = BT and C = −CT .Tensors 43 Exercise 3. ω 2 = c13 . Give the matrix forms corresponding to the cases of symmetric and antisymmetric tensors. Given an antisymmetric tensor C = cij ii ij in a Cartesian frame.1.11.13. Take B= 1 A + AT . Their signiﬁcance is also shown by the following Theorem 3. Exercise 3. 2 C= 1 A − AT .12. Proof. we can construct a vector ω = ω k ik according to the formulas ω 1 = c32 . 1 1 (ab + ba) + (ab − ba) 2 2 . ω 3 = c21 . x · C = x × ω. Demonstrate that the quadratic form x · A · x does not change if the secondorder tensor A is replaced by its symmetric part. 2 and check all the statements. It is easy to verify directly that C · x = ω × x. Exercise 3. The dyad ab can be decomposed into symmetric and antisymmetric parts as ab = for example. How many components can be independently speciﬁed for a symmetric tensor? For an antisymmetric tensor? Both symmetric and antisymmetric tensors arise naturally in the physical sciences. Show that if A is symmetric and B is antisymmetric then A ·· B = 0.
The crossproducts of a tensor A = aij ei ej and a vector x are deﬁned by the formulas A × x = aij ei (ej × x). These formulas are written in noncoordinate form so they hold in any frame. Equation (3.44 Tensor Analysis with Applications in Mechanics where x is an arbitrary vector.17) (x = 0) holds. Writing this as i (ai − λδk )xk = 0.17) may also be written in the form (A − λE) · x = 0. You should ﬁnd that this diﬀers from the previous eigenvector.15.14. 3. we extend these notions to tensors. which is derived as follows. Exercise 3. so it makes sense to introduce left and right eigenvectors. The eigenvalues of a secondorder tensor A are found as solutions of the characteristic equation for A. Show that C = E × ω = ω × E. In components (3. which is called the conjugate vector . x) is called an eigenpair if the equality A · x = λx (3. The pair (λ. As the analogous question in matrix theory relates to eigenvalues and eigenvectors. The reader can derive the formulas for ω. for an arbitrary frame. In the case of a symmetric tensor they coincide. x × A = aij (x × ei )ej . Now try to position an eigenvector on the left: x · ab = λx. Exercise 3. ·k .6 Eigenvalues and Eigenvectors of a SecondOrder Symmetric Tensor We now consider the question of which basis yields a tensor of simplest form. Find the eigenpairs of the dyad ab. Hence x is an eigenvector of A if A operates on x to give a vector proportional to x.17) becomes aij ei ej · xk ek = λxi ei or i aij gjk xk ei = λxk δk ei .
In nonlinear elasticity. has matrix ⎛ ⎞ 1 0 1 ⎝ 2 −1 0 ⎠ . when referred to a certain Cartesian basis. After representing the tensor in diagonal form it will be easy to see that I1 (A) and I3 (A) are. Writing it in the form −λ3 + I1 (A)λ2 − I2 (A)λ + I3 (A) = 0 (3. I1 (A) = tr A. Exercise 3. I2 (A) = λ1 λ2 + λ1 λ3 + λ2 λ3 . The coeﬃcients I1 (A). a23 · 3 a·3 − λ a13 · This is the characteristic equation1 for A. the invariants and eigenvalues of several tensors play important roles in the formulation of various constitutive laws. . and are expressed in terms of the eigenvalues by the Vi`te formulas e I1 (A) = λ1 + λ2 + λ3 . second. the trace and determinant of the tensor A. 1 Note that it is expressed in terms of mixed components of the tensor. However. A tensor A. Lurie (2005).16. In fact. I2 (A). λ2 . λ3 . 2 I3 (A) = det A. See. and third invariants of A. and third principal invariants of A. and I3 (A) are called the ﬁrst. A nontrivial solution exists if and only if the determinant of the coeﬃcient matrix vanishes: a11 − λ · a21 · a31 · a22 · a12 · −λ a32 · = 0. Ogden (1997)]. 0 1 2 Find the ﬁrst. respectively. it characterizes the properties of the tensor and has invariant properties since the eigenvalues of a tensor do not depend on the coordinate frame in which they are obtained.18) we note that it is cubic in λ. second. I2 (A) = 1 2 [tr A − tr A2 ]. hence there are at most three distinct eigenvalues λ1 . [Lurie (1990).Tensors 45 we have a system of three simultaneous equations in the three variables xk . I3 (A) = λ1 λ2 λ3 . for example.
Indeed. suppose that A · x1 = λ1 x1 . Proof. x · x = x2 is positive. we have a formula analogous to (3. and then (3. To prove the second part of the theorem. Moreover. eigenvectors corresponding to distinct eigenvalues are orthogonal. (3.2. These have special properties.19) This will be used below. The eigenvalues of a real symmetric tensor are real. (3. hence they are real as well. . and subtraction gives (λ1 − λ2 )x1 · x2 = x2 · A · x1 − x1 · A · x2 = 0. the most important secondorder tensors are the realvalued symmetric tensors.19)). Let A be a real symmetric tensor and λ an eigenvalue of A that corresponds to the eigenvector x = 0. we write out (A · x) · x = x · (A · x) = (A · x) · x (we have used a property of the inner product in a complex linear space. so A · x = λx.46 Tensor Analysis with Applications in Mechanics In applications. We recall that a bar over an expression denotes complex conjugation. Theorem 3.20) λ2 x2 · x1 = (A · x2 ) · x1 . Because the eigenvalues λ are real. For a realvalued tensor that is considered as an operator in the complex space C3 . the components of the eigenvectors satisfy a linear system of simultaneous equations having real coeﬃcients.16): (A · x) · y = x · (A · y). To see that (A · x) · x takes a real value. It follows that λ= (A · x) · x . It follows that x1 · x2 = 0. x·x Now we prove that λ is real. A · x2 = λ2 x2 . Dotmultiply both sides of this equality by x. From these we obtain λ1 x1 · x2 = (A · x1 ) · x2 . where λ2 = λ1 .
we may ﬁnd that there are three distinct solutions or fewer than three. and the process of referring the tensor to its principal axes is known as diagonalization. Hence the coeﬃcients of the dyads xi xj in A are nonzero only for those coeﬃcients that lie on the main diagonal of the matrix representation of A. moreover. Show that for any secondorder tensor A (not necessarily symmetric). Exercise 3. and we may calculate the components of A from aij = xi · A · xj . So when the eigenvalues are distinct we can compose a Cartesian frame from the orthonormal eigenvectors xk and then express the tensor A in terms of its components aij as A= aij xi xj . . We can therefore write 3 A= i=1 λi xi xi . Since the frame xk is Cartesian the reciprocal basis is the same. (3. This holds for the eigenvectors corresponding to diﬀerent eigenvalues of a real symmetric tensor A. and dot multiplication by xi from the left gives j xi · A · xj = xi · λj xj = λj δi .21) This is called the orthogonal representation of A. x2 . Since the xi are also eigenvectors we can write A · xj = λj xj . The eigenvectors composing the coordinate frame give us the principal axes of A.17.20) we may obtain another property of the eigenvectors x1 . In other words. eigenvectors corresponding to distinct eigenvalues are linearly independent. In solving the characteristic equation for λ. known as generalized orthogonality: x1 · A · x2 = 0. these diagonal entries are the eigenvalues of A.Tensors 47 From (3. and is useful in applications. an eigenvector is determined up to a constant multiple. Note that if x is an eigenvector of A then so is αx for any α = 0.
then the subspace is twodimensional and we can select an orthonormal pair that is orthogonal to the third eigenvector (since λ1 and λ3 are distinct).48 Tensor Analysis with Applications in Mechanics When the characteristic equation of a tensor has fewer than three distinct solutions for λ. and choosing a Cartesian frame we would have an eigenvalue of multiplicity three. this case arises only when the tensor under consideration is proportional to the unit tensor E. If A is symmetric. If λ1 = λ2 = λ3 then any vector is an eigenvector. Show directly that a secondorder tensor (not necessarily symmetric) having three distinct eigenvalues cannot have more than one linearly independent eigenvector corresponding to each eigenvalue.22) by A and thereby represent A4 in terms . Furthermore. Similarly. For example the 2 × 2 matrix A= has characteristic equation a−λ b = λ2 − (a + d)λ + (ad − bc) = 0.7 The Cayley–Hamilton Theorem The Cayley–Hamilton theorem states that every square matrix satisﬁes its own characteristic equation.22) This permits us to represent A in terms of lower powers of A. c d−λ and the Cayley–Hamilton theorem tells us that A itself satisﬁes A2 − (a + d)A + (ad − bc)I = 0. then the repeated eigenvalue is said to be degenerate. Such a tensor is called a ball tensor.18. 3. Such an eigenvalue corresponding to two linearly independent eigenvectors is regarded as a multiple eigenvalue (multiplicity two). Corresponding to a multiple root of the characteristic equation we have a subspace of eigenvectors. we may dot multiply (3. 3 a b c d (3. However. If λ1 = λ2 is a double root.18) states that A satisﬁes the equation −A3 + I1 (A)A2 − I2 (A)A + I3 (A)E = 0. Exercise 3. we may represent A in a Cartesian basis and apply facts from the theory of symmetric matrices. where I is the 2 × 2 identity matrix and the zero on the right side denotes the 2×2 zero matrix. the Cayley–Hamilton theorem for a secondorder tensor A whose characteristic equation is given by (3.
nonlinear elasticity) where functions of tensors are represented approximately by truncated Taylor series. Exercise 3. and A2 . Use the Cayley–Hamilton theorem to express A3 in terms of A2 . .23): A3 = 3 i=1 3 [I1 (A)λ2 − I2 (A)λi + I3 (A)]ii ii i λ2 ii ii − I2 (A) i 3 3 = I1 (A) i=1 λi ii ii + I3 (A) i=1 i=1 ii ii = I1 (A)A2 − I2 (A)A + I3 (A)E.19.g. Such a tensor A has the representation (3. i (3. It is clear that we could continue in this fashion and eventually express any desired power of A in terms of E. 3.21). and E if A = i1 i1 + i2 i1 + i2 i2 + i3 i2 .8 Other Properties of SecondOrder Tensors Tensors of rotation A tensor Q of order two is said to be orthogonal if it satisﬁes the equality Q · QT = QT · Q = E. as desired. This is useful in certain applications (e. i λ3 ii ii . A..23) Let us return to equation (3. A. It is easy to establish the Cayley–Hamilton theorem for the case of a symmetric tensor. where we redenote ii = xi because the eigenvectors xi constitute an orthonormal basis: 3 A= i=1 λi ii ii .18) written for the ith eigenvalue and put it into the expression (3. We get A2 = A · A = Similarly A3 = A · A · A = 3 i=1 3 3 3 3 i λi ii λj ij δj = i.j=1 i=1 λi ii ii · i=1 j=1 λj ij ij = λ2 ii ii .Tensors 49 of lower powers of A.
We now consider the orthogonal tensor Q as an operator in the space of all vectors. (c) Show that if Q is orthogonal then so is Qn for every integer n. Thus. Thus 1 = det E = det Q det QT = (det Q)2 as desired. we call Q an improper orthogonal tensor if det Q = −1. Furthermore.50 Tensor Analysis with Applications in Mechanics We see that QT = Q−1 for an orthogonal tensor. Because of the properties of the determinant of a matrix and the correspondence between tensors and matrices. angles are also preserved. Exercise 3. First we put y = x to see that Q preserves vector magnitudes: Q · x2 = x2 . The operator deﬁned by the orthogonal tensor Q preserves the magnitudes of vectors and the angles between them. Proof. Indeed. (a) Show that the tensor Q = −i1 i1 + i2 i2 + i3 i3 is orthogonal. Is it proper or improper? (b) Show that k qij q kj = δi if Q is orthogonal.20. det Q = ±1. Consider the result of application of Q to both the multipliers of the inner product x · y: (Q · x) · (Q · y) = (x · QT ) · (Q · y) = x · QT · Q · y = x · E · y = x · y. We call Q a proper orthogonal tensor if det Q = +1.3. det Q is determined by the determinant of the matrix of mixed components of Q. Theorem 3. . by deﬁnition of the dot product in terms of the cosine. we have for two tensors A and B of order two det(A · B) = det A det B and det A = det AT .
The equations of physics should often be introduced in such a way that they are invariant under rotation of the coordinate frame whose position is not determined in space. an improper tensor also causes an axis reﬂection that changes the “handedness” of the frame. in particular. Let us note that some quantities are always invariant under rotation. We see that this is a representation of the “rotated” tensor. The trace of the rotated tensor is given by ·j E ·· (Q · A · QT ) = E ·· (Q · ai ei ej · QT ) ·j = E ·· ai (Q · ei )(Q · ej ). This means that the unit tensor (the metric tensor!) can be represented.) The situation is analogous to the case of a solid body where the position of a point is deﬁned by a vector beginning at a ﬁxed origin of some frame and ending at the point. as E = (Q · ei )(Q · ei ) = (Q · ej )(Q · ej ). Any motion of the solid with a ﬁxed point is a rotation with respect to some axis by some angle. Under the action of Q the frame ei transforms to the frame Q · ei .Tensors 51 The action of Q amounts to a rotation of all vectors of the space. this is the case for a proper orthogonal tensor. also called the trace of the tensor and denoted tr(A). to which the reciprocal basis is Q · ei . One of them is the ﬁrst invariant I1 (A) of a tensor. (More precisely. Such invariance under rotation should be veriﬁed. derived as the result of applying Q to each vector of the dyadic components of A. This quantity. To show invariance we consider the tensor Q · A · QT = Q · (aij ei ej ) · QT = aij (Q · ei )(ej · QT ) = aij (Q · ei )(Q · ej ). is the sum of the diagonal mixed components of A: ·i tr(A) = ai . . and in large part this can be done by showing that the application of Q to all the vectors of the relation does not change the form of the relation. The trace can be equivalently determined in the noncoordinate form tr(A) = E ·· A = E • A (the reader should check this).
We show that Q · xi is an eigenvector of Q · A · QT : (Q · A · QT ) · (Q · xi ) = Q · A · QT · Q · xi = Q · A · E · xi = Q · (A · xi ) = Q · (λi xi ) = λi (Q · xi ).52 Tensor Analysis with Applications in Mechanics It follows that ·j E ·· (Q · A · QT ) = (Q · ek )(Q · ek ) ·· ai (Q · ei )(Q · ej ) ·j = ai (Q · ek ) · (Q · ei ) (Q · ek ) · (Q · ej ) ·j = ai (ek · QT ) · (Q · ei ) (ek · QT ) · (Q · ej ) ·j = ai e k · Q T · Q · e i ·j = ai e k · E · e i ·j = ai e k · e i ·j i k = ai δ k δ j · = ak k ek · QT · Q · ej ek · E · ej ek · ej = tr(A). Indeed. Polar decomposition A secondorder tensor A is nonsingular if det A = 0 where A is the matrix of mixed components of A. Let e be an axis of rotation deﬁned by an orthogonal tensor Q. and let ω be the angle of rotation about e. It can be shown that the proper orthogonal tensor has the representation Q = E cos ω + (1 − cos ω)ee − e × E sin ω. It is possible to express such a tensor as a product of a symmetric tensor and another tensor. let A · xi = λi xi . Another example demonstrates that under the transformation Q the eigenvalues of a tensor A remain the same but the eigenvectors xi rotate. A .
24) Here Q is an orthogonal tensor of order two. By orthogonal decomposition such a tensor S may be written in the form S = λ1 i1 i1 + λ2 i2 i2 + λ3 i3 i3 . Hence all the eigenvalues λ of A · AT are positive. requires that we introduce some additional terminology. A symmetric tensor is said to be positive deﬁnite if its eigenvalues are all positive. x2 The numerator is positive because x · (A · AT ) · x = (x · A) · (AT · x) = (x · A) · (x · A) = x · A2 . (3.4. known as the polar decomposition theorem. If A is nonsingular then the tensor A · AT is symmetric and positive deﬁnite.Tensors 53 statement of this result. With these facts in hand we may turn to our main result. we begin with the deﬁnition (A · AT ) · x = λx of an eigenvalue λ and dot with x from the left to get λ= x · (A · AT ) · x . . The decomposition may be done in two ways: as a left polar decomposition A= S·Q or as a right polar decomposition A=Q·S. The diagonalization of A · AT now shows that it is positive deﬁnite. and S and S are positive deﬁnite and symmetric.25) (3. To see positive deﬁniteness. Symmetry follows from the equation (A · AT )T = (AT )T · AT = A · AT . Theorem 3. where all the λi > 0 and the eigenvectors ik constitute an orthonormal basis. Any nonsingular tensor A of order two may be written as a product of an orthogonal tensor and a positive deﬁnite symmetric tensor.
we have A · AT = λ1 i1 i1 + λ2 i2 i2 + λ3 i3 i3 . Thus we have expressed A = S · Q as in (3. To see that Q is orthogonal.54 Tensor Analysis with Applications in Mechanics Proof. We deﬁne S ≡ (A · AT ) 1/2 = λ1 i1 i1 + λ2 i2 i2 + λ3 i3 i3 since the λi are positive. The validity of (3.24). We have called a tensor positive deﬁnite if its eigenvalues are all positive. Let x be the image of r under A: x = A · r. If r locates any point on the unit sphere then r · r = 1.26) . we write Q · QT = (S−1 · A) · (S−1 · A)T = (S−1 · A) · (AT · (S−1 )T ) = S−1 · (A · AT ) · (ST )−1 = S−1 · S2 · S−1 = E. (3. let us show that a nonsingular tensor A operates on the position vectors of the points on the unit sphere to produce position vectors deﬁning the points of an ellipsoid. An alternative deﬁnition is that A is positive deﬁnite if x · (A · x) > 0 for all x = 0. Exercise 3.21. Because A · AT is positive deﬁnite and symmetric. We see that S−1 exists and is equal to 1 1 1 S−1 = √ i1 i1 + √ i2 i2 + √ i3 i3 . λ1 λ2 λ3 Now we set Q ≡ S−1 · A. Explain. As an application of polar decomposition.25) follows from deﬁning S ≡ QT · S · Q. where the ik are orthonormal.
Because S= λ1 i1 i1 + λ2 i2 i2 + λ3 i3 i3 . Moreover. Polar decomposition provides the background for introducing measures of deformation in nonlinear continuum mechanics.Tensors 55 Then r = A−1 · x. The tensor dev A is deﬁned by the above equality. expansion in the Cartesian frame ik with use of 3 x= i=1 xi ii .24) gives 1 = [(S · Q)−1 · x] · [(S · Q)−1 · x] = x · {[(S · Q)−1 ]T · (S · Q)−1 } · x = x · {S−1 · S−1 } · x (the reader can supply the missing details). It has the same eigenvectors as A.26) along with (3. S−1 = 3 j=1 1 ij ij λj reduces the above equation x · {S−1 · S−1 } · x = 1 to the form 3 i=1 1 2 x = 1. it is used to formulate constitutive equations in the theories of plasticity. Deviator and ball tensor representation For A we can introduce the representation A= 1 I1 (A)E + dev A. creep. 3 Such a representation is found useful in the theory of elasticity. and substitution into (3. but eigenvalues that diﬀer from the eigenvalues of A by (1/3)I1 (A): 1 ˜ λi = λi − tr A. λi i This is the equation of an ellipsoid. and viscoelasticity. 3 .
·jk These necessitate the respective transformation laws ˜i ˜j ˜ rijk = rmnp Am An Ap ˜ k . We then consider a linear space whose basis is the set of 27 quantities ei ej ek and call it the space of thirdorder tensors.27) The property of objectivity of R remains paramount. b. leading to the requirement that the components transform appropriately when we change the frame. we introduce triad quantities of the type R = abc where a. · ·k R = r i ei ej ek . Expanding these vectors in terms of a basis ei we obtain R = a i b j ck e i e j e k . A general element of this space. a tensor R of order three.1.9 Extending the Dyad Idea Thirdorder tensors can be introduced in a way that parallels the introduction of dyads in § 3. and c are vectors.27) in contravariant components. has the representation R = rijk ei ej ek . we could use the covarianttype representation R = rijk ei ej ek or either of the mixed representations R = rij ei ej ek . As an alternative to the representation (3. The now familiar procedure of setting rijk ei ej ek = rmnp em en ep ˜ ˜˜ ˜ under the change of frame ei = Aj ej i˜ gives rijk = rmnp Ai Aj Ak ˜ m n p — a direct extension of (3.8). (3. We continue to refer to the numerical values ai bj ck as the tensor’s components.56 Tensor Analysis with Applications in Mechanics 3. Using the tensor product as before.
is a tensor of order one. which arose naturally in our discussion of the vector cross product. This quantity. for instance. ˜i · m j k as is easily veriﬁed. while the double dot product of a triad with a dyad is given by abc ·· xy = a(c · x)(b · y). Verify the following formulas for operations involving the Levi–Civita tensor: (a) E ··· zyx = x · (y × z). Dot products involving triads follow familiar rules. Verify that E = −E × E. (c) E · x = −x×. The reader has surmised by now that the order of a tensor is always equal to the number of free indices needed to specify its components.23. ˜ij · · m n k ˜ ˜ r·jk = rmnp Ai An Ap . .22. Note: The notation of (c) may require some explanation. it becomes equivalent to the cross product y × x. (b) E ·· xy = y × x. When E · x is applied to another vector y.Tensors 57 and ˜ r··k = rmnp Ai Aj Ap . In Chapter 2 we also met a quantity whose components are speciﬁed by three indices: ijk . The scalar product of triads is deﬁned by the rule abc • xyz = (a · x)(b · y)(c · z). and the development of a special notation is unwarranted. it is rare that the action of some tensor can be described as the action of two vectors. is known as the Levi–Civita tensor and is given by E= ijk e i j k e e . Exercise 3. One may also deﬁne a triple dot product of a triad with another triad: abc ··· xyz = (c · x)(b · y)(a · z). A vector. This is the meaning of the right side of (c). The dot product of a triad with a vector is given by the formula abc · x = ab(c · x). Exercise 3. Although the notation is awkward. The result of applying the thirdorder tensor E to a vector x is a secondorder tensor E · x.
A fourthorder tensor C can be represented by several types of components: cijkl . in elasticity one encounters double dot products between the tensor of elastic constants and the strain tensor. In generalized form Hooke’s law becomes σ = C ·· ε where σ is the stress tensor. for example.10 Tensors of the Fourth and Higher Orders We can obviously extend the present treatment to tensors of any desired order. while the other three are mixed with indices in various positions. respectively. 2 2 . Dot products with vectors can be taken as before: the rule is that we simply dot multiply the basis vectors positioned nearest to the dot. For example. A dot product of a fourthorder tensor with a vector gives. R =C·x = cijkl ei ej ek el · xm em m = cijkl ei ej ek δl xm = cijkl xl ei ej ek — a tensor of order three. Similarly. and ε is the strain tensor. Double dot products also appear in applications. C = cijkl ei ej ek el . a dot product between a thirdorder tensor and a vector gives a secondorder tensor. ·jkl ci . Let us illustrate the essential points using tensors of order four. As before these components represent C with respect to various bases. ·· · cijk l . for instance. it continues to enjoy the linearity properties stated earlier. The density of the function of internal (elastic) energy in linear elasticity is 1 1 σ ·· ε = (C ·· ε) ·· ε.58 Tensor Analysis with Applications in Mechanics 3. C is the tensor of elastic constants. Carrying out such operations we may obtain results of various kinds. · ·kl cij . Wherever the dot product is utilized. cijkl . The ﬁrst and last are purely contravariant and purely covariant.
the only isotropic vector is 0. a material is isotropic when its constitutive equations are invariant with respect to certain transformations: the rotations and mirror reﬂections. In mechanics. So ˜ A = aij ei ej = aij Q · ei Q · ej = Q · (aij ei ej ) · QT . It is impossible to tell whether a ball made of isotropic material has been rotated through some angle. e2 . the last expression can be put in a more symmetrical form 1 ε ·· C ·· ε 2 in which the result does not depend on the order of operations. In the new basis we have A = aij ei ej . Let A be a secondorder tensor so that in a basis e1 . is isotropic: it is equally transparent in all directions. for example. Recall that any rotation or mirror reﬂection of R3 is uniquely deﬁned by an orthogonal tensor Q. Clearly. ˜ ˜˜ This means that A is isotropic if and only if the equation A = Q · A · QT holds for any orthogonal tensor Q. By the above deﬁnition we must have aij = aij . Any scalar quantity is isotropic. Isotropic tensors In engineering. . ˜ ˜˜ Let A be isotropic. We say that a tensor is isotropic if its individual components are invariant under all possible rotations and mirror reﬂections in R3 . isotropic materials play an important role. Air. Let us apply Q to each vector of the basis: ˜ e k = Q · ek . These are the materials whose properties are the same in all directions. First we consider the question of when various tensorial quantities can be isotropic.Tensors 59 Because σ and ε are symmetric. e3 we have A = aij ei ej . material properties are expressed via constitutive relations. From a mathematical point of view.
3.29. The following is. unfortunately. Let us demonstrate this.28) where α. . Exercise 3. The quantities C. The proof is cumbersome and we omit it.24. . .27.60 Tensor Analysis with Applications in Mechanics Common sense tells us that the metric tensor E. . . the general form of the fourthorder isotropic tensor is αEE + βek Eek + γI. Under an orthogonal transformation Q of R3 . and 3.11 Functions of Tensorial Arguments The reader is familiar with the notion of a function f (x1 . ˜ By the general deﬁnition. A = λE for some scalar λ. If we regard xk as a Cartesian component of a vector x = (x1 . If λ is a scalar. 3. which is the unit tensor as well. .28. should be isotropic. β. γ are arbitrary scalars [Jeﬀreys (1931)]. . xn ) in n variables. . and ε will be considered further in Chapter 6. not a trivial exercise. For any orthogonal Q we have Q · QT = E. This can be rewritten as E = Q · E · QT . (3. This fact can be applied to the tensor of elastic constants for an isotropic material. It asserts that any isotropic secondorder tensor takes the form λE. xn ). Show that A is an isotropic tensor of order two if and only if it is a ball tensor: that is. σ. The properties of the tensors in the representation are exhibited in Exercises 3. then clearly the ball tensor λE is isotropic as well. . Hence E is isotropic. it is isotropic if cijmn = cijmn for all Q. a fourthorder tensor C = cijmn ei ej em en takes the form C = cijmn (Q · ei )(Q · ej )(Q · em )(Q · en ). ˜ In a Cartesian frame.
Tensors
61
then we can regard f as a function of the vectorial argument x: f = f (x). But a logically good deﬁnition of such a function dictated by physics should require that f be independent of the representation of x in a basis. Similarly, we can consider a function of one or more tensorial arguments. In a ﬁxed basis, such a function reduces to a function in many variables, the components of the tensorial arguments. Again, however, a true function of a tensorial argument cannot depend on the basis representations of its arguments. To extend this notion further, we can consider functions that take values in the set of vectors — or even tensors. Such functions arise in applications. For example, a force vector f = f (t) can be given as a function of time t. Later, we will encounter other functions that take values in the set of tensors of some order. Depending on this latter set, the function may be termed scalarvalued, vectorvalued, or tensorvalued. As for any function in many variables, we can apply the tools of calculus to tensorvalued functions. These include the notion of continuity, the ﬁrst diﬀerential, and derivatives. We will consider these topics later. Linear functions In linear elasticity and linear shell theory, linear relations and quadratic functions (such as occur in strain energy expressions) play central roles. We deﬁne a linear function of a tensorial variable as a function f which, for any tensors A, B and scalars λ, µ, satisﬁes the relation f (λA + µB) = λf (A) + µf (B). This mimics the deﬁnition of a linear matrix operator. From this point of view, the equation y = kx + b represents a linear function only if b = 0. Theorem 3.5. Let f be a scalarvalued function of a vectorial argument x. There is a unique c such that for all x, f (x) = c · x. (3.29)
Proof. We expand x = xk ek with respect to the basis ek . By linearity, f (x) = xk f (ek ). Equation (3.29) holds with c = f (ek )ek . Supposing the existence of two vectors c1 and c2 such that c1 · x = c2 · x, and putting x = c2 − c1 , we get c1 = c2 .
62
Tensor Analysis with Applications in Mechanics
Now let us consider a scalarvalued function whose argument is a secondorder tensor A. Clearly, for a secondorder tensor B, the function f (A) = tr(B · A) = B ·· AT is linear. Exercise 3.25. Show that B ·· AT = A ·· BT . Theorem 3.6. Let f be a scalarvalued function of a secondorder tensor A. There is a unique secondorder tensor B such that for all A, f (A) = tr(B · AT ). (3.30)
The proof is left to the reader. The representation of any tensorvalued function of a tensorial argument is similar. We use it to introduce the tensor of elastic constants in linear elasticity. Theorem 3.7. Let F = F(A) be a linear function from the set of secondorder tensors A to the same set of secondorder tensors. There is a unique fourthorder tensor C such that for all A, F(A) = C ·· AT . Proof. Write A = amn em en and introduce C by the formula C = F(em en )em en . Then C ·· AT = F(em en )em en ·· aij ej ei = F(em en )amn = F(A). So the representation is valid. Proof of uniqueness is left to the reader. Exercise 3.26. An operator on the set of vectors x is given by the formula y = B · x where B is a secondorder tensor. This can be extended to the set of secondorder tensors by the equation Y = B · XT . Show that by using the fourthorder tensor C = B · en Een , we get C ·· XT = B · XT for all X. Note that we cannot represent Hooke’s law using only this operation B · X. Looking back, we note that the general form of a fourthorder isotropic tensor contains three independent isotropic tensors. Their properties are exhibited in the following exercises. Exercise 3.27. Show that the identity operator from the representation (3.31) is I = ek em ek em ; that is, for all A we have I · ·AT = A. (3.31)
Tensors
63
Exercise 3.28. A linear function is deﬁned by the equality F(A) = AT . Show that the corresponding fourthorder tensor is ek Eek , i.e., show that ek Eek ·· AT = AT . Exercise 3.29. A linear function is deﬁned by F(A) = (tr A)E. Show that the corresponding fourthorder tensor is EE, i.e., show that EE ·· AT = (tr A)E. Isotropic scalarvalued functions A scalar function of a tensor is said to be isotropic if it retains its form under any orthogonal transformation of the space (or equivalently, of its basis). A scalarvalued function f (A) of a secondorder argument A is isotropic if and only if for any orthogonal tensor Q we have f (A) = f Q · A · QT . Let us demonstrate that any eigenvalue λ of a secondorder tensor A, when considered as a scalarvalued function of A (i.e., λ = f (A)), is isotropic. Indeed, an eigenvalue satisﬁes the characteristic equation det(A − λE) = 0. The eigenvalues of Q · A · QT satisfy det(Q · A · QT − λE) = 0. The equality det(Q · A · QT − λE) = det(Q · A · QT − λQ · E · QT ) = det Q · (A − λE) · QT = (det Q)2 det(A − λE) = det(A − λE) shows that the eigenvalues of A and Q · A · QT satisfy the same equation. Hence they coincide. Exercise 3.30. Show that the invariants I1 (A), I2 (A), I3 (A) are isotropic functions of A.
64
Tensor Analysis with Applications in Mechanics
It can be shown that any scalarvalued isotropic function of a secondorder tensor is a function of its invariants [Lurie (1990); Ogden (1997); Truesdell and Noll (2004)]. This is used in nonlinear elasticity to introduce the constitutive equations for isotropic bodies. By Theorem 3.6, a scalarvalued linear function has the representation f (A) = B ·· AT for some secondorder tensor B. Theorem 3.8. The function f (A) = B ·· AT is isotropic if and only if B is an isotropic tensor of order two, and hence a ball tensor: B = λE. Proof. Because f is isotropic, we have B ·· AT = B ·· (Q · A · QT )T for any orthogonal Q. Using B ·· AT = tr(B · A) we get B ·· (Q · A · QT )T = tr(B · Q · A · QT ) = tr(QT · B · Q · A) = (QT · B · Q) ·· AT . Hence for any A we have B ·· AT = (QT · B · Q) ·· AT . But this occurs if and only if the relation B = QT · B · Q holds for any Q. Therefore B is isotropic. Exercise 3.31. Show that a scalarvalued, linear, isotropic function of a secondorder tensor A is a linear function of tr A: that is, f (X) = λ tr X. If a scalar function maintains its form under some subgroup of orthogonal transformations, then one can ﬁnd functions that are invariant under these. The reader will ﬁnd applications of this idea in books on crystallography and elasticity.
Tensors
65
Isotropic tensorvalued functions Now we consider a function F whose domain and range are the set of secondorder tensors. We say that F(A) is isotropic if the components of its image value do not change from one Cartesian basis to another. So F is isotropic if F Q · A · QT = Q · F(A) · QT holds for any orthogonal tensor Q. An example of an isotropic tensorvalued function is F = λA. As with a scalarvalued function, it can be shown that a tensorvalued linear function F represented in terms of a fourthorder tensor C, F = C ·· AT , is isotropic if and only if C is isotropic and therefore is given by (3.28). So F takes the form F(A) = αE tr A + βAT + γA. Mechanicists employ functions whose domains and ranges can be sets of symmetric tensors. This imposes certain additional restrictions on the form of C. Indeed, let A = AT and F(A) = F(A)T . Using the representation F(A) = cijmn amn ii ij with a Cartesian basis ii , we see that cijmn = cjimn = cijnm . In the general case, C has 81 independent components. But, in view of symmetry, C has only 36 independent components. An isotropic fourthorder tensor C satisfying the symmetry conditions takes the form αEE + β(ek Eek + I). So the general form of an isotropic linear function satisfying the symmetry condition is F(A) = αE tr A + 2βA. We will introduce the elements of calculus for functions of tensorial arguments. First, we require a way to gauge the magnitude of a vector or tensor. Suitable norms exist for this purpose.
66
Tensor Analysis with Applications in Mechanics
3.12
Norms for Tensors, and Some Spaces
In § 2.7 we introduced a norm in the space R3 . It can be immediately extended to Rk for any k as x = (x · x)1/2 , where the inner product of x and y in the space is given by x · y. Similarly, we can introduce a norm and inner product in the set of secondorder tensors. We denote the inner product by (A, B) and deﬁne it using the dot product as (A, B) = A ·· BT = aij ei ej ·· bts es et = aij bts gjs git = aij bts g js g it = aij ei ej ·· bts es et
s t = aij bts δj δi
= aij bij . It is clear that in a Cartesian frame (A, A) is the sum of all the squared components of A, so this is quite similar to the scalar product of vectors. Using the same reasoning as above, we can show that the axioms of the scalar product hold here as well (note that using only a Cartesian frame, we could regard the components of a tensor as those of a ninedimensional vector, so this is another reason why the axioms hold). In this case the inner product axioms are written for arbitrary secondorder tensors A, B, C as (i) (A, A) ≥ 0, and (A, A) = 0 if and only if A = 0; (ii) (A, B) = (B, A); (iii) (αA + βB, C) = α(A, C) + β(B, C) for any real α, β. By linear algebra, the expression A = (A, A)1/2 is a norm in the set of secondorder tensors. It satisﬁes the following axioms. (i) A ≥ 0, with A = 0 if and only if A = 0; (ii) αA = α A for any real α; (iii) A + B ≤ A + B .
Tensors
67
We still have the Schwarz inequality (A, B) ≤ A B .
In linear algebra, many particular implementations of the vector and tensor norms can be introduced. We can do the same in any ﬁxed frame, but if we wish to change the frames under consideration we must remember that these norms should change in accordance with the tensor transformation rules. Note that the representation of a linear function in the previous section used AT as an argument. This reﬂects the form of the inner product on the set of secondorder tensors. It is worth noting two important properties of the norms we have introduced, as these are used in analysis. Let A and B be secondorder tensors and x a vector. Then the relations A·x ≤ A and A·B ≤ A hold. The latter implies that Ak ≤ A
k
x
B
.
(3.32)
Using this property, we can justify the introduction of tensorvalued functions like eA . Exercise 3.32. Using (3.32), prove convergence of the series eA = E + 1 1 A + · · · + Ak + · · · . 1! k!
It is easy to extend the notion of inner product to tensors of any order; we introduce the inner product (A, B) = ai1 i2 ···in bi1 i2 ···in . The reader can represent this in all its particular forms and verify the inner product axioms.
Consider. can be modiﬁed by replacing the absolute value with a suitable norm. etc. the ordinary deﬁnitions of limit. b] if for any ε > 0 there is a δ > 0. e2 . b].68 Tensor Analysis with Applications in Mechanics Some elements of calculus As in ordinary calculus. b] with values in R3 . Various norms can be used on R3 .. in linear algebra it is shown that on a ﬁnitedimensional space all norms are equivalent. e3 so that f (t) = f k (t)ek . we say that a function f (t) is continuous on [a. then the product rule holds in the form (f (t) · g(t)) = f (t) · g(t) + f (t) · g (t). such that for any t = t0 and t − t0  ≤ δ we have f (t) − a < ε. However. if f (t) and g(t) are both diﬀerentiable at t. Furthermore. For example. independently of the form taken by the norm. b] to the set of tensors of some order. Equivalence of two norms · 1 and · 2 means that there exist positive constants c1 and c2 such that for any element x of the space we have 0 < c1 ≤ x x 1 2 ≤ c2 < ∞. then df (t) df k (t) = ek . Consequently. When a = f (t0 ). If we expand f in a basis e1 . for example. a function on the segment [a. we say the function is continuous at t0 . The derivative of f (t) at t0 is given by f (t0 ) = df dt = lim t=t0 t→t0 f (t) − f (t0 ) . When a function takes values in R3 . b] if it is continuous at each point of [a. derivative. where c1 and c2 do not depend on x. dependent on ε. integral. we can introduce the notions of limit and continuity at a point t0 in the same way as in calculus: The function f : [a. This mapping pairs each point of [a. b] with at most one vector from R3 . dt dt . Such a function will be given on a certain set. We may similarly construct a function from [a. As in calculus. we can introduce the notion of a function in one or many variables that takes values in a set of vectors or tensors. either norm can be used in the deﬁnition of limit: the limit will exist or not. b] → R3 has limit a at t = t0 ∈ [a. t − t0 The derivative of a vectorvalued function has many properties familiar from ordinary calculus.
the norm of a vector function is usually introduced in a Cartesian frame.Tensors 69 Here we have assumed that the ek do not depend on t.33) This formula depends on the Cartesian frame ik .33).33) is an improper way to characterize the intensity of a vector ﬁeld. If we use a similar norm involving vector components in a frame having singular points in V — as may be the case with spherical coordinates — we obtain a norm that is not equivalent to (3. otherwise an application of the product rule would have been required (we will encounter this situation later). For example. Deﬁnite integration can also be carried out in componentwise fashion: b b f (t) dt = a a f k (t) dt ek . dt dt Some normed spaces In textbooks on functional analysis. the theory of functions in many variables is generalized in a similar way to the tensorial functions. . Moreover. The reader should remember that formally. Clearly.33). We record here only the formula for the derivative of a tensorial function F(t) = f mn (t)em en : df mn (t) dF(t) = em en .34) If we would like to use a norm of the type (3. the proper norm for a function given in curvilinear coordinates is based on the above norm: f (x) = f i (x)ri = max f i (x)fi (x) V 1/2 . the norm on the space C(V ) of continuous vector functions given on a compact region V is f (x) C = fk (x)ik C = max max fk (x) k V . The integral has all the properties familiar from calculus. (3. we should change the absolute value to the norm. we need to remember that during transformation of the frame we must change the form of the norm accordingly. However. This means that (3. (3. all this can be extended to the case of a function in one scalar variable taking values in the set of tensors of some order. in all the deﬁnitions of ordinary calculus.
Exercise 3.70 Tensor Analysis with Applications in Mechanics On the set of secondorder tensor functions continuous on a compact region V . The inner product is then (A. More complex relations occur in nonlinear elasticity or plasticity. Here one must diﬀerentiate tensorial functions. Finally. What is the inverse to E + A? 3. for example. . B) = V aij bij dV. ∂xk . ∆x→0 ∆x The ﬁrst diﬀerential is given by the formula df = f (x) dx. For a function in n variables we have n (3. and introduce all forms of the inner product and norm corresponding to (3. (3. Let A be a tensor of order two with A = q < 1. . Let A be a tensor of order two with A = q < 1. Demonstrate that E − A has the inverse (E − A)−1 that is equal to E + A + A2 + A3 + · · · + An + · · · . the stress tensor can be found as the derivative of the strain energy with respect to the strain tensor.33. xn ) = k=1 ∂f dxk .34. are related though the generalized form of Hooke’s law. Exercise 3. We recall that for an ordinary function f (x). The stress and strain tensors.34): A(x) = aij (x)ri rj = max aij (x)aij (x) V 1/2 . there are functions that relate two tensors.35) The reader should verify all the axioms of the inner product for this. . In elasticity. we can introduce a norm similar to (3. for example.36) df (x1 . note that instead of the norms for continuous vector and tensor functions we can introduce the norms and scalar products corresponding to the space of scalar functions L2 (V ). .13 Diﬀerentiation of Tensorial Functions In the linear theory of elasticity.35). the derivative is f (x) = lim f (x + ∆x) − f (x) .
. (3. the function f (x + ε dx) is a function in one variable ε. Later we will refer to this vector quantity as the gradient of f . . . write f (x1 . .37) it follows that df (x1 . From (3. xn ) = df (x + ε dx) dε . xm ) = f (x). and consider this as a function in a vectorial variable. . . we can formally represent this as n df (x1 . Let ε be a real variable. For ﬁxed x and dx. For the intermediate expression. ε=0 n = ε=0 k=1 ∂f ik ∂xk n · m=1 dxm im . xn ) = k=1 ∂f ik ∂xk n · m=1 dxm im . . . .37) (3.Tensors 71 Using a Cartesian basis i1 . . .x = k=1 ∂f ik ∂xk and call it the derivative of f with respect to x. xn ) = df (x + ε dx) dε = f. The chain rule formally applied to this function gives us df (x + ε dx) dε hence df (x1 . we introduce the notation n f. . . We began with a function in n variables. . . In the same way we introduce n dx = m=1 dxm im where the dxm are some quantities that are not necessarily inﬁnitesimal. Let us regard n x= m=1 xm im as a vector. . .38) The righthand side of this equality is termed the Gˆteaux derivative of a f at the point x in the direction dx. .38) and (3. . but could in fact consider a function of a vectorial argument f (x) and present the same operations in noncomponent form. in . .x · dx. ε=0 .
When we wish to have this in a form that does not include the basis vectors. Here. Note that in any basis. .) For f. The order of F.39) ε→0 ∂ε ε ε=0 for any tensor dX.x are partial derivatives of f with respect to the components of the expansion of x in the basis. we can repeat a the method used above for the ﬁrst diﬀerential written in Cartesian components. we must use the expansion of dx in the dual basis.X ·· dXT . the components of f. Again.X = ∂fij ii ij im in . and hence the summation convention applies in situations such as this.X (X) is four.X is called the derivative of f with respect to the tensor argument X. The reader may wish to verify this by direct calculation. ε (3.X (X) and F.X = im in . it becomes valid for any basis.X (X). First we consider a scalarvalued function f (X) whose argument X belongs to the secondorder tensors. In a ﬁxed basis. In component form it is F. the components of X. Then df = f. These ideas can be extended to tensorial functions of tensorial arguments in a straightforward manner. The expression f.X (X) is deﬁned as F. As above. in the Cartesian basis we have ∂f f. not necessarily inﬁnitesimal. it can be considered as a function in 3 × 3 = 9 variables.X ·· dXT ≡ lim (3. For a function that maps values X in the set of secondorder tensors into the same set.40) This is a particular case of the Gˆteaux derivative. Then we introduce the Gˆteaux derivative: a f (X + ε dX) − f (X) ∂ f (X + ε dX) = f. in a Cartesian basis in R3 we can introduce the ﬁrst diﬀerential df .X ·· dXT = ∂ F(X + εdX) ∂ε ≡ lim ε=0 ε→0 F(X + εdX) − F(X) . Although the last formula was derived in Cartesian coordinates. the derivative F. ∂xmn where the xmn are the Cartesian components of X. ∂xmn (We recall that in Cartesian bases ik = ik . it holds in any basis.72 Tensor Analysis with Applications in Mechanics When we derive a relation in Cartesian coordinates but present it in noncomponent form.
Let the Xi be secondorder tensors. . . . Xm ) in several tensorial arguments. . Let f (X1 . . Using the method applied to the previous exercise. is deﬁned by the equality ∂f ∂ f (X1 . . (k) (1) (m) . Then ∂f ∂f = i i . . In a similar way. (i) j k ∂Xi ∂xjk Exercise 3.36. Xm ) ·· YT = ∂Xi ∂ε ε=0 for any secondorder tensor Y. . Xm ) = f (xij . .Tensors 73 the notations df (X) dX and dF(X) dX are also used. . ε=0 As in case of a tensorial function of one variable. the components of the above partial derivatives can be expressed as ordinary partial derivatives of the components. . Find the derivative of f (X) = I1 (X) ≡ tr X with respect to X. . when calculating dF we get a linear function in dX. Exercise 3. . . xij ) be a function in 9m variables. When function F takes values in the set of secondorder tensors. . . . This is why we studied linear functions earlier. . the components xij of Xk . . the partial derivative ∂F/∂Xi is similarly deﬁned by equality ∂F ∂ F(X1 . . Find the derivative of f (X) = tr X2 . . . . Xm ) ·· YT = ∂Xi ∂ε . Exercise 3. . Xi + εY. We describe this representation in Cartesian coordinates. Xi + εY. Now we introduce a partial derivative for a scalarvalued function f (X1 . . show that the derivative of f (X) = tr X3 with respect to X is 3(XT )2 .37. denoted by ∂f /∂Xi . we can deﬁne the derivative of a tensorvalued function for tensors of any order. The partial derivative of f with respect to Xi . . Clearly. . .35.
That is.27 and 3. Using Exercises 3. I2 . ∂I3 We have presented examples of the derivatives of scalarvalued functions. Show that the derivative of f (X) = I3 (X) ≡ det X with respect to X is [X2 − I1 (X)X + I2 (X)E]T . demonstrate that its derivative is f. 3. f = f (I1 . Show that the derivative of 1 f (X) = I2 (X) ≡ [tr2 X − tr X2 ] 2 with respect to X is I2 (X). Exercise 3.X = I1 (X)E − XT .28. The strain energy of an isotropic elastic medium is a function of the invariants Ik of the strain tensor.40. and 3. The derivatives of tensorvalued functions are more complicated.74 Tensor Analysis with Applications in Mechanics Exercise 3. The formulas for diﬀerentiating the invariants of the strain tensor are used to write down the constitutive equation for a nonlinear elastic isotropic material under ﬁnite deformation. X. Note that the derivative of a linear function F(X) = C ·· XT is C: F. the derivative of a secondorder tensor X with respect to X is I. Using the results of Exercises 3.X = ek Eek .36.41. Exercise 3.38.X = C. then F.39.35.39.X = ∂f ∂f ∂f E− + I1 + I2 ∂I1 ∂I2 ∂I3 ∂f ∂f + I1 ∂I2 ∂I3 XT + ∂f T 2 X . . verify that if F(X) = X then F. Exercise 3. By the exercise above.X = I.X = I. a tensor of order four. I3 ). If F(X) = XT .
X to be symmetric? The sets of symmetric and antisymmetric secondorder tensors are subspaces of the space of all secondorder tensors. Let us consider the derivative of a scalarvalued function f (X) of a symmetric secondorder tensor X. Why do we require f. If in this deﬁnition we do not require f. then (3.X to be symmetric. the stress tensor σ and strain tensor ε.54). f.X ·· dX ε=0 (3. The component representation of f. deﬁnes f. holding only for all the symmetric tensors dX. So the subspaces of symmetric and antisymmetric tensors are mutually orthogonal.X = ∂f im in . 3. Note that the deﬁnition of derivative for a function with respect to a symmetric tensor argument is closely related to the problem of representing a linear function acting on the subspace of symmetric tensors. which diﬀers slightly from the general case (see Problems 3. In (3. we ﬁnd the derivative of the function f (X) = C ·· X. and 3.X into the deﬁnition: (f. ∂xmn So we have brought the symmetry of f. are symmetric (cf.41). As an example.X . as this speciﬁc case has some peculiarities. dX is arbitrary but symmetric.41) for any symmetric tensor dX. By deﬁnition. The derivative f.X nonuniquely up to an additive term B that can be any antisymmetric secondorder tensor. It is easy to see that A ·· B = 0.39) as follows.X ·· dX = C ·· dX .Tensors 75 On symmetric tensor functions The principal tensors of linear elasticity.41).X is a secondorder symmetric tensor that satisﬁes the condition ∂ f (X + ε dX) ∂ε = f. Chapter 6).X )T = f.X in a Cartesian basis is f..52.53. We modify the deﬁnition (3. where C is a secondorder tensor and XT = X. Let A be a symmetric tensor and B an antisymmetric secondorder tensor. So we consider the problem of diﬀerentiating a tensorvalued function of a symmetric tensorial argument.
Similar changes permit the deﬁnition of partial derivative for tensorial functions in many symmetric tensorial arguments.X becomes f. At ﬁrst glance it seems that f. we should modify the deﬁnition of the derivative of a tensorvalued function of a secondorder symmetric argument. 2 This expression is uniquely deﬁned.pt of F.X = C if and only if C = CT .X = C..X . the derivative with respect to X is denoted by F. The symmetry of F.pt = Fij. Let F(X) = C ·· X. we can represent C as a sum of symmetric and antisymmetric tensors: C= But 1 1 (C − CT ) ·· dX = (C − C) ·· dX = 0 2 2 and thus C ·· dX = 1 (C + CT ) ·· dX 2 1 1 (C + CT ) + (C − CT ) 2 2 as claimed.76 Tensor Analysis with Applications in Mechanics for all dX such that dX = dXT .tp . it takes values in the subset of tensors of order n + 2 which are symmetric in the two last indices and satisfy the equality F..X we have Fij. . We said that f.X must be symmetric. Indeed.X in the last two indices means that for the components Fij.X ·· dX = ∂ F(X + ε dX) ∂ε (3. so f..42) ε=0 for all symmetric secondorder tensors dX... Indeed the derivative must be a symmetric tensor. and this does hold for an arbitrary argument X. Exercise 3. In a similar fashion. A consequence is that f. But for a symmetric argument X the answer changes. Now we consider the case when C is not symmetric..42.X = 1 (C + CT ). If F takes values in the set of tensors of order n.
X = (C ·· X + C ·· X + X ·· C + X ·· C ). and C by transposing the ﬁrst two indices of the components. Suppose C = C = C and C ·· X = X ·· C for any symmetric tensor X.14 Problems In this section. this means the equalities cmnpt = cnmpt and cmnpt = cptmn hold for any sets of indices. Ω an antisymmetric secondorder tensor. Using the solution of Exercise 3.2 Find the components of the tensor i1 i2 − i2 i1 + 2i3 i3 in the Cartesian basis (i1 . i3 ). C = cmnpt in im ip it .. in terms of components. demonstrate that f. Exercise 3.e. i..1 Write out the components of the dyad i1 i2 in a Cartesian basis (i1 . we use A.e. Demonstrate that F. Exercise 3. Let f (X) = 1 X ·· C ·· X. B.X = 1 (C + C ). Let f (X) = 1 X ·· C ·· X. . 2 where C = cmnpt im in ip it and X is a secondorder symmetric tensor. X to denote secondorder tensors. 3.X = C ·· X.43.Tensors 77 where C = cmnpt im in ip it is a fourthorder tensor and X is a secondorder symmetric tensor. 2 where C is derived from C by transposing the last two indices: C = cmnpt im in it ip . C = cmnpt im in it ip .44. C. 3.43. i2 . i. i3 ). 4 where C is derived from C by transposing the last two indices of the components of C. 2 where C is a fourthorder tensor and X is a secondorder symmetric tensor. 3. Demonstrate that 1 f. λ a scalar. i2 . unless otherwise stated. and Q an orthogonal secondorder tensor.
(b) i1 i2 − i2 i1 + 2i3 i3 . Show that I1 (A−1 ) = I2 (A) . .3 Write out the components of the tensor that is the dyad a1 a2 in the Cartesian basis (i1 .78 Tensor Analysis with Applications in Mechanics 3.10 Find the left and right polar decompositions of the following tensors: (a) λE. (b) i1 i2 + i2 i1 .7 Show that I1 (X) = tr X.9 Let A be a symmetric. 2). I3 (A) I2 (A−1 ) = I1 (A) . (e) 2i1 i1 + 3i2 i2 + 4i3 i3 . 3. (e) i1 i1 + 2i1 i2 + 2i2 i1 + i3 i1 + i1 i3 . i2 . (b) aa + bb + cc if a. i3 ). −1. 3. 3. −2) and a2 = (1. (c) i1 i2 − 2i2 i1 + i1 i3 . where a1 = (−1. (d) i1 i2 + i2 i3 + i1 i3 .5 Determine the symmetric and antisymmetric parts of the following tensors: (a) i1 i2 . (d) λE. I2 (X) = 1 2 [tr X − tr X2 ]. 3. b.4 Show that a0 = 0b = 0. I3 (A) −1 I3 (A−1 ) = I3 (A). 2 I3 (X) = det X. 3. (d) aa.8 Find the invariants of the following tensors: (a) aa. (c) i1 i1 . (c) i1 i1 + i2 i2 . invertible tensor. 2. 3. c are mutually orthogonal. (e) i1 i1 + 2i1 i2 + 2i2 i1 + i3 i1 + i1 i3 . 3. (b) i1 i2 .6 Determine the ball and deviator parts of the following tensors: (a) i1 i2 .
(d) λE + ai1 i1 + bi2 i2 . Prove that A + ab is a nonsingular tensor if and only if 1 + b · A−1 · a = 0. Demonstrate that tr(A · Y) A X = aY − a tr A2 satisﬁes the equation tr(A · X) = 0.17 Let A be a symmetric tensor.14 Under what conditions will the equation aX + E tr X = 0 have a nonzero solution X? 3.11 Suppose tr A2 = 0. 3. ﬁnd a solution X and establish conditions for uniqueness. 3. Show that I2 (dev A) ≤ 0. 3. 3.Tensors 79 (c) λE + ai1 i1 .18 Let A be a nonsingular tensor and let a and b be arbitrary vectors. (c) X + a(tr X)A = B.12 Let a nonzero scalar a and secondorder tensors A. (A + ab)−1 = A−1 − 1 + b · A−1 · a . and Y is an arbitrary secondorder tensor.13 For the equations (a) X + tr(A · X)B = C.16 Demonstrate that in a Cartesian basis E = i 1 i2 i3 + i2 i3 i1 + i3 i1 i2 − i1 i3 i2 − i3 i2 i1 − i2 i1 i3 . (b) XT + tr(A · X)B = C. 3. a is an arbitrary scalar. (e) ai1 i1 + bi2 i2 + ci3 i3 . and 1 (A−1 · a)(b · A−1 ). B be given.15 Under what conditions will the equation aX + A tr X = 0 have a nontrivial solution? 3. Find a solution X of the equation aX + tr(A · X)E = B. 3.
(e) A if A = AT .80 Tensor Analysis with Applications in Mechanics 3. 3. 2 3.20 Show that (a) (E × ω)2n = (−1)n−1 (ω · ω)n−1 (ωω − Eω · ω). . (b) E × ω. 3. 3. Show that symmetrical tensors A and B are commutative if their sets of eigenvectors coincide. Find the vectorial invariants of the following tensors: (a) aa.19 Show that (E × ω)2 = ωω − Eω · ω. (c) a × E × b = ba − a · bE. (ab)× = a × b.23 Show that (a × A)× = A · a − a tr A. 3.29 Let the symmetric tensors A and B be commutative. the vectorial invariant was introduced by J. ω × A = −(AT × ω)T . (c) ab − ba.24 Show that (A × B) ·· BT = 0 holds for arbitrary secondorder tensors A and B. Demonstrate that (A · B)× = 0. For example. 3. (d) E.28 Let a · b = 0. 3. 3.25 Show that tr [b × (a × A)] = b · A · a − a · b tr A. 3.26 Show that A × ω = −(ω × AT )T . Demonstrate that the dyads aa and bb are commutative. Gibbs (1839–1903) as A× = amn em × en .W. 3.22 Show that A× = 1 (A − AT )× .21 For a secondorder tensor A = amn em en .27 Two secondorder tensors A and B are called commutative if A · B = B · A. (b) (E × ω)2n+1 = (−1)n (ω · ω)n E × ω.
41 Find the principal invariants of the tensor (E × ω)2 . 3.33 Show that if Ω is an antisymmetric tensor. 3. Show that the tensor Q = E − 2ee is orthogonal. 3. 3) and dm (m = 1. 3.30 Show that the tensor Q = i1 i2 − i2 i1 + i3 i3 is orthogonal. Establish the identity Q× = 2 sin ωe. 3.37 Let ek (k = 1. 3. We call θ a ﬁnite rotation vector. Q= 4 + θ2 where θ2 = θ · θ. then the tensor Q = (E + Ω) · (E − Ω)−1 is orthogonal. 3. A−1 = det A Use this to ﬁnd the inverses of the following tensors: .43 From the Cayley–Hamilton theorem for a nondegenerate tensor. it follows that 1 A2 − I1 (A)A + I2 (A)E .34 Show that the tensor Q = eΩ is orthogonal if Ω is antisymmetric. 3. 3.32 Let e · e = 1. where ω and e are the rotation angle and axis. 3.42 Show that I3 (A) = 1 tr3 A − 3 tr A tr A2 + 2 tr A3 .40 Find the principal invariants of the tensor E × ω.38 Let Q be a proper orthogonal tensor. 3. 2. is orthogonal.39 Find the principal invariants of the tensor αE + βee.36 Let Q be an orthogonal tensor.Tensors 81 3. 6 3.35 Demonstrate that the tensor 1 (4 − θ2 )E + 2θθ − 4E × θ . 3. respectively. 2. 3) be two orthonormal bases.31 Show that the principal invariants of an orthogonal tensor satisfy the 2 relations I1 I3 = I2 and I3 = 1. Establish the identity (Q − E) ·· (QT − E) = 6 − 2 tr Q. Verify that Q = ek dk is an orthogonal tensor.
Demonstrate that cof A = λ2 λ3 i1 i1 + λ1 λ3 i2 i2 + λ1 λ2 i3 i3 . that is.45 Let A be a secondorder tensor. suppose the representation ∞ f (x) = k=0 f (k) (0) k x k! holds and the series converges for x < r. In the ball. 3. (c) cof λA = λ2 cof A. the relation ∞ f (x) = k=1 f (k) (0) k−1 x (k − 1)! f (k) (0) k X . (b) ai1 i3 + bi2 i2 + ci3 i1 . 3. (d) tr(B · X). A tensor X is called the cofactor of A and denoted by X = cof A if X satisﬁes the equation A · XT = XT · A = (det A)E.82 Tensor Analysis with Applications in Mechanics (a) i1 i1 + 2i2 i2 + i3 i3 + i1 i2 + i2 i1 . (d) aE + bi1 i2 . (b) a · X · a. k! holds. 3.47 Let f (x) be an analytic function in an open ball x < r centered at x = 0. (d) I2 (A) = tr cof A. c = 0. Check the following properties of cofactor: (a) cof(AT ) = (cof A)T . (c) ai1 i1 + i2 i2 + i3 i3 + bi1 i2 . then cof A = (det A)A−T . 3. Let ∞ F(X) = k=0 .44 Find the derivative with respect to the tensor X of the following scalarvalued functions: (a) tr X4 . (c) a · X · b.46 Let A = λ1 i1 i1 + λ2 i2 i2 + λ3 i3 i3 . (b) cof(A · B) = cof A · cof B. b. If A is nonsingular. a.
C = cmnpt im in ip it . 3. for all A = AT .X = k=1 f (k) (0) T k−1 X = f (XT ) (k − 1)! for X < r. Show that there is a unique symmetric secondorder tensor B such that f (A) = tr(B · A) for any symmetric A. 3. Show that the derivative of f (X) with respect to XT is equal to the transpose of its derivative with respect to X. On S the inner product A ·· BT of the whole space takes the form A ·· BT = A ·· B. that is A = AT . i. Show that S is a linear subspace of the linear space of all the secondorder tensors. demonstrate that (det X). 3.53 Let F = F(A) be a linear function from the set of secondorder symmetric tensors A to the set of secondorder tensors. Show that there is a unique fourthorder tensor C such that F(A) = C ·· A. (b) tr(A · X)E.50 For a nonsingular tensor X.Tensors 83 Demonstrate that ∞ [tr(F(X)].XT = (f (X). 3..51 Let S be the set of all symmetric secondorder tensors. 3.e. cmnpt = cmntp .52 Let f be a scalarvalued function of a symmetric secondorder tensor A. (c) a · X · bcd.48 Let f (X) be a scalarvalued function.49 Find the derivative with respect to X of the following tensorvalued functions: (a) tr(A · XT )E.X = (det X)X−T . 3.X )T . f (X).
.. 3.pt = −cmn. Show that there is a unique fourthorder tensor C such that F(A) = C ·· A. T C = cmnpt im in ip it ..84 Tensor Analysis with Applications in Mechanics 3. cmnpt = cmntp . Show that Q is a linear subspace of the linear space of all secondorder tensors. where 0 is the zero tensor of order n − 2. cmnpt = cnmpt .56 Let C be a tensor of order n that is antisymmetric in the two last components: cmn..tp . whenever A = A . Let A be an arbitrary symmetric secondorder tensor. Demonstrate that C ·· A = 0.55 Let Q be the set of all antisymmetric secondorder tensors.54 Let F = F(A) be a linear function from the set of secondorder symmetric tensors A to the same set of symmetric secondorder tensors (F(A) = F(A)T ). On Q the inner product A ·· BT of the whole space takes the form A ·· BT = −A ·· B = −AT ·· BT . 3.
there is an aesthetic reason: in noncoordinate form many statements of mathematical physics look much less cumbersome than their counterparts stated in terms of coordinates. it is often convenient to introduce curvilinear coordinates in the body. In this frame. cylindrical. The position of a point in space is characterized by three numbers called coordinates of the point. Coordinate systems in common use include the Cartesian. Such equations are usually derived using a Cartesian space frame. and spherical systems. However.Chapter 4 Tensor Fields 4. most students would prefer a short and beautiful statement to a nightmarish formula taking half a page. Hence inside a spatial object there arises what we call a vector or tensor ﬁeld. and componentwise diﬀerentiation is not enough to characterize the change of the vector function. Another reason for introducing these tools is the objectivity of the laws of nature: we must be able to formulate frame independent statements of these laws.1 Vector Fields The state of a point (more precisely of an inﬁnitely small volume) within a natural object is frequently characterized by a vector or a tensor. the operations of calculus closely parallel those of one dimensional analysis: the diﬀerentiation of a vector function proceeds on a component by component basis. Finally. Thus we need to develop the apparatus of calculus for vector and tensor functions when the frame in the object is changeable. As a rule these ﬁelds are governed by simultaneous partial diﬀerential equations. It is said that beauty governs the world. for example. The ﬁrst of these diﬀers from the latter two in an important respect: the frame vectors of a Cartesian 85 . although this is not absolutely true. In this way we get a frame that changes from point to point. in terms of which the problem formulation is simpler.
while those for the curvilinear systems change from point to point. In this case the local onetoone correspondence is provided by the requirement that the Jacobian ∂xi = 0. We suppose the functions xi = xi (q 1 .2) Fixing some origin O. the correspondence (4. See Fig. q 2 . In a general coordinate system the position of a point in space is determined uniquely by three numbers q 1 .86 Tensor Analysis with Applications in Mechanics system are unique. q 2 . 3). q 2 . q 3 .1) Except for some set of singular points in space. q 3 ) .1. q 2 . When a point moves along a coordinate line. we characterize the position of a point P (q 1 .1) is one to one. x3 of a point can be determined through the general curvilinear coordinates by relations of the form xi = xi (q 1 . q 3 ) by a vector r connecting the points O and P : r = r(q 1 . (4. x2 . q 2 . 4. 2. q 2 . These are referred to as curvilinear coordinates if the frame is not Cartesian.1 Curvilinear coordinates. The Cartesian coordinates x1 . q 3 ). Fig. 4. ∂q j (4. the end of its position vector moves along this line and the diﬀerence vector ∆r = r(q 1 + ∆q 1 . q 2 . If we ﬁx two of the coordinates and change the third. along the q 1 line for instance. we get a line in space called a coordinate line. q 3 ) (i = 1. q 3 ) − r(q 1 . q 3 ) are smooth (continuously diﬀerentiable).
g ij = ri · rj . we obtain from this the same expression for the Jacobian (4. q 3 ) is not singular then the vectors ri are noncoplanar and therefore constitute a frame triad. 4. The smaller the value of ∆q 1 . respectively.2). The mixed product r1 · (r2 × r3 ) is the volume of the frame parallelepiped. r3 = . q 2 . q 2 . q 3 ) = r1 ∂q 1 tangent to the q 1 coordinate line.2 Generation of a local tangent vector to a coordinate line. In particular.Tensor Fields 87 Fig. 4. we now deal with frame vectors that change from point to point. The limit as ∆q 1 → 0 of the ratio ∆r/∆q 1 is a vector ∂r(q 1 . at each point we can repeat our prior reasoning. q 3 ) we can introduce two other vectors ∂r(q 1 . However. let us introduce the reciprocal basis ri by the relation i ri · rj = δj . By the previous chapter we deﬁne the metric coeﬃcients gij = ri · rj . the closer ∆r is to being tangent to the q 1 line. Let us denote it ∂xi √ . q 2 . renumbering the coordinates if necessary. is directed along the chord (Fig. At the same point (q 1 . q 2 . If the coordinate point (q 1 . g = r1 · (r2 × r3 ) = ∂q j Unlike the previous chapters in which all frame vectors were constant.2). . j j gi = ri · rj = δi . r2 = ∂q 2 ∂q 3 tangent to the q 2 and q 3 coordinate lines. q 2 . q 3 ) . q 3 ) ∂r(q 1 .
88 Tensor Analysis with Applications in Mechanics which are the components of the metric tensor E at each point in space. dt dt dt Indeed. Let e1 (t) and e2 (t) be continuously diﬀerentiable at some ﬁnite t. Note that the rules for diﬀerentiating vector functions are quite similar to those for differentiating ordinary functions. 2 there exist ei (t + ∆t) − ei (t) dei (t) = lim . ∆t→0 dt ∆t It is easily seen that d(e1 (t) + e2 (t)) de1 (t) de2 (t) = + . e1 (t + ∆t) − e1 (t) + e2 (t + ∆t) − e2 (t) d(e1 (t) + e2 (t)) = lim ∆t→0 dt ∆t e1 (t + ∆t) − e1 (t) e2 (t + ∆t) − e2 (t) + lim = lim ∆t→0 ∆t→0 ∆t ∆t de1 (t) de2 (t) + . q 3 ). q 2 . which means that for i = 1. dt dt dt The reader can adapt the proof for the scalar case almost word for word. the product of a scalar function f (t) and a vector function e(t) is diﬀerentiated by a rule similar to the formula for diﬀerentiating a product of scalar functions: d(f (t)e(t)) df (t) de(t) = e(t) + f (t) . dt dt Finally. dt d (e1 (t) × e2 (t)) = e1 (t) × e2 (t) + e1 (t) × e2 (t). = dt dt Similarly. Above we diﬀerentiated a vector function r(q 1 . The rules for partial diﬀerentiation of vector functions in several variables look quite similar and we leave their formulation to the reader. Representing vectors in Cartesian coordinates it is easy to see the validity of the following formulas: d (e1 (t) · e2 (t)) = e1 (t) · e2 (t) + e1 (t) · e2 (t). dt . For brevity we consider the case of ﬁelds depending on one variable t. for a constant c we have d(ce1 (t)) de1 (t) =c .
4. e3 (t)]. e3 (t)] + [e1 (t).3 Cylindrical coordinate system. e2 (t). 4. where ρ is the radial distance from the zaxis and φ is the azimuthal angle (Fig. For the mixed product [e1 (t). Cylindrical coordinates In the cylindrical coordinate system we have (q 1 .Tensor Fields 89 Exercise 4. e2 (t). e2 (t).1. (a) Diﬀerentiate e1 (t) · e2 (t) and e1 (t) × e2 (t) with respect to t if e1 (t) = i1 e−t + i2 . e2 (t) = −i1 sin2 t + i2 e−t . q 3 ) = (ρ. Using the expression ˆ ˆ r = xx + yy + ˆz z . (b) Show that [e(t) × e (t)] = e(t) × e (t) for any diﬀerentiable vector function e(t). Fig. e3 (t)] dt = [e1 (t). e3 (t)] + [e1 (t). e2 (t). q 2 .3). Exercise 4.2. φ. e3 (t)] = (e1 (t) × e2 (t)) · e3 (t) show that d [e1 (t). z). e2 (t).
φ. φ. For instance. These facts may be used to construct the reciprocal basis as r1 = r1 . It is easy to verify that the frame vectors r1 . By direct computation we also ﬁnd that r1  = 1. we have ˆ ˆ r(ρ. z) = xρ cos φ + yρ sin φ + ˆz. r3  = 1. r2 . r2  = ρ. 0 0 1 0 0 1 . φ. ∂ρ y = ρ sin φ. and are ⎛ ⎛ ⎞ ⎞ 1 0 0 1 0 0 (gij ) = ⎝ 0 ρ2 0 ⎠ . z) ˆ = −ˆ ρ sin φ + yρ cos φ. r3 are mutually perpendicular. z) ˆ ˆ = x cos φ + y sin φ. z) ˆ = z. r3 = r3 . (g ij ) = ⎝ 0 1/ρ2 0 ⎠ . x ∂φ r3 = ∂r(ρ. z = z. ∂z ∂r(ρ. The various metric coeﬃcients for the cylindrical frame are easily computed. z Then r1 = Similarly we compute r2 = Then √ g = r1 · (r2 × r3 ) ˆ ˆ ˆ = (ˆ cos φ + y sin φ) · [(−ˆ ρ sin φ + yρ cos φ) × z] x x ˆ ˆ = (ˆ cos φ + y sin φ) · (ˆ ρ sin φ + xρ cos φ) x y = ρ(cos2 φ + sin2 φ) = ρ. ˆ ˆ r1 · r2 = (ˆ cos φ + y sin φ) · (−ˆ ρ sin φ + yρ cos φ) x x = −ρ cos φ sin φ + ρ cos φ sin φ = 0. r2 = r2 /ρ2 . ∂r(ρ.90 Tensor Analysis with Applications in Mechanics for the position vector in rectangular coordinates and the coordinate transformation formulas x = ρ cos φ. φ.
They are (v 1 . N/m2 for a stress).. in spherical coordinates. vφ . Exercise 4. r has the dimension of length whereas θ and φ are dimensionless. vφ /ρ. φ) is given by d2 r d2 ρ ˆ =ρ −ρ 2 dt dt2 dφ dt 2 ˆ +φ ρ d2 φ dρ dφ +2 dt2 dt dt . φ). it is conventional in applications to write ˆ ˆ ˆ v = ρvρ + φvφ + zvz . But in applications many quantities have physical dimensions (e.3. and z are introduced in cylindrical coordinates. . and the frame vectors become dimensionless. These are unit vectors along the directions of r1 . whereas the frame vector for r is dimensionless.g. we can identify the contravariant and covariant components of v from our previous expressions. When we introduce coordinates in space some can also have dimensions.Tensor Fields 91 ˆ ˆ ˆ Commonly the unit basis vectors ρ. the physical components are the projections of the vector on the directions of the corresponding frame vectors. We see that expression of v in terms of the basis (ˆ . for instance. q 3 ) = (r. Show that the acceleration of a particle in plane polar coordinates (ρ. and r3 . φ. vz ) that are neither contravariant nor covariant. vφ ρ. v3 ) = (vρ . vφ . Given any vector v at a point (ρ. When one uses physical components of vectors and tensors they take the dimension of the corresponding tensor in full. (v1 . and vz are known as the physical components of v. In this way some components of the metric tensor may have dimensions and. v 3 ) = (vρ . when involved in transformation formulas. z) in space. φ.1 ) But since we can also express v in the forms v = v 1 r1 + v 2 r2 + v 3 r3 = v1 r1 + v2 r2 + v3 r3 . Spherical coordinates In the spherical coordinate system we have (q 1 . v2 . (For a vector. vz ). φ. The quantities vρ . respectively. 1 Mathematicians often prefer to deal with dimensionless quantities. vz ). the frame vectors corresponding to θ and φ must have the length dimension. ˆ) yields components ρ ˆ z (vρ . r2 . θ. v 2 . q 2 . at the point of interest. Since r has the length dimension. may thus introduce not only numerical changes but also dimensional changes in the terms so that it becomes possible to add such quantities.
give ˆ ˆ ˆ r(r. 4. φ) ˆ r3 = = −ˆ r sin θ sin φ + yr sin θ cos φ. ∂θ ∂r(r. θ. z = r cos θ In this case √ sin θ cos φ sin θ sin φ cos θ g = r cos θ cos φ r cos θ sin φ −r sin θ = r2 sin θ. Then r1 = ∂r(r. and θ is the polar angle (Fig. x ∂φ y = r sin θ sin φ. r3  = r sin θ.4). 4. φ) = xr sin θ cos φ + yr sin θ sin φ + zr cos θ. and the frame vectors are again mutually orthogonal. φ) ˆ ˆ ˆ = x sin θ cos φ + y sin θ sin φ + z cos θ. θ. φ is the azimuthal angle. φ) ˆ ˆ ˆ r2 = = xr cos θ cos φ + yr cos θ sin φ − zr sin θ.4 Spherical coordinate system. where r is the radial distance from the origin.92 Tensor Analysis with Applications in Mechanics Fig. r3 = r3 /r2 sin2 θ . −r sin θ sin φ r sin θ cos φ 0 In this system we ﬁnd that r1  = 1. The coordinate transformation formulas x = r sin θ cos φ. r2 = r2 /r2 . ∂r ∂r(r. r2  = r. Hence we have r1 = r1 . θ. θ.
3). We see that both the spherical and cylindrical frames are orthogonal. In § 2.4. j ˜ ˜ fi = Aj fj . q 2 . but have reciprocal lengths so that ri ri  = 1 for each i. vθ . The f i are still termed contravariant components of f . . v 2 . ir Aj = i ∂ qj ˜ . and so on. j ˜ fi = Aj fj . i for the transformation laws pertaining to the components of a vector ﬁeld f . r i ∂q j ˜ Aj = . q 2 . q 3 ). v2 . and φ. In this case the reciprocal basis vectors have the same directions as the vectors of the main basis.4 we obtained the transformation laws that apply to the components of a vector under a change of frame.Tensor Fields 93 for the vectors of the reciprocal basis. For the inverse transformation ˜ ˜i = Aj rj . Because ri = we can write ri = Aj ˜j . and we can write immediately ˜ f i = Ai f j . so q j = q j (q 1 . The transformation laws i for the components of higherorder tensor ﬁelds are written similarly. These laws are a consequence of ˜ the change of the old curvilinear coordinates q k to other coordinates q j . (gij ) = ⎝ 0 r2 (g ij ) = ⎝ 0 1/r2 0 0 2 2 2 2 0 0 r sin θ 0 0 1/r sin θ Exercise 4. (v1 . and (vr .4) The results of § 2.3) ˜ ∂ qj ˜ ∂r ∂r ∂ q j = j i = ˜j i r i ∂q ∂ q ∂q ˜ ∂q to describe the change of frame. ˜k . while the fi are still termed covariant components — the main diﬀerence is that now Aj i ˜ and Aj can change from point to point in space. The metric coeﬃcients are ⎛ ⎛ ⎞ ⎞ 1 0 0 1 0 0 ⎠. We now extend the transformation laws to the case of a vector ﬁeld in general coordinates. Give expressions relating the various sets of components r ˆ (v 1 . 2. but with the tilde above: ˜k .4 now apply. i ∂ qi ˜ ˜ ˜ f i = Ai f j . vφ ) of a vector v. ∂q i (4. we denote all the quantities in the same manner as ˜ r r for the coordinates q k . i (4. k = 1. q 3 ) q ˜ ˜ (j. In the coordinates q j . θ. ˜ ˜ q k = q k (˜1 . v 3 ). The unit basis vectors of spherical coordinates are denoted ˆ ˆ. ⎠. v3 ).
(ds)2 = (dr)2 + r2 (dθ)2 + r2 sin2 θ(dφ)2 .2 Diﬀerentials and the Nabla Operator Let us consider the inﬁnitesimal vector extending from point (q 1 . 2) and the reciprocal basis vectors ri in the (u. 4. Exercise 4.94 Tensor Analysis with Applications in Mechanics Exercise 4.5. q 2 + dq 2 .6. (4. In cylindrical and spherical frames we have (ds)2 = (dρ)2 + ρ2 (dφ)2 + (dz)2 . respectively.6) On the right we have a quadratic form with respect to the variables dq i . Use these to calculate the metric coeﬃcients. denoted by dr: dr = Hence we can deﬁne dq i as dq i = ri · dr = dr · ri . β are constants. However. v) system. Note that we only introduce spatial coordinates having superscripts. q 3 + dq 3 ). Find (ds)2 for the oblique system of Exercise 4. . ∂q i (Here there is no summation over i. y) by the transformation equations x = u cos α + v cos β. v) in the plane is related to a set of Cartesian coordinates (x. q 2 . (a) Sketch a few u and v coordinate curves superimposed on the xyplane.5. y = u sin α + v sin β. In a Cartesian frame (4. (b) Find the basis vectors ri (i = 1. where α.6) takes the familiar form (ds)2 = (dx1 )2 + (dx2 )2 + (dx3 )2 . A set of oblique rectilinear coordinates (u. In the oblique system ﬁnd the covariant components of z in terms of the contravariant components of z. the coeﬃcients of this quadratic form are the covariant components of the metric tensor.5) ∂r i dq = ri dq i . for Cartesian frames it is common to see subscripts used exclusively. (4. (c) Let z be a given vector.) The length of this inﬁnitesimal vector is deﬁned by (ds)2 = dr · dr = ri dq i · rj dq j = gij dq i dq j . q 3 ) to point (q 1 + dq 1 .
Its diﬀerential is df (q 1 .5) we can write df = ri ∂f · dr. Using (4. q 2 . q 3 ) i dq . The nabla operator is often referred to as the gradient operator. ∂q i The ﬁrst multiplier on the right is a vector ri ∂f /∂q i . q 3 ) be a scalar diﬀerentiable function of the variables q i .7. q 2 . (a) Show that the unit normal to the surface ϕ(q 1 . whose action on a function f is as given above: ∇f = ri (we repeat that there is summation over i here). q 2 . q 2 . q 2 . Let us introduce a symbolic vector ∇ = ri ∂ ∂q i ∂f ∂q i called the nabla operator. q 2 = c2 is given by cos θ12 = g 12 g 11 g 22 . (b) Show that the angle at a point of intersection of the surfaces ψ(q 1 . q 3 ) = ∂f (q 1 . is given by cos θ = ∂ϕ ∂ψ ∂q i ∂q j g mn ∂ϕ ∂ϕ g rt ∂ψ ∂ψ ∂q m ∂q n ∂q r ∂q t g ij rj . q 3 ) = c = constant is given by n= g ij ∂ϕ ∂q i g mn ∂ϕ ∂ϕ ∂q m ∂q n ϕ(q 1 .Tensor Fields 95 Let f (q 1 . . ∂q i Here and in similar situations the i in the denominator stands in the lower position and so we must sum over i. Exercise 4. q 3 ) = c2 . q 2 . (c) Show that the angle between the coordinate surfaces q 1 = c1 . q 3 ) = c1 .
+θ +φ ∇f = ˆ r ∂r r ∂θ r sin θ ∂φ ˆ ∇f = ρ respectively. q 2 . q 3 ) = ∂f (q 1 . is clearly a tensor of order two. The quantity ∇f . q 3 ) = (dr · ri ) ∂q i 1 2 3 ∂f (q . For a pair of vectors.96 Tensor Analysis with Applications in Mechanics (d) Derive the condition for orthogonality of surfaces g ij ∂ϕ ∂ψ = 0. +φ z ∂ρ ρ ∂φ ∂z ∂f ˆ 1 ∂f ˆ 1 ∂f . We can apply these to the pair consisting . q 3 ): df (q 1 .8. dr An application of the gradient to a tensor brings a new tensor whose order is one higher than that of the original tensor. q . q 2 . With the aid of transposition we can present df in the form df = ∇f T · dr. q 3 ) i (r · dr) ∂q i ∂f (q 1 . q 2 . known as the gradient of f . Sometimes ∇f T is called the gradient of f . it is also called the derivative of f in the direction of r and denoted as df = ∇f T . = dr · ri ∂q i ∂f (q 1 . Show that the gradient operation in the cylindrical and spherical frames is given by the formulas ∂f ˆ 1 ∂f + ˆ ∂f .5) again. q 2 . What are the expressions for these when using the components connected with the triads ri and ri ? Let us ﬁnd a formula for the diﬀerential of a vector function f (q 1 . we introduced two types of multiplication: the dot and cross product operations. q ) . q 2 . q 2 . Then df (q 1 . ∂q i Thus we can represent this as df = dr · ∇f . q 3 ) i dq . q 3 ) = We use (4. ∂q i ∂q j Exercise 4.
q 2 . ∇ · E. ∂q i Exercise 4. It can be shown that Ω= 1 ∇f T − ∇f . ∂q i ∂ r = ri · ri = 3. This tensor is antisymmetric (the reader should verify this). In this way we get two operations: the divergence div f = ∇ · f = ri · and the rotation rot f = ∇ × f = ri × The vector ω= 1 rot f 2 ∂f . q 3 ). ∂q i ∇ · r = ri · and ∇ × r = ri × ∂ r = ri × ri = 0. ∇E. Calculate ∇E. ∂q i is called the curl of f . In terms of the curl of f we can introduce a tensor Ω as Ω = E × ω.9. and ∇ × E. ∇ · E. 2 Similarly we can introduce the divergence and rotation operations for a tensor A of any order: ∇ · A = ri · ∂ A. ∂q i ∇ × A = ri × ∂ A. ∂q i Let us see what happens when we apply the nabla operator to the radius vector: ∇r = ri ∂ r = ri ri = E.Tensor Fields 97 of the nabla operator (which is regarded as a formal vector) and a vector function f = f (q 1 .7) ∂f . and is called the tensor of spin. ∇ × E. ∂q i (4. .
let f be a vector. ∇(f × g) = (∇f ) × g − (∇g) × f . x2 .10. x3 ) = f k (x1 . q 3 )ri . Exercise 4. ∇ · (fg) = (∇ · f )g + f ∇ · g. q 2 .8) Let us consider how to diﬀerentiate a vector function f that is written out in curvilinear coordinates: f (q 1 .11. ∇ × (f × g) = g · ∇f − g∇ · f − f · ∇g + f ∇ · g. But the components of f are not constant since the . 4. Show that the following identities hold (f and g are any vectors): ∇(f · g) = (∇f ) · g + f · ∇gT = (∇f ) · g + (∇g) · f . ∂xi ∂xi (4. In this case the frame vectors ri depend on q k as well. which means that simple diﬀerentiation of the components of f does not result in the needed formula. q 3 )/∂q i = 0. ∇(f Q) = (∇f )Q + f ∇Q. q 2 . q 3 ) = f i (q 1 . and let Q be a tensor of order two. x2 . x3 ) = ik . Let f and g be functions. q 2 . ∇(f f ) = (∇f )f + f ∇f . q 3 ) = c. By the general deﬁnition of the partial derivative we have ∂f (q 1 . x2 .98 Tensor Analysis with Applications in Mechanics Exercise 4. with the tacit understanding that the formulas necessary to do this resemble those from ordinary calculus. x3 )ik with respect to a Cartesian variable xi it is enough to diﬀerentiate each component of f with respect to this variable: ∂ ∂f k (x1 . ∇ · (f × g) = g · (∇ × f ) − f · (∇ × g). Moreover.3 Diﬀerentiation of a Vector Function We have diﬀerentiated vector functions with respect to q i . x3 ) f (x1 . Verify the following identities: ∇(f g) = g∇f + f ∇g. q 2 . To understand this let us consider a constant function f (q 1 . x2 . the reader certainly knows that to diﬀerentiate a vector function f (x1 .
4. 2.8). (4. however in the general case we exploit only the fact that the partial derivative of a frame vector is a vector as well. ij ji (4. The derivative of the product of a simple function and a vector function is taken by the product rule: ∂ ∂ f (q 1 . 3). j ∂q i ∂q ∂q ∂q ∂q (4. q 2 . q 2 . hence it can be expanded in the same frame ri (i = 1.12) .4 Derivatives of the Frame Vectors We would like to ﬁnd the value of the derivative ∂ ∂ r = j j i ∂q ∂q ∂ r ∂q i = ∂2r ∂2r ∂ = i j = i rj . q 3 ) ∂ri = ri + f i (q 1 . q 3 ) = k f i (q 1 . ∂q k ∂q So the derivative of a vector function written in component form consists of two terms. and the other contains the derivatives of the frame vectors. q 2 . Let us denote the coeﬃcients of the expansion by Γk : ij ∂ ri = Γk rk . q 3 )ri ∂q k ∂q ∂f i (q 1 .10) The quantities Γk are called Christoﬀel coeﬃcients of the second kind and ij are often denoted by Γk = ij k ij . The ﬁrst is the same as for the derivative in the Cartesian frame (4. ij ∂q j (4.11) By (4.Tensor Fields 99 rk are variable. q 2 . Thus we need to ﬁnd the latter.9) Of course if we have the expression for the Cartesian components of the frame vector we can compute these derivatives (and will do so in the exercises).9) there is symmetry in the subscripts of the Christoﬀel symbols: Γ k = Γk . q 3 ) k . hence the derivatives of the components of this function are not zero.
j i ∂q ∂q ∂q Relation (4. So the Christoﬀel symbols are not the components of a tensor. q 2 .13) This is called covariant diﬀerentiation. q 2 . and are denoted by ∇k f i = ∂f i + Γi f t . The coeﬃcients of ri are called the covariant derivatives of the contravariant components of f . q 3 ) = ri + Γj f i (q 1 . 4. q 3 )ri kt ∂q k ∂f i = + Γi f t ri . kt ∂q k Let us discuss some properties of the Christoﬀel coeﬃcients. q 2 .5 Christoﬀel Coeﬃcients and their Properties In a Cartesian frame the Christoﬀel symbols are zero. despite their notation. q 3 ) = ri + Γi f t (q 1 . q 2 .11). q 3 ) f (q 1 . kt ∂q k (4.10) can be written as rij = Γk rk ij from which it follows that rij · rt = Γk rk · rt = Γk gkt . q 2 . so it is impossible to obtain them for another frame by the usual transformation rules for tensor components.14) . and these rates do not appear in the transformation rules for tensors. q 3 )rj ki k ∂q ∂q k ∂f i (q 1 .100 Tensor Analysis with Applications in Mechanics Now we can write out the formula for the derivative of a vector function in full: ∂ ∂f i (q 1 . It is important to have formulas for computing the Christoﬀel symbols. Let us introduce the notation rij = ∂ ∂2r r = i j. This is easy to understand: the Christoﬀel symbols depend not only on the frame vectors themselves but on their rates of change from point to point. ij ij (4. This is why many authors prefer the notation shown on the right side of equation (4.
ij Using these we can obtain ∂rj = −Γj rt .21) (4.Tensor Fields 101 The lefthand side of this can be expressed in terms of the components of the metric tensor. jt ∂q t ∂q ∂q ∂q (4. It is clear that they are symmetric in the ﬁrst two subscripts: Γijk = Γjik . Thus we have Γk gkt = Γijt ij and it follows that Γk = g kt Γijt . They are denoted frequently as [ij.17) and dividing by 2: rij · rt = 1 2 ∂git ∂gtj ∂gji + − ∂q j ∂q i ∂q t = Γijt . it ∂q i Indeed 0= so ∂ri i · rj = −ri · Γk rk = −Γk δk = −Γi jt jt jt ∂q t and we have ∂ri = −Γi rj .15) and (4. ∂q t ∂q ∂ ∂ gtj = i rt · rj = rti · rj + rt · rji . (4.16) from the sum of (4.19) . (4.15) ∂q j ∂q Similarly ∂ ∂ gji = t rj · ri = rjt · ri + rj · rit . jt ∂q t ∂ri ∂ ∂ri ∂ i δj = t ri · rj = t · rj + ri · rjt = t · rj + ri · Γk rk . k] = Γijk .16) (4. Indeed ∂ ∂ git = j ri · rt = rij · rt + ri · rtj .20) (4.18) The quantities Γijk are called Christoﬀel coeﬃcients of the ﬁrst kind. ∂q i ∂q (4.17) Now we obtain rij · rt by subtracting (4.
∂q The presence of the ﬁrst term on the right means that the Christoﬀel coeﬃcient is not a tensor of order three.4).14) applied ˜ in the new system gives ˜ r Γk ˜k = ˜ij . Exercise 4.22) and expansion by the product rule gives ˜j ˜ij = Am r ˜ ∂ An i ˜i ∂rn rn + An m ∂q m ∂q ˜n ˜ ∂ Ai ˜ ˜ = Am m rn + Am An Γp rp j j i nm ∂q ˜n ˜ ∂ Ai ˜ ˜ = Am m Ak ˜k + Am An Γp Ak ˜k . r ij But ˜ij = r ∂ ∂q m ∂ ˜j ∂ ˜i ˜i = ˜ = Am m (An rn ). j nr j i nm p r ∂q Comparison with (4. This conﬁrms our earlier statement of this fact. q ). Show that the nonzero Christoﬀel coeﬃcients of the ﬁrst kind for spherical coordinates are Γ221 = −r.3)–(4. Γ233 = Γ323 = r2 sin θ cos θ. which was based on diﬀerent reasoning. Suppose the new coordinates q i are ˜ ˜ determined by the old coordinates q i through relations of the form q i = i 1 2 3 q (q .102 Tensor Analysis with Applications in Mechanics It is also useful to have formulas for transformation of the Christoﬀel symbols under change of coordinates. r r j j ∂q m i ∂q ˜ ∂q ˜ ∂q (4. Γ122 = Γ212 = ρ. Γ332 = −r2 sin θ cos θ. . q . and refer back to equations (4.22) shows that ˜n ˜j ∂ Ai n ˜j ˜i p nm ˜ ij Γk = Am m Ak + Am An Ak Γp . Γ313 = Γ133 = r sin2 θ. Γ331 = −r sin θ. 2 Γ122 = Γ212 = r. Show that the only nonzero Christoﬀel coeﬃcients of the ﬁrst kind for cylindrical coordinates are Γ221 = −ρ.12. Now (4.
33 Γ3 = Γ3 = cot θ.14. 12 21 Γ2 = − sin θ cos θ. by construction. nonEuclidean spaces When we derive the length of the elementary vector dr we write (ds)2 = dr · dr = ri dq i · rj dq j = gij dq i dq j . if it can be described by a set of Cartesian coordinates xt (t = 1. A system of plane elliptic coordinates (u. This is one of the main properties of the metric tensor for a real coordinate space. that is. . then the line element (ds)2 can be written as (ds)2 = 3 t=1 (dxt )2 . y = c sinh u sin v. where c is a constant.23) With respect to the variables dq i and dq j we have. (4. 23 32 Exercise 4. 12 21 Show that the nonzero Christoﬀel coeﬃcients of the second kind for spherical coordinates are Γ1 = −r. 22 Γ1 = −r sin2 θ.Tensor Fields 103 Exercise 4.24) Given a set of admissible transformation equations xt = xt (q n ). Euclidean vs. If a space is Euclidean. 31 13 Γ2 = Γ2 = 1/r. Find the Christoﬀel coeﬃcients. 22 Γ2 = Γ2 = 1/ρ. Show that the only nonzero Christoﬀel coeﬃcients of the second kind for cylindrical coordinates are Γ1 = −ρ. v) is introduced according to the transformation formulas x = c cosh u cos v.13. Since dxt = we have (ds)2 = 3 t=1 ∂xt n dq ∂q n 2 ∂xt n dq ∂q n . 33 Γ3 = Γ3 = 1/r. 2. 3). the same line element can be expressed in terms of general coordinates q n . (4. a positive deﬁnite quadratic form.
It can be shown that Rpijk = · ∂Γp ∂Γp ij ik − − Γp Γ m − Γ p Γ m . Given the gij in the q n system. Suppose a space is originally described in terms of a set of general coordinates q n . 2.104 Tensor Analysis with Applications in Mechanics and comparison with (4. Such a description must include a set of metric coeﬃcients gij having (4.24)? Put still another way.25) provides us with six equations for the three unknown functions xt (q n ). (4. · (4.23) positive deﬁnite at each point. j = 1. Will it be possible to introduce a set of Cartesian coordinates xt that also describe this space? That is. so (4. are we guaranteed the existence of functions xt (q n ) such that in the resulting coordinates xt = xt (q n ) the line element takes the form (4. as such considerations are important in certain application areas.26) The tensor Rpijk is known as the Riemann–Christoﬀel tensor.25) for the metric coeﬃcients in the q n system.26) · represents six conditions on the gij that must hold for the space described by the gij to be Euclidean [Sokolnikoﬀ (1994)]. . Rpijk actually has only six independent components.23) gives 3 gij = t=1 ∂xt ∂xt ∂q i ∂q j (i. We now pose the following question. (4. It is possible to formulate these restrictions neatly in terms of a certain fourthorder tensor Rpijk as · the equality Rpijk = 0. This implies that some additional conditions must be fulﬁlled by the given metric coeﬃcients. 3). and its as· sociated tensor Rnijk = gnp Rpijk · is called the curvature tensor for the space. is the space Euclidean? Not necessarily. mj ik mk ij ∂q k ∂q j Moreover. Further details and more rigorous formulations can be found in some of the more comprehensive references. We have included this information only for the reader’s background.
Tensor Fields 105 4.13) ∂ ∂ f = k (f i ri ) = ∂q k ∂q ∂f i + Γi f t ri kt ∂q k and observed that the coeﬃcients of this expansion denoted by ∇k f i = ∂f i + Γi f t kt ∂q k (4.15. ki ∂q k (4. .27) are called the covariant derivatives of contravariant components of vector function f . Similarly. (4. ∂q k ∂q ∂q ∂q Using (4.21) we get ∂fi ∂ f = k ri − fi Γi rt = kt k ∂q ∂q ∂fi − Γj fj ri . ki ∂q k The coeﬃcients of this expansion are called covariant derivatives of the covariant components and are denoted by ∇k fi = ∂fi − Γj f j .6 Covariant Diﬀerentiation Let us return to the problem of diﬀerentiation of a vector function. Write out the covariant derivatives of the contravariant components of a vector f in plane polar coordinates.29) ∂q i Because ∇f = ri ∂ (fj rj ) = ri (∇i fj )rj = ri rj ∇i fj . ∂q i ∂ (f j rj ) = ri rj ∇i f j ∂q i we see that ∇i fj is a covariant component of ∇f . We obtained the formula for the derivative (4. Let us express the same derivatives in terms of the covariant components: ∂ ∂fi ∂ ∂ f = k (fi ri ) = k ri + fi k ri .28) In these notations we can write out the formula for diﬀerentiation in the form ∂f = rk ∇i fk = rj ∇i f j . Exercise 4. ∇f = ri shows that ∇i f j is a mixed component of ∇f .
ik jk ∂q k Changing dummy indices we get ∂ A= ∂q k ∂aij + Γi asj + Γj ais ri rj . kt ∂q k 4.7 Covariant Derivative of a SecondOrder Tensor Let us ﬁnd a partial derivative of a tensor A of order two: ∂ ∂ A = k