Elementary Vector and Tensor Analysis

UNM REPORT
September 10, 2002 11:22 am
Elementary vector and tensor analysis

presented in a framework generalizable to higher-order applications in
material modeling
R. M. Brannon
University of New Mexico, Albuquerque
Copyright is reserved.
Individual copies may be made for personal use.

No part of this document may be reproduced for profit.
Contact author at rmbrann@sandia.gov
NOTE: When using Adobes acrobat reader to view this

document, the page numbers in acrobat will not coincide
with the page numbers shown at the bottom of each page
of this document.
Note to draft readers: The most useful textbooks are

the ones with fantastic indexes. The repor ts index is
rather new and still under construction.
It would really help if you all could send me a note
whenever you discover that an impor tant entr y is missing from this index. Ill be sure to add it.
This work is a community effor t. Lets tr y to make this
document helpful to others.
ELEMENTARY VECTOR AND TENSOR ANALYSIS
FT
presented in a framework generalizable to higherorder applications in material modeling

Rebecca M. Brannon
University
of New Mexico Adjunct professor

rmbrann@sandia.gov
Abstract
RA
Elementary vector and tensor analysis concepts are reviewed using a notation
that proves useful for higher-order tensor analysis of anisotropic media.
Acknowledgments
RA
FT
To be added. Stay tuned...
ii
Contents
ii
xiii
1
3
7
7
8
9
9
10
10
10
10
11
11
13
14
17
17
18
19
19
20
22
23
24
24
24
25
25
26
27
29
29
30
31
31
32
33
34
34
35
35
36
RA
FT
Acknowledgments ...............................................................................................
Preface.....................................................................................................................
Introduction ............................................................................................................
Terminology from functional analysis .............................................................
Matrix Analysis.......................................................................................................
Definition of a matrix...........................................................................................
The matrix product...............................................................................................
The transpose of a matrix.....................................................................................
The inner product of two column matrices ..........................................................
The outer product of two column matrices. .........................................................
The trace of a square matrix.................................................................................
The Kronecker delta.............................................................................................
The identity matrix...............................................................................................
The 3D permutation symbol ................................................................................
The - (E-delta) identity.....................................................................................
The - (E-delta) identity with multiple summed indices ...................................
Determinant of a square matrix............................................................................
Principal sub-matrices and principal minors........................................................
Matrix invariants..................................................................................................
Positive definite ...................................................................................................
The cofactor-determinate connection ..................................................................
Inverse..................................................................................................................
Eigenvalues and eigenvectors ..............................................................................
Similarity transformations .............................................................................
Finding eigenvectors by using the adjugate.........................................................
Vector/tensor notation........................................................................................
Ordinary engineering vectors ...........................................................................
Engineering laboratory base vectors ................................................................
Other choices for the base vectors .......................................................................
Basis expansion of a vector..................................................................................
Summation convention details........................................................................
Dont forget to remember what repeated indices really mean ......................
Indicial notation in derivatives ......................................................................
BEWARE: avoid implicit sums as independent variables..............................
Reading index STRUCTURE, not index SYMBOLS .........................................
Aesthetic (courteous) indexing ............................................................................
Suspending the summation convention ...............................................................
Combining indicial equations ..............................................................................
The index-changing property of the Kronecker delta ..........................................
Summing the Kronecker delta itself ....................................................................
The under-tilde notation ...................................................................................
Simple vector operations and properties ......................................................
Dot product between two vectors ........................................................................
Dot product between orthonormal base vectors...................................................
iii
36
37
37
38
38
39
39
41
41
42
43
44
44
46
47
47
48
49
49
50
RA
FT
Finding the i-th component of a vector................................................................

Even and odd vector functions.............................................................................
Homogeneous functions.......................................................................................
Vector orientation and sense................................................................................
Simple scalar components....................................................................................
Cross product .......................................................................................................
Cross product between orthonormal base vectors ...............................................
Triple scalar product ............................................................................................
Triple scalar product between orthonormal RIGHT-HANDED base vectors .....
Axial vectors ........................................................................................................
Glide plane expressions .................................................................................
Projections ..............................................................................................................
Orthogonal linear projections ..............................................................................
Rank-1 orthogonal projections.............................................................................
Rank-2 orthogonal projections.............................................................................
Basis interpretation of orthogonal projections.....................................................
Rank-2 oblique linear projection .........................................................................
Rank-1 oblique linear projection .........................................................................
Complementary projectors...................................................................................
Normalized versions of the projectors .................................................................
Expressing a vector as a linear combination of three arbitrary (not necessarily
orthonormal) vectors......................................................................................
Generalized projections .......................................................................................
Linear projections ................................................................................................
Nonlinear projections...........................................................................................
Self-adjoint projections........................................................................................
The projection theorem ........................................................................................
Tensors......................................................................................................................
Linear operators (transformations) ......................................................................
Dyads and dyadic multiplication .........................................................................
Simpler no-symbol dyadic notation .................................................................
The matrix associated with a dyad.......................................................................
The sum of dyads .................................................................................................
A sum of two or three dyads is NOT (generally) reducible ...............................
Scalar multiplication of a dyad ............................................................................
The sum of four or more dyads is reducible!.......................................................
The dyad definition of a second-order tensor ......................................................
Expansion of a second-order tensor in terms of basis dyads ...............................
Tensor operations..................................................................................................
Dotting a tensor from the right by a vector..........................................................
The transpose of a tensor .....................................................................................
Dotting a tensor from the left by a vector ............................................................
Dotting a tensor by vectors from both sides ........................................................
Extracting a particular tensor component ............................................................
Dotting a tensor into a tensor (tensor composition).............................................
Tensor analysis primitives ...................................................................................
51
53
53
53
54
55
57
58
61
62
63
64
64
64
65
66
66
69
69
69
70
71
71
71
73
iv
Copyright is reserved. Individual copies may be made for personal use. No part of this document may be reproduced for profit.
73
76
76
77
78
79
80
80
82
82
83
83
84
84
85
85
86
87
89
89
89
91
93
94
95
96
97
97
98
98
99
99
101
101
102
102
104
104
108
108
111
111
112
114
116
116
RA
FT
Three kinds of vector and tensor notation............................................................

There exists a unique tensor for each linear function ..........................................
Finding the tensor associated with a linear function............................................
The identity tensor ...............................................................................................
Tensor associated with composition of two linear transformations.....................
The power of heuristically consistent notation ....................................................
The inverse of a tensor.........................................................................................
The COFACTOR tensor ......................................................................................
Cofactor tensor associated with a vector .............................................................
Cramers rule for the inverse ...............................................................................
Inverse of a rank-1 modification..........................................................................
Derivative of a determinant .................................................................................
Projectors in tensor notation ..............................................................................
Linear orthogonal projectors expressed in terms of dyads ..................................
Finding a projection to a desired target space......................................................
Properties of complementary projection tensors..................................................
Self-adjoint (orthogonal) projectors.....................................................................
Generalized complementary projectors ...............................................................
More Tensor primitives.........................................................................................
Deviatoric tensors ................................................................................................
Orthogonal (unitary) tensors ................................................................................
Tensor associated with the cross product.............................................................
Physical application of axial vectors....................................................................
Symmetric and skew-symmetric tensors .............................................................
Positive definite tensors .......................................................................................
Faster way to check for positive definiteness ................................................
Negative definite and semi-definite tensors.........................................................
Isotropic and deviatoric tensors ...........................................................................
Tensor operations..................................................................................................
Second-order tensor inner product.......................................................................
Fourth-order tensor inner product........................................................................
Higher-order tensor inner product .......................................................................
The magnitude of a tensor or a vector .................................................................
Useful inner product identities.............................................................................
Distinction between an Nth-order tensor and an Nth-rank tensor.........................
Fourth-order oblique tensor projections...............................................................
Coordinate/basis transformations ...................................................................
Change of basis (coordinate transformation).......................................................
Tensor invariance..................................................................................................
Whats the difference between a matrix and a tensor? ........................................
Example of a scalar rule that satisfies tensor invariance..................................
Example of a scalar rule that violates tensor invariance ..................................
Example of a 3x3 matrix that does not correspond to a tensor............................
The inertia TENSOR............................................................................................
Scalar invariants and spectral analysis..........................................................
Invariants of vectors or tensors ............................................................................
117
118
118
119
120
121
121
122
122
122
122
123
125
126
128
128
129
130
130
131
132
132
132
133
135
135
136
138
139
140
140
142
143
144
146
146
147
147
149
150
151
151
155
155
157
158
RA
FT
Primitive invariants..............................................................................................
Trace invariants....................................................................................................
Characteristic invariants.......................................................................................
Direct notation definitions of the characteristic invariants...........................
The cofactor in the triple scalar product .......................................................
Invariants of a sum of two tensors .......................................................................
CASE: invariants of the sum of a tensor plus a dyad ....................................
The Cayley-Hamilton theorem: ...........................................................................
CASE: Expressing the inverse in terms of powers and invariants.................
Inverse of the sum of a tensor plus a dyad...........................................................
Eigenvalue problems............................................................................................
Algebraic and geometric multiplicity of eigenvalues ..........................................
Diagonalizable tensors .........................................................................................
Eigenprojectors ....................................................................................................
Geometrical entities ............................................................................................
Equation of a plane ..............................................................................................
Equation of a line .................................................................................................
Equation for a sphere ...........................................................................................
Equation for an ellipsoid......................................................................................
Example..........................................................................................................
Equation for a cylinder with an ellipse-cross-section ..........................................
Equation for a right circular cylinder...................................................................
Equation of a general quadric (including hyperboloid) .......................................
Generalization of the quadratic formula and completing the square................
Polar decomposition............................................................................................
The Q-R decomposition.......................................................................................
The polar decomposition theorem: ......................................................................
The polar decomposition is a nonlinear projection operation .............................
The *FAST* way to do a polar decomposition in two dimensions.....................
Material symmetry................................................................................................
Isotropic second-order tensors in 3D space .........................................................
Isotropic second-order tensors in 2D space .........................................................
Isotropic fourth-order tensors ..............................................................................
Transverse isotropy..............................................................................................
Abstract vectors/tensor algebra......................................................................
Definition of an abstract vector............................................................................
Inner product spaces ............................................................................................
Continuous functions are vectors! .......................................................................
Tensors are vectors! .............................................................................................
Vector subspaces..................................................................................................
Subspaces and the projection theorem.................................................................
Abstract contraction and swap (exchange) operators ..........................................
The contraction tensor ...................................................................................
The swap tensor .............................................................................................
Vector/tensor calculus........................................................................................
ASIDE #1: total and partial derivative notation...................................
vi
160
160
163
164
166
167
169
171
175
RA
FT
ASIDE #2: Right and left gradient operations...............................................

The nabla or del gradient operator ......................................................
Okay, if the above relation does not hold, does anything LIKE IT hold?......
Derivatives in reduced dimension spaces ............................................................
A more applied example.................................................................................
Series expansion of a nonlinear vector function ..................................................
Closing remarks ....................................................................................................
REFERENCES ............................................................................................................
INDEX........................................................................................................................
vii
FT
RA
D
viii
Figures
FT
Finding components via projections. .................................................

38
Cross product .....................................................................................
39
Vector decomposition ........................................................................
45
(a) Rank-1 orthogonal projection, and (b) Rank-2 orthogonal projection. 47
Oblique projection. ............................................................................
48
Rank-1 oblique projection..................................................................
49
Projections of two vectors along a an obliquely oriented line. ..........
51
Three oblique projections. .................................................................
52
Oblique projection. ............................................................................
55
Visualization of the polar decomposition. ......................................... 137
RA
Figure 5.1.
Figure 5.2.
Figure 6.1.
Figure 6.2.
Figure 6.3.
Figure 6.4.
Figure 6.5.
Figure 6.6.
Figure 6.7.
Figure 17.1.
ix
FT
RA
D
RA
FT
Tables
xi
FT
RA
D
xii
DRAF

Preface
Rebe
cca B
ranno
Preface
Math and science journals often have extremely restrictive page limits,
making it virtually impossible to present a coherent development of complicated concepts by working upward from basic concepts. Furthermore, scholarly journals are intended for the presentation of new results, so detailed
explanations of known results are generally frowned upon (even if those
results are not well-known or well-understood). Consequently, only those
readers who are already well-versed in a subject have any hope of effectively
reading the literature to further expand their knowledge. While this situation
is good for experienced researchers and specialists in a particular field of
study, it can be a frustrating handicap for less experienced people or people
whose expertise lies elsewhere. This report serves these individuals by presenting several known theorems or mathematical techniques that are useful
for the analysis material behavior. Most of these theorems are scattered willynilly throughout the literature. Several rarely appear in elementary textbooks. Most of the results in this report can be found in advanced textbooks on
functional analysis, but these books tend to be overly generalized, so the application to specific problems is unclear. Advanced mathematics books also tend
to use notation that might be unfamiliar to the typical research engineer. This
report presents derivations of theorems only where they help clarify concepts.
The range of applicability of theorems is also omitted in certain situations. For
example, describing the applicability range of a Taylor series expansion
requires the use of complex variables, which is beyond the scope of this document. Likewise, unless otherwise stated, I will always implicitly presume that
functions are well-behaved enough to permit whatever operations I perform.
For example, the act of writing df dx will implicitly tell the reader that I am
assuming that f can be written as a function of x and (furthermore) this function is differentiable. In the sense that much of the usual (but distracting)
mathematical provisos are missing, I consider this document to be a work of
engineering despite the fact that it is concerned principally with mathematics.
While I hope this report will be useful to a broader audience of readers, my
personal motivation is to establish a single bibliographic reference to which I
can point from my more stilted and terse journal publications.
Rebecca Brannon,
rmbrann@sandia.gov
Sandia National Laboratories
September 10, 2002 11:51 am.
xiii
T
F
A
R
D
cca
Rebe
Brann

Preface
on
xiv
DRAF
September 10, 2002 2:37 pm

Introduction
Rebe
cca B
ranno
ELEMENTARY VECTOR AND TENSOR ANALYSIS

presented in a framework generalizable to higherorder applications in material modeling
The discussion of tensor calculus will be

expanded at a later date
(my schedule is really packed!).
Please report errors to rmbrann@me.unm.edu
Things should be described as simply as possible,

but no simpler.
A. Einstein
1. Introduction
This report reviews tensor algebra (and a bit of tensor calculus) using a
notation that proves very useful when extending these basic ideas to higher
dimensions. Tensor notation unfortunately remains non-standardized, so its
important to at least scan this report to become familiar with our definitions
and overall approach to the field if you wish to move along to our other (more
sophisticated and contemporary) applications in materials modeling.
Many of the definitions and theorems provided in this report are not rigorous instead they are presented in a more physical engineering manner.
More careful expositions on these topics can be found in elementary textbooks
on matrix, vector, and tensor analysis. One may need to look to a functional
analysis text to find an equally developed discussion of projections. The reader
is presumed to have been exposed to vector analysis and matrix analysis at
the rudimentary level that is ordinarily covered in undergraduate calculus
courses.
1
T
F
A
R
D
cca
Rebe
Brann

Introduction
on
We present the information in this report in order to have a single reference in which all the concepts are presented together using a notation that is
consistent with that used in more advanced follow-up work that we intend to
published separately. Some of our other work explains that many theorems in
higher-dimensional realms have perfect analogs with the ordinary concepts
from 3D. For example, this elementary report discusses how to obliquely
project a vector onto plane, and we demonstrate in later (more advanced) work
that the act of solving viscoplasticity models by a return mapping algorithm is
perfectly analogous to vector projection.
Throughout this report, we use the term ordinary to refer to the three
dimensional physical space in which everyday engineering problems occur.
The term abstract will be used later when extending ordinary concepts to
higher dimensional spaces, which is the principal goal of generalized tensor
analysis. Except where otherwise stated, the basis { e 1, e 2, e 3 } used for vectors
orthonormal

and tensors in this report will be assumed regular (i.e.,
and righthanded). Thus, all indicial formulas in this report use what Malvern [12] calls
rectangular Cartesian components. Readers interested in irregular bases can
find a discussion of curvilinear coordinates at http://www.me.unm.edu/
~rmbrann/gobag.html (however, that report presumes that the reader is
already familiar with the notation and basic identities that are covered in this
report).
2
DRAF

Terminology from functional analysis
Rebe
cca B
ranno
Change isnt painful, but resistance to

change is.
anonymous(?)
2. Terminology from functional analysis
Vector, tensor, and matrix analysis are subsets of a more general area of
study called functional analysis. One purpose of this report is to specialize
several overly-general results from functional analysis into forms that are the
most convenient for real world physics applications. Functional analysis
deals with operators and their properties. For our purposes, an operator may
be regarded as a function f ( x ) . If the argument of the function is a vector and
if the result of the function is also vector, then the function is usually called a
transformation because it transforms one vector to become a new vector.
In this report, any non-underlined quantity is just an ordinary number (or,
using more fancy jargon, scalar or field member). Quantities with single
straight underlines (e.g., x or y ) might represent scalars, vectors, tensors, or
other abstract objects. We follow this convention throughout the text; namely,
when discussing a concept that applies equally well to a tensor of any order
(scalar, vector, second-order tensor), then we will use straight underlines.*
When discussing objects of a particular order, then we will use under-tildes, and the total number of under-tildes will equal the order of the object.
* At this point, the reader is not expected to already know what is meant by the term tensor, much
less the order of a tensor or the meaning of inner product. For now, consider this section to
apply to scalars and vectors. Just understand that the concepts reviewed in this section will also
apply in more general tensor settings, once learned.
3
T
F
A
R
D
cca
Rebe
Brann

on
Some basic terminology from functional analysis is defined very loosely

below. More mathematically correct definitions will be given later, or can be
readily found in the literature [e.g., Refs 31, 26, 27, 28, 29, 11]. Throughout
the following list, the analyst is presumed to be dealing with a set of objects
(scalars, vectors, or perhaps something more exotic) for which scalar multiplication and object addition have well-understood meanings. The single dot
multiplication symbol represents ordinary multiplication when the arguments are just scalars. Otherwise, it represents the appropriate inner product
depending on the arguments (e.g., its the vector dot product if the arguments
are vectors; its the tensor double dot product defined later when the
arguments are tensors).
A transformation f is linear if f ( x + y ) = f ( x ) + f ( y ) for all , ,
x , and y .
A transformation f is self-adjoint if y f ( x ) = x f ( y ) .
A transformation f is a projector if f ( f ( x ) ) = f ( x ) . The term
idempotent is also frequently used. A projector is a function that will
keep on returning the same result if it is applied more than once.
Any operator f must have a domain of admissible values of x for which
f ( x ) is well-defined. Throughout this report, the domain of a function
must be inferred by the reader so that the function makes sense. For
example, if f ( x ) = 1 x , then the reader is expected to infer that the
domain is the set of nonzero x . Furthermore, throughout this report, all
scalars, vectors and tensors are assumed to be real unless otherwise
stated.
The codomain of an operator is the set of all y values such that
y = f ( x ) . For example, if f ( x ) = x 2 , then the codomain is the set of
nonnegative numbers,* whereas the range is the set of reals. The term
range space will often be used to refer to the range of a linear operator.
A set S is said to be closed under a some particular operation if
application of that operation to a member of S always gives a result that
is itself a member of S. For example, the set of all symmetric matrices is
closed under matrix addition because the sum of two symmetric matrices
is itself a symmetric matrix. By contrast, set of all orthogonal matrices is
not closed under matrix addition because the sum of two orthogonal
matrices is not generally itself an orthogonal matrix.
The null space of an operator is the set of all x for which f ( x ) = 0 .
* This follows because we have already stated that x is to be presumed real.
Matrices are defined in the next section.
4
DRAF

Rebe
cca B
ranno
For each input x , a well-defined operator f must give a unique output

y = f ( x ) . In other words, a single x must never correspond to two or more
possible values of y . The operator is called one-to-one if the reverse
situation also holds. Namely, f is one-to-one if each y in the codomain of
f is obtained by a unique x such that y= f ( x ) . For example, the function
f ( x ) = x 2 is not one-to-one because a single value of y can be obtained by
two values of x (e.g., y=4 can be obtained by x=2 or x= 2 ).
If a function is one-to-one, then it is invertible. The inverse f 1 is defined
such that x= f 1 ( y ) .
A linear combination of two objects x and y is any object r that can be
expressed in the form r = x + y for some choice of scalars and . A
linear combination of three objects ( x , y , and z ) is any object r that can
be expressed in the form r = x + y + z . Of course, this definition
makes sense only if you have an unambiguous understanding of what the
objects represent. Moreover, you must have a definition for scalar
multiplication and addition of the objects. If, for example, the objects are
1 2 matrices, then scalar multiplication x of some matrix [ x 1, x 2 ]
would be defined [ x 1, x 2 ] and the linear combination x + y would be
a 1 2 matrix given by [ x 1 + y 1, x 2 + y 2 ] .
A set of objects is linearly independent if no member of the set can be
written as a linear combination of the other members of the set. If, for
example, the objects are 1 2 matrices, then the three-member set
{ [ 1, 2 ], [ 3, 4 ], [ 5, 6 ] } is not linearly independent because the third matrix
can be expressed as a linear combination of the first two matrices; namely,
[ 5, 6 ] = ( 1 ) [ 1, 2 ] + ( 2 ) [ 3, 4 ] .
The span of a collection of linearly independent vectors is the set of all
vectors that can be written as a linear combination of the vectors in the
collection. For example, the span of the two vectors { 1, 1, 0 } and { 1, 1, 0 }
is the set of all vectors expressible in the form 1 { 1, 1, 0 } + 2 { 1, 1, 0 } .
This set of vectors represents any vector { x 1, x 2, x 3 } for which x 3 =0 .
The dimension of a set or space equals the minimum number of
numbers that you would have to specify in order to uniquely identify a
member of that set minus the number of independent constraints that
those numbers must satisfy. For example, the set of ordinary engineering
vectors referenced to a commonly agreed upon laboratory basis in three
dimensions is three dimensional because each vector has three
components. However, the set of unit vectors is two-dimensional because
the three components of a unit vector n must satisfy the one constraint,
5
T
F
A
R
D
cca
Rebe
Brann

on
n 12 + n 22 + n 32 = 1 .
If a set is closed under vector addition and scalar multiplication (i.e., if
every linear combination of set members gives a result that is also in the
set), then the set is called a linear manifold, or a linear space. Otherwise,
the set is curvilinear. The set of all unit vectors is a curvilinear space
because a linear combination of unit vectors does not result in a unit
vector. Linear manifolds are like planes that pass through the origin,
though they might be hyperplanes, which is just a fancy word for a plane
of more than just two dimensions. Linear spaces can also be onedimensional. Any straight line that passes through the origin is a linear
manifold.
Zero must always be a member of a linear manifold, and this fact is often
a great place to start when considering whether or not a set is a linear
space. For example, we could assert that the set of unit vectors is not a
linear space by simply noting that the zero vector is not a unit vector.
A plane that does not pass through the origin must not be a linear space.
We know this simply because such a plane does not contain the zero
vector. This kind of plane is called an affine space. An affine space is a
set that would become a linear space if the origin were to be moved to any
single point in the set. For example, the point ( 0, b ) lies on the straight
line defined by the equation, y = mx + b . If we move the origin from
O = ( 0, 0 ) to a new location O * = ( 0, b ) , and introduce a change of
variables x * = x 0 and y * = y b , then the equation for this same line
described with respect to this new origin would become y * = mx * , which
does describe a linear space. Stated differently, a set S is affine if every
member x in that set is expressible in the form of a constant vector d plus
a vector x * that does belong to a linear space. Thus, learning about the
properties of linear spaces is sufficient to learn most of what you need to
know about affine spaces.
Given an n-dimensional linear space, a subset of members of that space is
basis if every member of the space can be expressed as a linear
combination of members of the subset. A basis always contains exactly as
many members as the dimension of the space.
A binary operation is simply a function or transformation that has two
arguments. For example, f ( x, y ) = x 2 y is a binary operation.
6
DRAF

Matrix Analysis
Rebe
There are a thousand hacking at the branches

of evil to one who is striking at the root.
Henry Thoreau
3. Matrix Analysis
Tensor analysis is neither a subset nor a superset of matrix analysis
tensor analysis complements matrix analysis. For the purpose of this report,
only the following concepts are required from matrix analysis:*
Definition of a matrix
A matrix is an ordered array of numbers that are typically arranged in the
form of a table having N rows and M columns. If one of the dimensions ( N
or M ) happens to equal 1, then the term vector is often used, although we
will often use the term array in order to avoid confusion with vectors in the
physical sense. A matrix is called square if M= N . We will usually typeset
matrices in plain text with brackets such as [ A ] .
For matrices of dimension N 1 , we may also use braces, as in { v } ;
namely, if N =3 , then
v1
{ v } = v2
(3.1)
v3
For matrices of dimension 1 M , we use angled brackets <v> ; Thus, if N =3 ,
then
<v> = [ v 1, v 2, v 3 ]
cca B
ranno
(3.2)
If attention must be called to the dimensions of a matrix, then they will be

shown as subscripts, for example, [ A ] M N . The number residing in the i th
row and j th column of [ A ] will be denoted A ij .
* Among the references listed in our bibliography, we recommend the following for additional reading, listed in order from easiest (therefore less rigorous) to most abstract (and therefore complete):
Refs. 23, 9,
Finish this list...
7
T
F
A
R
D
cca
Rebe
Brann

Matrix Analysis
on
In this report, vectors will be typeset in bold with one single under-tilde
(for example, v ) and the associated three components of the vector with
implicitly understood basis will be denoted { v } or <v > ,
respect to some
depending on whether those components are collected into a column

or row
matrix, respectively. Similarly, second-order tensors (to be defined later) will
be denoted in bold with two under-tildes (for example T ), and we will find that
tensors are often described in terms of an associated 3 3 matrix, which we

will denote by placing square brackets around the tensor symbol (for example,
[ T ] ). As was the case with vectors, the matrix of components is presumed ref
erenced
to some mutually understood underlying basis changing the basis
will not change the tensor T , but it will change its associated matrix [ T ] .
These comments will make more

sense later.
The matrix product

The matrix product of [ A ] M R times [ B ] R N is a new matrix [ C ] M N
written
[C] = [ A][ B]
(3.3)
Explicitly showing the dimensions,

[C]
MN
= [ A]
MR
[ B]
(3.4)
RN
Note that the dimension R must be common to both matrices on the righthand side, and this common dimension must reside at the inside or abutting position (the trailing dimension of [ A ] must equal the leading dimension of [ B ] )
The matrix product operation is defined
R
C ij =
A ik B kj ,
k=1
where i takes values from 1 to M ,

and j takes values from 1 to N .
(3.5)
The summation over k ranges from 1 to the common dimension, R .

As a special case, suppose that [ F ] is a square matrix of dimension N N .
Suppose that { v } is an array (i.e., column matrix) of dimension N 1 . Then
{u} = [ F ]{v}
(3.6)
must be an array of dimension N 1 with components given by

N
ui =
F ik v k ,
where i takes values from 1 to N
(3.7)
k=1
8
DRAF

Matrix Analysis
Rebe
The transpose of a matrix

The transpose of a matrix [ A ]
MN
is a new matrix [ B ]
reversed dimensions). The components of the transpose are
NM
(note the
B ij = A ji
where i takes values from 1 to N ,
and j takes values from 1 to M .
(3.8)
The transpose of [ A ] is written as [ A ] T , and the notation A ijT means the

component of [ A ] T . Thus, the above equation may be written
ij
A ijT = A ji
where i takes values from 1 to N ,
and j takes values from 1 to M .
(3.9)
The dimensions of [ A ] and [ A ] T are reverses of each other. Thus, for example, if { v } is an N 1 matrix, then { v } T is a 1 N matrix. In other words,
{ v } T = <v>
and
<v> T = { v }
(3.10)
The transpose of a product is the reverse product of the transposes. For example,
( [ A ] [ B ] ) T = [ B ] T [ A ] T , and
( <v> [ A ] ) T = [ A ] T <v> T = [ A ] T { v }
(3.11)
The inner product of two column matrices

The inner product of two column matrices, { v }
and { w }
, each
N1
N1
having the same dimension is defined
{ v }T { w } ,
or, using the angled-bracket notation, <v> { w }
(3.12)
Applying the definition of matrix multiplication, the result is a 1 1 matrix

(which is just a single number) given by
N
vk wk
cca B
ranno
(3.13)
k=1
If { v } and { w } contain component of two vectors v and w then the inner
product gives the same result as the vector dot product v w , defined later.

9
T
F
A
R
D
cca
Rebe
Brann

Matrix Analysis
on
The outer product of two column matrices.

The outer product of two column matrices, { a }
and { b }
, not necM1
N1
essarily of the same dimension is defined
{ a }{ b }T ,
or, using the angled-bracket notation, { a }<b>
(3.14)
For this case, the value of the adjacent dimension R in Eq. (3.5) is just 1, so
the summation ranges from 1 to 1 (which means that it is just a solitary term).
The result of the outer product is an M N matrix, whose ij component is
given by a i b j . If { a } and { b } contain components of two vectors a and b
then the outer product gives the matrix corresponding to the dyadic product
a b (also often denoted a b ), to be discussed in gory detail later.
The trace of a square matrix

A matrix [ A ] N N is called square because it has as many rows as it has
columns. The trace of a square matrix is simply the sum of the diagonal components:
N
tr [ A ] =
A kk
(3.15)
k=1
The trace operation satisfies the following properties:

tr ( [ A ] T ) = tr [ A ]
tr ( [ A ] [ B ] ) = tr ( [ B ] [ A ] )
(3.16)
(cyclic property)
(3.17)
The Kronecker delta

The so-called Kronecker delta is a symbol ij that is defined for different
values of the subscripts i and j . Specifically,
1 if i= j
ij =
0 if i j
(3.18)
The identity matrix

The identity matrix, denoted [ I ] , has all zero components except 1 on the
diagonal. For example, the 3 3 identity is
1 0 0
[I] = 0 1 0
0 0 1
(3.19)
The ij component of the identity is given by ij . Note that, for any array { v }
[ I ]{v} = {v}
(3.20)
10
DRAF

Matrix Analysis
Rebe
The 3D permutation symbol

The 3D permutation symbol (also known as the alternating symbol or the
Levi-Civita density) is defined
ijk
1 if ijk = 123, 231, or 312
= 1 if ijk = 321, 132, or 213
otherwise
0
+1
(3.21)
Note that the indices may be permuted cyclically without changing the value
of the result. Furthermore, inverting any two indices will change the sign of
the value. Thus, the permutation symbol has the following properties:
ijk = jki = kij = jik = ikj = kji
(3.22)
The term 3D is used to indicate that there are three subscripts on ijk each
of which take on values from 1 to 3.*
The - (E-delta) identity

If the alternating symbol is multiplied by another alternating symbol with
exactly one index being summed, a very famous and extraordinarily useful
result, called the - identity, applies. Namely,
3
ijn kln
= ik jl il jk .
(3.23)
n=1
Here, we have highlighted the index n in red to emphasize that it is

summed, while the other indices (i, j, k, and l) are free indices taking on values from 1 to 3. Later on, we are going to introduce the summation convention which states that expressions having one index appearing exactly twice
in a term should be understood summed over from 1 to 3 over that index.
Index symbols that appear exactly once in one term are called free indices,
taking values from 1 to 3, and they must appear exactly once in all of the other
terms. Using this convention, the above equation can be written as
ijn kln = ik jl il jk .
cca B
ranno
(3.24)
* Though not needed for our purposes, the 2D permutation symbol ij is defined to equal zero if i= j ,
+1 if ij=12 , and 1 if ij = 21 . The 4D permutation symbol ijkl is defined to equal zero if any of
the four indices are equal; it is +1 if ijkl is an even permutation of 1234 and 1 if ijkl is an odd
permutation. A permutation is simply a rearrangement. The permutation ijkl is even if rearranging
it back to 1234 can be accomplished by an even number of moves that exchange two elements at a
time. A cyclic permutation of an n-D permutation symbol will change sign if n is even, but remain
unchanged if n is odd. Thus, for our 3D permutation symbol, cyclic permutations dont change
sign, whereas cyclic permutations of the 4D permutation symbol will change sign.
11
T
F
A
R
D
cca
Rebe
Brann

Matrix Analysis
on
Because of the cyclic properties of the permutation symbol, the - identity

applies whenever any index on the first matches any index on the second .
For example, the above equation would apply to the expression nij kln
because nij = ijn . The negative of the - identity would also apply to the
expression inj kln because inj = ijn . Of course, if a negative permutation
is also required to place the summation index at the end of the second , then
the positive of the - identity would again apply.
To make an expression fit the index structure of Eq. (3.23), most people
laboriously apply the cyclic property to each alternating symbol until the
summed index is located at the trailing side on both of them. Keeping track of
whether or not these manipulations will require changing the final sign of the
right hand side of the - identity is one of the most common and avoidable
careless mistakes made when people use this identity. Even once the summation index has been properly positioned at the trailing end of each alternating
symbol, most people then apply a slow (and again error-prone process of figuring out where the free indices go). Typically people apply a left-right/outsideinside rule. By this, we mean that the free indices on the left sides of ijn and
kln are the indices that go on the first , then the right free indices go on the
second , then the outer free indices go on the third , and (finally) the inner
free indices go on the last . The good news is... you dont have to do it this
way! By thinking about the - identity in a completely different way, you can
avoid both the initial rearrangement of the indices on the alternating symbols
and the slow left-right-out-in placement of the indices. Lets suppose you want
to apply the - identity to the expression imk pin . First write a skeleton of
the identity as follows
imk pin = ?? ?? ?? ??
(3.25)
Our goal is to find a rapid and error-minimizing way to fill in the question
marks with the correct index symbols. Once you have written the skeleton,
look at the left-hand side to identify which index is summed. In this case, it is
the index i . Next say out loud the four free indices in an order defined by
cyclically moving forward from the summed index on each alternating symbol. Each alternating symbol has two free indices. To call out their names by
moving cyclically forward, you simply say the name of the two indices to the
right of the summed index, wrapping back around to the beginning if necessary. For example, the two indices cyclically forward from p in the sequence
pqr are qr; the two indices cyclically forward from q are rp; the two
indices forward from r are pq. For the first alternating symbol in our skeleton of Eq. (3.25), the two indices cyclically forward from the summed index i
are mk whereas the two indices cyclically forward from i in the second alternating symbol are np. You can identify these pairs quickly without ever hav-
12
DRAF

Matrix Analysis
Rebe
ing to rearrange anything, and you can (in your head) group the pairs together
to obtain a sequence of four free indices mknp. The final step is to write
these four indices onto the skeleton. If the indices are ordered 1234, then you
should write the first two indices (first and second) on the skeleton like this
1? 2? 1? 2?
(3.26)
You write the last pair (third and fourth) in order (34) on the first term and in
reverse order (43) on the last term:
13 24 14 23
(3.27)
Thus, for example, to place the free indices mknp onto the Kronecker deltas,
you would first take care of the mk by writing
m? k? m? k?
(3.28)
Then you just finish off with the last two np free indices by writing them
first in that order on the first term and in reverse order on the second term to
obtain the final result:
imk pin = mn kp mp kn .
(3.29)
This may seem a bit strange at first (especially if you are already stuck in the
left-right-out-in mind set), but this method is far quicker and less error-prone.
Give it a try until you become comfortable with it, and you probably wont
dream of going back to your old way.
The - (E-delta) identity with multiple summed indices

Recall that the - identity is given by
3
ijn kln
= ik jl il jk .
(3.30)
n=1
What happens if we now consider the case of two alternating symbols multiplied side-by-side with two indices being summed? This question is equivalent
to throwing a summation around the above equation in such a manner that
we add up only those terms for which j=l . Then
3
ijn kjn
j = 1n = 1
( ik jj ij jk )
j=1
= ik ( 11 + 22 + 33 )
= 3 ik ik
= 2 ik
cca B
ranno
( i1 1k + i2 2k + i3 3k )
(3.31)
Note that we simplified the first term by noting that 11 + 22 + 33 = 1 + 1 + 1 = 3 .

The second term was simplified by noting that i1 1k + i2 2k + i3 3k will be zero if
i k or it will equal 1 if i=k . Thus, it must be simply ik .
13
T
F
A
R
D
cca
Rebe
Brann

Matrix Analysis
on
Using similar logic, the - identity with all indices summed is equivalent
to setting i=k in the above equation, summing over each instance so that the
result is six. To summarize using the summation conventions,
ijn kjn = 2 ik
(3.32)
ijk ijk = 6
(3.33)
Determinant of a square matrix

The simplest way to explain what is meant by a determinant is to define it
recursively. In this section, we show how the determinant of a 3 3 matrix
can be alternatively defined by using the three-dimensional permutation symbol of Eq. (3.21).
A 1 1 matrix is just a single number. The determinant of a 1 1 matrix is
defined to equal its solitary component. Thus,
det [ A 11 ] A 11
(3.34)
The determinant of a 2 2 matrix is defined by

det
A 11 A 12
A 21 A 22
A 11 A 22 A 12 A 21
(3.35)
The determinant of a 3 3 matrix is defined by

A 11 A 12 A 13
det A 21 A 22 A 23
A 31 A 32 A 33
( A 11 A 22 A 33 + A 12 A 23 A 31 + A 13 A 21 A 32 )
( A 13 A 22 A 31 + A 11 A 23 A 32 + A 12 A 21 A 33 )
(3.36)
Note that we have arranged this formula such that the first indices in each
factor are 123. For the positive terms, the second indices are all the positive
permutations of 123. Namely: 123, 231, and 312. For the negative terms, the
second indices are all the negative permutations of 123. Namely: 321, 132, and
213. This relationship may be written compactly by using the permutation
symbol ijk from Eq. (3.21). Namely, if [ A ] is a 3 3 matrix, then
14
DRAF

Matrix Analysis
det [ A ] =
Rebe
3
ijk A1i A2 j A3k
(3.37)
i = 1 j = 1k = 1
This definition can be extended to square matrices of arbitrarily large dimension by using the n-dimensional permutation symbol (see footnote on page 11).
Alternatively, for square matrices of arbitrarily large dimension, the determinant can be defined recursively as
N
det [ A ] N N =
A ij A ijC
(no summation on index
i)
(3.38)
j=1
where i is a free index taking any convenient value from 1 to N (any choice
for i will give the same result). The quantity A ijC is called the cofactor of
A ij , and it is defined by
A ijC = ( 1 ) i + j det [ M ij ] ( N 1 ) ( N 1 )
(3.39)
Here [ M ij ] is the submatrix obtained by striking out the i th row and j th column of [ A ] . The determinant of [ M ij ] is called the minor associated with
A ij . By virtue of the ( 1 ) i + j , the cofactor component A ijC is often called the
signed minor. The formula in Eq. (3.38) is almost never used in numerical
calculations because it requires too many multiplications,* but it frequently
shows up in theoretical analyses.
The index i in Eq. (3.39) may be chosen for convenience (usually a row
with several zeros is chosen to minimize the number of sub-determinants that
must be computed). The above definition is recursive because det [ A ] N N is
defined in terms of smaller ( N 1 ) ( N 1 ) determinants, which may in turn
be expressed in terms of ( N 2 ) ( N 2 ) determinants, and so on until the
determinant is expressed in terms of only 1 1 determinants, for which the
determinant is defined in Eq. (3.34). As an example, consider using Eq. (3.38)
to compute the determinant of a 3 3 matrix. Choosing i=1 , Eq. (3.38) gives
A 11 A 12 A 13
det A 21 A 22 A 23 = A 11 det
A 31 A 32 A 33
A 22 A 23
A 32 A 33
A 12 det
A 21 A 23
A 31 A 33
+ A 13 det
A 21 A 22
A 31 A 32
cca B
ranno
(3.40)
Alternatively choosing i=2 , Eq. (3.38) gives

* Specifically, for large values of the dimension N , the number of multiplications required to evaluate the determinant using Crammers rule (as Eq. 3.38 is sometimes called) approaches ( e 1 ) N! ,
where e is the base of the natural logarithm. An ordinary personal computer would require a few
million years to compute a 20 20 determinant using Cramers rule! Far more efficient decomposition methods [__] can be used to compute determinants of large matrices.
15
T
F
A
R
D
cca
Rebe
Brann

Matrix Analysis
on
A 11 A 12 A 13
det A 21 A 22 A 23 = A 21 det
A 31 A 32 A 33
A 11 A 13
A 32 A 33
+ A 22 det
A 11 A 13
A 31 A 33
A 23 det
A 11 A 12
A 31 A 32
(3.41)
After using Eq. (3.35) to compute the 2 2 submatrices, both of the above
expressions give the same final result as Eq. (3.36).
Some key properties of the determinant are listed below:
det ( [ A ] T ) = det [ A ]
(3.42)
det ( [ A ] [ B ] ) = ( det [ A ] ) ( det [ B ] )
(3.43)
det ( [ A ] N N ) = N det [ A ]
(3.44)
1
det ( [ A ] 1 ) = ----------------det [ A ]
(3.45)
If [ B ] is obtained by swapping two rows (or two columns) of [ A ] ,

then det [ B ] = det [ A ] .
(3.46)
If any row of [ A ] can be written as a linear combination of the other
rows, then det [ A ]=0 . A special case is that det [ A ]=0 if any two rows of
(3.47)
[ A ] are equal.
For 3 3 determinants, the last two properties allow us to generalize
Eq. (3.37) to read
3
pqr det [ A ] =
ijk A pi Aqj Ark
(3.48)
i = 1 j = 1k = 1
or, using the summation convention in which repeated indices are understood
to be summed (and, for clarity, now shown in red),
pqr det [ A ] = ijk A pi A qj A rk
(3.49)
This expression is frequently cited in continuum mechanics textbooks as the

indicial definition of the determinant of a 3 3 matrix. Multiplying the above
formula by pqr and summing over p q and r (and using Eq. 3.33) reveals
that
1
det [ A ] = --- pqr A pi A qj A rk ijk
6
(3.50)
Here, there are implied summations over the indices i,j,k,p,q, and r. If it were
expanded out, the above expression would contain 729 terms, so it is obviously
not used to actually compute the determinant. However, it is not at all uncommon for expressions like this to show up in analytical analysis, and it is therefore essential for the analyst to recognize that the right-hand-side simplifies
so compactly.
16
DRAF

Matrix Analysis
Rebe
Principal sub-matrices and principal minors

A so-called n n principal submatrix of a square matrix [ A ] N N is any
n n submatrix (where n N ) whose diagonal components are also diagonal
components of the larger matrix. For example,
A 11 A 13
(3.51)
A 31 A 33
A 12 A 13
is a principal submatrix, whereas A22 A23 is not a principal submatrix. For a

3 3 matrix, there are three 1 1 principal submatrices (identically equal to
the diagonal components), three 2 2 principal submatrices, and only one
3 3 principal submatrix (equal to the matrix [ A ] itself). A sequence of 1 1 ,
2 2 , N N submatrices is nested if the 1 1 matrix is a submatrix of
the 2 2 matrix, the 2 2 matrix is a submatrix of the next larger submatrix,
and so forth.
A principal minor is the determinant of any principal submatrix. The term
nested minors means the determinants of a set of nested submatrices.
Matrix invariants
The k th characteristic invariant, denoted I k , of a matrix [ A ] is the sum
of all possible k k principal minors. For a 3 3 matrix, these three invariants are
I 1 = A 11 + A 22 + A 33
I 2 = det
A 11 A 12
A 21 A 22
+ det
(3.52)
A 11 A 13
A 31 A 33
+ det
A 22 A 23
(3.53)
A 32 A 33
A 11 A 12 A 13
I 3 = det A 21 A 22 A 23
cca B
ranno
(3.54)
A 31 A 32 A 33
Warning: if the matrix is non-symmetric, the characteristic invariants are not
a complete set of independent invariants. If all three characteristic invariants
of a symmetric matrix are zero, then the matrix itself is zero. However, as discussed later, it is possible for all three characteristic invariants of a non-symmetric matrix to be zero without the matrix itself being zero.
17
T
F
A
R
D
cca
Rebe
Brann

Matrix Analysis
on
Positive definite
A square matrix [ B ] N N is positive definite if and only if
{ v }T [ B ]{ v } > 0
for all { v }
In indicial notation, this requirement is
N
(3.55)
vi Bij v j > 0
(3.56)
i=1j=1
Written out explicitly for the special case of a 2 2 matrix

B 11 v 1 v 1 + B 12 v 1 v 2 + B 21 v 2 v 1 + B 22 v 2 v 2 > 0
(3.57)
B 12 + B 21
2 ------------------------- v 2 v 1 .
Note that the middle two terms can be combined and written as
B 11 + B 11
- . The third term can also be
Similarly, we can write the first term as 2 -----------------------
2
so written. Thus, the requirement for positive definiteness depends only on
the symmetric part of the matrix [ B ] . The non-symmetric part has no influence on whether or not a matrix is positive definite. Consequently, we may
replace Eq. (3.55) by the equivalent, but more carefully crafted statement:
[ B ] is positive definite if and only if { v } T [ A ] { v } > 0 for all { v } , where

[ A ] is the symmetric part of [ B ] .
It can be shown that, a matrix is positive definite if and only if the characteristic invariants of the symmetric part of the matrix are all positive.*
Fortunately, there is an even simpler test for positive definiteness: you only
have to verify that any nested set of principal minors are all positive! This calculation is easier than finding the invariants themselves because it requires
evaluation of only one principal minor determinant of each size (you dont
have to evaluate all of them).
* It is possible to construct a matrix that has all positive invariants, but whose symmetric part does
not have all positive invariants.
18
DRAF

Matrix Analysis
Rebe
The cofactor-determinate connection

Let [ A ] C denote the matrix of cofactors A ijC associated with a square
matrix [ A ] N N . This matrix is also sometimes called the adjugate matrix
(not to be confused with adjoint). Recall the definition of the cofactor given in
Eq. (3.39):
A ijC = ( 1 ) i + j det [ M ij ] ( N 1 ) ( N 1 )
(3.58)
By virtue of Eq. (3.42), we note that the transpose of the cofactor matrix is
identically equal to the cofactor matrix associated with [ A ] T . In other words,
the cofactor and transpose operations commute:
( [ A ]C )T = ( [ A ]T )C
(3.59)
As a short hand, we generally eliminate the parentheses and simply write

[ A ] CT to mean the transpose of the cofactor (or, equivalently, the cofactor of
the transpose). The generalization of Eq. (3.38) is
N
if
0
A ik A Cjk =
det [ A ] if
k=1
i j
i= j
(3.60)
Written more compactly,

N
A ik A Cjk = det [ A ] ij
(3.61)
k=1
Written in matrix form,

[ A ] [ A ] CT = ( det [ A ] ) [ I ]
(3.62)
It turns out that the location of the transpose and cofactor operations is inconsequential the result will be the same in all cases. Namely, the following
results also hold true:
[ A ] [ A ] CT = [ A ] C [ A ] T = [ A ] T [ A ] C = [ A ] CT [ A ] = ( det [ A ] ) [ I ]
(3.63)
Inverse
The inverse of a matrix [ A ] is the matrix denoted [ A ] 1 for which
[ A ] [ A ] 1 = [ A ] 1 [ A ] = [ I ]
(3.64)
If the inverse exists, then it is unique. If the inverse does not exist, then the
matrix [ A ] is said to be non-invertible or singular. A necessary and sufficient condition for the inverse to exist is that the determinant must be nonzero:
det [ A ]
cca B
ranno
(3.65)
19
T
F
A
R
D
cca
Rebe
Brann

Matrix Analysis
on
Comparing Eqs. (3.63) and (3.64), we note that the inverse may be readily
computed from the cofactor by
[ A ] CT
[ A ] 1 = ----------------det [ A ]
(3.66)
While this definition does uniquely define the inverse, it must never be used
as a definition of the cofactor matrix. The cofactor matrix is well-defined and
generally nonzero even if the matrix [ A ] is singular.
Eigenvalues and eigenvectors

As mentioned in Eq. (3.37), a nonzero vector (array) { p } is called an eigenvector of a square matrix [ A ] if there exists a scalar , called the eigenvalue,
such that [ A ] { p } = { p } . In order for this condition to have a non-trivial
solution, the determinant of the matrix [ A ] [ I ] must be zero. Setting this
determinant to zero results in a polynomial equation, called the characteristic
equation, for . If [ A ] is a 2 2 matrix, the equation will be quadratic. If [ A ]
is a 3 3 matrix, the equation will be cubic, and so forth. We highly recommend that you do not construct the matrix [ A ] [ I ] and then set its determinant equal to zero. While that would certainly work, it allows for too many
opportunities to make an arithmetic error. Instead, the fastest way to generate the characteristic equation is to first find all of the characteristic invariants of [ A ] . These invariants are the coefficients in the characteristic
equation, alternating sign, as follows
For [ A ] 2 2 , the characteristic equation is
2 I 1 + I 2 = 0 ,
where I 1 = A 11 + A 22 , and
I 2 = det
A 11 A 12
A 21 A 22
(3.67)
For [ A ] 3 3 , the characteristic equation is

3 I 12 + I 2 I 3 = 0 ,
where I 1 = A 11 + A 22 + A 33 ,
I 2 = det
A 11 A 12
A 21 A 22
+ det
A 11 A 13
A 31 A 33
+ det
A 22 A 23
A 32 A 33
, and
A 11 A 12 A 13
(3.68)
I 3 = det A 21 A 22 A 23
A 31 A 32 A 33
For [ A ] 4 4 , the characteristic equation is 4 I 1 3 + I 2 2 I 3 + I 4 = 0 .

Higher dimension matrices are similar.
20
DRAF

Matrix Analysis
Rebe
Because there the characteristic equation is a polynomial equation, an

N N matrix will have up to N possible eigenvalues. For each solution i
there exists at least one corresponding eigenvector { p } i , which is determined
by solving
[ A ] { p }i = i { p }i
(no sum on i ).
(3.69)
The solution for { p } i will have an undetermined magnitude and, for symmetric matrices, it is conventional to set the magnitude to one. For non-symmetric
matrices, however, the normalization convention is different, as explained
below.
If an eigenvalue i has algebraic multiplicity m (i.e., if the characteristic
equation gives a root i repeated m times), then there can be no more than a
total of m independent eigenvectors associated with that eigenvalue there
might be fewer (though there is always at least one). If the matrix [ A ] is symmetric, then it is well known [22] that it is always possible to find m independent eigenvectors. The directions of the eigenvectors when the multiplicity m
is greater than one are arbitrary. However, the one thing that is unique is the
span of these vectors (see page 5), and it is conventional to set the eigenvectors to any orthonormal set of vectors lying in the span. For non-symmetric
matrices, it might happen that an eigenvalue of multiplicity m corresponds to
a total of < m linearly independent eigenvectors, where is called the geometric multiplicity. For example, the matrix
53
05
(3.70)
Has an eigenvalue = 5 with algebraic multiplicity of two. To find the associate eigenvector(s), we must solve
5 3 p1 = 5 p1
p2
0 5 p2
cca B
ranno
(3.71)
Multiplying this out gives

5 p1 + 3 p2 = 5 p1
(3.72)
5 p2 = 5 p2
(3.73)
The second equation gives us no information, and the first equation gives the
constraint that p 2 = 0 . Therefore, we have only one eigenvector (geometric
multiplicity equals one) given by { 1, 0 } even though the eigenvalue had algebraic multiplicity of two. When the geometric multiplicity of an eigenvector is
less than the algebraic multiplicity, then there does still exist a subspace that
is uniquely associated with the multiple eigenvalue. However, characterizing
this subspace requires solving a generalized eigenproblem to construct additional vectors that will combine with the one or more ordinary eigenvectors to
21
T
F
A
R
D
cca
Rebe
Brann

Matrix Analysis
on
form a set of vectors that span the space. The process for doing this is onerous,
and we have not yet personally happened upon any engineering application
for which finding these generalized eigenvectors provides any useful information, so we will not cover the details. Instructions for the process can be found
in [23,21, 24]. If the generalized eigenvectors are truly sought, then they can
be found via the JordanDecomposition command in Mathematica [25] (see
discussion below to interpret the result).
Similarity transformations. Suppose that we have a set of eigenvalues
{ 1, 2, , N } for a matrix [ A ] , possibly with some of these eigenvalues having algebraic multiplicities greater than one. Let [ L ] denote the matrix whose
columns contain the corresponding eigenvectors (augmented, where necessary, to include generalized eigenvectors for the cases where the geometric
multiplicity is less than the algebraic multiplicity; the ordinary eigenvectors
corresponding to a given eigenvalue should always, by convention, be entered
into columns of [ L ] before the generalized eigenvectors). Then it can be shown
that the original matrix [ A ] satisfies the similarity transformation
[ A ] = [ L ] [ ] [ L ] 1
(3.74)
If there are no generalized eigenvectors contained in the matrix [ L ] , then the

matrix [ ] is diagonal, with the diagonal components being equal to the
eigenvalues. In this case, the original matrix [ A ] is said to be diagonalizable. If, on the other hand, [ L ] contains any generalized eigenvectors, then
[ ] still contains the eigenvalues on the diagonal, but it additionally will contain a 1 in the k 1, k position corresponding to each k th generalized eigenvector. In this form, the matrix [ ] is said to be in Jordan canonical form. For
example, the similarity transformation corresponding to Eq. (3.70) is
53 = 1 0 5 1 1 0
05
0 13 0 5 0 13
(3.75)
This result can be obtained in Mathematica [25] via the command JordanDecomposition[{{5,3},{0,5}}]. From this result, we note that the presence of
the 1 in the 12 position of the [ ] matrix implies that the second column of
[ L ] must contain a generalized eigenvector.
The matrix L will be orthogonal (i.e., [ L ] 1 = [ L ] T ) if and only if the original matrix [ A ] is symmetric. For symmetric matrices, there will never be any
generalized eigenvectors (i.e., the algebraic and geometric eigenvalue multiplicities will always be equal), and the [ ] matrix will therefore always be
fully diagonal (no 1 on any off-diagonal).
22
DRAF

Matrix Analysis
Rebe
Finding eigenvectors by using the adjugate

Recall that { p } is an eigenvector of [ A ] if [ A ] { p } = { p } , where is the
eigenvalue. This equation can be written
[ B ] { p } = 0 , where [ B ] = [ A ] [ I ]
(3.76)
Recall that we determined the eigenvalue by setting det [ B ] = 0 . We can likewise us the [ B ] matrix to determine the eigenvectors. Given that the determinant of [ B ] is zero, Eq. (3.63) tells us that
(3.77)
[ B ] [ B ] CT = 0 ,
from which it follows that if there is any nonzero column of the adjugate
matrix, [ B ] CT , then that column must be an eigenvector of [ B ] associated with
the eigenvalue that was used to construct [ B ] in Eq. (3.76). It turns out that,
for distinct eigenvalues (i.e., ones with multiplicity of 1), there will always be a
non-zero column of [ B ] CT , and therefore the adjugate matrix is a slick and
easy way to find the eigenvector. Unfortunately, the situation is not so great
when the eigenvalue is a double (or higher) root. For eigenvalues of multiplicity greater than one, it is still true that any nonzero column of [ B ] CT will be
an eigenvector, but the adjugate might turn out to be zero or it might not capture all of the possible eigenvectors. Consider, for example,
100
[ A] = 0 0 0
000
(3.78)
For the eigenvalue = 1 , the B matrix is

0 0 0
[ B ] = 0 1 0
0 0 1
for which the adjugate is
100
[ B ] CT = 0 0 0
000
the eigenvector is the only nonzero column,
1
0
0
cca B
ranno
, and therefore
(3.79)
However, for the double root eigenvalue, = 0 , the B matrix is just [ A ]

itself, and the adjugate is the zero matrix. Having no nonzero columns, this
adjugate matrix gives us no guidance to determine the eigenvector, and traditional methods must instead be used. Our recommendation is: use the adjugate to compute eigenvectors associated with distinct non-multiple
eigenvalues, but dont bother with it when the eigenvalue is a multiple root.
23
T
F
A
R
D
cca
Rebe
Brann

Vector/tensor notation
on
When two men in business always agree, one of

them is unnecessary. William Wrigler, Jr.
4. Vector/tensor notation
Ordinary engineering vectors
We use the term ordinary to indicate the notion of a vector as it is introduced in elementary calculus texts namely, something in 3D physical space
that has length and orientation. Physical examples include velocity, electric
field, and angular rotation. This definition is horribly ambiguous, but it is
adequate for reviewing basic concepts and we promise to atone for this transgression later. In general, we use the term engineering whenever a quantity
is defined from a physical rather than mathematical perspective. A more correct mathematicians definition of the term vector can be found on page 146,
where we will also introduce abstract vectors in higher-dimensional spaces.
Engineering laboratory base vectors

Let { e 1, e 2, e 3 } denote a set of mutually perpendicular vectors of unit mag we
will refer to as the laboratory basis or the laboratory
nitude, which
triad. Elementary calculus books often denote these vectors by { i, j, k } , but
practicing researchers just cant spare that many letters of the alphabet. The
unit base vectors are always presumed to exist as a reference from which all
directional quantities may be described. The orientation of the laboratory
triad should be regarded as mutually agreed upon by all observers (e.g., all
observers might agree that they point in the same directions as the edges of
the walls and floor in one corner of their laboratory, or they might define the
directions based on the stars it doesnt really matter what the choice is, so
long as everyone agrees on it). The laboratory base vectors are presumed to be
fixed in time and space. Unless otherwise indicated, all vector and tensor components cited in this report are referenced to the laboratory basis.
24
DRAF

Rebe
Other choices for the base vectors

We insist that the laboratory basis must exist and all observers must agree
upon its definition. However, other bases can be used as well. Different observers can, if they wish, opt to use supplemental basis triads. All observers will be
able to communicate effectively with each other by transforming their own
results into equivalent results expressed using the commonly agreed-upon lab
basis. The rules governing how to accomplish these transformations from one
basis to another are central to the study of vector and tensor analysis.
When explaining the meanings of various vector and tensor operations,
this report limits its scope to definitions that apply only to mutually perpendicular vectors, each of unit magnitude. We also require that the orthonormal
triad be right-handed (i.e., crossing the first lab base vector into the second
one gives the third one). Any basis that is both right-handed and orthonormal
is called regularized. All vector and tensor operations presented in this report
apply only to regularized bases; generalized definitions that apply to irregular
and curvilinear bases are provided in a sequel report [4].
Basis expansion of a vector

Let v be an ordinary vector (like velocity or electric field). Its expansion in
terms of an orthonormal basis { e 1, e 2, e 3 } is

3
v = v1 e1 + v2 e2 + v3 e3 =
vi e i
(4.1)
i=1
where v i are called the components of the vector with respect to the basis.
The zero vector 0 is defined to be the vector whose components are all zero.
We may arrange the components of a vector in a column matrix as

v1
{ v } = v2
(4.2)
v3
and the expansion of Eq. (4.1) is analogous to writing
v1
1
0
0
v2 = v1 0 + v2 1 + v3 0
0
0
1
v3
(4.3)
We may alternatively arrange the components of a vector into a row matrix:

<v> = v 1 v 2 v 3
cca B
ranno
(4.4)
The row representation is just the transpose of the column representation:
25
T
F
A
R
D
cca
Rebe
Brann

on
{ v } T = <v> ,
(4.5)
However there is no such thing as the transpose of a vector v T is meaning

less! If { v } are the components of a vector collected into a 3 1 matrix, then
{ v } T also represents the same vector even though the matrix { v } T happens
to be 1 3 .
Summation convention details

The expression in Eq. (4.1) is written more compactly by using Einsteins summation convention as
v = vi ei
(4.6)
The summation convention is defined such that any index (in this case i) that
is repeated exactly twice in a term is understood to be summed from 1 to 3.
The summation ranges up to 3 because ordinary engineering vectors are
always referenced to 3D physical space.
Later on, we will define quantities (like matrix components A ij ) that have
more that one index. Then, for example, the expression A ij v j , for which the
index j is repeated, would mean the same thing as
3
A ij v j
(4.7)
j=1
In this expression, note that the index i occurs exactly once and is not
repeated. Therefore the above expression is actually three expressions, corresponding to the index i taking the values from 1 to 3. For rectangular Cartesian components, the summation convention has two fundamental rules
(extra rules that apply for irregular bases can be found in Ref. [4]):
1. An index that occurs exactly twice in a term is called a dummy index, and
it is understood to be summed from 1 to 3, with the implied summation
symbol applied only to the term in which the repeated dummy index
appears.
2. An index that occurs exactly once in a term is called a free index, and it
must also appear exactly once in every other term.
The following expressions violate the summation convention:
ai + b j
(violates rule 2)
a i A ij b i
(violates rule 1)
(4.8)
26
DRAF

Rebe
The following expressions all satisfy the summation convention:

ai bi
b k A ik + a i + w jk U pkqi T jp v q
ai b j
a i A ij b k
(4.9a)
(4.9b)
(4.9c)
(4.9d)
The number of free indices indicates the number of expressions contained

implicitly in a single indicial expression. The first expression above has no
free indices, so it corresponds to a 0 th -order tensor (i.e., a scalar). The second
expression has one free index, so it corresponds to a 1 st -order tensor
(i.e., a vector). The next expression has two free indices, so it corresponds to a
2 nd -order tensor. The last expression also has two free indices, so it also corresponds to a 2 nd -order tensor. We will precisely define what is meant by the
term tensor later; for now, we are merely clarifying notation and terminology. Incidentally, whenever the term tensor is used, it usually refers to a second-order tensor. Many authors alternatively use the word rank or valence
instead of order. Hence, a tensor of rank 2 is the same thing as what we will
call a tensor of order 2. We prefer the term order to avoid confusion with the
meaning of rank from matrix analysis.
Incidentally, note that Eq. (4.9b) uses k as a dummy sum index in both the
first and last terms. This is perfectly legitimate. The rule that says a dummy
index must appear exactly twice applies to each individual term. Written out
in conventional form using summation symbols, Eq. (4.9b) is
3
k=1
( b k A ik ) + a i +
( w jk U pkqi T jp vq )
(4.10)
j = 1k = 1 p = 1q = 1
where the free index i ranges from 1 to 3.

Of course, summation is commutative (we can sum something from i=1 to N
and then from j=1 to M or vice versa without changing the final result). Said
differently, the summation signs can be applied in any order. For example, the
above equation could be written instead as
3
( bk Aik ) + ai + ( w jk U pkqi T jp vq )
k=1
cca B
ranno
(4.11)
q = 1k = 1 p = 1 j = 1
Moving the summation symbols from the jkpq order to this qkpj order has no
impact on the result.
Dont forget to remember what repeated indices really mean. Beginning students sometimes forget to recall that the summation rules are really
just a notational convenience. Sometimes its wise to go back to conventional
notation to simplify an indicial expression. Recall, for example, the definition
of the Kronecker delta:
27
T
F
A
R
D
cca
Rebe
Brann

on
1 if i= j
ij =
0 if i j
(4.12)
Now consider the indicial expression

s = kk
(4.13)
There are no free indices, so the result must be a scalar. New students almost
always get burned by using Eq. (4.12) to say that kk must be equal to 1. However, this conclusion is wrong. The index k is repeated, so it must be summed.
In other words, Eq. (4.13) really says
3
s =
kk
= 11 + 22 + 33 = 1 + 1 + 1 = 3
(4.14)
k=1
Recalling the true meaning of the indicial notation is also essential for simplifying other expressions where ij appears. Consider, for example,
A km mj
(4.15)
It is wise to mentally realize that this is really

3
A km mj
(4.16)
m=1
Using the definition of the Kronecker delta, note that

A kj if m= j
A km mj =
0 if m j
equation*
The index m is not intended to be summed in this
(4.17)
Consequently, only one of the three terms ( A k1 1 j , A k2 2 j , or A k3 3 j ) in

Eq. (4.16) will be nonzero (the one where m= j ). Therefore,
A km mj = A kj
(4.18)
This result is just one example of how a Kronecker delta may be removed from
an expression whenever one of its indices is a dummy sum index.
* Its important to inform your readers when you wish to temporarily suspend the summation conventions as we have done here. Some writers indicate that they do not wish for an index to be summed
by marking it with an overbar, as in A km mi , or by putting the index in parentheses, as in
A k ( m ) ( m )i or by typesetting non-summed indices with capital letters, as in A kM Mj .
28
DRAF

Rebe
Indicial notation in derivatives. The summation convention also applies to

derivatives. For example
y ik
----------z i
(4.19)
means the same thing as

3
y ik
---------z i
(4.20)
i=1
BEWARE: avoid implicit sums as independent variables. A word of caution is essential here. When you employ the summation convention in derivatives, you should be extra careful to be clear about what the independent
variable is. Consider for example
w = b 11 + b 22 + b 33
(4.21)
What would the expression w b kk mean?
There are two possible answers:
w
w
------------ = --------------------------------------------- = 1
b kk
( b 11 + b 22 + b 33 )
(4.22)
w
w
w
w
------------ = ----------- + ----------- + ----------- = 1 + 1 + 1 = 3
b kk
b 11 b 22 b 33
(4.23)
or
The two answers are not the same, so we need to establish a precedence rule.
Our experience in reading the literature is that most authors intend for the
expression w b kk to be interpreted as Eq. (4.23). Thus, the precedence rule
is to always apply summations after taking derivatives. In other words, imagine that w is a function of nine b ij components. After finding all nine w b ij
derivatives, then w b kk is obtained by summing the three derivatives corresponding to i = j . To minimize confusion, we recommend that you write
w
kk , where ij = ---------b ij
(4.24)
If you really want a derivative to be interpreted according to Eq. (4.22),

then you should write the expression more carefully as
w
------- , where s = b kk
s
cca B
ranno
(4.25)
29
T
F
A
R
D
cca
Rebe
Brann

on
Reading index STRUCTURE, not index SYMBOLS

This section may be skipped if desired. Here we try to give a little insight
to new students about how to read indicial formulas to maximize understanding and to avoid a queasy alphabet soup feeling when viewing expressions
with lots of indices.
The actual symbols selected to be free and dummy indices are entirely up
to you, so long as your choices do not violate the summation rules. Beginning
students often fail to read indicial formulas in the most instructive manner.
Experienced professionals know to instead look at the structure of where the
free and dummy indices appear. For example, to memorize some strange formula like
a jk a ki a ij
Y ijk = ----------+ ----------- ---------- ,
x i
x j x k
(4.26)
you would be wise to ignore the actual symbols used for the indices. Instead,
you should work to deduce the overall structure of the index placement. For
example, if you want to write down the formula for Y jim , then you could
always painstakingly convert Eq. (4.26) by replacing in all occurrences of i by
j , j by i , and k by m . However, doing it that way is error prone (especially
since you cannot do it step-by-step because you would violate the sum rules by
having four js in the expression after step 1 of replacing i by j ). People who
really understand indicial notation would construct the formula for Y jim by
spending a little time smelling the structure of the Eq. (4.26). If you look
carefully at that defining equation, you will note that the denominators in
the three terms on the right hand side have the same indices as those on the
left hand side and they also appear in the same order. Thus, your first step
to constructing the expression for Y jim would be to write a partial skeleton
formula as
a ?? a ?? a ??
Y jim = ---------+ ---------- ---------- ,
x j
x i x m
(4.27)
where the ?? stands for indices not yet inserted. Again looking at the structure of Eq. (4.26) you would note that the subscripts on each a ?? are simply
the other two indices not already in the denominator. Furthermore, those
subscripts are placed in an order that is a positive permutation of the free
indices moving clockwise from the index already placed in the denominator.
Specifically, the positive permutations of jim are: jim , imj , and mji .
Because the first term in Eq. (4.27) has x j in the denominator, we know that
the other two indices for the first term must be im . Similarly, the second
term has x i in the denominator, so the other two indices must be mj (not jm
because we need to select the same ordering as the positive permutation), etc.
Thus, the final expression is
30
DRAF

Rebe
a im a mj a ji
- + ------------ ---------Y jim = ----------x j
x i x m
(4.28)
In continuum mechanics, another common index structure is of the following

form
U ijkl = T ikjl + W iljk
(4.29)
To smell the structure of the index placement, you might find it useful to
mentally replace the indices with numbers representing the order in which
you should write them on the paper:
U 1234 = T 1324 + W 1423
(4.30)
Thus, for example, if you want to write the formula for U pqim , then you
should again start with a partially completed skeleton in which you place only
the first two indices:
U pq?? = T p?q? + W p?q?
(4.31)
Then you fill out the remaining two indices im by placing them in that order
on the first term and in the reverse order in the second term to obtain
U pqim = T piqm + W pmqi
(4.32)
Aesthetic (courteous) indexing

Once you have derived a final result, it is often helpful to your readers if
you change your choice of index symbols so that the free indices are ordered
alphabetically on the left hand side of the equation. For example, if your final
result is something weird like
u skm = A si B mik ,
(4.33)
then re-writing it in the equivalent form,

u ijk = A in B knj ,
cca B
ranno
(4.34)
is often appreciated by (certain) readers.
Suspending the summation convention

An equation that violates the summation convention is usually a mistake,
often indicating a typographical error or perhaps an error in which a dummy
index was not changed to a different symbol prior to a substitution. Occasionally, however, one produces a legitimate equation that violates the summation
convention, in which case, a temporary suspension of the summation convention must be indicated. We have already encountered one instance of needing
to suspend the convention in our Eq. (4.17).
31
T
F
A
R
D
cca
Rebe
Brann

on
The summation conventions do not allow an index to appear three or more

times in a term. Thus, for example, the expression
(4.35)
1 p1 ( p1 a ) + 2 p2 ( p2 a ) + 3 p3 ( p3 a )

would have to be written in traditional form by explicitly showing the summation sign:
3
k p k ( p k a )
(4.36)
k=1
The summation symbol cannot be removed because the summation index k

appears three times. Here we had a legitimate expression in which an index
occurred three times and a summation was desired. Below we consider the
case when an index appears exactly twice and a summation is not desired.
Sometimes a legitimate indicial expression cannot be written in a manner
that satisfies the summation convention. In such a situation, the summation
convention must be temporarily suspended. For example, the i th eigenvector
{ p } i corresponding to the i th eigenvalue i of a matrix [ A ] is defined such
that
[ A ] { p }i = i { p }i
(4.37)
(no sum on i ).
The phrase no sum on i tells the reader that the author wishes the index i to
be interpreted as a free index even though it appears exactly twice on the
right-hand side.
In tensor analysis, exceptions to the summation convention are rare, so it
is a very convenient notational tool, especially when an expression contains
numerous implied summations, as was the case in Eq. (3.50).
Combining indicial equations

The following equations all satisfy the summation rules:
a i = A ij v j
b i = B ij w j
{a} = [ A]{v}
{b} = [ B]{w}
s = ai bi
s = { a }T { b }
(4.38)
The expressions on the right show the equivalent matrix expression for the
operations. Note, in the last equation, that a transpose of { a } is required in
the matrix equation. There is no need for a transpose on a i in the indicial
expression a iT is meaningless.
In the first two equations, the dummy summation index is j and the free
index is i ; hence, those equations actually represent three separate equations
for each value of the free index i . In the last expression, the dummy summation index is i , and there are no free indices (indicating that the equation is
just a single equation for a single scalar).
32
DRAF

Rebe
It might be tempting to directly substitute the first two expressions in

Eq. (4.38) into the third equation to obtain s = A ij v j B ij w j . However, this
combined equation violates the summation convention because there are now
four occurrences of the index j on the right-hand-side. To properly combine
the equations in (4.38), the dummy index j must be changed to something
else in one of the equations before combining the equations. The symbol used
in place of j is arbitrary, except that it must not be i because i is already
being used as a free index. Replacing j by k in the first of Eq. (4.38) gives
a i = A ik v k , allowing the equations to be combined to give
s = { v }T [ A ]T [ B ]{ w }
s = A ik v k B ij w j
(4.39)
which does not violate the summation convention. Written out in traditional
form, this equation may be written
3
s =
A ik v k B ij w j
(4.40)
i = 1k = 1 j = 1
In the above analysis, we showed the equivalent matrix expressions in the

right-hand column. The above analysis could have been done entirely using
matrix manipulations, but the skill of changing index symbols becomes indispensable when working with operations that have no straightforward matrix
analogs.
The index-changing property of the Kronecker delta

The indicial form of Eq. (3.20) may be written
ij v j = v i
cca B
ranno
(4.41)
Note that the Kronecker delta ij acts in a way that appears to have
changed the dummy summation index j on v j to become an i on v i with the
ij removed. This property holds because, in the summation over j , only one
term is nonzero (the one where i= j ).
This index-changing property holds in general. For example, the expression A ij v k ip can be simplified by noting that the subscript i on ip is a
dummy summation subscript. Therefore, ip may be removed if the other
occurrence of i is changed to a p . The simplified expression is therefore
A pj v k .
We have already used the index-changing property when we simplified
Eq. (3.31). The index-changing property applies in more complicated expressions involving multiple Kronecker deltas. Consider, for example, the expression T qms pk v q km id . Here the subscript k appears exactly twice, so it is a
dummy summation index, and the pk may be removed from the expression if
the other occurrence of k is changed to a p . This gives T qms v q pm id . Now we
33
T
F
A
R
D
cca
Rebe
Brann

on
note that the index m is a dummy sum index, so pm may be removed if the
other occurrence of m is changed to a p to give T qps v q id . The id in this
final expression may not be removed because both i and d are free indices,
not summation indices.
Now consider the expression mk A km . Noting that m is a summation
index, this may be simplified by removing the Kronecker delta mk if the other
occurrence of m is replaced by k to give A kk , which means A 11 + A 22 + A 33
and which is therefore equivalent to the trace operation of Eq. (3.15).
Summing the Kronecker delta itself

Consider the expression ik ik . The index i is a dummy sum index, so
exactly one ik may be removed if the other occurrence of i is changed to a k .
Thus ik ik may be simplified to become kk . Recalling that ij equals 1 if
i= j , one might be tempted to write kk =1 , which would be wrong. The index
k is still a summation index, and therefore
kk = 11 + 22 + 33 = 1 + 1 + 1 = 3
(4.42)
Viewed differently, this merely says that the trace of the 3 3 identity matrix
is equal to 3.
The under-tilde notation

Conventional publishing practice typesets scalars in italic while vectors
and tensors are set in bold. To add clarity to our discussion, we extend this
practice to explicitly indicate the order of vectors and tensors by the number of
under-tildes. Thus, for example, s denotes a scalar, v denotes a vector, T
denotes a second-order tensor (to be defined later), g denotes a third-order
tensor, and so on.
34
DRAF

Simple vector operations and properties
Rebe
Blessed are those who expect nothing,

for they shall not be disappointed.
Carl Sandburg
5. Simple vector operations and properties
Dot product between two vectors
Using the summation convention, the dot product between two vectors v
and w is
(5.1)
v w vk wk

Written out in full,
v w = v1 w1 + v2 w2 + v3 w3

In matrix notation, this can be written
v1
(v w)

= v2
v3
(5.2)
w1
w2 = { v } T { w }
w3
w1
= v 1 v 2 v 3 w 2 = <v> { w }
(5.3)
w3
The magnitude of a vector v is given by
2
2
v = v = + v 1 + v 2 + v 32 = + v v

Geometrically, the dot product can be written
(5.4)
v w = vw cos vw ,
(5.5)
v

where v and w are the magnitudes of v and w , respec-
tively, and vw is the angle between v and w. The dot
product is commutative:
vw = wv

It also satisfies the inner product positivity rule
cca B
ranno
vw
(5.6)
35
T
F
A
R
D
cca
Rebe
Brann

on
v v > 0 if v 0

v v = 0 if and only if v = 0
(5.7)

This positivity property implies that the magnitude of a vector, as defined in

Eq. (5.4), will always be real.*
Dot product between orthonormal base vectors

If { e 1, e 2, e 3 } is an orthonormal basis, then

e i e j = ij ,

where (recall) the symbol ij is the Kronecker delta, defined
ij = 1 if i = j
0 if i j
(5.8)
(5.9)
Finding the i-th component of a vector

Recall that any vector v can be expressed as
v = v1 e1 + v2 e2 + v3 e3
(5.10)
Dotting both sides with e 1 gives v e 1 = v 1 . Similarly, v e 2 = v 2 and
of a vector with respect to an orthonorv e 3 = v 3 . Hence, the i th component

basis { e , e , e } can always be found by
mal
1 2 3
vi = v ei
(5.11)

This relationship can be derived using strictly indicial notation by noting that
Eq. (5.10) is equivalent to
v = vk ek
(5.12)
Dotting both sides by e i , invoking Eq. (5.8), and finally using the index-chang
ing property of the Kronecker
delta gives
v e i = v k e k e i = v k ki = v i
(5.13)

which is equivalent to Eq. (5.11). This method for finding the component of a
vector might seem at first glance to be trivial and of no obvious use. Suppose
however, that { e 1*, e 2*, e 3* } is a different basis. Further suppose that we do
v with respect to the original (unstarred) basis, but
of
know the components
we wish to determine the components of this vector with respect to this new
(starred) basis. Then v*i = v e*i . As a specific example, suppose that

* Keep in mind that we are considering only ordinary engineering vectors having real components. If
complex components were allowed, the inner product would be written v w v k w k , where the

overbar denotes the complex conjugate.
36
DRAF

Rebe
1
--- ( 3e
1
5
1
--- ( 4 e
1
5
e 1* =
+ 4e 2 ) , e 2* = e*3 , and e 3* =
+ 3e 2 )
(5.14)
Suppose that we know that, with respect to the unstarred basis the vector v is
given by
v = 2e 1 + 5e 2
(5.15)
Then the components with respect to the unstarred basis are just v 1 = 2 ,
v 2 = 5 and v 3 = 0 , whereas the components of this vector with respect to the
starred basis are
6 20
26
1
v 1* = v e 1* = ( 2e 1 + 5e 2 ) --- ( 3e 1 + 4e 2 ) = --- + ------ = -----5
5 5
5

(5.16)
v 2* = v e 2* = ( 2e 1 + 5e 2 ) ( e*3 ) = 0

(5.17)
8 15
7
1
v 3* = v e 3* = ( 2e 1 + 5e 2 ) --- ( 4 e 1 + 3e 2 ) = ------ + ------ = --5
5
5
5

(5.18)
The method presented in this section works only when the new basis is
orthonormal; see page 51 to learn how to find the components of a vector with
respect to an irregular (non-normalized and/or non-orthogonal) basis.
Even and odd vector functions

If a function f ( v ) has the property that f ( v ) = f ( v ) , then the function is
it is independent of the vectors directional sense. For
said to be even, and
example, the operation v v is even. On the other hand, if f ( v ) = f ( v ) ,
be
odd.
then the function is said to

Any function f ( v ) may be decomposed into its even part f e ( v ) plus its odd
part f o ( v ) as
f (v) = f e(v) + f o(v)

(5.19)
where
1
f e ( v ) --- [ f ( v ) + f ( v ) ]
2
1
f o ( v ) --- [ f ( v ) f ( v ) ]
2
(5.20)
Homogeneous functions
A function f is said to be homogenous of degree k if it satisfies
f ( v ) = k f ( v ) for all positive scalars .
(5.21)
For example, f ( x ) = x 3 is homogeneous of degree 3. The function

f ( x ) = x x is homogeneous of degree zero. The function f ( x ) = 1 + x is not
homogeneous of any degree.
When a function is homogenous of degree k, then can be shown [__] that
f ( v ) v = kf ( v )
cca B
ranno
(5.22)
37
T
F
A
R
D
cca
Rebe
Brann

on
where f ( v )= ( f v i )e i . Furthermore, the function f ( v ) is homogeneous of
That is,
degree k1.
f ( v ) = k 1 f ( v )
(5.23)
Vector orientation and sense

The term orientation will be used to describe the line that is parallel to a
given vector. The terms direction or sense will be used to describe the
direction that a vector points along its orientational line. Thus, for example,
the direction of v is opposite that of v whereas the orientations of the two
because they lie in parallel directions.
vectors are identical
Suppose that a function f ( v ) is homogeneous of degree zero. In other
we are supposing that a function satisfies
words, putting k=0 in Eq. (5.21),
the property that, for all scalars , f ( v ) = f ( v ) . Then the function f ( v )
on its magnitude or sense.
depends only on the vectors orientation, not
Simple scalar components

The vector inner (dot) product is extremely useful for extracting parts of vectors (or tensors) in particular directions.
v
w v

(a)
vw

v2
e2
e1
v1
(b)
(c)
Figure 5.1. Finding components via projections. (a) scalar component of v in the direction of w , (b) scalar component of w in the direction of v , and (c) the scalar components of
v in the directions of the base vectors.
The operation
w
---------where w
(5.24)
gives the (rectangular) component of v in the direction of a second vector w .
Since we are only talking about the direction

of w , the actual magnitude of w
that
is irrelevant thats why the formula depends only
on the unit vector w
points in the direction of w . The formula of Eq. (5.24) would not change if we
positive scalar . The result changes sign if w is
were to multiply w by any
replaced by w .
,
vw

38
DRAF

Rebe
Just as v can be projected onto w , the reverse operation of projecting w in

of v gives the component
of w in the direction of v .
the direction
v
(5.25)
where v -------w v ,
v

This operation depends on the direction of v , but not its magnitude.
Eq. (5.11) is a special case of Eq. (5.24) in which w is e i . Thus, v i is the
projection of v in the i th coordinate direction.
Cross product
The cross product between two vectors, a and b
is a vector defined as:

,
(5.26)
a b = ab ( sin ab )n

is perpendicular to both a and
The unit vector n
defined by the right hand
rule
b , with a direction
when
sweeping from a to b through the angle
ab . When defined in this

manner,
the angle ab
does not have to be less than 180 , though the
smaller angle between a and b is conventionally
used.
The parallelogram in the illustration has an
orientation perpendicular to a b . The area of

the parallelogram equals the magnitude
of a b .

If u = a b , then the components of u are
u1 = a2 b3 a3 b2
u2 = a3 b1 a1 b3
u3 = a1 b2 a2 b1
ab
a
Figure 5.2.
Cross product
(5.27)
Hence,
(5.28)
a b = ( a 2 b 3 a 3 b 2 )e 1 + ( a 3 b 1 a 1 b 3 )e 2 + ( a 1 b 2 a 2 b 1 )e 3

Heuristically, this equation is often written as the determinant of a matrix:
ab

e1 e2 e3

det a 1 a 2 a 3
cca B
ranno
(5.29)
b1 b2 b3
Cross product between orthonormal base vectors

Applying the definition of the cross product reveals that the nine possible
ways to cross the orthonormal base vectors with themselves give
39
T
F
A
R
D
cca
Rebe
Brann

on
e1 e1 = 0
e2 e1 = e3
e3 e1 = e2
e1 e2 = e3
e2 e2 = 0
e3 e2 = e1
e1 e3 = e2
e2 e3 = e1
e3 e3 = 0
(5.30)
These expressions may be written compactly using the summation notation as

e i e j = ijk e k
(5.31)
where ijk is the permutation symbol defined in Eq. (3.21).

We may use Eq. (5.31) to deduce the component (indicial) expression of the
cross product. Namely,
a b = ( a i e i ) ( b j e j ) = a i b j ( e i e j ) = a i b j ijk e k

(5.32)
In the last step, we applied Eq. (5.31). Noting that the final result is the sum
over k of an expression times e k , the k th component of the cross product oper
ation must therefore be
( a b ) k = a i b j ijk

(5.33)
This formula relies on our previously stated assumption that all component formulas in this report are referenced to an orthonormal right-handed
basis. If you use an orthonormal left-handed basis, then the above formula
would be ( a b ) k = a i b j ijk . Some authors take this situation as an implica
tion that the permutation symbol for left-handed base vectors should be
defined as the negative of the right-handed definition. This is misleading and
wrong-headed. Whats really going on is that the cross product operation can
be most generally written as a special third-order alternating tensor operat
ing on the vectors a and b . [Namely, using notation to be defined later
in this
report, a b = : ( a b ) .] The components of the alternating tensor can be

proved to equal = if the basis is orthonormal and right-handed, while
ijk
ijk
ijk = ijk if the basis is orthonormal but left-handed. The ijk components
have yet a different form if the basis is non-orthogonal or non-normalized.*
Note that we keep the definition of the permutation symbol unchanged only
the components of change form upon a change of basis. We reiterate that,
although the form of

the component formulas for a b depend on the nature

* Namely, if { g 1, g 2, g 3 } is a linearly independent set of vectors, then ijk = g i ( g j g k ) .
40
DRAF

Rebe
of the basis, the final result for a b is the same in all cases. In other words,

a b represents a particular and unique vector that does not depend on the

underlying basis. Only the method that you must use to determine this unique
vector varies depending on the choice of basis.
Triple scalar product

Given three vectors, a , b , and c , the triple scalar product is defined

[ a, b, c ] ( a b ) c
(5.34)

It can be shown that the triple scalar product has the same value if the vectors
are positively permuted and it changes sign if the vectors are negatively permuted. Specifically,
and
[ a, b, c ] = [ b, c, a ] = [ c, a, b ]

cca B
ranno
(5.35)
[ a, b, c ] = [ c, b, a ] = [ b, a, c ] = [ a, c, b ]
(5.36)

Physically, the absolute value of [ a, b, c ] equals the volume of the parallelepi of [ a, b, c ] will be positive if the vectors
ped formed by a , b , and c . The sign

meaning that when
the thumb of the right hand
form a right-handed
triad,
points in the direction of a and the index finger points in the direction of b
then the middle finger will point roughly in the direction of c (i.e., if the mid with c ). If the
dle finger were a vector, it would have a positive dot product
sign of [ a, b, c ] is negative, then the triad is said to be left-handed.

Triple scalar product between orthonormal RIGHT-HANDED

base vectors
Dotting Eq. (5.31) by e m gives
( e i e j ) e m = ijk km = ijm
(5.37)
where ijk is the permutation symbol defined in Eq. (3.21). Expressing the
above result with the free index m replaced by k gives
ijk = [ e i, e j, e k ]
(5.38)

The triple scalar product [ e i, e j, e k ] is certainly well-defined if the basis is
is
[ e , e , e ] = . As mentioned earlier, this
left-handed; in fact, the result
ijk
i j for
k redefining
fact should not be used as justification
the permutation symbol
when using a left-handed basis. We recommend always defining the permutation symbol such that 123 =+1 even if the basis is left-handed.
Even though the term tensor has not yet been defined, its worth mentioning here (for
future reference) that there is a straightforward formula for constructing a third-order permutation tensor in terms of any basis including left-handed and even non-orthonormal. The
permutation tensor components with respect to a non-orthonormal basis take yet a different
41
T
F
A
R
D
cca
Rebe
Brann

on
form (covered in the more advanced Ref. [4]), and the process of computing the triple scalar
product becomes more complicated. Nonetheless, the basic concept is the same: the triple scalar product of three vectors equals the triple inner product of the alternating tensor into the
dyadic multiplication (defined later) of the three vectors. It just happens that the components of the alternating tensor equal the permutation symbol when the basis is orthonormal
and right-handed. It is for this reason that we denote the alternating tensor by a symbol ijk
that differs from its right-handed orthonormal components ijk . A similar notational choice
is also made when we identity tensor, I , by a symbol that differs from its right-handed
orthonormal components (the Kronecker delta

ij ).
Axial vectors
Even though we have not yet covered the meaning of the word tensor,
this is still the most appropriate place (for future reference, once the meaning
of the word tensor is learned) to discuss axial vectors and axial tensors.
Given any vector a , we can always construct a skew-symmetric axial tensor
defined by
A a
(5.39)

We have not yet defined what is meant by the word tensor, nor have we
defined what is meant by the above tensor notation. For now, A should be
regarded as a 3 3 matrix whose components are constructed by the formula

A ij = ijk a k
(5.40)
Expanding out the implied summation over the index k (noting that all terms
for which i= j are zero, the matrix for the axial tensor is related to the components of the vector a according to
0 a3 a2
[ A] =
a3
0 a1
a2 a1
(5.41)
Conversely, given any tensor B , the associated axial vector can be constructed
by
1
a = --- : B
2
In component form, this notation means

1
a k = --- kij B ij
2
(5.42)
(5.43)
Expanding out the summations over i and j shows that

1
a 1 = --- ( B 32 B 23 )
2
(5.44)
42
DRAF

Rebe
1
a 2 = --- ( B 13 B 31 )
2
(5.45)
1
a 3 = --- ( B 21 B 12 )
2
If A is the skew-symmetric part of B , then
a 1 = A 32
cca B
ranno
(5.46)
(5.47)
a 2 = A 13
(5.48)
a 3 = A 21
(5.49)
Glide plane expressions. The mathematical description of any problem that

involves material shearing on a plane frequently ends up with dyadic expressions of the form b c c b . We have not yet defined what a dyad or a dyadic
this particular dyadic as a special skew symmetric
means. For now, regard
tensor, and its associated axial vector is the cross product between b and c .
That is,
(b c) = bc cb
(5.50)

We placed this result here because of its intimate connection with cross products. Expressions like this show up frequently in single crystal plasticity.
43
T
F
A
R
D
cca
Rebe
Brann

Projections
on
I am always ready to learn, although

I do not always like to be taught
Winston Churchill
6. Projections
Orthogonal linear projections
Equation (5.24) shows how to find the scalar component of v in the direc obtain a vec , then we would
tion of w . If we were to multiply the result by w
< 0 , a vector that is in the opposite
tor thatis in the direction of w or, if v w
we
multiply the component of v in the
direction of w . This operation, in which
by w
itself is called the orthogonal vector projection of v onto
direction of w
denote
this vector by v Dw (which may be read as the part
of v
w , and we will
the direction of w ). The mathematical
in
definition is
w
Dw
)w
---------where w
v
= (v w
(6.1)
For some applications, it is prudent to recognize that the projection operation involves two vectors, v and w , so it is sometimes advisable to instead
define a binary* operator D such that D ( v, w ) denotes the part of v in the
direction of w . In terms of this alternative notation,

Eq. (6.1) would read
vw
(6.2)
D ( v, w ) = ------------- ww

Clearly,
D ( v, w ) D ( w, v )

(6.3)
* The term binary is merely a fancy way of saying that the operator has two arguments.
44
DRAF

Projections
Rebe
v Dw
The part of the vector v that is left over i.e., v

is the part of the
denote by v Pw or by
vector v that is perpendicular
to w , which we will
P ( v, w ). It is defined by

)w
v Pw = v ( v w
(6.4)

or
vw
(6.5)
P ( v, w ) = v -------------
ww

Note that v Pw lies in the plane perpendicular to w . Also note that
P ( v, w ) + D ( v, w ) = v
(6.6)

This equation shows more clearly that the operators P and D decompose the
vector v into two parts, one parallel to w and the remainder perpendicular
to w .
To make a sketch of v Dw , simply draw a line parallel to w that passes

draw a line perpendicular to w (or, for 3D probthrough the tail of v . Then
lems, draw a plane perpendicular to w ) that passes through the tip of the vector v . These two lines (or, for 3D problems, the line and the plane) will
intersect
at a single point, which we will call A . Then v Dw is the vector
extending from the tail of v to A and v Pw is the vector extending

from A to
the tip of v . See Fig. .
(b)
(a)
tip
(c)
A
v Pw
w Dv
A
tail
v Dw
to
lines
parallel
w Pv
lines
parallel
to v
tail
tip
tip
v De2
e2
tail
e1
v De1
Figure 6.1. Vector decomposition (a) decomposition of v in directions parallel

and perpendicular to w , (b) decomposition of w into parts parallel and perpendic
ular to v , and (c) the projection

of v onto the unit
base vectors.
As mentioned earlier, D ( v, w ) D ( w, v ) . Nonetheless. The vector w may

that are
in parallel and perpendicular
to v
be similarly decomposed into parts
by the formulas
and
w Dv = D ( w, v ) = ( w v )v

Pv
w
= P ( w, v ) = w ( w v )v

v
where v -------v
cca B
ranno
(6.7)
45
T
F
A
R
D
cca
Rebe
Brann

Projections
on
Figure clearly shows that this decomposition of w results in completely different vectors as the decomposition of v .
The last sketch in Fig. shows the decomposition of a vector v into parts
that are aligned with the base vectors. In particular, note that we labeled the
vertical part of the vector as v De2 because, in three dimensions v De2 v Pe1 .
Suppose, for example, that the vector is v = 6e 1 + 3e 2 + 7e 3 . Then
(6.8)
v De2 = 3e 2
whereas
v Pe1 = 3e 2 + 7e 3 .
(6.9)
Rank-1 orthogonal projections

The notation a Db symbolizes the orthogonal projection of a vector a in the
direction of a second vector b . Since this operation involves two vectors, you
might alternatively prefer to denote it in a more standard way by defining a

binary operation D such that D ( a, b ) represents the projection of a onto a

second vector b . The result is always going to be a vector parallel to b . The
adjective orthogonal is used to indicate that each point on the vector a maps
to the closest point on a line parallel to b ; therefore, the projection direction is
a family of lines that are orthogonal to b . The set of all vectors parallel to b
forms a one-dimensional subspace because any member of the set may be written in the form b , which involves only one arbitrary scalar .
If n is a unit vector, then any vector x can be projected into a part that is
to n and simultaneously perpendicular
parallel
to the plane whose normal is
(6.10)
p = n(n x)

In other words, p is just x Dn . This projection has rank 1 because the target
line parallel to p . If one were to compute the
space is the one-dimensional

the result would be the vecaverage of x over all rotations of x about n , then
tor p .
46
DRAF

Projections
Rebe
Plane perpendicular to
(b)
(a)
Figure 6.2. (a) Rank-1 orthogonal projection, and (b) Rank-2 orthogonal projection.
The dashed lines show the projection direction. For the Rank-1 orthogonal projection, the
dashed lines point from a location on x to a target line that is parallel to the unit vector n
(the target line can be translated so that

it passes through any point in the plane without
changing the result for the projection vector p . For the Rank-2 projection, the dashed
lines point from a location on x toward the shortest

distance to the plane if the plane is
x . The projections are orthogonal because the path from
positioned at the tail of the vector
the vector x to the target line or plane

is the shortest distance.
Rank-2 orthogonal projections

As sketched in Fig. 6.2, any vector x can be projected onto a plane with
unit normal n by using the formula
q = x n(n x) .
(6.11)

In other words, q is just x Pn . This transformation is called an orthogonal pro

projected
jection because the
vector q represents the shadow cast by x onto
the plane perpendicular to n when the light rays (which are formally called
level lines) are coaxial with n . (Note: the term coaxial here means aligned,
sense.) The projection is second-rank because
but not necessarily of the same
the target space to which the vector is being projected is two dimensional (a
plane).
Basis interpretation of orthogonal projections

The vector n can have any orientation. In order to better interpret equations like (6.10) and (6.11), it is often useful to set up a basis that has, say, its
e 3 direction aligned with n . If we align e 3 with n , then the plane in Fig. 6.2,
would
be the 1-2 plane. Substituting
n=e 3 into Eqs. (6.10) and (6.11) and

applying Eq. (5.11) gives
p = e3 ( e3 x ) = x3 e3 .

q = x x3 e3 = x1 e1 + x2 e2 .
cca B
ranno
(6.12)
(6.13)
47
T
F
A
R
D
cca
Rebe
Brann

Projections
on
Thus, the projected vector p is the part of x that is in the 3-direction, and q
many engineering applications,
is the part of x that is in the 1-2 plane. In
there exist multiple

planes to which vectors are projected, so it is not always
possible to merely set up a single basis that is lined up with all of the planes of
interest thats why our direct notation equations (6.10) and (6.11) are more
convenient. Furthermore, the vector x and the planes normal n might be
notation formula of Eq. (6.11)
changing in time. In this case, the direct

is most
convenient because it allows direct computation of rates.
The concept of a projection will now be generalized to permit us to find the
late afternoon shadow of the vector x when the light rays are oblique to
the plane.
Rank-2 oblique linear projection

Figure 6.2(b) showed how to find the
shadow of a vector at high noon, when the
x
light rays were perpendicular to the target
plane. As sketched in Fig. 6.3, a more genb

a
eral kind of projection allows the light rays
to intersect the plane at an oblique angle.

This kind of projection can be characterized
q
via two vectors a and b . The target plane is

Plane perpendicular to b
defined to be perpendicular to b , and the
light rays are parallel to a . The magnitudes

Figure 6.3. Oblique projection.
of these vectors are inconsequential since The paths of projection (dashed

are all parallel to a , so they
they are only being used to define directions. lines)
obliquely intersect the plane.

The
q
result
for
is
unaffected
by transAs before, we seek an expression for the
plane.
lation of the
vector q that is the projection of x onto the
the projection direction (the light rays)
plane. This
time, however, we want
to be aligned with a different vector a . Referring to Fig. 6.3, we can see that
there must exist a scalar coefficient such

that the vector x can be written as
the sum of q plus some (as yet) unknown multiple of a :
(6.14)
x = q + a ,
To find the value of , we impose the condition that the vector q must lie in
the plane. In other words, b q must be zero if we want q to be perpendicular
to b . Dotting both sides of Eq.(6.14) by b (and setting b q = 0 ) gives

(6.15)
b x = 0 + (b a) .

Solving for and substituting the result into (6.14) gives the desired formula
for the oblique projection:
48
DRAF

Projections
Rebe
a(b x)
Q ( x ) = x --------------------(6.16)
- .
ab

Naturally, Eq. (6.11) is a special case of the more general Eq. (6.16), obtained
by choosing a =b =n . In other words, the projection is orthogonal only if a is
b
.
proportional to
q = Q ( x ) , where
Rank-1 oblique linear projection

The remaining part of x is obtained by a
p = x q . Thus,
rank 1 projection. Namely,

a(b x)
p = P ( x ) , where P ( x ) = -------------------- - . (6.17)
ab

The result of this operation will always be a
vector that is simply a scalar multiple of a .
This target space is one dimensional, so the

projector P has rank 1.
cca B
ranno
Note that the operator P has the prop- Figure 6.4. Rank-1 oblique proThe path obliquely intererty that P ( P ( x ) ) = P ( x ) . Physically, this jection.
sects
the
plane.
a vector that is
means that projecting
already in the target space will just give you back that vector unchanged (your
shadow has no shadow other than itself). This property is, in fact, used to
define the term projection. The operator Q has the similar property that
Q ( Q ( x ) ) = Q ( x ) , so it is therefore a projector.
Complementary projectors
The operator P is called the complement of Q . In general, two operators P
and Q are called complementary projectors if the following properties hold
P( P( x)) = P( x)
(6.18)
(6.19)
Q(Q( x)) = Q( x)
(6.20)
Q( P( x)) = P(Q( x)) = 0
P( x) + Q( x) = x
(6.21)
The last property states that the sum of P and Q must be the identity operation (this is an abstract way of saying that the P and Q operators permit an
additive decomposition of the vector x ). The rank of P plus the rank of Q
must sum to equal the dimension of the space, which is 3 for ordinary physical
vectors.
49
T
F
A
R
D
cca
Rebe
Brann

Projections
on
Normalized versions of the projectors

The projectors Q and P in Eqs. (6.16) and (6.17) are homogeneous of
degree zero with respect to a and/or b . (i.e., multiplying either of these vec not affect
the outcome of the operations). This
tors by any nonzero scalar does
property does not mean that the projection formulas depend on only the angle
between a and b . The scaling property simply implies that the formulas are
to the
magnitudes and senses of a and b only their orientainsensitive
tions matter.
Because the magnitudes of a and b are arbitrary, it is convenient in cer with a
and b , which have the same orientatain applications to replace them
satisfy
tions as a and b , but whose magnitudes
b =1
a
(6.22)

One possible way to construct a and b is
a
b
= ---------------a
b = ---------------(6.23)
and

ab
ab

and b are called dual vectors.
When the property of Eq. (6.22) holds, a
Using the dual vectors, Eq. (6.16) simplifiesto

( b x )
q = xa
(6.24)

The component form of Eq. (6.16) take a particularly simple form when
expressed using a nonorthogonal basis having the covariant* base vector
g 3 = a and the contravariant base vector g 3 = b . Namely
1
2
q = x g1 + x g2 ,
(6.25)
This result is strongly analogous to Eq. (6.13), with the key difference being
that the g k base vectors are nonorthogonal and not generally of unit length.
The projector
Q is a rank-2 projector because its range space is two dimensional.
The null space of the projector Q is the set of all x for which Q ( x ) = 0 .
Setting Q ( x ) = 0 in Eq. (6.16) reveals that the null space

of that particular
of vectors that are proportional to a . Thus, the null space

projector is the set
the target plane) is
of the projector is one-dimensional. The range space (i.e.,
* When a basis { g 1, g 2, g 3 } is nonorthonormal, there naturally arises a useful companion or dual

basis denoted { g 1, g 2, g 3 } which is defined such that g i g j = ij , where ij is the Kronecker

delta of Eq. (5.9). Thus, for example, g 1 is perpendicular to the plane formed by g 2 and g 3 . A
vector can be expressed in terms of either basis as v = v i g i = v i g i . Subscripted quantities are
called covariant and superscripted quantities are contravariant. (A cute mnemonic is cogo-below). Non-orthonormal tensor analysis is reviewed in Refs. [5,4].
50
DRAF

Projections
Rebe
two dimensional. Since P and Q are complementary projectors, the null space
of one is the range space of the other! Later on, projectors will be defined for
higher-dimensional spaces and the null spaces will be more interesting.
Figure (6.5) shows two vectors, x and y ,
that fall on the line defined by a . More

pre
cisely, for some scalar ,
x = y + a .
(6.26)
As seen in the sketch (or as verified by direct

substitution into Eq. (6.16)),
Q ( x )=Q ( y ) .
(6.27)
Conversely, if (6.27) holds, then so does

(6.26). This property of projections is the key
to validating radial or oblique return algorithms in the field of plasticity.
cca B
ranno
Q ( x )=Q ( y )
Figure 6.5. Projections of two

vectors along a an obliquely oriented line.
Expressing a vector as a linear combination of three

arbitrary (not necessarily orthonormal) vectors.
Any vector x can always be expressed as a linear combination of three
other vectors a , b , and c as long as those three vectors are linearly indepen as the triple scalar product is nonzero: [ a, b, c ] 0 ). If so,
dent (i.e., so long

then we know there exist scalar multiples , , and such that
x = a + b + c
(6.28)
The brute force simplest way to determine values of the scalar coefficients is to
simply dot both sides of this equation by a to obtain
x a = (a a) + (b a) + (c a)
(6.29)

We can similarly dot both sides of Eq. (6.28) by b and then separately by c to
(6.29) forms a set of three
obtain two additional equations which, with Eq.

equations and three unknowns (the { , , } coefficients). The system is solvable if and only if the original three vectors are linearly independent. We will
now discuss this identical process using a slight change of notation so that it
become more clear that each component actually corresponds to a rank-1
oblique projection of the vector x onto the three vectors. To simplify the nota c by three different symbols, { g , g , g } .
tion, we will now denote a , b , and
1
2
3

The right-hand side of Fig. 6.1 showed how an arbitrary vector x may be
decomposed into the sum of parts that are aligned with the orthonormal
{ e 1, e 2, e 3 } basis. However, a basis does not have to be orthonormal any
linearly
three
independent vectors { g 1, g 2, g 3 } may be alternatively used as a

51
T
F
A
R
D
cca
Rebe
Brann

Projections
on
basis. We wish to additively decompose a vector x into parts aligned with
these three non-orthonormal base vectors. Stated differently,

we wish to discover a path from the tail to the tip of x that always travels parallel (or anti Fig. 6.6).
parallel) to these three base vectors (See
P3 ( x )
g2
g3
P2 ( x )
angles 90
g1
P1 ( x )
Figure 6.6. Three oblique projections. The indicated plane contains g 1 and g 2 ,
to
while g 3 forms an oblique angle to the plane. The dual vector g 3 is proportional
g 1 g 2 , so it may be used to define the normal to the plane.
First, in order for the three { g 1, g 2, g 3 } vectors to be linearly independent

have
(and therefore form a basis), they must

a positive triple scalar product:
(6.30)
J o [ g 1, g 2, g 3 ] = ( g 1 g 2 ) g 3 > 0

Corresponding to the { g 1, g 2, g 3 } base vectors, there exists a unique

that

dual basis { g 1, g 2, g 3 } such

(where ij is an alternative symbol for the Kronecker delta).
(6.31)
g i g j = ij
In other words, g 1 must be perpendicular to g 2 and g 3 , so it must be propor ) . The constant of proportionality
must be set so that

tional to ( g 2 g
3
as can be verified by direct substitution, the dual basis
g 1 g 1 = 1 . Thus,
be given by
must
1
1
1
g 1 = ------ ( g 2 g 3 ) , g 2 = ------ ( g 3 g 1 ) , g 3 = ------ ( g 1 g 2 )
(6.32)
Jo
Jo
Jo
Without proof, we claim that the part of x in the direction of g k is given by an
oblique projection operation P k defined by

P k ( x ) = g k ( g k x ) , with no summation on k .
(6.33)

The three P k operators are complementary projections because they satisfy
the following properties:
P i ( x ) if i= j
Pi ( P j ( x ) ) =
if i j
0
(6.34)
P1 ( x ) + P2 ( x ) + P3 ( x ) = x
(6.35)
and
52
DRAF

Projections
Rebe
cca B
ranno
Generalized projections
A transformation P is called a projection (or idempotent) if and only if
P( P( x)) = P( x) ,
(6.36)
which merely states that projecting a vector that has already been projected
will give you the same projected vector right back. Consider, for example, the
function of Eq. (6.16):
a(b x)
P ( x ) = x --------------------(6.37)
ab

Physically, we know that this function is a projector. To prove it rigorously, we
operate on both sides of Eq. (6.37) by P to obtain
a(b x)
a b x -------------------- -
ab

a(b x)
P ( P ( x ) ) = x --------------------(6.38)
- -------------------------------------------------------ab
ab

Simplification shows that the right hand side reduces back to P ( x ) . Hence,
Eq. (6.36) holds and P is indeed a projection.
Linear projections
A projection is linear if and only if
(i) P ( x ) = P ( x ) , for all , and
(ii) P ( x 1 + x 2 ) = P ( x 1 ) + P ( x 2 ) for all x 1 and x 2 .

(6.39)
The projection of Eq. (6.37) is easily seen to be linear even though it is oblique.
Likewise, Eq. (6.33) is linear. Physically, a projection is linear if one can take
the projection of a linear combination of vectors and the result is the same as
first projecting the vectors and then taking their linear combination.
Nonlinear projections
An example of a nonlinear projection is
x
P ( x ) = ---------------(or P ( x )=0 if x =0 )
(6.40)

xx

This function is indeed a projection because Eq. (6.36) holds. The first linearity test (i) in Eq. (6.39) fails because, for example, P ( 2 x ) 2P ( x ) , so this pro
a vector to a
jection is nonlinear. Geometrically, the above operator transforms
unit vector in the same direction. Hence, the range set for this transformation
consists of vectors pointing from the origin to a point on the unit sphere. The
transformation of Eq. (6.40) is non-linear because the range (the target surface of the sphere) is geometrically curvilinear. The level lines (i.e., the light
rays) are straight lines that emanate radially from the origin.
53
T
F
A
R
D
cca
Rebe
Brann

Projections
on
Nonlinear projectors can also be constructed for which the range is a linear
manifold (i.e., a line or a plane) but the path to this range space is curvilinear.
For example, the transformation
[ x a ( x b ) 2 ]a
P ( x ) = -----------------------------------------------(6.41)

a (a b)

is a projection because P ( P ( x ) ) = P ( x ) . It has a linear range space (vectors
parallel to a ), but the projection is nonlinear because the paths to the range
space (i.e., the level lines) are curvilinear. The light rays are bending as they
approach the target space.
A very important nonlinear projection transformation from the field of continuum mechanics is the polar decomposition (see page 136).
Self-adjoint projections
A projection P is self-adjoint if and only if
u P ( v ) = v P ( u ) for all u and v .
Applying this definition to Eq. (6.37) gives
(6.42)
a(b v)
a(b u)
u v -------------------- = v u -------------------- ab
ab

Simplification gives
(6.43)
( u a ) ( b v ) = ( v a ) ( b u ) for all u and v .

(6.44)

This condition can be true only if a is proportional to b ,* in which case the
projection reduces to the orthogonal projection of Eq. (6.11). The P k projectors

defined in Eq. (6.33) are not self-adjoint unless the { g k } basis is orthonormal.
Physically, self-adjoint projections correspond to projections

to the nearest
point in the target space (like a shadow at high-noon), whereas non-selfadjoint projections are oblique (like a shadow in late afternoon).
Later on, we will see that linear projectors can be represented through the
use of projection tensors. Then we will find that a linear projector is selfadjoint only if its projection tensor is symmetric.
* To prove this claim, take u b and v =b to obtain ( u a ) ( b b ) = 0 or, since b 0 by implied

premise, u a = 0 . In other words, u a for all u b . Thus, a must be proportional
to b .
54
DRAF

Projections
Rebe
The projection theorem

Some of the most profoundly useful engineering theorems merely state intuitively
obvious concepts in a more useful (but less
obvious) mathematical form.*
The projection theorem is another mathematical statement of a physically obvious
concept. Namely, you can always break a
vector into a part in a desired direction plus
whatever is left over. Now lets state this
idea mathematically.
Given a projector P , the projection theorem states that any vector x can always be
decomposed uniquely as
xP
xQ
Figure 6.7. Oblique projection.

The path obliquely intersects the
plane.
x = x P + xQ
(6.45)
such that x P is in the range space of the projector P and x Q is in the null satisfy
space of the projector P . By this we mean that x P and x Q will
P( xP) = xP
(6.46)
P( xQ ) = 0
(6.47)
The vector x P is the projection of x , and x Q is simply the part of x that is
taking away x P . Specifically,

left over after
xP = P( x)
(6.48)
(6.49)
xQ = Q( x ) ,
where
Q( x) = x P( x)
(6.50)
The function Q is itself a projector. Some key properties of these complementary projectors are
P( P( x)) = P( x)
Q(Q( x)) = Q( x)
P(Q( x)) = Q( P( x)) = 0
P( x) + Q( x) = x
cca B
ranno
(6.51)
* For example, the principle of conservation of mass merely states that what goes in must come out
or stay there. The mathematical representation of this statement is less obvious.
55
T
F
A
R
D
cca
Rebe
Brann

Projections
on
Note that P ( Q ( x ) ) is not the same thing as x P x Q . If the projection is

oblique, then x P x Q will be nonzero, but P ( Q( x ) ) will always be zero for
complementary projectors.
The transformation P (x ) is an orthogonal projection if and only if P is self-adjoint, in which case x P x Q will also be zero.
The projection theorem is just the mathematical way to describe what is

fairly obvious from looking at Fig. 6.7 (i.e., a vector can be broken into its
shadow plus whatever is left over). The projection theorem becomes quite useful (and less obvious) in higher dimensional spaces. Later in this report, we
will introduce the notion of an abstract vector. Under this definition, we will
find that matrices, differentiable functions and many other seemingly unrelated things behave just like vectors. The fact that a scalar function f ( x ) can
be written as the sum of its even plus odd parts is an application of the projection theorem. The fact that a matrix can be written as the sum of its symmetric plus skew-symmetric parts is an application of the projection theorem. The
fact that a continuously differential function can be expressed as a Taylor
series is the projection theorem (the basis is the infinite set 1, x, x 2, x 3, ).
Most material constitutive laws are expressible in terms of projections. For
example, the projection theorem allows us to decompose stress and strain tensors into their deviatoric plus isotropic parts. Hookes law of linear elasticity
states that these individual parts are proportional to each other (i.e., pressure
is proportional to volumetric strain and the stress deviator is proportional to
the strain deviator). The proportionality constant is not necessarily equal for
the two parts, so stress itself is not proportional to strain.
56
DRAF

Tensors
Rebe
cca B
ranno
If you live to the age of a hundred, youve got it

made, because very few people die past the age
of a hundred
George Burns
7. Tensors
If youre sitting at a cocktail party with a bunch of engineers, physicists,
and mathematicians, and you want to start a heated debate, just ask out loud:
What is a tensor? One person will say that, for all practical purposes that a
tensor is just a fancy word for a matrix.* Then someone else will pipe up indignantly and insist that a tensor is a linear transformation from vectors to vectors. Yet another person will say that a tensor is an ordered set of numbers
that transform in a particular way upon a change of basis. In this chapter, we
will touch on each of these various perspectives and urge the reader to adopt
whatever definition of the term tensor they feel most comfortable with. As
long as your definition implies the other guys definition and vice versa, then
either definition is acceptable.
You cant study tensors without eventually (sometimes secretly) becoming
enamoured of one particular definition of tensors. We will ease the reader into
the notion of tensors by discussing how engineering tensors arise naturally
when working with linear vector transformations. We will then describe our
personal favorite (and less frequently adopted) definition of a tensor as being a
sum of new abstract objects called dyads. Admittedly, the dyad definition is
not very satisfying or particularly intuitive, but we like it anyway because this
definition requires no mention of a basis, per se, but it leads directly into the
supplemental concept of a tensor basis. From there, we will be poised to then
cover the definition of a tensor in terms of basis transformation rules.
* This is true for some practical purposes, but not for all practical purposes.
Often, determining if one definition implies the other is a difficult task. The person who defines a
tensor according to basis transformation rule is actually defining a particular class of tensors,
whereas the definition in terms of linear transformations has broader abstract applicability, but
becomes bogged down in the more subtle question: what is a vector? Answering this question for
engineering mechanics applications eventually also comes around to the need to introduce basis
transformation criteria, so in this arena, both definitions are equivalent. See page 146.
57
T
F
A
R
D
cca
Rebe
Brann

Tensors
on
Linear operators (transformations)

Let f denote a vector-valued function of a vector. The symbol f denotes a
set of instructions. The operator is not itself a vector its output is a vector.
Let x be the vector supplied as an argument to the function, and let y be the
of applying the function. Then we would write, y = f ( x ) .
result
The function is linear if and only if

f ( 1 x1 + 2 x2 ) = 1 f ( x1 ) + 2 f ( x2 )
Equivalently, the function f is called linear if and only if
(i)
for all scalars
f ( v ) = f ( v )
and (ii)
for all vectors u and v
f (u + v) = f (u) + f (v)
(7.1)

Linearity gives you freedom to apply the operator to a linear combination of

vectors in either of two equivalent ways: (1) you can take the linear combination of the vectors and directly apply the operator f or (2) you can apply the
operator separately to each of starting vectors and then take the linear combination of the individual results.
If we know how a linear function transforms the three base vectors, then
linearity lets us quickly determine how the function transforms an arbitrary
vector. To see why this is so, lets define three f i vectors to be the transforma
tions of the base vectors:
and
f 1 f ( e1 ) ,
f 2 f ( e2 ) ,
f 3 f ( e3 )
(7.2)
Importantly, these three vectors can be computed once and saved for all time.
Usually these vectors are stored as columns of a 3 3 matrix [ F ] so that
[ F ] = [{ f 1}{ f 2}{ f 3}]
(7.3)
In other words, F ij is the i th component of f j . More precisely,
F ij = e i f j
f j = F ij e i
(7.4)

Recall that an arbitrary vector x may be written
x = x1 e1 + x2 e2 + x3 e3
(7.5)
Operating on this vector by the linear function f and applying the linearity
property, shows that the action of the linear transformation on an arbitrary
vector v can be rapidly computed by
f ( x ) = x1 f ( e1 ) + x2 f ( e2 ) + x3 f ( e3 )
= x1 f 1 + x2 f 2 + x3 f 3
(7.6)
Using the summation convention,

f ( x) = x j f j
Substituting Eq. (7.4) into Eq (7.7) gives
(7.7)
58
DRAF

Tensors
Rebe
f ( x ) = x j ( F ij e i ) = ( F ij x j )e i
(7.8)
If we let the vector y denote the result of f ( x ) , then we write
y = f ( x ) = yi e .
(7.9)
Thus, the i th component of f ( x ) is obtained by the simple matrix multiplica

tion, y i = F ij x j , or
{ y} = [ F ]{ x}
(7.10)
The matrix [ F ] is called the matrix of components of the tensor associated with the linear transformation f . For any linear function, there exists an
associated tensor, and vice versa. Consequently, many authors [e.g., 19] define
a tensor to be any linear transformation of vectors to vectors.* We prefer to
keep the two ideas separate in our presentation. After all, we frequently know
the linear transformation before we know the tensor itself. For example,
when b is a known constant vector, the operation f ( x ) = b x is linear with
operation
is not immerespect to x . However, the tensor associated with this
diately obvious. The operation is linear, so we know that a tensor (representable by a matrix) must exist, but what is the tensor matrix that corresponds to
f ( x) = b x ?

It is critical to recognize that the components of the matrix F ij are, by construction, referenced your chosen basis. Thus, implicitly, a tensor must consist
of both a matrix of components and an associated basis. If the basis changes,
then the component matrix changes in a specific way (explained later), which
is why some people [e.g., 18] define a tensor to be a set of components that
transform in the necessary way upon a change of basis.
The intimate dependence of the F ij components on the underlying basis
{ e 1, e 2, e 3 } is well emphasized by using the following basis notation for ten
sors:
F =
+
+
F 11 ( e 1 e 1 ) + F 12 ( e 1 e 2 ) + F 13 ( e 1 e 3 )
F 21 ( e 2 e 1 ) + F 22 ( e 2 e 2 ) + F 23 ( e 2 e 3 )
F 31 ( e 3 e 1 ) + F 32 ( e 3 e 2 ) + F 33 ( e 3 e 3 )
cca B
ranno
(7.11)
* As clarified by Simmonds, ...To say that we are given a 2nd order tensor T means that we are told
Ts action on (i.e., where it sends) any vector v. Thus two 2nd order tensors S and T are said to be
equal if their action on all vectors is the same...
The distinction is analogous to the unique correspondence between animals and their DNA
sequences. Disregarding cloning and identical twins, there is exactly one animal for each DNA
sequence and vice versa, but this does not mean that animals and DNA sequences are the same
thing. Likewise, in tensor analysis, one often has a well-characterized linear transformation without
having an explicit expression for the associated tensor (even though we know it exists).
59
T
F
A
R
D
cca
Rebe
Brann

Tensors
on
The dyadic multiplication symbol is defined below. For now, the presence
of e i e j next to F ij should be regarded as a way to indicate to the reader
that the F ij matrix of components are defined relative to the { e 1, e 2, e 3 }

basis. In this form, F is often called a dyadic, but we will simply continue to
use the term tensor.

The above expansion may be written compactly using the summation convention as
F = F ij ( e i e j )
(7.12)
We will usually omit the symbol for dyadic multiplication so that two vectors written side by side are understood to be multiplied dyadically, and the
above equations would be written more compactly as
F =
F 11 e 1 e 1 + F 12 e 1 e 2 + F 13 e 1 e 3

F 21 e 2 e 1 + F 22 e 2 e 2 + F 23 e 2 e 3

F 31 e 3 e 1 + F 32 e 3 e 2 + F 33 e 3 e 3

+
+
(7.13)
and
F = F ij e i e j
(7.14)
If F is the tensor associated with a vector-valued linear transformation of
vectors, f ( x ) , then we introduce the following notation:
Fx

f ( x)
means the same thing as
(7.15)
Most authors do not use our double underline (under-tilde) convention to indicate the order of a tensor, so you will typically see tensors typeset simply in
bold, F. Furthermore, many authors do not use the raised dot notation (hence,
to them, Fx means the same thing as what we write as F x . As will be fur
ther explained later, our notation lends itself better to heuristic self-explanatory interpretations of the intended meanings of operations, which becomes
indispensable when working with higher order tensors in modern material
modeling applications.
Recall that many people define a tensor to be a linear operation taking vectors to vectors. The basis dyad e i e j is itself a tensor, and it has a component

matrix that contains all zeros except for a 1 at the ij position. Thus, the
matrix form of Eq. (7.13) is simply
60
DRAF

Tensors
Rebe
F 11 F 12 F 13
F 21 F 22 F 23
100
010
001
F 11 0 0 0 + F 12 0 0 0 + F 13 0 0 0
000
000
000
F 31 F 32 F 33
+
F 21
000
000
000
1 0 0 + F 22 0 1 0 + F 23 0 0 1
000
000
000
F 31
000
000
000
0 0 0 + F 32 0 0 0 + F 33 0 0 0
100
010
001
(7.16)
This basis/matrix correspondence is completely analogous to the act of writing

a vector in basis form as
v = v1 e1 + v2 e2 + v3 e3
and then interpreting this expression using matrix notation as

v1
1
0
0

v2 = v1 0 + v2 1 + v3 0

v3
0
0
1
(7.17)
(7.18)
Dyads and dyadic multiplication

The symbol represents dyadic multiplication. A dyad [32] between two
vectors a and b is an abstract mathematical construction denoted a b that

when it operates on an arbitrary vector v as follows:
takes on meaning
( a b ) v = a ( b v ) for all vectors v

(7.19)
A dyad has physical meaning only when operating on a vector. Alone, it is a

mathematical bookkeeping device that can be independently manipulated but
has no inherent meaning in its own right.
Two dyads a b and r s are said to be equal if and only if

( a b ) v = ( r s ) v for all vectors v

(7.20)

Note that the right hand side of Eq. (7.19) equals the vector a times a sca (b a) v
lar, ( b v ) . Thus, ( a b ) v is proportional to a . Similarly,
would be proportional to b . Thus, dyadic multiplication does not commute.
That is,
abba
in general
cca B
ranno
(7.21)
61
T
F
A
R
D
cca
Rebe
Brann

Tensors
on
Applying the definition of a dyad (namely, Eq. (7.19)) to both sides of

Eq. (7.20), we note that two dyads a b and r s are equal if and only if

a ( b v ) = r ( s v ) for all vectors v

(7.22)

It is straightforward (but somewhat tedious) to show that, for nonzero vectors,

a b = r s only if r = a and s = b where = 1 .
(7.23)

In other words, the vector r must be proportional to a and s must be propor
is equivalent to
tional to b . The condition on the proportionality constant
a b = r s . If a and/or b is zero, then a b = r s only if

requiring that

r and/or s is zero.
The above discussion reiterates that dyadic multiplication does not commute; i.e.,
abba
(7.24)
Referring to Eq. (7.23), we obtain the necessary and sufficient condition for
a b to commute:
(7.25)
a b = b a if and only if b = a for some .
In other words, the two vectors in the dyad would have to be multiples of each
other in order for a dyadic multiplication to commute.
Simpler no-symbol dyadic notation
Whenever two vectors are written side-by-side with no symbol between

them, they are understood to be multiplied dyadically. Thus
(7.26)
a b means the same as a b .
In addition to being a more compact notation, this convention for dyadic multiplication also has the appealing property that the definition of Eq. (7.19) can
be written
(ab) v = a(b v) ,
(7.27)
62
DRAF

Tensors
Rebe
allowing us to drop the parentheses without ambiguity.* We advocate retaining the explicit symbol only when it is needed for clarity.
Dyads are like defensive linemen in football essential, but rarely of
much interest per se. They are merely abstract mathematical objects, which
(up to this point) have no apparent practical use. After we define addition and
scalar multiplication for dyads, we will see that tensors (which are of great
practical importance) are always expressible as a linear combination of the
nine possible dyads between the orthonormal base vectors. These nine basis
dyads, { e 1 e 1, e 1 e 2, , e 3 e 3 } , will form a basis for tensors just as the three
{e , e , e } form a basis for ordinary vectors. With the no symbase vectors
2 3
1 (7.12)
may be written
bol notation, Eq.
F = F ij e i e j ,
(7.28)

In this expression, the indices i and j are summed from 1 to 3. Thus, the
above expression is a linear combination of dyads, the meaning of which is
described in more detail below.
The matrix associated with a dyad

The right-hand side of Eq. (7.27) is a vector whose i th component is
ai ( b j v j )
(7.29)
We can define a 3 3 matrix whose ij components are a i b j . Then the expression in Eq. (7.29) can be written
[ a i b j ]v j
(7.30)
This can be written as a matrix equation

a1 b1 a1 b2 a1 b3 v1

a2 b1 a2 b2 a2 b3 v2 ,

a3 b1 a3 b2 a3 b3 v3
cca B
ranno
(7.31)
which is the matrix representation of the left-hand side of Eq. (7.27). Using
conventional matrix notation, note that
* Some people strenuously object to this side-by-side notation for dyadic multiplication. They argue
in favor of the symbol because dyadic multiplication is different from scalar multiplication.
By the same logic, however, the identity ( + )v = v + v should also be notationally objection
able because addition between vectors is different from addition between scalars. Likewise, the
notation dy dx would be objectionable because derivatives are not really fractions. In both mathematics and engineering, we routinely overload operations for good reason: heuristic notational
advantages. The meanings of the overloaded operations are implied by the nature of the arguments.
63
T
F
A
R
D
cca
Rebe
Brann

Tensors
on
a1 b1 a1 b2 a1 b3
a1 b1
[ a b ] = a2 b1 a2 b2 a2 b3 = a2 b2
a3 b1 a3 b2 a3 b3
a3 b3
= { a }<b>
(7.32)
The result is a square matrix. The reader should contrast this operation with
the similar-looking scalar-valued operation of Eq. (5.3) in which the transpose
merely occurs in a different place!
The sum of dyads

The sum of two dyads, a b and c d , is an abstract notion defined such that
( a b + c d ) v = a ( b v ) + c ( d v ) for any vector v

(7.33)

The sum of three dyads, a b + c d + e f is defined such that

( a b + c d + e f ) v = a ( b v ) + c ( d v ) + e ( f v ) for all v
(7.34)

The sum of two or more dyads takes on meaning only when operating on a vector. The sum of dyads also has a matrix interpretation. Specifically, the matrix
[ a b + c d ] associated with the sum of two dyads is obtained by simply adding
matrices
the
for [ a b ] and [ c d ] .
A sum of two or three dyads is NOT (generally) reducible

By applying the definition of Eq. (7.33) its straightforward to prove that
a p + aq = a( p + q)
(7.35)

This is a very special situation in which the sum of two dyads can be written
as a single dyad (between a and p + q ). In general, however, the sum of two
dyad. One cannot generally find two
dyads cannot be expressed as a single
vectors r and s such that a b + c d = r s . Thus, the sum of two dyads is itself
The
sum of three dyads is also a distinct matha distinct new abstract
object.
ematical abstraction that in general cannot be written as either a single dyad
or the sum of two dyads. We will later demonstrate that the sum of four or
more dyads can always be reduced to the sum of three or fewer dyads (provided that the vectors belong to ordinary 3D space). Thus, the sum of four or
more dyads is not a new object.
Scalar multiplication of a dyad

The multiplication of a scalar with the dyad a b is a new dyad denoted
( a b ) , and is naturally defined such that
for all v
[ (ab)] v = a(b v)
(7.36)
64
DRAF

Tensors
Rebe
Using the dyad symbol for clarity, we note that the scalar multiple may be
regarded as an external multiplier or it may be absorbed into any one of the
vectors in the dyad:
(7.37)
(a b) = ( a) b = (a ) b = a ( b) = a (b )
If the scalar is placed on the right side of the dyad or between the two vectors,
it can be moved to the left side. In other words,
a b and a b mean the same thing as ( a b ) .

Thus, scalar multiplication requires no parentheses.
(7.38)
Scalar multiplication of a sum of dyads is defined as you might expect.

Namely,
(ab + cd) = ( ab + cd)

(ab + cd + e f ) = ( ab + cd + e f )

cca B
ranno
(7.39)
The sum of four or more dyads is reducible!

The sum of four or more dyads is defined by natural extension of the definitions of Eqs. (7.35) and (7.34). Thus, for example, the sum of four dyads is
defined so that, for all vectors v ,
( a b + c d + e f + g h ) v = a ( b v ) + c ( d v ) + e ( f v ) + g ( h v ) (7.40)

Throughout this report, our vectors are ordinary vectors in 3D physical
space. Consequently, any set of four vectors must be linearly dependent. Thus,
at least one of the vectors in the set, { b, d, f , h } , can be written as a linear
for illustration
combination of the other vectors. Suppose,

that the dependent
vector is h . Then there exist k scalars such that
h = 1 b + 2 d + 3 f
(7.41)
Hence, the dyad g h can be written

g h = ( 1 g )b + ( 2 g )d + ( 3 g ) f
(7.42)

and therefore,
a b + c d + e f + g h = ( a + 1 g )b + ( c + 2 g )d + ( e + 3 g ) f
(7.43)

The left-hand side is the sum of four dyads. The right hand side is the sum of
three dyads. This proves that any sum of four or more dyads can always be
reduced to three or fewer dyads. Consequently, the sum of four or more dyads
is not a new abstract object! The sum of any number of dyads is generally
referred to as a dyadic, but we will use the term tensor.
65
T
F
A
R
D
cca
Rebe
Brann

Tensors
on
The dyad definition of a second-order tensor

The function f [ v ] = a ( b v ) is linear in v . The dyad a b is the tensor

linear
associated with this
transformation. Dyads are the most rudimentary
tensors. As emphasized earlier, the sum of two dyads generally cannot be
reduced to a single dyad. Likewise, the sum of three dyads cannot be reduced
to the sum of fewer dyads. However, the sum of four or more dyads can always
be reduced to the sum three or fewer dyads. Consequently, a second-order tensor may be defined as any sum of dyads. The term dyadic is also used to
mean the same thing. This definition is equivalent to the more traditional definition of a tensor in terms of linear transformations from vectors to vectors in
the sense that one definition implies the other. The sum of dyads definition
is more useful in certain settings, especially when considering a tensor as a
higher-dimensional vector.
For clarity, we typeset second-order tensors in bold using two under-tildes.
Thus, A , B , U would denote second-order tensors. Tensors are extremely

useful in physical applications. For example, tensors represent stress, strain,
inertia, electrical permittivity, and many other important physical quantities.
Once a physical quantity is proved to be representable by a tensor, a wealth of
theorems from abstract tensor analysis then apply to the tensor, furthering its
physical interpretation.
Expansion of a second-order tensor in terms of basis dyads

A dyad is the most rudimentary form of a tensor. Many people think of a tensor as simply a 3 3 matrix. Most of the time, that view is adequate. However,
it is important to keep in mind that the components of a tensors matrix are
always referenced to a particular basis. If the basis changes, then the matrix
changes. As discussed later, the matrix corresponds to a tensor if and only if
the matrix changes in a particular manner upon a change of basis. Since the
preceding statement contained the phrase if and only if the component
transformation property is sometimes taken as a definition of the term tensor.
For now, however, we shall continue with our equivalent view of a tensor as a
sum of dyads.
For a three-dimensional vector space (i.e., for a space where vectors have
three components), the sum of two dyads cannot generally be written as a single dyad. Likewise, the sum of three dyads generally cannot be reduced to the
sum of fewer dyads. However, as proved earlier, any sum of more than three
dyads can always be reduced to the sum of three or fewer dyads. From this
observation, we offered the very abstract definition of a second-order tensor
to be any sum of dyads. Knowing that such a sum can always be written as the
sum of at most three dyads, one might be tempted to wrongly conclude that
66
DRAF

Tensors
Rebe
tensors form a three dimensional space; however the three dyads are not the
same for all tensors in fact, they are not even unique for any given tensor.
To determine the dimension of the space of tensors, we need to determine how
many scalars are needed to uniquely define a tensor.
Given a tensor A (i.e., given a sum of dyads), we know that there exist*
vectors such that

A = ab + cd + e f
(7.44)

Each of the vectors may be expanded in terms of a basis ( a = a i e i , etc.), in
which case, the first term would become

a b = ai ei b j e j
Using the property of Eq. (7.38), this can be written
(7.45)
(7.46)
a b = ai b j ei e j

Performing similar expansions for the other terms in Eq. (7.44), we see
that any sum of dyads (i.e. any tensor) can be written as a linear combination
of the nine possible basis dyads e i e j for i and j ranging from 1 to 3. That is, for
any tensor A , there exist scalar coefficients

(called components) A ij such that
A = A ij e i e j
(7.47)

where there are an implied sums of i and j ranging from 1 to 3 for a total of
nine terms. The principal advantage of the representation in Eq. (7.47) is that
the A ij components are unique for a given orthonormal basis, just as the components of an ordinary vector are unique for a given basis.
The A ij components of tensors are often displayed in 33 arrays like this:

A 11 A 12 A 13
[ A] =
A 21 A 22 A 23
(7.48)
A 31 A 32 A 33
Let A be a second-order tensor. In terms of an orthonormal basis
{ e 1, e 2, e 3 } , the tensor A may be written
A =

i=1j=1
A ij e i e j

(7.49)
or, using Einsteins summation convention,

A = A ij e i e j
cca B
ranno
(7.50)
* Keep in mind, we only need to assert that these vectors exist. In practice, the appropriate vectors are
almost never actually computed.
67
T
F
A
R
D
cca
Rebe
Brann

Tensors
on
Its crucial to realize that these components are referenced to a particular

basis. When working with more than one basis, the component matrix might
be subscripted with e e , as in the expression
A 11 A 12 A 13
[ A] =
A 21 A 22 A 23
A 31 A 32 A 33
(7.51)
ee
The e e subscript tells the reader that the ij component of the matrix is the
coefficient of e i e j in the basis expansion, A = A ij e i e j . Another way to indi

cate the associated basis is to append the name of the basis in angled brackets
after the listing of the matrix. Thus, for example,
A 11 A 12 A 13
[ A] =
A 21 A 22 A 23
<lab>
(7.52)
A 31 A 32 A 33
would indicate to the reader that the components are referenced to the laboratory basis. Finally, the notation
A 11 A 12 A 13
[ A] =
A 21 A 22 A 23
< e 3 =n >

(7.53)
A 31 A 32 A 33
would tell the reader that the component matrix [ A ] is being displayed with
respect to a special purpose basis in which the 3-direction has be aligned with
a known unit vector n . This sort of display would be used when the compo
nent matrix is particularly simple with respect some particular (cleverly
selected) basis.
So far, we have covered two important definitions of the term tensor. We

will delay discussing the definition that is cast in terms of basis transformation rules until we can cover some basic tensor operations using the definitions we have given so far. This discussion will lead us into a discussion of how
a tensors matrix is related to the underlying basis, at which time, the basis
transformation rules should make enough sense to understand why many
people define tensors according to these rules.
68
DRAF

Tensor operations
Rebe
Seriousness is the favorite refuge of

the shallow.
Oscar Wilde
8. Tensor operations
Dotting a tensor from the right by a vector
Keep in mind that a tensor is a mathematical bookkeeping device, which
takes on physical meaning only when operating on a vector. The notation
A v represents the linear transformation associated with the tensor. Specifi
cally, applying the definition of a linear combination of dyads, the notation
A v must be interpreted as

A v = ( A ij e i e j ) v = ( A ij e i ) ( e j v ) = ( A ij e i )v j = ( A ij v j )e i
(8.1)

In the second-to-last step, we have used Eq. (5.11) to write e j v = v j . Com

that
paring the far left and far right-hand sides of Eq. (8.1) we conclude
(8.2)
A v is a vector whose i th component is A ij v j .

Thus, the components of A v may be obtained by the matrix multiplication

[ A]{v} .
The transpose of a tensor

The transpose of a tensor A is a new tensor denoted A T and defined in
direct notation such that

u ( AT v ) =
The ij component of
tion becomes
v ( A u ) for all vectors u and v

(8.3)

A T is denoted A ijT . In indicial notation, the above defini
u i A ijT v j = v m A mn u n
cca B
ranno
(8.4)
We would like to change the dummy summation indices on the right hand side
so that we may compare it to the left hand side for arbitrary vectors u and v .
That means we want the index on components of u and v to be the same on
both sides. Replacing n by i and replacing m by j , the above equation

becomes
69
T
F
A
R
D
cca
Rebe
Brann

Tensor operations
on
u i A ijT v j = v j A ji u i
(8.5)
Rearranging gives
u i ( A ijT A ji )v j = 0
(8.6)
The only way this can hold for all vectors u and v is if
A ijT = A ji
(8.7)
The above equation is often cited as the definition of the transpose, but the
direct notation definition of Eq. (8.3) is more general since it makes no
assumption that the underlying basis is Cartesian.
It is straightforward to show that
( A )T = AT
(8.8)
and
(8.9)
( A + B )T = AT + BT
Thus, the transpose operation is linear. The transpose of a dyad a b is
obtained by simply swapping the order of the vectors. Namely,
( ab )T = ba
(8.10)
If a tensor is written in the form of Eq. (7.50), the transpose simply swaps
the base vectors. Specifically,
A T = ( A ij e i e j ) T = A ij ( e i e j ) T = A ij e j e i = A ji e i e j
(8.11)
change i to j
and j to i
swap
In writing the final form, we have merely emphasized that any symbol may be
used for the dummy subscripts. Namely, we may swap the symbols i and j
without loss. The ji component of A T is the coefficient of e j e i ; so the second
says that the ij
to-last expression says that A Tji = A ij . The final expression
component of A T is the coefficient of e i e j , namely A ji . Both statements are
index symbols to describe the same

equivalent they just use different free
result!
Dotting a tensor from the left by a vector

We will define the notation u A to be a vector defined such that

( u A ) v = u ( A v ) for all vectors v

Following an analysis similar to Eq. (8.1) we write
u A = u ( A ij e i e j ) = A ij ( u e i )e j = A ij ( u i )e j = ( u i A ij )e j
(8.12)
(8.13)
70
DRAF

Tensor operations
Rebe
Thus,
u A is a vector whose j th component is u i A ij .

Equivalently, changing the symbols for the indices,
(8.14)
u A is a vector whose i th component is u j A ji

(8.15)

For both Eqs. (8.2) and (8.15), the index that is summed is on the same side of
the [A] matrix as the dot product. The transpose of A is a new tensor A T
defined such that ( A T ) ij = A ji . Thus, we note that
T
uA = A u
(8.16)
Dotting a tensor by vectors from both sides

The notation u A v is defined in a notationally consistent manner.

Namely,
u A v = ( u A ) v = u ( A v ) = u i A ij v j

There are no free indices, so this result is a scalar.
(8.17)
For advanced analysis, it is often useful to recognize that

u A v = A :u v = u v : A

Where : is the tensor inner product defined later in Eq. (12.1).
(8.18)
Extracting a particular tensor component

Using techniques similar to those used to derive Eq. (5.11) one can prove
that
A ij = e i A e j = A :e i e j

By virtue of Eq. (8.12), parentheses are not needed in this formula.
cca B
ranno
(8.19)
Dotting a tensor into a tensor (tensor composition)

Heuristically, the reader may infer the meaning of A B by expanding

each tensor in its basis form as follows
A B = A ij e i e j B kl e k e l

Applying Eq. (7.38), this may be written
(8.20)
A B = A ij B kl e i e j e k e l

Applying Eq. (5.8), we note that e j e k = jk so that
A B = A ij B kl e i jk e l

Again applying Eq. (7.38), this may be written
(8.21)
(8.22)
71
T
F
A
R
D
cca
Rebe
Brann

Tensor operations
on
(8.23)
A B = A ij B kl jk e i e l

Finally, imposing the index changing property, we may remove the jk Kronecker delta if all occurrences of j are replaced with k , giving
(8.24)
A B = A ik B kl e i e l

This result is a linear combination of A ik B kl times the dyad e i e l . Therefore,
B . Later
A B must represent a tensor whose il component is given by A
ik kl

on, the linear transformation operator associated with A B will be seen to

represent the composition of the linear transformations associated individually with A and B .
72
DRAF

Tensor analysis primitives
Rebe
cca B
ranno
Everything comes to he who hustles while

he waits.
Thomas Edison
9. Tensor analysis primitives
Three kinds of vector and tensor notation
Three kinds of notations dominate the literature: direct, basis, and indicial. Direct notation uses underlined symbols with no reference to a basis. Indicial notation shows strictly the
component form of an expression. Basis notation is similar to indicial notation except that the
associated basis is shown explicitly.
Rather than giving formal definitions, we will merely show the distinctions by examples:
Example 1: Representations of a vector

direct: v (no indices, no base vectors)
basis: v i e i (no free indices, one base vector result is a vector)
indicial: v i (one free index result corresponds to a vector)

Example 2: A linear transformation of a vector
direct: A v (no indices, no base vectors, abstract operation symbol)

basis: ( A ij v j )e i (no free indices, one base vector result is a vector)
indicial: A ij v j (one free index result corresponds to a vector)

Example 3: The composition of two tensors
direct: A B (no indices, no base vectors, abstract operation symbol)

basis: ( A ij B jk )e i e k (no free indices, two base vectors result is a tensor)

indicial: A ij B jk (two free indices result corresponds to a tensor)
Example 4: The inner product of two vectors
direct: a b (no indices, no base vectors, abstract operation symbol)

basis: a k b k (no free indices, no base vectors result is a scalar)
indicial: a k b k (no free indices result corresponds to a scalar)
73
T
F
A
R
D
cca
Rebe
Brann

on
When counting base vectors in an expression given using basis notation, it is

important to understand that this count must occur after the expression has
been simplified into a form that is identical to the indicial form except also
multiplied by the base vectors corresponding to the free indices in the indicial
expression. For example, ( a i e i ) ( b k e k ) is expressed using basis notation, but
the presence of two base vectors

does not mean it corresponds to a tensor; the
dot product between e i and e k becomes simply ik so that the expression sim
plifies to a i b k ik or simply
ak b k , which has no base vectors and is therefore a
scalar.
Direct notation is useful for conceptual discussions. As long as the abstract
operations are well-defined, it is straightforward to convert from direct notation to any of the other notations. The principal disadvantage of direct notation is that the meanings of the abstract operations can vary widely in the
literature. For example, many authors define the notation Av to mean what
we have opted to denote by A v . Also, it is often more difficult to infer the

order of the result when using direct
notation. It is a bit easier, though, with
our under-tilde convention. To infer the order of the result, first you count
the number of under-tildes (for example, A v has three under-tildes); then

you subtract from this number by an amount appropriate
to operations in the
expression. For example, the single dot operation will always reduce the
count by 2. Three minus two equals one; therefore, A v must be a first-order

tensor (i.e., a vector). The cross product operation will reduce
the count by 1.
The double dot operation : (defined later) reduces the count by four.
An extremely important use for direct notation arises in vector and tensor
calculus, where the meaning of, say, the curl for an obscure curvilinear coordinate system can be simply looked up in a handbook. Likewise, the appropriate component form for, say, the dot product for a non-rectangular basis can be
looked up in a handbook. You can perform all of your physical analyses in Cartesian indicial form and then simply express the final result in direct notation.
As long as every step in your analysis involved valid vector and tensor operations (vector addition, dot products, cross products, dyadic products, gradient
operations, etc.), then straightforward rules exist to convert your final direct
notation result into any other notation or to any other coordinate system without error it is not necessary to re-do your entire analysis when you switch
notations or coordinates!
Direct notation formulas also tend to exhibit more obvious connections
with similar formulas for scalars. For example, the time rate of the inverse of
a tensor is given by
dF
d ( F 1 )
- = F 1 -------- F 1 ,
----------------dt
dt
which is clearly analogous to the scalar formula
(9.1)
74
DRAF

Rebe
1
d ---
x
dx dt
------------ = ---------------dt
x2
(9.2)
Tensor gradients are another good example. We can define direct notation
symbols for the derivative of scalars, vectors, or tensors with respect to other
scalars, vectors, or tensors as follows:
u i
du u i
du
ds s
------- --------- e k ,
--------------- ---------- e i e j ,
------------- e i e j e k , etc. (9.3)
d x x k
dw w j
d A A jk
By using a notation such as this, direct notation expressions for the chain rule
take forms that are very similar to what is already familiar for scalars. For
example, it can be shown that
du 1
dw
-------= -------(9.4)
dw
du
Suppose, as another example, that a scalar s is given by a function of

another scalar , a vector a , and a tensor F . If each of these arguments is
expressible as a function of position x then the gradient of s with respect to
position can be computed in direct notation

by using the chain rule:
s d s da s d F
ds
------- = ------- -------- + ------- -------- + ------- : -------(9.5)
d x a d x F d x
dx
This formula looks like an ordinary application of the chain rule except that
the appropriate inner product is used between factors. The indicial form of the
above equation would be
F mn
s s a k s
s
-------- = ------- -------- + --------- --------- + -------------- ------------- x i a k x i F mn x i
x i
cca B
ranno
(9.6)
Basis notation is useful when working with more than one basis, or as an
intermediate notation when converting a direct notation expression to indicial
form. Indicial notation is often the most clear, though an alphabet soup of indices can be distracting when discussing conceptual issues. Furthermore, the
constant task of looking after the summation rules (changing j s to k s, etc.)
can be wearisome and error-prone. There are times when even indicial notation is vague. For example, what does the expression f ii mean? Is it
f ( tr ) or is it tr [ f ] ? The two are not the same. Likewise, does A ij1
mean ( A 1 ) ij or 1 A ij . Sometimes these questions can be answered by defin

ing operator precedence* or simply from context. Nonetheless, these issues
* A good rule of thumb is that the indices take lowest precedence. Thus A ij1 would be found by taking the inverse of A and then finding the ij component.
75
T
F
A
R
D
cca
Rebe
Brann

on
are characteristic of potential flaws with indicial notation. The bottom line is
that all three notations are useful, and none is perfect. A competent analyst
should learn all notations and pick the one that is clearest for the application
at hand. A courteous analyst will always attempt to infer which notational
system their audience (or the majority of their audience) will be most comfortable with. Dont be one of those pompous idiots who insist that their preferred
system is the only way to go and therefore everyone else must have it
shoved down their throats.
There exists a unique tensor for each linear function

If a function f ( x ) is known to be linear, then there exists a second-order ten
sor F which is independent
of x such that the function f ( x ) can be written in
the form F x . Stated mathematically,

if y = f ( x )
, then there exists a tensor F such that
and f ( x ) is linear in x
y = F x and vice versa

The equation y = F x is expressed in indicial notation as

y i = F ij x j
(9.7)
(9.8)
where the associated base vector e i is understood to be multiplied by these

components. If the vectors x and y are written as 31 matrices and the
(9.8) can be written

tensor F is written as a 33 matrix,
Eq.
y1
F 11 F 12 F 13

y 2 = F 21 F 22 F 23

F 31 F 32 F 33
y3
x1

x2

x3
(9.9)
Finding the tensor associated with a linear function

So far, we have only asserted that the tensor F exists whenever the function
f ( x ) is linear. In practical applications, the fastest way to construct the F ij
matrix
is to recognize that the i th column of [ F ] contains the components of
the vector f ( e i ) .
For analytical applications, a direct notation form for the tensor is often
desired. The tensor F can be determined by simply differentiating the func
tion f ( x ) with respect to x .
76
DRAF

Rebe
if y = f ( x )
,
and
f
(
x
)
is
linear
in
x
then there exists a tensor F such that y = F x

yi
dy
where F = ------- = --------e i e j
(9.10)
x j
dx
An example of this statement is given in Eq. (11.23). Incidentally, the derivative ( ) x j denotes the partial derivative with respect to x j , holding the
other components of x constant. When applying Eq. (9.10) for finding F , it is
often important to recognize

that the derivative of a vector with respect to
itself is the identity tensor I , as defined below.
x i
dx
(9.11)
------- = --------e i e j = ij e i e j I
x j
dx
This formula is applied, for example, in Eq. (11.25).

Sometimes, if the component form f ( x ) is simple enough, one can easily
x . This approach is used in
deduce F ij by taking the coefficient of
j
Eqs. (11.21) and (11.22).
The identity tensor

Consider the following function
f ( x) = x
(9.12)
Applying Eq. (7.1), this function is seen to be linear in x . Therefore, applying
Eq. (9.10), there must exist a second order tensor, which

we will denote I ,
such that
I x = x for all x
(9.13)

Recalling Eq. (9.11), the ij components of I with respect to any orthonormal
basis are simply ij .

The 3 3 matrix for the second-order identity tensor I is just the identity
matrix
100
[I] = 0 1 0
001
with respect to any orthonormal basis.
(9.14)
Thus, in basis notation, the identity tensor may be written

I = ij e i e j

or, expanded out explicitly,
cca B
ranno
(9.15)
77
T
F
A
R
D
cca
Rebe
Brann

on
I = e1 e1 + e2 e2 + e3 e3 = ek ek
(9.16)

An alternative way to derive Eq. (9.14) is to recall that the i th column of the
tensor must contain f ( e i ) . Thus, the first column of the tensor must contain
the component array for f ( e 1 ) = e 1 . The component array for e 1 is
1
0 ,
0
(9.17)
so this must be the first column of the matrix. The second and third columns
are the component arrays for e 2 and e 3 , respectively. Thus we obtain the
same matrix as in Eq. (9.14).
Tensor associated with composition of two linear

transformations
Recall that for any linear transformation, f ( x ) , there exists a tensor F
such that
f ( x) = F x
(9.18)

Likewise, for any linear transformation, g ( x ) , there exists a tensor G such
that
g( x) = G x
(9.19)

Therefore, the composition of the two transformations must be given by
f ( g( x)) = F (G x)
(9.20)
The composition of two linear functions can be readily verified to be itself a

linear function. Hence there must exist a tensor, which we will denote F G

such that
f ( g ( x ) ) = ( F G ) x for all vectors x
(9.21)

The components of F G are

( F G ) ij = F ik G kj
(9.22)

Hence, the matrix for F G may be found by the simple matrix multiplica
tion, [ F ] [ G ] .
Eqs. (9.20) and (9.21) must both hold simultaneously for all x vectors.
Hence, the direct notation statement,

( F G ) x = F ( G x ) for all x
serves as a definition of what is meant by F G .

(9.23)
78
DRAF

Rebe
The power of heuristically consistent notation

In Eq. (9.23), we have chosen a heuristically appealing notation for F G

so that we may drop the parentheses to write
FGx
(9.24)

without ambiguity. Many authors prefer the more-easily typeset notation FG
to denote the composition, so they would not be able to drop the parentheses.
Our notation is more consistent because the reader can always heuristically
decipher what it means. For example, recalling that
F = F ij e i e j and G = G kl e k e l ,

the reader can heuristically write
(9.25)
F G = ( F ij e i e j ) ( G kl e k e l ) = F ij G kl ( e i e j ) ( e k e l )
(9.26)

Using Eq. (5.8), the user can note that e j e k = jk , so that the above equa
tion would become

F G = F ij G kl jk e i e l

Finally, noting that jk is zero unless, j=k , this becomes
(9.27)
F G = F ik G kl e i e l
(9.28)

The remaining dyad is e i e l . Hence, the il components of F G must be

( F G ) il = F ik G kl
(9.29)

Aside from our choice of dummy and free indices, this result is identical
to (9.22). Throughout this report, we define our notation in such a manner
that heuristic analyses like these will always give the correct interpretation of
the notation. Thus, for example, the cross product of a tensor with a vector
would be interpreted
A v = ( A ij e i e j ) ( v k e k )

= ( A ij v k e i ) ( e j e k )

Applying Eq. (5.31), the above equation becomes
cca B
ranno
(9.30)
A v = A ij v k pjk e i e p
(9.31)

This final form contains a dyad e i e p . The reader (who might never have heard
would be able to conclude that A v must
of crossing a tensor into a vector)

be a second-order tensor, with ip components given by
( A v ) ip = A ij v k pjk
(9.32)

Similarly, a reader would be able to deduce the meaning of two tensors written
side-by-side with no symbol between them as
79
T
F
A
R
D
cca
Rebe
Brann

on
A B = A ij B kl e i e j e k e l
(9.33)

We have already introduced the notion that any linear combination of e i s is a
Genvector, and any linear combination of e i e j dyads is a second-order tensor.
basis triads e e e which will be
eralizing this idea, we will later introduce
k
i j combination
identified with third order tensors. Likewise, any linear
of
e i e j e k e l must be a fourth-order tensor. Hence, the notation A B in Eq. (9.33)

represents
must
a fourth-order tensor whose ijkl components must be A ij B kl .
Recall that u v means the same thing as u v . Likewise, when publishing for
it might be wise to use the

notation A B instead of A B
a journal article,

because your readers might confuse A B with tensor composition, which we

here denote as A B .

The inverse of a tensor

Recall from Eq. (9.21) the two notations for the composition of two linear
operators f and g
f ( g ( x ) ) = ( F G ) x for all vectors x

Now suppose that f is defined to undo the action of g so that
(9.34)
f ( g ( x ) ) = x for all vectors x

(9.35)

Then f is called the inverse of g , and is denoted g 1 . The tensor associated
with the inverse of g is denoted G 1 . Hence, putting Eq. (9.35) into (9.34)
gives the tensor definition of the inverse:

x = ( G 1 G ) x for all vectors x
In other words, the tensor inverse is defined such that
(9.36)
G 1 G = I ,
(9.37)
which means that the matrix for G 1 is obtained by the inverse of the matrix
for G .
The COFACTOR tensor

In direct notation, the cofactor tensor G C is defined by

GC ( u v ) = ( G u ) ( G v )
(9.38)
This definition might seem quite strange until you think about it physically.
The magnitude of the cross product u v equals the area of the parallelogram
of the cross product is perpendicular to
formed by u and v and the direction
u and v . Suppose that all vectors are now transformed

the plane formed
by
the vector u becomes G u and v becomes
by a linear operator G . Then
of Eq. (9.38)
is the new
area vecG v . Thus, the vector on the right-hand-side
formed
tor
by the transformed vectors. Though not at all obvious that it should
80
DRAF

Rebe
be so, it turns out that this new area vector is linearly related to the old area
vector u v . Since the relationship between the old and new areas is linear,
exist a tensor, which we denote G C , that can act on the old area to
there must
give the new area.

In direct notation, Eq. (9.38) is written
G ijC ( jkm u k v m ) = irs ( G rk u k ) ( G rm v m )
(9.39)
or, since this must hold for all u k and v m ,

G ijC ( jkm ) = irs G rk G rm .
(9.40)
Multiplying both sides by pkm and using the - identity shows that
2G C
pj = irs G rk G rm pkm
(9.41)
or, rearranging and changing the symbols used for the free indices,
G ijC =
1
--2
cca B
ranno
ipq G pr G qs rsj
(9.42)
By writing this out explicitly, we see that the component matrix of the cofactor
tensor equals the cofactor of the component matrix for [ G ] (see Eq. 3.39). Inci
dentally, the cofactor tensor is often called the adjugate
tensor. Unfortunately, some writers also call it the adjoint, but this is a dangerous
misnomer (probably originating from mis-hearing the word adjugate). When
applied to tensors, the term adjoint should be defined to mean the complex
conjugate of the transpose at least thats how it should be defined if you
want your lexicon to agree with definitions used by mathematicians. We use
the term cofactor to side-step the whole ugly issue.
Examination of Eq. (9.42) reveals that the cofactor of the transpose equals
the transpose of the cofactor. Thus, we say that these two operations commute,
and we can write G CT = ( G C ) T = ( G T ) C = G TC .
The cofactor tensor G C is well defined even if G is non-invertible. How

ever, if G happens to be invertible, then
(9.43)
G C = det ( G ) G T

Incidentally, if we define A o u v to be an initial area vector and if we
the
transformed area vector, then substidefine A ( G u ) ( G v ) to be
tuting Eq. (9.43) into Eq. (9.38) gives

A = det ( G ) G T A o

In continuum mechanics, this equation is known as Nansons formula.
(9.44)
81
T
F
A
R
D
cca
Rebe
Brann

on
Cofactor tensor associated with a vector

The cofactor operation is really meant to be applied to an even-order tensor
(second-order, fourth-order, etc.). Consider, however, the most general form for
a skew-symmetric tensor W . To be skew-symmetric, this tensor must satisfy
W ji = W ij ,
(9.45)
from which it follows that W 11 = W 22 = W 33 , W 12 = W 21 , W 23 = W 32 , and
W 31 = W 13 . As will be discussed later, it is common practice to associate a
vector w with a skew-symmetric tensor, defined such that w 1 = W 23 ,
, and w = W . In terms of this vector, the most general form for
w2 = W
31
3
12
the 3 3 matrix of a skew-symmetric tensor is
w3 w2
0
[W ] =
w3
w2 w1
w1
(9.46)
Taking the cofactor of this tensor gives

w 12 w 1 w 2 w 1 w 3
[ W C ] = w 2 w 1 w 22 w 2 w 3
w 3 w 1 w 3 w 2 w 32
(9.47)
Referring to Eq. (7.32), we recognize that this matrix is simply the matrix
associated with the dyad w w . Therefore

If w is the axial vector associated with a skew-symmetric tensor W ,
then W C = w w
(9.48)

In many computational analysis codes, skew-symmetric tensors are not saved
as 3 3 matrices (doing so would be a waste of valuable memory). Instead,
when a subroutine requests a skew-symmetric tensor from the host code, it
will instead be given the three components of the axial vector w . If that rou
tine needs to compute the cofactor of the full skew-symmetric tensor
W , then
it must recognize that all it needs to do is construct the dyad w w .

Cramers rule for the inverse

The inverse may be expressed in terms of the cofactor tensor as
G CT
G 1 = ---------------(9.49)
det ( G )
Cramers rule is very inefficient for computing the inverse of a large matrix,
but it is perfectly adequate for the 3 3 matrices associated with second-order
tensors in 3D space.
82
DRAF

Rebe
Inverse of a rank-1 modification

In general, there is no direct notation formula for the inverse of the sum of
two matrices. But there is a simple formula for any rank-1 modification of a
tensor. Specifically, if
B = C + uv ,

then (as readily verified by back substitution)
C 1 u v C 1
B 1 = C 1 ------------------------------------
1 + v C 1 u

(9.50)
(9.51)
Derivative of a determinant
Let J denote the determinant of a tensor F :
J detF
(9.52)
Recall that the determinant of a tensor can be obtained by simply taking the
determinant of the matrix of components associated with that tensor:
F 11 F 12 F 13
J det F 21 F 22 F 23
(9.53)
F 31 F 32 F 33
From this we see that J = J ( F 11, , F 33 ) . In other words, J is a function of
the nine components of F . Now we seek to know the partial derivative of J
with respect to any one of the components F ij holding the other components
fixed. Recall from simple matrix analysis that the determinant in Eq. (9.53)
can be computed using cofactors:
3
J =
F ij F ijC
(no implied sum on i )
(9.54)
j=1
Now recall that the cofactor F ijC is the signed determinate of the submatrix
obtained by striking out the i th row and j th column of [ F ] . Consequently, F ijC
does not depend on the value of F ij , and the derivative of Eq. (9.54) gives simply
J
----------- = F ijC
F ij
cca B
ranno
(9.55)
dJ
In direct notation, this result is written as, -------- = F C or, if F is invertible,
dF
dJ
-------- = J F T
(9.56)
dF
83
T
F
A
R
D
cca
Rebe
Brann

Projectors in tensor notation
on
Human salvation lies in the hands of

the creatively maladjusted.
Martin Luther King, Jr.
10. Projectors in tensor notation
Linear orthogonal projectors expressed in terms of dyads
Recall that a projection transformation requires a target or range space
and a projection direction. In Fig. 6.3, the range space is the plane whose normal is b and the projection direction is in the direction of a . For an orthogo
Eq. (6.11), the
nal projection,
a is proportional to b and, referring to
projection transformation can be written

Q( x) = x n(n x)

Applying the definition of a dyad, this can be written
(10.1)
Q( x) = Q x
(10.2)
where
Q = I nn

The complementary orthogonal null projection of Eq. (6.49) is therefore
(10.3)
P( x) = P x ,

(10.4)
P = I Q = nn

(10.5)
where
If we set up a basis for which e 3 = n , then [recalling Eq. (9.16)],
P = e1 e1 + e2 e2
1 0 0
[ P] = 0 1 0
0 0 0
w.r.t. e 3 =n

(10.6)
84
DRAF

Q = e3 e3

Rebe
0 0 0
[Q] = 0 0 0
0 0 1
w.r.t. e 3 =n

(10.7)
Note that each projector is like a piece of the identity tensor.
Finding a projection to a desired target space

Sometimes the projection function P is not known in advance. Instead,
only the desired range space is known. Referring to Eq. (10.6a), note that P is
merely the sum of dyads of an orthonormal basis for the range space. In general, if a range space is known, all one has to do is obtain an orthonormal
basis for the range space and sum the diagonal basis dyads to obtain the projector. This concept is used extensively in the main part of this report. Keep in
mind that Q is also a projector. Its associated range space in this example is
simply the 3-axis.

The above statement also applies to constructing Q .
As discussed in great length on page 150, a subset of vector space is called

a subspace if every linear combination of members of the set is itself in that
set. In the above example, there are two subspaces, the 1-2 plane and the 3axis. For every subspace, a corresponding projector can and should be constructed so that the projection theorem can be invoked. Doing so usually
greatly reduces the complexity of engineering applications. Quite often, by
applying the projection theorem, a problem that originally presented itself
like an impassible forest can be discovered instead to be a well-ordered
orchard with easily passable avenues.
Properties of complementary projection tensors

Consider a projection transformation P ( x ) and its complement Q ( x ) as
then so is Q , and they there
defined in Eq. (6.50). If the function P is linear,
fore have associated tensors P and Q . For linear projectors, Eqs. (6.51) may
be written
QQ = Q
PP = P
PQ = QP = 0

P+Q = I
P
Q
x x = x PT Q x

For example, the projection tensor associated with Eq. (6.50) is
ab
Q = I -------------
ab

or, using the no-symbol dyadic notation,
cca B
ranno
(10.8)
(10.9)
85
T
F
A
R
D
cca
Rebe
Brann

on
ab
Q = I ----------- ab

The complementary projector is
(10.10)
ab
P = -----------(10.11)
ab

The general properties listed in Eq. (10.8) can be readily verified to hold for
this particular example.
Self-adjoint (orthogonal) projectors

A projection operator P ( x ) is self-adjoint if and only if y P ( x ) = x P ( y )
the projection is linear, this means
its associ
for all vectors x and y . If

that
ated tensor P must be symmetric:
A projection P is self-adjoint (orthogonal) if and only if P T = P .
Therefore, Q T = Q also holds.

(10.12)
For example, we see that the projector in Eq. (10.11) is self adjoint if and only
if a b = b a . Recalling Eq. (7.25), this is possible only if b = a , in which
the projection
(6.11)
in which
case
reduces to the orthogonal projection of Eq.
the unit vector n is just a a .

A linear self-adjoint projector transforms its argument to the nearest point
on the range space. This means that the null space is orthogonal to the range
space. For ordinary projections in 3D-space, this means that the vector is projected orthogonally, not obliquely. When Eq. (10.12) holds, the last property
listed in Eq. (10.8) becomes
x P xQ = 0
(10.13)
Suppose that two distinct vectors x and y are decomposed via the projection
theorem as
x = x P + xQ
y = y P + yQ
Then their inner product is given by
(10.14)
(10.15)
x y = x P y P + xQ y P + x P yQ + xQ yQ
If the projection is self-adjoint, then any vector in Q-space must be perpendicular to any vector in P-space. Thus, the middle two inner products in the
above equation are both zero, and we obtain
x y = x P y P + xQ yQ
Equivalently,
(10.16)
86
DRAF

Rebe
xy = xPy+xQy

(10.17)
Generalized complementary projectors

Given a projector P , its complement is also a projector defined by
Q = I P . Suppose that the range space associated with one of these projec
has
tors
a dimension greater than one. Then that projector may always be further decomposed into more projectors, all of which may be made
complementary to each other. For second-order tensors referenced to ordinary
3D space, there can be up to three projectors P 1 , P 2 , and P 3 . For second
order tensors referenced to N -dimensional space, there can be a total of M

orthogonal projections, where M N . These generalized complementary projectors must satisfy
P i P i = P i , where i is a free index (no summation)
P i P j = 0 if i j
k=1
Pk = I
(10.18)
The dimension of the range space for each projector is given by

(10.19)
dimP i = rank [ P i ] = trP i
If all of the projectors are self-adjoint (i.e., if they are all symmetric) then
dimP i may alternatively be computed by
dimP i = P i :P i
(10.20)

where the double dot operation is defined such that, for any tensors A and B ,
A : B = A ij B ij . For self-adjoint (symmetric) projectors, the generalization of

Eq. (10.17) is
M
x y = x P k y

k=1
(10.21)
For example, the projector P Eq. (10.6) represents a projection to the 12
plane. This projection can be decomposed further into more primitive projectors; namely e 1 e 1 and e 2 e 2 . Hence, if desired, we could define a set of three

generalized mutually
orthogonal
complementary projectors as
P1 = e1 e1
1 0 0
[ P] = 0 0 0
0 0 0
cca B
ranno
(10.22)
87
T
F
A
R
D
cca
Rebe
Brann

on
P2 = e2 e2
000
[ P2 ] = 0 1 0
000
(10.23)
P3 = e3 e3
000
[ P3 ] = 0 0 0
001
(10.24)
For this example, Eq. (10.21) becomes simply

xy

= x e1 e1 y + x e2 e2 y + x e3 e3 y

= x1 y1 + x2 y2 + x3 y3 ,
(10.25)
which is a comfortingly familiar formula! In the next section, we will define

the eigenproblem and show how each distinct eigenvalue has a corresponding
unique eigenprojector. If a complete set of eigenvectors exist, then they may be
combined as sums of dyads to generate a complementary projector set satisfying the properties of Eq. (10.18). If the original tensor is symmetric, then its
eigentensors are also symmetric.
88
DRAF

More Tensor primitives
Rebe
No matter how much cats fight, there always

seem to be plenty of kittens.
Abraham Lincoln
11. More Tensor primitives
Deviatoric tensors
Rebecca: add this section

Orthogonal (unitary) tensors
Consider a linear transformation f ( x ) in which the transformed vector
vector. In other words,
always has the same length as the original
f ( x) f ( x) = x x
(11.1)

Since the transformation is linear, there must exist a second-order tensor Q
such that
f ( x) = Q x ,

so Eq. (15.36) becomes
(11.2)
(Q x) (Q x) = x x

(11.3)
or
x ( QT Q I ) x = 0
(11.4)

This must hold for all x and (because the tensor in parentheses is symmetric),
the above equation therefore

implies that
QT Q I = 0
(11.5)

Thus, the transpose of the tensor Q must equal its inverse. A tensor is said to
be orthogonal if its inverse is the same

thing as its transpose:
Q 1 = Q T
Equivalently,
Q is orthogonal
QT Q = Q QT = I
By taking the determinant of this equation, we see that
cca B
ranno
(11.6)
(11.7)
89
T
F
A
R
D
cca
Rebe
Brann

on
( detQ ) 2 = 1
(11.8)
The term proper is used if detQ = +1 and improper if detQ = 1 . These
terms are unfortunate since theyimply there is something unsavory

about an
orthogonal tensor whose determinant is negative.
We shall soon see that a proper orthogonal tensor corresponds to a rigid
rotation about some axis, defined by a unit vector a , and, if we set up a basis
in which the 3-direction is aligned with the axis of rotation,

then
[Q] =
cos sin 0
sin cos 0 with respect to basis where e 3 = a .
0
0 1
(11.9)
Here, is the angle of rotation. This expression may be written

1 0 0
000
0 1 0
[ Q ] = cos 0 1 0 + 0 0 0 + sin 1 0 0
0 0 0
001
0 0 0
with respect to basis where e 3 = a .
Now we note that
(11.10)
100
0 1 0
000
[ I a a ] = 0 1 0 and [ a a ] = 0 0 0 , and [ A ] = 1 0 0

000
0 0 0
001
(11.11)
with respect to basis where e 3 = a .
where the tensor A is defined by A ij = ijk a k . Thus, we may write a direct
notation expression for any rotation as

Q = cos ( I a a ) + ( a a ) + sin A ,
which is sometimes called the Euler-Rodrigues formula for a rotation.
(11.12)
Many people wrongly claim that an improper orthogonal tensor corresponds physically to a reflection, but it generally represents a reflection in
combination with a rotation. In 3D, there are two types of simple reflections:
(1) a reflection about the origin, which merely reverses the direction of all vectors without affecting their magnitudes, or (2) a mirror-like operation that
will transform any vector x to become its mirror image across some plane
defined by a unit normal n. The first type of reflection fully inverts space
(east becomes west, north becomes

south, and up becomes down). The second
type of reflection inverts only one direction, leaving the others unchanged (left
becomes right, but up does not turn into down). Its easy enough to infer the
tensor that would induce the first type of reflection; namely, if every vector
turns around, then the associated tensor must simply be the negative of the
90
DRAF

Rebe
identity tensor, Q = I . For the second type of reflection, the part of the vec
tor x in the direction

of the mirror unit normal n is given by n n x ; the
become
transformed vector changes the sign of this part of the vector x to
n n x . However, the projection of the vector onto the plane of the mirror,
x n n x remains unchanged. Thus, the reflection is given by the sum of this
part plus the reversed part to give
unchanged
f ( x) = x 2(nn x)

The corresponding reflection tensor is
(11.13)
Q = I 2n n

(11.14)
or
1 0 0
[ Q ] = 0 1 0 with respect to basis where e 3 = n .
0 0 1
cca B
ranno
(11.15)
In 3D, that this second-type of reflection can always be regarded as a 180

rotation about n in combination with a type-one reflection about the origin:
1 0 0 1 0 0
[ Q ] = 0 1 0 0 1 0
0 0 1 0 0 1
(11.16)
with respect to basis where e 3 = n .
Note that the last matrix in this equation is identical to the matrix in
Eq. (11.19) with the rotation angle set to 180 . This is just one special case
of a more general statement. Namely, n 3D, if Q is an improper orthogonal
tensor, then Q = R , where R is proper orthogonal.
Tensor associated with the cross product

Using the permutation symbol, the component formulas for the cross product
given in Eq. (5.27) can be written compactly as
(11.17)
( a b ) i = ijk a j b k

Note that the operation, a b is linear with respect to b . Hence, there must

exist a tensor A that is independent

of b such that
ab = Ab
(11.18)

The i th component of A b is

( A b ) i = A ij b j
(11.19)

For Eq. (11.18) to hold, the right hand sides of Eqs. (11.17) and (11.19) must be
equal:
91
T
F
A
R
D
cca
Rebe
Brann

on
A ij b j = ijk a j b k
(11.20)
This relationship must hold for all b . Before we can eliminate b , however, we
subscripts so that the indices
must change the dummy summation

on the b s
are the same on both sides of the above equation. We have b j on the left hand
side of the equation. Consequently, on the right hand side of the equation, we
must change the dummy subscript k to j . Doing that, however, would result
in four instances of j , which would violate the summation conventions. Hence,
in order to change k to j , we must also change j to something else. An easy
way to do this is to change k to j and j to k . Hence, Eq. (11.20) is equivalent
to
A ij b j = ikj a k b j
(11.21)
Now that both b s have the same index, we may assert that this holds for all
b j and therefore
A ij = ikj a k = ijk a k ,
(11.22)
Lets quickly re-derive this result in an alternative manner by using

Eq. (9.10). Specifically, we know that A in is the derivative of the i th component of a b with respect to b n . We arbitrarily selected the index symbol n
because it is not already being used as an index anywhere in Eq. (11.17).
Applying Eq. (9.10) gives the in component of A :
( a b )i
b k
( ijk a j b k )
= ----------------------------- = ijk a j --------A in = ----------------------(11.23)
b n
b n
b n
In the last step, we have used the fact that neither ijk nor a j depends on b .
To simplify this result, we apply Eq. (9.10) to write b k b n = kn so that
A in = ijk a j kn
(11.24)
As usual, we note that kn equals zero unless k = n . Hence, terms in the

summation over k will be zero unless k = n . Thus
A in = ijn a j
(11.25)
The index n is a free index, so we could change n to k , thereby demonstrating

that Eq. (11.25) equivalent to the earlier result of Eq. (11.22).
Now, we wish to write Eq. (11.22) in direct notations. Noting that
ikj = ijk , Eq. (11.22) may be written
A ij = ijk a k
(11.26)
In direct notation, this result is written

A = a

In matrix form, the components of A are
(11.27)
92
DRAF

Rebe
0 a3 a2
[ A] =
a3
0 a1
a2 a1
(11.28)
Whenever a relationship of the form in Eq. (11.27) hold, we say that a is the
axial vector associated with the skew symmetric tensor A .
By the way, observe that

(11.29)
b a = ( a b ) = A b = b AT

However, we see from Eq. (11.28) that the tensor A has the skew-symmetry
property that
AT = A
Therefore, Eq. (11.29) becomes
(11.30)
ba = b A

(11.31)
Physical application of axial vectors

The operation of Eq. (11.27) occurs quite frequently in mechanics. Suppose, for example, that
r is the vector from the origin to a point P in a
rigid body. If the rigid body rotates about the origin with an angular velocity vector , then the
velocity of the point P is given by
P
v
rigid body
rota
t
axision
v = r
(11.32)

Equation (11.18) implies that this relationship
may be written alternatively in terms of an angular velocity tensor . Namely,
v = r,
(11.33)

where
0 3 2
[] =
0 1
2 1
cca B
ranno
(11.34)
The usefulness of instead of becomes apparent by considering rotation.
Rotation can be defined in terms of an axis and an angle of rotation, but multiple rotations about different axis do not commute (i.e., if you rotate about one
axis and then about the other, the result will not be the same if you switch the
order of rotation). Rotation can nevertheless be shown to be a linear transfor-
93
T
F
A
R
D
cca
Rebe
Brann

on
mation. In other words, given a collection of vectors, you can rotate them all
individually and then take a linear combination of the results or you can first
take the same linear combination of starting vectors and then rotate the result
either way, the final result will be the same. Hence, there exists a tensor R
such that rotating an initial vector r o to obtain a new vector r can be written
(11.35)
r = R ro

If the rotation varies in time, we can take time rates of this equation to obtain
the velocity, v = r . Specifically, noting that the initial vector r o is (obviously)
not a function of time, the rate of Eq. (11.35) gives

r
v = R
(11.36)
o
Noting from Eq. (11.35) that r o = R 1 r , Eq. (11.36) may be written
v = r , where R R 1
(11.37)
It can be shown that a rigid rotation tensor must be orthogonal* and hence
(11.38)
R 1 = R T
It can further be shown that this property in turn implies that must be
skew symmetric. The associated axial vector is the angular velocity vector .
Symmetric and skew-symmetric tensors

Note that the components in Eq. (11.28) have the following property:
(11.39)
or, in direct notation, A = A T
Any tensor with this property is said to be skew-symmetric. Here, the superscript T denotes the transpose.
A ij = A ji
A tensor A is symmetric if and only if it has the property that
A ij = A ji or, in direct notation, A = A T
(11.40)
Let A be any second-order tensor, neither symmetric nor skew symmetric in
general. The symmetric part of A is defined
1
sym A --- ( A + A T )
(11.41)
2
the skew-symmetric (or antisymmetric) part of A is defined
1
(11.42)
skw A --- ( A A T )
2
* For a detailed discussion of orthogonal tensors, including the definition of Euler angles, and how to
generate a rotation from an axis and angle, see Ref. [5].
94
DRAF

Rebe
With these definitions, we see that any tensor A can be decomposed addi
tively into symmetric and skew-symmetric parts:
(11.43)
A = sym A + skw A
Instead of sym and skw, many authors use superscripts s and a so that
Eq. (11.43) would be written somewhat more compactly as
s
(11.44)
A = A +A
Note that the act of taking the symmetric part of a tensor may be regarded as
a tensor function:
1
P ( A ) = --- ( A + A T )
2
(11.45)
This operation is a projection because P ( P ( A ) ) = P ( A ) . In other words, the
symmetric part of a tensor that is already symmetric is just the tensor itself.
We can also define a function for taking the skew-symmetric part of a tensor:
1
Q ( A ) = --- ( A A T )
2
(11.46)
The functions P and Q have the properties listed in Eqs. (6.51) except that,
this time, the argument of functions is a tensor rather than a vector. Hence,
the fact that any tensor can be split into symmetric and skew-symmetric parts
is merely a generalized application of the projection theorem!
Positive definite tensors

A tensor B is positive definite if and only if
x B x > 0 for all vectors x
(11.47)
This equation can be alternatively written as

B : x x > 0 for all vectors x ,
(11.48)

where : denotes the tensor inner product, soon to be defined in Eq. (12.1).
Noting that the dyad x x is a symmetric tensor, we can use the (yet to be pre
sented) result of Eq. (12.15)
to note that B : x x = ( sym B ): x x . Therefore, a

tensor B is positive definite if and only if sym B is positive definite.

The skew
symmetric part of B has no influence on positive definiteness.
A sym B
cca B
ranno
(11.49)
95
T
F
A
R
D
cca
Rebe
Brann

on
The tensor A (and therefore B ) is positive definite if and only if its associated
component matrix is positive definite. This will happen if and only if all three
characteristic invariants of A are positive. WARNING: It is possible to con
struct a tensor B that has all positive invariants, but the tensor is not positive
definite. It is imperative to test the invariants of the symmetric part of B to
check for definiteness.

Faster way to check for positive definiteness. Recall that the k th characteristic invariant of A is given by the sum of principal k k minors. It turns
out that computing all of these principal minors is not necessary in order to
test for positive definiteness. All you have to do is test one principal minor of
each size, making sure that, as you progress in size that the smaller principal
minors are nested inside of each larger minor. The concept of nesting makes
more sense if we explain it starting with the largest principal minor, which is
the determinant of the 3 3 matrix itself. For the next principal minor, select
any 2 2 submatrix whose diagonal components are also diagonal components of the larger matrix, and verify that the determinant of this 2 2 submatrix is positive. Then test the determinant of any 1 1 submatrix on the
diagonal of the 2 2 matrix (i.e., check any diagonal component of that
matrix). The following sequence of decreasing sized submatrices are all principal submatrices, but they are not nested:
A 11 A 12 A 13
Not nested:
A 21 A 22 A 23
A 31 A 32 A 33
A 11 A 12
A 21 A 22
, [ A 33 ]
(11.50)
These are not nested because the 1 1 submatrix [ A 33 ] is not a submatrix of

the larger 2 2 matrix. Of course, in practice, now that the meaning of nesting is clear, it is wiser to check the determinants of the submatrices starting
with the smallest one first. If, at any point in this process, you find a negative
determinant, then you can stop calculations with the conclusion that the original matrix is not positive definite.
Using the simplest nested set of principal minors (those that progress in
size starting from the 11 position), the simplest test for positive definiteness of
a 3 3 matrix involves merely verifying the following
A 11 > 0
det
A 11 A 12
A 21 A 22
(11.51)
>0
(11.52)
96
DRAF

Rebe
A 11 A 12 A 13
det A 21 A 22 A 23 > 0
(11.53)
A 31 A 32 A 33
Recalling that you may use any set of nested principal minors, it follows that a
positive definite tensor will always have positive diagonal components, so this
is a good thing to visually check before even looking at the larger determinants. There is nothing, that disallows the off-diagonals from being negative,
so long as the principal minors always evaluate to a positive number.
Negative definite and semi-definite tensors

A tensor C is negative definite if C is positive definite. Note that this means
invariants I and I must
that the odd

be negative, but the even invariant I 2
1
3
must be positive.
A tensor B is positive semi-definite if and only if
x B x 0 for all vectors x

(11.54)

from which it follows that its invariants (or principal subminors) will all be
simply non-negative (i.e., positive or zero).
A tensor C is negative semi-definite if C is positive semi-definite.
Isotropic and deviatoric tensors

Another similar-looking, but fundamentally different operation decomposes a
tensor into its deviatoric and isotropic parts. The deviatoric part of A is
dev
denoted variously as A
or A d or A or dev A , and it is defined
dev
1
A A A --- ( tr A )I
(11.55)
3
Here, tr A is a scalar called the trace of A defined
tr A A kk = A 11 + A 22 + A 33
(11.56)
iso
The isotropic part of A is denoted A or iso A , and is defined
iso 1
A --- ( tr A )I
(11.57)
3
Thus, just as a tensor can be decomposed into symmetric and skew-symmetric

parts, it can also be decomposed into deviatoric and isotropic parts:
dev
cca B
ranno
iso
A = A +A
(11.58)
This is also an application of the projection theorem! This equation is analogous to Eq. (__).
97
T
F
A
R
D
cca
Rebe
Brann

Tensor operations
on
Nothing pains some people more than having

to think.
Martin Luther King, Jr.
12. Tensor operations
Second-order tensor inner product
The inner product (also called the double dot product) between two secondorder tensors, A and B , is a scalar denoted A : B and defined

A : B A ij B ij
(12.1)

This is a summation of every component of A multiplied by the corresponding
component of B . Consequently, the tensor inner product is analogous to the
ordinary dot product between vectors in ordinary physical space. This is no

coincidence. Applying a mathematicians definition of a vector (page -146), the
set of all second-order tensors can be shown to be an abstract nine-dimensional vector space. In this space, the tensor inner product serves a role
exactly like the vector dot product in 3D physical space. We can define things
like the magnitude of a tensor and the angle between two tensors in a manner that is completely analogous to the definitions for ordinary vectors!
An important special case of Eq. (12.1) is the inner product between two
dyads:
(12.2)
a b :r s = a i b j r i s j = ( a r ) ( b s )

An even more specialized instance of this formula applies it to the tensor
inner product of basis dyads:
(12.3)
e i e j :e k e l = ( e i e k ) ( e j e l ) = ik jl

The double dot product is defined such that it must operate between tensors of at least second-order. There is no need for parentheses in Eq. (12.2)
because b :r would be meaningless the double dot is understood to reach
out until it is acting between two tensors (in this case, a b and r s ). Thus, for
example, a b c :T would be interpreted as the vector a multiplied

by the inner
product of b c with the second-order tensor T . The indicial

expression of this
operation would
be
a i b m c n T mn
(12.4)
98
DRAF

Tensor operations
Rebe
cca B
ranno
Fourth-order tensor inner product

The inner product between two fourth-order tensors, X and Y , is a scalar
denoted X ::Y and defined

X ::Y X ijkl Y ijkl
(12.5)

This is a summation of every component of X multiplied by the corresponding
component of Y . Consequently, the fourth-order

tensor inner product is analo
gous to the inner
products that we have defined for vectors and second-order
tensors. Again, this is no coincidence. By applying a mathematicians definition of a vector, the set of all fourth-order tensors can be shown to be an
abstract 81-dimensional vector space. Although this view is occasionally useful in applications, we will usually find that fourth-order tensors are most conveniently regarded as operations (such as material constitutive laws) that
transform second-order tensors to second-order tensors. Hence, fourth-order
tensors may be regarded as second-order tensors referenced to nine-dimensional tensor space.
Higher-order tensor inner product

Notationally, we have defined all of our inner products such that the number of dots indicates the number of contracted indices. Clearly, this notation
is not practical for higher-order tensors. An alternative notation for an n th order inner product may be defined as the order n surrounded by a circle.
Thus, for example,
X 4 Y means the same thing as X ::Y
(12.6)

Some writers [e.g., Ref. 17*] prefer always using a single raised dot to denote
all inner-products, regardless of the order. These writers demand that meaning of the single-dot operator must be inferred by the tensorial order of the
arguments. The reader is further expected to infer the tensorial order of the
arguments from the context of the discussion since most writers do not indicate tensor order by the number of under-tildes. These writers tend to define
the multiplication of two tensors written side by side (with no multiplication
symbol between them) to be the tensor composition. For example, when they
write AB between two tensors that have been identified as second-order, then
they mean what we would write as A B . When they write UV between two

tensors that have been identified as fourth-order, they mean what we would
write as U :V . Such notational conventions are undeniably easier to typeset,

and they work
fine whenever one restricts attention to the small set of conven* We call attention to this reference not because it is the only example, but because it a continuum
mechanics textbook that is in common use today and may therefore be familiar to a larger audience.
99
T
F
A
R
D
cca
Rebe
Brann

Tensor operations
on
tional tensor operations normally seen in trivial applications. However, more

exotic advanced tensor operations become difficult to define under this system. A consistent self-defining system such as the one used in this report is far
more convenient and flexible.
Throughout this report, our notation is self-defining in the sense that the
meaning of an expression can always be ascertained by expanding all arguments in basis form, as discussed on page 79. The following list shows several
indicial expressions along with their direct notation expressions under our
notation
U mnpq V mnpq
U ::V

U ijpq V pqkl
U :V

U ijkp V plmn
UV

U ijkl V mnpq
UV

A pq U pqij
A :U

A ip U pjkl
AU

A ij U klmn
AU

A ip pjq U qklm
AU
(12.7)
Writers who use inconsistent non-self-defining notational conventions would

be hard-pressed to come up with easily remembered direct notations for all of
the above operations. Their only recourse would be to meekly argue that such
operations would never be needed in real applications anyway. Before we come
off sounding too pompous, we acknowledge that there exist indicial expressions that do not translate elegantly into our system. For example, the equation
A ijkl e i e j e k e l = U ikjl e i e j e k e l

would have to be written under our notational system as
A = X 23 U

(12.8)
(12.9)
where the rather non-intuitive swap operator X 23 is defined in Eq. (19.20).

Even when using the notational scheme that we advocate, writers should
always provide indicial expressions to define their notations, especially when
the operations are rather unusual.
100
DRAF

Tensor operations
Rebe
The difficulties with direct notation might seem to suggest that perhaps
indicial notation would be the best choice. In some instances, this is true.
However, even indicial notation has its pitfalls, principally in regard to operator precedence. For example, the notation
f
----------- kk
(12.10)
is ambiguous. It could mean

f
f
--------------- or tr ------
( tr )
The two operations give different results.
cca B
ranno
(12.11)
Bottom line: always employ notation that seems most likely to accomplish
the desired interpretation by the educated reader.
The magnitude of a tensor or a vector

The magnitude of a second-order tensor A is a scalar denoted A defined
A A: A
(12.12)

This definition has exactly the same form as the more familiar definition of
the magnitude of a simple vector v:
v vv
(12.13)

Though rarely needed, the magnitude of a fourth-order tensor X is a sca
lar defined
(12.14)
X = X :: X

A vector is zero if and only if its magnitude is zero. Likewise, a tensor is zero if
and only if its magnitude is zero.
Useful inner product identities

The symmetry and deviator decompositions of tensors are often used in
conjunction with the following identities:
A : B = sym A :sym B + skw A :skw B
(12.15)

A : B = dev A :dev B + iso A :iso B
(12.16)

Decomposition of the tensors into its symmetric plus skew symmetric parts
( A = sym A + skw A and B = sym B + skw B ) represents an orthogonal projec
tion decomposition that is completely analogous to Eq. (10.14). Thus,

Eq. (12.15) is analogous to Eq. (10.16). A similar statement holds for the
decomposition of tensors into deviatoric plus isotropic parts.
101
T
F
A
R
D
cca
Rebe
Brann

Tensor operations
on
If B happens to be a symmetric tensor (i.e., if skw B = 0 ) then the inner
the symmetric
product between B any other tensor A will depend only on
part of A . Consequently, sometimes researchers will replace A by its sym
metric part without any loss in generality which can save on storage in
numerical computations.
Incidentally, note that the trace operation defined in Eq. (11.56) can be
written as an inner product inner product with the identity tensor:
(12.17)
tr A I : A

Also note that I :I = trI = 3 , so Eq. (12.16) may be alternatively written

1
A : B = A : B + --- ( tr A ) ( tr B )
(12.18)
3

Distinction between an N th-order tensor and an N th-rank

tensor
Many authors use the term N th -rank tensor to mean what we would call an
N th -order tensor. We reject this practice because the term rank has a specific meaning in matrix analysis that applies equally well for tensor analysis.
The rank of a second-order tensor is here defined to equal the rank of the
tensors Cartesian component matrix (i.e., the number of linearly independent
rows or columns).
When a second-order tensor is regarded as an operation that takes vectors
to vectors, then the rank of the second-order tensor is the dimension of the
range space. For example, if a second-order tensor projects a vector into its
part in the direction of some fixed unit vector, then the result of that tensor
operation will always be a vector that is a multiple of the unit vector. This
range space is one-dimensional, so the rank of this second-order tensor is 1.
Based on well-known matrix theory, we note that a second-order tensor is
invertible only if its rank is 3. We have already used the term rank in connection with projections. A rank-2 second-order projection is a tensor that
projects vectors to a 2-dimensional space. The rank of the component matrix
equals the rank of the projector.
Fourth-order oblique tensor projections

Second-order tensors are themselves 9-dimensional abstract vectors with
: denoting the inner product. Consequently, operations that are defined for
ordinary 3D vectors have analogs for tensors. Recall that Eq. (6.16) gave the
formula for the oblique projection of a vector x onto a plane perpendicular to a
given vector b . The light rays defining the projection

direction were parallel
to the vector a . The analog of Eq. (6.16) for tensors is
102
DRAF

Tensor operations
Rebe
cca B
ranno
A( B: X )
.
(12.19)
P ( X ) = X ---------------------A: B

As for the projection in 3-space, this operation represents a linear oblique projection in tensor space. The surface to which X is projected is orthogonal to
B and the oblique projection direction is aligned with A . This projection func
tion appears in the study of plasticity [6] in which a trial stress state is
returned to the yield surface via a projection of this form.
The fourth-order projection transformation can be readily verified to have
the following properties:
P ( X ) = P ( X ) for all scalars .
(12.20)
P ( X + Y ) = P ( X ) + P ( Y ) for all X and Y .

(12.21)
P( P( X )) = P( X ) .
(12.22)
The first two properties simply indicate that the projection operation is linear.
The last property says that projecting a tensor that has already been projected
merely gives the tensor back unchanged.
Finally, the analog of Eqs. (6.26) and (6.27) is the important identity that
P( X ) = P(Y )
X = Y + A.
if and only if
(12.23)
This identity is used, for example, to prove the validity of radial return algorithms in plasticity theory [6].
103
T
F
A
R
D
cca
Rebe
Brann

Coordinate/basis transformations
on
The chief function of your body is to carry your

brain around.
Thomas Edison
13. Coordinate/basis transformations
Change of basis (coordinate transformation)
Quite frequently, people speak of coordinate transformations as if they
were the same thing as basis transformations. Theyre not. The coordinates of
a point in 3D space comprise exactly three numbers. Each distinct set of coordinates uniquely identifies the location of the point in space. For example,
cylindrical coordinates are { r, , z } . In the vast majority of applications, the
base vectors are defined in a way that is coupled to the choice of coordinates.
For example, the base vectors for cylindrical coordinates are { e r, e , e z } , and
these orthonormal vectors are defined to point in the direction of increasing

values of the corresponding coordinate. The components of the position vector
x are the coefficients of the vector when it is expanded as a linear combina
tion
of the base vectors. For cylindrical coordinates, the position vector is given
by x = re r + ze z ; note that there is no term in which the second base vector
Therefore,
e appears.
even though x depends on three coordinates { r, , z } ,
it has only two nonzero components. Dependence of the position vector x on
the angular coordinate is contained implicitly in the dependence of the radial

base vector e r on . For rectangular cartesian coordinates, referenced to the
laboratory basis,
it does so happen that the vector components are identical to
the vector coordinates { x 1, x 2, x 3 } .
As mentioned above, the choice of basis is almost always motivated by the
choice of coordinates so that each base vector points in the direction of increasing values of the associated coordinate. However, there is no divine edict that
demands that the base vectors must be coupled in any way to the coordinates.
If, for example, you were studying the mechanics of a ferris wheel, then you
might favor using cylindrical coordinates to identify points in space (with the
z axis parallel to the wheels axis, but the laboratory basis (with, say, the e 2
direction perpendicular to the ground) to reference physical vectors such as

the acceleration of gravity. Using the cylindrical base vectors, the acceleration
of gravity would be g = ( sin )e r cos e , but using the lab basis, it is sim
effect of changing the basis. This
ply g = e 2 . In this section, we describe
the
is, in principal, completely unrelated to changing coordinates.
discussion
104
DRAF

Rebe
cca B
ranno
Consider two different orthonormal* triads: a hat basis { e 1, e 1, e 1 } and a
tilde basis { e 1, e 2, e 3 } . A vector v can be expressed in terms of either

basis
as follows:
v = v i e i
v = v i e i
Thus, it must be true that
(13.1)
(13.2)
(13.3)
v i e i = v i e i
The hat components { v 1, v 2, v 3 } are not generally equal to the tilde components { v 1, v 2, v 3 } , but they are related in a very specific way.
Because the bases are orthonormal, we have the usual identities that
e i e j = ij

(13.4)
and
e i e j = ij

We can also compute the mixed dot products between the base vectors:
(13.5)
Q ij = e i e j = e j e i

Referring to Eq. (5.5) we note that
(13.6)
Q ij = cos ij
(13.7)
where ij is the angle between e i and e j . For this reason, the Q ij are some that relate
times called the direction cosines

one basis to the other.
Dotting both sides of Eq. (13.1) by e j and applying Eq. (13.4) gives
v i Q ij = v j ,
(13.8)
which shows how the v j components can be computed if the v i components
are known.
Dotting both sides of Eq. (13.1) by e j and applying Eq. (13.5) gives
v j = v i Q ji ,
(13.9)
which shows how the v i components can be computed if the v j components

are known. These transformation relationships can be easily remembered by
noting that the first index on Q always matches the index on the hat components while the second index always matches the index on the tilde components. This index matching property also holds in Eq. (13.6). By the way,
many authors use the above transformation relations as a definition of the
term vector for ordinary 3D physical space.
These transformation relationships can be written in matrix form as
* For a discussion of general non-orthonormal bases, see Ref. [4].
105
T
F
A
R
D
cca
Rebe
Brann

on
{ v } = [ Q ] T { v } and { v } = [ Q ] { v }
(13.10)
The only way that these equations can both be true is if [ Q ] is orthogonal:
[ Q ] 1 = [ Q ] T
(13.11)
This property holds only for transformations between orthonormal bases. Furthermore, if the handedness of both bases is the same, then
det [ Q ] = +1
(13.12)
If the handedness of one basis is different from the other, then det [ Q ] = 1 .
The transformation relationships between tensors can be derived in a similar manner. Specifically, if T = T ij e i e j = T ij e i e j , then
(13.13)
T ij = Q ip Q jq Tpq
and
T ij = Q pi Q qj T pq
(13.14)
Again, note that the first indices on each Q match the indices on the hat
components while the second indices on each Q match the indices on the
tilde components. Many authors use these tensor transformation rules as a
definition of the term tensor.
The transformation relations for a third order tensor h are
h ijk = Q Q Q h
ip jq kr pqr
(13.15)
and
h ijk = Q pi Q qj Q rk hpqr
(13.16)
The transformation relations for tensors of higher order are constructed similarly.
As a final remark, we recall that
(13.17)
Q ij = e i e j

We can always construct a basis-dependent tensor Q as
(13.18)
Q = e k e k

This tensor is basis-dependent because it characterizes the relationship

between the hat and tilde bases. Dotting Eq. (13.18) from the left by various choices of base vectors gives e i and from the right by e j gives
e i Q e j = e i e k e k e j = ik kj = ij
(13.19)

e i Q e j = e i e k e k e j = ik Q kj = Q ij
(13.20)

(13.21)
e i Q e j = e i e k e k e j = Q ik kj = Q ij

106
DRAF

Rebe
Thus, the component matrix for Q is the identity matrix when referenced to a
mixed e i e j basis. However, the component

matrix for Q with respect to either
the hat or the tilde basis contains the direction cosines

Q ij . Hence, the Q ij
matrix may be easily constructed such that its columns are the components of
each base vector with respect to the other basis. We may write
Q = ij e i e j = Q ij e i e j = Q ij e i e j

From which it follows that
cca B
ranno
(13.22)
(13.23)
e k = Q e k
and e i = Q ijT e j
This means that the columns of the Q ij matrix are the components of the
tilde basis with respect to the hat basis. Furthermore, the rows of the Q ij
matrix are the components of the hat basis with respect to the tilde basis.
107
T
F
A
R
D
cca
Rebe
Brann

Tensor invariance
on
When Im working on a problem, I never think about beauty.

I think only how to solve the problem. But, when I have
finished, if the solution is not beautiful, I know its wrong.
Buckminster Fuller
14. Tensor invariance

This section may be skipped, but it does provide some useful insight into what
is meant by a tensor.
Whats the difference between a matrix and a tensor?

This question is one of the most common questions for new students of tensor analysis. The answer comes down to the fact that tensor matrices must
correspond to components with respect to an underlying basis, and the component matrix must change according to the tensor transformation rules of the
previous section upon a change of basis. If this does not hold true, then the
matrix is just a matrix, not a tensor component matrix.
Suppose that a tensor is defined by giving a rule for constructing its components with respect to any convenient orthonormal basis. We use the term
tensor invariance to mean that the following:
If you apply the rule in the two bases separately, then the two results
must be related to each other by the transformation rule appropriate for
the order of the result.
Stated in a slightly different way, if you want to know the result of the rule
in a second basis, you can do one of the following:
You can apply the rule in the first basis and then transform the result to
the second basis.
You can first transform the arguments of the rule to the second basis, and
then apply the rule directly in the second basis.
Either way, you should obtain the same result. If you dont, then the rule violates tensor invariance.
When we talk about tensor invariance or when we use the phrase valid
tensor rule, we are using the word tensor in its generalized sense for which
scalars are regarded as 0-order tensors, vectors are 1 st order tensors, etc.
Therefore, the above statements apply for the question What is the difference
between a vector and a set of three numbers? Suppose your next door neighbor
told you that (for whatever reason) he was working with a set of three num-
108
DRAF

Tensor invariance
Rebe
bers defined to be two times the index of the number. Then the first number
would be 2, the second 4, and the third 6. He would have an array of three
numbers given by { 2, 4, 6 } . To determine if your neighbor is actually working
with a vector, you would first have to ask him if the indices of his rule had anything to do with a particular set of directions. Since this tutorial limits its
scope to orthonormal bases, we will presume that he told you that the three
directions corresponded to east, north, and up (which are orthogonal). You
would then ask your neighbor if his index doubling rule would still apply if he
were to instead use three different orthogonal reference directions such as
north-east, north-west, and up. If his answer is yes, the doubling rule applies
for any set of orthonormal reference directions, then you would conclude that
he is merely working with an array of three numbers, not a vector. To prove
that your neighbors doubling rule does not correspond to a vector, you would
only have to find one counter example. Let { e 1, e 2, e 3 } denote the directions
the
directions {north-east,
{east, north, up}. Now let { E 1, E 2, E 3 } denote

north-west, up}. In other words, let
e1 + e2
-,
E 1 = ---------------
2
e1 + e2
-,
E 2 = --------------------
2
E3 = e3
(14.1)
Then, the doubling rule in this system would give a vector

2 E 1 + 4E 2 + 6E 3
(14.2)
4e 1 + 6e 2
+ 6 E3
---------------------------
2
(14.3)
or
Thus, the doubling rule applied in the second system does not result in the
same vector as would be obtained by applying that rule in the first system
(namely, 2e 1 + 4e 2 + 6e 3 ), and you must therefore conclude that your neigh
bors rule gives

a well-defined
array, but (alas) not a vector.
Suppose, on the other hand, that your neighbor had told you that the doubling rule would not apply in a different system. You would then have to ask
him to provide you with a general rule that applies in any system. If the general rule has the property that you will end up with the same vector, no matter
what system you use, then his general rule would correspond to a vector, not
just an array. In general, if you have a rule that gives you three numbers
v 1, v 2, v 3 with respect to { e 1, e 2, e 3 } and the definition of the rule is such that
numbers

V 1, V 2, V 3 when applied to any other
it gives you three different
orthonormal triad { E 1, E 2, E 3 } , then the rule corresponds to a vector if and

only if
v1 e1 + v2 e2 + v3 e3 = V 1 E1 + V 2 E2 + V 3 E3
cca B
ranno
(14.4)
109
T
F
A
R
D
cca
Rebe
Brann

Tensor invariance
on
In other words, even though the components will generally vary upon a change
of basis, the sum of components times base vectors does not change. Thats
why the term invariance is used. It says that a vector (the sum of components
times basis) is invariant under a change of basis. If the basis changes, then
the components must change in a particular way so that Eq. (14.4) remains
true. For second-order tensors, a similar statement applies. Namely, the set of
nine basis dyads ( e 1 e 1, e 1 e 2, e 1 e 3, e 2 e 1, , e 3 e 3 ) forms a basis for second order
numbers
( a ,a , a , a , , a ) corresponds to a sectensors. A set of nine
11 12 13 21
33
ond order tensor if and only if the rule that assigns these numbers changes
upon a change of basis in such a manner that
a ij e i e j = A ij E i E j
(14.5)

where
{ E 1, E 2, E 3 }
is
any
other
(orthonormal)
basis
and

( A 11, A 12, A 13, A 21, , A 33 ) are the nine new numbers resulting from applying
the rule in the new basis. If the above equation does not hold true, then the
rule, and any numbers it spits out, do not correspond to a tensor.
Thankfully, the triplets and ordered 3 3 arrays that pop up in physics
typically do satisfy the vector and tensor invariance requirements of
Eqs. (14.4) and (14.5). Thats no surprise, since any physical phenomenon
should be unaffected by the coordinate frame that we adopt to view it from.
The components do not mean much by themselves, but the sum of components
times bases (being invariant) does have strong physical significance. It is the
sum that means something as a single entity. Even though describing a velocity requires supplying three numbers, youll never hear a person say the
velocity are . Being a vector, velocity is (and should be) regarded as a single
entity. Physical phenomena are routinely described conceptually using direct
notation and practically (i.e., for particular problems or in computer codes)
using components.
Whenever a rule is proved to satisfy invariance, we always endow it with
a direct (Gibbs) notation. Whenever a new operation is defined exclusively in
terms of invariant operations, then the new operation is immediately known
to be itself invariant.
Any expression that can be written in direct notation

is automatically known to be invariant itself.
110
DRAF

Tensor invariance
Rebe
Example of a scalar rule that satisfies tensor invariance

Given a vector v with components { v 1, v 2, v 3 } with respect to a basis
constructing a scalar by applying the following rule
{ e 1, e 2, e 3 } , we propose

s = v 12 + v 22 + v 32
(14.6)
or, using summation notation,
s = v i v i
(14.7)
Now consider a second basis { e 1, e 2, e 3 } . If we apply the rule directly in this

second basis, we get
s = v i v i
(14.8)
To demonstrate that the rule satisfies tensor invariance, we must demonstrate that Eqs. (14.7) and (14.8) give the same answer for the scalar s . Recalling Eq. (13.9) we know there exists an orthogonal matrix [Q] such that
v j = v i Q ji
(14.9)
Substituting this into Eq. (14.7), being extra careful to introduce two distinct
dummy summation indices, gives
s = v m Q im v k Q ik
(14.10)
Since [ Q ] is orthogonal, we know that Q im Q ik = mk . Hence

s = v m mk v k = v m v m
(14.11)
which is identical to Eq. (14.8). Thus, this rule for constructing the scalar s
satisfies tensor invariance. Now that we know that this rule is invariant, we
are permitted to give it a direct notation symbol. Of course, the symbol for this
operation is already defined it is simply the dot product:
s = vv

cca B
ranno
(14.12)
Example of a scalar rule that violates tensor invariance

Scalars themselves are invariant under a coordinate transformation. Newcomers to tensor analysis therefore often jump to the conclusion that any
rule that produces a scalar must be invariant. This section shows why this is
not the case. If you look back at our definition of an invariant rule, you will see
that you must be able to apply the rule in any two coordinate systems, and the
results must differ from each other by the transformation between the systems. If the rule produces a scalar, tensor invariance demands that both systems must produce the same value for the scalar.
Given a vector v with components { v 1, v 2, v 3 } with respect to a basis
{ e 1, e 2, e 3 } , lets consider
the following scalar-valued rule:

s = v1 + v2 + v3 ,
(14.13)
111
T
F
A
R
D
cca
Rebe
Brann

Tensor invariance
on
In other words, the scalar is constructed by simply adding up the vectors components. To show that this rule is not invariant, all we need is a counterexample. Consider a vector
v = 3e 1 + 4e 2
Consider a second basis
(14.14)
1
1
(14.15)
e 1 = --- ( 3e 1 + 4e 2 ) , e 2 = --- ( 4e 1 + 3e 2 ) , and e 3 = e 3
5
5
In terms of this second basis, the vector v is just
v = 5e 1
(14.16)
Now, if we apply the rule of Eq. (14.13) to Eq. (14.14) we obtain s = 7 . However, if we apply the rule to Eq. (14.16), we get s = 5 . The final results do not
agree! Consequently, the component summation rule is not a valid tensor rule.
Note, by contrast, that Eqs. (14.7) and (14.8) do give the same result for this
sample vector.
Example of a 3x3 matrix that does not correspond to a tensor

If you look up the definition of the term tensor, many textbooks (especially older engineering books) will say that a tensor is a 3 3 matrix of components referenced to a basis such that the components will change in a
particular manner upon a change of basis. Specifically, the components must
change according to the basis transformation rule described Eq. (13.13). To
help clarify this concept, lets present something that can be defined as a 3 3
matrix, but which is not the component matrix for any tensor.
In rigid body dynamics, the moments of inertia of a body describe the
amount of resistance the body has against being rotated. For example, the
moment of inertia zz is defined by
zz =
( x 2 + y 2 ) dV
(14.17)
Here, { x, y, z } are the Cartesian coordinates of a point in the body. Thus, the
quantity ( x 2 + y 2 ) is the square of the distance of a point from the z axis.
Intuitively, the farther points are from the z axis, the harder it is to rotate
those points body about the z axis. Hence, zz is a good measure of the resistance of the body to rotation about the z axis. The moments of inertia about
the other two axes are defined similarly:
xx =
( y 2 + z 2 ) dV
(14.18)
yy =
( z 2 + x 2 ) dV
(14.19)
The products of inertia are defined in many books by
112
DRAF

Tensor invariance
xy =
( xy ) dV ,
Rebe
yz =
( yz ) dV ,
zx =
and
( zx ) dV
(14.20)
Though less intuitive, these quantities measure how much wobble you would
feel when rotating the body about an axis.
It would seem natural to arrange these inertias into a 3 3 matrix as
xx xy xz
[ ] = yx yy yz
(14.21)
zx zy zz
Here, we have defined yx = xy , etc. to construct a symmetric matrix. It is
perfectly legitimate for us to arrange the inertias into a matrix form. The
question arises, however:
QUESTION: Does the [] matrix correspond to a tensor?
ANSWER: No, it does not!
Whenever you wish to prove that something is false, all you need is a counterexample. Consider a body that is a tiny ball of mass m located on the x axis
at a distance L from the origin. Treating the ball as a point mass so that
x L , y z 0 , the inertia values are
yy = zz = mL 2 and all other ij 0
(14.22)
Thus,
[] =
mL 2
000
010
001
w.r.t. the { xyz } coordinate system.
Now consider a different hat coordiy

nate system as shown. In this coordinate system, x = y L 2 and z 0 .
Applying the definitions of the inertia
components in this hat system gives
1
x x = y y = --- mL 2
2
(14.24)
zz = mL 2
(14.25)
1
xy = --- mL 2
2
(14.26)
(14.23)
y
m
cca B
ranno
x
L
Thus,
113
T
F
A
R
D
cca
Rebe
Brann

Tensor invariance
on
[ ] = mL 2
12 12
12 12
0
0
1
w.r.t. the { x y z } coordinate system.
(14.27)
Now the question becomes: would we have obtained Eq. (3.46) if we had
merely performed a coordinate transformation of Eq. (14.23)? Lets show that
the answer is no. The direction cosine matrix for the transformation from
the { xyz } system to the { x y z } system is
1
------2
1
[ Q ] = -----2
1
------2
1
------2
(14.28)
0 0 1
If we assume (incorrectly) that the [ ] matrix corresponds to a tensor, then
the following relationship should be true
[]
w.r.t.
{ x y z }
= [ Q ] T [ ] w.r.t. [ Q ]
(14.29)
{ xyz }
Thus, if [ ] corresponds to a tensor,* then it should be true that

1
------- ------- 0
------- ------- 0
000
12 12 0
2
2 2
2
This equation is false!
1 1
1 1
0
1
0
1 2 1 2 0 = ------ ------- 0
------- ------- 0
2 2
2 2
001
0
0 1
0 0 1
0 0 1
(14.30)
Multiplying this out, however, shows that equality does not hold. Consequently, this counterexample proves that the matrix of inertias defined in
Eq. (14.21) is not the matrix associated with a tensor!
The inertia TENSOR

Now we are left with the question: Is there any way to organize the inertias
into a matrix form that does correspond to a tensor? The answer to this question is yes! Simply putting negatives on the diagonals will turn out to define
a properly invariant tensor matrix:
xx xy xz
[] =
yx yy yz
(14.31)
zx zy zz
This new definition corresponds to a tensor if and only if, for any two Cartesian coordinate systems,
* Keep in mind: we are trying to prove that it doesnt.
114
DRAF

Tensor invariance
[]
w.r.t.
{ x y z }
Rebe
[ Q ]T [ ]
w.r.t. [ Q ]
{ xyz }
(14.32)
Lets first check whether placing negatives on the diagonals fixes our previous
counterexample. With our new definition for the inertia tensor, we now ask
whether negatives on the diagonals of the inertia tensors in Eq. (14.30) will
make that equation true. In other words, is the following matrix equation
true?
1
------- ------- 0
------- ------- 0
0 0 0
1 2 1 2 0
2
2 2
2
=
This equation is true!
1
1
1
1
1 2 1 2 0
------- ------- 0 0 1 0 ------- ------- 0
2
2
2
2
0 0 1
0
0 1
0 0 1
0 0 1
(14.33)
Multiplying this out shows that equality holds. Passing this test is necessary,
but not sufficient. he fact that one example worked out okay does not prove
that our new definition really does correspond to a tensor. We dont know for
sure (yet) whether or not some other counterexample might prove that our
new definition isnt a tensor either. It is always harder to prove something is
true than to prove that it is false!
It turns out that our new definition does indeed transform properly for all
possible coordinate changes. One of the simplest ways to prove that something
is a proper invariant tensor rule is to find some way to write out the definition
in direct (Gibbs) notation using only operations (such as the dot product and
dyadic multiplication) that have already been proved to be tensor operations.
We recognize that the expression x 2 + y 2 that appears in the definition of
zz can be written as ( x 2 + y 2 + z 2 ) z 2 , or simply x x z 2 , where the posi observation lets us
tion vector is x = xe x + ye y + ze z = x e x + y e y + z e z . This
into
the formula. Note that the offintroduce at least
a bit of direct notation
diagonals of the dyad x x are identical to the off-diagonals of our new inertia
matrix. The diagonal components

of x x are x 2 , y 2 , and z 2 , which are
exactly the same as the terms we used to rewrite the diagonals of the new
inertia matrix. For example, the zz component x 2 + y 2 was re-written as
x x z 2 . With a bit of thought, we therefore recognize that the inertia tensor
be written in the following clean direct notation:
may
=
[ x x ( x x ) I ] dV
cca B
ranno
(14.34)
By discovering this direct notation formula, we have relieved ourselves of the

burden of proving invariance under coordinate transformations. The direct
notation operations themselves (namely: dyadic multiplication, tensor addition, scalar multiplication, vector dot product) have all been individually
proven invariant in separate analyses. Any object constructed using these
invariant operations must itself be invariant!
115
T
F
A
R
D
cca
Rebe
Brann

Scalar invariants and spectral analysis
on
Distrust any enterprise that requires new

clothes.
Henry Thoreau
15. Scalar invariants and spectral analysis
Invariants of vectors or tensors
A scalar invariant of a vector or tensor is any scalar valued function of the
vector or tensor that gives the same result even upon a change in basis.
The sum v 1 + v 2 + v 3 is not an invariant of a vector v because the sum
v 1 + v 2 + v 3 with will generally give a different result when

computed with
respect to a different basis. However, the sum of the squares is an invariant.
We claim that
v 12 + v 12 + v 12 = v 12 + v 22 + v 32
(15.1)
or, using summation notation,

v k v k = v k v k
(15.2)
To prove that the sum of squares is a vector invariant, we must prove that
Eq. (15.2) holds true whenever the v k components are related to v j components through a regular basis transformation, namely v k = Q ik v i where Q jk
are the components of any proper orthogonal matrix. Making this substitution
for v k in the expression v k v k gives
v k v k = Q ik v i Q mk v m = v i Q ik Q mk v m = v i v i ,
(15.3)
where, in the last step that we have noted that Q ik Q mk = im by virtue of

orthogonality of [ Q ] . The above result proves that Eq. (15.2) is true. In other
words, the sum of the squares of vector components (referenced to an
orthonormal basis) is an invariant of any vector. We recognize this invariant
to be the square of the vectors magnitude.
2 + T 2 + T 2 of the squares of the diagonal components of a
The sum T 11
22
33
tensor, T is not an invariant of a tensor because the result is generally differ
ent for different bases. However, the sum of the diagonal components
T 11 + T 22 + T 33 alone does turn out to be an invariant because, for any orthogonal Q ij direction cosine matrix,
116
DRAF

T 11 + T 22 + T 33
= T ij Q ki Q kj
Rebe
= T kk
(now use orthogonality)
= T ij ij
= T ii
= T 11 + T 22 + T 33
(15.4)
Primitive invariants
As mentioned earlier, the magnitude of a vector is an invariant. Likewise,
the square of the magnitude is also an invariant. These two invariants are not
independent. One can always be computed from the other.
Primitive invariants of a vector or tensor are any minimal set of invariants
such that all other invariants may be expressed as functions of the primitive
invariants.
A vector has only one primitive invariant its magnitude. Symmetric tensors (referenced to 3D space) have three primitive invariants. This follows
because symmetric tensors have three eigenvalues. Since an invariant may be
computed in any basis with the same result, all invariants of symmetric tensors must be expressible as functions of the tensors eigenvalues { 1, 2, 3 } .
For example, the magnitude of a symmetric tensor is an invariant that may be
written as 12 + 22 + 32 . Nonsymmetric tensors have more than three primitive
invariants. For nonsymmetric tensors, the magnitude is itself an independent
invariant that cannot be expressed as a function of the eigenvalues. To prove
this statement, consider a tensor whose component matrix with respect to
some particular basis is given by
1 W 0
0 2 0
cca B
ranno
(15.5)
0 0 3
The eigenvalues are { 1, 2, 3 } , but the tensor magnitude is
12 + 22 + 32 + W 2 , which depends on the 12 component W . Different values
of W will give different magnitudes. Hence, the magnitude must not be
expressible as a function of the eigenvalues. Tensor magnitude is an independent fourth invariant for nonsymmetric tensors! Thus, dont let anyone tell
you that a tensor is zero if all of its eigenvalues are zero that statement is
true only for symmetric tensors!
117
T
F
A
R
D
cca
Rebe
Brann

on
Trace invariants
Three popular invariants of a second-order tensor are
I = trT
II = tr ( T 2 )
III = tr ( T 3 )
(15.6)
For symmetric tensors, these form an acceptable set of primitive invariants. A

symmetric tensor is zero if and only if these three invariants are all zero. This
statement if false for nonsymmetric tensors, as can be readily verified by putting 1 = 2 = 3 = 0 and a = 1 in Eq. (15.5). If the eigenvalues are known,
then the trace invariants are given by
I = 1 + 2 + 3
II = 12 + 22 + 32
III = 13 + 23 + 33
(15.7)
Characteristic invariants
The characteristic invariants of a second-order tensor T are defined as fol
lows:
I k = the sum of all possible k k principal subminors
(15.8)
Here, a k k principal subminor is the determinant of a matrix formed by

striking out rows and columns of the component matrix [ T ] until all that is
left is a k k matrix whose diagonal components are also

diagonal components of the original tensor [ T ] . The I k principal invariant requires you to
find all of the possible ways to construct a k k principal subminor and to

sum their determinants.
For a tensor referenced to 3D space, the component matrix is 3 3 and the
characteristic invariants are
I 1 = T 11 + T 22 + T 33
I 2 = det
T 11 T 12
T 21 T 22
+ det
T 22 T 23
T 32 T 33
+ det
T 11 T 13
T 31 T 33
T 11 T 12 T 13
I 3 = det T 21 T 22 T 23
(15.9)
T 31 T 32 T 33
118
DRAF

Rebe
These are called the characteristic invariants because they are coefficients in
the characteristic equation that gives the eigenvalues of T . Namely,
3
2
(15.10)
I1 + I2 I3 = 0
This equation is true even if the tensor T is nonsymmetric. For symmetric
tensors, the characteristic invariants form

an acceptable set of primitive
invariants, meaning that all other invariants may be expressed as functions of
the characteristic invariants this statement holds only for symmetric tensors. A symmetric tensor is zero if and only if its characteristic invariants are
all zero. This statement if false for nonsymmetric tensors, as can be readily
verified by putting 1 = 2 = 3 = 0 and a = 1 in Eq. (15.5).
If the eigenvalues of the tensor are known, then the characteristic invariants are given by
I 1 = 1 + 2 + 3
I 2 = 1 2 + 2 3 + 3 1
I 3 = 1 2 3
(15.11)
Thus, comparing with Eq. (15.7), we note that the characteristic invariants
are related to the trace invariants by
tr ( T ) = I 1
tr ( T 2 ) = I 12 2I 2
tr ( T 3 ) = I 13 3I 1 I 2 + 3I 3
(15.12)
These relationships hold only for 3D tensors. However, the basic definition of
the characteristic invariants and their relationship with the characteristic
equation extends analogously to other dimensions. For 4-dimensional space,
the characteristic equation is 4 I 1 3 + I 2 2 I 3 + I 4 = 0 . For two-dimensional space, the component matrices are only 2 2 and the characteristic the
characteristic equation is
2 I 1 + I 2 = 0 ,
where
I 1 = T 11 + T 22
and
I 2 = det
T 11 T 12
T 21 T 22
cca B
ranno
(15.13)
Applying this formula is the fastest and least error-prone way to write down
the characteristic equation of a 2 2 matrix.
Direct notation definitions of the characteristic invariants. The
invariant definitions that we have given so far demand that the underlying basis
must be orthonormal. Some more advanced texts (e.g., [20]) cite more
abstract, and therefore more broadly applicable, definitions of the characteristic invariants. These definitions are phrased in terms of the triple scalar product, [ ( ), ( ), ( ) ] , defined in Eq. (5.34). Namely, in 3D, I 1 , I 2 , and I 3 are
defined such that, for any vectors, u , v , and w ,

119
T
F
A
R
D
cca
Rebe
Brann

on
[ T u, v, w ] + [ u, T v, w ] + [ u, v, T v ] = I 1 [ u, v, w ]
(15.14)

[ u, T v, T w ] + [ T u , v, T w ] + [ T u, T v, w ] = I 2 [ u, v, w ]
(15.15)

(15.16)
[ T u , T v, T w ] = I 3 [ u, v, w ]

Note that the left hand side in the equation for I 1 sums over every possible
way for the tensor T to operate on exactly one of the vectors in the triple sca
lar product. Similarly, I 2 involves sums over every possible way for the tensor
to act on two of the vectors and I 3 involves sums over every way (only one
way) for the tensor to act on all three vectors in the triple scalar product.
Admittedly, these definitions are rather strange looking, and we have committed the proof that the right hand side can, in every instance simplify to one
invariant number times the triple scalar product [ u, v, w ] . Nonetheless,
it is essential to
expressions like these show up occasionally in analyses, and
recognize how beautifully they simplify. Furthermore, these definitions are a
nice place to start when attempting to deduce how to compute the tensor
invariants for irregular (non-orthonormal) bases.
The cofactor in the triple scalar product. We close this discussion with a
final identity, involving the cofactor tensor. In particular, recall the direct
notation definition of the cofactor, given in Eq. (9.38):
GC ( u v ) = ( G u ) ( G v )
Dotting both sides of this equation by a third arbitrary vector w gives
[ GC ( u v ) ] w = [ ( G u ) ( G v ) ] w
(15.17)
(15.18)
The right-hand side is the triple scalar product [ G u, G v, w ] . The left-hand

side can be written [ ( u v ) G CT ] w or ( u v ) [ G CT w ] , which is the triple scalar
product [ u, v, G CT w ] . Thus

[ G u, G v, w ] = [ u, v, G CT w ]

(15.19)
This result shows that, when a tensor acts on two arguments in a triple scalar
product, it can be recast as the transpose of the cofactor acting on the previously untransformed third vector. By the cyclic property of the triple scalar
product, we can assert that this statement holds true when the tensor is acting on any two vectors in a triple scalar product.
Applying the above identity to the left-hand-side of Eq. (15.15), and then
applying the definition of trace given in Eq. (15.14) shows that
I 2 = trT c
(15.20)
120
DRAF

Rebe
Invariants of a sum of two tensors

Consider the sum T = A + B of two tensors. The first invariant is simple
enough because the trace of a sum is the sum of the traces. Thus
I 1A + B = tr A + tr B = I 1A + I 1B
Solving Eq. (15.12b) for I 2 gives a formula for the second invariant:
1
I 2 = --- ( I 12 tr ( T 2 ) )
2
When applied to T = A + B , the result is
1
I 2A + B = --- ( ( tr A + tr B ) 2 tr ( A + B ) 2 )
2
which simplifies to
I 2A + B = I 2A + I 2B + I 1A I 1B tr ( A B )

Without proof, the determinant of a sum can be written
I 3A + B = det ( A + B ) = I 3A + I 3B + Ac : B + A : B c
cca B
ranno
(15.21)
(15.22)
(15.23)
(15.24)
(15.25)
CASE: invariants of the sum of a tensor plus a dyad. Now suppose that
the tensor B is actually a dyad:
B = uv
(15.26)
Then
I 1B = u v

B
I2 = 0
(15.27)
I 3B = 0
(15.29)
(15.28)
121
T
F
A
R
D
cca
Rebe
Brann

on
The Cayley-Hamilton theorem:

A tensor satisfies its own characteristic equation. That is,
T 3 I1T 2 + I2T I3 I = 0 ,
(15.30)
where I is the identity tensor, and 0 is the zero tensor. This theorem is trivial
the proof is quite subtle if the tensor is

to prove if the tensor is symmetric, but
non-symmetric [see Ref. __].
The above theorem provides a means of writing powers of a tensor in terms
of second or lower powers. For example, multiplying Eq. (15.30) by T and solv
ing for T 4 gives
T4
= I1T 3 I2T 2 + I3T
= I1( I1T 2 I2T + I3 I ) I2T 2 + I3T
= ( I 12 I 2 )T 2 + ( I 3 I 1 I 2 )T + I 1 I 3 I
(15.31)
where, in the second step, we used Eq. (15.30) to substitute for T 3 .
CASE: Expressing the inverse in terms of powers and invariants.

We can multiply both sides of Eq. (15.30) by T 1 and then solve for T 1 to
obtain
T 2 I1T + I2 I
T 1 = ------------------------------------I3
(15.32)
Inverse of the sum of a tensor plus a dyad

( A 1 u v A 1 )

( A + u v ) 1 = A 1 ------------------------------------------

1 + v A 1 u

(15.33)
Eigenvalue problems
Consider a general tensor, A . A vector p is called a right eigenvector of
A if and only if A p is proportional to p .In other words, there must exist a

scalar , called the eigenvalue, such that

A p = p
(15.34)

Eigenvectors are also called principal directions. Note that if an eigenvector

can be found, its magnitude is arbitrary because any scalar multiple of an
eigenvector would also be an eigenvector. Thus, if we say that an eigenvector
is unique, we really mean that its direction is unique.
Rewriting Eq. (15.34), we seek p and such that
[ A I ] p = 0
(15.35)
122
DRAF

Rebe
The only way that this equation can hold is if the tensor in brackets is singular. Hence, must satisfy
det [ A I ] = 0
(15.36)
Since we are speaking of ordinary tensors in ordinary 3D physical space,

expansion of this determinant gives a cubic equation for the eigenvalue .
Specifically, it can be shown that this characteristic equation is
3 I 12 + I 2 I 3 = 0
(15.37)
where I i are the invariants of Eq. (15.9). Once an eigenvalue is found, the corresponding eigenvector is determined by enforcing the definition of
Eq. (15.34). For two-dimensional problems, a graphical tool called Mohrs circle, is also very useful for performing eigenvalue analyses.*
Because the equation is cubic, there are up to three distinct eigenvalues,
{ 1, 2, 3 } . Hence there is the potential for having up to three associated
eigenvectors, { p 1, p 2, p 3 } , but there might be fewer (as explained below). Any
contains the proof that the eigenvectors corresponding
linear algebra textbook
to distinct eigenvalues will be linearly independent. In other words,
If i j , then p i and p j are linearly independent.
For symmetric tensors, the following stronger statement holds
cca B
ranno
(15.38)
For symmetric tensors,

(15.39)
if i j , then p i and p j are orthogonal.
Since the magnitudes of eigenvectors are arbitrary, it is customary to scale

them to unit length so that distinct eigenvectors for symmetric tensors are not
only orthogonal, but also orthonormal. As discussed later, a different convention applies for non-symmetric tensors.
Algebraic and geometric multiplicity of eigenvalues

If all three eigenvalues are distinct (i.e., if there are no repeated roots),
then the three eigenvectors are linearly independent and may be used as a
basis for the 3D physical space.
When two or more eigenvalues are equal, however, the situation becomes
more complicated. The algebraic multiplicity of an eigenvalue is the number of
times the eigenvalue repeats as a solution to Eq. (15.37). The geometric multiplicity of an eigenvalue is the number of linearly independent eigenvectors
there are associated with the eigenvalue. There is always at least one eigen* An extremely detailed exposition on Mohrs circle, including its definition for nonsymmetric tensors is provided in Ref. [7]. More elementary discussions may be found in virtually any undergraduate text on mechanics of materials.
123
T
F
A
R
D
cca
Rebe
Brann

on
vector associated with each eigenvalue. For repeat roots, the geometric multiplicity is always less than or equal to the algebraic multiplicity. The
distinction between geometric and algebraic multiplicity is important only for
nonsymmetric tensors because (as any good matrix analysis will prove),
For a symmetric tensor, the geometric multiplicity always
equals the algebraic multiplicity.
Consequently, a symmetric tensor will always have a complete set of linearly
independent eigenvectors that may be used as a basis for 3D space. Such a
basis is sometimes called an eigenbasis.
For repeated eigenvalues, the associated eigenvectors do not have unique
directions. If the geometric multiplicity is two, then there are exactly two associated linearly independent eigenvectors. These two eigenvectors may be used
to define a plane in 3D space. The eigenvectors associated with the eigenvalue
of geometric multiplicity two are unique to the degree that they must lie in
that plane, but their directions (and, of course, magnitudes) are arbitrary. If
desired, they may be scaled such that they are any two orthonormal vectors in
the plane. These statements apply even to nonsymmetric tensors. For symmetric tensors, in light of Eq. (15.39), we note that
A symmetric tensor always has an orthonormal eigenbasis.
(15.40)
A normalized eigenbasis is often denoted with different symbols, { 1, 2, 3 }

to emphasize that it has been selected such that
i j = ij

(15.41)
Although not guaranteed, it may so happen that a particular non-symmetric tensor of interest also has all of its geometric multiplicities equal to the
algebraic multiplicities. In such a case, that tensor also possesses a spanning
set of eigenvectors { p 1, p 2, p 3 } that may serve as a basis for 3D space. How
ever,
If a non-symmetric tensor happens to have an eigenbasis,
then the eigenbasis will not be orthogonal.
(15.42)
A non-orthogonal eigenbasis { p 1, p 2, p 3 } always has a dual basis

{ p 1, p 2, p 3 } that is defined such that

p i p j = ij ,
(15.43)
where the ij is an alternative symbol for the Kronecker delta. Comparing

Eqs. (15.43) with (15.41) shows that the dual basis associated with { 1, 2, 3 }

is just { 1, 2, 3 } itself, so there is no need to worry about a dual basis for

symmetric tensors.
124
DRAF

Rebe
Diagonalizable tensors
Whenever a tensor possesses a spanning set of eigenvectors, { p 1, p 2, p 3 }

(i.e., when a tensor has an eigenbasis), then the tensor is said to be diagonalizable. In matrix analysis, this means that the matrix [ A ] ee containing the
components of A is with respect to the laboratory basis { e 1, e 2, e 3 } is similar
diagonal.

to a diagonal matrix [ ] containing the eigenvalues on the
In other
words, there exists a matrix [ L ] such that
[ A ] ee = [ L ] [ ] [ L ] 1
(15.44)
where
1 0 0
[] =
0 2 0
(15.45)
0 0 3
Comparing Eq. (15.34) with (15.44), we note that the columns of the matrix
[ L ] must contain the components of the three eigenvectors { p 1, p 2, p 3 } with

respect to the orthonormal laboratory basis. If the tensor A is symmetric,
the
eigenvectors will be orthonormal and [ L ] will be orthogonal.

Using vector basis notation, Eq. (15.44) implies that the tensor A can be
written as the sum of three eigen-dyads:

A = 1 p1 p 1 + 2 p2 p 2 + 3 p3 p 3 ,
(15.46)

where the eigenvectors are here presumed to satisfy the normalization of
Eq. (15.43) so that, indeed, A p i = i p i (no sum on i ). Incidentally, note

that
pi A = i pi
(15.47)
The dual vectors p i , are sometimes called the left eigenvectors by virtue of
the above relationship.

It is apparent that the left eigenvalues are the same as
the right eigenvalues, and the left eigenvectors are the right eigenvectors
of A T .
If the tensor is symmetric, theres no distinction between the two types of

base vectors, and Eq. (15.46) can be written as simply
If A is symmetric, then A = 1 1 1 + 2 2 2 + 3 3 3
cca B
ranno
(15.48)
125
T
F
A
R
D
cca
Rebe
Brann

on
Eigenprojectors
Consider a diagonalizable tensor A for which two eigenvalues are repeat
roots. Suppose, for example, 2 = 3 . Then Eq. (15.46) can be written

or
A = 1 p1 p 1 + 2 ( p2 p 2 + p3 p 3 ) ,
(15.49)
A = 1 P1 + 2 P2 ,
(15.50)
where
P1 p1 p 1

P2 p2 p 2 + p3 p 3 .

By virtue of Eq. (15.43) we note that
(15.51)
and
P1 P1 = P1
P2 P2 = P2
(15.52)
Consequently, the P k tensors are projectors, called eigenprojectors. Further

more,
P1 P2 = 0 ,
Finally, since the eigenvectors form a basis, it can be shown that
(15.53)
(15.54)
P1 + P2 = I
Comparing these results with Eq. (10.18) reveals that the eigenprojectors are
complementary projectors and the number M in Eq. (10.18) is equal to the
number of distinct eigenvalues. Because 2 = 3 , the directions of eigenvectors p 2 and p 3 are not unique they only need to reside in the plane defined
However, even though the directions of p and p are not

by their
span.
2
3
unique, the sum, p 2 p 2 + p 3 p 3 , is unique! In other words, the eigenprojector

double
root is unique. Furthermore, note that the
associated with the
2
dimension of the range space for each eigenprojector equals the multiplicity of
the eigenvalue.
If all of the eigenvalues of the tensor instead had been distinct, we could
have performed a similar analysis to obtain
A = 1 P1 + 2 P2 + 3 P3 ,
(15.55)
where
P1 p1 p 1

P2 p2 p 2

P3 p3 p 3
(15.56)

and the properties of Eq. (10.18) would again hold, this time with M=3 .
126
DRAF

Rebe
cca B
ranno
P i P j = P i if i= j so each P i is a projector
(15.57)
P i P j = 0 if i j
(15.58)
(15.59)
P1 + P2 + P3 = I
Likewise, if all three eigenvalues had been equal (triple root), then the number of distinct eigenvalues would be M=1 , and the same process would (trivially) give
A = I ,
(15.60)
where we note that the identity tensor itself is a (degenerate) projector.

These concepts can be generalized to fourth-order tensors, where (instead
of eigenvectors) we will find eigentensors. The idea of eigenprojectors is
extremely useful in this case, especially in the study of material anisotropy.
127
T
F
A
R
D
cca
Rebe
Brann

Geometrical entities
on
Always do right. This will gratify

some people and astonish the rest.
Mark Twain
16. Geometrical entities
Equation of a plane
A plane in 3D space can be described completely by specifying the planes
orientation by a unit normal n and the distance d of the plane from the ori
gin. Physically, we know that any
position vector x that points from the origin
to a location in the plane must have a projection onto

the planes normal that
is identically equal to equal to the distance d . Thus, the direct notation
equation for a plane is
xn = d

Written out in component form,
(16.1)
x1 n1 + x2 n2 + x3 n3 = d
(16.2)
Even though a plane is a two dimensional object, it takes three independent

parameters to specify it. The unit normal has two independent components
(the third being restricted to ensure that n = 1 ) and the distance d makes
the third parameter.

If the plane does not pass through the origin (i.e., if d 0 ), then the equation for the plane may be written
n
x u = 1 , where u ---
d
(16.3)
In this case, u i equals the intercept of the plane on the i th axis of the Cartesian coordinate triad.
If the plane does pass through the origin, then the equation of the plane is
just
xn = 0
(16.4)

The equation of a plane passing through a particular point p is therefore
( x p) n = 0
(16.5)

128
DRAF

Rebe
Equation of a line
Consider a straight line that passes through one point x 1 and a second
x that extend
point x 2 . The easiest way to define the set of position vectors
in parametfrom the origin to some point on the line is to write the equations
ric form by simply expressing the fact that the vector from x 1 to any point on
the line must be parallel to the vector from x 1 to x 2 . Letting the proportional
for a line can be written
ity factor be denoted by t, the parametric equation
( x x1 ) = t ( x2 x1 )
(16.6)

Since this is a vector expression, it represents a set of three separate equations. If the parameter t equals zero, then x is located at x 1 . If t=1 , then the
point x is located at x 2 . Of course, values of t outside the interval

from 0 to 1
and will result in position vectors outside the line segare permissible
as well,
ment from x 1 to x 2 .
An alternative parametric equation for a line is
where
y = n
cca B
ranno
(16.7)
x2 x1
- , and L = x 2 x 1
(16.8)
y x 1--2- ( x 1 + x 2 ) , n ----------------L

The parameter varies from L/2 to +L/2 as x varies from x 1 to x 2 .
More generally, one can define a line by any point p on the line and the
be written
orientation n of the line. Then the equation of the line can
(16.9)
y = n ,
where
y = x p,

(16.10)
and n defines the orientation of the line
Dotting both sides of Eq. (16.9) by n shows that
= yn
(16.11)

This expression for may be substituted back into Eq. (16.9) to give a nonparametric version of the equation of a line:
y = ( y n )n
(16.12)

Physically, this says that the projection of y onto n equals y itself. Even
equation,
a vector
so it really
though Eq. (16.12) is non-parametric, it is still
represents three simultaneous equations. These three equations are not independent. Given arbitrary choices for y 1 and y 2 , it is not generally possible to
129
T
F
A
R
D
cca
Rebe
Brann

on
solve Eq. (16.12) for y 3 . Certain solvability conditions must be met in order
for a solution to exist; namely, the above equation may be solved for y 3 if and
only if y 1 n 2 = y 2 n 1 . This solvability condition is expressing the requirement
that y 1 and y 2 must fall on the shadow cast by the line onto the 1-2 plane.
Equation (16.12) is non-parametric, but it is still a set of three distinct
equations. It can be expressed as a single equation by noting that two vectors
a and b are equal if and only if ( a b ) ( a b ) = 0 . Therefore, an equiva

lent
version
of Eq. (16.12) is
y y = ( y n )2
(16.13)

Any of the above boxed equations may be used as an equation of a line in
3D space. All but the last equation are vector expressions, so they actually
represent a set of three equations, each of which involves a linear combination
of the position vector components. The last equation is a single (scalar-valued)
expression for a line, but the price paid for having a single equation is that the
position components appear in quadratic form pretty odd given that it represents the equation for a linear (straight) line!
Equation for a sphere

A sphere is defined as the set of points that are equidistant from the
spheres center. Therefore, if a sphere of radius R is centered at a point p ,
then its equation is simply

( x p ) ( x p ) = R2

(16.14)
Equation for an ellipsoid

Suppose you wish to define an ellipsoid in 3D space that is centered at the
origin and whose principle axes are parallel to three vectors { a 1, a 2, a 3 } and
With
whose principal radii are given by the magnitudes of these vectors.
respect to a coordinate system that is aligned with the principle directions, the
equation of the ellipsoid is
x 12 x 22 x 32
-----2- + -----2- + -----2- = 1 , where k = a k
1 2 3
(16.15)
If, however, the principle axes are not coincident with your basis, then the
equation for the ellipsoid takes on a more general quadratic form. Namely,
xBx = 1 ,

where the tensor B is defined by the inverse of the dyad sum
A = ak ak
(16.16)
(16.17)
130
DRAF

Rebe
Since the a k are (by premise) mutually perpendicular, we know that they
must be eigenvectors
of the tensor A . Consequently, the eigenvalues of the
tensor are k a k and the inverse of the tensor is
3
ak ak

B = ------------(16.18)
k4
k=1
Example. Suppose that

a1 = e1 + e2
a2 = 4 ( e1 + e2 )
a 3 = 3e 3
The magnitudes are

1 =
(16.19)
(16.20)
(16.21)
(16.22)
2 = 4 2
(16.23)
3 = 3
(16.24)
The dyads are

a1 a1 = e1 e1 + e1 e1 + e1 e2 + e2 e2

-->
a 2 a 2 = 16 ( e 1 e 1 e 1 e 1 e 1 e 2 + e 2 e 2 )

a 1 a 1 = 9e 3 e 3

-->
000
000
009
110
110
000
-->
(16.25)
16 16 0
16 16 0
0 0 0
(16.26)
(16.27)
Thus,
110
16 16 0
000
1
1
1
[ A 1 ] = -------------4- 1 1 0 + -----------------4- 16 16 0 + ----4- 0 0 0
3
( 2)
(4 2)
000
0 0 0
009
(16.28)
or
17
-----64
1
[ A ] = 15
-----
64
15
-----64
17
-----64
(16.29)
cca B
ranno
1
--9
Thus, the equation for the ellipsoid is
131
T
F
A
R
D
cca
Rebe
Brann

on
15
17 2 17 2 1 2 15
------ x 1 + ------ x 2 + --- x 3 + ------ x 1 x 2 + ------ x 2 x 1 = 1
64
9
64
64
64
(16.30)
Of course, the last two terms, which come from the off-diagonals, may be combined, but we left them separated to emphasize the structure of the solution.
Equation for a cylinder with an ellipse-cross-section

The equation for a cylinder whose cross section is an ellipse could be
regarded as a degenerate case of the above ellipsoid for which one of the specified vectors, say a 3 , is infinite in magnitude. In this case, only the first two
direction vectors, a 1 and a 2 are needed to specify the geometry of the cylin Eq. (16.18) becomes
drical ellipsoid. In this case,
a1 a1 a2 a2
- + ----------- B = -----------
14
24
(16.31)
Equation for a right circular cylinder

The special case of a circular cylinder of radius r follows from Eq. (16.31) by
taking
a1 = a2
(16.32)
1 = 2 = r
(16.33)
where
Then Eq. (16.31) becomes

I nn
e1 e1 + e2 e2

- = ---------------B = ---------------------------
r2
r2
(16.34)
Here, e i is the unit vector in the direction of a i and n is the unit vector in the
of the cylinder axis. The final step in Eq. (16.34)

direction
followed from the
fact that the cylinder axis is perpendicular to the first two directions. With
that result, Eq. (16.16) becomes simply
x x ( n x )2 = r2

(16.35)
Equation of a general quadric (including hyperboloid)

A quadric surface is governed by an equation of the form
xBx = 1
(16.36)

As already discussed, this form is capable of describing an ellipsoid. However,
that is just one class of surfaces described by the form. Without loss, the B
may be presumed symmetric (if not, then it must be symmetrized in order for
the comments in this section to remain true).
132
DRAF

Rebe
An ellipsoid corresponds to the case that the B tensor is positive definite.
If the B tensor possesses negative eigenvalues, then the quadric surface is a
hyperboloid.
Generalization of the quadratic formula and completing the

square
Consider a spatially varying field given by an equation of the following
form:
= x Ax+bx+c

(16.37)
As long as A is invertible, we claim that there exists a symmetric tensor B
such that this field can be written in the form

= ( x p) B ( x p) d
(16.38)

Expanding this out, recognizing the B may be presumed symmetric without
loss, gives
= x B x 2p B x + p B p d

Equating this expression with Eq. (16.37) gives
(16.39)
B = A
(16.40)
2 p B = b
(16.41)
pB pd = c
(16.42)

The last two equations can be solved for p and d . Thus we have the result:
If = x A x + b x + c

then = ( x p ) A ( x p ) + d ,

1 1
1
where p = --- A b and d = --- b A 1 b c
2
4
cca B
ranno
(16.43)
Written out in full,
1
1
1
= x + --- A 1 b A x + --- A 1 b + --- b A 1 b c ,
2
4
2
(16.44)
133
T
F
A
R
D
cca
Rebe
Brann

on
We will be interested in describing the shape of surfaces of constant . Such

surfaces can be described by considering only the case that =0 , because nonzero values of can always absorbed into the parameter c. For the case
that =0 , Eq. (16.37) is like a tensor version of a quadratic formula, so inquiring minds want to know what is the analog of the quadratic formula. In particular, what is the analog of the discriminant? To answer this question, lets
define
y 2A x + b

or, solving for x ,
1
x = --- A 1 ( y b )
2
(16.45)
(16.46)
With this change of variables, Eq. (16.44) becomes

4 = A 1 : [ y y + b b ] 4c
(16.47)
Suppose that =0 . Then, if this result were for scalars rather than tensors, it
would read
1
0 = --- [ y 2 + b 2 ] 4c ,
a
y = b 2 4ac
or
(16.48)
Thus, we see that the vector y plays a role similar to that played by the dis As a matter of fact, note that Eq. (16.46) can
criminant in scalar quadratics!
be written
x = ( 2 A ) 1 ( b + y ) ,
(16.49)

b b 2 4ac
which is the analog of the scalar quadratic formula, x = ---------------------------------------- . In
2a
the scalar quadratic equation, there were only two values of y that would
make =0 . By contrast, for the tensor case, setting =0 gives
0 = A 1 : [ y y + b b ] 4c
(16.50)
or, written in a somewhat different manner

y
y * A 1 y * = 1 , where y * -------------------------------------------
b A 1 b 4c
(16.51)
which describes a quadric surface (i.e., a family of y vectors exists that will
make =0 ).
134
DRAF

Polar decomposition
Rebe
Our ability to delude ourselves may be

an important survival tool.
Lily Tomlin
17. Polar decomposition
The Q-R decomposition
A basic theorem from matrix analysis states that any matrix [ F ] can be
decomposed into the form
[ F ] = [Q][][ R] ,
(17.1)
where [ ] is diagonal, and the matrices [ Q ] and [ R ] are orthogonal (a.k.a.

unitary). Since [ Q ] is orthogonal, we know that
[ Q ]T [ Q ] = [ I ]
(17.2)
Therefore, Eq. (17.1) may be written without loss in the form

[ F ] = [ Q ][ ][ Q ]T [ Q ][ R ] ,
(17.3)
[ F ] = [ V ] [ r ] , where [ V ] [ Q ] [ ] [ Q ] T and [ r ] [ Q ] [ R ]
(17.4)
or
Similarly, since [ R ] is orthogonal, we know that

[ R ][ R ]T = [ I ]
(17.5)
Therefore, Eq. (17.1) may be written without loss in the form

[ F ] = [ Q ][ R ][ R ]T [ ][ R ] ,
(17.6)
[ F ] = [ r ] [ U ] , where [ U ] [ R ] T [ ] [ R ] and [ r ] [ Q ] [ R ]
(17.7)
or
Note that [ r ] is itself orthogonal and that [ U ] and [ V ] are symmetric. Thus,
we conclude the Q-R decomposition guarantees the existence of symmetric
matrices [ U ] and [ V ] as well as an orthogonal matrix [ r ] such that
[ F ] = [r][U ] = [V ][r]
cca B
ranno
(17.8)
135
T
F
A
R
D
cca
Rebe
Brann

Polar decomposition
on
The polar decomposition theorem:

In tensor analysis (especially in continuum mechanics), the Q-R decomposition theorem is called the polar decomposition theorem. Specifically, for
each invertible tensor F there exists a unique orthogonal tensor R , a unique
symmetric positive definite right stretch U , and a unique symmetric

posi
tive definite left stretch V such that
F = RU = VR
(17.9)

A tensor is called a stretch if it is symmetric and positive definite. Physically, a
stretch describes a deformation in which there is an orthonormal triad of
material vectors that change length but not orientation. These special vectors
are the principal directions of the stretch and the corresponding eigenvalues
are the ratios of deformed length to undeformed length. A physical interpretation of the polar decomposition is shown in Fig. 17.1. Note that R U repre
sents a stretch U followed by a rotation R .
In the polar decomposition theorem, the descriptors symmetric and positive definite are requirements, not consequences. A stretch must be both
symmetric and positive definite merely showing that a tensor F is symmet
ric is not sufficient to prove that it is a stretch. For example, any rotation of
exactly 180 will produce a symmetric F tensor, but it will not be positive def
inite.
The procedure for finding R , U , and V is as follows:

Construct a tensor C = F T F , and perform a spectral analysis on it to
find its eigenvalues C k and eigenprojectors P k .* Then
1
/
2
C = C 1 P 1 + C 2 P 2 + C 3 P 3 and U = C
= C1 P1 + C2 P2 + C3 P3 .
The positive square root must be taken to make U positive definite.
1
Compute R = F U .

Compute V = R U R T .

Alternatively, if only V is desired, it may be computed by V = ( F F T ) 1 / 2 .

and then the rotation would

be given by R = V 1 F .
* If p k is the eigenvector associated with C k , then the eigenprojector P equals the dyad
k
p k p k , where there is no sum on k.
By computing U in the principal basis of C , we have ensured that it will be symmetric. In general,
an N N positive definite symmetric matrix like C can have an infinite number of square roots,
N
of which only 2 are symmetric and only one is symmetric and positive definite.
136
DRAF

Polar decomposition
Rebe
Compress the
material by half
along the ellipse
1-axis.
E2
1U
2U
First rotate
the material
counterclockwise
by 60 degrees.
then rotate
the material
counterclockwise
by 60 degrees.
g
1
g
2
E1
Then compress the

material by half
along the ellipse
1-axis.
1V
2V
Figure 17.1. Visualization of the polar decomposition. This figure shows that the deformation F can be visualized as two step process. The upper R U path first applies the stretch
of the vector labeled U and
U to compress the material by a factor of 1 2 in the direction
1
then
the rotation tensor R rotates counterclockwise by 60 . The same deformation
is
achieved on the bottom path by first rotating by 60 and then compressing by a factor of 1 2
in the direction of the vector labeled 1V . In these figures, we have painted a circle (or
sphere in 3D) on the reference cube to show how it becomes an ellipse (or ellipsoid in 3D).
The vectors kU and kV lie on the major axes of the ellipse.
cca B
ranno
137
T
F
A
R
D
cca
Rebe
Brann

Polar decomposition
on
Incidentally, because R is orthogonal, it follows that R = R T . Therefore,
the rotation can be alternatively

be computed by
R = F T U
or by
R = V F T

which is sometimes more computationally convenient.
(17.10)
The polar decomposition is a nonlinear projection operation

Recall that the reference stretch tensor is computed by
U = P( F ) ,
(17.11)
where the operator P is defined for any argument X by
1
/
2
T
P( X ) = +( X X )
(17.12)
We have inserted the + sign in front of this equation to reiterate that the
square root of a tensor is not unique, but the positive definite square root is
unique. The transformation P is a projection in the sense defined in Eq. (6.36)
because P ( P ( X ) ) = P ( X ) . In other words, if F is already a symmetric posi
tive definite tensor, then its associated polar stretch is itself.

Similarly, the rotation tensor, can be regarded purely as a function of F as
follows:
R = F U 1 = F [ + ( X T X ) 1 / 2 ] 1 = Q ( F )
(17.13)

where the operator Q is defined for any argument X by
1
/
2
T
Q( X ) = X ( X X )
(17.14)
The operator Q is a projector because Q ( Q ( X ) ) = Q ( X ) . In other words, if F
is already a rotation, then its associated polar rotation tensor is itself.

Neither of the projectors, P or Q , are linear operators. The codomain is not
a linear manifold for either of these operators. The codomain for P is the set of
all positive definite symmetric tensors. Its true that the sum of two positive
definite tensors is itself positive definite; however, an arbitrary linear combination of positive definite tensors (i.e., a general combination that permits
negative coefficients) will not be itself positive definite. The codomain of Q is
the set of all orthogonal rotation tensors. This is not a linear space because
adding two orthogonal tensors will not result in an orthogonal tensor.
138
DRAF

Polar decomposition
Rebe
The *FAST* way to do a polar decomposition in two

dimensions
For planar deformations, the deformation gradient tensor has the form
F 11 F 12 0
[ F ] = F 21 F 22 0
0
, where F 33 > 0
(17.15)
0 F 33
We claim by serendipity* that the polar decomposition can be performed

rapidly by the following formula:
[ R] =
cos sin 0
sin cos 0 ,
0
0 1
[ U ] = [ R ]T [ F ]
and
(17.16)
where
F 11 + F 22
cos = ------------------------------------------------------------------------( F 11 + F 22 ) 2 + ( F 21 F 12 ) 2
F 21 F 12
sin = ------------------------------------------------------------------------( F 11 + F 22 ) 2 + ( F 21 F 12 ) 2
(17.17)
Beware! You must define cos and sin separately in order to uniquely determine the rotation angle. It is certainly true that
tan = ( F 21 F 12 ) ( F 11 + F 22 ) ,
cca B
ranno
(17.18)
but this relation does not uniquely define the rotation angle because there are
always two angles in the range from 0 to 2 that satisfy the above equation.
By contrast there is only one angle in the range from 0 to 2 that satisfies
Eq. (17.17). The following code fragment may be used to compute a polar rotation tensor in 2-D:
c = F(1,1)+F(2,2); s = F(2,1)-F(1,2);
d = Sqrt(c*c+ s*s);
c=c/d;s=s/d;
R(1,1)=c;R(2,1)=-s;R(1,2)=s;R(2,2)=c;
* Its easy to verify that our formulas yield an orthogonal [R] and a symmetric [U]. It is straightforward, but tedious, to also prove that our formula gives a positive definite [U] matrix. This property
is essential in order to confirm the validity of our serendipitous formulas.
139
T
F
A
R
D
cca
Rebe
Brann

Material symmetry
on
Advice is what we ask for when we already

know the answer, but wish we didnt.
Erica Jong
18. Material symmetry
The symmetry of a material is measured by how its properties vary with
material orientation. If the material properties are unaffected by material orientation, then the material is said to be isotropic. If the material properties
are unaffected by rotation about some given vector a , as for fiber-reinforced
plastics, then the material is transversely isotropic.
Similar concepts apply to tensor analysis. If the components of a tensor are
unchanged upon an orthonormal right-handed change of basis, then the tensor is said to be isotropic. If the components of a tensor are unchanged for any
rotation of the basis about some given vector a , then the tensor is trans
versely isotropic.
Material and tensor theory are covered in a rigorous and elegant manner
through the use of group theory. Here will only give a simple overview of the
results of greatest use in material modeling.
Isotropic second-order tensors in 3D space

In earlier sections, we introduced the notion of vector and tensor invariants. Ordinarily, vector and tensor components will change upon a change in
basis and hence components are numbers, but not scalar invariants. Now we
ask what relationships must be satisfied if the components themselves are to
remain unchanged upon a change of basis. Such relationships define isotropic
tensors. A vector or tensor is called isotropic if its components are the same
with respect to any orthonormal basis. Any linear combination of isotropic
tensors is itself an isotropic tensor. Therefore, the set of all isotropic tensors
(of a given order) forms a subspace. By this, we mean that there exists a finite
number of primitive isotropic tensors that serve as a basis such that any isotropic tensor may be expressed as a linear combination of the primitive isotropic base tensors.
First we will prove that the only isotropic vector is the zero vector. Referring to Eq. (13.9), a vector is isotropic if and only if
140
DRAF

Material symmetry
Rebe
v j = v i Q ji for all orthogonal Q ji

or, in matrix form, { v } = [ Q ] { v } for all orthogonal [ Q ]
(18.1)
Since this must hold for all orthogonal [ Q ] , it must also hold for any particular choice for an orthogonal [ Q ] . Choosing
0 1 0
[ Q ] = 1 0 0
0 0 1
(18.2)
shows that one restriction on the vector components must be v 1 =v 2 . Other

choices for the orthogonal tensor eventually reveals that not only must the
components of the vector all be equal, the only way to satisfy Eq. (18.1) is for
all of the components to be zero. Hence, the only isotropic vector is the zero
vector.
Referring to Eq. (13.13), a second-order tensor in 3D space is isotropic if
and only if
[ T ] = [ Q ] [ T ] [ Q ] T for all orthogonal [ Q ]
(18.3)
This condition implies that the tensor must be a scalar times the identity
tensor I . In other words, the most general form for an isotropic tensor refer
enced to 3D space is
0 0
0 0
0 0
(18.4)
where is an arbitrary scalar. Isotropic tensors in 3D space is therefore onedimensional since only one scalar is needed. The identity tensor I is a basis
for this space. Any general tensor B may be projected to its isotropic
part by
the operation
I ( I : B)

iso B = -----------------I :I

Note that I : B = tr B and I :I = trI = 3 . Hence,

1
iso B = --- ( tr B )I
3
cca B
ranno
(18.5)
(18.6)
This is a very familiar result. The idea of finding the isotropic part by projecting to the space of isotropic tensors becomes less obvious when considering
tensors in spaces of different dimensions.
141
T
F
A
R
D
cca
Rebe
Brann

Material symmetry
on
Isotropic second-order tensors in 2D space

The structure of isotropic tensor components depends on the dimension of
the space. The most general form for an isotropic tensor referenced to 2D
space is

(18.7)
The most general form for an isotropic second-order tensor in two dimensions
therefore contains two arbitrary parameters. Any isotropic tensor in 2D space
may be expressed as a linear combination of the following two primitive base
tensors:
[I] = 1 0
01
and
[] = 0 1
1 0
(18.8)
Note that is the 2D version of the permutation symbol; namely, ij is zero if

if ij=12 , and it is 1 if ij=21 . In two dimensions, the isotropic part
i= j , it is +1
of a second-order tensor F would be obtained by projecting the tensor onto the
space spanned by the basis

in Eq. (18.8). This basis is orthogonal, but not normalized, so the appropriate projection operation is
I ( I :F ) ( : F )
isoF = ---------------- - + --------------- I:I
:

1
1
= --- I ( I :F ) + --- ( :F )
2
2
I ( F 11 + F 22 ) ( F 12 F 21 )
- + ------------------------------= -------------------------------2
2
(18.9)
In component form,
1 F + F 22 F 12 F 21
isoF = --- 11
2 F F F +F
21
12
11
22
(18.10)
Incidentally, the 2D polar rotation tensor associated with any 2D F tensor is
proportional to iso F . Specifically,
F + F 22 F 12 F 21
1
R = -------------------------------------------------------------------------- 11
( F 11 + F 22 ) 2 + ( F 12 F 21 ) 2 F 21 F 12 F 11 + F 22
for 2D space only!.
(18.11)
This formula provides a particularly convenient (and barely known) method

for finding the polar decomposition for 2D problems.
142
DRAF

Material symmetry
Rebe
cca B
ranno
Isotropic fourth-order tensors

The most general form for a fourth-order isotropic tensor c ijrs is
c ijrs = ( ij rs ) + ( ir js ) + ( is jr )
(18.12)
For a proof, see, for example, Ref. [1]. The above expression has three arbitrary parameters, so the space of isotropic fourth-order tensors is threedimensional,* with base tensors given by the parentheses above. The basis
tensors in the parentheses of Eq. (18.12) are only one choice for the basis. In
material modelling, a more convenient basis for the space of isotropic fourthorder tensors is
iso = 1
--- ij kl
P ijkl
3
1
symdev = 1
--- ( ik jl + il jk ) --- ij kl
P ijkl
2
3
skew = 1
--- ( ik jl il jk )
P ijkl
2
(18.13a)
(18.13b)
(18.13c)
Note that these are all constructed from linear combinations of the primitive
bases of Eq. (18.12). Even though the component formulas for this basis are
considerably more complicated, the properties of this basis are irresistible.
Specifically, the basis of Eq. (18.13) consists of complementary projectors! By
this we mean
iso P iso
iso
P ijkl
klmn = P ijmn ,
iso P symdev = 0 ,
P ijkl
klmn
iso P skew = 0
P ijkl
klmn
(18.14)
iso
symdev P iso
P ijkl
klmn = 0 ,
symdev P symdev = P symdev ,

P ijkl
klmn
ijmn
symdev P skew = 0
P ijkl
klmn
(18.15)
skew P iso
P ijkl
klmn = 0 ,
skew P symdev = 0 ,
P ijkl
klmn
skew P skew = P skew

P ijkl
klmn
ijmn
(18.16)
Recall that second-order tensors dont really take on any meaning until
they act on a vector. Likewise, the meaning of a fourth-order tensor should be
inferred by operating on a second order tensor. Specifically, for any tensor B ,
we note
1
iso B
-B
P ijkl
kl = -3 kk ij
(18.17)
1
symdev = 1
--- ( B ij + B ji ) --- B kk ij
P ijkl
2
3
(18.18)
1
skew B
- ( B B ji )
P ijkl
kl = -2 ij
(18.19)
* The space is two dimensional if one imposes a minor symmetry restriction that c ijrs =c jirs =c ijsr .
Thats why isotropic elastic materials have only two independent stiffness moduli.
143
T
F
A
R
D
cca
Rebe
Brann

Material symmetry
on
iso returns the isotropic part of B , P symdev returns the symmetricThus, P ijkl
ijkl
skew returns the skew-symmetric part of skew-symdeviatoric part of B , and P ijkl
metric part of B .
Transverse isotropy
A material property is said to be transversely isotropic about a given unit
vector a if the material can be rotated about this vector without changing the
property. If the property is a scalar invariant, then by definition it
material
cannot change when the material is rotated. The concept of transverse isotropy becomes increasingly complicated for vector and higher order tensor
properties.
In what follows, we often consider a material orthogonal basis
{ p 1, p 2, p 3 } such that p 3 = a . This basis will not generally coincide with the
basis { e , e , e } used in general calculations. Components with
laboratory
1 2 basis
3
respect to the material
are called material components. Components
with respect to the laboratory basis are called laboratory components. Our
goal in what follows is to show the simplified material components of transversely isotropic vectors and tensors and then to show how those tensors may
be expressed directly in the laboratory components without having to perform
a coordinate transformation.
A vector u is transversely isotropic if and only if its material components
are of the form

0

{u} = 0
u
3
in the material basis { p i }
(18.20)
In other words, the only vector that remains unchanged when rotated about
the material symmetry direction a is one that is itself proportional to a .
Thus, since a is a unit vector, the most

general form for a transversely isotro
pic vector u is
u = u3 a
Thus, in the laboratory basis,

a1

u = u3 a2

a3
(18.21)
in the laboratory basis { e i }
(18.22)
Now consider a general second-order tensor A . To be transversely-isotro

pic, the material components of must be of the form
144
DRAF

Material symmetry
Rebe
At As 0
[ A] =
As At
Aa
(18.23)
Here, the subscripts t and a indicate transverse and axial (relative to the
symmetry axis), and the subscript s indicates shear. If the tensor is symmetric, then it cannot have shear terms and its most general form is
At 0
[ A] =
0 At 0
0 0 Aa
(18.24)
In other words, the tensor must be diagonal in the material basis, with the 11
and 22 transverse components equal. Thus, transverse isotropy means the
tensor must be isotropic in the transverse plane. In dyadic notation, the above
equation may be written
(18.25)
A = At ( p1 p1 + p2 p2 ) + Aa p3 p3

Recall that Eq. (9.16) is true in any orthogonal system; therefore it holds for
the material coordinate system and Eq. (18.25) can therefore be written
(18.26)
A = At ( I p3 p3 ) + Aa p3 p3

or, since p 3 = a
A = At ( I a a ) + Aa a a
(18.27)
Rearranging a bit gives the most general direct notation form for a transversely isotropic second-order tensor.
(18.28)
The tremendous advantage of Eq. (18.28) over the matrix representation of

Eq. (18.23) is that Eq. (18.28) can be computed directly in the laboratory basis
without having to perform a coordinate transformation. Namely, in the laboratory basis,
a1 a1 a1 a2 a1 a3
100
[ A ] = At 0 1 0 + ( Aa At ) a a a2 a2 a a
2 1
2 3
001
a3 a1 a3 a2 a3 a3
cca B
ranno
(18.29)
Furthermore, the Eq. (18.28) is more useful when the transverse axis is
changing dynamically because Eq. (18.28) can be differentiated with respect
to time.
145
T
F
A
R
D
cca
Rebe
Brann

Abstract vectors/tensor algebra
on
I find the harder I work, the more luck

I seem to have.
Thomas Jefferson
19. Abstract vectors/tensor algebra
Definition of an abstract vector
Mathematicians define vectors as being members of a vector space. A vector space consists of certain basic components:
A1.
A field R must exist. (For this report, the field is the set of reals; for an excellent
definition of the term field, see Ref. [28].)
A2.
There must be a discerning definition of membership in a set V.
There must be a rule for multiplying a scalar times a member v of V. Further

more, this multiplication rule must be proved closed in V :
If R and v V then v V
A4. There must be a rule for adding two members of V.

Furthermore, this vector addition rule must be proved closed in V :
If v V and w V then v + w V

A5. There must be a well defined process for determining whether two members of
V are equal.
A3.
A6.
The multiplication and addition rules must satisfy the following rules:
v + w = w + v and v = v

u + (v + w) = (u + v) + w

There must exist a zero vector 0 V such that v + 0 = v .

Unfortunately, textbooks seem to fixate on item A6, completely neglecting the

far more subtle and difficult items A2, A3, A4, and A5. Engineering vectors
are more than something with length and direction. Likewise, engineering
vectors are more than simply an array of three numbers. When people define
vectors according to the way their components change upon a change of basis,
they are implicitly addressing axiom A2. In general, axiom A2 is the most difficult axiom to satisfy when discussing specific vector spaces. Verifying
axiom A2 demands introducing rather awkward specialized tests that apply to
particular sets, so dealing with axiom A2 apparently does not appeal to mathematicians. Engineers, on the other hand, must focus principally on this
axiom!
146
DRAF

Rebe
Inner product spaces

An inner product (here denoted ( a, b ) ) is a scalar-valued binary* operation
satisfy the following rules:
between two vectors, a and b , that must
A7. ( a, b ) = ( b, a )

A8. ( a, a ) > 0 if a 0 and ( a, a ) = 0 only if a = 0 .

An inner product space is just a vector space that has an inner product.
Ordinary engineering vectors form an inner product space. For engineering vectors, the inner product is computed by
( a, b ) a b = a 1 b 1 + a 2 b 2 + a 3 b 3 = a k b k

For second-order tensors, the inner product is
(19.1)
( A, B ) A : B = A ij B ij
(19.2)

The choice of inner product is not unique. For many engineering applications,
it is often convenient to define an alternative energy inner product that uses a
symmetric positive definite weighting tensor W :
( a, b ) a W b
(19.3)

Note that it is essential that W be symmetric in order to satisfy A7 and posi
tive definite to satisfy A8. Physically,
the word energy is an apt descriptor for
the second-order analog of Eq. (19.3). For continuum elasticity problems, a
material having a strain and a stress has an internal energy given by
u = 1--2- : E : ,
(19.4)

where E is the fourth-order elastic stiffness tensor. Thus, 1--2- E plays a weigh
ing rolemuch like the tensor W in Eq. (19.3).
Continuous functions are vectors!

Many seemingly unrelated things can be shown to be vectors by the above
abstract, axiom-based, definition. Consider, for example, the following discerning definition for a set: V is the set of real continuous functions. Applying
the above axioms shows that this is a vector space. Furthermore, if f and g
are two continuous functions, then their inner product may be defined as
fg =
f ( x ) g ( x ) dx
cca B
ranno
(19.5)
* The term binary is just an obnoxiously fancy way of saying that the function has two arguments.
147
T
F
A
R
D
cca
Rebe
Brann

on
Note the similar structures: In (19.1), there is a summation over the dummy
subscript, k. In (19.5), there is an integration over the dummy argument, x.
Once something is discovered to be an inner product vector space, a whole
wealth of already-proved theorems becomes available. For example, the angle
between two vectors is defined by
ab
(19.6)
cos = ---------------------------------
aa bb

We could likewise use Eq. (19.5) to define the angle between two continuous
functions. Recall that the dyad between two ordinary vectors a and b is
defined to be a new object such that ( a b ) v = a ( b v ) for all vectors v .
two
real
continuous

Likewise, you can define a dyad between

functions f and
g to be a new object such that
( f g) h = f ( g h)
(19.7)
Recall that a dyad a b is a rudimentary tensor i.e., a new object that
needs two indices to describe

it so that its indicial expression is a i b j . Likewise, the dyad between continuous functions is interpreted as a special kind of
new object: functions of two variables. Thus ( f g ) ( x, y ) = f ( x ) g ( y ) . Recall
that dyads were just special kinds of tensors. Likewise f g is a special kind
of function of two variables separable. Recall that the vector equation
z = W v is written in indicial form as

zi =
W ij v j
(19.8)
j=1
Analogously, given continuous functions z ( x ) , v ( y ) , and a binary function

W ( x, y ) the operation z = W v would be interpreted as
z( x) =
W ( x, y )v ( y ) d y
(19.9)
Once you recognize this parallel structure between continuous functions

and ordinary engineering vectors, you will find that reading advanced calculus
textbooks becomes a lot easier. You will see that you already know many of the
things that those books teach. The major distinction is that ordinary engineering vectors belong to finite-dimensional spaces (they only need 3 base vectors)
whereas the space of continuous functions is infinite dimensional. As with
ordinary vectors, you can define a basis for continuous functions, but the basis
contains an infinite number of members. For example, a Taylor series is an
expansion of a continuous function in terms of the power basis { 1, x, x 2, } . A
Fourier series is an expansion in terms a different (trigonometric) basis.
148
DRAF

Rebe
Tensors are vectors!

For applications in advanced continuum mechanics, it is extremely useful
to recognize that second-order tensors form a nine-dimensional inner-product
vector space. Using the rules for scalar multiplication and tensor addition, one
can readily verify that the axioms A1 through A6 are satisfied by tensors. The
inner product between two tensors is defined in Eq. (12.1) as
A : B A ij B ij
(19.10)

which satisfies the inner product axioms, A7 and A8. This definition of the
tensor inner product sums pairwise over all components just like Eq. (19.1).
By pairwise, we mean that, for example, A 12 is multiplied by B 12 , not by B 21 .
There are an infinite number of ways to generate a scalar from a binary operation between two tensors. For example, some authors wrongly refer to
A B = tr ( A B ) = A ij B ji as an inner product. However, this operation is

not an inner product because it fails the positivity axiom A8. Specifically, consider
1 3 0
[ A] = 4 2 0 .
0 0 1
(19.11)
Then
A :A = A 11 A 11 + A 12 A 12 + + A 33 A 33 = 1 + 9 + 16 + 4 + 1 = 31 > 0

but
(19.12)
A A = A 11 A 11 + A 12 A 21 + + A 33 A 33 = 1 12 12 + 4 + 1 = 18 < 0 , (19.13)

which proves A B is not acceptable as an inner product.

Both A :A and A A are scalar invariants of the tensor A . Note that

A A = tr ( A 2 ) , which may be written in terms of the characteristic invariants

as I 12 2I 2 . For symmetric tensors, A A = A :A . However, for nonsymmetric

tensors, the inner product invariant A :A is not generally expressible as a

function of the characteristic invariants, and it is therefore a fourth independent invariant. To prove this last statement, consider
00
[A] = 0 0 0 ,
0 0 0
cca B
ranno
(19.14)
149
T
F
A
R
D
cca
Rebe
Brann

on
where is an arbitrary scalar. The characteristic invariants of this tensor are

all equal to zero. However, the inner product invariant is A :A = 2 , which

can take on different values, depending on the value of and which cannot
therefore be expressible as a function f ( I 1, I 2, I 3 ) , of the characteristic invariants.
Vector subspaces
Consider a set S that is a proper subset of a vector set V . The set S is
called a subspace if any linear combination of members of S is itself in S . A
subspace S is itself a vector space because all of the vector space axioms that
were proved for the larger vector set V continue to hold for set S . Proving that
something is a subspace really only requires proving axioms A3 and A4 with
axiom A2 replaced by the more restrictive discerning definition for membership in S .
Again, a subspace is itself a vector space. It is generally of lower dimension
than V . In ordinary 3D space, a subspace is like a plane any linear combination of vectors in a plane is itself in the plane. By contrast, consider the set
of all unit vectors is not a subspace because a linear combination of two unit
vectors is not itself a unit vector. Recalling that second-order tensors are
themselves nine-dimensional vectors, we encounter tensor subspaces as well.
The set of all symmetric second-order tensors is a tensor subspace because
any linear combination of symmetric tensors is itself a symmetric tensor. The
set of all orthogonal tensors is not a tensor subspace because a linear combination of orthogonal tensors is not itself orthogonal. The set of all transversely
isotropic tensors is not a subspace, but the set (which we will denote TIa) consisting of all tensors that are transversely isotropic about a particular direction a does form a subspace. Specifically, referring to Eq. (18.28) we note that
the tensors, I and a a form a basis for TIa. This basis, is not however, orthog
onal because I : ( a a ) 0 . Referring instead to Eq. (18.27), we could alterna

tively define
P = I aa

Q = aa
(19.15)
These tensors form an orthogonal basis for TIa. The basis is orthogonal
because, recalling that a is a unit vector, P :Q = 0 . Furthermore, note that

P :P = 2 and Q :Q = 1

(19.16)
150
DRAF

Rebe
cca B
ranno
Hence, this basis is orthogonal, but not normalized. We could naturally divide
P by 2 to normalize the basis, but doing so would make the base tensor no
longer a projection. Normalizing the basis would be wise if our applications

frequently required the tensor inner product of transversely isotropic tensors.
Most engineering applications require just the single dot (composition) product between two transversely isotropic tensors, in which case leaving the basis
as projectors would be more convenient to utilize the fact that P P = P .

Subspaces and the projection theorem

Whenever a subspace S is discovered, it is almost always wise to identify
the operation that will project vectors in V into parts that are in S and orthogonal to S . Many seemingly unrelated theorems are all in fact applications of
the projection theorem. Consider, for example, the set of all even continuous
functions i.e., they have the property f ( x ) = f ( x ) . This set is subspace
because any linear combination of even functions will also be even. A famous
theorem states that any function can be written as the sum of an even plus an
odd function. This theorem is an application of the projection theorem. Note
that odd functions are orthogonal to even functions because the integral in Eq.
(19.5) will be zero if f is even and g is odd.
The set of all matrices that have a zero trace is a subspace. An oft-used theorem states that any matrix can be written as the sum of its isotropic and
deviatoric (traceless) parts. This is the projection theorem again. Likewise any
matrix can be written as the sum of its symmetric plus antisymmetric parts.
This is the projection theorem.
Abstract contraction and swap (exchange) operators *

Virtually all operations in tensor algebra can be written using a rather
arcane but nevertheless conceptually useful generalized notation in
which the arguments are first multiplied dyadically and then operated on by
either contraction or swap operators, as defined below. These operators are
almost never used in practice, but it is often useful to visualize simpler operations in terms of contractions and swaps in order to determine properties of
the operations.
Any vector v is expressible as components v i times base vectors e j . Any
T is expressible as components T ij times base dyads

second order tensor
e i e j . In general, any N th -order tensor may be expressed as a sum of compo (having a total of N indices) times N dyadically multiplied base vectors.
nents
* This section is rather mathematical and it may be skipped without loss in continuity.
151
T
F
A
R
D
cca
Rebe
Brann

on
The contraction operator C IJ is defined to dot the I th base vector with the
base vector. Here, I and J are not tensor indices they merely indicate
the positional location of a dyadically multiplied base vectors. The contraction
operation is well-defined only when operating on second- and higher-order
tensors. For example
J th
C 12 ( a b ) = C 12 ( a i b j e i e j )

C 12 ( T ) = C 12 ( T ij e i e j ) =
2
2
C 1 = C 1 ( ijk e i e j e k )

C 13 = C 13 ( ijk e i e j e k )

= a i b j e i e j = a i b j ij = a i b i = a b

T ij e i e j = T ij ij = T ii = trT

= ijk e i e j e k = iik e k

= ijk ( e i e k )e j = iji e j

C 24 H = C 24 ( H ijkl e i e j e k e l ) = H ijkl ( e j e l )e i e k = H ijkj e i e k

(19.17)
The contraction operator reduces the order of the operand by two. If, for
example, the operand is a second-order tensor, then the result is a zero-order
tensor, which is a scalar. As seen above, the contraction of a third-order tensor
gives a first-order tensor, which is a vector. The contraction of a fourth-order
tensor is a second-order tensor.
JL dots the I th base vector with the J th vector
The generalized operator C IK
and the K th base vector with the L th vector. This operation reduces the order
by four. To be well defined, all four of the indices, I, J , K , and L , must have
distinct values. The operator is well-defined for tensors of fourth-and higher
order. For example,
42 ( H ) = C 42 ( H
(19.18)
C 13
13
ijkl e i e j e k e l ) = H ijkl ( e i e l ) ( e j e k ) = H ikki
42 ( ) is not the same thing as C 4 ( C 2 ( ) ) . In particuIncidentally, note that C 13

1
1
lar,
42 ( H ) = C 2 ( C 4 ( H ) ) = C 2 ( C 2 ( H ) ) .
C 13
(19.19)
1
1
1
3
JLN are similarly defined.

Higher order contraction operators such as C IKM
The swap (or exchange) operator X IJ is defined to swap the I th base vector with the J th vector. This operation does not affect the order of the operand.
The swap operator is well-defined only when operating on second- and higherorder tensors. For example
152
DRAF

Rebe
cca B
ranno
X 12 ( T ) = X 12 ( T ij e i e j ) = T ij e j e i = T T
2
2
X 1 = X 1 ( ijk e i e j e k ) = ijk e j e i e k

X 13 = X 13 ( ijk e i e j e k ) = ijk e k e j e i

(19.20)
JL swaps the I th base vector with the J th

The generalized swap operator X IK
th
vector and the K base vector with the L th vector. To be well defined, all four
of the indices, I, J , K , and L , must have distinct values. The operator is welldefined for tensors of fourth-and higher order. For example,
42 H = X 42 ( H
X 13
13
ijkl e i e j e k e l ) = H ijkl e l e k e j e i

(19.21)
42 ( ) is not the same thing as X 4 ( X 2 ( ) ) . In particular,

Note that note that X 13
1
1
42 H = X 2 S 4 H = X 4 X 2 H .
X 13
3 1
1 3

(19.22)
JLN are similarly defined.

Higher order swap operators such as X IKM
The contraction operator and the swap operator are almost never mentioned explicitly in publications but they are nonetheless very useful in generating theorems because they are both linear operations. Namely,
C IJ ( A + B ) = C IJ ( A ) + C IJ ( B ) for all scalars
and for all tensors A and B of arbitrary (but equal) order (>2).
(19.23)
X IJ ( A + B ) = X IJ ( A ) + X IJ ( B ) for all scalars

and for all tensors A and B of arbitrary (but equal) order.
(19.24)
The last property is particularly useful for performing integrals of quantities

that differ from each other only in the ordering of their indices. Noting that
any permutation of indices may be obtained through a sequence of swap operations, we recognize that only one of our integrals need be computed. Linearity of the swap operation allows us to immediately write down the solution for
all of the integrals once we have found the solution for the first one.
At the beginning of this section, we mentioned that most tensor operations
may be written as a sequence of either swap or contraction operations acting
on dyadically multiplied operands. For example
A x = C 23 ( A x ) = C 13 ( x A ) = C 12 ( x X 12 ( A ) )

(19.25)
45 ( u v )
u v = C 23

(19.26)
and
153
T
F
A
R
D
cca
Rebe

on
Brann
In the above equation, is the permutation tensor given by ijk e i e j e k , where

u v is a
ijk is the permutation symbol defined in Eq. (3.21). The operand
fifth order tensor having ijklm components given by ijk u l v m . The contrac45 reduces the order from five down to one by contracting the
tion operator C 23
2nd index with the 4th index and the 3rd index with the 5th index to obtain
ijk u j v k , which is indeed the indicial expression for the i th component of
u v . It may appear that we have taken a fairly simple operation (the cross

product)
and written it in an insanely complicated manner.
Keep in mind: we are not advocating the use of the contraction and swap
operations in published analyses. We merely want to drive home the point
that virtually all tensor operations can be expressed using contraction and
swap operators. Contractions and swaps are particularly useful when applied
to tensor-integral calculus. Specifically, suppose that f ( x ) denotes some n th
order tensor valued function of a vector x . If C and X denote
any contraction
and swap operation applicable to f ( x ) , then
C f ( x ) dV = C [ f ( x ) ] dV
(19.27)
(19.28)
X f ( x ) dV = X [ f ( x ) ] dV
In other words, contraction and swap operations commute with integration.

This means that the best context in which to derive new theorems is to do it
for dyadically multiplied operands whenever possible. From this generalized
context, a host of new theorems fall out as corollaries. For example, a very
general form of the Gauss-Ostrogradski theorem states that the integral over
a closed surface of the outward normal n multiplied dyadically by a tensor of
volume integral of the gradient of
arbitrary order can be transformed to the
that tensor.
n (
)d A =
) dV
(19.29)
Here, the notation n ( ) denotes the outward unit normal n to the integra some quantity ( ) of arbitrary tensororder, and ( )
tion surface multiplied
denotes the gradient of that tensor.
Because the contraction operation commutes with integration, one can
immediately write down the classic divergence theorem by taking the C 12
contraction of both sides of Eq. (19.29) to obtain
n (
)d A =
) dV
(19.30)
Another corollary is obtained by inserting the cross product into Eq. (19.29),
which we know we can do because we know that a cross product is expressible
in terms of contractions (which in turn commute with integration). Thus,
n (
)d A =
) dV
(19.31)
154
DRAF

Rebe
Incidentally, despite the presence of the cross product, this theorem is not
equivalent to Stokes theorem.
Consider a particular indicial form of Eq. (19.29) in which the operand is a
second-order tensor:
ni T jk d A
xi T jk dV
(19.32)
The swap operator X IJ is linear and therefore commutes with integration.

Thus, in addition to the above equation, we also have an immediate set of corollaries in which the indices in the above equation may be arbitrarily permuted. Likewise, the contraction operator C IJ is linear and may be commuted
with integration. Thus, any of the indices in the above equation may be contracted.
The contraction tensor. Since the contraction operator C 12 ( T ) is linear, we
know there must exist a second order tensor C such that
2
C 1 ( T ) = C :T = C ij T ij
(19.33)

We know from Eq. (19.17) that
C 12 ( G ) = trT = T kk = ij T ij
(19.34)
Comparing the above two equations, we note that C is simply the identity
as the identity tensor
tensor. Hence, the contraction operation can be viewed
double dotted into two specified base vectors. This point of view is useful when
considering differentiation. The identity is a constant tensor and its derivative
is therefore zero. It is often useful conceptually utilize the contraction in gradient identities. For example,
(19.35)
w = I : w

The contraction is extremely useful in integration identities. For example
w dV
I : ( w ) dV
= I : w d V

(19.36)
These formulas show that knowing the complete gradient tensor w is really
the most useful. You can just take the trace of this total gradient tensor to
immediately obtain the divergence. Likewise, you can take the (negative)
axial part of the total gradient tensor to obtain the curl:
axial w = : w = w
cca B
ranno
(19.37)
The swap tensor. Recall from Eq. (19.20) that the effect of the swap operation
on a second-order tensor T is to simply produce the transpose of that tensor:
155
T
F
A
R
D
cca
Rebe
Brann

on
X 12 ( T ) = T T
(19.38)
This is a linear operation transforming a tensor T into a new tensor T T .
Therefore, this simple operation must be expressible in terms of a fourth-order

swap tensor S such that
(19.39)
T T = S :T

Writing this out in indicial notation reveals that
S = in jm e i e j e m e n
(19.40)
156
DRAF

Vector/tensor calculus
Rebe
You shouldnt put down a loser, Cindy...

you might be one yourself some day.
Carol Brady, The Brady Bunch
20. Vector/tensor calculus
Up to now, this report has focused on vector and tensor algebra, not calculus. Rather than following a standard rigorous (and therefore stilted) development which is available in any good tensor analysis textbook, we will define
our calculus notation through a series of examples. Whenever we write
dy
------- ,
(20.1)
dx
we are implicitly telling the reader that the vector y can be expressed as a
function of a single vector x . Consequently each y i component

must be a func
tion of the three { x 1, x 2, x 3 } components. By the chain rule,*
yi
d y = d y i e i = --------dx j e i
x j
(20.2)
Noting that dx j = d x e j , this may be written

yi
yi
d y = -------- ( d x e j )e i = --------e i e j d x
x j
x j
cca B
ranno
(20.3)
Even if y is a nonlinear function of x , note that the infinitesimal increment

we have now demonstrated that there
d y is linear
with respect to d x . Thus,
exists
a second order tensor, which
we will denote d y d x such that
d
y
d y = ------- d x
(20.4)
d x
* In Eq. (20.2) we have written y i x j on the reasonable assumption that our readers will understand this to mean the partial derivative with respect to x j , holding all other components of x con
stant. Except in this very special case, we strongly recommend explicitly showing what variables
are held constant in partial derivatives whenever there is even the slightest chance of misinterpretation or confusion by your readers. This issue pops up continually in thermodynamics. The partial
derivative of pressure P with respect to volume V takes on different values depending on the conditions at which the measurement is made. If, for example, the variation of pressure with volume is
measured at constant temperature T , then the derivative should be written ( P V ) T . However, if
the measurement is made under reversible insulated (no heat flow) conditions, then the derivative
should be written ( P V ) s , where s is entropy. The two derivatives arent equal!
157
T
F
A
R
D
cca
Rebe
Brann

on
Comparing with Eq. (20.3) we see that

yi
dy
------- = --------e i e j
x j
dx
(20.5)
ASIDE #1: total and partial derivative notation. Note that we have
quietly introduced our notational preference to use a d when writing d y d x
the
even though we must, by necessity, use a in the partial derivative on
right-hand-side. This choice has been made quite deliberately in order to maximize the analogies that tensor calculus has with ordinary scalar calculus. To
get a better appreciation of why we have made this choice, consider the scalar
calculus situation in which some variable y is expressible solely as a function
of an independent variable x . Then, of course, you would naturally write
dy dx , not y x . The very act of writing dy dx tells the reader that y
really can be written solely in terms of x . If it turns out that y can be alternatively written as a function of two other variables, u and v, then we could use
the chain rule to write
y du
y dv
dy
------- = ------ ------- + ------ ------ u v dx v u dx
dx
(20.6)
Note that we use d on the left-hand-side, and on the right. Our situation
with vector derivatives is very similar. We have a vector y that can be computed if we know the single vector x . We say this object is a single vector
because it really is a single entity even

though it consists of multiple parts
(components, base vectors). A velocity vector has three components, but you
wouldnt say the velocity ARE fast. A vector IS the sum (not are the sum) of
components times base vectors. A vector IS a single object. When y = f ( x ) ,
we can regard y to be a function of the three components x 1 , x 2, and x 3 .
Then, in analog with Eq. (20.6) we would write
y dx 1 y dx 2 y dx 3
dy
(20.7)
------- = -------- --------- + -------- --------- --------- + -------x 1 d x x 2 d x x 3 d x
dx
Note the distinctions in the use of d and in this equation. Its not hard to
show that
dx j
--------- = e j
(20.8)
dx
Likewise, since the base vectors themselves are fixed when using rectangular
Cartesian coordinates,
( yi ei )
yi
y
- = --------e
------- = ----------------x j i
x j
x j
(20.9)
Putting Eqs. (20.8) and (20.9) into (20.7), taking care to preserve the order of
dyadic multiplication of the base vectors, gives us back our result of Eq. (20.5).
158
DRAF

Rebe
There are other analogs with conventional scalar calculus that motivate
our use of d on the left hand side of Eq. (20.5). In scalar calculus, we know
that
dx
1
------- = -----------------dy
dy dx
(20.10)
This identity holds if dy dx 0 (otherwise, we would be dividing by zero). The

analog for tensors is
d y 1
dx
(20.11)
------- = ------- ,
dx
dy
which holds only if d y d x is invertible.
Eq. (20.10) was applicable because y was expressible solely as a function

of x . In general,
x
1
------ ---------------y y x
(20.12)
Equality holds only if both partial derivatives hold the same variable constant
in the derivative (in which case we are implicitly restricting our attention to a
world in which that variable is always constant, and in such a world, y
would be expressible solely as a function of x ). To convince yourself that
Eq. (20.12) is correct, consider polar coordinates in which x and y are expressible as functions of r and . Namely
x = r cos
y = r sin
and
(20.13)
from which it follows that

x
------
r
x
---- r
y
------
r
y
---- r
cos r sin
sin r cos
(20.14)
These equations can, in principal be inverted to obtain r and in terms of x

and y . However, to compute ( r x ) y , we can alternatively use implicit differentiation, and we will find that
r
---- x y
r
---- y x

------
x y

------
y x
x
------
r
x
---- r
y
------
r
y
---- r
cos r sin
sin r cos
cos sin
= sin cos
-------------- -----------r
r
(20.15)
Note from Eq. (20.14) that

x
------ = cos
r
cca B
ranno
(20.16)
and Eq. (20.15) tells us that
159
T
F
A
R
D
cca
Rebe
Brann

on
r
----- = cos
x y
(20.17)
Therefore,
r
1
----- ---------------------- x y ( x r )
(20.18)
If (for some strange reason) we wished to know the variation in r with respect
to x holding constant, then we would be able to write
r
1
----- = ---------------------- x
( x r )
(20.19)
The only distinction between the above two equations is what variable is
being held constant. In one case equality holds; in the other it doesnt thats
why it is so essential to explicitly show whats being held constant unless you
are 100% sure there is no chance of confusion. Dont be so sure. In classic multivariable, scalar, calculus, the very act of writing y x tells your reader that
y can not be written solely in terms of x it must be expressible in terms of
x and whatever variable is being held constant in the partial derivative.
Thats why we do not use . When we write d y d x , we want our readers to
know implicitly that y can be written purely as a function

of x and nothing
else. We would write y x only when y depends additionally

on other
things, and we would be diligent to show what variable is held constant in the
derivative if there is any chance for misinterpretation of what this second
variable is.
ASIDE #2: Right and left gradient operations. Note that the ij component
in Eq. (20.5) is y i x j . Some books define the gradient such that the ij component is y j x i . When they do this, they are required to redefine their definition of the vector-vector gradient to be
dy
d y = d x -------
(20.20)
d x
This is fine, although we frown at instances where the gradient conventions are mixed inconsistently.
The nabla or del gradient operator. Researchers who use y j x i
typically do so because this convention permits them to use the alternative
notation y , where the nabla or del operator is heuristically defined such
that
( )
( ) = ----------- e i ,
x i
(20.21)
160
DRAF

Rebe
With this right-operating del, the differentiation with respect to x acts on

quantities to the right. Thus, the ij component of y has indices that are
heuristically similar to the ij component of an ordinary

dyad a y , namely
a i y j . In other words, y would have to be defined such that
y
d y = d x ( y ) , which implies that ( y ) ij = ---------j
(20.22)
x i
To use a heuristic del operator with our definition, however, we would

have to define
( ) = ( ) x j e j ,
which operates on arguments to its left. Then we could write
(20.23)
d y d x = y
(20.24)
Then the ij component of y would be heuristically similar to the ij compo be defined such that
nents of y b . Thus, y would
y
d y = ( y ) d x , which implies that ( y ) ij = --------i
(20.25)
x j
Comparing this with Eq. (20.22) shows that the two types of del operators
are related by a simple transpose
( y ) = ( y )T
(20.26)
Both the backward and forward operating del are used by Malvern and we
also find that they are both useful, depending on the application at hand.
However, for the applications in this report, a del operator is never needed.
We will always use the fraction-like notation of Eq. (20.5), which is technically a left-operating derivative. In general, the indices are placed in order
first on the dependent variable y i and then on the independent variable x j .
Since we will always follow this convention, we may write d y d x as
dy
------(20.27)
dx
without ambiguity.
Our convention of placing indices first on the dependent variable and then
on the independent variable extends to define higher order derivatives as well.
For example, the notation
dU
-------dY
cca B
ranno
(20.28)
would be taken to denote a sixth-order tensor defined by
161
T
F
A
R
D
cca
Rebe
Brann

on
U ij
------------------e e e e e e
Y klmn i j k l m n
(20.29)
Incidentally, note that we use d rather than in our derivative in

Eq. (20.27). By using d , we are implicitly telling the reader that y is a
else.
proper function of the single vector x , and it is not a function of anything
The fact that x itself a function of its three components is irrelevant.* By

than , we are telling the reader that generalizations of
using d rather
familiar differentiation identities continue to hold true for this new type of
derivative. For example, the scalar identity
dx
1
------- = -----------------dy
dy dx
(20.30)
has its analog for tensors as

d y 1
dx
------- = ------- ,
(20.31)
d x
dy
where the superscript 1 is the tensor inverse. Of course, this property holds
only if the inverse exists (i.e., only if det ( d y d x ) 0 ), just as Eq. (20.30) holds
only if dy dx 0 .
In contrast to Eq. (20.27), when we write a derivative with the symbol
instead of the d symbol, we are telling the reader that the dependent variable depends on more than one tensor or vector quantity. For example, by
writing the notation
y
-----(20.32)
x z

we are implicitly telling the reader that y can be written as a proper function
of x and z . The subscript also indicates that the vector z is being held con in the
derivative. Thus, the above derivative quantifies
how the vector y
stant
would change if x were to be varied without changing z . For general varia
tions in both independent variables, the total increment in y must include
contributions from both d x and dz . Specifically, the above partial

derivative
must be defined such that

y
y
d y = ------- d x + ------ dz
(20.33)
x z
z x

As before, if y is a function of x and z , then y i depends on the x j and z k
components. Hence,
* The situation is like having y= f ( x ) and x = g ( u, v ) . By the chain rule,
y
y
dy = ------ du + ------ dv . On the left side, we use d, while on the right side we use .
v u
u v
162
DRAF

Rebe
y
y
------ = --------i e i e j .
x z
x j z

By the way, it is true that
(20.34)
y 1
x
------
= -------
(20.35)
y z
x z

This is true because the same quantity namely, z is being held constant
on both sides of the equation. The property is not true

if different things are
held constant. Thus,
y
x
------ -------
x w
y z

in general
(20.36)
Okay, if the above relation does not hold, does anything LIKE IT hold?
If you really wish to relate ( x y ) z to ( y x ) w , then you must be extra

seeking
careful with your logic. The basic motivation
for
such a relationship in
the first place is that you must not have the actual function for x as a function
some other
of y and z available. Instead, you have y as a function of x and
variable
w . Recall that the very act of writing
( x y ) z tells your reader that
there does exist some function for x as a function of y and

z even if you dont
happen to have that function available at hand. Now, if x is a function of y

to obtain y as a
and z , then it might be possible to solve this relationship
function
of x and z . Now, the very act of writing the expression ( y x ) w ,

implies that y must be expressible as a function of x and some other variable
w . Thus, we have simultaneous implied functions
y = y ( x, z ) and y = y ( x, w )
(20.37)

The only way these can both be true is if
w = w ( x, z )

Now we can apply the chain rule to write
y
y
y
x
w
------ = ------- ------- + ------- ------
x z
x w x z w x x z

or, noting that ( x x ) z is just the identity tensor,
y
y
w
------- = ------- + ------- ------ x z
x w w
x z
x

Putting this into Eq. (20.35) gives
y
y
x
w
------ = ------- + ------
- ------ y z
x w w
x z
x

cca B
ranno
(20.38)
(20.39)
(20.40)
(20.41)
163
T
F
A
R
D
cca
Rebe
Brann

on
Derivatives in reduced dimension spaces

As an introduction to this important topic, suppose that a one-dimensional
curve is defined in 3D space. Such a curve can be defined parametrically such
that for each value of a scalar s there exists a unique point x on the curve. As
a specific example, consider

x = sE 1 + e s E 2
(20.42)
The parameter s is here assumed to vary freely from to . The non-parametric equation for this space curve is simply x 2 = e x1 and x 3 = 0 .
The derivative of x with respect to s is
dx
------- = E 1 + e s E 2
ds
(20.43)
Because the curve is one-dimensional, we know that each position x corresponds to a unique value of the parameter s . Conversely, each value of s is
associated with exactly one position vector x . Hence, s may be regarded as a
proper function of x . However, what exactly

is this function. Inspecting
Eq. (20.22), we can identify

s = x1
(20.44)
but we can equally well identify

s = ln x 2
(20.45)
Both of these expressions will give the correct value for s as long as x lies on
the space curve. However, directly differentiating s with respect to x gives
different answers. We will call these derivatives the nominal derivatives,

and we will denote them with a star:
ds
(20.46)
Eq. (20.44) gives ------- = E 1
d x
E2
ds
Eq. (20.45) gives ------- = -----(20.47)
d x
x2
These two expressions are perpendicular to each other, so they cannot possibly
be equal. The discrepancy must be resolved by using a projection of the increment d x onto the space of allowable increments. From Eq. (20.42), note that
d x = ( E 1 + e s E 2 ) ds
(20.48)
Therefore, the increment in position must always be parallel to the vector

E1 + e s E2

We can define the projector onto this vector as
(20.49)
164
DRAF

Rebe
E1 E1 +
E +
E +
E

21
12
2 2P = ------------(20.50)
- = ----------------------------------------------------------------------------------------------
1 + e 2s

If d x is an arbitrary increment in x that may or may not lie on the space
we can write the set of all allowable
curve,
increments in the form
es E
es E
cca B
ranno
e 2s E
( d x ) = P d x
Therefore, we define the increment in s so that
(20.51)
ds
ds
ds = ------- ( d x ) = ------- P d x
(20.52)
d x
d x
from which it follows that the proper expression for the derivative of s with
respect to allowable increments in position must be given by
ds
ds
------- = ------- P
d x
dx
When applied to Eq. (20.46) this formula gives

E1 + e s E2
ds
------- = ------------------------dx
1 + e 2s
and when applied to Eq. (20.47), it gives
(20.53)
(20.54)
E1 + e s E2
ds
------- = ------------------------(20.55)
dx
1 + e 2s
The above two results are identical and should be used whenever the derivative of s with respect to x is needed. Using this expression for the derivative
also leads to intuitive results

for chain rule applications. For example, the
chain rule tells us that
dx
dx ds
------- = ------- ------ds d x
dx
Substituting Eqs. (20.53) and (20.54) into the right hand side leads to
(20.56)
dx
------- = P
(20.57)
dx
Thus, for this case of constrained increments in x , the derivative of x with

projection operator
that
respect to itself is not the identity tensor it is the
takes you to the space of allowable increments in x . For allowable increments
( d x ) , the projector has the property
P ( d x ) = ( d x )
(20.58)
Thus, the projection operator is the identity tensor along the space curve!
165
T
F
A
R
D
cca
Rebe
Brann

on
A more applied example. In materials modeling, tensors are frequently

symmetric. The projection operator that transforms any tensor into its symmetric part is given by
sym = 1
--- ( ik jl + il jk )
P ijkl
(20.59)
2
For nonlinear elasticity, the stress tensor pq is taken to be a function of the
strain tensor mn . In rate form, nonlinear elasticity is therefore of the form
pq = E pqmn mn
(20.60)
where the tangent stiffness tensor is defined

pq
E pqmn = ----------- mn
(20.61)
Because the strain tensor is symmetric, the tangent stiffness tensor can (and
should) be replaced by
pq
sym
E pqmn = ------------ P ijmn
ij
(20.62)
Doing this will ensure consistent results for derivatives.
166
DRAF

Rebe
Series expansion of a nonlinear vector function

The concept of expanding a nonlinear vector function as a series expansion
is so useful that we will now close this report with it. In undergraduate calculus, the expansion of a function of one variable f ( x ) about a point x o is given
as
1
1
f ( x ) = f ( x o ) + f ( x o ) ( x x o ) + ----- f ( x o ) ( x x o ) 2 + ----- f ( x o ) ( x x o ) 3 +
3!
2!
Now consider a function of two variables, f ( x, y ) . The expansion of this
function about the point ( x o, y o ) is given by
1 o
f ( x, y ) = f o + f xo ( x x o ) + ----- f xx
( x xo ) 2
2!
1
+ f oy ( y y o ) + ----- f oyy ( y y o ) 2
2!
1 o
+ ----- f xy ( x x o ) ( y y o )
2!
1
+ ----- f oyx ( y y o ) ( x x o ) +
2!
(20.63)
where
f o f ( x o, y o )
f ( x, y )
f xo = --------------------- evaluated at ( x, y ) =
x
f ( x, y )
f oy = --------------------- evaluated at ( x, y ) =
y
2 f ( x, y )
o =
f xx
----------------------- evaluated at ( x, y )
x 2
2 f ( x, y )
f oyy = ----------------------- evaluated at ( x, y )
y2
2 f ( x, y )
o = ----------------------f xy
- evaluated at ( x, y )
xy
2 f ( x, y )
o = ----------------------- evaluated at ( x, y )
f xy
xy
( x o, y o )
( x o, y o )
= ( x o, y o )
= ( x o, y o )
= ( x o, y o )
= ( x o, y o )
f oyx
cca B
ranno
(20.64)
Note that the expansion is similar in structure to the expansion of a function

of a single variable except that there is one derivative of every order for every
possible number of ways to select the differentiation variables. Even though
o = f o , this mixed partial derivative still counts twice in the expansion.
f xy
yx
Now consider a field variable f ( x ) . The vector x has three components, so
this is a function of three variables. In direct notation,

the expansion of this
field variable about a point x o is
1
f ( x ) = f o + f o ( x x o ) + ----- ( x x o ) f ( x x o ) +
(20.65)
2!
where
167
T
F
A
R
D
cca
Rebe
Brann

on
f o f ( xo )
df ( x )
f
f o = -------------- - = -------- e i evaluated at x = x o
x
2
2
d f ( x)
f
f = ----------------(20.66)
- = ----------------- e i e j evaluated at x = x o
x i x j
d xd x
o

In Eq. (20.65) the first order derivative f o is a vector and it appears in a vec
tor inner product with ( x x o ) . The second-order
derivative f is a second
o is the same
order tensor, and it is dotted from both sides by ( x x o ) , which

thing as taking the second-order tensor inner product
f : [ ( x x o ) ( x x o ) ]
(20.67)

Following along with this pattern, the next term in Eq. (20.65) would be
1
----- f [ ( x x o ) ( x x o ) ( x x o ) ] ,
(20.68)
3! o

where would denote the third-order tensor inner product* so that the component form of the above expression would be
1 3 f
----- --------------------------3! x i x j x k
[ ( x i x io ) ( x j x oj ) ( x k x ko ) ]
(20.69)
xo
In this case, the component form is more clear than the direct notation form,
so it might be a better choice in publications, especially given that the triple
dot notation is highly nonstandard and might only confuse your readers.
The tensor calculus section will be expanded at a later

date to include, for example, more derivative identities
and the essential integral theorems.
* Clearly, indicating the order of the inner product by the number of stacked dots could lead to
some very weird notation for higher order tensors. We have seen the notation :: used for the fourthorder inner product. A circled number can always be used for higher order inner products.
168
DRAF

Closing remarks
Rebe
cca B
ranno
The ability to quote is a serviceable substitute

for wit.
W. Somerset Maugham
21. Closing remarks
Someday, I will use this space to write something clever or profound. Until
then, this is all you get. Surely youre sick of this report by now?
Please, if you liked this report and if you want it to serve others, then take
a moment to send me an email with your suggestions and corrections at
rmbrann@me.unm.edu.
Thanks!
Rebecca Brannon
169
T
F
A
R
D
cca
Rebe
Brann

Closing remarks
on
170
DRAF

REFERENCES
Rebe
REFERENCES
1Aris,
R., Vectors, Tensors, and the Basic Equations of Fluid Machanics. Englewood Cliffs, N.J.: Prentice-Hall, 1962.
2
Bird, R.B., The dynamics of polymeric liquids. Finish
this.
3Budiansky,
Bernard, (1990) Tensors, in: Handbook of Applied Mathematics 2nd Ed, Edited by Carl E. Pearson. Van Nostrand Reinhold.
4Brannon,
R. M., http://me.unm.edu/~rmbrann/curvilinear.pdf
5Brannon,
R. M., http://me.unm.edu/~rmbrann/Rotation.pdf
6Brannon,
R. M., http://me.unm.edu/~rmbrann/RadialReturn.pdf
7Brannon,
R. M., http://me.unm.edu/~rmbrann/MohrsCircle.pdf
8Brannon,
Rebecca M. and Drugan, Walter J., Influence of nonclassical

elastic-plastic features on shock wave existence and spectral solutions. J. Mech.
Phys. Solids, 41 (2), pp. 297-330.
9Wylie,
C. Ray, and Barrett, Louis C., Advanced Engineering Mathematics, McGraw-Hill (1982).
10Brannon,
Rebecca M. Caveats concerning conjugate stress and strain

measures for anisotropic elasticity. Acta Mechanica, Accepted 1997.
11Stakgold,
cca B
ranno
Ivar, Greens Functions and Boundary Value Problems,
Wiley (1979).
12
Malvern, Lawrence E., Introduction to the Mechanics of a Continuous Medium, Prentice-Hall (1969).
This is one of the best known references on elementary continuum mechanics. HomeLibrary
13Smart,
W.M., Textbook on Spherical Astronomy. 6th Ed. revised by

R. M. Green., Cambridge University Press, 1977.
Contains theorems that are useful in generating the mesh of the sphere. PZT:000322
14Smith,
G.F., Smith, M.M., and Rivlin, R.S., Integrity bases for a symmetric tensor and a vector The crystal classes. Arch. Rational Mech. Anal.,
Vol 12 (1963).
The title says it all. This paper shows the most general quadratic form that can be constructed
from the isotropic invariants for a crystal class. 970806
171
T
F
A
R
D
cca
Rebe
Brann

REFERENCES
on
15
Spencer, A.J.M., Theory of Invariants, in Continuum Physics I, edited by A.C. Eringen Academic, New York (1971).
Derives the integrity bases for arbitrary numbers of vectors and tensors for (i) full orthogonal
group (ii) proper orthogonal (rotation) group, (iii) transverse isotropy, and (iv) crystal classes.
16
Varadan, Vasundara V.; Jeng, Jiann-Hwa; and Varadan, Vijay K., Form
invariant constitutive relations for transversely isotropic piezoelectric materials. J Acoust. Soc. Am. 82 (1) pp. 337-342 (1987).
This paper shows how to derive transversely isotropic constitutive relations using integrity
basis for the transverse orthogonal group of transformations. 970722A
17Continuum
mechanics textbook
from my class.
18Tai,
Chen-To (1997) Generalized Vector and Dyadic Analysis. Applied

Mathematics in Field Theory, 2nd Ed. IEEE Press, NY.
19Simmonds,
James G. A Brief on Tensor Analysis, Springer-Verlag,
NY (1994).
20Chadwick
Continuum mechanics book
21Stewart,
G.W., Introduction to Matrix Computations, in Computer

Science and Applied Mathematics series of monographs and textbooks. Academic Press (1973).
22Arfken,
George B. and Weber, Hans, J., Mathematical Methods for

Physicists, Academic Press, San Diego (1995).
23Greenberg
applied math book.
24
Horn, Roger A., and Johnson, Charles R., Matrix Analysis, Cambridge University Press, Melbourne Australia (1985).
25Mathematica
reference
26
Friedman, Avner, Foundations of Modern Analysis, Dover (1970).
27
Rudin, Walter, Real and Complex Analysis, McGraw-Hill (1987).
28Ramkrishna,
Doraiswami, and Amundson, Neal R., Linear Operator

Methods in Chemical Engineering, Prentice-Hall (1985).
29Folland,
Gerald B., Real Analysis, Wiley-Interscience (1984).
172
DRAF

REFERENCES
30
Rebe
cca B
ranno
Buck, R. Creighton, Advanced Calculus, McGraw-Hill (1978).
31Boccara,
Nino, Functional Analysis: An introduction for Physicists, Academic Press (1990).

32
Rebeccas math handbook that contains the discussion of tensors in terms of dyads.
173
T
F
A
R
D
cca
Rebe
Brann

REFERENCES
on
174
INDEX
A
adjoint 19
adjoint, not the same thing as adjugate 81
adjugate 19
affine 6
alternating (permutation) symbol vs. alternating TENSOR 40
alternating symbol (also called the permutation symbol or
Levi-Civita density) 11
alternating symbol, cyclic property 11
alternating TENSOR 40
alternating tensor 42
angle between two vectors 35
area, computed using cross products 39
axial tensor 42
axial vectors 42
axial vectors, applied to rotation 93
B
basis 6
basis transformations 104
basis, effect of changing 8
basis, for a tensor 60
basis, style of notation 73
binary operation 6
C
calculus 157
Cayley-Hamilton theorem 122
characteristic invariant 17
closed 4
codomain 4
cofactor 15
COFACTOR tensor 80
cofactor-determinant connection 19
complementary projectors 85
completing the square, generalized to vectors/tensors 133
component, of a tensor 71
component, of a vector in direction of another vector 38
components of a vector 25
components, of a vector 36
components, of a vector with respect to NON-orthonormal
basis 37
composition (dot product) between two tensors 78
composition of two tensors 71
constraints 5
Continuous functions are vectors 147
contraction operator 151
contraction tensor 155
coordinate transformations 104
Cramers rule for finding the inverse 82
cross product 39
cross product, associated tensor 91
cross product, between orthonormal base vectors 39
cross product, between two tensors 100
cross product, handedness issues 40
curvilinear 6
cyclic property of the alternating symbol 11
cylinder, equation for 132
D
definition of a matrix 7
definition of a vector 24
definition of a vector, generalized 146
definition of the 3x3 determinant 16
definition of the cofactor 20
definition of the inverse 20
definition of the laboratory basis 25
definition of the permutation symbol 40
derivative, of a vector with respect to a vector 157
Determinant 14
determinant, definition 16
determinant, derivative of 83
determinate, using cofactors 19
Deviatoric tensors 89
deviatoric tensors 97
diagonalizable tensors 125
difference between a matrix and a tensor 112
difference between a tensor and a matrix 108
dimension 5
dimension, of a matrix 7
direct notation 110
direct, style of notation 73
direction cosines 105
domain 4
dot product, between a tensor and a vector 69
dot product, between a vector and a tensor 70
dot product, between orthonormal base vectors 36
dot product, between two tensors 71
dot product, between vectors 35
dot product, dotting a tensor from both sides 71
double dot product 98
double-bar (vector/tensor magnitude) 101
double-dot product 98
dummy index 26
dyad "no-symbol" notation 62
dyad, component matrix 63
dyad, scalar multiplication rule 64
dyadic 66
dyadic multiplication 61
dyadic sums, reducibility 64
DYADS 61
dyads, REDUCIBILITY 65
dyads, sum of two 64
E
e-delta identity 11
e-delta identity, easier mnemonic 12
e-delta identity, fast way 12
e-delta identity, multiple summed indices 13
e-delta identity, three summed indices 14
e-delta identity, two summed indices 14
Eigenprojectors 126
Eigenvalue problems 122
eigenvalues, algebraic vs. geometric multiplicity 123
eigenvector, right and left 122
Einstein, quote 1
ellipsoid, equation of 130
Equation for a cylinder 132
Equation for a sphere 130
Equation for an ellipsoid 130
Equation of a line 129
Equation of a plane 128
Euler angles 94
even functions 37
F
free index 11, 26
functional analysis 3
G
Glide plane 43
gradient 157
gradients, direct notation 75
H
handedness 41
homogeneous functions 37
hyperboloid, equation of 132
hyperplanes 6
I
idempotent 53
identity matrix 10
identity tensor 77
implicit sums as independent variables 29
improper orthogonal tensors 90
indexing, aesthetic/courteous 31
Indicial notation in derivatives 29
indicial, style of notation 73
inertia tensor 114
Inner product spaces 147
inner product, between tensors, identities 101
inner product, of two column matrices 9
invariance 108
invariants, characteristic 118
invariants, more than three 118
Invariants, of a sum of two tensors 121
invariants, primitive 117
invariants, scalar 116
invariants, traces of powers 118
Inverse 19
inverse, Cramers rule 82
inverse, of a function/transformation 5
Inverse, of a tensor plus a dyad 122
inverse, of a tensor plus a dyad 83
irregular 37
Isotropic fourth-order tensors 143
isotropic materials 140
Isotropic second-order tensors, in 2D 142
Isotropic second-order tensors, in 3D 140
isotropic tensors 97
K
Kronecker delta 10
Kronecker delta, index-changing property 33
Kronecker delta, summing 34
L
laboratory base vectors 24
laboratory basis 24, 25
laboratory triad 24
left-handed 41
Levi-Civita density (also called the permutation symbol or
alternating symbol) 11
line, equation of 129
LINEAR 58
linear 4
linear combination 5
linear manifold 6
linear operators/transformations 58
linear space 6
linear transformations, existence of an associated tensor 76
linear transformations, finding the associated tensor 76
linearly independent 5
M
magnitude of a tensor 101
magnitude of a vector 101
Material symmetry 140
matrices, not the same as tensors 112
matrix vs. tensor 108
matrix, definition 7
matrix, dimensions 7
matrix, product 8
minor 15
minor, signed 15
Mohrs circle (footnote) 123
N
nested submatrices 17
notation
three styles (direct, basis, indicial) 73
notation, under-tilde 34
notation, WHY FRET OVER IT? 79
null space 4
O
odd functions 37
one-to-one 5
operators, linear 58
order of a tensor 102
order vs. rank 102
orientation of a vector 38
orthogonal tensors 89
orthogonal tensors, improper 90
orthogonal tensors, proper 90
orthogonal, improper 90
orthogonal, projections 46
outer product, of two column matrices 10
P
permutation (alternating) symbol, handedness issues 40
permutation symbol (also called the alternating symbol or
Levi-Civita density) 11
permutation symbol, cyclic property 11
plane, equation of 128
polar decomposition 135
polar decomposition, as a nonlinear projection 138
polar decomposition, fast method in 2D 139
principal directions 122
principal minors 17
principal submatrix 17
projection decomposition, finding magnitudes 101
projection of a vector onto another vector 38
projection theorem 55
projection, and subspaces 151
projection, of a vector onto another vector 44
projection, to a desired target space 85
PROJECTIONS 44
projections, fourth-order oblique 102
projections, generalized 53
projections, linear 53
projections, nonlinear 53
projections, orthogonal linear 44
projections, orthogonal, basis interpretation 47
projections, rank-1 oblique 49

projections, rank-1 orthogonal 46
projections, rank-2 oblique 48
projections, rank-2 orthogonal 47
projections, self-adjoint 54
projector 4
projectors, complementary 49, 85
projectors, complementary, generalized 87
PROJECTORS, linear, tensor representation 84
projectors, oblique, "normalized" 50
projectors, self-adjoint 86
proper orthogonal tensors 90
Q
Q-R decomposition 135
quadratic formula, generalized to vectors/tensors 133
quadric 132
quadric surface 134
quadruple dot product 99
R
range 4
range space 4
rank of a tensor 102
rank vs. order 102
reflection 90
reflection about the origin 90
reflection in a mirror 90
regularized basis 25
right hand rule 39
right-handed 41
rigor, philosophy 1
ROTATION (footnote reference) 94
rotation of 180 degrees 91
S
self-adjoint 4
sense of a vector 38
signed minor 15
skew-symmetric tensors 94
skew-symmetric tensors, relationship to axial vectors 42
spectral analysis 122
spectral analysis, diagonalizable tensors 125
sphere, equation of 130
Subspaces and the projection theorem 151
SUMMATION CONVENTION 26
summation convention 11
summation convention, suspending 31
swap (exchange) operator 151
symmetric tensors 94
T
tensor inner product (double-dot) 98
tensor inner product, fourth-order 99
tensor invariance 108
TENSOR OPERATIONS 69, 98
Tensor primitives (classification of tensors) 89
tensor primitives, the identity 77
tensor vs. matrix 108
tensor, "defined" as a matrix 57
tensor, basis 66
TENSOR, cat fights over the definition 57
tensor, decomposing into sym + skew parts 95
tensor, defined according to basis transformation rules 57
tensor, defined as a linear transformation from vectors to

vectors 57
tensor, definition as a linear transformation 66
tensor, definition as any sum of dyads 66
tensor, definition in terms of change of basis transformation
rules 106
tensor, definition in terms of component transformation
properties 66
tensor, DEFINITION in terms of dyads 66
tensor, dotting by a vector 69
tensor, dotting from the left by a vector 70
tensor, extracting components 71
tensor, single dotting with another tensor 71
tensor, testing for whether a tensor is zero 118
tensor, transpose 69
TENSORS 57
Tensors are vectors 149
tensors, deviatoric 97
tensors, isotropic 97
tensors, not the same as matrices 112
tensors, orthogonal (unitary) 89
tensors, symmetric and skew-symmetric 94
TENSORS, three kinds of notation (direct, basis, indicial) 73
tilde 8
trace of a matrix 10
transformation 3
transformations, linear 58
transpose 69
transpose of a matrix 9
Transverse isotropy 144
transversely isotropic materials 140
triad, laboratory 24
triple scalar product 41
triple scalar product between orthonormal RIGHT-HANDED
base vectors 41
U
under-tilde 8, 34
unitary tensors 89
V
vector operations 35
vector orientation and sense 38
Vector subspaces 150
vector, abstract definition 56
vector, expressing as a linear combination of three other
vectors 51
vectors, ordinary 24
volume, computed using the triple scalar product 41
Z
Zero, as a test for linear manifolds 6
zero, identifying if a tensor is zero 118
zero, verifying vanishing characteristic invariants is NOT
sufficient to conclude that a matrix itself is zero 17

Elementary Vector and Tensor Analysis - Brannon

Uploaded by

Copyright:

Available Formats

Elementary Vector and Tensor Analysis - Brannon

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Elementary Vector and Tensor Analysis - Brannon

Uploaded by

Copyright:

Available Formats

UNM REPORT

September 10, 2002 11:22 am

Individual copies may be made for personal use.

Contact author at rmbrann@sandia.gov

NOTE: When using Adobes acrobat reader to view this

Note to draft readers: The most useful textbooks are

ELEMENTARY VECTOR AND TENSOR ANALYSIS

presented in a framework generalizable to higherorder applications in material modeling

of New Mexico Adjunct professor

To be added. Stay tuned...

Finding the i-th component of a vector................................................................

Three kinds of vector and tensor notation............................................................

ASIDE #2: Right and left gradient operations...............................................

Finding components via projections. .................................................

September 10, 2002 11:51 am

September 10, 2002 11:51 am

September 10, 2002 2:37 pm

ELEMENTARY VECTOR AND TENSOR ANALYSIS

The discussion of tensor calculus will be

Things should be described as simply as possible,

September 10, 2002 2:37 pm

September 10, 2002 2:37 pm

Change isnt painful, but resistance to

September 10, 2002 2:37 pm

Some basic terminology from functional analysis is defined very loosely

September 10, 2002 2:37 pm

For each input x , a well-defined operator f must give a unique output

September 10, 2002 2:37 pm

September 10, 2002 2:37 pm

There are a thousand hacking at the branches

If attention must be called to the dimensions of a matrix, then they will be

Finish this list...

September 10, 2002 2:37 pm

depending on whether those components are collected into a column

tensors are often described in terms of an associated 3 3 matrix, which we

These comments will make more

The matrix product

Explicitly showing the dimensions,

where i takes values from 1 to M ,

The summation over k ranges from 1 to the common dimension, R .

must be an array of dimension N 1 with components given by

where i takes values from 1 to N

September 10, 2002 2:37 pm

The transpose of a matrix

reversed dimensions). The components of the transpose are

The transpose of [ A ] is written as [ A ] T , and the notation A ijT means the

The inner product of two column matrices

or, using the angled-bracket notation, <v> { w }

Applying the definition of matrix multiplication, the result is a 1 1 matrix

If { v } and { w } contain component of two vectors v and w then the inner

September 10, 2002 2:37 pm

The outer product of two column matrices.

or, using the angled-bracket notation, { a }<b>

The trace of a square matrix

The trace operation satisfies the following properties:

The Kronecker delta

The identity matrix

September 10, 2002 2:37 pm

The 3D permutation symbol