Advanced Calculus Notes

Lecture Notes for Advanced Calculus
James S. Cook
Liberty University
Department of Mathematics
Fall 2011
2
introduction and motivations for these notes
There are many excellent texts on portions of this subject. However, the particular path I choose
this semester is not quite in line with any particular text. I required the text on Advanced Cal-
culus by Edwards because it contains all the major theorems that traditionally are covered in an
Advanced Calculus course.
My focus diers signicantly. If I had students who had already completed a semester in real
analysis then we could delve into the more analytic aspects of the subject. However, real analy-
sis is not a prerequisite so we take a dierent path. Generically the story is as follows: a linear
approximation replaces a complicated object very well so long as we are close to the base-point
for the approximation. The rst level of understanding is what I would characterize as algebraic,
beyond that is the analytic understanding. I would argue that we must rst have a rm grasp of
the algebraic before we can properly attack the analytic aspects of the subject.
Edwards covers both the algebraic and the analytic. This makes his text hard to read in places
because the full story is at some points technical. My goal is to focus on the algebraic. That said,
I will try to at least point the reader to the section of Edward where the proof can be found.
Linear algebra is not a prerequisite for this course. However, I will use linear algebra. Matrices,
linear transformations and vector spaces are necessary ingredients for a proper discussion of ad-
vanced calculus. I believe an interested student can easily assimilate the needed tools as we go so I
am not terribly worried if you have not had linear algebra previously. I will make a point to include
some baby
1
linear exercises to make sure everyone who is working at this course keeps up with the
story that unfolds.
Doing the homework is doing the course. I cannot overemphasize the importance of thinking
through the homework. I would be happy if you left this course with a working knowledge of:
set-theoretic mapping langauge, bers and images and how to picture relationships diagra-
matically.
continuity in view of the metric topology in n-space.
the concept and application of the derivative and dierential of a mapping.
continuous dierentiability
inverse function theorem
implicit function theorem
tangent space and normal space via gradients
1
if you view this as an insult then you havent met the right babies yet. Baby exercises are cute.
3
extrema for multivariate functions, critical points and the Lagrange multiplier method
multivariate Taylor series.
quadratic forms
critical point analysis for multivariate functions
dual space and the dual basis.
multilinear algebra.
metric dualities and Hodge duality.
the work and ux form mappings for
3
.
basic manifold theory
vector elds as derivations.
Lie series and how vector elds generate symmetries
dierential forms and the exterior derivative
integration of forms
generalized Stokess Theorem.
surfaces
fundmental forms and curvature for surfaces
dierential form formulation of classical dierential geometry
some algebra and calculus of supermathematics
Before we begin, I should warn you that I assume quite a few things from the reader. These notes
are intended for someone who has already grappled with the problem of constructing proofs. I
assume you know the dierence between and . I assume the phrase i is known to you.
I assume you are ready and willing to do a proof by induction, strong or weak. I assume you
know what , , , and denote. I assume you know what a subset of a set is. I assume you
know how to prove two sets are equal. I assume you are familar with basic set operations such
as union and intersection (although we dont use those much). More importantly, I assume you
have started to appreciate that mathematics is more than just calculations. Calculations without
context, without theory, are doomed to failure. At a minimum theory and proper mathematics
allows you to communicate analytical concepts to other like-educated individuals.
Some of the most seemingly basic objects in mathematics are insidiously complex. Weve been
taught theyre simple since our childhood, but as adults, mathematical adults, we nd the actual
4
denitions of such objects as or are rather involved. I will not attempt to provide foundational
arguments to build numbers from basic set theory. I believe it is possible, I think its well-thought-
out mathematics, but we take the existence of the real numbers as an axiom for these notes. We
assume that exists and that the real numbers possess all their usual properties. In fact, I assume
, , , and all exist complete with their standard properties. In short, I assume we have
numbers to work with. We leave the rigorization of numbers to a dierent course.
The format of these notes is similar to that of my calculus and linear algebra and advanced calculus
notes from 2009-2011. However, I will make a number of denitions in the body of the text. Those
sort of denitions are typically background-type denitions and I will make a point of putting them
in bold so you can nd them with ease.
I have avoided use of Einsteins implicit summation notation in the majority of these notes. This
has introduced some clutter in calculations, but I hope the student nds the added detail helpful.
Naturally if one goes on to study tensor calculations in physics then no such luxury is granted, you
will have to grapple with the meaning of Einsteins convention. I suspect that is a minority in this
audience so I took that task o the to-do list for this course.
The content of this course diers somewhat from my previous oering. The presentation of ge-
ometry and manifolds is almost entirely altered. Also, I have removed the chapter on Newtonian
mechanics as well as the later chapter on variational calculus. Naturally, the interested student is
invited to study those as indendent studies past this course. If interested please ask.
I should mention that James Callahans Advanced Calculus: a geometric view has inuenced my
thinking in this reformulation of my notes. His discussion of Morses work was a useful addition to
the critical point analysis.
I was inspired by Flanders text on dierential form computation. It is my goal to implement some
of his nicer calculations as an addition to my previous treatment of dierential forms. In addition,
I intend to encorporate material from Burns and Gideas Dierential Geometry and Topology with
a View to Dynamical Systems as well as Munkrese Analysis on Manifolds. These additions should
greatly improve the depth of the manifold discussion. I intend to go signicantly deeper this year
so the student can perhaps begin to appreciate manifold theory.
I plan to take the last few weeks of class to discuss supermathematics. This will serve as a sideways
review for calculus on
. In addition, I hope the exercise of abstracting calculus to supernumbers

gives you some ideas about the process of abstraction in general. Abstraction is a cornerstone of
modern mathematics and it is an essential skill for a mathematician. We may also discuss some of
the historical motivations and modern applications of supermath to supersymmetric eld theory.
NOTE To BE DELETED:
-add pictures from 2009 notes.
5
-change equations to numbered equations.
6
Contents
1 set-up 11
1.1 set theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2 vectors and geometry for -dimensional space . . . . . . . . . . . . . . . . . . . . . . 13
1.2.1 vector algebra for three dimensions . . . . . . . . . . . . . . . . . . . . . . . . 19
1.2.2 compact notations for vector arithmetic . . . . . . . . . . . . . . . . . . . . . 20
1.3 functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.4 elementary topology and limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2 linear algebra 37
2.1 vector spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.2 matrix calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3 linear transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.3.1 a gallery of linear transformations . . . . . . . . . . . . . . . . . . . . . . . . 49
2.3.2 standard matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.3.3 coordinates and isomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.4 normed vector spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3 dierentiation 67
3.1 the dierential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.2 partial derivatives and the existence of the dierential . . . . . . . . . . . . . . . . . 73
3.2.1 directional derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.2.2 continuously dierentiable, a cautionary tale . . . . . . . . . . . . . . . . . . 78
3.2.3 gallery of derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.3 additivity and homogeneity of the derivative . . . . . . . . . . . . . . . . . . . . . . . 86
3.4 chain rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.5 product rules? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.5.1 scalar-vector product rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.5.2 calculus of paths in
3
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.5.3 calculus of matrix-valued functions of a real variable . . . . . . . . . . . . . . 93
3.5.4 calculus of complex-valued functions of a real variable . . . . . . . . . . . . . 95
3.6 complex analysis in a nutshell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
3.6.1 harmonic functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
7
8 CONTENTS
4 inverse and implicit function theorems 103
4.1 inverse function theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.2 implicit function theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.3 implicit dierentiation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5 geometry of level sets 119
5.1 denition of level set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
5.2 tangents and normals to a level set . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.3 method of Lagrange mulitpliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
6 critical point analysis for several variables 133
6.1 multivariate power series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.1.1 taylors polynomial for one-variable . . . . . . . . . . . . . . . . . . . . . . . . 133
6.1.2 taylors multinomial for two-variables . . . . . . . . . . . . . . . . . . . . . . 135
6.1.3 taylors multinomial for many-variables . . . . . . . . . . . . . . . . . . . . . 138
6.2 a brief introduction to the theory of quadratic forms . . . . . . . . . . . . . . . . . . 141
6.2.1 diagonalizing forms via eigenvectors . . . . . . . . . . . . . . . . . . . . . . . 144
6.3 second derivative test in many-variables . . . . . . . . . . . . . . . . . . . . . . . . . 151
6.3.1 morse theory and future reading . . . . . . . . . . . . . . . . . . . . . . . . . 154
7 multilinear algebra 155
7.1 dual space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
7.2 multilinearity and the tensor product . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.2.1 bilinear maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.2.2 trilinear maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
7.2.3 multilinear maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
7.3 wedge product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
7.3.1 wedge product of dual basis generates basis for . . . . . . . . . . . . . . . 167
7.3.2 the exterior algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
7.3.3 connecting vectors and forms in
3
. . . . . . . . . . . . . . . . . . . . . . . . 175
7.4 bilinear forms and geometry; metric duality . . . . . . . . . . . . . . . . . . . . . . . 177
7.4.1 metric geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
7.4.2 metric duality for tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
7.4.3 inner products and induced norm . . . . . . . . . . . . . . . . . . . . . . . . . 182
7.5 hodge duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
7.5.1 hodge duality in euclidean space
3
. . . . . . . . . . . . . . . . . . . . . . . 183
7.5.2 hodge duality in minkowski space
4
. . . . . . . . . . . . . . . . . . . . . . . 185
7.6 coordinate change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
7.6.1 coordinate change for
0
2
( ) . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
CONTENTS 9
8 manifold theory 193
8.1 manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
8.1.1 embedded manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
8.1.2 manifolds dened by charts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
8.1.3 dieomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
8.2 tangent space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
8.2.1 equivalence classes of curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
8.2.2 contravariant vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
8.2.3 derivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
8.2.4 dictionary between formalisms . . . . . . . . . . . . . . . . . . . . . . . . . . 212
8.3 the dierential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
8.4 cotangent space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
8.5 tensors at a point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
8.6 tensor elds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
8.7 metric tensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
8.7.1 classical metric notation in
. . . . . . . . . . . . . . . . . . . . . . . . . . 220
8.7.2 metric tensor on a smooth manifold . . . . . . . . . . . . . . . . . . . . . . . 221
8.8 on boundaries and submanifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
9 dierential forms 229
9.1 algebra of dierential forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
9.2 exterior derivatives: the calculus of forms . . . . . . . . . . . . . . . . . . . . . . . . 231
9.2.1 coordinate independence of exterior derivative . . . . . . . . . . . . . . . . . . 232
9.2.2 exterior derivatives on
3
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
9.3 pullbacks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
9.4 integration of dierential forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
9.4.1 integration of -form on
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
9.4.2 orientations and submanifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
9.5 Generalized Stokes Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
9.6 poincares lemma and converse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
9.6.1 exact forms are closed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
9.6.2 potentials for closed forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
9.7 classical dierential geometry in forms . . . . . . . . . . . . . . . . . . . . . . . . . . 257
9.8 E & M in dierential form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
9.8.1 dierential forms in Minkowski space . . . . . . . . . . . . . . . . . . . . . . . 258
9.8.2 exterior derivatives of charge forms, eld tensors, and their duals . . . . . . 263
9.8.3 coderivatives and comparing to Griths relativitic E & M . . . . . . . . . . 265
9.8.4 Maxwells equations are relativistically covariant . . . . . . . . . . . . . . . . 266
9.8.5 Electrostatics in Five dimensions . . . . . . . . . . . . . . . . . . . . . . . . . 268
10 supermath 273
10 CONTENTS
Chapter 1
set-up
In this chapter we settle some basic terminology about sets and functions.
1.1 set theory
Let us denote sets by capital letters in as much as is possible. Often the lower-case letter of the
same symbol will denote an element; is to mean that the object is in the set . We can
abbreviate
1
and
2
by simply writing
1
,
2
, this is a standard notation. The union
of two sets and is denoted = { }. The intersection of two sets is
denoted = { }. If a set has no elements then we say is the empty set
and denote this by writing = . It sometimes convenient to use unions or intersections of several
sets:
= { there exists with
= { for all we have
}
we say is the index set in the denitions above. If is a nite set then the union/intersection
is said to be a nite union/interection. If is a countable set then the union/intersection is said
to be a countable union/interection
1
. Suppose and are both sets then we say is a subset
of and write i implies for all . If then we also say is a
superset of . If then we say i = and = . Recall, for sets , we dene
= i implies for all and conversely implies for all . This
is equivalent to insisting = i and . The dierence of two sets and is
denoted and is dened by = { such that / }
2
.
1
recall the term countable simply means there exists a bijection to the natural numbers. The cardinality of such
a set is said to be
2
other texts somtimes use =
11
12 CHAPTER 1. SET-UP
We often make use of the following standard sets:
natural numbers (positive integers); = {1, 2, 3, . . . }.
natural numbers up to the number ;
= {1, 2, 3, . . . , 1, }.
integers; = {. . . , 2, 1, 0, 1, 2, . . . }. Note,
>0
= .
non-negative integers;
0
= {0, 1, 2, . . . } = {0}.
negative integers;
<0
= {1, 2, 3, . . . } = .
rational numbers; = {
, , = 0}.
irrational numbers; = { / }.
open interval from to ; (, ) = { < < }.
half-open interval; (, ] = { < } or [, ) = { < }.
closed interval; [, ] = { }.
The nal, and for us the most important, construction in set-theory is called the Cartesian product.
Let , , be sets, we dene:
= {(, ) and }
By a slight abuse of notation
3
we also dene:
= {(, , ) and and }
In the case the sets comprising the cartesian product are the same we use an exponential notation
for the construction:
2
= ,
3
=
We can extend to nitely many sets. Suppose
is a set for = 1, 2, . . . then we denote the

Cartesian product by
=1
and dene
=1
i = (
1
,
2
, . . . ,
) where
for each = 1, 2, . . . . An element

as above is often called an n-tuple.
We dene
2
= {(, ) , }. I refer to
2
as R-two in conversational mathematics. Like-
wise, R-three is dened by
3
= {(, , ) , , }. We are ultimately interested in studying
R-n where
= {(
1
,
2
, . . . ,
for = 1, 2, . . . , }. In this course if we consider
3
technically ( ) = ( ) since objects of the form (, (, )) are not the same as ((, ), ), we
ignore these distinctions and map both of these to the triple (, , ) without ambiguity in what follows
1.2. VECTORS AND GEOMETRY FOR -DIMENSIONAL SPACE 13
it is assumed from the context that .
In terms of cartesian products you can imagine the -axis as the number line then if we paste
another numberline at each value the union of all such lines constucts the plane; this is the
picture behind
2
= . Another interesting cartesian product is the unit-square; [0, 1]
2
=
[0, 1] [0, 1] = {(, ) 0 1, 0 1}. Sometimes a rectangle in the plane with its edges
included can be written as [
1
,
2
] [
1
,
2
]. If we want to remove the edges use (
1
,
2
) (
1
,
2
).
Moving to three dimensions we can construct the unit-cube as [0, 1]
3
. A generic rectangu-
lar solid can sometimes be represented as [
1
,
2
] [
1
,
2
] [
1
,
2
] or if we delete the edges:
(
1
,
2
) (
1
,
2
) (
1
,
2
).
1.2 vectors and geometry for -dimensional space
Denition 1.2.1.
Let , we dene
= {(
1
,
2
, . . . ,
= 1, 2, . . . , }. If
then we say is an n-vector. The numbers in the vector are called the components;
= (
1
,
2
, . . . ,
) has -th component
.
Notice, a consequence of the denition above and the construction of the Cartesian product
4
is that
two vectors and are equal i
for all
. Equality of two vectors is only true if all

components are found to match. It is therefore logical to dene addition and scalar multiplication
in terms of the components of vectors as follows:
Denition 1.2.2.
Dene functions + :
and :
by the following rules: for each

,
and :
(1.) ( +)
(2.) ()
for all {1, 2, . . . , }. The operation + is called vector addition and it takes two
vectors ,
and produces another vector +
. The operation is called scalar

multiplication and it takes a number and a vector
and produces another

vector
. Often we simply denote by juxtaposition .

If you are a gifted at visualization then perhaps you can add three-dimensional vectors in your
mind. If youre mind is really unhinged maybe you can even add 4 or 5 dimensional vectors. The
beauty of the denition above is that we have no need of pictures. Instead, algebra will do just
ne. That said, lets draw a few pictures.
4
see my Math 200 notes or ask me if interested, its not entirely trivial
Notice these pictures go to show how you can break-down vectors into component vectors which
point in the direction of the coordinate axis. Vectors of length one which point in the coordinate
directions make up what is called the standard basis
5
It is convenient to dene special notation
for the standard basis. First I dene a useful shorthand,
Denition 1.2.3.
The symbol
=
_
1 , =
0 , =
is called the Kronecker delta.
For example,
22
= 1 while
12
= 0.
Denition 1.2.4.
Let

1
be dened by (
. The size of the vector
is determined by context.
We call
the -th standard basis vector.

Example 1.2.5. Let me expand on what I mean by context in the denition above:
In we have
1
= (1) = 1 (by convention we drop the brackets in this case)
5
the term basis is carefully developed in the linear algebra course. In a nutshell we need two things: (1.) the
basis has to be big enough that we can add togther the basis elements to make any thing in the set (2.) the basis is
minimal so no single element in the basis can be formed by adding togther other basis elements
In
2
we have
1
= (1, 0) and
2
= (0, 1).
In
3
we have
1
= (1, 0, 0) and
2
= (0, 1, 0) and
3
= (0, 0, 1).
In
4
we have
1
= (1, 0, 0, 0) and
2
= (0, 1, 0, 0) and
3
= (0, 0, 1, 0) and
4
= (0, 0, 0, 1).
A real linear combination of {
1
,
2
, ,
} is simply a nite weighted-sum of the objects from

the set;
1
1
+
2
2
+
where
1
,
2
,
. If we take coecients
1
,
2
,
then
is is said to be a complex linear combination. I invite the reader to verify that every vector in
is a linear combination of
1
,
2
, . . . ,
6
. It is not dicult to prove the following properties for
vector addition and scalar multiplication: for all , ,
and , ,
(.) + = +, (.) ( +) + = + ( +)
(.) + 0 = , (.) = 0
(.) 1 = , (.) () = (),
(.) ( +) = +, (.) ( +) = +
(.) +
(.)
These properties of
are abstracted in linear algebra to form the denition of an abstract vector

space. Naturally
is a vector space, in fact it is the quintessial model for all other vector spaces.
Fortunately
also has a dot-product. The dot-product is a mapping from
to . We take
in a pair of vectors and output a real number.
Denition 1.2.6. Let ,
we dene by
=
1
1
+
2
2
+ +
.
Example 1.2.7. Let = (1, 2, 3, 4, 5) and = (6, 7, 8, 9, 10)
= 6 + 14 + 24 + 36 + 50 = 130
Example 1.2.8. Suppose we are given a vector
. We can select the -th component by

taking the dot-product of with
. Observe that
and consider,

=
_

=1
=1
=1
=
1
1
+ +
+ +
.
The dot-product with
has given us the length of the vector in the -th direction.

The length or norm of a vector and the angle between two vectors are induced from the dot-product:
Denition 1.2.9.
6
the calculation is given explicitly in my linear notes
The length or norm of
is a real number which is dened by =

.
Furthermore, let , be nonzero vectors in
we dene the angle between and by

cos
1
_

. together with these dentions of length and angle forms a Euclidean

Geometry.
Technically, before we make this denition we should make sure that the formulas given above even
make sense. I have not shown that is nonnegative and how do we know that argument of
the inverse cosine is within its domain of [1, 1]? I now state the propositions which justify the
preceding denition.(proofs of the propositions below are found in my linear algebra notes)
Proposition 1.2.10.
Suppose , ,
and then
1. =
2. ( +) = +
3. ( ) = () = ()
4. 0 and = 0 i = 0
The formula cos
1
_

is harder to justify. The inequality that we need for it to be reasonable

is
1, otherwise we would not have a number in the (

1
) = () = [1, 1].
An equivalent inequality is which is known as the Cauchy-Schwarz inequality.
Proposition 1.2.11.
If ,
then
Example 1.2.12. Let = [1, 2, 3, 4, 5]
and = [6, 7, 8, 9, 10]
nd the angle between these vectors

and calculate the unit vectors in the same directions as and . Recall that, = 6 +14 +24 +
36 + 50 = 130. Furthermore,
=
1
2
+ 2
2
+ 3
2
+ 4
2
+ 5
2
=
1 + 4 + 9 + 16 + 25 =
55
=
6
2
+ 7
2
+ 8
2
+ 9
2
+ 10
2
=
36 + 49 + 64 + 81 + 100 =
330
We nd unit vectors via the standard trick, you just take the given vector and multiply it by the
reciprocal of its length. This is called normalizing the vector,
=
1
55
[1, 2, 3, 4, 5]
=
1
330
[6, 7, 8, 9, 10]
The angle is calculated from the denition of angle,

= cos
1
_
130
55
330
_
= 15.21
Its good we have this denition, 5-dimensional protractors are very expensive.
Proposition 1.2.13.
Let ,
and suppose then

1. =
2. + + (triangle inequality)
3. 0
4. = 0 i = 0
The four properties above make
paired with :
a normed linear space.

Well see how dierentiation can be dened given this structure. It turns out that we can dene a
reasonable concept of dierentiation for other normed linear spaces. In this course well study how
to dierentiate functions to and from
, matrix-valued functions and complex-valued functions of

a real variable. Finally, if time permits, well study dierentiation of functions of functions which
is the central task of variational calculus. In each case the underlying linear structure along
with the norm is used to dene the limits which are necessary to set-up the derivatives. The focus
of this course is the process and use of derivatives and integrals so I have not given proofs of the
linear algebraic propositions in this chapter. The proofs and a deeper view of the meaning of these
propositions is given at length in Math 321. If you havent had linear then youll just have to trust
me on these propositions
7
7
or you could just read the linear notes if curious
Denition 1.2.14.
The distance between
and
is dened to be (, ) .
If we draw a picture this denition is very natural. Here we are thinking of the points , as vectors
from the origin then is the vector which points from to (this is algebraically clear since
+( ) = ). Then the distance between the points is the length of the vector that points from
one point to the other. If you plug in two dimensional vectors you should recognize the distance
formula from middle school math:
((
1
,
2
), (
1
,
2
)) =
(
1
1
)
2
+ (
2
2
)
2
Proposition 1.2.15.
Let :
be the distance function then

1. (, ) = (, )
2. (, ) 0
3. (, ) = 0 i = 0
4. (, ) +(, ) (, )
In real analysis one studies a set paired with a distance function. Abstractly speaking such a pair
is called a metric space. A vector space with a norm is called a normed linear space. Because
we can always induce a distance function from the norm via the formula (, ) = every
normed linear space is a metric space. The converse fails. Metric spaces need not be vector spaces,
a metric space could just be formed from some subset of a vector space or something more exotic
8
.
The absolute value function on denes distance function (, ) = . In your real analysis
8
there are many texts to read on metric spaces, one nice treatment is Rosenlichts Introduction to Analysis, its a
good read
course you will study the structure of the metric space (, : ) in great depth. I
include these comments here to draw your attention to the connection between this course and the
real analysis course. I primarily use the norm in what follows, but it should be noted that many
things could be written in terms of the distance function.
1.2.1 vector algebra for three dimensions
Every nonzero vector can be written as a unit vector scalar multiplied by its magnitude.

such that = 0 = where =

1
.
You should recall that we can write any vector in
3
as
=< , , >= < 1, 0, 0 > + < 0, 1, 0 > + < 0, 0, 1 >=
where we dened the

=< 1, 0, 0 >,

=< 0, 1, 0 >,

=< 0, 0, 1 >. You can easily verify that
distinct Cartesian unit-vectors are orthogonal. Sometimes we need to produce a vector which is
orthogonal to a given pair of vectors, it turns out the cross-product is one of two ways to do that
in
3
. We will see much later that this is special to three dimensions.
Denition 1.2.16.
If =<
1
,
2
,
3
> and =<
1
,
2
,
3
> are vectors in
3
then the cross-product
of and is a vector which is dened by:

=<
2
2
,
3
3
,
1
1
> .
The magnitude of

can be shown to satisfy
sin() and the direction can

be deduced by right-hand-rule. The right hand rule for the unit vectors yields:
=

,

=

,

If I wish to discuss both the point and the vector to which it corresponds we may use the notation
= (
1
,
2
, . . . ,
)

=<
1
,
2
, . . . ,
>
With this notation we can easily dene directed line-segments as the vector which points from one
point to another, also the distance bewtween points is simply the length of the vector which points
from one point to the other:
Denition 1.2.17.
Let ,
. The directed line segment from to is

=

. This vector is
drawn from tail to the tip where we denote the direction by drawing an arrowhead.
The distance between and is (, ) =
.
1.2.2 compact notations for vector arithmetic
I prefer the following notations over the hat-notation of the preceding section because this notation
generalizes nicely to -dimensions.
1
=< 1, 0, 0 >
2
=< 0, 1, 0 >
3
=< 0, 0, 1 > .
Likewise the Kronecker delta and the Levi-Civita symbol are at times very convenient for abstract
calculation:
=
_
1 =
0 =
1 (, , ) {(1, 2, 3), (3, 1, 2), (2, 3, 1)}

1 (, , ) {(3, 2, 1), (2, 1, 3), (1, 3, 2)}
0 if any index repeats
An equivalent denition for the Levi-civita symbol is simply that
123
= 1 and it is antisymmetric
with respect to the interchange of any pair of indices;
.
Now let us restate some earlier results in terms of the Einstein repeated index conventions
9
, let
and then
standard basis expansion
orthonormal basis
(
+

)
vector addition
(
vector subtraction
(
scalar multiplication
dot product
(
cross product.
All but the last of the above are readily generalized to dimensions other than three by simply
increasing the number of components. However, the cross product is special to three dimensions.
I cant emphasize enough that the formulas given above for the dot and cross products can be
utilized to yield great eciency in abstract calculations.
Example 1.2.18. . .
9
there are more details to be seen in the Appendix if youre curious
1.3. FUNCTIONS 21
1.3 functions
Suppose and are sets, we say : is a function if for each the function
assigns a single element () . Moreover, if : is a function we say it is a -valued
function of an -variable and we say = () whereas = (). For example,
if :
2
[0, 1] then is real-valued function of
2
. On the other hand, if :
2
then
wed say is a vector-valued function of a complex variable. The term mapping will be used
interchangeably with function in these notes
10
. Suppose : and and then
we may consisely express the same data via the notation : .
Sometimes we can take two given functions and construct a new function.
1. if : and : then
: is the composite of with .

2. if , : and is a set with an operation of addition then we dene :
pointwise by the natural assignment ( )() = () () for each . We say that
is the sum(+) or dierence() of and .
3. if : and where there is an operation of scalar multiplication by on then
: is dened pointwise by ()() = () for each . We say that is scalar
multiple of by .
Usually we have in mind = or = and often the addition is just that of vectors, however
the denitions (2.) and (3.) apply equally well to matrix-valued functions or operators which is
another term for function-valued functions. For example, in the rst semester of calculus we study
/ which is a function of functions; / takes an input of and gives the output /. If we
write = 3/ we have a new operator dened by (3/)[] = 3/ for each function in
the domain of /.
Denition 1.3.1.
10
in my rst set of advanced calculus notes (2010) I used the term function to mean the codomain was real numbers
whereas mapping implied a codomain of vectors. I was following Edwards as he makes this convention in his text. I
am not adopting that terminology any longer, I think its better to use the term function as we did in Math 200 or
250. A function is an abstract construction which allows for a vast array of codomains.
Suppose : . We dene the image of
1
under as follows:
(
1
) = { there exists
1
with () = }.
The range of is (). The inverse image of
1
under is dened as follows:
1
(
1
) = { ()
1
}.
The inverse image of a single point in the codomain is called a ber. Suppose : .
We say is surjective or onto
1
i there exists
1
such that (
1
) =
1
. If a function
is onto its codomain then the function is surjective. If (
1
) = (
2
) implies
1
=
2
for all
1
,
2

1
then we say f is injective on
1
or 1 1 on
1
. If a function
is injective on its domain then we say the function is injective. If a function is both
injective and surjective then the function is called a bijection or a 1-1 correspondance.
Example 1.3.2. Suppose :
2
and (, ) = for each (, )
2
. The function is not
injective since (1, 2) = 1 and (1, 3) = 1 and yet (1, 2) = (1, 3). Notice that the bers of are
simply vertical lines:
1
(
) = {(, ) () (, ) =
} = {(
, ) } = {
}
Example 1.3.3. Suppose : and () =
2
+ 1 for each . This function is not
surjective because 0 / (). In contrast, if we construct : [1, ) with () = () for each
then can argue that is surjective. Neither nor is injective, the ber of
is {
}
for each
= 0. At all points except zero these maps are said to be two-to-one. This is an
abbreviation of the observation that two points in the domain map to the same point in the range.
Denition 1.3.4.
Suppose :
and suppose further that for each ,

() = (
1
(),
2
(), . . . ,
()).
Then we say that = (
1
,
2
, . . . ,
) and for each
the functions
are
called the component functions of . Furthermore, we dene the projection
to be the map
() =
for each = 1, 2, . . . . This allows us to express each of the

component functions as a composition

.
3
2
and (, , ) = (
2
+
2
, ) for each (, , )
3
. Identify
that
1
(, , ) =
2
+
2
whereas
2
(, , ) = . You can easily see that () = [0, ] .
Suppose
2
[0, ) and
then
1
({(
2
,
)}) =
1
() {
}
where
1
() denotes a circle of radius . This result is a simple consequence of the observation
that (, , ) = (
2
,
) implies
2
+
2
=
2
and =
.
1.3. FUNCTIONS 23
Example 1.3.6. Let , , be particular constants. Suppose :
3
and (, , ) =
+ + for each (, , )
3
. Here there is just one component function so we could say that
=
1
but we dont usually bother to make such an observation. If at least one of the constants
, , is nonzero then the bers of this map are planes in three dimensional space with normal
, ,
1
({}) = {(, , )
3
+ + = }
If = = = 0 then the ber of is simply all of
3
and the () = {0}.
The denition below explains how to put together functions with a common domain. The codomain
of the new function is the cartesian product of the old codomains.
Denition 1.3.7.
Let :
1

and :
1

be a mappings then (, ) is a
mapping from
1
to
1
2
dened by (, )() = ((), ()) for all
1
.
Theres more than meets the eye in the denition above. Let me expand it a bit here:
(, )() = (
1
(),
2
(), . . . ,
(),
1
(),
2
(), . . . ,
()) where = (
1
,
2
, . . . ,
)
You might notice that Edwards uses for the identity mapping whereas I use . His notation is
quite reasonable given that the identity is the cartesian product of all the projection maps:
= (
1
,
2
, . . . ,
)
Ive had courses where we simply used the coordinate notation itself for projections, in that nota-
tion have formulas such as (, , ) = ,
() =
and
) =
.
Another way to modify a given function is to adjust the domain of a given mapping by restriction
and extension.
Denition 1.3.8.
Let :
be a mapping. If then we dene the restriction of

to to be the mapping
: where
() = () for all . If and

then we say a mapping : is an extension of i
()
= .
When I say
()
= this means that these functions have matching domains and they agree at
each point in that domain;
()
() = () for all (). Once a particular subset is chosen
the restriction to that subset is a unique function. Of course there are usually many susbets of
() so you can imagine many dierent restictions of a given function. The concept of extension
is more vague, once you pick the enlarged domain and codomain it is not even necessarily the case
that another extension to that same pair of sets will be the same mapping. To obtain uniqueness
for extensions one needs to add more stucture. This is one reason that complex variables are
interesting, there are cases where the structure of the complex theory forces the extension of a
complex-valued function of a complex variable to be unique. This is very surprising. Similarly a
linear transformation is uniquely dened by its values on a basis, it extends uniquely from that
nite set of vectors to the innite number of points in the vector space. This is very restrictive on
the possible ways we can construct linear mappings. Maybe you can nd some other examples of
extensions as you collect your own mathematical storybook.
Denition 1.3.9.
Let :
be a mapping, if there exists a mapping : () such that

=
()
and =
then is the inverse mapping of and we denote =

1
.
If a mapping is injective then it can be shown that the inverse mapping is well dened. We dene
1
() = i () = and the value must be a single value if the function is one-one. When a
function is not one-one then there may be more than one point which maps to a particular point
in the range.
Notice that the inverse image of a set is well-dened even if there is no inverse mapping. Moreover,
it can be shown that the bers of a mapping are disjoint and their union covers the domain of the
mapping:
() = ()
1
{}
1
{} =
()
1
{} = ().
This means that the bers of a mapping partition the domain.
Example 1.3.10. . .
Denition 1.3.11.
Let :
be a mapping. Furthermore, suppose that : is a

mapping which is constant on each ber of . In other words, for each ber
1
{}
we have some constant such that (
1
{}) = . The subset
1
() is called a
cross section of the ber partition of .
How do we construct a cross section for a particular mapping? For particular examples the details
of the formula for the mapping usually suggests some obvious choice. However, in general if you
accept the axiom of choice then you can be comforted in the existence of a cross section even in
the case that there are innitely many bers for the mapping.
1.4. ELEMENTARY TOPOLOGY AND LIMITS 25
Example 1.3.12. . .
Proposition 1.3.13.
Let :
be a mapping. The restriction of to a cross section

of is an injective function. The mapping

: () is a surjection. The mapping
: () is a bijection.
The proposition above tells us that we can take any mapping and cut down the domain and/or
codomain to reduce the function to an injection, surjection or bijection. If you look for it youll see
this result behind the scenes in other courses. For example, in linear algebra if we throw out the
kernel of a linear mapping then we get an injection. The idea of a local inverse is also important
to the study of calculus.
Example 1.3.14. . .
Denition 1.3.15.
Let :
be a mapping then we say a mapping is a local inverse of

i there exits such that = (
)
1
.
Usually we can nd local inverses for functions in calculus. For example, () = sin() is not 1-1
therefore it is not invertible. However, it does have a local inverse () = sin
1
(). If we were
more pedantic we wouldnt write sin
1
(). Instead we would write () =
_
sin
[
2
,
2
]
_
1
() since
the inverse sine is actually just a local inverse. To construct a local inverse for some mapping we
must locate some subset of the domain upon which the mapping is injective. Then relative to that
subset we can reverse the mapping. The inverse mapping theorem (which well study mid-course)
will tell us more about the existence of local inverses for a given mapping.
1.4 elementary topology and limits
In this section we describe the metric topology for
. In the study of functions of one real variable

we often need to refer to open or closed intervals. The denition that follows generalizes those
concepts to -dimensions. I have included a short discussion of general topology in the Appendix
if youd like to learn more about the term.
Denition 1.4.1.
An open ball of radius centered at
is the subset all points in
which are less

than units from , we denote this open ball by
() = {
< }
The closed ball of radius centered at
is likewise dened
() = {
}
Notice that in the = 1 case we observe an open ball is an open interval: let ,
() = { < } = { < } = ( , +)
In the = 2 case we observe that an open ball is an open disk: let (, )
2
,
((, )) =
_
(, )
2
(, ) (, ) <
_
=
_
(, )
2
( )
2
+ ( )
2
<
_
For = 3 an open-ball is a sphere without the outer shell. In contrast, a closed ball in = 3 is a
solid sphere which includes the outer shell of the sphere.
Denition 1.4.2.
Let
. We say is an interior point of i there exists some open ball

centered at which is completely contained in . We say
is a limit point of i
every open ball centered at contains points in {}. We say
is a boundary
point of i every open ball centered at contains points not in and other points which
are in {}. We say is an isolated point of if there exist open balls about
which do not contain other points in . The set of all interior points of is called the
interior of . Likewise the set of all boundary points for is denoted . The closure
of is dened to be = {
a limit point}
If youre like me the paragraph above doesnt help much until I see the picture below. All the terms
are aptly named. The term limit point is given because those points are the ones for which it is
natural to dene a limit.
Example 1.4.3. . .
Denition 1.4.4.
Let
is an open set i for each there exists > 0 such that
() and
() . Let
is an closed set i its complement
= {
/ }
is an open set.
Notice that [, ] = (, ) (, ). It is not hard to prove that open intervals are open hence
we nd that a closed interval is a closed set. Likewise it is not hard to prove that open balls are
open sets and closed balls are closed sets. I may ask you to prove the following proposition in the
homework.
Proposition 1.4.5.
A closed set contains all its limit points, that is
is closed i = .
Example 1.4.6. . .
In calculus I the limit of a function is dened in terms of deleted open intervals centered about the
limit point. We can dene the limit of a mapping in terms of deleted open balls centered at the
limit point.
Denition 1.4.7.
Let :
be a mapping. We say that has limit
at limit point
of i for each > 0 there exists a > 0 such that
with 0 < < implies

() < . In such a case we can denote the above by stating that
lim
() = .
In calculus I the limit of a function is dened in terms of deleted open intervals centered about the
limit point. We just dened the limit of a mapping in terms of deleted open balls centered at the
limit point. The term deleted refers to the fact that we assume 0 < which means we
do not consider = in the limiting process. In other words, the limit of a mapping considers
values close to the limit point but not necessarily the limit point itself. The case that the function
is dened at the limit point is special, when the limit and the mapping agree then we say the
mapping is continuous at that point.
Example 1.4.8. . .
Denition 1.4.9.
Let :
be a mapping. If is a limit point of then we say that

is continuous at i
lim
() = ()
If is an isolated point then we also say that is continous at . The mapping is
continous on i it is continous at each point in . The mapping is continuous i
it is continuous on its domain.
Notice that in the = = 1 case we recover the denition of continuous functions from calc. I.
Proposition 1.4.10.
Let :
be a mapping with component functions

1
,
2
, . . . ,
hence
= (
1
,
2
, . . . ,
). If is a limit point of then

lim
() = lim
() =
for each = 1, 2, . . . , .
.
Proof: () Suppose lim
() = . Then for each > 0 choose > 0 such that 0 < <
implies () < . This choice of suces for our purposes as:
()
()
)
2
=1
(
()
)
2
= () < .
Hence we have shown that lim
() =
for all = 1, 2, . . . .
() Suppose lim
() =
for all = 1, 2, . . . . Let > 0. Note that / > 0 and therefore

by the given limits we can choose
> 0 such that 0 < < implies
()
<
/.
Choose = {
1
,
2
, . . . ,
} clearly > 0. Moreoever, notice 0 < <
hence
requiring 0 < < automatically induces 0 < <
for all . Suppose that
and 0 < < it follows that

() =
=1
(
()
=1
()
2
<
=1
(
/)
2
<
=1
/ = .
Therefore, lim
() = and the proposition follows.

We can analyze the limit of a mapping by analyzing the limits of the component functions:
Example 1.4.11. . .
The following follows immediately from the preceding proposition.
Proposition 1.4.12.
Suppose that :
is a mapping with component functions

1
,
2
, . . . ,
.
Let be a limit point of then is continous at i
is continuous at for
= 1, 2, . . . , . Moreover, is continuous on i all the component functions of are
continuous on . Finally, a mapping is continous i all of its component functions are
continuous. .
The proof of the proposition is in Edwards, its his Theorem 7.2. Its about time I proved something.
Proposition 1.4.13.
The projection functions are continuous. The identity mapping is continuous.
Proof: Let > 0 and choose = . If
such that 0 < < then it follows that

< .. Therefore, lim
= which means that lim
() = () for all
.
Hence is continuous on
which means is continuous. Since the projection functions are

component functions of the identity mapping it follows that the projection functions are also con-
tinuous (using the previous proposition).
Denition 1.4.14.
The sum and product are functions from
2
to dened by
(, ) = + (, ) =
Proposition 1.4.15.
The sum and product functions are continuous.
Preparing for the proof: Let the limit point be (, ). Consider what we wish to show: given a
point (, ) such that 0 < (, ) (, ) < we wish to show that
(, ) ( +) < or for the product (, ) () <
follow for appropriate choices of . Think about the sum for a moment,
(, ) ( +) = + +
I just used the triangle inequality for the absolute value of real numbers. We see that if we could
somehow get control of and then wed be getting closer to the prize. We have control
of 0 < (, ) (, ) < notice this reduces to
( , ) <
( )
2
+ ( )
2
<
it is clear that ( )
2
<
2
since if it was otherwise the inequality above would be violated as
adding a nonegative quantity ( )
2
only increases the radicand resulting in the squareroot to be
larger than . Hence we may assume ()
2
<
2
and since > 0 it follows < . Likewise,
< . Thus
(, ) ( +) = + < + < 2
We see for the sum proof we can choose = /2 and it will work out nicely.
Proof: Let > 0 and let (, )
2
. Choose = /2 and suppose (, )
2
such that
(, ) (, ) < . Observe that
(, ) (, ) < ( , )
2
<
2

2
+
2
<
2
.
It follows < and < . Thus
(, ) ( +) = + + < + = 2 = .
Therefore, lim
(,)(,)
(, ) = +. and it follows that the sum function if continuous at (, ).
But, (, ) is an arbitrary point thus is continuous on
2
hence the sum function is continuous. .
Preparing for the proof of continuity of the product function: Ill continue to use the same
notation as above. We need to study (, ) () = < . Consider that
= + = ( ) +( ) +
We know that < and < . There is one less obvious factor to bound in the expression.
What should we do about ?. I leave it to the reader to show that:
< < +
Now put it all together and hopefully well be able to solve for .
= + < ( +) + =
2
+( +) =
I put solve in quotes because we have considerably more freedom in our quest for nding . We
could just as well nd which makes the = become an <. That said lets pursue equality,
2
+( +) = 0 =

( +)
2
+ 4
2
Since , , > 0 it follows that
( +)
2
+ 4 <
( +)
2
= + hence the (+) solution
to the quadratic equation yields a positive namely:
=
+
( +)
2
+ 4
2
Yowsers, I almost made this a homework. There may be an easier route. You might notice we have
run across a few little lemmas (Ive boxed the punch lines for the lemmas) which are doubtless
useful in other proofs. We should collect those once were nished with this proof.
Proof: Let > 0 and let (, )
2
. By the calculations that prepared for the proof we know that
the following quantity is positive, hence choose
=
+
( +)
2
+ 4
2
> 0.
Note that
11
,
= + = ( ) +( ) algebra
+ triangle inequality
< ( +) + by the boxed lemmas
=
2
+( +) algebra
=
where we know that last step follows due to the steps leading to the boxed equation in the proof
preparation. Therefore, lim
(,)(,)
(, ) = . and it follows that the product function if con-
tinuous at (, ). But, (, ) is an arbitrary point thus is continuous on
2
hence the product
function is continuous. .
Lemma 1.4.16.
Assume > 0.
1. If , then < < +.
2. If ,
then <
< for = 1, 2, . . . .
The proof of the proposition above is mostly contained in the remarks of the preceding two pages.
Example 1.4.17. . .
11
my notation is that when we stack inequalities the inequality in a particular line refers only to the immediate
vertical successor.
Proposition 1.4.18.
Let :
and :
be mappings. Suppose that

lim
() = and suppose that is continuous at then

lim
(

)() =
_
lim
()
_
.
The proof is in Edwards, see pages 46-47. Notice that the proposition above immediately gives us
the important result below:
Proposition 1.4.19.
Let and be mappings such that

is well-dened. The composite function

is
continuous for points (

) such that the following two conditions hold:
1. is continuous at
2. is continuous at ().
I make use of the earlier proposition that a mapping is continuous i its component functions are
continuous throughout the examples that follow. For example, I know (, ) is continuous since
was previously proved continuous.
Example 1.4.20. Note that if =
(, ) then () =
_
(, )
_
() =
_
(, )()
_
=
(, ) =
2
. Therefore, the quadratic function () =
2
is continuous on as it is the composite
of continuous functions.
Example 1.4.21. Note that if =
(, ), ) then () = (
2
, ) =
3
. Therefore, the
cubic function () =
3
is continuous on as it is the composite of continuous functions.
Example 1.4.22. The power function is inductively dened by
1
= and
=
1
for all
. We can prove () =
is continous by induction on . We proved the = 1 case

previously. Assume inductively that () =
1
is continuous. Notice that
=
1
= () = (, ()) = (
(, ))().
Therefore, using the induction hypothesis, we see that () =
is the composite of continuous

functions thus it is continuous. We conclude that () =
is continuous for all .

We can play similar games with the sum function to prove that sums of power functions are
continuous. In your homework you will prove constant functions are continuous. Putting all of
these things together gives us the well-known result that polynomials are continuous on .
Proposition 1.4.23.
Let be a limit point of mappings , :
and suppose . If
lim
() =
1
and lim
() =
2
then
1. lim
(() +()) = lim
() + lim
().
2. lim
(()()) =
_
lim
()
__
lim
()
_
.
3. lim
(()) = lim
().
Moreover, if , are continuous then +, and are continuous.
Proof: Edwards proves (1.) carefully on pg. 48. Ill do (2.) here: we are given that If lim
() =
1
and lim
() =
2
thus by Proposition 1.4.10 we nd lim
(, )() = (
1
,
2
).
Consider then,
lim
(()()) = lim
_
(, )
_
defn. of product function
=
_
lim
(, )
_
since is continuous
= (
1
,
2
) by Proposition 1.4.10.
=
1
2
denition of product function
=
_
lim
()
__
lim
()
_
.
In your homework you proved that lim
= thus item (3.) follows from (2.). .

The proposition that follows does follow immediately from the proposition above, however I give a
proof that again illustrates the idea we used in the examples. Reinterpreting a given function as a
composite of more basic functions is a useful theoretical and calculational technique.
Proposition 1.4.24.
Assume , :
are continuous functions at and suppose .

1. + is continuous at .
2. is continuous at
3. is continuous at .
Moreover, if , are continuous then +, and are continuous.
Proof: Observe that ( + )() = (
(, ))() and ()() = (
(, ))(). Were given that

, are continuous at and we know , are continuous on all of
2
thus the composite functions
(, ) and
(, ) are continuous at and the proof of items (1.) and (2.) is complete. To
prove (3.) I refer the reader to their homework where it was shown that () = for all is a
continuous function. We then nd (3.) follows from (2.) by setting = (function multiplication
commutes for real-valued functions). .
We can use induction arguments to extend these results to arbitrarily many products and sums of
power functions.To prove continuity of algebraic functions wed need to do some more work with
quotient and root functions. Ill stop here for the moment, perhaps Ill ask you to prove a few more
fundamentals from calculus I. I havent delved into the denition of exponential or log functions
not to mention sine or cosine. We will assume that the basic functions of calculus are continuous
on the interior of their respective domains. Basically if the formula for a function can be evaluated
at the limit point then the function is continuous.
Its not hard to see that the comments above extend to functions of several variables and map-
pings. If the formula for a mapping is comprised of nite sums and products of power func-
tions then we can prove such a mapping is continuous using the techniques developed in this
section. If we have a mapping with a more complicated formula built from elementary func-
tions then that mapping will be continuous provided its component functions have formulas which
are sensibly calculated at the limit point. In other words, if you are willing to believe me that
sin(), cos(),
, ln(), cosh(), sinh(),
,
1
, . . . are continuous on the interior of their domains

then its not hard to prove:
(, , ) =
_
sin() +
cosh(
2
) +
, cosh(),
+
1
)
_
is a continuous mapping at points where the radicands of the square root functions are nonnegative.
It wouldnt be very fun to write explicitly but it is clear that this mapping is the Cartesian product
of functions which are the sum, product and composite of continuous functions.
Denition 1.4.25.
A polynomial in -variables has the form:
(
1
,
2
, . . . ,
) =
1
,
2
,...,
=0
1
,
2
,...,
1
1

2
2

where only nitely many coecients
1
,
2
,...,
= 0. We denote the set of multinomials in
-variables as (
1
,
2
, . . . ,
).
Polynomials are (). Polynomials in two variables are (, ), for example,
(, ) = + () = 1, linear function
(, ) = + + () = 1, ane function
(, ) =
2
+ +
2
deg(f)=2, quadratic form
(, ) =
2
+ +
2
+ + + deg(f)=2
If all the terms in the polynomial have the same number of variables then it is said to be homo-
geneous. In the list above only the linear function and the quadratic form were homogeneous.
Remark 1.4.26.
There are other topologies possible for
. For example, one can prove that
1
=
1
+
2
+ +
gives a norm on
and the theorems we proved transfer over almost without change by

just trading for
1
. The unit ball becomes a diamond for the 1-norm. There are
many other norms which can be constructed, innitely many it turns out. However, it has
been shown that the topology of all these dierent norms is equivalent. This means that
open sets generated from dierent norms will be the same class of sets. For example, if
you can t an open disk around every point in a set then its clear you can just as well t
an open diamond and vice-versa. One of the things that makes innite dimensional linear
algebra more fun is the fact that the topology generated by distinct norms need not be
equivalent for innite dimensions. There is a dierence between the open sets generated by
the Euclidean norm verses those generated by the 1-norm. Incidentally, my thesis work is
mostly built over the 1-norm. It makes the supernumbers happy.
Chapter 2
linear algebra
Our goal in the rst section of this chapter is to gain conceptual clarity on the meaning of the
central terms from linear algebra. This is a birds-eye view of linear, my selection of topics here is
centered around the goal of helping you to see the linear algebra in calculus. Once you see it then
you can use the theory of linear algebra to help organize your thinking. Our ultimate goal is that
organizational principle. Our goal here is not to learn all of linear algebra, rather we wish to use it
as a tool for the right jobs as they arise this semester.
In the second section we summarize the tools of matrix computation. We will use matrix addition,
multiplication and throughout this course. Inverse matrices and the noncommuative nature of ma-
trix multiplication are illustrated. It is assumed that the reader has some previous experience in
matrix computation, at least in highschool you should have spent some time.
In the third section the concept of a linear transformation is formalized. The formula for any
linear transformation from
to
can be expressed as a matrix multiplication. We study this

standard matrix in enough depth as to understand its application in for dierentiation. A number
of examples to visualize the role of a linear transformation are oered for breadth. Finally, isomor-
phisms and coordinate maps are discussed.
In the fourth section we dene norms for vector spaces. We study how the norm allows us to dene
limits for an abstract vector space. This is important since it allows us to quantify continuity for
abstract linear transformations as well as ultimately to dene dierentiation on a normed vector
space in the chapter that follows.
37
38 CHAPTER 2. LINEAR ALGEBRA
2.1 vector spaces
Suppose we have a set paired with an addition and scalar multiplication such that for all , ,
and , , :
(.) + = +, (.) ( +) + = + ( +)
(.) + 0 = , (.) = 0
(.) 1 = , (.) () = (),
(.) ( +) = +, (.) ( +) = +
(.) +
(.)
then we say that is a vector space over . To be a bit more precise, by (iii.) I mean to say
that there exist some element 0 such that + 0 = for each . Also, (iv.) should be
understood to say that for each there exists another element such that +() = 0.
Example 2.1.1.
is a vector space with respect to the standard vector addition and scalar
multiplication.
Example 2.1.2. = {+ , } is a vector space where the usual complex number addition
provides the vector addition and multiplication by a real number (+) = +() clearly denes
a scalar multiplication.
Example 2.1.3. The set of all matrices is a vector space with respect to the usual matrix
addition and scalar multiplication. We will elaborate on the details in an upcoming section.
Example 2.1.4. Suppose is the set of all functions from a set to a vector space then is
naturally a vector space with respect to the function addition and multiplication by a scalar. Both
of those operations are well-dened on the values of the function since we assumed the codomain of
each function in is the vector space .
There are many subspaces of function space which provide interesting examples of vector spaces.
For example, the set of continuous functions:
Example 2.1.5. Let
0
(
) be a set of continuous functions from
to
then
0
(
)
is a vector space with respect to function addition and the usual multiplication. This fact relies on
the sum and scalar multiple of continuous functions is once more continuous.
Denition 2.1.6.
We say a subset of a vector space is linearly independent (LI) i for scalars
1
,
2
, . . . ,
1
+
2
2
+
= 0
1
=
2
= = 0
for each nite subset {
1
,
2
, . . . ,
} of .
In the case that is nite it suces to show the implication for a linear combination of all the
vectors in the set. Notice that if any vector in the set can be written as a linear combination of
the other vectors in then that makes fail the test for linear independence. Moreover, if a set
is not linearly independent then we say is linearly dependent.
2.1. VECTOR SPACES 39
Example 2.1.7. The standard basis of
is denoted {
1
,
2
, . . . ,
}. We can show linear inde-

pendence easily via the dot-product: suppose that
1
1
+
2
2
+
= 0 and take the dot-product

of both sides with
to obtain
(
1
1
+
2
2
+
= 0
(
1
+
2
= 0
(1) = 0
but, was arbitrary hence it follows that
1
=
2
= =
= 0 which establishes the linear

independence of the standard basis.
Example 2.1.8. Consider = {1, } . We can argue is LI as follows: suppose
1
(1)+
2
() =
0. Thus
1
+
2
= 0 for some real numbers
1
,
2
. Recall that a basic property of complex numbers is
that if
1
=
2
then (
1
) = (
2
) and (
1
) = (
2
) where
= (
)+(
). Therefore,
the complex equation
1
+
2
= 0 yields two real equations
1
= 0 and
2
= 0.
Example 2.1.9. Let
0
() be the vector space of all continuous functions from to . Suppose
is the set of monic
1
monomials = {1, ,
2
,
3
, . . . }. This is an innite set. We can argue LI
as follows: suppose
1
1
+
2
2
+ +
= 0. For convenience relable the powers

1
,
2
, . . . ,
by
1
,
2
, . . . ,
such that 1 <
1
<
2
< <
. This notation just shues the terms in the

nite sum around so that the rst term has the lowest order: consider
1
+
2
+ +
= 0
If
1
= 0 then evaluate at = 0 to obtain
1
= 0. If
1
> 0 then dierentiate
1
times and
denote this new equation by
1
. Evaluate
1
at = 0 to nd
1
(
1
1) 3 2 1
1
= 0
hence
1
= 0. Since we set-up
1
<
2
it follows that after
1
-dierentiations the second summand
is still nontrivial in
1
. However, we can continue dierentiating until we reach
2
and
then constant term is
2
!
2
so evaluation will show
2
= 0. We continue in this fashion until we
have shown that
= 0 for = 1, 2, . . . . It follows that is a linearly independent set.

We spend considerable eort in linear algebra to understand LI from as many angles as possible.
One equivalent formulation of LI is the ability to equate coecients. In other words, a set of objects
is LI i whenever we have an equation with thos objects we can equate coecients. In calculus
when we equate coecients we implicitly assume that the functions in question are LI. Generally
speaking two functions are LI if their graphs have distinct shapes which cannot be related by a
simple vertical stretch.
Example 2.1.10. Consider = {2
, 3(1/2)
} as a subset the vector space

0
(). To show linear
dependence we observe that
1
2
+
2
3(1/2)
= 0 yields (
1
+3
2
)2
= 0. Hence
1
+3
2
= 0 which
means nontrivial solutions exist. Take
2
= 1 then
1
= 3. Of course the heart of the matter is
that 3(1/2)
= 3(2
) so the second function is just a scalar multiple of the rst function.

1
monic means that the leading coecient is 1.
If youve taken dierential equations then you should recognize the concept of LI from your study
of solution sets to dierential equations. Given an -th order linear dierential equation we always
have a goal of calculating -LI solutions. In that context LI is important because it helps us
make sure we do not count the same solution twice. The general solution is formed from a linear
combination of the LI solution set. Of course this is not a course in dierential equations, I include
this comment to make connections to the other course. One last example on LI should suce to
make certain you at least have a good idea of the concept:
Example 2.1.11. Consider
3
as a vector space and consider the set = {,
i,
j,
k} where we
could also denote

i =
1
,
j =
2
,
k =
3
but Im aiming to make your mind connect with your
calculus III background. This set is clearly linearly dependent since we can write any vector as
a linear combination of the standard unit-vectors: moreover, we can use dot-products to select the
, and components as follows:
= (
i)
i + (
j)
j + (
k)
k
Linear independence helps us quantify a type of redundancy for vectors in a given set. The next
denition is equally important and it is sort of the other side of the coin; spanning is a criteria
which helps us insure a set of vectors will cover a vector space without missing anything.
Denition 2.1.12.
We say a subset of a vector space is a spanning set for i for each there
exist scalars
1
,
2
, . . . ,
and vectors
1
,
2
, . . . ,
such that =
1
1
+
2
2
+
.
We denote {
1
,
2
, . . . ,
} = {
1
1
+
2
2
+

1
,
2
, . . . ,
}.
If and is a vector space then it is immediately obvious that () . If is a
spanning set then it is obvious that (). It follows that when is a spanning set for
we have () = .
Example 2.1.13. It is easy to show that if
then =
1
1
+
2
2
+ +
. It follows
that
= {
=1
.
Example 2.1.14. Let 1, where
2
= 1. Clearly = {1, }.
Example 2.1.15. Let be the set of polynomials. Since the sum of any two polynomials and
the scalar multiple of any polynomial is once more a polynomial we nd is a vector space with
respect to function addition and multiplication of a function by a scalar. We can argue that the set
of monic monomials {1, ,
2
, . . . } a spanning set for . Why? Because if () then that means
there are scalars
0
,
1
, . . . ,
such that () =
0
+
1
+
2
2
+ +
Denition 2.1.16.
We say a subset of a vector space is a basis for i is a linearly independent
spanning set for . If is a nite set then is said to be nite dimensional and the
number of vectors in is called the dimension of . That is, if = {
1
,
2
, . . . ,
} is a
basis for then ( ) = . If no nite basis exists for then is said to be innite
dimensional.
2.2. MATRIX CALCULATION 41
The careful reader will question why this concept of dimension is well-dened. Why can we not
have bases of diering dimension for a given vector space? I leave this question for linear algebra,
the theorem which asserts the uniqueness of dimension is one of the deeper theorems in the course.
However, like most everything in linear, at some level it just boils down to solving some particular
set of equations. You might tell Dr. Sprano its just algebra. In any event, it is common practice
to use the term dimension in courses where linear algebra is not understood. For example,
2
is a
two-dimensional space. Or well say that
3
is a three-dimensional space. This terminology agrees
with the general observation of the next example.
Example 2.1.17. The standard basis {
=1
for
is a basis for
and (
) = . This
result holds for all . The line is one-dimensional, the plane is two-dimensional, three-space
is three-dimensional etc...
Example 2.1.18. The set {1, } is a basis for . It follows that () = 2. We say that the
complex numbers form a two-dimensional real vector space.
Example 2.1.19. The set of polynomials is clearly innite dimensional. Contradiction shows this
without much eort. Suppose had a nite basis . Choose the polynomial of largest degree (say
) in . Notice that () =
+1
is a polynomial and yet clearly () / () hence is not a
spanning set. But this contradicts the assumption is a basis. Hence, by contradiction, no nite
basis exists and we conclude the set of polynomials is innite dimensional.
There is a more general use of the term dimension which is beyond the context of linear algebra.
For example, in calculus II or III you may have heard that a circle is one-dimensional or a surface
is two-dimensional. Well, circles and surfaces are not usually vector spaces so the terminology is
not taken from linear algebra. In fact, that use of the term dimension stems from manifold theory.
I hope to discuss manifolds later in this course.
2.2 matrix calculation
An matrix is an array of numbers with -rows and -columns. We dene

to be the
set of all matrices. The set of all -dimensional column vectors is
1
=
2
. The set of
all -dimensional row vectors is
1
. A given matrix

has -components
. Notice
that the components are numbers;
for all , such that 1 and 1 . We

should not write =
because it is nonesense, however = [
] is quite ne.
Suppose

, note for 1 we have
()
1
whereas for 1 we nd
()
1
. In other words, an matrix has columns of length and rows of length .
2
We will use the convention that points in
are column vectors. However, we will use the somewhat subtle

notation (1, 2, . . . ) = [1, 2, . . . ]
. This helps me write
rather than
1
and I dont have to pepper
transposes all over the place. If youve read my linear algebra notes youll appreciate the wisdom of our convention.
Two matrices and are equal i
for all , . Given matrices , with components
and constant we dene

(+)
()
, for all , .
The zero matrix in

is denoted 0 and dened by 0
= 0 for all , . The additive inverse

of

is the matrix such that + () = 0. The components of the additive inverse
matrix are given by ()
for all , . Likewise, if

and

then the
product

is dened by
3
:
()
=1
for each 1 and 1 . In the case = = 1 the indices , are omitted in the equation
since the matrix product is simply a number which needs no index. The identity matrix in

is the square matrix whose components are the Kronecker delta;
=
_
1 =
0 =
.
The notation
is sometimes used. For example,

2
=
_
1 0
0 1
_
. If the size of the identity matrix
needs emphasis otherwise the size of the matrix is to be understood from the context.
Let

. If there exists

such that = and = then we say that
is invertible and
1
= . Invertible matrices are also called nonsingular. If a matrix has no
inverse then it is called a noninvertible or singular matrix.
Let

then

is called the transpose of and is dened by (
for all 1 and 1 . It is sometimes useful to know that ()
and
(
)
1
= (
1
)
. It is also true that ()

1
=
1
1
. Furthermore, note dot-product of
,
is given by =
.
The -th standard basis matrix for

is denoted
for 1 and 1 . The

matrix
is zero in all entries except for the (, )-th slot where it has a 1. In other words, we
dene (
. I invite the reader to show that the term basis is justied in this context
4
.
Given this basis we see that the vector space

has (

) = .
Theorem 2.2.1.
3
this product is dened so the matrix of the composite of a linear transformation is the product of the matrices
of the composed transformations. This is illustrated later in this section and is proved in my linear algebra notes.
4
the theorem stated below contains the needed results and then some, you can nd the proof is given in my linear
algebra notes. It would be wise to just work it out in the 2 2 case as a warm-up if you are interested
If

and then
(.) =
=1
(.) =
=1
=1
(.) [
] =
() (.) [
] =
()
(.)
(.)
(.)
.
You can look in my linear algebra notes for the details of the theorem. Ill just expand one point
here: Let

then
=
11

12

1
21

22

2
.
.
.
.
.
.
.
.
.
1

2

=
11
1 0 0
0 0 0
.
.
.
.
.
. 0
0 0 0
+
12
0 1 0
0 0 0
.
.
.
.
.
. 0
0 0 0
+ +
0 0 0
0 0 0
.
.
.
.
.
. 0
0 0 1
=
11
11
+
12
12
+ +
.
The calculation above follows from repeated -applications of the denition of matrix addition
and another -applications of the denition of scalar multiplication of a matrix.
Example 2.2.2. Suppose = [
1 2 3
4 5 6
]. We see that has 2 rows and 3 columns thus
23
.
Moreover,
11
= 1,
12
= 2,
13
= 3,
21
= 4,
22
= 5, and
23
= 6. Its not usually possible to
nd a formula for a generic element in the matrix, but this matrix satises
= 3( 1) + for
all ,
5
. The columns of are,
1
() =
_
1
4
_
,
2
() =
_
2
5
_
,
3
() =
_
3
6
_
.
The rows of are
1
() =
_
1 2 3
,
2
() =
_
4 5 6

1 2 3
4 5 6
] then
=
_
1 4
2 5
3 6
_
. Notice that
1
() =
1
(
),
2
() =
2
(
)
5
In the statement for all , it is to be understood that those indices range over their allowed values. In the
preceding example 1 2 and 1 3.
and
1
() =
1
(
),
2
() =
2
(
),
3
() =
3
(
)
Notice (
= 3( 1) + for all , ; at the level of index calculations we just switch the

indices to create the transpose.
Example 2.2.4. Let = [
1 2
3 4
] and = [
5 6
7 8
]. We calculate
+ =
_
1 2
3 4
_
+
_
5 6
7 8
_
=
_
6 8
10 12
_
.
1 2
3 4
] and = [
5 6
7 8
]. We calculate
=
_
1 2
3 4
_
_
5 6
7 8
_
=
_
4 4
4 4
_
.
Now multiply by the scalar 5,
5 = 5
_
1 2
3 4
_
=
_
5 10
15 20
_
Example 2.2.6. Let ,

be dened by
= 3 + 5 and
=
2
for all , . Then we
can calculate (+)
= 3 + 5 +
2
for all , .
Example 2.2.7. Solve the following matrix equation,
0 =
_

_
+
_
1 2
3 4
_

_
0 0
0 0
_
=
_
1 2
3 4
_
The denition of matrix equality means this single matrix equation reduces to 4 scalar equations:
0 = 1, 0 = 2, 0 = 3, 0 = 4. The solution is = 1, = 2, = 3, = 4.
The denition of matrix multiplication (()
=1
) is very nice for general proofs, but

pragmatically I usually think of matrix multiplication in terms of dot-products. It turns out we can
view the matrix product as a collection of dot-products: suppose

and

then
=
1
()
1
()
1
()
2
()
1
()
()
2
()
1
()
2
()
2
()
2
()
()
.
.
.
.
.
.
.
.
.
()
1
()
()
2
()
()
()
Let me explain how this works. The formula above claims ()
()
() for all , .
Recall that (
())
and (
())
thus
()
=1
=1
(
())
())
Hence, using denition of the dot-product, ()
()
(). This argument holds for

all , therefore the dot-product formula for matrix multiplication is valid.
Example 2.2.8. The product of a 3 2 and 2 3 is a 3 3
1 0
0 1
0 0
_
4 5 6
7 8 9
_
=
[1, 0][4, 7]
[1, 0][5, 8]
[1, 0][6, 9]
[0, 1][4, 7]
[0, 1][5, 8]
[0, 1][6, 9]
[0, 0][4, 7]
[0, 0][5, 8]
[0, 0][6, 9]
4 5 6
7 8 9
0 0 0
Example 2.2.9. The product of a 3 1 and 1 3 is a 3 3
1
2
3
_
4 5 6
4 1 5 1 6 1
4 2 5 2 6 2
4 3 5 3 6 3
4 5 6
8 10 12
12 15 18

1 2
3 4
] and = [
5 6
7 8
]. We calculate
=
_
1 2
3 4
_ _
5 6
7 8
_
=
_
[1, 2][5, 7]
[1, 2][6, 8]
[3, 4][5, 7]
[3, 4][6, 8]
_
=
_
5 + 14 6 + 16
15 + 28 18 + 32
_
=
_
19 22
43 50
_
Notice the product of square matrices is square. For numbers , it we know the product of
and is commutative ( = ). Lets calculate the product of and in the opposite order,
=
_
5 6
7 8
_ _
1 2
3 4
_
=
_
[5, 6][1, 3]
[5, 6][2, 4]
[7, 8][1, 3]
[7, 8][2, 4]
_
=
_
5 + 18 10 + 24
7 + 24 14 + 32
_
=
_
23 34
31 46
_
Clearly = thus matrix multiplication is noncommutative or nonabelian.
When we say that matrix multiplication is noncommuative that indicates that the product of two
matrices does not generally commute. However, there are special matrices which commute with
other matrices.
1 0
0 1
] and =
_

. We calculate
=
_
1 0
0 1
_ _

_
=
_

_
Likewise calculate,
=
_

_ _
1 0
0 1
_
=
_

_
Since the matrix was arbitrary we conclude that = for all
22
.
Example 2.2.12. Consider , , from Example ??.
+ =
_
5
7
_
+
_
6
8
_
=
_
11
15
_
Using the above we calculate,
( +) =
_
1 2
3 4
_ _
11
15
_
=
_
11 + 30
33 + 60
_
=
_
41
93
_
.
In constrast, we can add and ,
+ =
_
19
43
_
+
_
22
50
_
=
_
41
93
_
.
Behold, ( +) = + for this example. It turns out this is true in general.
I collect all my favorite properties for matrix multiplication in the theorem below. To summarize,
matrix math works as you would expect with the exception that matrix multiplication is not
commutative. We must be careful about the order of letters in matrix expressions.
Theorem 2.2.13.
If , ,

, ,

,

and
1
,
2
then
1. (+) + = + ( +),
2. () = (),
3. + = +,
4.
1
(+) =
1
+
2
,
5. (
1
+
2
) =
1
+
2
,
6. (
1
2
) =
1
(
2
),
7. (
1
) =
1
() = (
1
) = ()
1
,
8. 1 = ,
9.
= =
,
10. ( + ) = + ,
11. (
1
+
2
) =
1
+
2
,
12. (+) = +,
Proof: I will prove a couple of these primarily to give you a chance to test your understanding
of the notation. Nearly all of these properties are proved by breaking the statement down to
components then appealing to a property of real numbers. Just a reminder, we assume that it is
known that is an ordered eld. Multiplication of real numbers is commutative, associative and
distributes across addition of real numbers. Likewise, addition of real numbers is commutative,
associative and obeys familar distributive laws when combined with addition.
Proof of (1.): assume , , are given as in the statement of the Theorem. Observe that
((+) +)
= (+)
defn. of matrix add.

= (
) +

=
+ (
) assoc. of real numbers

=
+ ( +)

= (+ ( +))

for all , . Therefore (+) + = + ( +).
Proof of (6.): assume
1
,
2
, are given as in the statement of the Theorem. Observe that
((
1
2
))
= (
1
2
)
defn. scalar multiplication.

=
1
(
2
) assoc. of real numbers

= (
1
(
2
))

for all , . Therefore (
1
2
) =
1
(
2
).
Proof of (10.): assume , , are given as in the statement of the Theorem. Observe that
((( + ))
( + )
defn. matrix multiplication,

=
) defn. matrix addition,

=
) dist. of real numbers,

=
) prop. of nite sum,

= ()
+ ( )
defn. matrix multiplication( 2),

= ( + )
defn. matrix addition,

for all , . Therefore ( + ) = + .
The proofs of the other items are similar, we consider the , -th component of the identity and then
apply the denition of the appropriate matrix operations denition. This reduces the problem to
a statement about real numbers so we can use the properties of real numbers at the level of
components. Then we reverse the steps. Since the calculation works for arbitrary , it follows the
the matrix equation holds true.
2.3 linear transformations
Denition 2.3.1.
Let and be vector spaces over . If a mapping : satises
1. ( +) = () +() for all , ; this is called additivity.
2. () = () for all and ; this is called homogeneity.
then we say is a linear transformation. If = then we may say that is a linear
transformation on .
Proposition 2.3.2.
If

and :
is dened by () = for each
then is a linear
transformation.
Proof: Let

and dene :
by () = for each
. Let ,
and
,
( +) = ( +) = + = () +()
and
() = () = = ()
thus is a linear transformation.
Obviously this gives us a nice way to construct examples. The following proposition is really at the
heart of all the geometry in this section.
2.3. LINEAR TRANSFORMATIONS 49
Proposition 2.3.3.
Let = { + [0, 1], ,
with = 0} dene a line segment from to + in
. If :
is a linear transformation then () is a either a line-segment from

() to ( +) or a point.
Proof: suppose and are as in the proposition. Let () then by denition there exists
such that () = . But this implies there exists [0, 1] such that = + so
( +) = . Notice that
= ( +) = () +() = () +().
which implies {() + () [0, 1]} =
2
. Therefore, ()
2
. Conversely, suppose

2
then = () +() for some [0, 1] but this yields by linearity of that = (+)
hence (). Since we have that ()
2
and
2
() it follows that () =
2
. Note
that
2
is a line-segment provided that () = 0, however if () = 0 then
2
= {()} and the
proposition follows.
2.3.1 a gallery of linear transformations
My choice of mapping the unit square has no particular signcance in the examples below. I
merely wanted to keep it simple and draw your eye to the distinction between the examples.
In each example well map the four corners of the square to see where the transformation takes
the unit-square. Those corners are simply (0, 0), (1, 0), (1, 1), (0, 1) as we traverse the square in a
counter-clockwise direction.
Example 2.3.4. Let =
_
0
0
_
for some > 0. Dene () = for all
2
. In particular
this means,
(, ) = (, ) =
_
0
0
_ _

_
=
_

_
.
We nd (0, 0) = (0, 0), (1, 0) = (, 0), (1, 1) = (, ), (0, 1) = (0, ). This mapping is called
a dilation.
_
1 0
0 1
_
. Dene () = for all
2
. In particular this means,
(, ) = (, ) =
_
1 0
0 1
_ _

_
=
_

_
.
We nd (0, 0) = (0, 0), (1, 0) = (1, 0), (1, 1) = (1, 1), (0, 1) = (0, 1). This mapping is
called an inversion.
_
1 2
3 4
_
. Dene () = for all
2
(, ) = (, ) =
_
1 2
3 4
_ _

_
=
_
+ 2
3 + 4
_
.
We nd (0, 0) = (0, 0), (1, 0) = (1, 3), (1, 1) = (3, 7), (0, 1) = (2, 4). This mapping shall
remain nameless, it is doubtless a combination of the other named mappings.
1
2
_
1 1
1 1
_
. Dene () = for all
2
. In particular this
means,
(, ) = (, ) =
1
2
_
1 1
1 1
_ _

_
=
1
2
_

+
_
.
We nd (0, 0) = (0, 0), (1, 0) =
1
2
(1, 1), (1, 1) =
1
2
(0, 2), (0, 1) =
1
2
(1, 1). This mapping
is a rotation by /4 radians.
_
1 1
1 1
_
. Dene () = for all
2
(, ) = (, ) =
_
1 1
1 1
_ _

_
=
_

+
_
.
We nd (0, 0) = (0, 0), (1, 0) = (1, 1), (1, 1) = (0, 2), (0, 1) = (1, 1). This mapping is a
rotation followed by a dilation by =
2.
_
cos() sin()
sin() cos()
_
. Dene () = for all
2
. In particular
this means,
(, ) = (, ) =
_
cos() sin()
sin() cos()
_ _

_
=
_
cos() sin()
sin() + cos()
_
.
We nd (0, 0) = (0, 0), (1, 0) = (cos(), sin()), (1, 1) = (cos()sin(), cos()+sin()) (0, 1) =
(sin(), cos()). This mapping is a rotation by in the counter-clockwise direction. Of course you
could have derived the matrix from the picture below.
_
1 0
0 1
_
. Dene () = for all
2
(, ) = (, ) =
_
1 0
0 1
_ _

_
=
_

_
.
We nd (0, 0) = (0, 0), (1, 0) = (1, 0), (1, 1) = (1, 1), (0, 1) = (0, 1). This mapping is a
rotation by zero radians, or you could say it is a dilation by a factor of 1, ... usually we call this
the identity mapping because the image is identical to the preimage.
Example 2.3.11. Let
1
=
_
1 0
0 0
_
. Dene
1
() =
1
for all
2
means,
1
(, ) =
1
(, ) =
_
1 0
0 0
_ _

_
=
_

0
_
.
We nd
1
(0, 0) = (0, 0),
1
(1, 0) = (1, 0),
1
(1, 1) = (1, 0),
1
(0, 1) = (0, 0). This mapping is a
projection onto the rst coordinate.
Let
2
=
_
0 0
0 1
_
. Dene () =
2
for all
2
2
(, ) =
2
(, ) =
_
0 0
0 1
_ _

_
=
_
0
_
.
We nd
2
(0, 0) = (0, 0),
2
(1, 0) = (0, 0),
2
(1, 1) = (0, 1),
2
(0, 1) = (0, 1). This mapping is
projection onto the second coordinate.
We can picture both of these mappings at once:
_
1 1
1 1
_
. Dene () = for all
2
(, ) = (, ) =
_
1 1
1 1
_ _

_
=
_
+
+
_
.
We nd (0, 0) = (0, 0), (1, 0) = (1, 1), (1, 1) = (2, 2), (0, 1) = (1, 1). This mapping is not a
projection, but it does collapse the square to a line-segment.
Remark 2.3.13.
The examples here have focused on linear transformations from
2
to
2
. It turns out that
higher dimensional mappings can largely be understood in terms of the geometric operations
weve seen in this section.
0 0
1 0
0 1
. Dene () = for all

2
(, ) = (, ) =
0 0
1 0
0 1
_
=
.
We nd (0, 0) = (0, 0, 0), (1, 0) = (0, 1, 0), (1, 1) = (0, 1, 1), (0, 1) = (0, 0, 1). This mapping
moves the -plane to the -plane. In particular, the horizontal unit square gets mapped to vertical
unit square; ([0, 1] [0, 1]) = {0} [0, 1] [0, 1]. This mapping certainly is not surjective because
no point with = 0 is covered in the range.
_
1 1 0
1 1 1
_
. Dene () = for all
3
means,
(, , ) = (, , ) =
_
1 1 0
1 1 1
_
=
_
+
+ +
_
.
Lets study how maps the unit cube. We have 2
3
= 8 corners on the unit cube,
(0, 0, 0) = (0, 0), (1, 0, 0) = (1, 1), (1, 1, 0) = (2, 2), (0, 1, 0) = (1, 1)
(0, 0, 1) = (0, 1), (1, 0, 1) = (1, 2), (1, 1, 1) = (2, 3), (0, 1, 1) = (1, 2).
This mapping squished the unit cube to a shape in the plane which contains the points (0, 0), (0, 1),
(1, 1), (1, 2), (2, 2), (2, 3). Face by face analysis of the mapping reveals the image is a parallelogram.
This mapping is certainly not injective since two dierent points get mapped to the same point. In
particular, I have color-coded the mapping of top and base faces as they map to line segments. The
vertical faces map to one of the two parallelograms that comprise the image.
I have used terms like vertical or horizontal in the standard manner we associate such terms
with three dimensional geometry. Visualization and terminology for higher-dimensional examples is
not as obvious. However, with a little imagination we can still draw pictures to capture important
aspects of mappings.
_
1 0 0 0
1 0 0 0
_
. Dene () = for all
4
means,
(, , , ) = (, , , ) =
_
1 0 0 0
1 0 0 0
_
=
_

_
.
Lets study how maps the unit hypercube [0, 1]
4

4
. We have 2
4
= 16 corners on the unit
hypercube, note (1, , , ) = (1, 1) whereas (0, , , ) = (0, 0) for all , , [0, 1]. Therefore,
the unit hypercube is squished to a line-segment from (0, 0) to (1, 1). This mapping is neither
surjective nor injective. In the picture below the vertical axis represents the , , -directions.
Example 2.3.17. Suppose (, ) = (
,
2
+ ) note that (1, 1) = (1, 2) and (4, 4) = (2, 20).
Note that (4, 4) = 4(1, 1) thus we should see (4, 4) = (4(1, 1)) = 4(1, 1) but that fails to be true
so is not a linear transformation.
Example 2.3.18. Let (, ) =
2
+
2
dene a mapping from
2
to . This is not a linear
transformation since
((, )) = (, ) = ()
2
+ ()
2
=
2
(
2
+
2
) =
2
(, ).
We say is a nonlinear transformation.
Example 2.3.19. Suppose : is dened by () = + for some constants , .
Is this a linear transformation on ? Observe:
(0) = (0) + =
thus is not a linear transformation if = 0. On the other hand, if = 0 then is a linear
transformation.
A mapping on
which has the form () = + is called a translation. If we have a mapping of

the form () = + for some

and then we say is an ane tranformation
on
. Technically, in general, the line = + is the graph of an ane function on . I invite

the reader to prove that ane transformations also map line-segments to line-segments (or points).
2.3.2 standard matrices
Denition 2.3.20.
Let :
be a linear transformation, the matrix

such that () =
for all
is called the standard matrix of . We denote this by [] = or more

compactly, [
] = , we say that
is the linear transformation induced by . Moreover,

the components of the matrix are found from
= ((
)))
.
Example 2.3.21. Given that ([, , ]
) = [+2, 3+4, 5+6]
for [, , ]

3
nd the the
standard matrix of . We wish to nd a 33 matrix such that () = for all = [, , ]

3
.
Write () then collect terms with each coordinate in the domain,
+ 2
3 + 4
5 + 6
1
0
5
2
3
0
0
4
6
Its not hard to see that,
1 2 0
0 3 4
5 0 6
= [] =
1 2 0
0 3 4
5 0 6
Notice that the columns in are just as youd expect from the proof of theorem ??. [] =
[(
1
)(
2
)(
3
)]. In future examples I will exploit this observation to save writing.
Example 2.3.22. Suppose that ((, , , )) = ( + + +, , 0, 3 ), nd [].
(
1
) = ((1, 0, 0, 0)) = (1, 0, 0, 3)
(
2
) = ((0, 1, 0, 0)) = (1, 1, 0, 0)
(
3
) = ((0, 0, 1, 0)) = (1, 0, 0, 0)
(
4
) = ((0, 0, 0, 1)) = (1, 1, 0, 1)
[] =
1 1 1 1
0 1 0 1
0 0 0 0
3 0 0 1
.
I invite the reader to check my answer here and see that () = [] for all
4
as claimed.
Proposition 2.3.23.
Suppose :
and :
are linear transformations then + and are

linear transformations and
(1.) [ +] = [] + [], (2.) [ ] = [] [], (3.) [] = [].
In words, the standard matrix of the sum, dierence or scalar multiple of linear transfor-
mations is the sum, dierence or scalar multiple of the standard matrices of the respsective
linear transformations.
Example 2.3.24. Suppose (, ) = ( +, ) and (, ) = (2, 3). Its easy to see that
[] =
_
1 1
1 1
_
and [] =
_
2 0
0 3
_
[ +] = [] + [] =
_
3 1
1 2
_
Therefore, ( + )(, ) =
_
3 1
1 2
_ _

_
=
_
3 +
+ 2
_
= (3 + , + 2). Naturally this is the
same formula that we would obtain through direct addition of the formulas of and .
Proposition 2.3.25.
1
:
and
2
:
are linear transformations then

2

1
:
is a
linear transformation with matrix [
2

1
] such that
[
2

1
]
=1
[
2
]
[
1
]
for all = 1, 2, . . . and = 1, 2 . . . , .

Example 2.3.26. Let :
21
21
be dened by
([, ]
) = [ +, 2 ]
for all [, ]

21
. Also let :
21
31
be dened by
([, ]
) = [, , 3 + 4]
for all [, ]

21
. We calculate the composite as follows:
(
)([, ]
) = (([, ]
))
= ([ +, 2 ]
)
= [ +, +, 3( +) + 4(2 )]
= [ +, +, 11 ]
Notice we can write the formula above as a matrix multiplication,

(
)([, ]
) =
1 1
1 1
11 1
_
[
] =
1 1
1 1
11 1
.
Notice that the standard matrices of and are:
[] =
1 0
1 0
3 4
[] =
_
1 1
2 1
_
Its easy to see that [
] = [][] (as we should expect since these are linear operators)

Notice that

is not even dened since the dimensions of the codomain of do not match
the domain of . Likewise, the matrix product [][] is not dened since there is a dimension
mismatch; (2 2)(3 2) is not a well-dened product of matrices.
2.3.3 coordinates and isomorphism
Let be a nite dimensional vector space with basis = {
1
,
2
, . . .
}. The coordinate map
is dened by
(
1
1
+
2
2
+ +
) =
1
1
+
2
2
+ +
for all =
1
1
+
2
2
+ +
. Sometimes we have to adjust the numbering a bit for

double-indices. For example:

be dened by
(
) = (
11
, . . . ,
1
,
21
, . . . ,
2
, . . . ,
1
, . . . ,
)
This map simply takes the entries in the matrix and strings them out to a vector of length .
2
be dened by (+) = (, ). This is the coordinate map for
the basis {1, }.
Matrix multiplication is for vectors in
. Direct matrix multiplication of an abstract vector makes

no sense (how would you multiply a polynomial and a matrix?), however, since we can use the
coordinate map to change the abstract vector to a vector in
. The diagram below illustrates the

idea for a linear transformation from an abstract vector space with basis to another abstract
vector space with basis

:

//

OO
[]
,
//

Lets walk through the formula []

,
((
1
())): we begin on the RHS with a column

vector , then
1
lifts the column vector up to the abstract vector

1
() in . Next we operate
by which moves us over to the vector (
1
()) which is in . Finally the coordinate map
pushes the abstract vector in back to a column vector
((
1
())) which is in
1
. The
same journey is accomplished by just multiplying by the matrix []
,
.
Example 2.3.29. Let = {1, ,
2
} be the basis for
2
and consider the derivative mapping
:
2

2
. Find the matrix of assuming that
2
has coordinates with respect to on both
copies of
2
. Dene and observe
(
) =
+1
whereas
1
(
) =
1
for = 0, 1, 2. Recall (
2
+ +) = 2 +.
1
([]
,
) =
((
1
(
1
))) =
((1)) =
(0) = 0
2
([]
,
) =
((
1
(
2
))) =
(()) =
(1) =
1
3
([]
,
) =
((
1
(
3
))) =
((
2
)) =
(2) = 2
2
Therefore we nd,
[]
,
=
0 1 0
0 0 2
0 0 0
.
Calculate
3
. Is this surprising?
A one-one correspondence is a map which is 1-1 and onto. If we can nd such a mapping between
two sets then it shows those sets have the same cardnality. Cardnality is a crude idea of size, it
turns out that all nite dimensional vector spaces over have the same cardnality. On the other
hand, not all vector spaces have the same dimension. Isomorphisms help us discern if two vector
spaces have the same dimension.
Denition 2.3.30.
Let , be vector spaces then : is an isomorphism if it is a 1-1 and onto
mapping which is also a linear transformation. If there is an isomorphism between vector
spaces and then we say those vector spaces are isomorphic and we denote this by
.
Other authors sometimes denote isomorphism by equality. But, Ill avoid that custom as I am
reserving = to denote set equality. Details of the rst two examples below can be found in my
linear algebra notes.
3
and =
2
. Dene a mapping :
2

3
by
(
2
+ +) = (, , )
for all
2
+ +
2
. As vector spaces,
3
and polynomials of upto quadratic order are the
same.
Example 2.3.32. Let
2
be the set of 2 2 symmetric matrices. Let :
3
2
be dened by
(, , ) =
_

_
.
Example 2.3.33. Let (
) denote the set of all linear transformations from
to
.
(
) forms a vector space under function addition and scalar multiplication. There is a
natural isomorphism to matrices. Dene : (
)

by () = [] for all
linear transformations (
). In other words, linear transformations and matrices are

the same as vector spaces.
2.4. NORMED VECTOR SPACES 61
The quantication of same is a large theme in modern mathematics. In fact, the term iso-
morphism as we use it here is more accurately phrased vector space isomorphism. The are other
kinds of isomorphisms which preserve other interesting stuctures like Group, Ring or Lie Algebra
isomorphism. But, I think weve said more than enough for this course.
2.4 normed vector spaces
Denition 2.4.1.
Suppose is a vector space. If : is a function such that for all ,
and :
1. =
2. + + (triangle inequality)
3. 0
4. = 0 i = 0
then we say (, ) is a normed vector space. When there is no danger of ambiguity we
also say that is a normed vector space.
The norms below are basically thieved from the usual Euclidean norm on
.
Example 2.4.2.
can be given the Euclidean norm which is dened by =

for each

.
Example 2.4.3.
can also be given the 1-norm which is dened by

1
=
1
+
2
+ +
for each
.
We use the Euclidean norm by default.
Example 2.4.4. Consider as a two dimensional real vector space. Let + and dene
+ =
2
+
2
. This is a norm for .
Example 2.4.5. Let

. For each = [
] we dene
=
2
11
+
2
12
+ +
2
=1
=1
.
This is the Frobenius norm for matrices.
Each of the norms above allows us to dene a distance function and hence open sets and limits for
functions. An open ball in (,
) is dened
) = {
< }.
We dene the deleted open ball by removing the center from the open ball
){
} =
{ 0 <
< }. We say
is a limit point of a function i there exists a deleted open

ball which is contained in the (). We say is an open set i for each there exists
an open ball
() . Limits are also dened in the same way as in
, if : is a func-
tion from normed space (,
) to normed vector space (,
) then we say lim
() =
i for each > 0 there exists > 0 such that for all subject to 0 <
< it fol-
lows ()(
< . If lim
() = (
) then we say that is a continuous function at
.
Let (,
) be a normed vector space, a function from to is a called a sequence. Suppose

{
} is a sequence then we say lim
= i for each > 0 there exists such that
< for all with > . If lim
= then we say {
} is a convergent
sequence. We say {
} is a Cauchy sequence i for each > 0 there exists such that
< for all , with , > . In other words, a sequence is Cauchy if the
terms in the sequence get arbitarily close as we go suciently far out in the list. Many concepts
we cover in calculus II are made clear with proofs built around the concept of a Cauchy sequence.
The interesting thing about Cauchy is that for some spaces of numbers we can have a sequence
which converges but is not Cauchy. For example, if you think about the rational numbers we
can construct a sequence of truncated decimal expansions of :
{
} = {3, 3.1, 3.14, 3.141, 3.1415 . . . }

note that
for all and yet the
/ . When spaces are missing their limit points

they are in some sense incomplete. For this reason we say a metric space which contains all its
limit points is known as a complete space. Moreover, a normed vector space which is complete
is known as a Banach space. Fortunately all the main examples of this course are built on the
real numbers which are complete, this induces completeness for ,
and

. I may guide you
through the proof that , ,
and

are Banach spaces in a homework. When you take
real analysis youll spend some time thinking through the Cauchy concept.
Proposition 1.4.23 was given for the specic case of functions whose range is in . We might be able
to mimick the proof of that proposition for the case of normed spaces. We do have a composition
of limits theorem and I bet the sum function is continuous on a normed space. Moreover, if the
range happens to be a Banach algebra
6
then I would wager the product function is continuous.
Put these together and we get the normed vector space version of Prop. 1.4.23. That said, a direct
proof works nicely here so Ill just forego the more clever route here.
Proposition 2.4.6.
6
if is a Banach space that also has a product : such that 12 12 then is a
Banach algebra.
Let , be normed vector spaces. Let be a limit point of mappings , :
and suppose . If lim
() =
1
and lim
() =
2
then
1. lim
(() +()) = lim
() + lim
().
2. lim
(()) = lim
().
Moreover, if , are continuous then + and are continuous.
Proof: Let > 0 and suppose lim
() =
1
and lim
() =
2
. Choose
1
,
2
> 0
such that 0 < <
1
implies ()
1
< /2 and 0 < <
2
implies ()
2
< /2.
Choose = (
1
,
2
) and suppose 0 < <
1
,
2
hence
( +)() (
1
+
2
) = ()
1
+()
2
()
1
+()
2
< /2 +/2 = .
Item (2.) follows. To prove (2.) note that if = 0 the result is clearly true so suppose = 0.
Suppose > 0 and choose > 0 such that ()
1
< /. Note that if 0 < < then
()()
1
= (()
1
) = ()
1
< / = .
The claims about continuity follow immediately from the limit properties and that completes the
proof .
Perhaps you recognize these arguments from calculus I. The logic used to prove the basic limit
theorems on is essentially identical.
Proposition 2.4.7.
Suppose
1
,
2
,
3
are normed vector spaces with norms
1
,
2
,
3
respective. Let
: ()
2

3
and : ()
1

2
be mappings. Suppose that
lim
() =
and suppose that is continuous at
then
lim
(

)() =
_
lim
()
_
.
Proof: Let > 0 and choose > 0 such that 0 <
2
< implies () (
)
3
< . We
can choose such a since Since is continuous at
thus it follows that lim
() = (
).
Next choose > 0 such that 0 <
1
< implies ()
2
< . We can choose such
a because we are given that lim
() =
. Suppose 0 <
1
< and let = ()
note ()
2
< yields
2
< and consequently () (
)
3
< . Therefore, 0 <
1
< implies (())(
)
3
< . It follows that lim
((()) = (lim
()).
The squeeze theorem relies heavily on the order properties of . Generally a normed vector space
has no natural ordering. For example, is 1 > or is 1 < in ? That said, we can state a squeeze
theorem for functions whose domain reside in a normed vector space. This is a generalization of
what we learned in calculus I. That said, the proof oered below is very similar to the typical proof
which is not given in calculus I
7
Proposition 2.4.8. squeeze theorem.
Suppose : () , : () , : () where is a
normed vector space with norm . Let () () () for all on some > 0 ball
of
8
then we nd that the limits at
follow the same ordering,

lim
() lim
() lim
().
Moreover, if lim
() = lim
() = then lim
() = .
Proof: Suppose () () for all
1
()
for some
1
> 0 and also suppose lim
() =
and lim
() =
. We wish to prove that
. Suppose otherwise towards a

contradiction. That is, suppose
>
. Note that lim
[() ()] =
by the linearity
of the limit. It follows that for =
1
2
(
) > 0 there exists

2
> 0 such that
2
()
implies
() () (
) < =
1
2
(
). Expanding this inequality we have
1
2
(
) < () () (
) <
1
2
(
)
adding
yields,
3
2
(
) < () () <
1
2
(
) < 0.
Thus, () > () for all
2
()
. But, () () for all
1
()
so we nd a contradic-
tion for each
() where = (
1
,
2
). Hence
. The same proof can be applied to

and thus the rst part of the theorem follows.
Next, we suppose that lim
() = lim
() = and () () () for all

1
() for some
1
> 0. We seek to show that lim
() = . Let > 0 and choose

2
> 0
such that () < and () < for all
()
. We are free to choose such a
2
> 0 because the limits of and are given at = . Choose = (
1
,
2
) and note that if

()
then
() () ()
hence,
() () ()
but () < and () < imply < () and () < thus
< () () () < .
7
this is lifted word for word from my calculus I notes, however here the meaning of open ball is considerably more
general and the linearity of the limit which is referenced is the one proven earlier in this section
Therefore, for each > 0 there exists > 0 such that
()
implies () < so
lim
() = .
Our typical use of the theorem above applies to equations of norms from a normed vector space.
The norm takes us from to so the theorem above is essential to analyze interesting limits. We
shall make use of it in the next chapter.
Proposition 2.4.9. norm is continuous with respect to itself.
Suppose has norm then : dened by () = denes a continuous
function on .
Proof: Suppose
and
. We wish to show that for each > 0 there exists a

> 0 such that 0 <
< implies
< . Let > 0 and choose = then

0 <
< ... stuck.XXX .

Finally, we should like to nd a vector of the limits is the limit of the vector proposition for the
context of normed spaces. It is generally true for normed vector spaces that if we know the limits
of all the component functions converge to particular limits then the limit of a vector function is
simply the vector of those limits. The converse is not so simple because the basis expansion for a
normed vector space could fail to follow the pattern we expect from our study of
.
Let
2
have basis = {
1
,
2
} = {
1
, 3
1
+
2
} note the vector = 3
1
+
2
= 3
1
+(3
1
+
2
) =
2
. With respect to the basis we nd
1
= 3 and
2
= 1. The concept of length is muddled in these
coordinates. If we tried (incorrectly) to use the pythagorean theorem wed nd =
9 + 1 =
10
and yet the length of the vector is clearly just 1 since =
2
= (0, 1). The trouble with is that
it has dierent basis elements which overlap. To keep clear the euclidean idea of distance we must
insist on the use of an orthonormal basis.
Id rather not explain what that means at this point. Sucient to say that if is an orthonormal
basis then the coordinates preserve essentially the euclidean idea of vector length. In particular,
we can expect that if
=
=1
then
2
=
=1
.
Proposition 2.4.10.
Let , be normed vector spaces and suppose has basis = {
=1
such that when
=
=1
then
2
=
=1
. Suppose that : () is a mapping

with component functions
1
,
2
, . . . ,
with respect to the basis ( =
=1
). Let
be a limit point of then
lim
() = =
=1
lim
() =
for all = 1, 2, . . . .
Proof: Suppose lim
() = =
=1
. Since we assumed the basis of was orthonor-

mal we have that:
()
=1
()
2
= ()
2
where in the rst equality I simply added nonzero terms. With the inequality above in mind,
let > 0 and choose > 0 such that 0 < < implies () < . It follows that
()
2
<
2
and hence
()
< . The index is arbitrary therefore, lim
() =
for all
.
Conversely suppose lim
() =
for all
. Let = {
1
,
2
, . . . ,
}.
Let > 0 and choose, by virtue of the given limits for the component functions,
> 0 such
that 0 < <
implies
()
<

. Choose = {
1
,
2
, . . . ,
} and suppose
0 < < . Consider
() =
=1
(
()
=1
(
()
=1
()
However,
for all = 1, 2, . . . hence

()
=1
()
<
=1
=1
= .
Therefore, lim
() = .
I leave the case of non-orthonormal bases to the reader. In all the cases we consider it is possible
and natural to choose orthogonal bases to describe the vector space. Ill avoid the temptation to
do more here (there is more).
9
9
add a couple references for further reading here XXX
Chapter 3
dierentiation
Our goal in this chapter is to describe dierentiation for functions to and from normed linear spaces.
It turns out this is actually quite simple given the background of the preceding chapter. The dif-
ferential at a point is a linear transformation which best approximates the change in a function at
a particular point. We can quantify best by a limiting process which is naturally dened in view
of the fact there is a norm on the spaces we consider.
The most important example is of course the case :
. In this context it is natural to write

the dierential as a matrix multiplication. The matrix of the dierential is what Ewards calls the
derivative. Partial derivatives are also dened in terms of directional derivatives. The directional
derivative is sometimes dened where the dierential fails to exist. We will discuss how the criteria
of continuous dierentiability allows us to build the dierential from the directional derivatives.
Well see how the Cauchy-Riemann equations of complex analysis are really just an algebraic result
if we already have the theorem for continuously dierentiability. We will see how this general con-
cept of dierentiation recovers all the derivatives youve seen previously in calculus and much more.
On the other hand, I postpone implicit dierentiation for a future chapter where we have the
existence theorems for implicit and inverse functions. I also postpone discussion of the geometry of
the dierential. In short, existence of the dierential and the tangent space are essentially two sides
of the same problem. In fact, the approach of this chapter is radically dierent than my rst set of
notes on advanced calculus. Last year I followed Edwards a bit more and built up to the denition
of the dierential on the basis of the directional derivative and geometry. I dont think students
appreciate geometry or directional dierentiation well enough to make that approach successful.
Consquently, I begin with the unjustied denition of the derivative and then spend the rest of the
chapter working out precise implicationa and examples that ow fromt he dentition. I essentially
ignore the question of motivating the dentiion we begin with. If you want motivation, think
backward with this chapter or prehaps read Edwards or my old notes.
67
68 CHAPTER 3. DIFFERENTIATION
3.1 the dierential
The denition
1
below says that = ( +) ()
() when is close to zero.

Denition 3.1.1.
Let (,
) and (,
) be normed vector spaces. Suppose that is open and

: is a function the we say that is dierentiable at i there exists
a linear mapping : such that
lim
0
_
( +) () ()
_
= 0.
In such a case we call the linear mapping the dierential at and we denote =
.
In the case =
and =
are given the standard euclidean norms, the matrix of

the dierential is called the derivative of at and we denote [
] =
()

which means that
() =
() for all
.
Notice this denition gives an equation which implicitly denes
. For the moment the only way

we have to calculate
is educated guessing.
Example 3.1.2. Suppose : is a linear transformation of normed vector spaces and .
I propose = . In other words, I think we can show the best linear approximation to the change
in a linear function is simply the function itself. Clearly is linear since is linear. Consider the
dierence quotient:
( +) () ()
=
() +() () ()
=
0
.
Note = 0 implies
= 0 by the denition of the norm. Hence the limit of the dierence quotient
vanishes since it is identically zero for every nonzero value of . We conclude that
= .
Example 3.1.3. Let : where and are normed vector spaces and dene () =
for all . I claim the dierential is the zero transformation. Linearity of () = 0 is trivially
veried. Consider the dierence quotient:
( +) () ()
=
0
.
Using the arguments to the preceding example, we nd
= 0.
Typically the dierence quotient is not identically zero. The pair of examples above are very special
cases. Ill give a few more abstract examples later in this section. For now we turn to the question
of how this general denition recovers the concept of dierentiation we studied in calculus.
1
Some authors might put a norm in the numerator of the quotient. That is an equivalent condition since a function
: has lim
0
() = 0 i lim
0
() = 0
3.1. THE DIFFERENTIAL 69
Example 3.1.4. Suppose : () is dierentiable at . It follows that there exists a
linear function
: such that
2
lim
0
( +) ()
()
= 0.
Since
: is linear there exists a constant matrix such that
() = . In this silly
case the matrix is a 1 1 matrix which otherwise known as a real number. Note that
lim
0
( +) ()
()
= 0 lim
0
( +) ()
()
= 0.
In the left limit 0
we have < 0 hence = . On the other hand, in the right limit 0

+
we have > 0 hence = . Thus, dierentiability suggests that lim
0
(+)()()
= 0.
But we can pull the minus out of the left limit to obtain lim
0
(+)()()
= 0. Therefore,
lim
0
( +) ()
()
= 0.
We seek to show that lim
0
(+)()
= .
= lim
0
= lim
0
()
A theorem from calculus I states that if lim( ) = 0 and lim() exists then so must lim() and
lim() = lim(). Apply that theorem to the fact we know lim
0
()
exists and
lim
0
_
( +) ()
()
_
= 0.
It follows that
lim
0
()
= lim
0
( +) ()
.
Consequently,
() = lim
0
( +) ()
dened
() in calc. I.
Therefore,
() =
() . In other words, if a function is dierentiable in the sense we dened

at the beginning of this chapter then it is dierentiable in the terminology we used in calculus I.
Moreover, the derivative at is precisely the matrix of the dierential.
2
unless we state otherwise,
is assumed to have the euclidean norm, in this case
2
= .
2
3
is dened by (, ) = (,
2
, +3) for all (, )
2
.
Consider the dierence function at (, ):
= ((, ) + (, )) (, ) = ( +, +) (, )
Calculate,
=
_
( +)( +), ( +)
2
, + + 3( +)
_
_
,
2
, + 3
_
Simplify by cancelling terms which cancel with (, ):
=
_
+, 2 +
2
, + 3)
_
Identify the linear part of as a good candidate for the dierential. I claim that:
(, ) =
_
+, 2, + 3
_
.
is the dierential for at (x,y). Observe rst that we can write
(, ) =

2 0
1 3
_
.
therefore :
2
3
is manifestly linear. Use the algebra above to simplify the dierence quotient
below:
lim
(,)(0,0)
_
(, )
(, )
_
= lim
(,)(0,0)
_
(0,
2
, 0)
(, )
_
Note (, ) =
2
+
2
therefore we fact the task of showing that (0,
2
/
2
+
2
, 0) (0, 0, 0)
as (, ) (0, 0). Recall from our study of limits that we can prove the vector tends to (0, 0, 0)
by showing the each component tends to zero. The rst and third components are obviously zero
however the second component requires study. Observe that
0

2
2
+
2

2
2
=
Clearly lim
(,)(0,0)
(0) = 0 and lim
(,)(0,0)
= 0 hence the squeeze theorem for multivariate
limits shows that lim
(,)(0,0)
2
+
2
= 0. Therefore,
(,)
(, ) =

2 0
1 3
_
.
Computation of less trivial multivariate limits is an art wed like to avoid if possible. It turns out
that we can actually avoid these calculations by computing partial derivatives. However, we still
need a certain multivariate limit to exist for the partial derivative functions so in some sense its
unavoidable. The limits are there whether we like to calculate them or not. I want to give a few
more abstract examples before I get into the partial dierentiation. The purpose of this section is
to showcase the generality of the denition for dierential.
Example 3.1.6. Suppose () = ()+ () for all () and both and are dierentiable
functions on (). By the arguments given in Example 3.1.4 it suces to nd : such
that
lim
0
_
( +) () ()
_
= 0.
I propose that on the basis of analogy to Example 3.1.4 we ought to have
() = (
()+
()).
Let () = (
() +
()). Observe that, using properties of , (

1
+
2
) =
= (
() +
())(
1
+
2
) = (
() +
())
1
+(
() +
())
2
= (
1
) +(
2
).
for all
1
,
2
and . Hence : is linear. Moreover,
(+)()()
=
1
_
( +) + ( +) () + () (
() +
())
_
=
1
_
( +) ()
()
_
+
1
_
( +) ()
()
_
Consider the problem of calculating lim
0
(+)()()
. We use a lemma that a complex

function converges to zero i the real and imaginary parts of the function separately converge to
zero (this might be a homework). By dierentiability of and we nd again using Example 3.1.4
lim
0
1
_
( +) ()
()
_
= 0 lim
0
1
_
( +) ()
()
_
= 0.
Therefore,
() = (
() +
()). Note that the quantity
() +
() is not a real matrix

in this case. To write the derivative in terms of a real matrix multiplication we need to construct
some further notation which makes use of the isomorphism between and
2
. Actually, its pretty
easy if you agree that + = (, ) then
() = (
(),
()) so the matrix of the dierential

is (
(),
())
12
which makes since as :
2
.
Generally constructing the matrix for a function : where , = involves a fair
number of relatively ad-hoc conventions because the constructions necessarily involving choosing
coordinates. The situation is similar in linear algebra. Writing abstract linear transformations in
terms of matrix multiplication takes a little thinking. If you look back youll notice that I did not
bother to try to write a matrix Examples 3.1.2 or 3.1.3. The same is true for the nal example of
this section.

is dened by () =
2
. Notice
= ( +) () = ( +)( +)
2
= + +
2
I propose that is dierentiable at and () = +. Lets check linearity,
(
1
+
2
) = (
1
+
2
) + (
1
+
2
) =
1
+
1
+(
2
+
2
)
Hence :

is a linear transformation. By construction of the linear terms in the
numerator cancel leaving just the quadratic term,
lim
0
( +) () ()
= lim
0
.
It suces to show that lim
0
= 0 since lim() = 0 i lim() = 0 in a normed vector

space. Fortunately the normed vector space

is actually a Banach algebra. A vector space
with a multiplication operation is called an algebra. In the current context the multiplication is
simply matrix multiplication. A Banach algebra is a normed vector space with a multiplication that
satises . Thanks to this inequality we can calculate our limit via the squeeze
theorem. Observe 0

2
. As 0 it follows 0 hence lim

0
= 0. We
nd
() = +.
XXX- need to adjust example below to reect orthonormality assumption.
Example 3.1.8. Suppose is a normed vector space with basis = {
1
,
2
, . . . ,
}. Futhermore,
let : be dened by
() =
=1
()
where
: is dierentiable on for = 1, 2, . . . , . I claim that if =
=1
:
then lim
0
() =
=1
i lim
0
() =
for all = 1, 2, . . . , . In words, the limit of a

vector-valued function can be parsed into a vector of limits. Weve not proved this, I may make it
a homework. With this in mind consider (again we can trade for as we explained in-depth
in Example 3.1.4) the dierence quotient lim
0
_
(+)()
=1
_
, factoring out the basis
yields:
lim
0
_
=1
[
( +)
()
_
=
=1
_
lim
0
( +)
()
The expression on the left is the limit of a vector whereas the expression on the right is a vector of
limits. I make the equality by applying the claim. In any event, I hope you are not surprised that:
() =
=1
The example above encompasses a number of cases at once:

1. = , functions on , :
2. =
, space curves in , :
3. = , complex-valued functions of a real variable, = + :

3.2. PARTIAL DERIVATIVES AND THE EXISTENCE OF THE DIFFERENTIAL 73
4. =

, matrix-valued functions of a real variable, :

.
In short, when we dierentiate a function which has a real domain then we can dene the derivative
of such a function by component-wise dierentiation. It gets more interesting when the domain has
several independent variables. We saw this in Examples 3.1.5 and 3.1.7.
Remark 3.1.9.
I have deliberately dened the derivative in slightly more generality than we need for this
course. Its probably not much trouble to continue to develop the theory of dierentiation
for a normed vector space, however I will for the most part stop here. The theorems that
follow are not terribly complicated in the notation of
and traditionally this type of

course only covers continuous dierentiability, inverse and implicit function theorems in
the context of mappings from
to
. For the reader interested in generalizing these

results to the context of an abstract normed vector space feel free to discuss it with me
sometime. This much we can conclude from our brief experience thus far, if we study
functions whose domain is in then dierentiation is accomplished component-wise in the
range. This is good news since in all your previous courses I simply dened dierentiation
by the component-wise rule. This section at a minimum shows that idea is consistent with
the larger theory we are working out in this chapter. It is likely I will have you work out
the calculus of complex or matrix-valued functions of a real variable in the homework.
3.2 partial derivatives and the existence of the dierential
In the preceding section we calculated the dierential at a point via educated guessing. We should
like to nd better method to derive dierentials. It turns out that we can systematically calculate
the dierential from partial derivatives of the component functions. However, certain topological
conditions are required for us to properly paste together the partial derivatives of the component
functions. We describe how the criteria of continuous dierentiability achieves this goal. Much of
this section was covered in calculus III but we do bring new generality and vision to the calculations
described here.
3.2.1 directional derivatives
The directional derivative of a mapping at a point () along is dened to be the
derivative of the curve () = ( + ). In other words, the directional derivative gives you the
instantaneous vector-rate of change in the mapping at the point along . In the case that
= 1 then : ()
and the directional derivative gives the instantaneous rate of

change of the function at the point along . You probably insisted that = 1 in calculus III
but we make no such demand here. We dene the directional derivative for mappings and vectors
of non-unit length.
Denition 3.2.1.
Let : ()
and suppose the limit below exists for () and
then we dene the directional derivative of at along to be
()
where
() = lim
0
( +) ()
One great contrast we should pause to note is that the denition of the directional derivative is
explicit whereas the denition of the dierential was implicit. Many similarities do exist. For
example: the directional derivative
() and the dierential
() are both is homogenous in

.
Proposition 3.2.2.
Let : ()
then if
() exists in
then
() =
()
Proof: Let : ()
and suppose
()
. This means we are given that

lim
0
(+)()
()
. If = 0 then the proposition is clearly true. Consider, for

nonzero ,
lim
0
( +()) ()
= lim
0
( +()) ()
Hence by the substitution = we nd,

lim
0
( +()) ()
= lim
0
( +()) ()
Therefore, the limit on the left of the equality exists as the limit on the right of the equality is
given and we conclude
() =
() for all .
If were given the derivative of a mapping then the directional derivative exists. The converse is
not so simple as we shall discuss in the next subsection.
Proposition 3.2.3.
If :
is dierentiable at then the directional derivative
() exists
for each
and
() =
().
Proof: Suppose such that
is well-dened then we are given that

lim
0
( +) ()
()
= 0.
This is a limit in
, when it exists it follows that the limits that approach the origin along
particular paths also exist and are zero. In particular we can consider the path for = 0
and > 0, we nd
lim
0, >0
( +) ()
()
=
1
lim
0
+
( +) ()
()
= 0.
Hence, as = for > 0 we nd
lim
0
+
( +) ()
= lim
0
()
().
Likewise we can consider the path for = 0 and < 0
lim
0, <0
( +) ()
()
=
1
lim
0
( +) ()
()
= 0.
Note = thus the limit above yields
lim
0
( +) ()
= lim
0
()
lim
0
( +) ()
().
Therefore,
lim
0
( +) ()
()
and we conclude that
() =
() for all
since the = 0 case follows trivially.

Lets think about the problem we face. We want to nd a nice formula for the dierential. We
now know that if it exists then the directional derivatives allow us to calculate the values of the
dierential in particular directions. The natural thing to do is to calculate the standard matrix
for the dierential using the preceding proposition. Recall that if :
then the standard

matrix was simply [] = [(
1
)(
2
) (
)] and thus the action of is expressed nicely as a

matrix multiplication; () = []. Similarly,
is linear transformation and thus
() = [
] where [
] = [
(
1
)
(
2
)
)]. Moreover, by the preceding proposition

we can calculate
) =
() for = 1, 2, . . . , . Clearly the directional derivatives in the

coordinate directions are of great importance. For this reason we make the following denition:
Denition 3.2.4.
Suppose that :
is a mapping the we say that is has partial derivative
() at i the directional derivative in the
direction exists at . In this case we

denote,
() =
().
Also we may use the notation
() =
() or
when convenient. We also

construct the partial derivative mapping
as the mapping dened

pointwise for each where
() exists.
Lets expand this denition a bit. Note that if = (
1
,
2
, . . . ,
) then
() = lim
0
( +
) ()
()]
= lim
0
( +
()
for each = 1, 2, . . . . But then the limit of the component function
is precisely the directional

derivative at along
hence we nd the result
in other words,
= (
1
,
2
, . . . ,
).
Proposition 3.2.5.
If :
is dierentiable at then the directional derivative
() can
be expressed as a sum of partial derivative maps for each =<
1
,
2
, . . . ,
>
() =
=1
()
Proof: since is dierentiable at the dierential
exists and
() =
() for all
.
Use linearity of the dierential to calculate that
() =
(
1
1
+ +
) =
1
(
1
) + +
).
Note
) =
() =
() and the prop. follows.

My primary interest in advanced calculus is the dierential
3
. I discuss the directional derivative
here merely to connect with your past calculations in calculus III where we explored the geometric
and analytic signicance of the directional derivative. I do not intend to revisit all of that here
once more. Our focus is elsewhere. That said its probably best to include the example below:
3
then = [
and we can write the directional

derivative in terms of
= [
=
if we insist that = 1 then we recover the standard directional derivative we discuss in calculus
III. Naturally the () yields the maximum value for the directional derivative at if we
limit the inputs to vectors of unit-length. If we did not limit the vectors to unit length then the
directional derivative at can become arbitrarily large as
() is proportional to the magnitude

of . Since our primary motivation in calculus III was describing rates of change along certain
directions for some multivariate function it made sense to specialize the directional derivative to
vectors of unit-length. The denition used in these notes better serves the theoretical discussion. If
you read my calculus III notes youll nd a derivation of how the directional derivative in Stewarts
calculus arises from the general denition of the derivative as a linear mapping. Look up page 305g.
Incidentally, those notes may well be better than these in certain respects.
3
this is why I have yet to give an example in this section, you should get out your calculus III notes if you need a
refresher on directional derivatives
Proposition 3.2.7.
If :
is dierentiable at then the dierential
has derivative
matrix
() and it has components which are expressed in terms of partial derivatives of

the component functions:
[
for 1 and 1 .
Perhaps it is helpful to expand the derivative matrix explicitly for future reference:
() =
1
()
2
1
()
1
()
2
()
2
2
()
2
()
.
.
.
.
.
.
.
.
.
.
.
.
()
2
()
()
Lets write the operation of the dierential for a dierentiable mapping at some point in
terms of the explicit matrix multiplication by
(). Let = (
1
,
2
, . . .
() =
() =
1
()
2
1
()
1
()
2
()
2
2
()
2
()
.
.
.
.
.
.
.
.
.
.
.
.
()
2
()
()
2
.
.
.
You may recall the notation from calculus III at this point, omitting the -dependence,
= (
) =
_

1
,
2
, ,
So if the derivative exists we can write it in terms of a stack of gradient vectors of the component
functions: (I used a transpose to write the stack side-ways),
=
_
Finally, just to collect everything together,
1

2
2

2
2
.
.
.
.
.
.
.
.
.
.
.
.
=
_
1

2

(
1
)
(
2
)
.
.
.
(
Example 3.2.8. Recall that in Example 3.1.5 we showed that :

2

3
dened by (, ) =
(,
2
, + 3) for all (, )
2
was dierentiable. In fact we calculated that
(,)
(, ) =

2 0
1 3
_
.
If you recall from calculus III the mechanics of partial dierentiation its simple to see that
(,
2
, + 3) = (, 2, 1) =
2
1
(,
2
, + 3) = (, 0, 3) =
0
3
Thus [] = [
] (as we expect given the derivations in this section!)

3.2.2 continuously dierentiable, a cautionary tale
We have noted that dierentiablility on some set implies all sorts of nice formulas in terms of
the partial derivatives. Curiously the converse is not quite so simple. It is possible for the partial
derivatives to exist on some set and yet the mapping may fail to be dierentiable. We need an extra
topological condition on the partial derivatives if we are to avoid certain pathological
4
examples.
Example 3.2.9. I found this example in Hubbards advanced calculus text(see Ex. 1.9.4, pg. 123).
It is a source of endless odd examples, notation and bizarre quotes. Let () = 0 and
() =

2
+
2
sin
1
for all = 0. I can be shown that the derivative
(0) = 1/2. Moreover, we can show that
()
exists for all = 0, we can calculate:
() =
1
2
+ 2sin
1
cos
1
Notice that (
) = . Note then that the tangent line at (0, 0) is = /2.

4
pathological as in, your clothes are so pathological, whered you get them?
You might be tempted to say then that this function is increasing at a rate of 1/2 for near zero.
But this claim would be false since you can see that
() oscillates wildly without end near zero.

We have a tangent line at (0, 0) with positive slope for a function which is not increasing at (0, 0)
(recall that increasing is a concept we must dene in a open interval to be careful). This sort of
thing cannot happen if the derivative is continuous near the point in question.
The one-dimensional case is really quite special, even though we had discontinuity of the derivative
we still had a well-dened tangent line to the point. However, many interesting theorems in calculus
of one-variable require the function to be continuously dierentiable near the point of interest. For
example, to apply the 2nd-derivative test we need to nd a point where the rst derivative is zero
and the second derivative exists. We cannot hope to compute
) unless
is continuous at
.
The next example is sick.
Example 3.2.10. Let us dene (0, 0) = 0 and
(, ) =

2
2
+
2
for all (, ) = (0, 0) in
2
. It can be shown that is continuous at (0, 0). Moreover, since
(, 0) = (0, ) = 0 for all and all it follows that vanishes identically along the coordinate
axis. Thus the rate of change in the
1
or
2
directions is zero. We can calculate that
=
2
3
(
2
+
2
)
2
and

=

4
2
(
2
+
2
)
2
Consider the path to the origin (, ) gives
(, ) = 2
4
/(
2
+
2
)
2
= 1/2 hence
(, ) 1/2
along the path (, ), but
(0, 0) = 0 hence the partial derivative
is not continuous at (0, 0).

In this example, the discontinuity of the partial derivatives makes the tangent plane fail to exist.
XXX need to include graph of this thing.
Denition 3.2.11.
A mapping :
is continuously dierentiable at i the partial

derivative mappings
exist on an open set containing and are continuous at .

The dention above is interesting because of the proposition below. The import of the proposition
is that we can build the tangent plane from the Jacobian matrix provided the partial derivatives
are all continuous. This is a very nice result because the concept of the linear mapping is quite
abstract but partial dierentiation of a given mapping is easy.
Proposition 3.2.12.
If is continuously dierentiable at then is dierentiable at
Well follow the proof in Edwards on pages 72-73.
- need to include the continuously di. discuss here then do the examples.
- need to eliminate discussion of directional derivative except where absolutely necessary
- think about where to dene the dierential as a form. Consider peppering dierential at a point
many places...
3.2.3 gallery of derivatives
Our goal here is simply to exhbit the Jacobian matrix and partial derivatives for a few mappings.
At the base of all these calculations is the observation that partial dierentiation is just ordinary
dierentiation where we treat all the independent variable not being dierentiated as constants.
The criteria of indepedence is important. Well study the case the variables are not independent
in a later section.
Remark 3.2.13.
I have put remarks about the rank of the derivative in the examples below. Of course this
has nothing to do with the process of calculating Jacobians. Its something to think about
once we master the process of calculating the Jacobian. Ignore the red comments for now
if you wish
Example 3.2.14. Let () = (,
2
,
3
) then
() = (1, 2, 3
2
). In this case we have
() = [
] =
1
2
3
2
The Jacobian here is a single column vector. It has rank 1 provided the vector is nonzero. We
see that
() = (0, 0, 0) for all . This corresponds to the fact that this space curve has a
well-dened tangent line for each point on the path.
Example 3.2.15. Let (, ) = be a mapping from
3
3
. Ill denote the coordinates
in the domain by (
1
,
2
,
3
,
1
,
2
,
3
) thus (, ) =
1
1
+
2
2
+
3
3
. Calculate,
[
(,)
] = (, )
= [
1
,
2
,
3
,
1
,
2
,
3
]
The Jacobian here is a single row vector. It has rank 6 provided all entries of the input vectors are
nonzero.
Example 3.2.16. Let (, ) = be a mapping from
. Ill denote the coordinates

in the domain by (
1
, . . . ,
,
1
, . . . ,
) thus (, ) =
=1
. Calculate,
=1
_
=
=1
=1
Likewise,
=1
_
=
=1
=1
Therefore, noting that = (
1
, . . . ,
1
, . . . ,
),
[
(,)
]
= ()(, ) = = (
1
, . . . ,
,
1
, . . . ,
)
The Jacobian here is a single row vector. It has rank 2n provided all entries of the input vectors
are nonzero.
Example 3.2.17. Suppose (, , ) = (, , ) we calculate,
= (, 0, 0)

= (, 1, 0)

= (, 0, 1)
Remember these are actually column vectors in my sneaky notation; (
1
, . . . ,
) = [
1
, . . . ,
.
This means the derivative or Jacobian matrix of at (, , ) is
(, , ) = [
(,,)
] =

0 1 0
0 0 1
Note, (
(, , )) = 3 for all (, , )
3
such that , = 0. There are a variety of ways to
see that claim, one way is to observe [
(, , )] = and this determinant is nonzero so long

as neither nor is zero. In linear algebra we learn that a square matrix is invertible i it has
nonzero determinant i it has linearly indpendent column vectors.
Example 3.2.18. Suppose (, , ) = (
2
+
2
, ) we calculate,
= (2, 0)

= (0, )

= (2, )
The derivative is a 2 3 matrix in this example,
(, , ) = [
(,,)
] =
_
2 0 2
0
_
The maximum rank for
is 2 at a particular point (, , ) because there are at most two linearly

independent vectors in
2
. You can consider the three square submatrices to analyze the rank for
a given point. If any one of these is nonzero then the rank (dimension of the column space) is two.
1
=
_
2 0
0
_

2
=
_
2 2
0
_

3
=
_
0 2

_
Well need either (
1
) = 2 = 0 or (
2
) = 2 = 0 or (
3
) = 2
2
= 0. I believe
the only point where all three of these fail to be true simulataneously is when = = = 0. This
mapping has maximal rank at all points except the origin.
2
+
2
, , +) we calculate,
= (2, , 1)

= (2, , 1)
The derivative is a 3 2 matrix in this example,
(, ) = [
(,)
] =
2 2

1 1
The maximum rank is again 2, this time because we only have two columns. The rank will be two
if the columns are not linearly dependent. We can analyze the question of rank a number of ways
but I nd determinants of submatrices a comforting tool in these sort of questions. If the columns
are linearly dependent then all three sub-square-matrices of
will be zero. Conversely, if even one

of them is nonvanishing then it follows the columns must be linearly independent. The submatrices
for this problem are:
1
=
_
2 2

_

2
=
_
2 2
1 1
_

3
=
_

1 1
_
You can see (
1
) = 2(
2
2
), (
2
) = 2( ) and (
3
) = . Apparently we have
(
(, , )) = 2 for all (, )
2
with = . In retrospect this is not surprising.
Example 3.2.20. Suppose (, , ) = (
,
1
) = (
1
2
2
+
1
2
2
, ) for some constant . Lets
calculate the derivative via gradients this time,
= (
/,
/,
/) = (, ,
1
2
2
)
1
= (
1
/,
1
/,
1
/) = (0, , )
Therefore,
(, , ) =
_

1
2
2
0
_
Example 3.2.21. Let (, ) = ( cos , sin ). We calculate,
= (cos , sin ) and
= ( sin , cos )
Hence,
(, ) =
_
cos sin
sin cos
_
We calculate (
(, )) = thus this mapping has full rank everywhere except the origin.
Example 3.2.22. Let (, ) = (
2
+
2
, tan
1
(/)). We calculate,
=
_

2
+
2
,

2
+
2
_
and
=
_

2
+
2
,

2
+
2
_
Hence,
(, ) =
_

2
+
2
2
+
2
2
+
2
2
+
2
_
=
_

2
_
_
using =
2
+
2
_
We calculate (
(, )) = 1/ thus this mapping has full rank everywhere except the origin.
Example 3.2.23. Let (, ) = (, ,
2
) for a constant . We calculate,
2
=
_

2
,

2
_
Also, = (1, 0) and = (0, 1) thus
(, ) =
1 0
0 1
This matrix clearly has rank 2 where is is well-dened. Note that we need
2
2
> 0 for the
derivative to exist. Moreover, we could dene (, ) = (
2
, , ) and calculate,
(, ) =
1 0
2
0 1
.
Observe that
(, ) exists when
2

2

2
> 0. Geometrically, parametrizes the sphere
above the equator at = 0 whereas parametrizes the right-half of the sphere with > 0. These
parametrizations overlap in the rst octant where both and are positive. In particular, (
)
(
) = {(, )
2
, > 0 and
2
+
2
<
2
}
Example 3.2.24. Let (, , ) = (, , ,
2
) for a constant . We calculate,
2
=
_

2
,

2
,

2
_
Also, = (1, 0, 0), = (0, 1, 0) and = (0, 0, 1) thus
(, , ) =
1 0 0
0 1 0
0 0 1
This matrix clearly has rank 3 where is is well-dened. Note that we need
2
2
> 0 for the
derivative to exist. This mapping gives us a parametrization of the 3-sphere
2
+
2
+
2
+
2
=
2
for > 0. (drawing this is a little trickier)
Example 3.2.25. Let (, , ) = ( +, +, +, ). You can calculate,
[
(,,)
] =
1 1 0
0 1 1
1 0 1

This matrix clearly has rank 3 and is well-dened for all of

3
.
Example 3.2.26. Let (, , ) = . You can calculate,
[
(,,)
] =
_

This matrix fails to have rank 3 if , or are zero. In other words,
(, , ) has rank 3 in
3
provided we are at a point which is not on some coordinate plane. (the coordinate planes are
= 0, = 0 and = 0 for the , and coordinate planes respective)
Example 3.2.27. Let (, , ) = (, 1 ). You can calculate,
[
(,,)
] =
_

1 1 0
_
This matrix has rank 3 if either = 0 or ( ) = 0. In contrast to the preceding example, the
derivative does have rank 3 on certain points of the coordinate planes. For example,
(1, 1, 0) and
(0, 1, 1) both give (
) = 3.
3
3
be dened by () = for a xed vector = 0. We denote
= (
1
,
2
,
3
) and calculate,
( ) =

_

,,
_
=
,,
,,
It follows,
1
( ) =
=
2
2
= (0,
3
,
2
)
2
( ) =
=
3
3
= (
3
, 0,
1
)
3
( ) =
=
1
1
= (
2
,
1
, 0)
Thus the Jacobian is simply,
[
(,)
] =
0
3

2
3
0
1
2

1
0
In fact,
() = () = for each
3
. The given mapping is linear so the dierential of
the mapping is precisely the mapping itself.
Example 3.2.29. Let (, ) = (, , 1 ). You can calculate,
[
(,,)
] =
1 0
0 1
1 1
Example 3.2.30. Let (, ) = (, , ) where , , denote functions of , and I prefer to omit

the explicit depedendence to reduce clutter in the equations to follow.
= (
) and

= (
)
Then the Jacobian is the 3 2 matrix
_
(,)
The matrix
_
(,)
has rank 2 if at least one of the determinants below is nonzero,
_

_

_

_

_
Example 3.2.31. . .
Example 3.2.32. . .
3.3 additivity and homogeneity of the derivative
Suppose
1
:
and
2
:
. Furthermore, suppose both of these are

dierentiable at
. It follows that (
1
)
=
1
and (
2
)
=
2
are linear operators from
to
which approximate the change in

1
and
2
near , in particular,
lim
0
1
( +)
1
()
1
()
= 0 lim
0
2
( +)
2
()
2
()
= 0
To prove that =
1
+
2
is dierentiable at
we need to nd a dierential at for .

Naturally, we expect
= (
1
+
2
)
= (
1
)
+(
2
)
. Let = (
1
)
+(
2
)
and consider,
lim
0
(+)()()
= lim
0
1
(+)+
2
(+)
1
()
2
()
1
()
2
()
= lim
0
1
(+)
1
()
1
()
+ lim
0
2
(+)
2
()
2
()
= 0 + 0
= 0
Note that breaking up the limit was legal because we knew the subsequent limits existed and
were zero by the assumption of dierentiability of
1
and
2
at . Finally, since =
1
+
2
we
know is a linear transformation since the sum of linear transformations is a linear transformation.
Moreover, the matrix of is the sum of the matrices for
1
and
2
. Let and suppose =
1
then we can also show that
= (
1
)
= (
1
)
, the calculation is very similar except we just

pull the constant out of the limit. Ill let you write it out. Collecting our observations:
Proposition 3.3.1.
Suppose
1
:
and
2
:
are dierentiable at then
1
+
2
is dierentiable at and
(
1
+
2
)
= (
1
)
+ (
2
)
or (
1
+
2
)
() =
1
() +
2
()
Likewise, if then
(
1
)
= (
1
)
or (
1
)
() = (
1
())
These results suggest that the dierential of a function is a new object which has a vector space
structure. There is much more to say here later.
3.4 chain rule
The proof in Edwards is on 77-78. Ill give a heuristic proof here which captures the essence of the
argument. The simplicity of this rule continues to amaze me.
3.4. CHAIN RULE 87
Proposition 3.4.1.
If :
is dierentiable at and :
is dierentiable at
() then
is dierentiable at and
(
= ()
()

or, in matrix notation, (
() =
(())
()
Proof Sketch:
In calculus III you may have learned how to calculate partial derivatives in terms of tree-diagrams
and intermediate variable etc... We now have a way of understanding those rules and all the
other chain rules in terms of one over-arching calculation: matrix multiplication of the constituent
Jacobians in the composite function. Of course once we have this rule for the composite of two
functions we can generalize to -functions by a simple induction argument. For example, for three
suitably dened mappings , , ,
(

() =
((()))
(())
()
Example 3.4.2. . .
Example 3.4.3. . .
Example 3.4.4. . .
Example 3.4.5. . .
3.4. CHAIN RULE 89
Example 3.4.6. . .
Example 3.4.7. . .
3.5 product rules?
What sort of product can we expect to nd among mappings? Remember two mappings have
vector outputs and there is no way to multiply vectors in general. Of course, in the case we have
two mappings that have equal-dimensional outputs we could take their dot-product. There is a
product rule for that case: if

,

:
then

) = (

)

) +

(

)
Or in the special case of = 3 we could even take their cross-product and there is another product
rule in that case:

) = (

)

+

(

)
What other case can we multiply vectors? One very important case is
2
= where is is
customary to use the notation (, ) = + and = + . If our range is complex numbers
then we again have a product rule: if :
and :
then
() = (
) +(
)
I have relegated the proof of these product rules to the end of this chapter. One other object worth
dierentiating is a matrix-valued function of
. If we dene the partial derivative of a matrix to

be the matrix of partial derivatives then partial dierentiation will respect the sum and product of
matrices (we may return to this in depth if need be later on):
(+) =
() = (
) +(
)
Moral of this story? If you have a pair mappings whose ranges allow some sort of product then it
is entirely likely that there is a corresponding product rule
5
.
3.5.1 scalar-vector product rule
There is one product rule which we can state for arbitrary mappings, note that we can always
sensibly multiply a mapping by a function. Suppose then that :
and :
are dierentiable at . It follows that there exist linear transformations
and
where
lim
0
( +) ()
()
= 0 lim
0
( +) ()
()
= 0
Since ( +) () +
() and ( +) () +
() we expect
( +) (() +
())(() +
())
()() +()
() +()
()
. .
linear in
+
()
()
. .
2
order in
5
In my research I consider functions on supernumbers, these also can be multiplied. Naturally there is a product
rule for super functions, the catch is that super numbers , do not necessarily commute. However, if theyre
homogeneneous = (1)
. Because of this the super product rule is () = () + (1)
()
3.5. PRODUCT RULES? 91
Thus we propose: () = ()
() +()
() is the best linear approximation of .

lim
0
()( +) ()() ()
=
= lim
0
( +)( +) ()() ()
() ()
()
= lim
0
( +)( +) ()() ()
() ()
()
+
+ lim
0
()( +) ( +)()
+ lim
0
( +)() ()( +)
+ lim
0
()() ()()
= lim
0
_
()
( +) ()
()
+
( +) ()
()
()+
+
_
( +) ()
_
( +) ()
_
= ()
_
lim
0
( +) ()
()
_
+
_
lim
0
( +) ()
()
_
()
= 0
Where we have made use of the dierentiability and the consequent continuity of both and at
. Furthermore, note
( +) = ()
( +) +()
( +)
= ()(
() +
()) +()(
() +
())
= ()
() +()(
() +(()
() +()
())
= () +()
for all ,
and hence = ()
+()
is a linear transformation. We have proved

(most of) the following proposition:
Proposition 3.5.1.
If :
and :
are dierentiable at then is

dierentiable at and
()
= ()
() +()
()
() =
()() +()
()
The argument above covers the ordinary product rule and a host of other less common rules. Note
again that () and
() are vectors.
3.5.2 calculus of paths in
3
A path is a mapping from to
. We use such mappings to model position, velocity and

acceleration of particles in the case = 3. Some of these things were proved in previous sections
of this chapter but I intend for this section to be self-contained so that you can read it without
digging through the rest of this chapter.
Proposition 3.5.2.
If , :
are dierentiable vector-valued functions and : is a

dierentiable real-valued function then for each ,
1. ( +)
() =
() +
().
2. ()
() =
().
3. ()
() =
()() +()
().
4. ( )
() =
() () +()
().
5. provided = 3, ( )
() =
() () +()
().
6. provided () (
), (

)
() =
()(()).
We have to insist that = 3 for the statement with cross-products since we only have a standard
cross-product in
3
. We prepare for the proof of the proposition with a useful lemma. Notice this
lemma tells us how to actually calculate the derivative of paths in examples. The derivative of
component functions is nothing more than calculus I and one of our goals is to reduce things to
those sort of calculations whenever possible.
Lemma 3.5.3.
If :
is dierentiable vector-valued function then for all ,
() = (
1
(),
2
(), . . . ,
())
We are given that the following vector limit exists and is equal to
(),
() = lim
0
( +) ()
then by Proposition 1.4.10 the limit of a vector is related to the limits of its components as follows:
()
= lim
0
( +)
()
.
Thus (
())
() and the lemma follows

6
.
6
this notation I rst saw in a text by Marsden, it means the proof is partially completed but you should read on
to nish the proof
Proof of proposition: We use the notation =
= (
1
, . . . ,
) and =
=
(
1
, . . . ,
) throughout the proofs below. The
is understood to range over 1, 2, . . . . Begin

with (1.),
[( +)
[( +)
] using the lemma

=

] using def. ( +)
] +

] by calculus I, ( +)
.
= [
def. of vector addition for
and
Hence ( )
.The proofs of 2,3,5 and 6 are similar. Ill prove (5.),

[( )
[( )
] using the lemma

=

] using def.
=
] repeatedly using, ( +)
] repeatedly using, ()
] property of nite sum
= (
+ (

) def. of cross product

=
_

def. of vector addition

Notice that the calculus step really just involves calculus I applied to the components. The ordinary
product rule was the crucial factor to prove the product rule for cross-products. Well see the same
for the dot product of mappings. Prove (4.)
( )
() =

] using def.
=
] repeatedly using, ( +)
] repeatedly using, ()
. def. of dot product

The proof of (3.) follows from applying the product rule to each component of ()(). The proof
of (2.) follow from (3.) in the case that () = so
() = 0. Finally the proof of (6.) follows

from applying the chain-rule to each component.
3.5.3 calculus of matrix-valued functions of a real variable
Denition 3.5.4.
A matrix-valued function of a real variable is a function from to

. Suppose
:

is such that
: is dierentiable for each , then we

dene
=
_
which can also be denoted (
. We likewise dene
= [
] for with
integrable components. Denite integrals and higher derivatives are also dened component-
wise.
Example 3.5.5. Suppose () =
_
2 3
2
4
3
5
4
_
. Ill calculate a few items just to illustrate the
denition above. calculate; to dierentiate a matrix we dierentiate each component one at a time:
() =
_
2 6
12
2
20
3
_

() =
_
0 6
24 60
2
_

(0) =
_
2 0
0 0
_
Integrate by integrating each component:
() =
_

2
+
1

3
+
2
4
+
3

5
+
4
_
2
0
() =
2
0

3
2
0
2
0

5
2
0
=
_
4 8
16 32
_
Proposition 3.5.6.
Suppose , are matrix-valued functions of a real variable, is a function of a real variable,
is a constant, and is a constant matrix then
1. ()
(product rule for matrices)

2. ()
3. ()
4. ()
5. ()
6. (+)
where each of the functions is evaluated at the same time and I assume that the functions
and matrices are dierentiable at that value of and of course the matrices , , are such
that the multiplications are well-dened.
Proof: Suppose ()

and ()

consider,
()
(()
) defn. derivative of matrix

=

) defn. of matrix multiplication

=
) linearity of derivative
=
ordinary product rules

=
algebra
= (
+ (
defn. of matrix multiplication

= (
defn. matrix addition

this proves (1.) as , were arbitrary in the calculation above. The proof of (2.) and (3.) follow
quickly from (1.) since constant means
= 0. Proof of (4.) is similar to (1.):

()
(()
) defn. derivative of matrix

=

) defn. of scalar multiplication

=

ordinary product rule

= (
defn. matrix addition

= (

The proof of (5.) follows from taking () = which has
= 0. I leave the proof of (6.) as an

exercise for the reader. .
To summarize: the calculus of matrices is the same as the calculus of functions with the small
qualier that we must respect the rules of matrix algebra. The noncommutativity of matrix mul-
tiplication is the main distinguishing feature.
3.5.4 calculus of complex-valued functions of a real variable
Dierentiation of functions from to is dened by splitting a given function into its real and
imaginary parts then we just dierentiate with respect to the real variable one component at a
time. For example:
(
2
cos() +
2
sin()) =

(
2
cos()) +

(
2
sin())
= (2
2
cos()
2
sin()) +(2
2
sin() +
2
cos()) (3.1)
=
2
(2 +)(cos() + sin())
= (2 +)
(2+)
where I have made use of the identity
7
+
=
(cos() + sin()). We just saw that

which seems obvious enough until you appreciate that we just proved it for = 2 +.
7
or denition, depending on how you choose to set-up the complex exponential, I take this as the denition in
calculus II
3.6 complex analysis in a nutshell
Dierentiation with respect to a real variable can be reduced to the slogan that we dierentiate
componentwise. Dierentiation with respect to a complex variable requires additional structure.
They key distinguishing ingredient is complex linearity:
Denition 3.6.1.
If we have some function : such that
(1.) ( +) = () +() for all , (2.) () = () for all ,
then we would say that is complex-linear.
Condition (1.) is additivity whereas condition (2.) is homogeneity. Note that complex linearity
implies real linearity however the converse is not true.
Example 3.6.2. Suppose () = where if = + for , then = is the complex
conjugate of . Consider for = + where , ,
() = (( +)( +))
= ( +( +))
= ( +)
= ( )( )
= ()
hence this map is not complex linear. On the other hand, if we study mutiplication by just ,
() = (( +)) = ( +) = = ( ) = ( +) = ()
thus is homogeneous with respect to real-number multiplication and it is also additive hence is
real linear.
Suppose that is a linear mapping from
2
to
2
. It is known from linear algebra that there exists
a matrix =
_

_
such that () = for all
2
. In this section we use the notation
+ = (, ) and
(, ) (, ) = ( +)( +) = +( +) = ( , +).
This construction is due to Gauss in the early nineteenth century, the idea is to use two component
vectors to construct complex numbers. There are other ways to construct complex numbers
8
.
Notice that ( + ) =
_

__
_
= ( + , + ) = + + ( + ) denes a real
linear mapping on for any choice of the real constants , , , . In contrast, complex linearity
puts strict conditions on these constants:
8
the same is true for real numbers, you can construct them in more than one way, however all constructions agree
on the basic properties and as such it is the properties of real or complex numbers which truly dened them. That
said, we choose Gauss representation for convenience.
3.6. COMPLEX ANALYSIS IN A NUTSHELL 97
Theorem 3.6.3.
The linear mapping () = is complex linear i the matrix will have the special form
below:
_

_
To be clear, we mean to identify
2
with as before. Thus the condition of complex
homogeneity reads ((, ) (, )) = (, ) (, )
Proof: assume is complex linear. Dene the matrix of as before:
(, ) =
_

__
_
This yields,
( +) = + +( +)
We can gain conditions on the matrix by examining the special points 1 = (1, 0) and = (0, 1)
(1, 0) = (, ) (0, 1) = (, )
Note that (
1
,
2
) (1, 0) = (
1
,
2
) hence ((
1
+
2
)1) = (
1
+
2
)(1) yields
(
1
+
2
) +(
1
+
2
) = (
1
+
2
)( +) =
1

2
+(
1
+
2
)
We nd two equations by equating the real and imaginary parts:
1
+
2
=
1

2

1
+
2
=
1
+
2
Therefore,
2
=
2
and
2
=
2
for all (
1
,
2
) . Suppose
1
= 0 and
2
= 1. We nd
= and = . We leave the converse proof to the reader. The proposition follows.
In analogy with the real case we dene
() as follows:
Denition 3.6.4.
Suppose : () and () then we dene
() by the limit below:
() = lim
0
( +) ()
.
The derivative function
is dened pointwise for all such () that the limit above

exists.
Note that
() = lim
0
()
hence
lim
0
()
= lim
0
( +) ()
lim
0
( +) ()
()
= 0
Note that the limit above simply says that () =
() gives the is the best complex-linear

approximation of = ( +) ().
Proposition 3.6.5.
If is a complex dierentiable at
then linearization () =
) is a complex linear
mapping.
Proof: let , and note () =
)() =
) = ().
It turns out that complex dierentiability automatically induces real dierentiability:
Proposition 3.6.6.
If is a complex dierentiable at
then is (real) dierentiable at
with () =
).
Proof: note that lim
0
(+)()
()
= 0 implies
lim
0
( +) ()
()
= 0
but then = and we know () =
) is real-linear hence is the best linear approxi-

mation to at
and the proposition follows.

Lets summarize what weve learned: if : () is complex dierentiable at
and
= + then,
1. () =
) is complex linear.
2. () =
) is the best real linear approximation to viewed as a mapping on

2
.
The Jacobian matrix for = (, ) has the form
) =
_

)
_
Theorem 3.6.3 applies to
) since is a complex linear mapping. Therefore we nd the Cauchy

Riemann equations:
and
. We have proved the following theorem:

Theorem 3.6.7.
If = + is a complex function which is complex-dierentiable at
then the partial

derivatives of and exist at
and satisfy the Cauchy-Riemann equations at
.
Example 3.6.8. Let () =
where the denition of the complex exponential function is given

by the following, for each , and = +
( +) =
+
=
(cos() + sin()) =
cos() +
sin()
Identify for = + we have (, ) =
cos() and (, ) =
sin(). Calculate:
cos()
cos() &

cos()
sin(),
sin()
sin() &

sin()
cos().
Thus satises the CR-equations

and

. The complex exponential function is

complex dierentiable.
The converse of Theorem 3.6.7 is not true in general. It is possible to have functions , :
2
that satisfy the CR-equations at
and yet = + fails to be complex dierentiable

at
.
Example 3.6.9. Counter-example to converse of Theorem 3.6.7. Suppose (+) =
_
0 if = 0
1 if = 0
.
Clearly is identically zero on the coordinate axes thus along the -axis we can calculate the partial
derivatives for and and they are both zero. Likewise, along the -axis we nd
and
exist and
are zero. At the origin we nd
all exist and are zero. Therefore, the Cauchy-Riemann

equations hold true at the origin. However, this function is not even continuous at the origin, thus
it is not real dierentiable!
The example above equally well serves as an example for a point where a function has partial
derivatives which exist at all orders and yet the dierential fails to exist. Its not a problem of
complex variables in my opinion, its a problem of advanced calculus. The key concept to reverse
the theorem is continuous dierentiability.
Theorem 3.6.10.
If , ,
are continuous functions in some open disk of
and
) =
)
and
) =
) then = + is complex dierentiable at
.
Proof: we are given that a function :
)
2

2
is continuous with continuous partial
derivatives of its component functions and . Therefore, by Theorem ?? we know is (real)
dierentiable at
. Therefore, we have a best linear approximation to the change in near
which can be induced via multiplication of the Jacobian matrix:

(
1
,
2
) =
_

)
_ _

1
2
_
.
Note then that the given CR-equations show the matrix of has the form
[] =
_

_
where =
) and =
). Consequently we nd is complex linear and it follows that

is complex dierentiable at
since we have a complex linear map such that

lim
0
( +) () ()
= 0
note that the limit with in the denominator is equivalent to the limit above which followed directly
from the (real) dierentiability at
. (the following is not needed for the proof of the theorem, but
perhaps it is interesting anyway) Moreover, we can write
(
1
,
2
) =
_

_ _

1
2
_
=
_

1
+
1
+
2
_
=
1
+
2
+(
1
+
2
)
= (
)(
1
+
2
)
Therefore we nd
) =
gives () =
).
In the preceding section we found necessary and sucient conditions for the component functions
, to construct an complex dierentiable function = + . The denition that follows is the
next logical step: we say a function is analytic
9
at
if it is complex dierentiable at each point in

some open disk about
.
Denition 3.6.11.
Let = + be a complex function. If there exists > 0 such that is complex
dierentiable for each
) then we say that is analytic at
. If is analytic for
each
then we say is analytic on . If is not analytic at
then we say that
is a singular point. Singular points may be outside the domain of the function. If is
analytic on the entire complex plane then we say is entire. Analytic functions are
also called holomorphic functions
If you look in my complex variables notes you can nd proof of the following theorem (well, partial
proof perhaps, but this result is shown in every good complex variables text)
Theorem 3.6.12.
If : is a function and

: is an extension of which is analytic then
is unique. In particular, if there is an analytic extension of sine, cosine, hyperbolic sine

or hyperbolic cosine then those extensions are unique.
This means if we demand analyticity then we actually had no freedom in our choice of the exponen-
tial. If we nd a complex function which matches the exponential function on a line-segment ( in
9
you may recall that a function on was analyic at if its Talyor series at converged to the function in some
neighborhood of . This terminology is consistent but I leave the details for your complex analysis course
particular a closed interval in viewed as a subset of is a line-segment ) then there is just one
complex function which agrees with the real exponential and is complex dierentiable everywhere.
() =
extends uniquely to

() =
()
(cos(()) + sin(())).
Note

( + 0) =
(cos(0) + sin(0)) =
thus

= . Naturally, analyiticity is a desireable

property for the complex-extension of known functions so this concept of analytic continuation is
very nice. Existence aside, we should rst construct sine, cosine etc... then we have to check they
are both analytic and also that they actually agree with the real sine or cosine etc... If a function
on has vertical asymptotes, points of discontinuity or points where it is not smooth then the
story is more complicated.
3.6.1 harmonic functions
Weve discussed in some depth how to determine if a given function = + is in fact analytic.
In this section we study another angle on the story. We learn that the component functions
, of an analytic function = + are harmonic conjugates and they satisfy the phyically
signicant Laplaces equation
2
= 0 where
2
=
2
/
2
+
2
/
2
. In addition well learn
that if we have one solution of Laplaces equation then we can consider it to be the of some
yet undetermined analytic function = + . The remaining function is then constructed
through some integration guided by the CR-equations. The construction is similar to the problem
of construction of a potential function for a given conservative force in calculus III.
Proposition 3.6.13.
If = + is analytic on some domain then and are solutions of Laplaces
equation
= 0 on .
Proof: since = + is analytic we know the CR-equations hold true;
and
.
Moreover, is continuously dierentiable so we may commute partial derivatives by a theorem
from multivariate calculus. Consider
= (
+ (
= (
+ (
= 0
Likewise,
= (
+ (
= (
+ (
= 0
Of course these relations hold for all points inside and the proposition follows.
Example 3.6.14. Note () =
2
is analytic with =
2
2
and = 2. We calculate,
= 2,
= 2
= 0
Note
= 0 so is also a solution to Laplaces equation.

Now lets see if we can reverse this idea.
Example 3.6.15. Let (, ) = +
1
notice that solves Laplaces equation. We seek to nd a
harmonic conjugate of . We need to nd such that,
= 1

= 0
Integrate these equations to deduce (, ) = +
2
for some constant
2
. We thus construct
an analytic function (, ) = +
1
+ ( +
2
) = + +
1
+
2
. This is just () = + for
=
1
+
2
.
Example 3.6.16. Suppose (, ) =
cos(). Note that
= whereas
= hence
= 0. We seek to nd such that
cos()

sin()
Integrating
cos() with respect to and
sin() with respect to yields (, ) =
sin(). We thus construct an analytic function (, ) =
cos() +
sin(). Of course we
should recognize the function we just constructed, its just the complex exponential () =
.
Notice we cannot just construct an analytic function from any given function of two variables. We
have to start with a solution to Laplaces equation. This condition is rather restrictive. There
is much more to say about harmonic functions, especially where applications are concerned. My
goal here was just to give another perspective on analytic functions. Geometrically one thing we
could see without further work at this point is that for an analytic function = + the families
of level curves (, ) =
1
and (, ) =
2
are orthogonal. Note () =<
> and
() =<
> have
() () =
= 0
This means the normal lines to the level curves for and are orthogonal. Hence the level curves
of and are orthogonal.
Chapter 4
inverse and implicit function theorems
It is tempting to give a complete and rigourous proof of these theorems here, but I will resist the
temptation in lecture. Im actually more interested that the student understand what the theorem
claims. I will sketch the proof and show many applications. A nearly complete proof is found in
Edwards where he uses an iterative approximation technique founded on the contraction mapping
principle. All his arguments are in some sense in vain unless you have some working knowledge
of uniform convergence. Its hidden in his proof, but we cannot conclude the limit of his sequence
of function has the properties we desire unless the sequence of functions is uniformly convergent.
Sadly that material has its home in real analysis and I dare not trespass in lecture. That said, if
you wish Id be happy to show you the full proof if you have about 20 extra hours to develop the
material outside of class. Alternatively, as a course of future study, return to the proof after you
have completed Math 431 here at Liberty
1
. Some other universities put advanced calculus after
the real analysis course so that more analytical depth can be built into the course
2
4.1 inverse function theorem
Consider the problem of nding a local inverse for : () . If we are given a point
() such that there exists an open interval containing with
a one-one function then

we can reasonably construct an inverse function by the simple rule
1
() = i () = for
and (). A sucient condition to insure the existence of a local inverse is that the
derivative function is either strictly positive or strictly negative on some neighborhood of . If we
are give a continuously dierentiable function at then it has a derivative which is continuous on
some neighborhood of . For such a function if
() = 0 then there exists some interval centered at

for which the derivative is strictly positive or negative. It follows that such a function is strictly
monotonic and is hence one-one thus there is a local inverse at . Therefore, even in calculus I we
nd the derivative informs us about the local invertibility of a function.
1
often read incorrectly as LU
2
If you think this would be worthwhile then by all means say as much in your exit interview, I believe we should
value the opinions of students, especially when they are geared towards academic excellence
103
104 CHAPTER 4. INVERSE AND IMPLICIT FUNCTION THEOREMS
The arguments I just made are supported by theorems that are developed in calculus I. Let me shift
gears a bit and give a direct calculational explaination based on the linearization approximation.
If then () () +
()( ). To nd the formula for the inverse we solve = () for

:
() +
()( ) +
1
()
_
()
Therefore,
1
() +
1
()
_
()
for near ().

Example 4.1.1. Just to help you believe me, consider () = 3 2 then
() = 3 for all .
Suppose we want to nd the inverse function near = 2 then the discussion preceding this example
suggests,
1
() = 2 +
1
3
( 4).
I invite the reader to check that (
1
()) = and
1
(()) = for all , .
In the example above we found a global inverse exactly, but this is thanks to the linearity of the
function in the example. Generally, inverting the linearization just gives the rst approximation to
the inverse.
Consider : ()
. If is dierentiable at
then we can write ()

() +
()( ) for . Set = () and solve for via matrix algebra. This time we need
to assume
() is an invertible matrix in order to isolate ,

() +
()( ) + (
())
1
_
()
Therefore,
1
() + (
())
1
_
()
for near (). Apparently the condition to nd a

local inverse for a mapping on
is that the derivative matrix is nonsingular

3
in some neighbor-
hood of the point. Experience has taught us from the one-dimensional case that we must insist the
derivative is continuous near the point in order to maintain the validity of the approximation.
Recall from calculus II that as we attempt to approximate a function with a power series it takes
an innite series of power functions to recapture the formula exactly. Well, something similar is
true here. However, the method of approximation is through an iterative approximation procedure
which is built o the idea of Newtons method. The product of this iteration is a nested sequence of
composite functions. To prove the theorem below one must actually provide proof the recursively
generated sequence of functions converges. See pages 160-187 of Edwards for an in-depth exposition
of the iterative approximation procedure. Then see pages 404-411 of Edwards for some material
on uniform convergence
4
The main analytical tool which is used to prove the convergence is called
the contraction mapping principle. The proof of the principle is relatively easy to follow and
3
nonsingular matrices are also called invertible matrices and a convenient test is that is invertible i () = 0.
4
actually that later chapter is part of why I chose Edwards text, he makes a point of proving things in
in such
a way that the proof naturally generalizes to function space. This is done by arguing with properties rather than
formulas. The properties oen extend to innite dimensions whereas the formulas usually do not.
4.1. INVERSE FUNCTION THEOREM 105
interestingly the main non-trivial step is an application of the geometric series. For the student
of analysis this is an important topic which you should spend considerable time really trying to
absorb as deeply as possible. The contraction mapping is at the base of a number of interesting
and nontrivial theorems. Read Rosenlichts Introduction to Analysis for a broader and better
organized exposition of this analysis. In contrast, Edwards uses analysis as a tool to obtain results
for advanced calculus but his central goal is not a broad or well-framed treatment of analysis.
Consequently, if analysis is your interest then you really need to read something else in parallel to
get a better ideas about sequences of functions and uniform convergence. I have some notes from
a series of conversations with a student about Rosenlicht, Ill post those for the interested student.
These notes focus on the part of the material I require for this course. This is Theorem 3.3 on page
185 of Edwards text:
Theorem 4.1.2. ( inverse function theorem )
Suppose :
is continuously dierentiable in an open set containing and the

derivative matrix
() is invertible. Then is locally invertible at . This means that

there exists an open set containing and a open set containing = () and
a one-one, continuously dierentiable mapping : such that (()) = for all
and (()) = for all . Moreover, the local inverse can be obtained as the
limit of the sequence of successive approximations dened by
() = and
+1
() =
() [
()]
1
[(
()) ] for all .

The qualier local is important to note. If we seek a global inverse then other ideas are needed.
If the function is everywhere injective then logically () = denes
1
() = and
1
so
constructed in single-valued by virtue of the injectivity of . However, for dierentiable mappings,
one might wonder how can the criteria of global injectivity be tested via the dierential. Even in
the one-dimensional case a vanishing derivative does not indicate a lack of injectivity; () =
3
has
1
() =
3
and yet
(0) = 0 (therefore
(0) is not invertible). One the other hand, well see

in the examples that follow that even if the derivative is invertible over a set it is possible for the
values of the mapping to double-up and once that happens we cannot nd a single-valued inverse
function
5
Remark 4.1.3. James R. Munkres Analysis on Manifolds good for a dierent proof.
Another good place to read the inverse function theorem is in James R. Munkres Analysis
on Manifolds. That text is careful and has rather complete arguments which are not entirely
the same as the ones given in Edwards. Munkres text does not use the contraction mapping
principle, instead the arguments are more topological in nature.
5
there are scientists and engineers who work with multiply-valued functions with great success, however, as a point
of style if nothing else, we try to use functions in math.
To give some idea of what I mean by topological let be give an example of such an argument.
Suppose :
is continuously dierentiable and
() is invertible. Heres a sketch of the

argument that
() is invertible for all near as follows:

1. the function :
dened by () = (
()) is formed by a multinomial in the

component functions of
(). This function is clearly continuous since we are given that the
partial derivatives of the component functions of are all continuous.
2. note we are given
() is invertible and hence (
()) = 0 thus the continuous function

is nonzero at . It follows there is some open set containing for which 0 / ()
3. we have (
()) = 0 for all hence
() is invertible on .
I would argue this is a topological argument because the key idea here is the continuity of .
Topology is the study of continuity in general.
Remark 4.1.4. James J. Callahans Advanced Calculus: a Geometric View, good reading.
James J. Callahan recently authored Advanced Calculus: a Geometric View. This text has
great merit in both visualization and well-thought use of linear algebraic techniques. In
addition, many student will enjoy his staggered proofs where he rst shows the proof for a
simple low dimensional case and then proceeds to the general case. I almost used his text
this semester.
_
sin() +1, sin() +2
_
for (, )
2
. Clearly is contin-
uously dierentiable as all its component functions have continuous partial derivatives. Observe,
(, ) = [
] =
_
0 cos()
cos() 0
_
Hence
(, ) is invertible at points (, ) such that (
(, )) = cos() cos() = 0. This

means we may not be able to nd local inverses at points (, ) with =
1
2
(2 + 1) or =
1
2
(2 + 1) for some , . Points where
(, ) are singular are points where one or both

of sin() and sin() reach extreme values thus the points where the Jacobian matrix are singular
are in fact points where we cannot nd a local inverse. Why? Because the function is clearly not
1-1 on any set which contains the points of singularity for . Continuing, recall from precalculus
that sine has a standard inverse on [/2, /2]. Suppose (, ) [/2, /2]
2
and seek to solve
(, ) = (, ) for (, ):
(, ) =
_
sin() + 1
sin() + 2
_
=
_

_

_
sin() + 1 =
sin() + 2 =
_

_
= sin
1
( 1)
= sin
1
( 2)
_
It follows that
1
(, ) =
_
sin
1
( 2), sin
1
( 1)
_
for (, ) [0, 2] [1, 3] where you should
note ([/2, /2]
2
) = [0, 2] [1, 3]. Weve found a local inverse for on the region [/2, /2]
2
.
In other words, we just found a global inverse for the restriction of to [/2, /2]
2
. Technically
we ought not write
1
, to be more precise we should write:
(
[/2,/2]
2)
1
(, ) =
_
sin
1
( 2), sin
1
( 1)
_
.
4.1. INVERSE FUNCTION THEOREM 107
It is customary to avoid such detail in many contexts. Inverse functions for sine, cosine, tangent
etc... are good examples of this slight of langauge.
A coordinate system on
is an invertible mapping of
to
. However, in practice the term

coordinate system is used with less rigor. Often a coordinate system has various degeneracies. For
example, in polar coordinates you could say = /4 or = 9/4 or generally = 2 + /4 for
any . Lets examine polar coordinates in view of the inverse function theorem.
Example 4.1.6. Let (, ) =
_
cos(), sin()
_
for (, ) [0, ) (/2, /2). Clearly
is continuously dierentiable as all its component functions have continuous partial derivatives.
To nd the inverse we seek to solve (, ) = (, ) for (, ). Hence, consider = cos() and
= sin(). Note that
2
+
2
=
2
cos
2
() +
2
sin
2
() =
2
(cos
2
() + sin
2
()) =
2
and
=
sin()
cos()
= tan().
It follows that =
2
+
2
and = tan
1
(/) for (, ) (0, ) . We nd
1
(, ) =
_

2
+
2
, tan
1
(/)
_
.
Lets see how the derivative ts with our results. Calcuate,
(, ) = [
] =
_
cos() sin()
sin() cos()
_
note that (
(, )) = hence we the inverse function theorem provides the existence of a local

inverse around any point except the origin. Notice the derivative does not detect the defect in the
angular coordinate. Challenge, nd the inverse function for (, ) =
_
cos(), sin()
_
with
() = [0, ) (/2, 3/2). Or, nd the inverse for polar coordinates in a neighborhood of
(0, 1).
3
3
is dened by (, , ) = (, , ) for constants , ,
where = 0. Clearly is continuously dierentiable as all its component functions have
continuous partial derivatives. We calculate
(, , ) = [
] = [
1
3
]. Thus
(
(, , )) = = 0 for all (, , )
3
hence this function is locally invertible everywhere.
Moreover, we calculate the inverse mapping by solving (, , ) = (, , ) for (, , ):
(, , ) = (, , ) (, , ) = (/, /, /)
1
(, , ) = (/, /, /).
is dened by () = + for some matrix

and
vector
. Under what conditions is such a function invertible ?. Since the formula for
this function gives each component function as a polynomial in the -variables we can conclude the
function is continuously dierentiable. You can calculate that
() = . It follows that a sucient

condition for local inversion is () = 0. It turns out that this is also a necessary condition as
() = 0 implies the matrix has nontrivial solutions for = 0. We say () i
= 0. Note if () then () = + = . This is not a problem when () = 0
for in that case the null space is contains just zero; () = {0}. However, when () = 0 we
learn in linear algebra that () contains innitely many vectors so is far from injective. For
example, suppose () = {
1
} then you can show that (
1
,
2
, . . . ,
) = (,
2
, . . . ,
)
for all . Hence any point will have other points nearby which output the same value under .
Suppose () = 0, to calculate the inverse mapping formula we should solve () = for ,
= + =
1
( )
1
() =
1
( ).
Remark 4.1.9. inverse function theorem holds for higher derivatives.
In Munkres the inverse function theorem is given for -times dierentiable functions. In
short, a
function with invertible dierential at a point has a
inverse function local

to the point. Edwards also has arguments for > 1, see page 202 and arguments and
surrounding arguments.
4.2 implicit function theorem
Consider the problem of solving
2
+
2
= 1 for as a function of .
2
+
2
= 1
2
= 1
2
=
1
2
.
A function cannot have two outputs for a single input, when we write in the expression above
it simply indicates our ignorance as to which is chosen. Once further information is given then we
may be able to choose a + or a . For example:
1. if
2
+
2
= 1 and we want to solve for near (0, 1) then =
1
2
is the correct choice
since > 0 at the point of interest.
2. if
2
+
2
= 1 and we want to solve for near (0, 1) then =
1
2
is the correct choice
since < 0 at the point of interest.
3. if
2
+
2
= 1 and we want to solve for near (1, 0) then its impossible to nd a single
function which reproduces
2
+
2
= 1 on an open disk centered at (1, 0).
What is the defect of case (3.) ? The trouble is that no matter how close we zoom in to the point
there are always two -values for each given -value. Geometrically, this suggests either we have a
discontinuity, a kink, or a vertical tangent in the graph. The given problem has a vertical tangent
and hopefully you can picture this with ease since its just the unit-circle. In calculus I we studied
4.2. IMPLICIT FUNCTION THEOREM 109
implicit dierentiation, our starting point was to assume = () and then we dierentiated
equations to work out implicit formulas for /. Take the unit-circle and dierentiate both sides,
2
+
2
= 1 2 + 2
= 0

.
Note

is not dened for = 0. Its no accident that those two points (1, 0) and (1, 0) are
precisely the points at which we cannot solve for as a function of . Apparently, the singularity
in the derivative indicates where we may have trouble solving an equation for one variable as a
function of the remaining variable.
We wish to study this problem in general. Given -equations in (+)-unknowns when can we solve
for the last -variables as functions of the rst -variables. Given a continuously dierentiable
mapping = (
1
,
2
, . . . ,
) :
study the level set: (here

1
,
2
, . . . ,
are
constants)
1
(
1
, . . . ,
,
1
, . . . ,
) =
1
2
(
1
, . . . ,
,
1
, . . . ,
) =
2
.
.
.
(
1
, . . . ,
,
1
, . . . ,
) =
We wish to locally solve for

1
, . . . ,
as functions of
1
, . . .
. That is, nd a mapping :
such that (, ) = i = () near some point (, )
such that (, ) = . In
this section we use the notation = (
1
,
2
, . . .
) and = (
1
,
2
, . . . ,
).
Before we turn to the general problem lets analyze the unit-circle problem in this notation. We
are given (, ) =
2
+
2
and we wish to nd () such that = () solves (, ) = 1.
Dierentiate with respect to and use the chain-rule:
= 0
We nd that / =
= /. Given this analysis we should suspect that if we are

given some level curve (, ) = then we may be able to solve for as a function of near
if () = and
() = 0. This suspicion is valid and it is one of the many consequences of the

implicit function theorem.
We again turn to the linearization approximation. Suppose (, ) = where
and
and suppose :
is continuously dierentiable. Suppose (, )
has
(, ) = . Replace with its linearization based at (, ):
(, ) +
(, )( , )
here we have the matrix multiplication of the ( + ) matrix
(, ) with the ( + ) 1
column vector ( , ) to yield an -component column vector. It is convenient to dene
partial derivatives with respect to a whole vector of variables,
1

1
.
.
.
.
.
.
1

1
.
.
.
.
.
.
In this notation we can write the ( + ) matrix
(, ) as the concatenation of the

matrix

(, ) and the matrix

(, )
(, ) =
_
(, )
(, )
_
With this notation we have
(, ) +

(, )( ) +

(, )( )
If we are near (, ) then (, ) thus we are faced with the problem of solving the following
equation for :
+

(, )( ) +

(, )( )
Suppose the square matrix

(, ) is invertible at (, ) then we nd the following approximation

for the implicit solution of (, ) = 0 for as a function of :
=
_
(, )
_
1
_
(, )( )
_
.
Of course this is not a formal proof, but it does suggest that
_
(, )
= 0 is a necessary
condition for solving for the variables.
As before suppose :
. Suppose we have a continuously dierentiable function

:
such that () = and (, ()) = . We seek to nd the derivative of in terms

of the derivative of . This is a generalization of the implicit dierentiation calculation we perform
in calculus I. Im including this to help you understand the notation a bit more before I state the
implicit function theorem. Dierentiate with respect to
for
_
(, ())
_
=
=1
=1
=1
= 0
we made use of the identity

to squash the sum of to the single nontrivial term and the

zero on the r.h.s follows from the fact that

() = 0. Concatenate these derivatives from = 1

up to = :
_
1
+
=1
2
+
=1
=1
_
= [00 0]
Properties of matrix addition allow us to parse the expression above as follows:
_
_
+
_

=1
=1
=1
_
= [00 0]
But, this reduces to
+
_
_
= 0

The concatenation property of matrix multiplication states [
1
] = [
1
]
we use this to write the expression once more,
_
= 0

= 0

where in the last implication we made use of the assumption that

is invertible.
Theorem 4.2.1. (Theorem 3.4 in Edwardss Text see pg 190)
Let : ()
be continuously dierentiable in a open ball about the

point (, ) where (, ) = (a constant vector in
). If the matrix

(, ) is invertible
then there exists an open ball containing in
and an open ball containing (, )

in
and a continuously dierentiable mapping :
such that (, ) =
i = () for all (, ) . Moreover, the mapping is the limit of the sequence of
successive approximations dened inductively below
() = ,
+1
=
() [
(, )]
1
(,
()) for all .

We will not attempt a proof of the last sentence for the same reasons we did not pursue the details
in the inverse function theorem. However, we have already derived the rst step in the iteration in
our study of the linearization solution.
Proof: Let : ()
be continuously dierentiable in a open ball about

the point (, ) where (, ) = (
a constant). Furthermore, assume the matrix

(, )
is invertible. We seek to use the inverse function theorem to prove the implicit function theorem.
Towards that end consider :
dened by (, ) = (, (, )). To begin,

observe that is continuously dierentiable in the open ball which is centered at (, ) since
and have continuous partials of their components in . Next, calculate the derivative of
= (, ),
(, ) = [
] =
_

_
=
_

_
The determinant of the matrix above is the product of the deteminant of the blocks
and
; (
(, ) = (
)(
) =
. We are given that

(, ) is invertible and hence

(
(, )) = 0 thus (
(, ) = 0 and we nd
(, ) is invertible. Consequently, the inverse

function theorem applies to the function at (, ). Therefore, there exists
1
:
such that
1
is continuously dierentiable. Note (, ) and contains the
point (, ) = (, (, )) = (, ).
Our goal is to nd the implicit solution of (, ) = . We know that
1
((, )) = (, ) and (
1
(, )) = (, )
for all (, ) and (, ) . As usual to nd the formula for the inverse we can solve
(, ) = (, ) for (, ) this means we wish to solve (, (, )) = (, ) hence = . The
formula for is more elusive, but we know it exists by the inverse function theorem. Lets say
= (, ) where :
and thus
1
(, ) = (, (, )). Consider then,
(, ) = (
1
(, ) = (, (, )) = (, (, (, ))
Let = thus (, ) = (, (, (, )) for all (, ) . Finally, dene () = (, ) for
all (, ) and note that = (, ()). In particular, (, ) and at that point we nd
() = (, ) = by construction. It follows that = () provides a continuously dierentiable
solution of (, ) = near (, ).
Uniqueness of the solution follows from the uniqueness for the limit of the sequence of functions
described in Edwards text on page 192. However, other arguments for uniqueness can be oered,
independent of the iterative method, for instance: see page 75 of Munkres Analysis on Manifolds.
Remark 4.2.2. notation and the implementation of the implicit function theorem.
We assumed the variables were to be written as functions of variables to make explicit
a local solution to the equation (, ) = . This ordering of the variables is convenient to
argue the proof, however the real theorem is far more general. We can select any subset of
input variables to make up the so long as

is invertible. I will use this generalization

of the formal theorem in the applications that follow. Moreover, the notations and are
unlikely to maintain the same interpretation as in the previous pages. Finally, we will for
convenience make use of the notation = () to express the existence of a function such
that = () when appropriate. Also, = (, ) means there is some function for which
= (, ). If this notation confuses then invent names for the functions in your problem.
Example 4.2.3. Suppose (, , ) =
2
+
2
+
2
. Suppose we are given a point (, , ) such
that (, , ) =
2
for a constant . Problem: For which variable can we solve? What, if
any, inuence does the given point have on our answer? Solution: to begin, we have one
equation and three unknowns so we should expect to nd one of the variables as functions of the
remaining two variables. The implicit function theorem applies as is continuously dierentiable.
1. if we wish to solve = (, ) then we need
(, , ) = 2 = 0.
(, , ) = 2 = 0.
(, , ) = 2 = 0.
The point has no local solution for if it is a point on the intersection of the -plane and the
sphere (, , ) =
2
. Likewise, we cannot solve for = (, ) on the = 0 slice of the sphere
and we cannot solve for = (, ) on the = 0 slice of the sphere.
Notice, algebra veries the conclusions we reached via the implicit function theorem:
=
2
=
2
=
2
When we are at zero for one of the coordinates then we cannot choose + or since we need both on
an open ball intersected with the sphere centered at such a point
6
. Remember, when I talk about
local solutions I mean solutions which exist over the intersection of the solution set and an open
ball in the ambient space (
3
in this context). The preceding example is the natural extension of
the unit-circle example to
3
. A similar result is available for the -sphere in
. I hope you get

the point of the example, if we have one equation then if we wish to solve for a particular variable in
terms of the remaining variables then all we need is continuous dierentiability of the level function
and a nonzero partial derivative at the point where we wish to nd the solution. Now, the implicit
function theorem doesnt nd the solution for us, but it does provide the existence. In the section
that follows, existence is really all we need since focus our attention on rates of change rather than
actually solutions to the level set equation.
Example 4.2.4. Consider the equation
+
3
= 2. Can we solve this equation for
= (, ) near (0, 0, 1)? Let (, , ) =
+
3
and note (0, 0, 1) =
0
+1+0 = 2 hence
(0, 0, 1) is a point on the solution set (, , ) = 2. Note is clearly continuously dierentiable
and
(, , ) = 3
2

(0, 0, 1) = 3 = 0
therefore, there exists a continuously dierentiable function : ()
2
which solves
(, , (, )) = 2 for (, ) near (0, 0) and (0, 0) = 1.
Ill not attempt an explicit solution for the last example.
Example 4.2.5. Let (, , ) i + + = 2 and + = 1. Problem: For which
variable(s) can we solve? Solution: dene (, , ) = ( + + , + ) we wish to study
(, , ) = (2, 1). Notice the solution set is not empty since (1, 0, 1) = (1 +0 +1, 0 +1) = (2, 1)
Moreover, is continuously dierentiable. In this case we have two equations and three unknowns
so we expect two variables can be written in terms of the remaining free variable. Lets examine
the derivative of :
(, , ) =
_
1 1 1
0 1 1
_
6
if you consider (, , ) =
2
as a space then the open sets on the space are taken to be the intersection with
the space and open balls in
3
. This is called the subspace topology in topology courses.
Suppose we wish to solve = () and = () then we should check invertiblility of
(, )
=
_
1 1
0 1
_
.
The matrix above is invertible hence the implicit function theorem applies and we can solve for
and as functions of . On the other hand, if we tried to solve for = () and = () then
well get no help from the implicit function theorem as the matrix
(, )
=
_
1 1
1 1
_
.
is not invertible. Geometrically, we can understand these results from noting that (, , ) = (2, 1)
is the intersection of the plane + + = 2 and + = 1. Subsituting + = 1 into + + = 2
yields +1 = 2 hence = 1 on the line of intersection. We can hardly use as a free variable for
the solution when the problem xes from the outset.
The method I just used to analyze the equations in the preceding example was a bit adhoc. In
linear algebra we do much better for systems of linear equations. A procedure called Gaussian
elimination naturally reduces a system of equations to a form in which it is manifestly obvious how
to eliminate redundant variables in terms of a minimal set of basic free variables. The of the
implicit function proof discussions plays the role of the so-called pivotal variables whereas the
plays the role of the remaining free variables. These variables are generally intermingled in
the list of total variables so to reproduce the pattern assumed for the implicit function theorem we
would need to relable variables from the outset of a calculation. The calculations in the examples
that follow are not usually possible. Linear equations are particularly nice and basically what Im
doing is following the guide of the linearization derivation in the context of specic examples.
Example 4.2.6. XXX
Example 4.2.7. XXX
Example 4.2.8. XXX
4.3. IMPLICIT DIFFERENTIATION 115
4.3 implicit dierentiation
Enough theory, lets calculate. In this section I apply previous theoretical constructions to specic
problems. I also introduce standard notation for constrained partial dierentiation which is
also sometimes called partial dierentiation with a side condition. The typical problem is the
following: given equations:
1
(
1
, . . . ,
,
1
, . . . ,
) =
1
2
(
1
, . . . ,
,
1
, . . . ,
) =
2
.
.
.
(
1
, . . . ,
,
1
, . . . ,
) =
calculate partial derivative of dependent variables with respect to independent variables. Contin-
uing with the notation of the implicit function discussion well assume that will be dependent
on . I want to recast some of our arguments via dierentials
7
. Take the total dierential of each
equation above,
1
(
1
, . . . ,
,
1
, . . . ,
) = 0
2
(
1
, . . . ,
,
1
, . . . ,
) = 0
.
.
.
(
1
, . . . ,
,
1
, . . . ,
) = 0
Hence,
1
+ +
1
+ +
= 0
1
+ +
1
+ +
= 0
.
.
.
1
+ +
1
+ +
= 0
Notice, this can be nicely written in column vector notation as:
1
+ +
1
+ +
= 0
Or, in matrix notation:
[
1
.
.
.
+ [
1
.
.
.
= 0
7
in contrast, In the previous section we mostly used derivative notation
Finally, solve for , we assume [
]
1
exists,
1
.
.
.
= [
]
1
[
1
.
.
.
Given all of this we can calculate

by simply reading the coeent
in the -th row. I will

make this idea quite explicit in the examples that follow.
Example 4.3.1. Lets return to a common calculus III problem. Suppose (, , ) = for some
constant . Find partial derivatives of , or with repsect to the remaining variables.
Solution: Ill use the method of dierentials once more:
=
= 0
We can solve for , or provided
or
is nonzero respective and these dierential

expressions reveal various partial derivatives of interest:
=
&

&

&

In each case above, the implicit function theorem allows us to solve for one variable in terms of the
remaining two. If the partial derivative of in the denominator are zero then the implicit function
theorem does not apply and other thoughts are required. Often calculus text give the following as a
homework problem:
= 1.
In the equation above we have appear as a dependent variable on , and also as an independent
variable for the dependent variable . These mixed expressions are actually of interest to engineering
and physics. The less mbiguous notation below helps better handle such expressions:
_
= 1.
In each part of the expression we have clearly denoted which variables are taken to depend on the
others and in turn what sort of partial derivative we mean to indicate. Partial derivatives are not
taken alone, they must be done in concert with an understanding of the totality of the indpendent
variables for the problem. We hold all the remaining indpendent variables xed as we take a partial
derivative.
4.3. IMPLICIT DIFFERENTIATION 117
The explicit independent variable notation is more important for problems where we can choose
more than one set of indpendent variables for a given dependent variables. In the example that
follows we study = (, ) but we could just as well consider = (, ). Generally it will not
be the case that
_
is the same as
_
. In calculation of
_
we hold constant as we
vary whereas in
_
we hold constant as we vary . There is no reason these ought to be the

same
8
.
Example 4.3.2. Suppose +++ = 3 and
2
2+
3
= 5. Calculate partial derivatives
of and with respect to the independent variables , . Solution: we begin by calculation
of the dierentials of both equations:
+ + + = 0
(2 2) 2 2 + 3
2
= 0
We can solve for (, ). In this calculation we can treat the dierentials as formal variables.
+ =
2 + 3
2
= (2 2) + 2
I nd matrix notation is often helpful,
_
1 1
2 3
2
_ _

_
=
_

(2 2) + 2
_
Use Kramers rule, multiplication by inverse, substitution, adding/subtracting equations etc... what-
ever technique of solving linear equations you prefer. Our goal is to solve for and in terms
of and . Ill use Kramers rule this time:
=
_
1
(2 2) + 2 3
2
_
_
1 1
2 3
2
_ =
3
2
( ) + (2 2) 2
3
2
+ 2
Collecting terms,
=
_
3
2
+ 2 2
3
2
+ 2
_
+
_
3
2
2
3
2
+ 2
_
From the expression above we can read various implicit derivatives,

_
=
3
2
+ 2 2
3
2
+ 2
&
_
=
3
2
2
3
2
+ 2
The notation above indicates that is understood to be a function of independent variables , .
_
means we take the derivative of with respect to while holding xed. The appearance
8
a good exercise would be to do the example over but instead aim to calculate partial derivatives for , with
respect to independent variables ,
of the dependent variable can be removed by using the equations (, , , ) = (3, 5). Similar
ambiguities exist for implicit dierentiation in calculus I. Apply Kramers rule once more to solve
for :
=
_
1
2 (2 2) + 2
_
_
1 1
2 3
2
_ =
(2 2) + 2 2( +)
3
2
+ 2
Collecting terms,
=
_
2 + 2 2
3
2
+ 2
_
+
_
2 2
3
2
+ 2
_
We can read the following from the dierential above:

_
=
2 + 2 2
3
2
+ 2
&
_
=
2 2
3
2
+ 2
You should ask: where did we use the implicit function theorem in the preceding example? Notice
our underlying hope is that we can solve for = (, ) and = (, ). The implicit function
theorem states this is possible precisely when

(,)
=
_
1 1
2 3
2
_
is non singular. Interestingly
this is the same matrix we must consider to isolate and . The calculations of the example
are only meaningful if the
_
1 1
2 3
2
_
= 0. In such a case the implicit function theorem
applies and it is reasonable to suppose , can be written as functions of , .
Example 4.3.3. Suppose the temperature in a room is given by (, , ) = 70+10(
2
2
).
Find how the temperature varies on a sphere
2
+
2
+
2
=
2
. We can choose any one
variable from (, , ) and write it as a function of the remaining two on the sphere. However, we
do need to a
Chapter 5
geometry of level sets
Our goal in this chapter is to develop a few tools to analyze the geometry of solution sets to equa-
tion(s) in
. These solution sets are commonly called level sets. I assume the reader is already
familiar with the concept of level curves and surfaces from multivariate calculus. We go much fur-
ther in this chapter. Our goal is to describe the tangent and normal spaces for a -dimensional level
set in
. The dimension of the level set is revealed by its tangent space and we discuss conditions
which are sucient to insure the invariance of this dimension over the entirety of the level set. In
contrast, the dimension of the normal space to a -dimensional level set in
is . The theory
of orthogonal complements is borrowed from linear algebra to help understand how all of this ts
together at a given point on the level set. Finally, we use this geometry and a few simple lemmas
to justify the method of Lagrange multipliers. Lagranges technique and the theory of multivariate
Taylor polynomials form the basis for analyzing extrema for multivariate functions. In short, this
chapter deals with the question of extrema on the edges of a set whereas the next chapter deals
with the interior point via the theory of quadratic forms applied to the second-order approximation
to a function of several variables. Finally, we should mention that -dimensional level sets provide
examples of -dimensional manifolds, however, we defer careful discussion of manifolds for a later
chapter.
5.1 denition of level set
A level set is the solution set of some equation or system of equations. We conne our interest to
level sets of
. For example, the set of all (, ) that satisfy

(, ) =
is called a level curve in
2
. Often we can use to label the curve. You should also recall level
surfaces in
3
are dened by an equation of the form
(, , ) = .
119
120 CHAPTER 5. GEOMETRY OF LEVEL SETS
The set of all (
1
,
2
,
3
,
4
)
4
which solve (
1
,
2
,
3
,
4
) = is a level volume in
4
. We
can obtain lower dimensional objects by simultaneously imposing several equations at once. For
example, suppose
1
(, , ) = = 1 and
2
(, , ) =
2
+
2
+
2
= 5, points (, , ) which solve
both of these equations are on the intersection of the plane = 1 and the sphere
2
+
2
+
2
= 5.
Let = (
1
,
2
), note that (, , ) = (1, 5) describes a circle in
3
. More generally:
Denition 5.1.1.
Suppose : ()
. Let be a vector of constants in
and suppose
= {
() = } is non-empty and is continuously dierentiable on an open

set containing . We say is an ( )-dimensional level set i
() has linearly
independent rows at each .
The condition of linear independence of the rows is give to eliminate possible redundancy in the
system of equations. In the case that = 1 the criteria reduces to the conditon level function has
() = 0 over the level set of dimension 1. Intuitively we think of each equation in () =

as removing one of the dimensions of the ambient space
. It is worthwhile to cite a useful result

from linear algebra at this point:
Proposition 5.1.2.
Let

. The number of linearly independent columns in is the same as the
number of linearly independent rows in . This invariant of is called the rank of .
Given the wisdom of linear algebra we see that we should require a ( )-dimensional level set
=
1
() to have a level function :
whose derivative is of rank over all of . We

can either analyze linear independence of columns or rows.
Example 5.1.3. Consider (, , ) =
2
+
2
2
and suppose =
1
{0}. Calculate,
(, , ) = [2, 2, 2]
Notice that (0, 0, 0) and
(0, 0, 0) = [0, 0, 0] hence is not rank one at the origin. At all

other points in we have
(, , ) = 0 which means this is almost a 3 1 = 2-dimensional

level set. However, almost is not good enough in math. Under our denition the cone is not a
2-dimensional level set since it fails to meet the full-rank criteria at the point of the cone.
Example 5.1.4. Let (, , ) = (, ) and dene =
1
(, ) for some xed pair of constants
, . We calculate that
(, , ) =
2

22
. We clearly have rank two at all points in
hence is a 3 2 = 1-dimensional level set. Perhaps you realize is the vertical line which passes
through (, , 0) in the -plane.
5.2. TANGENTS AND NORMALS TO A LEVEL SET 121
5.2 tangents and normals to a level set
There are many ways to dene a tangent space for some subset of
. One natural denition is

that the tangent space to is simply the set of all tangent vectors to curves on which pass
through the point . In this section we study the geometry of curves on a level-set. Well see how
the tangent space is naturally a vector space in the particular context of level-sets in
.
Throughout this section we assume that is a -dimensional level set dened by :
where
1
() = . This means that we can apply the implicit function theorem to and for
any given point = (
) where
and
. There exists a local continuously

dierentiable solution :
such that (
) =
and for all we have (, ()) =

. We can view (, ) = for near as the graph of = () for . With the set-up above
in mind, suppose that : . If we write = (
) then it follows = (
)
over the subset () of . More explicitly, for all such that () () we have
() = (
(), (
())). Therefore, if (0) = then (0) = (
, (
)). Dierentiate, use the

chain-rule in the second factor to obtain:
() = (
(),
())
()). We nd that the tangent

vector to of has a rather special form which was forced on us by the implicit function
theorem:
(0) = (
(0),
(0)). Or to cut through the notation a bit, if
(0) = = (
)
then = (
). The second component of the vector is not free of the rst, it essentially
redundant. This makes us suspect that the tangent space to at is -dimensional.
Theorem 5.2.1.
Let :
be a level-mappping which denes a -dimensional level set

by
1
() = . Suppose
1
,
2
: are dierentiable curves with
1
(0) =
1
and
2
(0) =
2
then there exists a dierentiable curve : such that
(0) =
1
+
2
and
(0) = . Moreover, there exists a dierentiable curve : such that
(0) =
1
and
(0) = .
Proof: It is convenient to dene a map which gives a local parametrization of at . Since
we have a description of locally as a graph = () (near ) it is simple to construct the
parameterization. Dene :
by () = (, ()). Clearly () = () and

there is an inverse mapping
1
(, ) = is well-dened since = () for each (, ) ().
Let
and observe that

() = (
1
() +) = (
+) = (
+, (
+))
is a curve from to such that (0) = (
, (
)) = (
) = and using the chain rule on

the nal form of ():
(0) = (,
)).
The construction above shows that any vector of the form (
) is the tangent vector of a

particular dierentiable curve in the level set (dierentiability of follows from the dierentiability
of and the other maps which we used to construct ). In particular we can apply this to the
case =
1
+
2
and we nd () = (
1
() + (
1
+
2
)) has
(0) =
1
+
2
and (0) = .
Likewise, apply the construction to the case =
1
to write () = (
1
() + (
1
)) with
(0) =
1
and (0) = .
The idea of the proof is encapsulated in the picture below. This idea of mapping lines in a at
domain to obtain standard curves in a curved domain is an idea which plays over and over as you
study manifold theory. The particular redundancy of the and sub-vectors is special to the
discussion level-sets, however anytime we have a local parametrization well be able to construct
curves with tangents of our choosing by essentially the same construction. In fact, there are in-
nitely many curves which produce a particular tangent vector in the tangent space of a manifold.
XXX - read this section again for improper, premature use of the term manifold
Theorem 5.2.1 shows that the denition given below is logical. In particular, it is not at all obvious
that the sum of two tangent vectors ought to again be a tangent vector. However, that is just what
the Theorem 5.2.1 told us for level-sets
1
.
Denition 5.2.2.
Suppose is a -dimensional level-set dened by =
1
{} for :
. We
dene the tangent space at to be the set of pairs:
= {(, ) there exists dierentiable : and (0) = where =
(0)}
Moreover, we dene (i.) addition and (ii.) scalar multiplication of vectors by the rules
(.) (,
1
) + (,
2
) = (,
1
+
2
) (.) (,
1
) = (,
1
)
for all (,
1
), (,
2
)
and .
When I picture
in my mind I think of vectors pointing out from the base-point . To make

an explicit connection between the pairs of the above denition and the classical geometric form
of the tangent space we simply take the image of
under the mapping (, ) = + thus

(
) = { + (, )
}. I often picture
as (
)
2
1
technically, there is another logical gap which I currently ignore. I wonder if you can nd it.
2
In truth, as you continue to study manifold theory youll nd at least three seemingly distinct objects which are
all called tangent vectors; equivalence classes of curves, derivations, contravariant tensors.
We could set out to calculate tangent spaces in view of the denition above, but we are actually
interested in more than just the tangent space for a level-set. In particular. we want a concrete
description of all the vectors which are not in the tangent space.
Denition 5.2.3.
Suppose is a -dimensional level-set dened by =
1
{} for :
and
is the tangent space at . Note that
where
= {}
is given the
natural vector space structure which we already exhibited on the subspace
. We dene
the inner product on
as follows: for all (, ), (, )
,
(, ) (, ) = .
The length of a vector (, ) is naturally dened by (, ) = . Moreover, we say two
vectors (, ), (, )
are orthogonal i = 0. Given a set of vectors
we
dene the orthogonal complement by
= {(, )
(, ) (, ) for all (, ) }.
Suppose
1
,
2

then we say
1
is orthogonal to
2
i
1

2
= 0 for all
1

1
and
2

2
. We denote orthogonality by writing
1

2
. If every
can be written
as =
1
+
2
for a pair of
1

1
and
2

2
where
1

2
then we say that
is
the direct sum of
1
and
2
which is denoted by
=
1
2
.
There is much more to say about orthogonality, however, our focus is not in that vein. We just
need the langauge to properly dene the normal space. The calculation below is probably the most
important calculation to understand for a level-set. Suppose we have a curve : where
=
1
() is a -dimensional level-set in
. Observe that for all ,

(()) =
(())
() = 0.
In particular, suppose for = 0 we have (0) = and =
(0) which makes (, )
with
() = 0.
Recall :
has an derivative matrix where the -th row is the gradient vector
of the -th component function. The equation
() = 0 gives us -independent equations as

we examine it componentwise. In particular, it reveals that (, ) is orthogonal to
() for
= 1, 2, . . . , . We have derived the following theorem:
Theorem 5.2.4.
Let :
be a level-mappping which denes a -dimensional level set by
1
() = . The gradient vectors
() are perpendicular to the tangent space at ; for

each
(, (
())
) (
.
Its time to do some counting. Observe that the mapping :
dened by () = (, )
is an isomorphism of vector spaces hence (
) = . But, by the same isomorphism we can

see that
= (
) hence (
) = + . In linear algebra we learn that if we have a

-dimensional subspace of an -dimensional vector space then the orthogonal complement
is a subspace of with codimension . The term codimension is used to indicate a loss

of dimension from the ambient space, in particular (
) = . We should note that the

direct sum of and
covers the whole space;
= . In the case of the tangent space,

the codimension of
is found to be + = . Thus (
= . Any basis for

this space must consist of linearly independent vectors which are all orthogonal to the tangent
space. Naturally, the subset of vectors {(, (
())
=1
forms just such a basis since it is given
to be linearly independent by the (
()) = condition. It follows that:

(
())
where equality can be obtained by the slightly tedious equation (
= ((
()
)) . That
equation simply does the following:
1. transpose
() to swap rows to columns

2. construct column space by taking span of columns in
()
3. adjoin to make pairs of vectors which live in
many wiser authors wouldnt bother. The comments above are primarily about notation. Certainly
hiding these details would make this section prettier, however, would it make it better? Finally, I
once more refer the reader to linear algebra where we learn that (())
= (
). Let me
walk you through the proof: let

. Observe (
) i
= 0 for
= 0 i
() = 0 for = 1, 2, . . . , i
() = 0 for = 1, 2, . . . , i ()
.
Another useful identity for the perp is that (
= . With those two gems in mind consider

that:
(
())
())
= (
()
)
Let me once more replace by a more tedious, but explicit, procedure:
= ((
()
))
Theorem 5.2.5.
Let :
be a level-mappping which denes a -dimensional level set by
1
() = . The tangent space
and the normal space at are given by
= {} (
()
) &
= {} (
()
).
Moreover,
. Every vector can be uniquely written as the sum of a tangent

vector and a normal vector.
The fact that there are only tangents and normals is the key to the method of Lagrange multipliers.
It forces two seemingly distinct objects to be in the same direction as one another.
4
be dened by (, , , ) = +
2
+
2
2
2
note that (, , , ) = 0
gives a three dimensional subset of
4
, lets call it . Notice =< 2, 2, 4, 1 > is nonzero
everywhere. Lets focus on the point (2, 2, 1, 0) note that (2, 2, 1, 0) = 0 thus the point is on .
The tangent plane at (2, 2, 1, 0) is formed from the union of all tangent vectors to = 0 at the
point (2, 2, 1, 0). To nd the equation of the tangent plane we suppose : is a curve with
= 0 and (0) = (2, 2, 1, 0). By assumption (()) = 0 since () for all . Dene
(0) =< , , , >, we nd a condition from the chain-rule applied to
= 0 at = 0,
()
_
=
_
_
(())
() = 0 (2, 2, 1, 0) < , , , >= 0

< 4, 4, 4, 1 > < , , , >= 0
4 + 4 4 + = 0
Thus the equation of the tangent plane is 4( 2) + 4( 2) 4( 1) + = 0. In invite the
reader to nd a vector in the tangent plane and check it is orthogonal to (2, 2, 1, 0). However,
this should not be surprising, the condition the chain rule just gave us is just the statement that
< , , , > ((2, 2, 1, 0)
) and that is precisely the set of vector orthogonal to (2, 2, 1, 0).

4
2
be dened by (, , , ) = ( +
2
+
2
2, +
2
+
2
2). In
this case (, , , ) = (0, 0) gives a two-dimensional manifold in
4
lets call it . Notice that
1
= 0 gives +
2
+
2
= 2 and
2
= 0 gives +
2
+
2
= 2 thus = 0 gives the intersection of
both of these three dimensional manifolds in
4
(no I cant see it either). Note,
1
=< 2, 2, 1, 0 >
2
=< 0, 2, 1, 2 >
It turns out that the inverse mapping theorem says = 0 describes a manifold of dimension 2 if
the gradient vectors above form a linearly independent set of vectors. For the example considered
here the gradient vectors are linearly dependent at the origin since
1
(0) =
2
(0) = (0, 0, 1, 0).
In fact, these gradient vectors are colinear along along the plane = = 0 since
1
(0, , , 0) =
2
(0, , , 0) =< 0, 2, 1, 0 >. We again seek to contrast the tangent plane and its normal at
some particular point. Choose (1, 1, 0, 1) which is in since (1, 1, 0, 1) = (0 + 1 + 1 2, 0 +
1 + 1 2) = (0, 0). Suppose that : is a path in which has (0) = (1, 1, 0, 1) whereas
(0) =< , , , >. Note that

1
(1, 1, 0, 1) =< 2, 2, 1, 0 > and
2
(1, 1, 0, 1) =< 0, 2, 1, 1 >.
Applying the chain rule to both
1
and
2
yields:
(
1

)
(0) =
1
((0)) < , , , >= 0 < 2, 2, 1, 0 > < , , , >= 0
(
2

)
(0) =
2
((0)) < , , , >= 0 < 0, 2, 1, 1 > < , , , >= 0
This is two equations and four unknowns, we can solve it and write the vector in terms of two free
variables correspondant to the fact the tangent space is two-dimensional. Perhaps its easier to use
matrix techiques to organize the calculation:
_
2 2 1 0
0 2 1 1
_
=
_
0
0
_
We calculate,
_
2 2 1 0
0 2 1 1
_
=
_
1 0 0 1/2
0 1 1/2 1/2
_
. Its natural to chose , as free vari-
ables then we can read that = /2 and = /2 /2 hence
< , , , >=< /2, /2 /2, , >=

2
< 0, 1, 2, 0 > +
2
< 1, 1, 0, 2 >
We can see a basis for the tangent space. In fact, I can give parametric equations for the tangent
space as follows:
(, ) = (1, 1, 0, 1) + < 0, 1, 2, 0 > + < 1, 1, 0, 2 >
Not surprisingly the basis vectors of the tangent space are perpendicular to the gradient vectors
1
(1, 1, 0, 1) =< 2, 2, 1, 0 > and
2
(1, 1, 0, 1) =< 0, 2, 1, 1 > which span the normal plane
to the tangent plane
at = (1, 1, 0, 1). We nd that
is orthogonal to
. In summary
and
=
4
. This is just a fancy way of saying that the normal and the tangent
plane only intersect at zero and they together span the entire ambient space.
5.3 method of Lagrange mulitpliers
Let us begin with a statement of the problem we wish to solve.
Problem: given an objective function :
and continuously dierentiable

constraint function :
, nd extreme values for the objective function

relative to the constraint () = .
Note that () = is a vector notation for -scalar equations. If we suppose (
()) =
then the constraint surface () = will form an ( )-dimensional level set. Let us make that
supposition throughout the remainder of this section.
In order to solve a problem it is sometimes helpful to nd necessary conditions by assuming an
answer exists. Let us do that here. Suppose
maps to the local extrema of (
) on =
1
{}.
This means there exists an open ball around
for which (
) is either an upper or lower bound

of all the values of over the ball intersected with . One clear implication of this data is that
if we take any continuously dierentiable curve on which passes through
, say :
with (0) =
and (()) = for all , then the composite

is a function on which takes
an extreme value at = 0. Fermats theorem from calculus I applies and as

is dierentiable
near = 0 we nd (

)
(0) = 0 is a necessary condition. But, this means we have two necessary

conditions on :
5.3. METHOD OF LAGRANGE MULITPLIERS 127
1. (()) =
2. (

)
(0) = 0
Let us expand a bit on both of these conditions:
1.
(0) = 0
2.
(0) = 0
The rst of these conditions places
(0)
but then the second condition says that
) =
()(
is orthogonal to
(0) hence ()(
. Now, recall from the last section that

the gradient vectors of the component functions to span the normal space, this means any vector
in
can be written as a linear combination of the gradient vectors. In particular, this means
there exist constants
1
,
2
, . . . ,
such that
()(
=
1
(
1
)(
+
2
(
2
)(
+ +
)(
We may summarize the method of Lagrange multipliers as follows:

1. choose -variables which aptly describe your problem.
2. identify your objective function and write all constraints as level surfaces.
3. solve =
1
1
+
2
2
+ +
subject to the constraint () = .

4. test the validity of your proposed extremal points.
The obvious gap in the method is the supposition that an extrema exists for the restriction
.
Well examine a few examples before I reveal a sucient condition. Well also see how absence of
that sucient condition does allow the method to fail.
Example 5.3.1. Suppose we wish to nd maximum and minimum distance to the origin for points
on the curve
2

2
= 1. In this case we can use the distance-squared function as our objective
(, ) =
2
+
2
and the single constraint function is (, ) =
2

2
. Observe that =<
2, 2 > whereas =< 2, 2 >. We seek solutions of = which gives us < 2, 2 >=
< 2, 2 >. Hence 2 = 2 and 2 = 2. We must solve these equations subject to the
condition
2

2
= 1. Observe that = 0 is not a solution since 0
2
= 1 has no real solution.
On the other hand, = 0 does t the constraint and
2
0 = 1 has solutions = 1. Consider
then
2 = 2 and 2 = 2 (1 ) = 0 and (1 +) = 0
Since = 0 on the constraint curve it follows that 1 = 0 hence = 1 and we learn that
(1 +1) = 0 hence = 0. Consequently, (1, 0 and (1, 0) are the two point where we expect to nd
extreme-values of . In this case, the method of Lagrange multipliers served its purpose, as you
can see in the graph. Below the green curves are level curves of the objective function whereas the
particular red curve is the given constraint curve.
The picture below is a screen-shot of the Java applet created by David Lippman and Konrad
Polthier to explore 2D and 3D graphs. Especially nice is the feature of adding vector elds to given
objects, many other plotters require much more eort for similar visualization. See more at the
website: http://dlippman.imathas.com/g1/GrapherLaunch.html.
Note how the gradient vectors to the objective function and constraint function line-up nicely at
those points.
In the previous example, we actually got lucky. There are examples of this sort where we could get
false maxima due to the nature of the constraint function.
Example 5.3.2. Suppose we wish to nd the points on the unit circle (, ) =
2
+
2
= 1 which
give extreme values for the objective function (, ) =
2

2
. Apply the method of Lagrange
multipliers and seek solutions to = :
< 2, 2 >= < 2, 2 >
We must solve 2 = 2 which is better cast as (1) = 0 and 2 = 2 which is nicely written
as (1 +) = 0. On the basis of these equations alone we have several options:
1. if = 1 then (1 + 1) = 0 hence = 0
2. if = 1 then (1 (1)) = 0 hence = 0
But, we also must t the constraint
2
+
2
= 1 hence we nd four solutions:
1. if = 1 then = 0 thus
2
= 1 = 1 (1, 0)
2. if = 1 then = 0 thus
2
= 1 = 1 (0, 1)
We test the objective function at these points to ascertain which type of extrema weve located:
(0, 1) = 0
2
(1)
2
= 1 & (1, 0) = (1)
2
0
2
= 1
When constrained to the unit circle we nd the objective function attains a maximum value of 1 at
the points (1, 0) and (1, 0) and a minimum value of 1 at (0, 1) and (0, 1). Lets illustrate the
answers as well as a few non-answers to get perspective. Below the green curves are level curves of
the objective function whereas the particular red curve is the given constraint curve.
The success of the last example was no accident. The fact that the constraint curve was a circle
which is a closed and bounded subset of
2
means that is is a compact subset of
2
. A well-known
theorem of analysis states that any real-valued continuous function on a compact domain attains
both maximum and minimum values. The objective function is continuous and the domain is
compact hence the theorem applies and the method of Lagrange multipliers succeeds. In contrast,
the constraint curve of the preceding example was a hyperbola which is not compact. We have
no assurance of the existence of any extrema. Indeed, we only found minima but no maxima in
Example 5.3.1.
The generality of the method of Lagrange multipliers is naturally limited to smooth constraint
curves and smooth objective functions. We must insist the gradient vectors exist at all points of
inquiry. Otherwise, the method breaks down. If we had a constraint curve which has sharp corners
then the method of Lagrange breaks down at those corners. In addition, if there are points of dis-
continuity in the constraint then the method need not apply. This is not terribly surprising, even in
calculus I the main attack to analyze extrema of function on assumed continuity, dierentiability
and sometimes twice dierentiability. Points of discontinuity require special attention in whatever
context you meet them.
At this point it is doubtless the case that some of you are, to misquote an ex-student of mine, not-
impressed. Perhaps the following examples better illustrate the dangers of non-compact constraint
curves.
Example 5.3.3. Suppose we wish to nd extrema of (, ) = when constrained to = 1.
Identify (, ) = = 1 and apply the method of Lagrange multipliers and seek solutions to
= :
< 1, 0 >= < , > 1 = and 0 =
If = 0 then 1 = is impossible to solve hence = 0 and we nd = 0. But, if = 0 then
= 1 is not solvable. Therefore, we nd no solutions. Well, I suppose we have succeeded here
in a way. We just learned there is no extreme value of on the hyperbola = 1. Below the
green curves are level curves of the objective function whereas the particular red curve is the given
constraint curve.
Example 5.3.4. Suppose we wish to nd extrema of (, ) = when constrained to
2
2
= 1.
Identify (, ) =
2

2
= 1 and apply the method of Lagrange multipliers and seek solutions to
= :
< 1, 0 >= < 2, 2 > 1 = 2 and 0 = 2
If = 0 then 1 = 2 is impossible to solve hence = 0 and we nd = 0. If = 0 and
2
2
= 1
then we must solve
2
= 1 whence = 1. We are tempted to conclude that:
1. the objective function (, ) = attains a maximum on
2
2
= 1 at (1, 0) since (1, 0) = 1
2. the objective function (, ) = attains a minimum on
2
2
= 1 at (1, 0) since (1, 0) =
1
But, both conclusions are false. Note
2
2
1
2
= 1 hence (
2, 1) are points on the constraint

curve and (
2, 1) =
2 and (
2, 1) =
2. The error of the method of Lagrange multipliers

in this context is the supposition that there exists extrema to nd, in this case there are no such
points. It is possible for the gradient vectors to line-up at points where there are no extrema. Below
the green curves are level curves of the objective function whereas the particular red curve is the
given constraint curve.
Incidentally, if you want additional discussion of Lagrange multipliers for two-dimensional prob-
lems one very nice source I certainly protted from was the YouTube video by Edward Frenkel of
Berkley. See his website http://math.berkeley.edu/ frenkel/ for links.
Example 5.3.5. XXX- add example from last notes. (chapter 8)
XXXneed to polish the notation for normal space. XXXadd examples for level sets.
Chapter 6
critical point analysis for several
variables
In the typical calculus sequence you learn the rst and second derivative tests in calculus I. Then
in calculus II you learn about power series and Taylors Theorem. Finally, in calculus III, in many
popular texts, you learn an essentially ad-hoc procedure for judging the nature of critical points
as minimum, maximum or saddle. These topics are easily seen as disconnected events. In this
chapter, we connect them. We learn that the geometry of quadratic forms is ellegantly revealed by
eigenvectors and more than that this geometry is precisely what elucidates the proper classications
of critical points of multivariate functions with real values.
6.1 multivariate power series
We set aside the issue of convergence for another course. We will suppose the series discussed in
this section exist on and converge on some domain, but we do not seek to treat that topic here and
now. Our focus is computational. How should we calculate the Taylor series for (, ) at (, )?
Or, what about () at
?.
6.1.1 taylors polynomial for one-variable
If : is analytic at
then we can write

() = (
) +
)(
) +
1
2
)(
)
2
+ =
=0
()
(
)
!
(
We could write this in terms of the operator =

and the evaluation of =
() =
_

=0
1
!
( )
()
_
=
=
133
134 CHAPTER 6. CRITICAL POINT ANALYSIS FOR SEVERAL VARIABLES
I remind the reader that a function is called entire if it is analytic on all of , for example
, cos()
and sin() are all entire. In particular, you should know that:
= 1 + +
1
2
2
+ =
=0
1
!
cos() = 1
1
2
2
+
1
4!
4
=
=0
(1)
(2)!

2
sin() =
1
3!
3
+
1
5!
5
=
=0
(1)
(2 + 1)!
2+1
Since
= cosh() + sinh() it also follows that

cosh() = 1 +
1
2
2
+
1
4!
4
=
=0
1
(2)!
2
sinh() = +
1
3!
3
+
1
5!
5
=
=0
1
(2 + 1)!
2+1
The geometric series is often useful, for , with < 1 it is known
+ +
2
+ =
=0
=

1
This generates a whole host of examples, for instance:
1
1 +
2
= 1
2
+
4
6
+
1
1
3
= 1 +
3
+
6
+
9
+
3
1 2
=
3
(1 + 2 + (2)
2
+ ) =
3
+ 2
4
+ 4
5
+
Moreover, the term-by-term integration and dierentiation theorems yield additional results in
conjuction with the geometric series:
tan
1
() =

1 +
2
=
=0
(1)
2
=
=0
(1)
2 + 1
2+1
=
1
3
3
+
1
5
5
+
ln(1 ) =
ln(1 ) =
1
1
=
=0
=0
1
+ 1
+1
6.1. MULTIVARIATE POWER SERIES 135
Of course, these are just the basic building blocks. We also can twist things and make the student
use algebra,
+2
=
2
=
2
(1 + +
1
2
2
+ )
or trigonmetric identities,
sin() = sin( 2 + 2) = sin( 2) cos(2) + cos( 2) sin(2)
sin() = cos(2)
=0
(1)
(2 + 1)!
( 2)
2+1
+ sin(2)
=0
(1)
(2)!
( 2)
2
.
Feel free to peruse my most recent calculus II materials to see a host of similarly sneaky calculations.
6.1.2 taylors multinomial for two-variables
Suppose we wish to nd the taylor polynomial centered at (0, 0) for (, ) =
sin(). It is a
simple as this:
(, ) =
_
1 + +
1
2
2
+
__

1
6
3
+
_
= + +
1
2
2

1
6
3
+
the resulting expression is called a multinomial since it is a polynomial in multiple variables. If
all functions (, ) could be written as (, ) = ()() then multiplication of series known
from calculus II would often suce. However, many functions do not possess this very special
form. For example, how should we expand (, ) = cos() about (0, 0)?. We need to derive the
two-dimensional Taylors theorem.
We already know Taylors theorem for functions on ,
() = () +
()( ) +
1
2
()( )
2
+ +
1
!
()
()( )
and... If the remainder term vanishes as then the function is represented by the Taylor
series given above and we write:
() =
=0
1
!
()
()( )
.
Consider the function of two variables :
2
which is smooth with smooth partial
derivatives of all orders. Furthermore, let (, ) and construct a line through (, ) with
direction vector (
1
,
2
) as usual:
() = (, ) +(
1
,
2
) = ( +
1
, +
2
)
for . Note (0) = (, ) and
() = (
1
,
2
) =
(0). Construct =

: and
choose () such that () for (). This function is a real-valued function of a
real variable and we will be able to apply Taylors theorem from calculus II on . However, to
dierentiate well need tools from calculus III to sort out the derivatives. In particular, as we
dierentiate , note we use the chain rule for functions of several variables:
() = (

)
() =
(())
()
= (()) (
1
,
2
)
=
1
( +
1
, +
2
) +
2
( +
1
, +
2
)
Note
(0) =
1
(, )+
2
(, ). Dierentiate again (I omit (()) dependence in the last steps),
() =
1
( +
1
, +
2
) +
2
( +
1
, +
2
)
=
1
(()) (
1
,
2
) +
2
(()) (
1
,
2
)
=
2
1
+
1
+
2
+
2
2
=
2
1
+ 2
1
+
2
2
Thus, making explicit the point dependence,
(0) =
2
1
(, ) +2
1
(, ) +
2
2
(, ). We
may construct the Taylor series for up to quadratic terms:
(0 +) = (0) +
(0) +
1
2
(0) +
= (, ) +[
1
(, ) +
2
(, )] +

2
2
_
2
1
(, ) + 2
1
(, ) +
2
2
(, )
+
Note that () = ( +
1
, +
2
) hence (1) = ( +
1
, +
2
) and consequently,
( +
1
, +
2
) = (, ) +
1
(, ) +
2
(, )+
+
1
2
_
2
1
(, ) + 2
1
(, ) +
2
2
(, )
_
+
Omitting point dependence on the 2
derivatives,
( +
1
, +
2
) = (, ) +
1
(, ) +
2
(, ) +
1
2
_
2
1
+ 2
1
+
2
2
+
Sometimes wed rather have an expansion about (, ). To obtain that formula simply substitute
=
1
and =
2
. Note that the point (, ) is xed in this discussion so the derivatives
are not modied in this substitution,
(, ) = (, ) + ( )
(, ) + ( )
(, )+
+
1
2
_
( )
2
(, ) + 2( )( )
(, ) + ( )
2
(, )
_
+
At this point we ought to recognize the rst three terms give the tangent plane to = (, ) at
(, , (, )). The higher order terms are nonlinear corrections to the linearization, these quadratic
terms form a quadratic form. If we computed third, fourth or higher order terms we will nd that,
using =
1
and =
2
as well as =
1
and =
2
,
(, ) =
=0
2
1
=0
2
2
=0

2
=0
1
!
()
(
1
,
2
)
1
)(
2
) (
)
Example 6.1.1. Expand (, ) = cos() about (0, 0). We calculate derivatives,
= sin()
= sin()
=
2
cos()
= sin() cos()
=
2
cos()
=
3
sin()
= cos() cos() +
2
sin()
= cos() cos() +
2
sin()
=
3
sin()
Next, evaluate at = 0 and = 0 to nd (, ) = 1 + to third order in , about (0, 0). We
can understand why these derivatives are all zero by approaching the expansion a dierent route:
simply expand cosine directly in the variable (),
(, ) = 1
1
2
()
2
+
1
4!
()
4
+ = 1
1
2
2
+
1
4!
4
+ .
Apparently the given function only has nontrivial derivatives at (0, 0) at orders 0, 4, 8, .... We can
deduce that
(0, 0) = 0 without furthter calculation.

This is actually a very interesting function, I think it dees our analysis in the later portion of this
chapter. The second order part of the expansion reveals nothing about the nature of the critical
point (0, 0). Of course, any student of trigonometry should recognize that (0, 0) = 1 is likely
a local maximum, its certainly not a local minimum. The graph reveals that (0, 0) is a local
maxium for restricted to certain rays from the origin whereas it is constant on several special
directions (the coordinate axes).
6.1.3 taylors multinomial for many-variables
Suppose : ()
is a function of -variables and we seek to derive the Taylor series

centered at = (
1
,
2
, . . . ,
). Once more consider the composition of with a line in ().

In particular, let :
be dened by () = + where = (
1
,
2
, . . . ,
) gives the
direction of the line and clearly
() = . Let : () be dened by () = (())

for all such that () (). Dierentiate, use the multivariate chain rule, recall here
that =
1
1
+
2
2
+ +
=1
() = (())
() = (()) =
=1
)(())
If we omit the explicit dependence on () then we nd the simple formula
() =
=1
.
Dierentiate a second time,
() =

=1
(())
_
=
=1
_
_
_
(())
_
=
=1
_
(())
()
Omitting the () dependence and once more using
() = we nd
() =
=1

Recall that =
=1
and expand the expression above,
() =
=1
=1
_
=
=1
=1
where we should remember
depends on (). It should be clear that if we continue and take

-derivatives then we will obtain:
()
() =
1
=1
2
=1

=1
More explicitly,
()
() =
1
=1
2
=1

=1
)(())
Hence, by Taylors theorem, provided we are suciently close to = 0 as to bound the remainder
1
() =
=0
1
!
_

1
=1
2
=1

=1
)(())
_
1
there exist smooth examples for which no neighborhood is small enough, the bump function in one-variable has
higher-dimensional analogues, we focus our attention to functions for which it is possible for the series below to
converge
Recall that () = (()) = ( +). Put
2
= 1 and bring in the
1
!
to derive
( +) =
=0
1
=1
2
=1

=1
1
!
_
_
()
.
Naturally, we sometimes prefer to write the series expansion about as an expresssion in = +.
With this substitution we have = and
= ( )
thus
() =
=0
1
=1
2
=1

=1
1
!
_
_
() (
1
)(
2
) (
).
3
lets unravel the Taylor series centered at (0, 0, 0) from the
general formula boxed above. Utilize the notation =
1
, =
2
and =
3
in this example.
() =
=0
3
1
=1
3
2
=1

3
=1
1
!
_
_
(0)
.
The terms to order 2 are as follows:
() = (0) +
(0) +
(0) +
(0)
+
1
2
_

(0)
2
+
(0)
2
+
(0)
2
+
+
(0) +
(0) +
(0) +
(0) +
(0) +
(0)
_
+
Partial derivatives commute for smooth functions hence,
() = (0) +
(0) +
(0) +
(0)
+
1
2
_

(0)
2
+
(0)
2
+
(0)
2
+ 2
(0) + 2
(0) + 2
(0)
_
+
1
3!
_

(0)
3
+
(0)
3
+
(0)
3
+ 3
(0)
2
+ 3
(0)
2
+3
(0)
2
+ 3
(0)
2
+ 3
(0)
2
+ 3
(0)
2
+ 6
(0)
_
+
Example 6.1.3. Suppose (, , ) =
. Find a quadratic approximation to near (0, 1, 2).

Observe:
= ()
2
= ()
2
= ()
2
+
2
+
2
+
2
2
if = 1 is not in the domain of then we should rescale the vector so that = 1 places (1) in (), if is
smooth on some neighborhood of then this is possible
Evaluating at = 0, = 1 and = 2,
(0, 1, 2) = 2
(0, 1, 2) = 0
(0, 1, 2) = 0
(0, 1, 2) = 4
(0, 1, 2) = 0
(0, 1, 2) = 0
(0, 1, 2) = 2
(0, 1, 2) = 0
(0, 1, 2) = 1
Hence, as (0, 1, 2) =
0
= 1 we nd
(, , ) = 1 + 2 + 2
2
+ 2( 1) + 2( 2) +
Another way to calculate this expansion is to make use of the adding zero trick,
(, , ) =
(1+1)(2+2)
= 1 +( 1 + 1)( 2 + 2) +
1
2
_
( 1 + 1)( 2 + 2)
2
+
Keeping only terms with two or less of , ( 1) and ( 2) variables,
(, , ) = 1 + 2 +( 1)(2) +(1)( 2) +
1
2
2
(1)
2
(2)
2
+
Which simplies once more to (, , ) = 1 + 2 + 2( 1) +( 2) + 2
2
+ .
6.2. A BRIEF INTRODUCTION TO THE THEORY OF QUADRATIC FORMS 141
6.2 a brief introduction to the theory of quadratic forms
Denition 6.2.1.
Generally, a quadratic form is a function :
whose formula can be written

() =
for all
where

such that
= . In particular, if = (, )
and =
_

_
then
() =
=
2
+ + +
2
=
2
+ 2 +
2
.
The = 3 case is similar,denote = [
] and = (, , ) so that
() =
=
11
2
+ 2
12
+ 2
13
+
22
2
+ 2
23
+
33
2
.
Generally, if [
]

and = [
then the associated quadratic form is

() =
=1
<
2
.
In case you wondering, yes you could write a given quadratic form with a dierent matrix which
is not symmetric, but we will nd it convenient to insist that our matrix is symmetric since that
choice is always possible for a given quadratic form.
It is at times useful to use the dot-product to express a given quadratic form:
= () = () =
Some texts actually use the middle equality above to dene a symmetric matrix.
Example 6.2.2.
2
2
+ 2 + 2
2
=
_

_
2 1
1 2
_ _

_
Example 6.2.3.
2
2
+ 2 + 3 2
2
2
=
_

2 1 3/2
1 2 0
3/2 0 1
Proposition 6.2.4.
The values of a quadratic form on
{0} is completely determined by its values on

the ( 1)-sphere
1
= {
= 1}. In particular, () =
2
( ) where
=
1
.
Proof: Let () =
. Notice that we can write any nonzero vector as the product of its
magnitude and its direction =
1
,
() = ( ) = ( )
=
2

=
2
( ).
Therefore () is simply proportional to ( ) with proportionality constant
2
.
The proposition above is very interesting. It says that if we know how works on unit-vectors then
we can extrapolate its action on the remainder of
. If : then we could say () > 0

i () > 0 for all . Likewise, () < 0 i () < 0 for all . The proposition below
follows from the proposition above since
2
ranges over all nonzero positive real numbers in the
equations above.
Proposition 6.2.5.
If is a quadratic form on
and we denote
{0}
1.(negative denite) (
) < 0 i (
1
) < 0
2.(positive denite) (
) > 0 i (
1
) > 0
3.(non-denite) (
) = {0} i (
1
) has both positive and negative values.
Before I get too carried away with the theory lets look at a couple examples.
Example 6.2.6. Consider the quadric form (, ) =
2
+
2
. You can check for yourself that
= (, ) is a cone and has positive outputs for all inputs except (0, 0). Notice that () =
2
so it is clear that (
1
) = 1. We nd agreement with the preceding proposition. Next, think about
the application of (, ) to level curves;
2
+
2
= is simply a circle of radius
or just the
origin. Heres a graph of = (, ):
Notice that (0, 0) = 0 is the absolute minimum for . Finally, lets take a moment to write
(, ) = [, ]
_
1 0
0 1
_ _

_
in this case the matrix is diagonal and we note that the e-values are
1
=
2
= 1.
2
2
2
. You can check for yourself
that = (, ) is a hyperboloid and has non-denite outputs since sometimes the
2
term
dominates whereas other points have 2
2
as the dominent term. Notice that (1, 0) = 1 whereas
(0, 1) = 2 hence we nd (
1
) contains both positive and negative values and consequently we
nd agreement with the preceding proposition. Next, think about the application of (, ) to level
curves;
2
2
2
= yields either hyperbolas which open vertically ( > 0) or horizontally ( < 0)
or a pair of lines =
2
in the = 0 case. Heres a graph of = (, ):
The origin is a saddle point. Finally, lets take a moment to write (, ) = [, ]
_
1 0
0 2
_ _

_
in this case the matrix is diagonal and we note that the e-values are
1
= 1 and
2
= 2.
Example 6.2.8. Consider the quadric form (, ) = 3
2
. You can check for yourself that =
(, ) is parabola-shaped trough along the -axis. In this case has positive outputs for all inputs
except (0, ), we would call this form positive semi-denite. A short calculation reveals that
(
1
) = [0, 3] thus we again nd agreement with the preceding proposition (case 3). Next, think
about the application of (, ) to level curves; 3
2
= is a pair of vertical lines: =
/3 or
just the -axis. Heres a graph of = (, ):
Finally, lets take a moment to write (, ) = [, ]
_
3 0
0 0
_ _

_
in this case the matrix is
diagonal and we note that the e-values are
1
= 3 and
2
= 0.
Example 6.2.9. Consider the quadric form (, , ) =
2
+2
2
+3
2
. Think about the application
of (, , ) to level surfaces;
2
+ 2
2
+ 3
2
= is an ellipsoid. I cant graph a function of three
variables, however, we can look at level surfaces of the function. I use Mathematica to plot several
below:
Finally, lets take a moment to write (, , ) = [, , ]
1 0 0
0 2 0
0 0 3
_
in this case the matrix
is diagonal and we note that the e-values are
1
= 1 and
2
= 2 and
3
= 3.
6.2.1 diagonalizing forms via eigenvectors
The examples given thus far are the simplest cases. We dont really need linear algebra to un-
derstand them. In contrast, e-vectors and e-values will prove a useful tool to unravel the later
examples
3
Denition 6.2.10.
Let

. If
1
is nonzero and = for some then we say is an
eigenvector with eigenvalue of the matrix .
Proposition 6.2.11.
Let

then is an eigenvalue of i () = 0. We say () = ()
the characteristic polynomial and () = 0 is the characteristic equation.
Proof: Suppose is an eigenvalue of then there exists a nonzero vector such that =
which is equivalent to = 0 which is precisely ( ) = 0. Notice that ( )0 = 0
3
this is the one place in this course where we need eigenvalues and eigenvector calculations, I include these to
illustrate the structure of quadratic forms in general, however, as linear algebra is not a prerequisite you may nd some
things in this section mysterious. The homework and study guide will elaborate on what is required this semester
thus the matrix ( ) is singular as the equation ( ) = 0 has more than one solution.
Consequently () = 0.
Conversely, suppose ( ) = 0. It follows that ( ) is singular. Clearly the system
( ) = 0 is consistent as = 0 is a solution hence we know there are innitely many solu-
tions. In particular there exists at least one vector = 0 such that () = 0 which means the
vector satises = . Thus is an eigenvector with eigenvalue for .
_
3 1
3 1
_
nd the e-values and e-vectors of .
() =
_
3 1
3 1
_
= (3 )(1 ) 3 =
2
4 = ( 4) = 0
We nd
1
= 0 and
2
= 4. Now nd the e-vector with e-value
1
= 0, let
1
= [, ]
denote the
e-vector we wish to nd. Calculate,
(0)
1
=
_
3 1
3 1
_ _

_
=
_
3 +
3 +
_
=
_
0
0
_
Obviously the equations above are redundant and we have innitely many solutions of the form
3 + = 0 which means = 3 so we can write,
1
=
_

3
_
=
_
1
3
_
. In applications we
often make a choice to select a particular e-vector. Most modern graphing calculators can calcu-
late e-vectors. It is customary for the e-vectors to be chosen to have length one. That is a useful
choice for certain applications as we will later discuss. If you use a calculator it would likely give
1
=
1
10
_
1
3
_
although the
10 would likely be approximated unless your calculator is smart.

Continuing we wish to nd eigenvectors
2
= [, ]
such that ( 4)
2
= 0. Notice that ,
are disposable variables in this context, I do not mean to connect the formulas from the = 0 case
with the case considered now.
(4)
1
=
_
1 1
3 3
_ _

_
=
_
+
3 3
_
=
_
0
0
_
Again the equations are redundant and we have innitely many solutions of the form = . Hence,
2
=
_

_
=
_
1
1
_
is an eigenvector for any such that = 0.
Theorem 6.2.13.
A matrix

is symmetric i there exists an orthonormal eigenbasis for .
There is a geometric proof of this theorem in Edwards
4
(see Theorem 8.6 pgs 146-147) . I prove half
of this theorem in my linear algebra notes by a non-geometric argument (full proof is in Appendix C
of Insel,Spence and Friedberg). It might be very interesting to understand the connection between
the geometric verse algebraic arguments. Well content ourselves with an example here:
0 0 0
0 1 2
0 2 1
. Observe that () = ( +1)( 3) thus

1
=
0,
2
= 1,
3
= 3. We can calculate orthonormal e-vectors of
1
= [1, 0, 0]
,
2
=
1
2
[0, 1, 1]
and
3
=
1
2
[0, 1, 1]
. I invite the reader to check the validity of the following equation:
1 0 0
0
1
2
1
2
0
1
2
1
0 0 0
0 1 2
0 2 1
1 0 0
0
1
2
1
2
0
1
2
1
0 0 0
0 1 0
0 0 3
Its really neat that to nd the inverse of a matrix of orthonormal e-vectors we need only take the
transpose; note
1 0 0
0
1
2
1
2
0
1
2
1
1 0 0
0
1
2
1
2
0
1
2
1
1 0 0
0 1 0
0 0 1
.
XXX remove comments about e-vectors and e-value before this section and put them here as
motivating examples for the proposition that follows.
Proposition 6.2.15.
If is a quadratic form on
with matrix and e-values

1
,
2
, . . . ,
with orthonormal
e-vectors
1
,
2
, . . . ,
then
(
) =
2
for = 1, 2, . . . , . Moreover, if = [
1
] then
() = (
=
1
2
1
+
2
2
2
+ +
where we dened =
.
Let me restate the proposition above in simple terms: we can transform a given quadratic form to
a diagonal form by nding orthonormalized e-vectors and performing the appropriate coordinate
transformation. Since is formed from orthonormal e-vectors we know that will be either a
rotation or reection. This proposition says we can remove cross-terms by transforming the
quadratic forms with an appropriate rotation.
Example 6.2.16. Consider the quadric form (, ) = 2
2
+ 2 + 2
2
. Its not immediately
obvious (to me) what the level curves (, ) = look like. Well make use of the preceding
4
think about it, there is a 1-1 correspondance between symmetric matrices and quadratic forms
proposition to understand those graphs. Notice (, ) = [, ]
_
2 1
1 2
_ _

_
. Denote the matrix
of the form by and calculate the e-values/vectors:
() =
_
2 1
1 2
_
= ( 2)
2
1 =
2
4 + 3 = ( 1)( 3) = 0
Therefore, the e-values are
1
= 1 and
2
= 3.
()
1
=
_
1 1
1 1
_ _

_
=
_
0
0
_

1
=
1
2
_
1
1
_
I just solved + = 0 to give = choose = 1 then normalize to get the vector above. Next,
(3)
2
=
_
1 1
1 1
_ _

_
=
_
0
0
_

2
=
1
2
_
1
1
_
I just solved = 0 to give = choose = 1 then normalize to get the vector above. Let
= [
1
2
] and introduce new coordinates = [ , ]
dened by =
. Note these can be

inverted by multiplication by to give = . Observe that
=
1
2
_
1 1
1 1
_

=
1
2
( + )
=
1
2
( + )
or
=
1
2
( )
=
1
2
( +)
The proposition preceding this example shows that substitution of the formulas above into yield
5
:
( , ) =
2
+ 3
2
It is clear that in the barred coordinate system the level curve (, ) = is an ellipse. If we draw
the barred coordinate system superposed over the -coordinate system then youll see that the graph
of (, ) = 2
2
+ 2 + 2
2
= is an ellipse rotated by 45 degrees. Or, if you like, we can plot
= (, ):
5
technically

( , ) is (( , ), ( , ))
2
+2+
2
. Its not immediately obvious
(to me) what the level curves (, ) = look like. Well make use of the preceding proposition to
understand those graphs. Notice (, ) = [, ]
_
1 1
1 1
_ _

_
. Denote the matrix of the form by
and calculate the e-values/vectors:
() =
_
1 1
1 1
_
= ( 1)
2
1 =
2
2 = ( 2) = 0
1
= 0 and
2
= 2.
(0)
1
=
_
1 1
1 1
_ _

_
=
_
0
0
_

1
=
1
2
_
1
1
_
(2)
2
=
_
1 1
1 1
_ _

_
=
_
0
0
_

2
=
1
2
_
1
1
_
= [
1
2
dened by =
. Note these can be

=
1
2
_
1 1
1 1
_

=
1
2
( + )
=
1
2
( + )
or
=
1
2
( )
=
1
2
( +)
The proposition preceding this example shows that substitution of the formulas above into yield:
( , ) = 2
2
It is clear that in the barred coordinate system the level curve (, ) = is a pair of paralell
lines. If we draw the barred coordinate system superposed over the -coordinate system then youll
see that the graph of (, ) =
2
+ 2 +
2
= is a line with slope 1. Indeed, with a little
algebraic insight we could have anticipated this result since (, ) = (+)
2
so (, ) = implies
+ =
thus =
. Heres a plot which again veries what weve already found:

Example 6.2.18. Consider the quadric form (, ) = 4. Its not immediately obvious (to
me) what the level curves (, ) = look like. Well make use of the preceding proposition to
understand those graphs. Notice (, ) = [, ]
_
0 2
0 2
_ _

_
. Denote the matrix of the form by
and calculate the e-values/vectors:
() =
_
2
2
_
=
2
4 = ( + 2)( 2) = 0
1
= 2 and
2
= 2.
(+ 2)
1
=
_
2 2
2 2
_ _

_
=
_
0
0
_

1
=
1
2
_
1
1
_
(2)
2
=
_
2 2
2 2
_ _

_
=
_
0
0
_

2
=
1
2
_
1
1
_
= [
1
2
dened by =
. Note these can be

=
1
2
_
1 1
1 1
_

=
1
2
( + )
=
1
2
( + )
or
=
1
2
( )
=
1
2
( +)
( , ) = 2
2
+ 2
2
It is clear that in the barred coordinate system the level curve (, ) = is a hyperbola. If we
draw the barred coordinate system superposed over the -coordinate system then youll see that
the graph of (, ) = 4 = is a hyperbola rotated by 45 degrees. The graph = 4 is thus a
hyperbolic paraboloid:
The fascinating thing about the mathematics here is that if you dont want to graph = (, ),
but you do want to know the general shape then you can determine which type of quadraic surface
youre dealing with by simply calculating the eigenvalues of the form.
Remark 6.2.19.
I made the preceding triple of examples all involved the same rotation. This is purely for my
lecturing convenience. In practice the rotation could be by all sorts of angles. In addition,
you might notice that a dierent ordering of the e-values would result in a redenition of
the barred coordinates.
6
We ought to do at least one 3-dimensional example.
Example 6.2.20. Consider the quadric form dened below:
(, , ) = [, , ]
6 2 0
2 6 0
0 0 5
Denote the matrix of the form by and calculate the e-values/vectors:

() =
6 2 0
2 6 0
0 0 5
= [( 6)
2
4](5 )
= (5 )[
2
12 + 32](5 )
= ( 4)( 8)(5 )
1
= 4,
2
= 8 and
3
= 5. After some calculation we nd the following
orthonormal e-vectors for :
1
=
1
1
1
0

2
=
1
1
1
0

3
=
0
0
1
Let = [
1
3
] and introduce new coordinates = [ , , ]
dened by =
. Note these
can be inverted by multiplication by to give = . Observe that
=
1
1 1 0
1 1 0
0 0

=
1
2
( + )
=
1
2
( + )
=
or
=
1
2
( )
=
1
2
( +)
=
( , , ) = 4
2
+ 8
2
+ 5
2
6.3. SECOND DERIVATIVE TEST IN MANY-VARIABLES 151
It is clear that in the barred coordinate system the level surface (, , ) = is an ellipsoid. If we
draw the barred coordinate system superposed over the -coordinate system then youll see that
the graph of (, , ) = is an ellipsoid rotated by 45 degrees around the . Plotted below
are a few representative ellipsoids:
In summary, the behaviour of a quadratic form () =
is governed by its set of eigenvalues

7
{
1
,
2
, . . . ,
}. Moreover, the form can be written as () =

1
2
1
+
2
2
2
+ +
by choosing
the coordinate system which is built from the orthonormal eigenbasis of (). In this coordinate
system the shape of the level-sets of becomes manifest from the signs of the e-values. )
Remark 6.2.21.
If you would like to read more about conic sections or quadric surfaces and their connection
to e-values/vectors I reccommend sections 9.6 and 9.7 of Antons linear algebra text. I
have yet to add examples on how to include translations in the analysis. Its not much
more trouble but I decided it would just be an unecessary complication this semester.
Also, section 7.1,7.2 and 7.3 in Lays linear algebra text show a bit more about how to
use this math to solve concrete applied problems. You might also take a look in Gilbert
Strangs linear algebra text, his discussion of tests for positive-denite matrices is much
more complete than I will give here.
6.3 second derivative test in many-variables
There is a connection between the shape of level curves (
1
,
2
, . . . ,
) = and the graph

+1
=
(
1
,
2
, . . . ,
) of . Ill discuss = 2 but these comments equally well apply to = (, , ) or

higher dimensional examples. Consider a critical point (, ) for (, ) then the Taylor expansion
about (, ) has the form
( +, +) = (, ) +(, )
where (, ) =
1
2
(, ) +
(, ) +
1
2
(, ) = [, ][](, ). Since []
= [] we can
nd orthonormal e-vectors
1
,
2
for [] with e-values
1
and
2
respective. Using = [
1
2
] we
7
this set is called the spectrum of the matrix
can introduce rotated coordinates (
) = (, ). These will give

(
) =
1
2
+
2
2
Clearly if
1
> 0 and
2
> 0 then (, ) yields the local minimum whereas if
1
< 0 and
2
< 0
then (, ) yields the local maximum. Edwards discusses these matters on pgs. 148-153. In short,
supposing () + , if all the e-values of are positive then has a local minimum of ()
at whereas if all the e-values of are negative then reaches a local maximum of () at .
Otherwise has both positive and negative e-values and we say is non-denite and the function
has a saddle point. If all the e-values of are positive then is said to be positive-denite
whereas if all the e-values of are negative then is said to be negative-denite. Edwards
gives a few nice tests for ascertaining if a matrix is positive denite without explicit computation
of e-values. Finally, if one of the e-values is zero then the graph will be like a trough.
2
2
+ 2 1) expand about the point (0, 1):
(, ) = (
2
)(
2
+ 2 1) = (
2
)(( 1)
2
)
expanding,
(, ) = (1
2
+ )(1 ( 1)
2
+ ) = 1
2
( 1)
2
+
Recenter about the point (0, 1) by setting = and = 1 + so
(, 1 +) = 1
2
2
+
If (, ) is near (0, 0) then the dominant terms are simply those weve written above hence the graph
is like that of a quadraic surface with a pair of negative e-values. It follows that (0, 1) is a local
maximum. In fact, it happens to be a global maximum for this function.
Example 6.3.2. Suppose (, ) = 4(1)
2
+(2)
2
+((1)
2
(2)
2
)+2(1)(2)
for some constants , . Analyze what values for , will make (1, 2) a local maximum, minimum
or neither. Expanding about (1, 2) we set = 1 + and = 2 + in order to see clearly the local
behaviour of at (1, 2),
(1 +, 2 +) = 4
2
2
+(
2
2
) + 2
= 4
2
2
+(1
2
2
) + 2
= 4 +(+ 1)
2
+ 2 (+ 1)
2
+
There is no nonzero linear term in the expansion at (1, 2) which indicates that (1, 2) = 4 +
may be a local extremum. In this case the quadratic terms are nontrivial which means the graph of
this function is well-approximated by a quadraic surface near (1, 2). The quadratic form (, ) =
(+ 1)
2
+ 2 (+ 1)
2
has matrix
[] =
_
(+ 1)
(+ 1)
2
_
.
6.3. SECOND DERIVATIVE TEST IN MANY-VARIABLES 153
The characteristic equation for is
([] ) =
_
(+ 1)
(+ 1)
2
_
= ( ++ 1)
2
2
= 0
We nd solutions
1
= 1 + and
2
= 1 . The possibilities break down as follows:
1. if
1
,
2
> 0 then (1, 2) is local minimum.
2. if
1
,
2
< 0 then (1, 2) is local maximum.
3. if just one of
1
,
2
is zero then is constant along one direction and min/max along another
so technically it is a local extremum.
4. if
1
2
< 0 then (1, 2) is not a local etremum, however it is a saddle point.
In particular, the following choices for , will match the choices above
1. Let = 3 and = 1 so
1
= 3 and
2
= 1;
2. Let = 3 and = 1 so
1
= 3 and
2
= 5
3. Let = 3 and = 2 so
1
= 0 and
2
= 4
4. Let = 1 and = 3 so
1
= 1 and
2
= 5
Here are the graphs of the cases above, note the analysis for case 3 is more subtle for Taylor
approximations as opposed to simple quadraic surfaces. In this example, case 3 was also a local
minimum. In contrast, in Example 6.2.17 the graph was like a trough. The behaviour of away
from the critical point includes higher order terms whose inuence turns the trough into a local
minimum.
Example 6.3.3. Suppose (, ) = sin() cos() to nd the Taylor series centered at (0, 0) we can
simply multiply the one-dimensional result sin() =
1
3!
3
+
1
5!
5
+ and cos() = 1
1
2!
2
+
1
4!
4
+ as follows:
(, ) = (
1
3!
3
+
1
5!
5
+ )(1
1
2!
2
+
1
4!
4
+ )
=
1
2
2
+
1
24
1
6
1
12
2
+
= +
The origin (0, 0) is a critical point since
(0, 0) = 0 and
(0, 0) = 0, however, this particular

critical point escapes the analysis via the quadratic form term since = 0 in the Taylor series
for this function at (0, 0). This is analogous to the inconclusive case of the 2nd derivative test in
calculus III.
Example 6.3.4. Suppose (, , ) = . Calculate the multivariate Taylor expansion about the
point (1, 2, 3). Ill actually calculate this one via dierentiation, I have used tricks and/or calculus
II results to shortcut any dierentiation in the previous examples. Calculate rst derivatives
= ,
and second derivatives,
= 0
= 0
= 0,
and the nonzero third derivatives,
= 1.
It follows,
( +, +, +) =
= (, , ) +
(, , ) +
(, , ) +
(, , ) +
1
2
(
) +
Of course certain terms can be combined since
etc... for smooth functions (we assume

smooth in this section, moreover the given function here is clearly smooth). In total,
(1 +, 2 +, 3 +) = 6 + 6 + 3 + 2 +
1
2
_
3 + 2 + 3 + + 2 +
_
+
1
3!
(6)
Of course, we could also obtain this from simple algebra:
(1 +, 2 +, 3 +) = (1 +)(2 +)(3 +) = 6 + 6 + 3 + + 3 + 2 + +.
6.3.1 morse theory and future reading
Chapter 7
multilinear algebra
7.1 dual space
Denition 7.1.1.
Suppose is a vector space over . We dene the dual space to to be the set of all
linear functions from to . In particular, we denote:
= { : ( +) = () +() and () = () , and }

If
then we say is a dual vector.

I oer several abstract examples to begin, however the majority of this section concerns
.
Example 7.1.2. Suppose denotes the set of continuous functions on . Dene () =
1
0
() .
The mapping : is linear by properties of denite integrals therefore we identify the denite
integral denes a dual-vector to the vector space of continuous functions.
Example 7.1.3. Suppose = (, ) denotes a set of functions from a vector space to .
Note that is a vector space with respect to point-wise dened addition and scalar multiplication
of functions. Let
and dene () = (
). The mapping : is linear since

( + ) = ( + )(
) = (
) + (
) = () + () for all , and . We nd

that the evaluation map denes a dual-vector
.
Example 7.1.4. The determinant is a mapping from

to but it does not dene a dual-vector
to the vector space of square matrices since (+) = () +().
Example 7.1.5. Suppose () = for a particular vector
. We argue
where we
recall =
is a vector space. Additivity follows from a property of the dot-product on
,
( +) = ( +) = + = () +()
for all ,
. Likewise, homogeneity follows from another property of the dot-product: observe

() = () = ( ) = ()
155
156 CHAPTER 7. MULTILINEAR ALGEBRA
for all
and .
Example 7.1.6. Let (, ) = 2 + 5 dene a function :
2
. Note that
(, ) = (, ) (2, 5)
hence by the preceding example we nd (
2
)
.
The preceding example is no accident. It turns out there is a one-one correspondance between row
vectors and dual vectors on
. Let
then we dene
() = . We proved in Example
7.1.5 that
. Suppose (
we see to nd
such that =
. Recall that a
linear function is uniquely dened by its values on a basis; the values of on the standard basis
will show us how to choose . This is a standard technique. Consider:
with
1
=
=1
() = (
=1
) =
=1
. .
) =
=1
)
. .
=
where we dene = ((
1
), (
2
), . . . , (
))
. The vector which corresponds naturally

2
to
is simply the vector of of the values of on the standard basis.
The dual space to
is a vector space and the correspondance
gives an isomorphism of
and (
. The image of a basis under an isomorphism is once more a basis. Dene :
()
by () =
to give the correspondance an explicit label. The image of the standard basis under
is called the standard dual basis for (
. Consider (
), let
and calculate
(
)() =
() =
In particular, notice that when =
then (
)(
) =
. Dual vectors are linear

transformations therefore we can dene the dual basis by its values on the standard basis.
Denition 7.1.7.
The standard dual basis of (
is denoted {
1
,
2
, . . . ,
} where we dene
to be the linear transformation such that
) =
for all ,
. Generally, given a
vector space with basis = {
1
,
2
, . . . ,
} we say the basis
= {
1
,
2
, . . . ,
} is
dual to i
) =
for all ,
.
The term basis indicates that {
1
,
2
, . . . ,
} is linearly independent
3
and {
1
,
2
, . . . ,
} =
(
. The following calculation is often useful: if
with =
=1
then
() =
=1
_
=
=1
) =
=1
() =
.
1
the super-index is not a power in this context, it is just a notation to emphasize
is the component of a vector.

2
some authors will say
1
is dual to
1
since () =
and
is a row vector, I will avoid that langauge

in these notes.
3
direct proof of LI is left to the reader
7.2. MULTILINEARITY AND THE TENSOR PRODUCT 157
The calculation above is a prototype for many that follow in this chapter. Next, suppose (
and suppose
with =
=1
. Calculate,
() =
_

=1
_
=
=1
(
() =
=1
(
this shows every dual vector is in the span of the dual basis {
=1
.
7.2 multilinearity and the tensor product
A multilinear mapping is a function of a Cartesian product of vector spaces which is linear with
respect to each slot. The goal of this section is to explain what that means. It turns out the set
of all multilinear mappings on a particular set of vector spaces forms a vector space and well show
how the tensor product can be used to construct an explicit basis by tensoring a bases which are
dual to the bases in the domain. We also examine the concepts of symmetric and antisymmetric
multilinear mappings, these form interesting subspaces of the set of all multilinear mappings. Our
approach in this section is to treat the case of bilinearity in depth then transition to the case of
multilinearity. Naturally this whole discussion demands a familarity with the preceding section.
7.2.1 bilinear maps
Denition 7.2.1.
Suppose
1
,
2
are vector spaces then :
1
2
is a binear mapping on
1
2
i
for all ,
1
, ,
2
and :
(1.) ( +, ) = (, ) +(, ) (linearity in the rst slot)
(2.) (, +) = (, ) +(, ) (linearity in the second slot).
bilinear maps on
When
1
=
2
= we simply say that : is a bilinear mapping on . The set of
all bilinear maps of is denoted
2
0
. You can show that
2
0
forms a vector space under
the usual point-wise dened operations of function addition and scalar multiplication
4
. Hopefully
you are familar with the example below.
Example 7.2.2. Dene :
by (, ) = for all ,
. Linearity in each slot

follows easily from properties of dot-products:
( +, ) = ( +) = + = (, ) +(, )
(, +) = ( +) = + = (, ) +(, ).
4
sounds like homework
We can use matrix multiplication to generate a large class of examples with ease.
Example 7.2.3. Suppose

and dene :
by (, ) =
for all
,
. Observe that, by properties of matrix multiplication,

( +, ) = ( +)
= (
) =
= (, ) +(, )
(, +) =
( +) =
= (, ) +(, )
for all , ,
and . It follows that is bilinear on
.
Suppose : is bilinear and suppose = {
1
,
2
, . . . ,
} is a basis for whereas
= {
1
,
2
, . . . ,
} is a basis of
with
) =
(, ) =
_

=1
=1
_
(7.1)
=
,=1
(
)
=
,=1
)
=
,=1
(
()
()
Therefore, if we dene
= (
) then we may compute (, ) =
,=1
. The calculation
above also indicates that is a linear combination of certain basic bilinear mappings. In particular,
can be written a linear combination of a tensor product of dual vectors on .
Denition 7.2.4.
Suppose is a vector space with dual space
. If ,
then we dene :
by ( )(, ) = ()() for all , .
Given the notation
5
preceding this denition, we note (
)(, ) =
()
() hence for all

, we nd:
(, ) =
,=1
(
)(
)(, ) therefore, =
,=1
(
We nd
6
that
2
0
= {
,=1
. Moreover, it can be argued
7
that {
,=1
is a linearly
independent set, therefore {
,=1
forms a basis for
2
0
. We can count there are
2
vectors
5
perhaps you would rather write (
)(, ) as
(, ), that is also ne.

6
with the help of your homework where you will show {
,=1

2
0

7
yes, again, in your homework
in {
,=1
hence (
2
0
) =
2
.
If =
and if {
=1
denotes the standard dual basis, then there is a standard notation for
the set of coecients found in the summation for . In particular, we denote = [] where
= (
) hence, following Equation 7.1,

(, ) =
,=1
) =
=1
=1
Denition 7.2.5.
Suppose : is a bilinear mapping then we say:
1. is symmetric i (, ) = (, ) for all ,
2. is antisymmetric i (, ) = (, ) for all ,
Any bilinear mapping on can be written as the sum of a symmetric and antisymmetric bilinear
mapping, this claim follows easily from the calculation below:
(, ) =
1
2
_
(, ) +(, )
_
. .
+
1
2
_
(, ) (, )
_
. .
.
We say
is symmetric in , i
for all , . Likewise, we say
is antisymmetric in
, i
for all , . If is a symmetric bilinear mapping and is an antisymmetric bilinear

mapping then the components of are symmetric and the components of are antisymmetric.
Why? Simply note:
(
) = (
and
(
) = (
.
You can prove that the sum or scalar multiple of an (anti)symmetric bilinear mapping is once more
(anti)symmetric therefore the set of antisymmetric bilinear maps
2
( ) and the set of symmetric
bilinear maps
0
2
are subspaces of
0
2
. The notation
2
( ) is part of a larger discussion on
the wedge product, we will return to it in a later section.
Finally, if we consider the special case of =
once more we nd that a bilinear mapping

:
has a symmetric matrix []
= [] i is symmetric whereas it has an antisymmetric

matric []
= [] i is antisymmetric.
bilinear maps on
Suppose :
is bilinear then we say

2
0
. In addition, suppose = {
1
,
2
, . . . ,
}
is a basis for whereas
= {
1
,
2
, . . . ,
} is a basis of
with
) =
. Let ,
(, ) =
_

=1
=1
_
(7.2)
=
,=1
(
)
=
,=1
)
=
,=1
(
)(
)(
)
Therefore, if we dene
= (
) then we nd the nice formula (, ) =
,=1
. To
further rene the formula above we need a new concept.
The dual of the dual is called the double-dual and it is denoted
. For a nite dimensional vector

space there is a cannonical isomorphism of and
. In particular, :
is dened by
()() = () for all
. It is customary to replace with
wherever the context allows.

For example, to dene the tensor product of two vectors , as follows:
Denition 7.2.6.
Suppose is a vector space with dual space
. We dene the tensor product of vectors

, as the mapping :
by ( )(, ) = ()() for all , .

We could just as well have dened = () () where is once more the cannonical
isomorphism of and
. Its called cannonical because it has no particular dependendence on

the coordinates used on . In contrast, the isomorphism of
and (
was built around the

dot-product and the standard basis.
All of this said, note that (
)(
) =
(, ) thus,
(, ) =
,=1
(
(, ) =
,=1
(
We argue that {
,=1
is a basis
8
Denition 7.2.7.
8
2
0
is a vector space and weve shown
2
0
( ) { }
,=1
but we should also show
2
0
and
check for LI of { }
,=1
.
Suppose :
is a bilinear mapping then we say:

1. is symmetric i (, ) = (, ) for all ,
2. is antisymmetric i (, ) = (, ) for all ,
The discussion of the preceding subsection transfers to this context, we simply have to switch some
vectors to dual vectors and move some indices up or down. I leave this to the reader.
bilinear maps on
Suppose :
is bilinear, we say
1
1
(or, if the context demands this detail

1
1
). We dene
1
1
( ) by the natural rule; ( )(, ) = ()() for all
(, )
. We nd, by calculations similar to those already given in this section,

(, ) =
,=1
and =
,=1
where we dened

= (
).
bilinear maps on
Suppose :
is bilinear, we say
1
1
(or, if the context demands this detail

1
1
). We dene
1
1
by the natural rule; ( )(, ) = ()() for all
(, )
. We nd, by calculations similar to those already given in this section,

(, ) =
,=1
and =
,=1
where we dened
= (
).
7.2.2 trilinear maps
Denition 7.2.8.
Suppose
1
,
2
,
3
1
3
is a trilinear mapping on
3
i for all ,
1
, ,
2
. ,
3
and :
(1.) ( +, , ) = (, , ) +(, , ) (linearity in the rst slot)
(2.) (, +, ) = (, , ) +(, , ) (linearity in the second slot).
(3.) (, , +) = (, , ) +(, , ) (linearity in the third slot).
If : is trilinear on then we say is a trilinear mapping on and
we denote the set of all such mappings
0
3
. The tensor product of three dual vectors is dened
much in the same way as it was for two,
( )(, , ) = ()()()
Let {
=1
is a basis for with dual basis {
=1
for
. If is trilinear on it follows
(, , ) =
,,=1
and =
,,=1
where we dened
= (
) for all , ,
.
Generally suppose that
1
,
2
,
3
are possibly distinct vector spaces. Moreover, suppose
1
has basis
{
1
=1
,
2
has basis {
2
=1
and
3
has basis {
3
=1
. Denote the dual bases for
1
,
2
,
3
in
the usual fashion: {
1
=1
, {
1
=1
, {
1
=1
. With this notation, we can write a trilinear mapping
on
1
3
as follows: (where we dene
= (
))
(, , ) =
=1
=1
=1
and =
=1
=1
=1
However, if
1
,
2
,
3
happen to be related by duality then it is customary to use up/down indices.
For example, if :
is trilinear then we write

9
=
,,=1
and say
1
2
. On the other hand, if :
is trilinear then wed write

=
,,=1
and say
2
1
. Im not sure that Ive ever seen this notation elsewhere, but perhaps it could
be useful to denote the set of trinlinear maps :
as
1
1 1
. Hopefully we will
not need such silly notation in what we consider this semester.
There was a natural correspondance between bilinear maps on
and square matrices. For a

trilinear map we would need a three-dimensional array of components. In some sense you could
picture :
as multiplication by a cube of numbers. Dont think too hard

about these silly comments, we actually already wrote the useful formulae for dealing with trilinear
objects. Lets stop to look at an example.
9
we identify
with its double-dual hence this tensor product is already dened, but to be safe let me write it out
in this context
(, , ) =
()
()(
).
Example 7.2.9. Dene :
3

3

3
by (, , ) = (). You may not have
learned this in your linear algebra course
10
but a nice formula
11
for the determinant is given by the
Levi-Civita symbol,
() =
3
,,=1
3
note that
1
() = [
1
],
2
() = [
2
] and
3
() = [
3
]. It follows that
(, , ) =
3
,,=1
Multilinearity follows easily from this formula. For example, linearity in the third slot:
(, , +) =
3
,,=1
( +)
(7.3)
=
3
,,=1
) (7.4)
=
3
,,=1
+
3
,,=1
(7.5)
= (, , ) +(, , ). (7.6)
Observe that by properties of determinants, or the Levi-Civita symbol if you prefer, swapping a pair
of inputs generates a minus sign, hence:
(, , ) = (, , ) = (, , ) = (, , ) = (, , ) = (, , ).
If : is a trilinear mapping such that
(, , ) = (, , ) = (, , ) = (, , ) = (, , ) = (, , )
for all , , then we say is antisymmetric. Likewise, if : is a trilinear
mapping such that
(, , ) = (, , ) = (, , ) = (, , ) = (, , ) = (, , ).
for all , , then we say is symmetric. Clearly the mapping dened by the determinant
is antisymmetric. In fact, many authors dene the determinant of an matrix as the antisym-
metric -linear mapping which sends the identity matrix to 1. It turns out these criteria unquely
10
maybe you havent even taken linear yet!
11
actually, I take this as the denition in linear algebra, it does take considerable eort to recover the expansion
by minors formula which I use for concrete examples
dene the determinant. That is the motivation behind my Levi-Civita symbol denition. That
formula is just the nuts and bolts of complete antisymmetry.
You might wonder, can every trilinear mapping can be written as a the sum of a symmetric and
antisymmetric mapping? The answer is no. Take the following trilinear mapping on
3
for example:
(, , ) = [
3
] +
You can verify this is linear in each slot however, it is antisymetric in the rst pair of slots
(, , ) = [
3
] + = [
3
] + = (, , )
and symmetric in the last pair,
(, , ) = [
3
] + = [
3
] + = (, , ).
Generally, the decomposition of a multilinear mapping into more basic types is a problem which
requires much more thought than we intend here. Representation theory is concerned with precisely
this problem: how can we decompose a tensor product into irreducible pieces. Their idea of tensor
product is not precisely the same as ours, however algebraically the problems are quite intertwined.
Ill leave it at that unless youd like to do an independent study on representation theory. Ideally
youd already have linear algebra and abstract algebra complete before you attempt that study.
7.2.3 multilinear maps
Denition 7.2.10.
Suppose
1
,
2
, . . .

1
is a -multilinear
mapping on
1
i for each and

1
,
1

1
,
2
,
2

2
, . . . ,
(
1
, . . . ,
, . . . ,
) = (
1
, . . . ,
, . . . ,
) +(
1
, . . . ,
, . . . ,
)
for = 1, 2, . . . , . In other words, we assume is linear in each of its -slots. If is
multilinear on
then we say that
and we say is a type (, ) tensor

on .
The denition above makes a dual vector a type (1, 0) tensor whereas a double dual of a vector a
type (0, 1) tensor, a bilinear mapping on is a type (2, 0) tensor and a bilinear mapping on
is
a type (0, 2) tensor with respect to .
We are free to dene tensor products in this context in the same manner as we have previously.
Suppose
1

1
,
2

2
, . . . ,
and
1

1
,
2

2
, . . . ,
then
(
1
,
2
, . . . ,
) =
1
(
1
)
2
(
2
)
)
It is easy to show the tensor produce of -dual vectors as dened above is indeed a -multilinear
mapping. Moreover, the set of all -multilinear mappings on
1

2

clearly forms a
vector space of dimension (
1
)(
2
) (
) since it naturally takes the tensor product of

the dual bases for
1
,
2
, . . . ,
as its basis. In particular, suppose for = 1, 2, . . . , that
has
basis {
=1
which is dual to {
=1
the basis for
. Then we can derive that a -multilinear

mapping can be written as
=
1
=1
2
=1

=1
2
...
1
1

2
2

If is a type (, ) tensor on then there is no need for the ugly double indexing on the basis
since we need only tensor a basis {
=1
for and its dual {
=1
for
in what follows:
=
1
,...,=1
1
,...,=1
2
...
2
...

.
permutations
Before I dene symmetric and antisymmetric for -linear mappings on I think it is best to discuss
briey some ideas from the theory of permutations.
Denition 7.2.11.
A permutation on {1, 2, . . . } is a bijection on {1, 2, . . . }. We dene the set of permutations
on {1, 2, . . . } to be
. Further, dene the sign of a permutation to be () = 1 if is

the product of an even number of transpositions whereas () = 1 if is the product
of a odd number transpositions.
Let us consider the set of permutations on {1, 2, 3, . . . }, this is called
the symmetric group,

its order is ! if you were wondering. Let me remind
12
you how the cycle notation works since it
allows us to explicitly present the number of transpositions contained in a permutation,
=
_
1 2 3 4 5 6
2 1 5 4 6 3
_
= (12)(356) = (12)(36)(35) (7.7)
recall the cycle notation is to be read right to left. If we think about inputing 5 we can read from
the matrix notation that we ought to nd 5 6. Clearly that is the case for the rst version of
written in cycle notation; (356) indicates that 5 6 and nothing else messes with 6 after that.
Then consider feeding 5 into the version of written with just two-cycles (a.k.a. transpositions ),
rst we note (35) indicates 5 3, then that 3 hits (36) which means 3 6, nally the cycle (12)
doesnt care about 6 so we again have that (5) = 6. Finally we note that () = 1 since it is
made of 3 transpositions.
It is always possible to write any permutation as a product of transpositions, such a decomposition
is not unique. However, if the number of transpositions is even then it will remain so no matter
12
or perhaps, more likely, introduce you to this notation
how we rewrite the permutation. Likewise if the permutation is an product of an odd number of
transpositions then any other decomposition into transpositions is also comprised of an odd number
of transpositions. This is why we can dene an even permutation is a permutation comprised by
an even number of transpositions and an odd permutation is one comprised of an odd number of
transpositions.
Example 7.2.12. Sample cycle calculations: we rewrite as product of transpositions to deter-
min if the given permutation is even or odd,
= (12)(134)(152) = (12)(14)(13)(12)(15) = () = 1
= (1243)(3521) = (13)(14)(12)(31)(32)(35) = () = 1
= (123)(45678) = (13)(12)(48)(47)(46)(45) = () = 1
We will not actually write down permutations in the calculations the follow this part of the notes.
I merely include this material as to give a logically complete account of antisymmetry. In practice,
if you understood the terms as the apply to the bilinear and trilinear case it will usually suce for
concrete examples. Now we are ready to dene symmetric and antisymmetric.
Denition 7.2.13.
A -linear mapping : is completely symmetric if
(
1
, . . . , , . . . , , . . . ,
) = (
1
, . . . , , . . . , , . . . ,
)
for all possible , . Conversely, if a -linear mapping on has
(
1
, . . . , , . . . , , . . . ,
) = (
1
, . . . , , . . . , , . . . ,
)
for all possible pairs , then it is said to be completely antisymmetric or alter-
nating. Equivalently a -linear mapping L is alternating if for all
1
,
2
, . . . ,
) = ()(
1
,
2
, . . . ,
)
The set of alternating multilinear mappings on is denoted , the set of -linear alter-
nating maps on is denoted
. Often an alternating -linear map is called a -form.

Moreover, we say the degree of a -form is .
Similar terminology applies to the components of tensors. We say
2
...
is completely symmetric
in
1
,
2
, . . . ,
2
...
(1)
(2)
...
()
for all
. On the other hand,
2
...
is completely
antisymmetric in
1
,
2
, . . . ,
2
...
= ()
(1)
(2)
...
()
for all
. It is a simple
exercise to show that a completely (anti)symmetric tensor
13
has completely (anti)symmetric com-
ponents.
13
in this context a tensor is simply a multilinear mapping, in physics there is more attached to the term
7.3. WEDGE PRODUCT 167
The tensor product is an interesting construction to discuss at length. To summarize, it is asso-
ciative and distributive across addition. Scalars factor out and it is not generally commutative.
For a given vector space we can in principle generate by tensor products multilinear mappings
of arbitrarily high order. This tensor algebra is innite dimensional. In contrast, the space of
forms on is a nite-dimensional subspace of the tensor algebra. We discuss this next.
7.3 wedge product
We assume is a vector space with basis {
=1
throughout this section. The dual basis is denoted
{
=1
as is our usual custom. Our goal is to nd a basis for the alternating maps on and explore
the structure implicit within its construction. This will lead us to call the exterior algebra
of after the discussion below is complete.
7.3.1 wedge product of dual basis generates basis for
Suppose : is antisymmetric and =
.=1
, it follows that
for all
,
. Notice this implies that
= 0 for = 1, 2, . . . , . For a given pair of indices , either

< or < or = hence,
=
<
<
<
<
<
<
<
<
<
). (7.8)
Therefore, {
and < } spans the set of antisymmetric bilinear maps on

. Moreover, you can show this set is linearly independent hence it is a basis fo
2
. We dene
the wedge product of
. With this notation we nd that the alternating

bilinear form can be written as
=
<
,=1
1
2
where the summation on the r.h.s. is over all indices

14
. Notice that
is an antisymmetric
bilinear mapping because
(, ) =
(, ), however, there is more structure here than

14
yes there is something to work out here, probably in your homework
just that. It is also true that
. This is a conceptually dierent antisymmetry, it

is the antisymmetry of the wedge produce .
Suppose : is antisymmetric and =
,,=1
, it follows that
and
for all , ,
. Notice this implies that
= 0 for = 1, 2, . . . , . A calculation similar to the one just oered for the case of a bilinear
map reveals that we can write as follows:
=
<<
_
(7.9)
Dene
thus
=
<<
,,=1
1
3!
(7.10)
and it is clear that {
, ,
and < < } forms a basis for the set of alternating

trilinear maps on .
Following the patterns above, we dene the wedge product of dual basis vectors,
()
(1)
(2)

()
(7.11)
If , we would like to show that
(. . . , , . . . , , . . . ) =
(. . . , , . . . , , . . . ) (7.12)
follows from the complete antisymmetrization in the denition of the wedge product. Before we
give the general argument, lets see how this works in the trilinear case. Consider,
=
=
.
Calculate, noting that
(, , ) =
()
()
() =
hence
(, , ) =
Thus,
(, , ) =
and you can check that
(, , ) =
(, , ). Similar tedious calculations prove

antisymmetry of the the interchange of the rst and second or the rst and third slots. Therefore,
is an alternating trilinear map as it is clearly trilinear since it is built from the sum of
tensor products which we know are likewise trilinear.
The multilinear case follows essentially the same argument, note
(. . . ,
, . . . ,
, . . . ) =
()
(1)
1

()
()
()
(7.13)
whereas,
(. . . ,
, . . . ,
, . . . ) =
()
(1)
1

()
()
()
. (7.14)
Suppose we take each permutation and subsitute
such that () = () and () = ()

and otherwise and agree. In cycle notation, () = . Substitution into Equation 7.14:
(. . . ,
, . . . ,
, . . . )
=
(())
(1)
1

()
()
()
()
(1)
1

()
()
()
(. . . ,
, . . . ,
, . . . ) (7.15)
Here the of a permutation is (1)
where is the number of cycles in . We observed

that () has one more cycle than hence (()) = (). Therefore, we have shown that
.
Recall that
in the = 2 case. There is a generalization of that result to the

> 2 case. In words, the wedge product is antisymetric with respect the interchange of any two
dual vectors. For = 3 we have the following identities for the wedge product:
. .
. .
. .
. .
. .
Ive indicated how these signs are consistent with the = 2 antisymmetry. Any permutation of
the dual vectors can be thought of as a combination of several transpositions. In any event, it is
sometimes useful to just know that the wedge product of three elements is invariant under cyclic
permutations of the dual vectors,
and changes by a sign for anticyclic permutations of the given object,
Generally we can argue that, for any permutation
= ()
(1)

(2)

()
This is just a slick formula which says the wedge product generates a minus whenever you ip two
dual vectors which are wedged.
7.3.2 the exterior algebra
The careful reader will realize we have yet to dene wedge products of anything except for the dual
basis. But, naturally you must wonder if we can take the wedge product of other dual vectors or
morer generally alternating tensors. The answer is yes. Let us dene the general wedge product:
Denition 7.3.1. Suppose
and
. We dene
to be the set of all increasing lists

of -indices, this set can be empty if ( ) is not suciently large. Moreover, if = (
1
,
2
, . . . ,
)
then introduce notation
hence:
=
1
,
2
,...,=1
1
!
2
...
1
!
and
=
1
,
2
,...,=1
1
!
2
...
1
!
Naturally,
and we dened this carefully in the

preceding subsection. Dene
+
as follows:
=
1
!!
.
Again, but with less slick notation:
=
1
,
2
,...,=1
1
,
2
,...,=1
1
!!
2
...
2
...
All the denition above really says is that we extend the wedge product on the basis to distribute
over the addition of dual vectors. What this means calculationally is that the wedge product obeys
the usual laws of addition and scalar multiplication. The one feature that is perhaps foreign is the
antisymmetry of the wedge product. We must take care to maintain the order of expressions since
the wedge product is not generally commutative.
Proposition 7.3.2.
Let , , be forms on and then
() ( +) = + distributes across vector addition
() ( +) = + distributes across vector addition
() () = () = ( ) scalars factor out
() ( ) = ( ) associativity
I leave the proof of this proposition to the reader.
Proposition 7.3.3. graded commutivity of homogeneous forms.
Let , be forms on of degree and respectively then
= (1)

Proof: suppose =
1
!
is a -form on and =
1
!
is a -form on . Calculate:
=
1
!!
by defn. of ,
=
1
!!
coecients are scalars,

= (1)
1
!!
(details on sign given below)

= (1)

Lets expand in detail why
= (1)
. Suppose = (
1
,
2
, . . . ,
) and =
(
1
,
2
, . . . ,
), the key is that every swap of dual vectors generates a sign:
= (1)
= (1)
(1)
.
.
.
.
.
.
.
.
.
= (1)
(1)
(1)
. .
= (1)
.
Example 7.3.4. Let be a 2-form dened by
=
1

2
+
2

3
And let be a 1-form dened by
= 3
1
Consider then,
= (
1

2
+
2

3
) (3
1
)
= (3
1

2

1
+ 3
2

3

1
= 3
1

2

3
.
(7.16)
whereas,
= 3
1
(
1

2
+
2

3
)
= (3
1

1

2
+ 3
1

2

3
= 3
1

2

3
.
(7.17)
so this agrees with the proposition, (1)
= (1)
2
= 1 so we should have found that = .
This illustrates that although the wedge product is antisymmetric on the basis, it is not always
antisymmetric, in particular it is commutative for even forms.
The graded commutivity rule = (1)
has some suprising implications. This rule is

ultimately the reason is nite dimensional. Lets see how that happens.
Proposition 7.3.5. linear dependent one-forms wedge to zero:
If ,
and = for some then = 0.

Proof: to begin, note that = hence 2 = 0 and it follows that = 0. Note:
= = (0) = 0
therefore the proposition is true.
Proposition 7.3.6.
Suppose that
1
,
2
, . . . ,
are linearly dependent 1-forms then
1

2

= 0.
Proof: by assumption of linear dependence there exist constants
1
,
2
, . . . ,
not all zero such

that
1
+
2
2
+
= 0.
Suppose that
is a nonzero constant in the sum above, then we may divide by it and consequently
we can write
in terms of all the other 1-forms,
=
1
1
+ +
1
1
+
+1
+1
+ +
_
Insert this sum into the wedge product in question,
1

2
. . .
=
1

2

= (
1
/
)
1

2

1

+(
2
/
)
1

2

2

+
+(
1
/
)
1

2

1

+(
+1
/
)
1

2

+1

+
+(
)
1

2

= 0.
(7.18)
We know all the wedge products are zero in the above because in each there is at least one 1-form
repeated, we simply permute the wedge products till they are adjacent and by the previous propo-
sition the term vanishes. The proposition follows.
Let us pause to reect on the meaning of the proposition above for a -dimensional vector space
. The dual space
is likewise -dimensional, this is a general result which applies to all nite-

dimensional vector spaces
15
. Thus, any set of more than dual vectors is necessarily linearly
dependent. Consquently, using the proposition above, we nd the wedge product of more than
one-forms is trivial. Therefore, while it is possible to construct
for > we should understand

that this space only contains zero. The highest degree of a nontrivial form over a vector space of
dimension is an -form.
Moreover, we can use the proposition to deduce the dimension of a basis for
, it must consist
of the wedge product of distinct linearly independent one-forms. The number of ways to choose
distinct objects from a list of distinct objects is precisely n choose p,
_
_
=
!
( )!!
for 0 . (7.19)
Proposition 7.3.7.
If is an -dimensional vector space then
is an
_
_
-dimensional vector space of -
forms. Moreover, the direct sum of all forms over has the structure
=
1

1

and is a vector space of dimension 2
Proof: dene
0
= then it is clear
forms a vector space for = 0, 1, . . . , . Moreover,
= {0} for = hence the term direct sum is appropriate. It remains to show
( ) = 2
where ( ) = . A natural basis for is found from taking the union of the
bases for each subspace of -forms,
= {1,
1
,
2
, . . . ,
1
1
<
2
< <
}
But, we can count the number of vectors in the set above as follows:
= 1 + +
_
2
_
+ +
_

1
_
+
_
_
Recall the binomial theorem states
( +)
=0
_
+
1
+ +
1
+
.
Recognize that = (1 + 1)

15
however, in innite dimensions, the story is not so simple
We should note that in the basis above the space of -forms is one-dimensional because there is
only one way to choose a strictly increasing list of integers in
. In particular, it is useful to note
= {
1
}. The form
1
is sometimes called the the top-form

16
.
Example 7.3.8. exterior algebra of
2
Let us begin with the standard dual basis {
1
,
2
}. By
denition we take the = 0 case to be the eld itself;
0
, it has basis 1. Next,
1
=
(
1
,
2
) =
and
2
= (
1

2
) is all we can do here. This makes a 2
2
= 4-
dimensional vector space with basis
{1,
1
,
2
,
1

2
}.
3
1
,
2
,
3
}.
By denition we take the = 0 case to be the eld itself;
0
1
=
(
1
,
2
,
3
) =
. Now for something a little more interesting,
2
= (
1

2
,
1

3
,
2

3
)
and nally,
3
= (
1

2

3
).
This makes a 2
3
= 8-dimensional vector space with basis
{1,
1
,
2
,
3
,
1

2
,
1

3
,
2

3
,
1

2

3
}
it is curious that the number of independent one-forms and 2-forms are equal.
4
1
,
2
,
3
,
4
}.
By denition we take the = 0 case to be the eld itself;
0
1
=
(
1
,
2
,
3
,
4
) =
. Now for something a little more interesting,
2
= (
1

2
,
1

3
,
1

4
,
2

3
,
2

4
,
3

4
)
and three forms,
3
= (
1

2

3
,
1

2

4
,
1

3

4
,
2

3

4
).
and
3
= (
1
3
). Thus a 2
4
= 16-dimensional vector space. Note that, in contrast
to
3
, we do not have the same number of independent one-forms and two-forms over
4
.
Lets explore how this algebra ts with calculations we already know about determinants.
1
2
]. I propose the determinant of is given by the top-form
on
2
via the formula () = (
1

2
)(
1
,
2
). Suppose =
_

_
then
1
= (, ) and
16
or volume form for reasons we will explain later, other authors begin the discussion of forms from the consideration
of volume, see Chapter 4 in Bernard Schutz Geometrical methods of mathematical physics
2
= (, ). Thus,
_

_
= (
1

2
)(
1
,
2
)
= (
1
1
)((, ), (, ))
=
1
(, )
2
(, )
2
(, )
1
(, )
= .
I hope this is not surprising!
1
3
]. I propose the determinant of is given by the top-
form on
3
via the formula () = (
1
3
)(
1
,
2
,
3
). Lets see if we can nd the expansion
by cofactors. By the denition we have
1

2

3
=
=
1
3
+
2
1
+
3
2
=
1
(
2
2
)
2
(
1
1
) +
3
(
1
1
)
=
1
(
2

3
)
2
(
1

3
) +
3
(
1

2
).
I submit to the reader that this is precisely the cofactor expansion formula with respect to the rst
column of . Suppose =
then
1
= (, , ),
2
= (, , ) and
3
= (, , ).
Calculate,
() =
1
(
1
)(
2

3
)(
2
,
3
)
2
(
1
)(
1

3
)(
2
,
3
) +
3
(
1
)(
1

2
)(
2
,
3
)
= (
2

3
)(
2
,
3
) (
1

3
)(
2
,
3
) +(
1

2
)(
2
,
3
)
= ( ) ( ) +( )
which is precisely my claim.
7.3.3 connecting vectors and forms in
3
There are a couple ways to connect vectors and forms in
3
. Mainly we need the following maps:
Denition 7.3.13.
Given =< , , >
3
we can construct a corresponding one-form
=
1
+
2
+
3
or we can construct a corresponding two-form
=
2

3
+
3

1
+
1

2
Recall that (
1
3
) = (
2
3
) = 3 hence the space of vectors, one-forms, and also two-
forms are isomorphic as vector spaces. It is not dicult to show that
1
+
2
=
1
+
2
and
1
+
2
=
1
+
2
for all
1
,
2

3
and . Moreover,
= 0 i = 0 and
= 0 i
= 0 hence () = {0} and () = {0} but this means that and are injective and since
the dimensions of the domain and codomain are 3 and these are linear transformations
17
it follows
and are isomorphisms.
It appears we have two ways to represent vectors with forms in
3
. Well see why this is important
as we study integration of forms. It turns out the two-forms go with surfaces whereas the one-
forms attach to curves. This corresponds to the fact in calculus III we have two ways to integrate
a vector-eld, we can either calculate ux or work. Partly for this reason the mapping is called
the work-form correspondence and is called the ux-form correspondence. Integration
has to wait a bit, for now we focus on algebra.
Example 7.3.14. Suppose =< 2, 0, 3 > and =< 0, 1, 2 > then
= 2
1
+3
3
and
=
2
+2
3
.
Calculate the wedge product,
= (2
1
+ 3
3
) (
2
+ 2
3
)
= 2
1
(
2
+ 2
3
) + 3
3
(
2
+ 2
3
)
= 2
1

2
+ 4
1

3
+ 3
3

2
+ 6
3

3
= 3
2

3
4
3

1
+ 2
1

2
=
<3,4,2>
=
(7.20)
Coincidence? Nope.
Proposition 7.3.15.
Suppose ,
3
then
where denotes the cross-product which is

dened by =
3
,,=1
.
Proof: Suppose =
3
=1
and =
3
=1
then
3
=1
and
3
=1
.
Calculate,
=
_
3
=1
_
3
=1
_
=
3
=1
3
=1
In invite the reader to show
= (
3
=1
) where Im using
= () to make the
argument of the ux-form mapping easier to read, hence,
=
3
=1
3
=1
(
3
=1
) = (
3
,,=1
)
. .

= ( ) =
Of course, if you dont like my proof you could just work it out like the example that precedes this
proposition. I gave the proof to show o the mappings a bit more.
17
this is not generally true, note () =
2
has () = 0 i = 0 and yet is not injective. The linearity is key.
7.4. BILINEAR FORMS AND GEOMETRY; METRIC DUALITY 177
Is the wedge product just the cross-product generalized? Well, not really. I think theyre quite
dierent animals. The wedge product is an associative product which makes sense in any vector
space. The cross-product only matches the wedge product after we interpret it through a pair of
isomorphisms ( and ) which are special to
3
. However, there is debate, largely the question
comes down to what you think makes the cross-product the cross-product. If you think it must
pick a unique perpendicular direction to a pair of given directions then that is only going to work
in
3
since even in
4
there is a whole plane of perpendicular vectors to a given pair. On the other
hand, if you think the cross-product in
4
should be pick the unique perpendicular to a given triple
of vectors then you could set something up. You could dene =
1
((
))
where :
3
4
is an isomorphism well describe in a upcoming section. But, you see its
no longer a product of two vectors, its not a binary operation, its a tertiary operation. In any
event, you can read a lot more on this if you wish. We have all the tools we need for this course.
The wedge product provides the natural antisymmetric algebra for -dimensiona and the work and
ux-form maps naturally connect us to the special world of three-dimensional mathematics.
There is more algebra for forms on
3
however we defer it to a later section where we have a few
more tools. Chief among those is the Hodge dual. But, before we can discuss Hodge duality we
need to generalize our idea of a dot-product just a little.
7.4 bilinear forms and geometry; metric duality
The concept of a metric goes beyond the familar case of the dot-product. If you want a more
strict generalization of the dot-product then you should think about an inner-product. In contrast
to the denition below, the inner-product replaces non-degeneracy with the stricter condition of
positive-denite which would read (, ) > 0 for = 0. I included a discussion of inner products
at the end of this section for the interested reader although we are probably not going to need all
of that material.
7.4.1 metric geometry
A geometry is a vector space paired with a metric. For example, if we pair
with the dot-

product you get Euclidean space. However, if we pair
4
with the Minkowski metric then we
obtain Minkowski space.
Denition 7.4.1.
If is a vector space and : is
1. bilinear:
0
2
,
2. symmetric: (, ) = (, ) for all , ,
3. nondegenerate: (, ) = 0 for all implies = 0.
the we call a metric on .
If =
then we can write (, ) =
where [] = . Moreover, (, ) = (, ) implies
= . Nondegenerate means that (, ) = 0 for all
i = 0. It follows that = 0
has no non-trivial solutions hence
1
exists.
for all ,
. This denes a metric for
, it is just
the dot-product. Note that (, ) =
hence we see [] = where denotes the identity

matrix in

.
Example 7.4.3. Suppose = (
0
,
1
,
2
,
3
), = (
0
,
1
,
2
,
3
)
4
then dene the Minkowski
product of and as follows:
(, ) =
0
0
+
1
1
+
2
2
+
3
3
It is useful to write the Minkowski product in terms of a matrix multiplication. Observe that for
,
4
,
(, ) =
0
0
+
1
1
+
2
2
+
3
3
=
_
3
_
1 0 0 0
0 1 0 0
0 0 1 0
0 0 0 1
where we have introduced the matrix of the Minkowski product. Notice that
= and () =
1 = 0 hence (, ) =
makes a symmetric, nondegenerate bilinear form on

4
. The
formula is clearly related to the dot-product. Suppose = (
0
, ) and = (
0
, ) then note
(, ) =
0
0
+
For vectors with zero in the zeroth slot this Minkowski product reduces to the dot-product. However,
for vectors which have nonzero entries in both the zeroth and later slots much diers. Recall that
any vectors dot-product with itself gives the square of the vectors length. Of course this means that
= 0 i = 0. Contrast that with the following: if = (1, 1, 0, 0) then
(, ) = 1 + 1 = 0
Yet = 0. Why study such a strange generalization of length? The answer lies in physics. Ill give
you a brief account by dening a few terms: Let = (
0
,
1
,
2
,
3
)
4
then we say
1. is a timelike vector if < , > < 0
2. is a lightlike vector if < , > = 0
3. is a spacelike vector if < , > > 0
If we consider the trajectory of a massive particle in
4
that begins at the origin then at any later
time the trajectory will be located at a timelike vector. If we consider a light beam emitted from
the origin then at any future time it will located at the tip of a lightlike vector. Finally, spacelike
vectors point to points in
4
which cannot be reached by the motion of physical particles that pass
throughout the origin. We say that massive particles are conned within their light cones, this
means that they are always located at timelike vectors relative to their current position in space
time. If youd like to know more I can reccomend a few books.
At this point you might wonder if there are other types of metrics beyond these two examples.
Surprisingly, in a certain sense, no. A rather old theorem of linear algebra due to Sylvester states
that we can change coordinates so that the metric more or less resembles either the dot-product or
something like it with some sign-ips. Well return to this in a later section.
7.4.2 metric duality for tensors
Throughout this section we consider a vector space paired with a metric : .
Moreover, the vector space has basis {
=1
which has a -dual basis {
=1
. Up to this point
we always have used a -dual basis where the duality was oered by the dot-product. In the
context of Minkowski geometry that sort of duality is no longer natural. Instead we must follow
the denition below:
Denition 7.4.4.
If is a vector space with metric and basis {
=1
then we say the basis {
=1
is -dual
i
Suppose
) =
and consider =
,=1
. Furthermore, suppose
are the com-

ponents of the inverse matrix to (
) this means that
=1
. We use the components

of the metric and its inverse to raise and lower indices on tensors. Here are the basic conven-
tions: given an object
which has the contravariant index we can lower it to be covariant by

contracting against the metric components as follows:
On the other hand, given an object
which has a covariant index we can raise it to be con-

travariant by contracting against the inverse components of the metric:
I like to think of this as some sort of conservation of indices. Strict adherence to the notation
drives us to write things such as
=1
just to keep up the up/down index pattern.

I should mention that these formulas are much more beautiful in the physics literature, you can
look at my old Math 430 notes from NCSU if youd like a healthy dose of that notation
18
. I use
Einsteins implicit summation notation throughout those notes and I discuss this index calculation
more in the way a physicist typically approaches it. Here I am trying to be careful enough that
these equations are useful to mathematicians. Let me show you some examples:
Example 7.4.5. Specialize for this example to =
4
with (, ) =
. Suppose =
4
=0
the components
are called contravariant components. The metric allows us

to dene covariant components by
=
4
=0
.
For the minkowski metric this just adjoins a minus to the zeroth component: if (
) = (, , , )
then
= (, , , ).
Example 7.4.6. Suppose we are working on
with the Euclidean metric
and it follows
that
or to be a purist for a moment
. In this case
. The covariant and contravariant components are the same. This is why is was ok
to ignore up/down indices when we work with a dot-product exclusively.
What if we raise an index and the lower it back down once more? Do we really get back where we
started? Given
we lower the index by
then we raise it once more by
and the last summation squishes down to
once more. It would seem this procedure of raising

and lowering indices is at least consistent.
Example 7.4.7. Suppose we raise the index on the basis {
} and formally obtain {
}
on the other hand suppose we lower the index on the dual basis {
} to formally obtain {
}. Im curious, are these consistent? We should get
) =
, Ill be nice an look at
) in the following sense:
_
=
) =
Interesting, but what does it mean?

I used the term formal in the preceding example to mean that the example makes sense in as much
as you accept the equations which are written. If you think harder about it then youll nd it was
rather meaningless. That said, this index notation is rather forgiving.
18
just a taste: =
or
or

Ok, but what are we doing? Recall that I insisted on using lower indices for forms and upper
indices for vectors? The index conventions Im toying with above are the reason for this strange
notation. When we lower an index we might be changing a vector to a dual vector, or vice-versa
when we raise an index we might be changing a dual vector into a vector. Let me be explicit.
1. given we create
by the rule
() = (, ).
2. given
we create
by the rule
() =
1
(, ) where
1
(, ) =
.
Recall we at times identify and
. Lets work out the component structure of
and see how

it relates to ,
) = (,
) = (
) =
) =
Thus,
where
. When we lower the index were actually using an

isomorphism which is provided by the metric to map vectors to forms. The process of raising the
index is just the inverse of this isomorphism.
) =
1
(,
) =
1
(
) =
thus
where
.
Suppose we want to change a type (0, 2) tensor to a type (2, 0) tensor. Were given :
where =
. Dene

: as follows:
(, ) = (
)
What does this look like in components? Note
) = (
) =
hence
and
) = (
) =
_
_
=
) =
Or, as is often customary, we could write
. However, this is an abuse of notation

since
are not technically components for . If we have a metric we can recover either from

or vice-versa. Generally, if we are given two tensors, say

1
of rank (, ) and the
2
of rank (
),
then these might be equilvalent if + =
. It may be that through raising and lowering

indices (a.k.a. appropriately composing with the vectordual vector isomorphisms) we can convert
1
to
2
. If you read Gravitation by Misner, Thorne and Wheeler youll nd many more thoughts
on this equivalence. Challenge: can you nd the explicit formulas like

(, ) = (
) which
back up the index calculations below?
or
I hope Ive given you enough to chew on in this section to put these together.
7.4.3 inner products and induced norm
There are generalized dot-products on many abstract vector spaces, we call them inner-products.
Denition 7.4.8.
Suppose is a vector space. If <, >: is a function such that for all , ,
and :
1. < , >=< , > (symmetric)
2. < +, >=< , > + < , > (additive in the rst slot)
3. < , >= < , > (homogeneity in the rst slot)
4. < , > 0 and < , >= 0 i = 0
then we say (, <, >) is an inner-product space with inner product <, >.
Given an inner-product space (, <, >) we can easily induce a norm for by the formula =
< , > for all . Properties (1.), (3.) and (4.) in the denition of the norm are fairly obvious
for the induced norm. Lets think throught the triangle inequality for the induced norm:
+
2
= < +, + > def. of induced norm
= < , + > + < , + > additive prop. of inner prod.
= < +, > + < +, > symmetric prop. of inner prod.
= < , > + < , > + < , > + < , > additive prop. of inner prod.
=
2
+ 2 < , > +
2
At this point were stuck. A nontrivial identity
19
called the Cauchy-Schwarz identity helps us
proceed; < , > . It follows that +
2

2
+ 2 +
2
= ( + )
2
.
However, the induced norm is clearly positive
20
so we nd + +.
Most linear algebra texts have a whole chapter on inner-products and their applications, you can
look at my notes for a start if youre curious. That said, this is a bit of a digression for this course.
19
I prove this for the dot-product in my linear notes, however, the proof is written in such a way it equally well
applies to a general inner-product
20
note: if you have (5)
2
< (7)
2
it does not follow that 5 < 7, in order to take the squareroot of the inequality
we need positive terms squared
7.5. HODGE DUALITY 183
7.5 hodge duality
We can prove that
_
_
=
_

_
. This follows from explicit computation of the formula for
_
_
or
from the symmetry of Pascals triangle if you prefer. In any event, this equality suggests there is
some isomorphism between and ( )-forms. When we are given a metric on a vector space
(and the notation of the preceding section) it is fairly simple to construct the isomorphism.
Suppose we are given
and following our usual notation:

=
1
,
2
,...,=1
1
!
2
...
Then, dene the hodge dual to be the ( )-form given below:

=
1
,
2
,...,=1
1
!( )!
2
...
2
...
1
2
...
I should admit, to prove this is a reasonable denition wed need to do some work. Its clearly a
linear transformation, but bijectivity and coordinate invariance of this denition might take a little
work. I intend to omit those details and instead focus on how this works for
3
or
4
. My advisor
taught a course on ber bundles and there is a much more general and elegant presentation of the
hodge dual over a manifold. Ask if interested, I think I have a pdf.
7.5.1 hodge duality in euclidean space
3
To begin, consider a scalar 1, this is a 0-form so we expect the hodge dual to give a 3-form:
1 =
,,
1
0!3!
=
1

2

3
Interesting, the hodge dual of 1 is the top-form on
3
. Conversely, calculate the dual of the top-
form, note
1
3
=
1
6
reveals the components of the top-form are precisely
thus:
(
1

2

3
) =
3
,,=1
1
3!(3 3)!
=
1
6
(1 + 1 + 1 + (1)
2
+ (1)
2
+ (1)
2
) = 1.
Next, consider
1
, note that
1
=
hence the components are

1
. Thus,
1
=
,,
1
1!2!
=
1
2
=
1
2
(
231
2

3
+
321
3

2
) =
2

3
Similar calculations reveal
2
=
3
1
and
3
=
1
2
. What about the duals of the two-forms?
Begin with =
1

2
note that
1

2
=
1

2

2

1
thus we can see the components are
=
1
1
. Thus,
(
1

2
) =
,,
1
2!1!
(
1
1
)
=
1
2
_
12
21
_
=
1
2
(
3
(
3
)) =
3
.
Similar calculations show that (
2

3
) =
1
and (
3

1
) =
2
. Put all of this together and we
nd that
(
1
+
2
+
3
) =
2

3
+
3

1
+
1

2
and
(
2

3
+
3

1
+
1

2
) =
1
+
2
+
3
Which means that
and
. Hodge duality links the two dierent form-representations

of vectors in a natural manner. Moveover, for
3
we should also note that = for all
3
.
In general, for other metrics, we can have a change of signs which depends on the degree of .
We can summarize hodge duality for three-dimensional Euclidean space as follows:
1 =
1

2

3
(
1

2

3
) = 1
1
=
2

3
(
2

3
) =
1
2
=
3

1
(
3

1
) =
2
3
=
1

2
(
1

2
) =
3
A simple rule to calculate the hodge dual of a basis form is as follows
1. begin with the top-form
1

2

3
2. permute the forms until the basis form you wish to hodge dual is to the left of the expression,
whatever remains to the right is the hodge dual.
For example, to calculate the dual of
2

3
note
1

2

3
=
2

3
. .

1
..

(
2

3
) =
1
.
Consider what happens if we calculate , since the dual is a linear operation it suces to think
about the basis forms. Let me sketch the process of
where is a multi-index:
1. begin with
1

2

3
2. write
1

2

3
= (1)
and identify
= (1)
.
3. then to calculate the second dual once more begin with
1

2

3
and note
1

2

3
= (1)
since the same transpositions are required to push
to the left or
to the right.
4. It follows that
for any multi-index hence = for all

3
.
I hope that once you get past the index calculation you can see the hodge dual is not a terribly
complicated construction. Some of the index calculation in this section was probably gratutious,
but I would like you to be aware of such techniques. Brute-force calculation has its place, but a
well-thought index notation can bring far more insight with much less eort.
7.5.2 hodge duality in minkowski space
4
The logic here follows fairly close to the last section, however the wrinkle is that the metric here
demands more attention. We must take care to raise the indices on the forms when we Hodge dual
them. First lets list the basis forms, we have to add time to the mix ( again = 1 so
0
= = if
you worried about it ) Remember that the Greek indices are dened to range over 0, 1, 2, 3. Here
Name Degree Typical Element Basis for
4
function = 0 1
one-form = 1 =
0
,
1
,
2
,
3
two-form = 2 =
,
1
2
2

3
,
3

1
,
1

2
0

1
,
0

2
,
0

3
three-form = 3 =
,,
1
3!
1

2

3
,
0

2

3
0

1

3
,
0

1

2
four-form = 4
0

1

2

3
0

1

2

3
the top form is degree four since in four dimensions we can have at most four dual-basis vectors
without a repeat. Wedge products work the same as they have before, just now we have
0
to play
with. Hodge duality may oer some surprises though.
Denition 7.5.1. The antisymmetric symbol in at
4
is denoted
and it is dened by the

value
0123
= 1
plus the demand that it be completely antisymmetric.
We must not assume that this symbol is invariant under a cyclic exhange of indices. Consider,
0123
=
1023
ipped (01)
= +
1203
ipped (02)
=
1230
ipped (03).
(7.21)
In four dimensions well use antisymmetry directly and forego the cyclicity shortcut. Its not a big
deal if you notice it before it confuses you.
Example 7.5.2. Find the Hodge dual of =
1
with respect to the Minkowski metric
, to begin
notice that has components
=
1
as is readily veried by the equation

1
=
. Lets
raise the index using as we learned previously,
=
1
=
1
Starting with the denition of Hodge duality we calculate
(
1
) =
,,,
1
!
1
()!
,,,
(1/6)
1
,,
(1/6)
1
= (1/6)[
1023
0

2

3
+
1230
2

3

0
+
1302
3

0

2
+
1320
3

2

0
+
1203
2

0

3
+
1032
0

3

2
]
= (1/6)[
0

2

3
2

3

0
3

0

2
+
3

2

0
+
2

0

3
+
0

3

2
]
=
2

3

0
=
0

2

3
.
(7.22)
the dierence between the three and four dimensional Hodge dual arises from two sources, for one
we are using the Minkowski metric so indices up or down makes a dierence, and second the
antisymmetric symbol has more possibilities than before because the Greek indices take four values.
I suspect we can calculate the hodge dual by the following pattern: suppose we wish to nd the
dual of where is a basis form for
4
with the Minkowski metric
1. begin with the top-form
0

1

2

3
2. permute factors as needed to place to the left,
3. the form which remains to the right will be the hodge dual of if no
0
is in otherwise the
form to the right multiplied by 1 is .
Note this works for the previous example as follows:
1. begin with
0

1

2

3
2. note
0

1

2

3
=
1

0

2

3
=
1
(
0

2

3
)
3. identify
1
=
0

2

3
(no extra sign since no
0
appears in
1
)
Follow the algorithm for nding the dual of
0
,
1. begin with
0

1

2

3
2. note
0

1

2

3
=
0
(
1

2

3
)
3. identify
0
=
1

2

3
( added sign since
0
appears in form being hodge dualed)
Lets check from the denition if my algorithm worked out right.
0
, to begin
notice that
0
has components
=
0
as is readily veried by the equation

0
=
. Lets
raise the index using as we learned previously,
=
0
=
0
the minus sign is due to the Minkowski metric. Starting with the denition of Hodge duality we
calculate
(
0
) =
,,,
1
!
1
()!
,,,
(1/6)
0
,,
(1/6)
0
,,
(1/6)
0
,,
(1/6)
1

2

3
sneaky step
=
1

2

3
.
(7.23)
Notice I am using the convention that Greek indices sum over 0, 1, 2, 3 whereas Latin indices sum
over 1, 2, 3.
0

1
, to
begin notice the following identity, it will help us nd the components of
0

1
=
,
1
2
2
0
now we antisymmetrize to get the components of the form,
0

1
=
,
1
2
0
[
1
]
where
0
[
1
]
=
0
and the factor of two is used up in the antisymmetrization. Lets raise

the index using as we learned previously,
0
[
1
]
=
0
1
+
0
1
=
[0
]1
the minus sign is due to the Minkowski metric. Starting with the denition of Hodge duality we
calculate
(
0

1
) =
1
!
1
()!
= (1/4)(
[0
]1
)
= (1/4)(
01
10
)
= (1/2)
01
= (1/2)[
0123
2

3
+
0132
3

2
]
=
2

3
(7.24)
Note, the algorithm works out the same,
0

1

2

3
=
0

1
. .

0
(
2

3
) (
0

1
) =
2

3
The other Hodge duals of the basic two-forms calculate by almost the same calculation. Let us make
a table of all the basic Hodge dualities in Minkowski space, I have grouped the terms to emphasize
1 =
0

1

2

3
(
0

1

2

3
) = 1
(
1

2

3
) =
0
0
=
1

2

3
(
0

2

3
) =
1
1
=
2

3

0
(
0

3

1
) =
2
2
=
3

1

0
(
0

1

2
) =
3
3
=
1

2

0
(
3

0
) =
1

2
(
1

2
) =
3

0
(
1

0
) =
2

3
(
2

3
) =
1

0
(
2

0
) =
3

1
(
3

1
) =
2

0
the isomorphisms between the one-dimensional
0
4
and
4
4
, the four-dimensional
1
4
and
4
, the six-dimensional
2
4
and itself. Notice that the dimension of
4
is 16 which we have
explained in depth in the previous section. Finally, it is useful to point out the three-dimensional
work and ux form mappings to provide some useful identities in this 1 + 3-dimensional setting.
=
0

=
0

_
=
I leave verication of these formulas to the reader ( use the table). Finally let us analyze the process
of taking two hodge duals in succession. In the context of
3
we found that = , we seek to
discern if a similar formula is available in the context of
4
with the minkowksi metric. We can
calculate one type of example with the identities above:
=
0

= (
0

) =
7.6. COORDINATE CHANGE 189

Perhaps this is true in general?
If we accept my algorithm then its not too hard to sort through using multi-index notation: since
hodge duality is linear it suces to consider a basis element
where is a multi-index,
1. transpose dual vectors so that
0

1

2

3
= (1)
2. if 0 / then
= (1)
and 0 since = {0, 1, 2, 3}. Take a second dual by

writing
0

1

2

3
= (1)
but note ((1)
) =
since 0 . We nd

for all not containing the 0-index.

3. if 0 then
= (1)
and 0 / since = {0, 1, 2, 3}. Take a second dual by

writing
0
3
= (1)
) and hence ((1)
) =
since 0 / . We
nd
for all containing the 0-index.

4. it follows that = for all
4
with the minkowski metric.
To conclude, I would warn the reader that the results in this section pertain to our choice of notation
for
4
. Some other texts use a metric which is relative to our notation. This modies many
signs in this section. See Misner, Thorne and Wheelers Gravitation or Bertlmanns Anomalies in
Field Theory for future reading on Hodge duality and a more systematic explaination of how and
when these signs arise from the metric.
7.6 coordinate change
Suppose has two bases

= {
1
,

2
, . . . ,

} and = {
1
,
2
, . . . ,
}. If then we can write

in as a linear combination of the

basis or the basis:
=
1
1
+
2
2
+ +
and =
1

1
+
2

2
+ +
given the notation above, we dene coordinate maps as follows:
() = (
1
,
2
, . . . ,
) = and
() = (
1
,
2
, . . . ,
) =
We sometimes use the notation
() = []
= whereas
() = []
= . A coordinate map
takes an abstract vector and maps it to a particular representative in
. A natural question
to ask is how do dierent representatives compare? How do and compare in our current
notation? Because the coordinate maps are isomorphisms it follows that
is an
isomorphism and given the domain and codomain we can write its formula via matrix multiplication:
() =
( ) =
However,
1
( ) = hence
() = and consequently, = . Conversely, to switch to

barred coordinates we multiply the coordinate vectors by
1
; =
1
.
Continuing this discussion we turn to the dual space. Suppose

= {
=1
is dual to

= {
=1
and
= {
=1
is dual to = {
=1
. By denition we are given that
) =
and
) =
for all ,
. Suppose
is a dual vector with components
with respect
to the
basis and components
with respect to the

basis. In particular this means we can

either write =
=1
or =
=1

. Likewise, given a vector we can either write

=
=1
or =
=1

. With these notations in mind calculate:

() =
_

=1
_
_

=1
_
=
,=1
) =
=1
=1
=1
and by the same calculation in the barred coordinates we nd, () =
=1

. Therefore,
=1
=1

.
Recall, = . In components,
=1
. Substituting,
=1
=1
=1

.
But, this formula holds for all possible vectors and hence all possible coordinate vectors . If we
consider =

then
hence
=1

=1

. Moreover,
=1
=1
=1
=1
=1
. Thus,
=1
. Compare how vectors and dual vectors

transform:

=1
verses
=1
(
1
)
.
It is customary to use lower-indices on the components of dual-vectors and upper-indices on the
components of vectors: we say =
=1
has contravariant components whereas

=
=1
has covariant components. These terms arise from the coordinate

change properties we derived in this section. The convenience of the up/down index notation will
be more apparent as we continue our study to more complicated objects. It is interesting to note
the basis elements tranform inversely:
=1
(
1
)
verses

=1
.
The formulas above can be derived by arguments similar to those we already gave in this section,
7.6. COORDINATE CHANGE 191
however I think it may be more instructive to see how these rules work in concert:
=
=1

=1
=1
(
1
)
(7.25)
=
=1
=1
(
1
)
=1
=1
=1
=1
(
1
)
=1
=1
=1
.
7.6.1 coordinate change for
0
2
( )
For an abstract vector space, or for
with a nonstandard basis, we have to replace , with their

coordinate vectors. If has basis

= {
1
,

2
, . . . ,

} with dual basis
= {
1
,

2
, . . . ,

} and
, have coordinate vectors , (which means =
=1

and =
=1

) then,
(, ) =
,=1

where

= (

). If = {
1
,
2
, . . . ,
} is another basis on with dual basis
then we
dene
= (
) and we have
(, ) =
,=1
.
Recall that

=1
. With this in mind calculate:
= (

) =
_

=1
=1
_
=
,=1
) =
,=1
We nd the components of a bilinear map transform as follows:
,=1
XXX- include general coordinate change and metrics with sylvesters theorem.
Chapter 8
manifold theory
In this chapter I intend to give you a fairly accurate account of the modern denition of a manifold
1
.
In a nutshell, a manifold is simply a set which allows for calculus locally. Alternatively, many people
say that a manifold is simply a set which is locally at, or it locally looks like
. This covers
most of the objects youve seen in calculus III. However, the technical details most closely resemble
the parametric view-point.
1
the denitions we follow are primarily taken from Burns and Gideas Dierential Geometry and Topology With
a View to Dynamical Systems, I like their notation, but you should understand this denition is known to many
authors
193
194 CHAPTER 8. MANIFOLD THEORY
8.1 manifolds
Denition 8.1.1.
We dene a smooth manifold of dimension as follows: suppose we are given a set ,
a collection of open subsets
of
, and a collection of mappings

which satises the following three criteria:
1. each map
is injective
2. if
= then there exists a smooth mapping
:
1
)
1
)
such that
3. =
)
Moreover, we call the mappings
the local parametrizations or patches of and the

space
is called the parameter space. The range
together with the inverse

1
is
called a coordinate chart on . The component functions of a chart (,
1
) are usually
denoted
1
= (
1
,
2
, . . . ,
) where
: for each = 1, 2, . . . , . .
We could add to this denition that is taken from an index set (which could be an innite
set). The union given in criteria (3.) is called a covering of . Most often, we deal with nitely
covered manifolds. You may recall that there are innitely many ways to parametrize the lines
or surfaces we dealt with in calculus III. The story here is no dierent. It follows that when we
consider classication of manifolds the denition we just oered is a bit lacking. We would also like
to lump in all other possible compatible parametrizations. In short, the denition we gave says a
manifold is a set together with an atlas of compatible charts. If we take that atlas and adjoin
to it all possible compatible charts then we obtain the so-called maximal atlas which denes a
dierentiable structure on the set . Many other authors dene a manifold as a set together
with a dierentiable structure. That said, our less ambtious denition will do.
We should also note that
=
1
hence
1
= (
1
)
1
=
1
. The functions
are called the transition functions of . These explain how we change coordinates locally.
I now oer a few examples so you can appreciate how general this denition is, in contrast to the
level-set denition we explored previously. We will recover those as examples of this more general
denition later in this chapter.
and suppose :
is the identity mapping ( () = for

all
) denes the collection of paramterizations on . In this case the collection is just one
mapping and = =
, clearly is injective and covers
. The remaining overlap criteria

is trivially satised since there is only one patch to consider.
8.1. MANIFOLDS 195
and suppose
then :
dened by () =
makes () = an -dimensional manifold. Again we have no overlap and the covering criteria
is clearly satised so that leaves injectivity of . Note () = (
) implies
hence
=
.
Example 8.1.4. Suppose is an -dimensional vector space over with basis = {
=1
.
Dene :
as follows, for each = (

1
,
2
, . . . ,
() =
1
1
+
2
2
+ +
.
Injectivity of the map follows from the linear independence of . The overlap criteria is trivially
satised. Moreover, () = thus we know that (
) = which means the vector space is

covered. All together we nd is an -dimensional manifold. Notice that the inverse of of the
coordinate mapping
from out earlier work and so we nd the coordinate chart is a coordinate

mapping in the context of a vector space. Of course, this is a very special case since most manifolds
are not spanned by a basis.
You might notice that there seems to be little contact with criteria two in the examples above.
These are rather special cases in truth. When we deal with curved manifolds we cannot avoid it
any longer. I should mention we can (and often do) consider other coordinate systems on
.
Moreover, in the context of a vector space we also have innitely many coordinate systems to
use. We will have to analyze compatibility of those new coordinates as we adjoin them. For the
vector space its simple to see the transition maps are smooth since theyll just be invertible linear
mappings. On the other hand, it is more work to show new curvelinear coordinates on
are
compatible with Cartesian coordinates. The inverse function theorem would likely be needed.
Example 8.1.5. Let = {(cos(), sin()) [0, 2)}. Dene
1
() = (cos() sin()) for all
(0, 3/2) =
1
. Also, dene
2
() = (cos() sin()) for all (, 2) =
2
. Injectivity
follows from the basic properties of sine and cosine and covering follows from the obvious geometry
of these mappings. However, overlap we should check. Let
1
=
1
(
1
) and
2
=
2
(
2
). Note
1

2
= {(cos(), sin()) < < 3/2}. We need to nd the formula for
12
:
1
2
(
1

2
)
1
1
(
1

2
)
In this example, this means
12
: (, 3/2) (, 3/2)
Example 8.1.6. Lets return to the vector space example. This time we want to allow for all
possible coordinate systems. Once more suppose is an -dimensional vector space over . Note
that for each basis = {
=1
. Dene
as follows, for each = (

1
,
2
, . . . ,
() =
1
1
+
2
2
+ +
.
Suppose ,
are bases for which dene local parametrizations
respective. The transition

functions :
are given by
=
1

Note is the composition of linear mappings and is therefore a linear mapping on
. It follows
that () = for some () = {

() = 0}. It follows that is a smooth
mapping since each component function of is simply a linear combination of the variables in
.
Lets take a moment to connect with linear algebra notation. If =
1
then
=
1
hence
as we used
as the coordinate chart in linear algebra and

1
.
Thus,
() =
() implies []
= []
. This matrix is the coordinate change matrix from

linear algebra.
The contrast of Examples 8.1.3 and 8.1.6 stems in the allowed coordinate systems. In Example
8.1.3 we had just one coordinate system whereas in Example 8.1.6 we allowed ininitely many. We
could construct other manifolds over the set . We could take all coordinate systems that are of a
particular type. If =
then it is often interesting to consider only those coordinate systems for

which the Pythagorean theorem holds true, such coordinates have transition functions in the group
of orthogonal transformations. Or, if =
4
then we might want to consider only inertially related
coordinates. Inertially related coordinates on
4
preserve the interval dened by the Minkowski
product and the transition functions form a group of Lorentz transformations. Orthogonal matrices
and Lorentz matrices are simply the matrices of the aptly named transformations. In my opinion
this is one nice feature of saving the maximal atlas concept for the dierentiable structure. Manifolds
as we have dened them give us a natural mathematical context to restrict the choice of coordinates.
From the viewpoint of physics, the maximal atlas contains many coordinate systems which are
unnatural for physics. Of course, it is possible to take a given theory of physics and translate
physically natural equations into less natural equations in non-standard coordinates. For example,
look up how Newtons simple equation

= is translated into rotating coordinate systems.
8.1. MANIFOLDS 197
8.1.1 embedded manifolds
The manifolds in Examples 8.1.2, 8.1.3 and 8.1.5 were all dened as subsets of euclidean space.
Generally, if a manifold is a subset of Euclidean space
then we say the manifold is embedded

in
. In contrast, Examples 8.1.4 and 8.1.6 are called abstract manifolds since the points in the
manifold were not found in Euclidean space
2
. If you are only interested in embedded manifolds
3
then the denition is less abstract:
Denition 8.1.7. embedded manifold.
We say is a smooth embedded manifold of dimension i we are given a set

such that at each there is a set an open subsets
of
and open subsets
of containing such that the mapping
satises the following

criteria:
1. each map
is injective
2. each map
is smooth
3. each map
1
is continuous
4. the dierential
has rank for each
.
You may identify that this denition more closely resembles the parametrized objects from your
multivariate calculus course. There are two key dierences with this denition:
1. the set
is assumed to be open in where
. This means that for each point
there exists and open -ball
such that contains . This is called the subspace

topology for induced from the euclidean topology of
. No topological assumptions were

given for
in the abstract denition. In practice, for the abstract case, we use the charts
to lift open sets to , we need not assume any topology on since the machinery of the
manifold allows us to build our own. However, this can lead to some pathological cases so
those cases are usually ruled out by stating that our manifold is Hausdor and the covering
has a countable basis of open sets
4
. I will leave it at that since this is not a topology course.
2. the condition that the inverse of the local parametrization be continuous and
be smooth
were not present in the abstract denition. Instead, we assumed smoothness of the transition
functions.
One can prove that the embedded manifold of Dentition 8.1.7 is simply a subcase of the abstract
manifold given by Denition 8.1.1. See Munkres Theorem 24.1 where he shows the transition
2
a vector space could be euclidean space, but it could also be a set of polynomials, operators or a lot of other
rather abstract objects.
3
The detion I gave for embedded manifold here is mostly borrowed from Munkres excellent text Analysis on
Manifolds where he primarily analyzes embedded manifolds
4
see Burns and Gidea page 11 in Dierential Geometry and Topology With a View to Dynamical Systems
functions of an embedded manifold are smooth. In fact, his theorem is given for the case of a
manifold with boundary which adds a few complications to the discussion. Well discuss manifolds
with boundary at the conclusion of this chapter.
Example 8.1.8. A line is a one dimensional manifold with a global coordinate patch:
() =
+
for all . We can think of this as the mapping which takes the real line and glues it in
along
some line which points in the direction and the new origin is at
. In this case :
and
has matrix which has rank one i = 0.

Example 8.1.9. A plane is a two dimensional manifold with a global coordinate patch: suppose
,

are any two linearly independent vectors in the plane, and
is a particular point in the plane,

(, ) =
for all (, )
2
. This amounts to pasting a copy of the -plane in
where we moved the

origin to
. If we just wanted a little paralellogram then we could restrict (, ) [0, 1] [0, 1],
then we would envision that the unit-square has been pasted on to a paralellogram. Lengths and
angles need not be maintained in this process of gluing. Note that the rank two condition for says
the derivative
(, ) = [
] = [
] must have rank two. But, this amounts to insisting the

vectors

,

are linearly independent. In the case of
3
this is conveniently tested by computation
of

which happens to be the normal to the plane.
Example 8.1.10. A cone is almost a manifold, dene
(, ) = ( cos(), sin(), )
for [0, 2] and 0. What two problems does this potential coordinate patch :
2
3
suer from? Can you nd a modication of which makes () a manifold (it could be a subset
of what we call a cone)
The cone is not a manifold because of its point. Generally a space which is mostly like a manifold
except at a nite, or discrete, number of singular points is called an orbifold. Recently, in the
past decade or two, the orbifold has been used in string theory. The singularities can be used to
t various charge to elds through a mathematical process called the blow-up.
Example 8.1.11. Let (, ) = (cos() cosh(), sin() cosh(), sinh()) for (0, 2) and .
This gives us a patch on the hyperboloid
2
+
2
2
= 1
Example 8.1.12. Let (, , , ) = (, , , cos(), sin()) for (0, 2) and (, , )
3
.
This gives a copy of
3
inside
5
where a circle has been attached at each point of space in the two
transverse directions of
5
. You could imagine that is nearly zero so we cannot traverse these
extra dimensions.
8.1. MANIFOLDS 199
Example 8.1.13. The following patch describes the Mobius band which is obtained by gluing a
line segment to each point along a circle. However, these segments twist as you go around the circle
and the structure of this manifold is less trivial than those we have thus far considered. The mobius
band is an example of a manifold which is not oriented. This means that there is not a well-dened
normal vectoreld over the surface. The patch is:
(, ) =
_
_
1 +
1
2
cos(
2
)
cos(),
_
1 +
1
2
sin(
2
)
sin(),
1
2
sin(
2
)
_
for 0 2 and 1 1. To understand this mapping better try studying the map evaluated
at various values of ;
(0, ) = (1 +/2, 0, 0), (, ) = (1, 0, /2), (2, ) = (1 /2, 0, 0)
Notice the line segment parametrized by (0, ) and (2, ) is the same set of points, however the
orientation is reversed.
Example 8.1.14. A regular surface is a two-dimensional manifold embedded in
3
. We need

2

3
such that, for each ,
,
has rank two for all (, )
. Moreover, in
this case we can dene a normal vector eld (, ) =
and if we visualize these vectors

as attached to the surface they will point in or out of the surface and provide the normal to the
tangent plane at the point considered. The surface is called orientable i the normal vector eld
is non-vanishing on . XXX, clean up normal idea
Example 8.1.15. Graphs
Example 8.1.16. ( )-dimensional level set.
8.1.2 manifolds dened by charts
Given patches which dene a manifold you can derive charts, conversely, given properly constructed
charts you can just as well nd patches and hence the manifold structure. I include these examples
to illustrate why charts are a nice starting point for certain examples. Recall that if :
is a local parametrization then

1
: is called a local coordinate chart. We seek
to translate the denition which we gave in patch-notation into chart notation for this subsection.
Throughout this subsection we denote
1
then we must insist, for all ,
where
is open in
satises
1. each map
is bijective
2. if
)
such that
.
3. =

For convenience of discussion we suppose the local parametrizations are also bijective. There is
no loss of generality since we can always make an injective map bijective by simply shrinking the
codomain. The original denition only assumed injectivity since that was sucient. Now that we
talk about inverse mappings it is convenient to add the supposition of surjectvity.
Example 8.1.17. Let = {(, )
2

2
+
2
= 1}.
1. Let
+
= {(, ) > 0} = (
+
) and dene
+
(, ) =
2. Let
= {(, ) < 0} = (
) and dene
(, ) =
3. Let
= {(, ) > 0} = (
) and dene
(, ) =
4. Let
= {(, ) < 0} = (
) and dene
(, ) =
The set of charts = {(
+
,
+
), (
), (
), (
)} forms an atlas on which gives

the circle a dierentiable structure
5
. It is not hard to show the transition functions are smooth on
the image of the intersection of their respect domains. For example,
+

=
+
= {(, )
, > 0}, its easy to calculate that
1
+
() = (,
1
2
) hence
(
1
+
)() =
(,
1
2
) =
1
2
for each
(
+
). Note
(
+
) implies 0 < < 1 hence it is clear the transition
function is smooth. Similar calculations hold for all the other overlapping charts. This manifold is
usually denoted =
1
.
A cylinder is the Cartesian product of a line and a circle. In other words, we can create a cylinder
by gluing a copy of a circle at each point along a line. If all these copies line up and dont twist
around then we get a cylinder. The example that follows here illustrates a more general pattern,
we can take a given manifold an paste a copy at each point along another manifold by using a
Cartesian product.
Example 8.1.18. Let = {(, , )
3

2
+
2
= 1}.
1. Let
+
= {(, , ) > 0} = (
+
) and dene
+
(, , ) = (, )
2. Let
= {(, , ) < 0} = (
) and dene
(, , ) = (, )
3. Let
= {(, , ) > 0} = (
) and dene
(, , ) = (, )
4. Let
= {(, , ) < 0} = (
) and dene
(, , ) = (, )
The set of charts = {(
+
,
+
), (
), (
), (
)} forms an atlas on which gives the

cylinder a dierentiable structure. It is not hard to show the transition functions are smooth on the
5
meaning that if we adjoin the innity of likewise compatible charts that denes a dierentiable structure on
8.1. MANIFOLDS 201
image of the intersection of their respective domains. For example,
+

=
+
= {(, , )
, > 0}, its easy to calculate that
1
+
(, ) = (,
1
2
, ) hence
(
1
+
)(, ) =
(,
1
2
, ) = (
1
2
, )
for each (, )
(
+
). Note (, )
(
+
) implies 0 < < 1 hence it is clear the
transition function is smooth. Similar calculations hold for all the other overlapping charts.
Generally, given two manifolds and we can construct by taking the Cartesian product
of the charts. Suppose
and
then you can dene

the product chart :
as =
. The Cartesian product together

with all such product charts naturally is given the structure of an ( + )-dimensional manifold.
For example, in the preceding example we took =
1
and = to consruct =
1
.
Example 8.1.19. The 2-torus, or donut, is constructed as
2
=
1
1
. The -torus is constructed
by taking the product of -circles:
=
1
1

1
. .

The atlas on this space can be obtained by simply taking the product of the
1
charts -times.
One of the surprising discoveries in manifold theory is that a particular set of points may have many
dierent possible dierentiable structures. This is why mathematicians often say a manifold is a
set together with a maximal atlas. For example, higher-dimensional spheres (
7
,
8
, ...) have more
than one dierentiable structure. In contrast,
for 6 has just one dierentiable structure.

XXX-add reference.
The most familar example of a manifold is just
2
or
3
itself. One may ask which coordinates
are in the atlas which contains the standard Cartesian coordinate chart. The most commonly used
charts other than Cartesian would probably be the spherical and cylindrical coordinate systems for
3
or the polar coordinate system for
2
. Technically, certain restrictions must be made on the
domain of these non-Cartesian coordinates if we are to correctly label them coordinate charts.
Interestingly, applications are greedier than manifold theorists, we do need to include those points
in
which spoil the injectivity of spherical or cylindrical coordinates. On the other hand, those
bad points are just the origin and a ray of points which do not contribute noticable in the calcula-
tion of a surface or volume integral.
I will not attempt to make explicit the domain of the coordinate charts in the following two examples
( you might nd them in a homework):
Example 8.1.20. Dene
(, , ) = (, , ) implicitly by the coordinate transformations

= cos() sin(), = sin() sin(), = cos()
These can be inverted,
=
2
+
2
+
2
, = tan
1
_
_
, = cos
1
_

2
+
2
+
2
_
To show compatibility with the standard Cartesian coordinates we would need to select a subset of
3
for which
is 1-1 and the since
= the transition functions are just

1
.
Example 8.1.21. Dene
(, , ) = (, , ) implicitly by the coordinate transformations

= cos(), = sin(), =
These can be inverted,
=
2
+
2
, = tan
1
_
_
, =
You can take (
) = {(, , ) 0 < < 2, } {(0, 0, 0)}

Remark 8.1.22.
I would encourage you to read Burns and Gidea and/or Munkres for future study. Youll
nd much more material, motivation and depth in those texts and if you were interested in
an independent study after youve completed real analysis feel free to ask.
8.1.3 dieomorphism
At the outset of this study I emphasized that the purpose of a manifold was to give a natural
languague for calculus on curved spaces. This denition begins to expose how this is accomplished.
Denition 8.1.23. smoothness on manifolds.
Suppose and are smooth manifolds and : is a function then we say is
smooth i for each there exists local parametrizations

and
such that
and
1
is a smooth mapping
from
to
. If : is a smooth bijection then we say is a dieomorphism.

Moreover, if is a dieomorphism then we say and are dieomorphic.
In other words, is smooth i its local coordinate representative is smooth. It suces to check one
representative since any other will be related by transition functions which are smooth: suppose
we have patches

and

such that

. .
.
=
1
. .
. .
. .
.
. .
. .
follows from the chain rule for mappings. This formula shows that if is smooth with respect to a
particular pair of coordinates then its representative will likewise be smooth for any other pair of
compatible patches.
8.1. MANIFOLDS 203
Example 8.1.24. Recall in Example 8.1.3 we studied = {
. Recall we have one

parametrization :
which is dened by () =
. Clearly
1
(
, ) = for all
(
, ) . Let
have Cartesian coordinates so the identity map is the patch for
. Consider
the function = :
, we have only the local coordinate representative

1

to
consider. Let

=
1
= .
Hence, is a smooth bijection from
to and we nd is dieomorphic to
The preceding example naturally generalizes to an arbitrary coordinate chart. Suppose is a

manifold and
1
:
is a coordinate chart around the point . We argue that

1
is a
dieomorphism. Once more take the Cartesian coordinate system for
and suppose

:

is a local parametrization with

. The local coordinate representative of
1
is simply the
transition function since:
1
.
We nd
1
is smooth on

. It follows that
1
is a dieomorphism since we know transition
functions are smooth on a manifold. We arrive at the following characterization of a manifold: a
manifold is a space which is locally dieomorphic to
.
However, just because a manifold is locally dieomorphic to
that does not mean it is actually

dieomorphic to
. For example, it is a well-known fact that there does not exist a smooth
bijection between the 2-sphere and
2
. The curvature of a manifold gives an obstruction to making
such a mapping.
8.2 tangent space
Since a manifold is generally an abstract object we would like to give a denition for the tangent
space which is not directly based on the traditional geometric meaning. On the other hand, we
should expect that the denition which is given in the abstract reduces to the usual geometric
meaning for the context of an embedded manifold. It turns out there are three common viewpoints.
1. a tangent vector is an equivalence class of curves.
2. a tangent vector is a contravariant vector.
3. a tangent vector is a derivation.
I will explain each case and we will nd explicit isomorphisms between each language. We assume
that is an -dimensional smooth manifold throughout this section.
8.2.1 equivalence classes of curves
I essentially used case (1.) as the denition for the tangent space of a level-set. Suppose :
is a smooth curve with (0) = . In this context, this means that all the local
coordinate representatives of are smooth curves on
; that is, for every parametrization :

containing the mapping

6
is a smooth mapping from to
. Given the coordinate system dened by

1
, we dene two smooth curves
1
,
2
on with
1
(0) =
2
(0) = to be similar at i (
1
1
)
(0) = (
1
2
)
(0). If smooth curves
1
,
2
are similar at then we denote this by writing
1

2
. We insist the curves be
parametrized such that they reach the point of interest at the parameter = 0, this is not a severe
restriction since we can always reparametrize a given curve which reaches at =
by replacing
the parameter with
. Observe that
denes an equivalence relation on the set of smooth

curves through which reach at = 0 in their domain.
(i) reexive:
i (0) = and (
1
(0) = (
1
(0). If is a smooth curve on

with (0) = then clearly the reexive property holds true.
(ii) symmetric: Suppose
1

2
then (
1
1
)
(0) = (
1
2
)
(0) hence (
1
2
)
(0) =
(
1
1
)
(0) and we nd
2

1
thus
is a symmetric relation.
(iii) transitive: if
1

2
and
2

3
then (
1
1
)
(0) = (
1
2
)
(0) and (
1
2
)
(0) =
(
1
3
)
(0) thus (
1
1
)
(0) = (
1
3
)
(0) which shows

1

3
.
The equivalence classes of
partition the set of smooth curves with (0) = . Each equivalence

class = { :
} corresponds uniquely to a particular vector (

1
(0) in
6
Note, we may have to restrict the domain of
1
such that the image of falls inside , keep in mind this

poses no threat to the construction since we only consider the derivative of the curve at zero in the nal construction.
That said, keep in mind as we construct composites in this section we always suppose the domain of a curve includes
some nbhd. of zero. We need this assumption in order that the derivative at zero exist.
8.2. TANGENT SPACE 205
. Conversely, given = (
1
,
2
, . . . ,
we can write the equation for a line in
with
direction and base-point
1
():
() =
1
() +.
We compose with to obtain a smooth curve through which corresponds to the vector .
In invite the reader to verify that =
has
(1.) (0) = (2.) (
1
(0) = .
Notice that the correspondence is made between a vector in
and a whole family of curves.

There are naturally many curves that share the same tangent vector to a given point.
Moreover, we show these equivalence classes are not coordinate dependent. Suppose
rel-
ative to the chart
1
: , with . In particular, we suppose (0) = (0) = and
(
1
(0) = (
1
(0). Let

1
:

, with

, we seek to show
relative to the
chart

1
. Note that

hence, by the chain rule,

(
(0) = (
(
1
())(
1
(0)
Likewise, (
(0) = (
(
1
())(
1
(0). Recognize that (
(
1
()) is an in-
vertible matrix since it is the derivative of the invertible transition functions, label (
(
1
()) =
to obtain:
(
(0) = (
1
(0) and (
(0) = (
1
(0)
the equality (
(0) = (
(0) follows and this shows that
relative to the

coordi-
nate chart. We nd that the equivalence classes of curves are independent of the coordinate system.
With the analysis above in mind we dene addition and scalar multiplication of equivalence classes
of curves as follows: given a coordinate chart
1
: with , equivalence classes
1
,
2
at and
, if
1
has (
1
1
)
(0) =
1
in
and
2
has (
1
2
)
(0) =
2
in
then we
dene
(i)
1
+
2
= where (
1
(0) =
1
+
2
(ii)
1
=

where (
1
(0) =
1
.
We know and exist because we can simply push the lines in
based at
1
() with directions
1
+
2
and
1
up to to obtain the desired curve and hence the required equivalence class.
Moreover, we know this construction is coordinate independent since the equivalence classes are
indpendent of coordinates.
Denition 8.2.1.
Suppose is an -dimensional smooth manifold. We dene the tangent space at
to be the set of
-equivalence classes of curves. In particular, denote:
= { smooth and (0) = }

Keep in mind this is just one of three equivalent denitions which are commonly implemented.
8.2.2 contravariant vectors
If (
(0) = and (
1
(0) = then = where = (
(
1
()). With this in
mind we could use the pair (, ) or (, ) to describe a tangent vector at . The cost of using (, )
is it brings in questions of coordinate dependence.
The equivalence class viewpoint is at times quite useful, but the denition of vector oered here is
a bit easier in certain respects. In particular, relative to a particular coordinate chart
1
: ,
with , we dene (temporary notation)
= {(, )
}
Vectors are added and scalar multiplied in the obvious way:
(,
1
) + (,
2
) = (,
1
+
2
) and (,
1
) = (,
1
)
for all (,
1
, (,
2
)
and . Moreover, if we change from the

1
chart to the

1
co-
ordinate chart then the vector changes form as indicated in the previous subsection; (, ) (, )
where = and = (
(
1
()). The components of (, ) are said to transform
contravariantly.
Technically, this is also an equivalence class construction. A more honest notation would be to
replace (, ) with (, , ) and then we could state that (, , ) (,

, ) i = and =
(
(
1
()). However, this notation is tiresome so we do not pursue it further. I prefer the
notation of the next viewpoint.
8.2.3 derivations
To begin, let us dene the set of locally smooth functions at :
() = { : is smooth on an open set containing }

In particular, we suppose
() to mean there exists a patch : such that is

smooth on . Since we use Cartesian coordinates on by convention it follows that :
smooth indicates the local coordinate representative

: is smooth (it has continuous
partial derivatives of all orders).
Denition 8.2.2.
Suppose
() is a linear transformation which satises the Leibniz rule then

we say
is a derivation on
(). Moreover, we denote
( + ) =
() +
() and
() = ()
() +
()() for all ,
() and .
Example 8.2.3. Let = and consider
= /
. Clearly is a derivation on smooth

functions near
.
Example 8.2.4. Consider =
2
. Pick = (
) and dene
and
.
Once more it is clear that
()
2
. These derivations action is accomplished by partial
dierentiation followed by evaluation at .
Example 8.2.5. Suppose =
. Pick
and dene =

. Clearly this is a
derivation for any
.
Are the other types of derivations? Is the only thing a derivation is is a partial derivative operator?
Before we can explore this question we need to dene partial dierentiation on a manifold. We
should hope the denition is consistent with the langauge we already used in multivariate calculus
(and the preceding pair of examples) and yet is also general enough to be stated on any abstract
smooth manifold.
Denition 8.2.6.
Let be a smooth -dimensional manifold and let : be a local parametrization
with . The -th coordinate function
: is the -component function of
1
: . In other words:
1
() = () = (
1
(),
2
(), . . . ,
())
These
are manifold coordinates. In constrast, we will denote the standard Cartesian

coordinates in
via
so a typical point has the form (

1
,
2
, . . . ,
) and viewed
as functions
where
() =
() =
. We dene the partial derivative

with respect to
at for
() as follows:
() =

_
(

)()
_
=
1
()
=

1
_
()
.
The idea of the dention is simply to take the function with domain in then pull it back to
a function

1
:
on
. Then we can take partial derivatives of

1
in the same way we did in multivariate calculus. In particular, the partial derivative w.r.t.
is
calculated by:
() =

_
_
_
(() +
)
_
=0
which is precisely the directional derivative of

1
in the -direction at (). In fact, Note
_
_
(() +
) = (
1
(() +
)).
The curve
1
(() +
) is the curve on through where all coordinates are xed except

the -coordinate. It is a coordinate curve on .
XXX-need a picture of coordinate curves on manifold!
Notice in the case that =
is given Cartesian coordinate = then

1
= as well and
the
1
(() +
) reduces to +
which is just the -th coordinate curve through on
. It follows that the partial derivative dened for manifolds naturally reduces to the ordinary
partial derivative in the context of =
with Cartesian coordinates. The beautiful thing is

that almost everything we know for ordinary partial derivatives equally well transfers to

.
Theorem 8.2.7. Partial dierentiation on manifolds
Let be a smooth -dimensional manifold with coordinates
1
,
2
, . . . ,
near . Fur-
thermore, suppose coordinates
1
,
2
, . . . ,
are also dened near . Suppose ,
()
and then:
1.

_
+
2.

3.

= ()

()
4.

5.
=1
6.

=1
Proof: The proof of (1.) and (2.) follows from the calculation below:
( +)
() =

_
( +)
1
_
()
=

1
+
1
_
()
=

1
_
()
+

1
_
()
=

() +

() (8.1)
The key in this argument is that composition ( + )
1
=

1
+
1
along side the
linearity of the partial derivative. Item (3.) follows from the identity ()
1
= (

1
)(
1
)
in tandem with the product rule for a partial derivative on
. The reader may be asked to complete

the argument for (3.) in the homework. Continuing to (4.) we calculate from the denition:
_
(
1
)()
_
()
=

()
=
.
where the last equality is known from multivariate calculus. In invite the reader to prove it from
the denition if unaware of this fact. Before we prove (5.) it helps to have a picture and a bit
more notation in mind. Near the point we have two coordinate charts :
and
:
, we take the chart domain to be small enough so that both charts are
dened. Denote Cartesian coordinates on by
1
,
2
, . . . ,
and for we likewise use Cartesian

coordinates
1
,
2
, . . . ,
. Let us denote patches , as the inverses of these charts;

1
=
and
1
= . Transition functions
1
1
are mappings from
to
and
we note
_
(
1
)()
_
=

Likewise, the inverse transition functions

1
1
are mappings from
to
_
(
1
)()
_
=

Recall that if , :
and

= then
= by the chainrule, hence (
)
1
=
.
Apply this general fact to the transition functions, we nd their derivative matrices are inverses.
Item (5.) follows. In matrix notation we item (5.) reads

= . Item (6.) follows from:
_
(

1
)()
_
()
=

_
(

1
)()
_
()
=

_
_
1
_
(
1
(), . . . ,
())
_
()
: where
() = (
1
)
()
=
=1
(
1
)
()
(

1
)
1
)(())
: chain rule
=
=1
(
1
)
()
(

1
)
()
=
=1

The key step was the multivariate chain rule.
This theorem proves we can lift calculus on
to in a natural manner. Moreover, we should

note that items (1.), (2.) and (3.) together show

is a derivation at . Item (6.) should remind

the reader of the contravariant vector discussion. Removing the from the equation reveals that
=1
A notation convenient to the current discussion is that a contravariant transformation is (,
)
(,
) where
and = (
1
)
(()) =

()
. Notice this is the inverse of what we see
in (6.). This suggests that the partial derivatives change coordinates like as a basis for the tangent
space. To complete this thought we need a few well-known propositions for derivations.
Proposition 8.2.8. derivations on constant function gives zero.
If
() is a constant function and
then
() = 0.
Proof: Suppose () = for all , dene () = 1 for all and note = on . Since
is a derivation is satises the Leibniz rule hence
() =
() = ()
() +()() =
() +
()
() = 0.
Moreover, by homogeneity of
, note
() =
() =
(). Thus,
() = 0.
Proposition 8.2.9.
If ,
() and () = () for all and
then
() =
().
Proof: Note that () = () implies () = () () = 0 for all . Thus, the previous
proposition yields
() = 0. Thus,
( ) = 0 and by linearity
()
() = 0. The
proposition follows.
Proposition 8.2.10.
Suppose
and is a chart dened near ,
=1
Proof: this is a less trivial proposition. We need a standard lemma before we begin.
Lemma 8.2.11.
Let be a point in smooth manifold and let : be a smooth function. If
: is a chart with and () = 0 then there exist smooth functions
:
whose values at satisfy
() =

(). In addition, for all near enough to we have

() = () +
=1
()
()
Proof: follows from proving a similar identity on
then lifting to the manifold. I leave this as a

nontrivial exercise for the reader. This can be found in many texts, see Burns and Gidea page 29
for one source.
Consider
(), and use the lemma, we assume () = 0 and
() =

():
() =
_
() +
=1
()
()
_
=
(()) +
=1
()
()
_
=
=1
_
() +
()
())
_
=
=1
().
The calculation above holds for arbitrary
() hence the proposition follows.

Weve answered the question posed earlier in this section. It is true that every derivation of a
manifold is simply a linear combination of partial derivatives. We can say more. The set of deriva-
tions at naturally forms a vector space under the usual addition and scalar multiplication of
operators: if
then we dene
by (
)() =
() +
() and
by
(
)() =
() for all ,
() and . It is easy to show
is a vector space under

these operations. Moreover, the preceding proposition shows that
= {
=1
hence
is an -dimensional vector space

7
.
Finally, lets examine coordinate change for derivations. Given two coordinate charts , at
we have two ways to write the derivation
=1
or
=1
7
technically, we should show the coordinate derivations

are linearly independent to make this conclusion. I

dont suppose weve done that directly at this juncture. You might nd this as a homework
It is simple to connect these formulas. Whereas, for -coordinates,
) =
=1
(8.2)
This is the contravariant transformation rule. In contrast, recall

=1
. We
should have anticipated this pattern since from the outset it is clear there is no coordinate depen-
dence in the denition of a derivation.
8.2.4 dictionary between formalisms
We have three competing views of how to characterize a tangent vector.
1.
= { smooth and (0) = }

2.
= {(, )
}
3.
Perhaps it is not terribly obvious how to get a derivation from an equivalence class of curves.
Suppose is a tangent vector to at and let ,
(). Dene an operator
associated to
via
() = (

)
(0). Consider, ( +)
)() = ( +)(()) = (())+(()) dierentiate

at set = 0 to verify that
( +)() = ( +)
(0) =
()()+
(). Furthermore, observe

that (()
)() = (())(()) therefore by the product rule from calculus I,

(()
() = (

)
()(()) +(())(
()
hence, noting (0) = we verify the Leibniz rule for
() = (()
(0) = (

)
(0)() +()(
(0) =
()() +()
()
In view of these calculations we nd that :
dened by ( ) =
is
well-dened. Moreover, we can show is an isomorphism. To be clear, we dene:
( )() =
() = (

)
(0).
Ill begin with injectivity. Suppose ( ) = (
) then for all
() we have ( )() = (
)()
hence (

)
(0) = (

)
(0) for all smooth functions at . Take = : and it follows

that
hence =

and we have shown is injective. Linearity of must be judged on
the basis of our denition for the addition of equivalence classes of curves. I leave linearity and
surjectivity to the reader. Once those are established it follows that is an isomorphism and
.
The isomorphism between
and
was nearly given in the previous subsection. Es-

sentially we can just paste the components from
onto the partial derivative basis for

derivations. Dene :
for each (,
, relative to coordinates
at ,
_
,
=1
_
=
=1
Note that if we used a dierent chart then (,
) (,
) and consequently
_
,
=1
_
=
=1
=1
.
Thus is single-valued on each equivalence class of vectors. Furthermore, the inverse mapping is
simple to write: for a chart at ,
1
(
) = (,
=1
)
and the value of the mapping above is related contravariantly if we were to use a dierent chart
1
(
) = (,
=1
).
See Equation 8.2 and the surrounding discussion if you forgot. It is not hard to verify that
is bijective and linear thus is an isomorphism. We have shown
. Let us
summarize:
Sorry to be anticlimatic here, but we choose the following for future use:
Denition 8.2.12. tangent space
We denote
.
8.3 the dierential
In this section we generalize the concept of the dierential to the context of manifolds. Recall that
for :
the dierential
was a linear transformation which best

approximated the change in near . Notice that while the domain of could be a mere subset
of
the dierential always took all of
as its domain. This suggests we should really think

of the dierential as a mapping which transports tangent vectors to to tangent vectors at .
Often
is called the push-forward by at because it pushes tangent vectors along side the
mapping.
Denition 8.3.1. dierential for manifolds.
Suppose and are smooth manifolds of dimension and respective. Furthermore,
suppose : is a smooth mapping. We dene
()
as follows: for
each
and
(())
)() =
).
Notice that : () and consequently
: and it follows
() and it is natural to nd
in the domain of
. In addition, it is not hard to show
)
()
. Observe:
1.
)( +) =
(( +)
) =
) =
) +
)
2.
)() =
(()
) =
)) =
)) =
)()
The proof of the Leibniz rule is similar. Now that we have justied the denition lets look at an
interesting application to the study of surfaces in
3
.
Suppose
3
is an embedded two-dimensional manifold. In particular suppose is a regular
surface which means that for each parametrization : the normal vector eld (, ) =
(
)(, ) is a smooth non-vanishing vector eld on . Recall that the unit-sphere

2
= {
3
= 1} is also manifold, perhaps you showed this in a homework. In any event, the mapping
:
2
dened by
(, ) =

provides a smooth mapping from the surface to the unit sphere. The change in measures how
the normal deects as we move about the surface . One natural scalar we can use to quantify
that curving of the normal is called the Gaussian curvature which is dened by = ().
Likewise, we dene = () which is the mean curvature of . If
1
,
2
are the eigen-
values the operator
then it is a well-known result of linear algebra that (
) =
1
2
and
(
) =
1
+
2
. The eigenvalues are called the principal curvatures. Moreover, it can be
shown that the matrix of
is symmetric and a theorem of linear algebra says that the eigenvalues

are real and we can select an orthogonal basis of eigenvectors for
.
Example 8.3.2. Consider the plane with base point
and containing the vectors

,

, write
(, ) =
to place coordinates , on the plane . Calculate the Gauss map,

(, ) =

This is constant on hence
= 0 for each . The curvatures (mean, Gaussian and

principles) are all zero for this case. Makes sense, a plane isnt curved!
Let me outline how to calculate the curvature directly when is not trivial. Calculate,
_
(
) =

_
=
(
Thus, using the discussion of the preceding section,
_
=
2
=1
(
Therefore, the matrix of
is the 22 matrix
_
(
_
with respect to the choice of coordinates
1
,
2
on and
1
,
2
on the sphere.
Example 8.3.3. Suppose (, ) = ( , ,
2
) parameterizes part of a sphere
of
radius > 0. You can calculate the Gauss map and the result should be geometrically obvious:
(, ) =
1
_
, ,
2
_
Then the and components of (, ) are simply / and / respective. Calculate,
[
] =
_

]
_
=
_
1
0
0
1
_
Thus the Gaussian curvature of the sphere = 1/
2
. The principle curvatures are
1
=
2
= 1/
and the mean curvature is simply = 2/. Notice that as we nd agreement with the
curvature of a plane.
Example 8.3.4. Suppose is a cylinder which is parametrized by (, ) = (cos , sin , ).
The Gauss map yields (, ) = (cos , sin , 0). I leave the explicit details to the reader, but it can
be shown that
1
= 1/,
2
= 0 and hence = 0 whereas = 1/.
The dierential is actually easier to frame in the equivalence class curve formulation of
. In
particular, suppose = [] as a more convenient notation for what follows. In addition, suppose
: is a smooth function and []
then we dene
()
as follows:
([]) = [

]
There is a chain-rule for dierentials. Its the natural rule youd expect. If : and
: then, denoting = (),
) =
.
The proof is simple in the curve notation:
_
_
([]) =
([])
_
=
([

]) = [
(

)] =
)[].
You can see why the curve formulation of tangent vectors is useful. It does simply certain questions.
That said, we will insist
in sequel.
The push-forward need not be an abstract
8
exercise.
2
,

2
,
is the polar coordinate transformation. In particular,
(, ) = ( cos , sin )
Lets examine where pushes

and

. We use (, ) as coordinates in the codomain and

the problem is to calculate, for
or
, the values of
)() and
)()
as we know
) =
)()
)()
where = ().
) =
( cos )
( sin )
= cos
+ sin
A similar calculation follows for
). However, let me do the calculation in the traditional

notation from multivariate calculus. If we denote = (, ) and drop the point dependence on the
partials we nd the formulas for / below:
= sin

+ cos

.
Therefore, the push-forward is a tool which we can use to change coordinates for vectors. Given
the coordinate transformation on a manifold we just push the vector of interest presented in one
coordinate system to the other through the formulas above. In multivariate calculus we simply
thought of this as changing notation on a given problem. I would be good if you came to the same
understanding here.
8
its not my idea of abstract that is wrong... think about that.
8.4. COTANGENT SPACE 217
8.4 cotangent space
The tangent space to a smooth manifold is a vector space of derivations and we denote it by
. The dual space to this vector space is called the cotangent space and the typical elements
are called covectors.
Denition 8.4.1. cotangent space
Suppose is a smooth manifold and
is the tangent space at . We dene,
= {
}.
If is a local coordinate chart at and

, . . . ,

is a basis for
then we denote
the dual basis
1
,
2
, . . . ,
where
_
=
. Moreover, if is a covector at then

9
:
=
=1
where
=
_

_
and
is a short-hand for
. We should understand that covectors

are dened at a point even if the point is not explicitly indicated in a particular context. This
does lead to some ambiguity in the same way that the careless identication of the function and
its value () does throughout calculus. That said, an abbreviated notation is often important to
help us see through more dicult patterns without getting distracted by the minutia of the problem.
You might worry the notation used for the dierential and our current notation for the dual basis
of covectors is not consistent. After all, we have two rather dierent meanings for
at this time:
1.
: is a smooth function hence
()
is dened as a push-forward,
)() =
)
2.
where
_
=
It is customary to identify
()
with hence there is no trouble. Let us examine how the
dual-basis condition can be derived for the dierential, suppose : hence
: ,
_
() =

((
)) =

()
. .

=
()
_

_
=
Where, weve made the identication 1 =

()
( which is the nut and bolts of writing
()
=
) and hence have the beautiful identity:
_
=
.
9
we explained this for an arbitrary vector space and its dual
in a previous chapter, we simply apply those

results once more here in the particular context =
In contrast, there is no need to derive this for case (2.) since in that context this serves as the
denition for the object. Personally, I nd the multiple interpretations of objects in manifold theory
is one of the most dicult aspects of the theory. On the other hand, the notation is really neat
once you understand how subtly it assumes many theorems. You should understand the notation
we enjoy at this time is the result of generations of mathematical thought. Following a similar
derivation for an arbitrary vector
and : we nd
) =
()
This notation is completely consistent with the total dierential as commonly discussed in multi-
variate calculus. Recall that if :
then we dened
=

1
+

2
+ +

.
Notice that the -th component of is simply

. Notice that the identity
) =
()
gives us the same component if we simply evaluate the covector
on the coordinate basis

_
=

8.5 tensors at a point

Given a smooth -dimensional manifold and a point we have a tangent space
and
a cotangent space
. The set of tensors at is simply the set of all multilinear mappings

on the tangent and cotangent space at . We again dene the set of all type (, ) tensors to be
meaning
i is a multilinear mapping of the form

:
. .

. .

.
Relative to a particular coordinate chart at we can build a basis for
via the tensor

product. In particular, for each
there exist constants
2
...
2
...
such that
1
,...,,
1
,...,=1
(
2
...
2
...
)()
.
The components can be calculated by contraction with the appropriate vectors and covectors:
(
2
...
2
...
)() =
_

, . . . ,

1
, . . . ,
_
.
We can summarize the equations above with multi-index notation:
and

8.6. TENSOR FIELDS 219

Consequently,
()
. We may also construct wedge products and build the

exterior algebra as we did for an arbitrary vector space. Given a metric

0
2
we can calculate
hodge duals in
. All these constructions are possible at each point in a smooth manifold

10
.
8.6 tensor elds
Since the tangent and cotangent space are dened at each point in a smooth manifold we can
construct the tangent bundle and cotangent bundle by simply taking the union of all the tangent
or cotangent spaces:
Denition 8.6.1. tangent and cotangent bundles.
Suppose is a smooth manifold the we dene the tangent bundle and the cotan-
gent bundle
as follows:
=
and
.
The cannonical projections , tell us where a particular vector or covector are found on the
manifold:
(
) = and (
) =
I usually picture this construction as follows:
XXX add projection pictures
Notice the bers of and are
1
() =
and
1
() =
. Generally a ber bundle

(, , ) consists of a base manifold , a bundle space and a projection
: . A local section of is a mapping : such that
is injective.
In other words, the image of a section hits each ber over its domain just once. A section selects
a particular element of each ber. Heres an abstract picture of section, I sometimes think of the
section as its image although technically the section is actually a mapping:
XXX add section picture
Given the language above we nd a natural langauge to dene vector and covector-elds on a
manifold. However, for reasons that become clear later, we call a covector-eld a dierential one-
form.
Denition 8.6.2. tensor elds.
10
I assume is Hausdor and has a countable basis, see Burns and Gidea Theorem 3.2.5 on page 116.
Let , we dene:
1. is a vector eld on i is a section of on
2. is a dierential one-form on i is a section of
on .
3. is a type (, ) tensor-eld on i is a section of
on .
We consider only smooth sections and it turns out this is equivalent
11
to the demand that the
component functions of the elds above are smooth on .
8.7 metric tensor
Ill begin by discussing briey the informal concept of a metric. The calculations given in the rst
part of this section show you how to think for nice examples that are embedded in
. In such
cases the metric can be deduced by setting appropriate terms for the metric on
to zero. The
metric is then used to set-up arclength integrals over a curved space, see my Chapter on Varitional
Calculus from the previous notes if you want examples.
In the second part of this chapter I give the careful denition which applies to an arbitrary manifold.
I include this whole section mostly for informational purposes. Our main thrust in this course is
with the calculus of dierential forms and the metric is actually, ignoring the task of hodge duals,
not on the center stage. That said, any student of dierential geometry will be interested in the
metric. The problem of paralell transport
12
, and the denition and calculation of geodesics
13
are
fascinating problems beyond this course.
8.7.1 classical metric notation in
Denition 8.7.1.
The Euclidean metric is
2
=
2
+
2
+
2
. Generally, for orthogonal curvelinear
coordinates , , we calculate
2
=
1
2
+
1
2
+
1
2
.
The beauty of the metric is that it allows us to calculate in other coordinates, consider
= cos() = sin()
For which we have implicit inverse coordinate transformations
2
=
2
+
2
and = tan
1
(/).
From these inverse formulas we calculate:
= < /, / > = < /
2
, /
2
>
11
all the bundles above are themselves manifolds, for example is a 2-dimensional manifold, and as such the
term smooth has already been dened. I do not intend to delve into that aspect of the theory here. See any text on
manifold theory for details.
12
how to move vectors around in a curved manifold
13
curve of shortest distance on a curved space, basically they are the lines on a manifold
8.7. METRIC TENSOR 221
Thus, = 1 whereas = 1/. We nd that the metric in polar coordinates takes the form:
2
=
2
+
2
2
Physicists and engineers tend to like to think of these as arising from calculating the length of
innitesimal displacements in the or directions. Generically, for , , coordinates
=
1
=
1
=
1
and
2
=
2
+
2
+
2
. So in that notation we just found
= and
= . Notice then
that cylindircal coordinates have the metric,
2
=
2
+
2
2
+
2
.
For spherical coordinates = cos() sin(), = sin() sin() and = cos() (here 0 2
and 0 , physics notation). Calculation of the metric follows from the line elements,
= sin()
=
Thus,
2
=
2
+
2
sin
2
()
2
+
2
2
.
We now have all the tools we need for examples in spherical or cylindrical coordinates. What about
other cases? In general, given some -manifold embedded in
how does one nd the metric on

that manifold? If we are to follow the approach of this section well need to nd coordinates on
such that the manifold is described by setting all but of the coordinates to a constant.
For example, in
4
we have generalized cylindircal coordinates (, , , ) dened implicitly by the
equations below
= cos(), = sin(), = , =
On the hyper-cylinder = we have the metric
2
=
2
2
+
2
+
2
. There are mathemati-
cians/physicists whose careers are founded upon the discovery of a metric for some manifold. This
is generally a dicult task.
8.7.2 metric tensor on a smooth manifold
A metric on a smooth manifold is a type (2, 0) tensor eld on which is at each point a
metric on
. In particular, is a metric i makes the assignment
for each where

the mapping
is a metric. Recall the means
is a symmetric, nondegenerate
bilinear form on
. Relative to a particular coordinate system at we write

=
In this context
: are assumed to be smooth functions, the values may vary from point to
point in . Furthermore, we know that
for all ,
and the matrix [
] is invertible
by the nondegneracy of . Recall we use the notation
for components of the inverse matrix, in

particular we suppose that
=1
.
Recall that according to Sylvesters theorem we can choose coordinates at some point which
will diagonalize the metric and leave (
) = {1, 1, . . . , 1, 1, 1, . . . , 1}. In other words, we

can orthogonalize the coordinate basis at a paricular point . The interesting feature of a curved
manifold is that as we travel away from the point where we straightened the coordinates it is
generally the case the components of the metric will not stay diagonal and constant over the whole
coordinate chart. If it is possible to choose coordinates centered on such that the coordinates are
constantly orthogonal with respect the metric over then the manifold is said to be at on .
Examples of at manifolds include
, cylinders and even cones without their point. A manifold

is said to be curved if it is not at. The denition I gave just now is not probably one youll nd
in a mathematics text
14
. Instead, the curvature of a manifold is quantied through various tensors
which are derived from the metric and its derivatives. In particular, the Ricci and Riemann tensors
are used to carefully characterize the geometry of a manifold. It is very tempting to say more
about the general theory of curvature, but I will resist. If you would like to do further study I can
recommend a few books. We will consider some geometry of embedded two-dimensional manifolds
in
3
. That particular case was studied in the 19-th century by Gauss and others and some of the
notation below goes back to that time.
Example 8.7.2. Consider a regular surface which has a global parametrization :
2

3
. In the usual notation in
3
,
(, ) = ((, ), (, ), (, ))
Consider a curve : [0, 1] we can calculate the arclength of via the usal calculation in
3
.
The magnitude of velocity
() is
() and naturally this gives us

hence =
() and
the following integral calculates the length of ,
1
0
()
Since [0, 1] it follows there must exist some two-dimesional curve ((), ()) for which
() = ((), ()). Observe by the chain rule that
() =
_
_
We can calculate the square of the speed in view of the formula above, let

= and

= ,
()
2
=
_

2

2
+ 2
+
2

2
,

2
+ 2
+
2

2
,

2
+ 2
+
2

2
_
(8.3)
14
this was the denitin given in a general relativity course I took with the physicisist Martin Rocek of SUNY Stony
Brook. He then introduced non-coordinate form-elds which kept the metric constant. I may nd a way to show you
some of those calculations at the end of this course.
8.7. METRIC TENSOR 223
Collecting together terms which share either
2
, or
2
and noting that
2
+
2
+
2
and
2
+
2
+
2
we obtain:
()
2
=

2
+

2
Or, in the notation of Gauss,
= ,
= and
= hence the arclength on is

given by
1
0

2
+ 2 +
2
We discover that on there is a metric induced from the ambient euclidean metric. In the current
coordinates, using (, ) =
1
,
= + 2 +
hence the length of a tangent vector is dened via =
(, ), we calcate the length of a

curve by integrating its speed along its extent and the speed is simply the magnitude of the tangent
vector at each point. The new thing here is that we judge the magnitude on the basis of a metric
which is intrinsic to the surface.
If arclength on is given by Gauss , , then what about surface area?. We know the magnitude
of the cross product of the tangent vectors
on will give us the area of a tiny paralellogram

corresponding to a change in and in . Thus:
=
However, Lagranges identity says
2
=
hence =
and we can calculate surface area (if this integral exists!) via
() =
2
.
I make use of the standard notation for double integrals from multivariate calculus and the integra-
tion is to be taken over the domain of the parametrization of .
Many additional formulas are known for , , and there are entire texts devoted to exploring
the geometric intracies of surfaces in
3
. For example, John Opreas Dierential Geometry and
its Applications. Theorem 4.1 of that text is the celebrated Theorem Egregium of Gauss which
states the curvature of a surface depends only on the metric of the surface as given by , , . In
particular,
=
1
2
_
+

__
.
Where curvature at is dened by () = (
) and
is the shape operator is dened

by the covariant derivative
() =
= ((
1
), (
2
), (
3
)) and is simply the normal
vector eld to dened by (, ) =
in our current notation.

It turns out there is an easier way to calculate curvature via wedge products. I will hopefully show
how that is done in the next chapter. However, I do not attempt to motivate why the curvature is
called curvature. You really should read something like Oprea if you want those thoughts.
4
and choose an atlas of charts which are all intertially related to
the standard Cartesian coordinates on
4
. In other words, we allow coordinates which can be
obtained from a Lorentz transformation; = and
44
such that
= . Dene
=
3
,=0
for the standard Cartesian coordinates on

4
. We can show that the
metric is invariant as we change coordinates, if you calculate the components of in some other
coordinate system then you will once more obtain
as the components. This means that if we

can write the equation for the interval between events in one coordinate system then that inter-
val equation must also hold true in any other inertial coordinate system. In particle physics this
is a very useful observation because it means if we want to analyze an relativistic interaction then
we can study the problem in the frame of reference which makes the problem simplest to understand.
In physics a coordinate system if also called a frame of reference, technically there is something
missing from our construction of from a relativity perspective. As a mathematical model of
spacetime
4
is not quite right. Why? Because Einsteins rst axiom or postulate of special relativity
is that there is no preferred frame of reference. With
4
there certainly is a preferred frame, its
impicit within the very denition of the set
4
, we get Cartesian coordinates for free. To eliminate
this convenient set of, according to Einstein, unphysical coordinates you have to consider an ane
space which is dieomorphic to
4
. If you take modern geometry youll learn all about ane space.
I will not pursue it further here, and as a bad habit I tend to say paired with is minkowski
space. Technically this is not quite right for the reasons I just explained.
8.8 on boundaries and submanifolds
A manifold with boundary
15
is basically just a manifold which has an edge which is also a manifold.
The boundary of a disk is a circle. In fact, in general, a closed ball in (+1)-dimensional euclidean
space has a boundary which is the
sphere.
The boundary of quadrants I and II of the -plane is the -axis. Or, to generalize this example,
we dene the upper-half of
as follows:

= {(
1
,
2
, . . . ,
1
,
0}.
The boundary of

is the
1
2

1
-hyperplane which is the solution set of
= 0 in
; we
can denote the boundary by

hence,

=
1
{0}. Furthermore, we dene

+
= {(
1
,
2
, . . . ,
1
,
> 0}.
15
I am glossing over some analytical details here concerning extensions and continuity, smoothness etc... see section
24 of Munkres a bit more detail in the embedded case.
8.8. ON BOUNDARIES AND SUBMANIFOLDS 225
It follows that

=

+

1
{0}. Note that a subset of

is said to be open in

i there exists some open set
such that

= . For example, if we consider
3
then the open sets in the -plane are formed from intesecting open sets in
3
with the plane; an
open ball intersects to give an open disk on the plane. Or for
2
an open disks intersected with
the -axis give open intervals.
Denition 8.8.1.
We say is a smooth -dimensional manifold with boundary i there exists a family
{
} of open subsets of
or

and local parameterizations
such
that the following criteria hold:
1. each map
is injective
2. if
:
1
)
1
)
such that
3. =
)
We again refer to the inverse of a local paramterization as a coordinate chart and often
use the notation
1
() = (
1
(),
2
(), . . . ,
()). If there exists open in
such that
: is a local parametrization with then is an interior point. Any point
which is not an interior point is a boundary point. The set of all boundary points
is called boundary of is denoted .
A more pragmatic characterization
16
of a boundary point is that i there exists a chart
at such that
() = 0. A manifold without boundary is simply a manifold in our denition

since the denitions match precisely if there are no half-space-type charts. In the case that is
nonempty we can show that it forms a manifold without boundary. Moreover, the atlas for is
naturally induced from that of by restriction.
Proposition 8.8.2.
Suppose is a smooth -dimensional manifold with boundary = then is a
smooth manifold of dimension 1. In other words, is an 1 dimensional manifold
with boundary and () = .
Proof: Let and suppose :

is a local parametrization containing
. It follows
1
= (
1
,
2
, . . . ,
1
,
) : is a chart at with
() = 0. De-
ne the restriction of
1
to
= 0 by :
() by () = (, 0) where
=
{(
1
, . . . ,
1
)
1
(
1
, . . . ,
1
,
) }. It follows that
1
: ()

1
is just the rst 1 coordinates of the chart
1
which is to say
1
= (
1
,
2
, . . . ,
1
). We
16
I leave it to the reader to show this follows from the words in green.
construct charts in this fashion at each point in . Note that
is open in
1
hence the man-
ifold only has interior points. There is no parametrization in which takes a boundary-type
subset half-plane as its domain. It follows that () = . I leave compatibility and smoothness
of the restricted charts on to the reader.
Given the terminology in this section we should note that there are shapes of interest which simply
do no t our terminology. For example, a rectangle = [, ] [, ] is not a manifold with bound-
ary since if it were we would have a boundary with sharp edges (which is not a smooth manifold!).
I have not included a full discussion of submanifolds in these notes. However, I would like to
give you some brief comments concerning how they arise from particular functions. In short, a
submanifold is a subset of a manifold which also a manifold in a natural manner. Burns and Gidea
dene for a smooth mapping from a manifold to another manifold that
a is a critical point of if

()
is not surjective. Moreover, the image
() is called the critical value of .
b is a regular point of if is not critical. Moreover, is called a regular value
of i
1
{} contains no critical points.
It turns out that:
Theorem 8.8.3.
If : is a smooth function on smooth manifolds , of dimensions ,
respective and is a regular value of with nonempty ber
1
{} then the ber
1
{} is a submanifold of of dimension ().
Proof: see page 46 of Burns and Gidea. .
The idea of this theorem is a variant of the implicit function theorem. Recall if we are given
:
then the local solution = () of (, ) = exists provided

is invertible.
But, this local solution suitably restricted is injective and hence the mapping () = (, ()) is a
local parametrization of a manifold in
. In fact, the graph = () gives -dimensional

submanifold of the manifold
. (think of =
hence = + and = so
we nd agreement with the theorem above at least in the concrete case of level-sets)
Example 8.8.4. Consider :
2
dened by (, ) =
2
+
2
. Calculate = 2 + 2
we nd that the only critical value of is (0, 0) since otherwise either or is nonzero and as a
consequence is surjective. It follows that
1
{
2
} is a submanifold of
2
for any > 0. I think
youve seen these submanifolds before. What are they?
Example 8.8.5. Consider :
3
dened by (, , ) =
2

2

2
calculate that =
2 2 +2. Note (0, 0, 0) is a critical value of . Furthermore, note
1
{0} is the cone
2
=
2
+
2
which is not a submanifold of
3
. It turns out that in general just about anything
8.8. ON BOUNDARIES AND SUBMANIFOLDS 227
can arise as the inverse image of a critical value. It could happen that the inverse image is a
submanifold, its just not a given.
Theorem 8.8.6.
If be a smooth manifold without boundary and : is a smooth function with a
regular value then
1
(, ] is a smooth manifold with boundar
1
{}.
is dened by () =
2
then = 0 is the only critical value
of and we nd
1
(,
2
] is a submanifold with boundary
1
{
2
}. Note that
1
(, 0) =
in this case. However, perhaps you also see
=
1
[0,
2
] is the closed -ball and
1
() is the (1)-sphere of radius .
Theorem 8.8.8.
Let be a smooth manifold with boundary and a smooth manifold without bound-
ary. If : and
: have regular value then

1
{} is a smooth
()-dimensional manifold with boundary
1
{} .
This theorem would seem to give us a generalization of the implicit function theorem for some
closed sets. Interesting. Finally, I should mention that it is customary to also allow use the set
1
= { 0} as the domain of a parametrization in the case of one-dimensional manifolds.
Chapter 9
dierential forms
9.1 algebra of dierential forms
In this section we apply the results of the previous section on exterior algebra to the vector space
=
. Recall that {

} is a basis of
and thus the basis {
} of utilized throughout
the previous section on exterior algebra will be taken to be
, 1
in this section. Also recall that the set of covectors {
} is a basis of
which is dual to {

}
and consequently the {
} in the previous section is taken to be
, 1
in the present context. With these choices the machinery of the previous section takes over and
one obtains a vector space
) for each 1 and for arbitrary . We write
for the set of ordered pairs (, ) where and
) and we refer to
() as the
k-th exterior power of the tangent bundle . There is a projection :
() dened by
(, ) = for (, )
(). One refers to (
, ) as a vector bundle for reasons we do not

pursue at this point. To say that is a section of this vector bundle means that :
()
is a (smooth) function such that ()
) for all . Such functions are also called

dierential forms, or in this case, k-forms.
Denition 9.1.1. vector eld on open subset of
.
To say that is a vector eld on an open subset of means that
=
1

1
+
2

2
+
where
1
,
2
, ,
are smooth functions from into R.

229
230 CHAPTER 9. DIFFERENTIAL FORMS
Note that in this context we implicitly require that dierential forms be smooth. To explain this
we write out the requirements more fully below.
If is a function with domain such that for each , ()
) then is called a
dierential k-form on if for all local vector elds
1
,
2
, ,
dened on an arbitrary open

subset of it follows that the map dened by

(
1
(),
2
(), ,
())
is smooth on . For example if (
1
,
2
, ,
) is a chart then its domain is open in

and the map

is a dierential 1-form on . Similarly the map

is a dierential 2-form on . Generally if is a 1-form and (

1
,
2
, ,
) is a chart then there

are functions (
) dened on the domain of such that

() =
()
for all in the domain of .

Similarly if is a 2-form on and = (
1
,
2
, ,
) is any chart on then there are smooth

functions
on () such that
=
1
2
,=1
()(
)
and such that
() =
() for all ().

Generally if is a -form and is a chart then on ()
1
!
()(
)
where the {
} are smooth real-valued functions on = () and
= ()
,
for every permutation . (this is just a fancy way of saying if you switch any pair of indices it
generates a minus sign).
The algebra of dierential forms follows the same rules as the exterior algebra we previously dis-
cussed. Remember, a dierential form evaluated a particular point gives us a wedge product of a
bunch of dual vectors. It follows that the dierential form in total also follows the general properties
of the exterior algebra.
Theorem 9.1.2.
9.2. EXTERIOR DERIVATIVES: THE CALCULUS OF FORMS 231
If is a -form, is a -form, and is a -form on then
1. ( ) = ( )
2. = (1)
( )
3. ( +) = ( ) +( ) ,
.
Notice that in
3
the set of dierential forms
= {1, , , , , , , }
is a basis of the space of dierential forms in the sense that every form on
3
is a linear combination
of the forms in with smooth real-valued functions on
3
as coecients.
Example 9.1.3. Let = + and let = 3 + where , are functions. Find ,
write the answer in terms of the basis dened in the Remark above,
= ( +) (3 +)
= (3 +) + (3 +)
= 3 + + 3 +
= 3
(9.1)
Example 9.1.4. Top form: Let = and let be any other form with degree > 0.
We argue that = 0. Notice that if > 0 then there must be at least one dierential inside
so if that dierential is
we can rewrite =
for some . Then consider,

=
(9.2)
now has to be either 1, 2 or 3 therefore we will have
repeated, thus the wedge product will be

zero. (can you prove this?).
9.2 exterior derivatives: the calculus of forms
The operation depends only on the values of the forms point by point. We dene an operator
on dierential forms which depends not only on the value of the dierential form at a point but on
its value in an entire neighborhood of the point. Thus if ia -form then to dene at a point
we need to know not only the value of at but we also need to know its value at every in a
neighborhood of .
You might note the derivative below does not directly involve the construction of dierential forms
from tensors. Also, the rule given below is easily taken as a starting point for formal calculations.
In other words, even if you dont understant the nuts and bolts of manifold theory you can still
calculate with dierential forms. In the same sense that highschool students do calculus, you can
do dierential form calculations. I dont believe this is a futile exercise so long as you understand
you have more to learn. Which is not to say we dont know some things!
Denition 9.2.1. the exterior derivative.
If is a -form and = (
1
,
2
, ,
) is a chart and =
1
!
and we dene a
( + 1)-form to be the form
=
1
!
.
Where
is dened as it was in calculus III,
=1
.
Note that
is well-dened as
is just a real-valued function on (). The denition in an expanded form is given by
=
1
!
1
=1
2
=1

=1
(
where
=
1
!
1
=1
2
=1

=1
()
.
9.2.1 coordinate independence of exterior derivative
The Einstein summation convention is used in this section and throughout the remainder
of this chapter, please feel free to email me if it confuses you somewhere. When an index is
repeated in a single summand it is implicitly assumed there is a sum over all values of that
index
It must be shown that this denition is independent of the chart used to dene . Suppose for
example, that
()(
)
for all in the domain of a chart (
1
,
2
,
) where
() (), = .
We assume, of course that the coecients {
()} are skew-symmetric in for all . We will

have dened in this chart by
=
.
9.2. EXTERIOR DERIVATIVES: THE CALCULUS OF FORMS 233
We need to show that
for all () () if this denition is

to be meaningful. Since is given to be a well-dened form we know
()
()
.
Using the identities
we have
so that
_
.
Consequently,
) =

_
](
_
(
)
+
1

2
_
(
)
=

1
_

_
_
=

__
where in (*) the sum
is zero since:
) =

2
[(
] = 0.
It follows that is independent of the coordinates used to dene it.
Consequently we see that for each the operator maps
() into
+1
() and has the following
properties:
Theorem 9.2.2. properties of the exterior derivative.
If
(),
() and , R then
1. ( +) = () +()
2. ( ) = ( ) + (1)
( )
3. () = 0
Proof: The proof of (1) is obvious. To prove (2), let = (
1
, ,
) be a chart on then
(ignoring the factorial coecients)
( ) = (
= (
)
+
)
=
(1)
))
+
((
)
= ( (1)
) +
)
= + (1)
( ) .
9.2.2 exterior derivatives on
3
We begin by noting that vector elds may correspond either to a one-form or to a two-form.
Denition 9.2.3. dictionary of vectors verses forms on
3
.
Let

= (
1
,
2
,
3
) denote a vector eld in
3
. Dene then,
which we will call the work-form of

. Also dene
=
1
2
) =
1
2
)
which we will call the ux-form of

.
If you accept the primacy of dierential forms, then you can see that vector calculus confuses two
separate objects. Apparently there are two types of vector elds. In fact, if you have studied coor-
dinate change for vector elds deeply then you will encounter the qualiers axial or polar vector
elds. Those elds which are axial correspond directly to two-forms whereas those correspondant
to one-forms are called polar. As an example, the magnetic eld is axial whereas the electric eld
is polar.
Example 9.2.4. Gradient: Consider three-dimensional Euclidean space. Let :
3
then
=

which gives the one-form corresponding to .

Example 9.2.5. Curl: Consider three-dimensional Euclidean space. Let

be a vector eld and
let
be the corresponding one-form then

= (
) + (
) + (
)
=
.
9.3. PULLBACKS 235
Thus we recover the curl.
Example 9.2.6. Divergence: Consider three-dimensional Euclidean space. Let

be a vector
eld and let
=
1
2
be the corresponding two-form then
= (
1
2
=
1
2
=
1
2

=
1
2
2
)
=

= (

)
and in this way we recover the divergence.
9.3 pullbacks
Another important operation one can perform on dierential forms is the pull-back of a form
under a map
1
. The denition is constructed in large part by a sneaky application of the push-
forward (aka dierential) discussed in the preceding chapter.
Denition 9.3.1. pull-back of a dierential form.
If : is a smooth map and
() then
is the form on dened by

(
(
1
, ,
) =
()
(
(
1
),
(
2
), ,
)) .
for each and
1
,
2
, . . . ,
. Moreover, in the case = 0 we have a smooth

function : and the pull-back is accomplished by composition (
)() = (
)()
for all .
This operation is linear on forms and commutes with the wedge product and exterior derivative:
Theorem 9.3.2. properties of the pull-back.
If : is a
1
-map and
(),
() then
1.
( +) = (
) +(
) ,
2.
( ) =
)
3.
() = (
)
1
thanks to my advisor R.O. Fulp for the arguments that follow
Proof: The proof of (1) is clear. We now prove (2).
( )]
(
1
, ,
+
) = ( )
()
(
(
1
), ,
(
+
))
=
(sgn)( )
()
(
1
), ,
(
(+)
))
=
sgn()(
(
(1)
),
(
()
))((
(+1)

(+)
)
=
sgn()(
(
(1)
, ,
()
)(
)(
(+1)
, ,
(+)
)
= [(
) (
)]
(
1
,
2
, ,
(+)
)
Finally we prove (3).
()]
(
1
,
2
,
+1
) = ()
()
((
1
), (
+1
))
= (
)
()
((
1
), , (
+1
))
=
_
()
_
(
)
()
((
1
), , (
+1
))
=
_
()
_
[
)](
1
, ,
+1
)
= [(
) (
)](
1
, ,
+1
)
= [(
)](
1
, ,
+1
)
= (
(
1
, ,
+1
) .
The theorem follows. .
We saw that one important application of the push-forward was to change coordinates for a given
vector. Similar comments apply here. If we wish to change coordinates on a given dierential form
then we can use the pull-back. However, given the direction of the operation we need to use the
inverse coordinate transformation to pull forms forward. Let me mirror the example from the last
chapter for forms on
2
. We wish to convert from , to , notation.
2
,

2
,
is the polar coordinate transformation. In particular,
(, ) = ( cos , sin )
The inverse transformation, at least for appropriate angles, is given by
1
(, ) =
_
2
+
2
, tan
1
(/)
_
.
Let calculate the pull-back of under
1
: let =
1
()
1
()
Again, drop the annoying point-dependence to see this clearly:
1
() = (
) +(
) =

9.4. INTEGRATION OF DIFFERENTIAL FORMS 237

Likewise,
1
() = (
) +(
) =

Note that =
2
+
2
and = tan
1
(/) have the following partial derivatives:
2
+
2
=

and

2
+
2
=

2
+
2
=

2
and

2
+
2
=

2
Of course the expressions using are pretty, but to make the point, we have changed into , -
notation via the pull-back of the inverse transformation as advertised. We nd:
=
+
2
+
2
and =
+
2
+
2
.
Once again we have found results with the pull-back that we might previously have chalked up to
substitution in multivariate calculus. Thats often the idea behind an application of the pull-back.
Its just a formal langauge to be precise about a substitution. It takes us past simple symbol
pushing and helps us think about where things are dened and how we may recast them to work
together with other objects. I leave it at that for here.
9.4 integration of dierential forms
The general strategy is generally as follows:
(i) there is a natural way to calculate the integral of a -form on a subset of
(ii) given a -form on a manifold we can locally pull it back to a subset of
provided the
manifold is an oriented
2
-dimensional and thus by the previous idea we have an integral.
(iii) globally we should expect that we can add the results from various local charts and arrive at
a total value for the manifold, assuming of course the integral in each chart is nite.
We will only investigate items (.) and (.) in these notes. There are many other excellent texts
which take great eort to carefully expand on point (iii.) and I do not wish to replicate that eort
here. You can read Edwards and see about pavings, or read Munkres where he has at least 100
pages devoted to the careful study of multivariate integration. I do not get into those topics in my
notes because we simply do not have sucient analytical power to do them justice. I would encour-
age the student interested in deeper ideas of integration to nd time to talk to Dr. Skoumbourdis,
he has thought a long time about these matters and he really understands integration in a way we
dare not cover in the calculus sequence. You really should have that conversation after youve taken
2
we will discuss this as the section progresses
real analysis and have gained a better sense of what analysis purpose is in mathematics. That
said, what we do cover in this section and the next is fascinating whether or not we understand all
the analytical underpinnings of the subject!
9.4.1 integration of -form on
Note that on
a -form is the top form thus there exists some smooth function on
such that
= ()
1

2

for all . If is a subset of then we dene the

integral of over via the corresponding intgral of -variables in
. In particular,
()
where on the r.h.s. the symbol
is meant to denote the usual integral of -variables on
. It
is sometimes convenient to write such an integral as:
()
()
1
but, to be more careful, the integration of over is a quantity which is independent of the
particular order in which the variables on
are assigned. On the other hand, the order of the

variables in the formula for certainly can introuduce signs. Note
= ()
2

1

.
How the can we reasonably maintain the integral proposed above? Well, the answer is to make
a convention that we must write the form to match the standard orientation of
. The stan-
dard orientation of
is given by
=
1

2

. If the given form is written
= ()
then we dene
()
. Since it is always possible to write a -form as

a function multiplying
on
this denition suces to cover all possible -forms. I expand a

few basic cases below:
Suppose
= () on some subset = [, ] of ,
() =
().
Or, if
(,)
= (, ) then for a aubset of
2
,
(, ) =
.
If
(,,)
= (, , ) then for a aubset of
3
,
(, , ) =
.
In practice we tend to break the integrals above down into an interated integral thanks to Fubinis
theorems. The integrals
and
are not in and of themselves dependent on orientation.

However the set may be oriented the value of those integrals are the same for a xed function
. The orientation dependence of the form integral is completely wrapped up in our rule that the
form must be written as a multiple of the volume form on the given space.
9.4.2 orientations and submanifolds
Given a -manifold we say it is an oriented manifold i all coordinates on are consistently
oriented. If we make a choice and say
0
:
0

0
is right-handed then any overlapping patch
1
:
1

1
is said to be right-handed i (
01
) > 0. Otherwise, if (
01
) < 0 then the
patch
1
:
1

1
is said to be left-handed. If the manifold is orientable then as we continue to
travel across the manifold we can choose coordinates such that on each overlap the transition func-
tions satisfy (
) > 0. In this way we nd an atlas for an orientable which is right-handed.

We can also say is oriented is there exists a nonzero volume-form on . If is -dimensional
then a volume form is simply a nonzero -form. At each point we can judge if a given
coordinate system is left or right handed. We have to make a convention to be precise and I do so
at this point. We assume is positive and we say a coordinate system with chart (
1
,
2
, . . . ,
)
is positively oriented i (
1
,
2
, . . . ,
) > 0. If a coordinate system is not positively

oriented then it is said to be negatively oriented and we will nd (
1
,
2
, . . . ,
) < 0 in
that case. It is important that we suppose
= 0 at each since that is what allows us to

demarcate coordinate systems as positive or negatively oriented.
Naturally, you are probably wondering: is a positively oriented coordinate system is the same idea
as a right-handed coordinate system as dened above? To answer that we should analyze how the
changes coordinates on an overlap. Suppose we are given a positive volume form and
a point where two coordinate systems and are both dened. There must exist some
function such that

= ()
1

2

To change coordinates recall
=1
and subsitute,
=
1
,...,
=1
(

1
)()

1
= (

1
)()
_
1

2

(9.3)
If you calculate the value of on
, . . . ,
youll nd (
) = (().
Whereas, if you evaluate on
, . . . ,
then the value is (
) =
(())
_
()
. But, we should recognize that

_
= (
) hence two coordinate sys-

tems which are positively oriented must also be consistently oriented. Why? Assume (
) =
(()) > 0 then (
) = (())
_
()
> 0 i
_
()
> 0 hence is positively ori-

ented if we are given that is positively oriented and
_
> 0.
Let be an oriented -manifold with orientation given by the volume form and an associated
atlas of positively oriented charts. Furthermore, let be a -form dened on . Suppose
there exists a local parametrization :
and then there is a smooth

function such that
= ()
1

2

for each . We dene the integral of

over as follows:
1
()
(())
]
Is this denition dependent on the coordinate system :
? If we instead used
coordinate system :

where coordinates
1
,
2
, . . . ,
on

then the given
form has a dierent coecient of
1

2

= ()
1

2

= (
1
)()
_
1

2

Thus, as we change over to coordinates the function picks up a factor which is precisely the
determinant of the derivative of the transition functions.
1
)()
_
1

2

1
()
(
1
)()
_
]
We need
in order for the integral
to be well-dened. Fortunately, the needed equality

is almost provided by the change of variables theorem for multivariate integrals on
. Recall,
()
()
where

is more pedantically written as

=

1
, notation aside its just the function written
in terms of the new -coordinates. Likewise,

limits -coordinates so that the corresponding
-coordinates are found in . Applying this theorem to our pull-back expression,
1
()
(())
1
()
(
1
)()
.
Equality of
and
follows from the fact that is oriented and has transition functions
3
which satisfy (
) > 0. We see that this integral to be well-dened only for oriented manifolds.
To integrate over manifolds without an orientation additional ideas are needed, but it is possible.
3
once more recall the notation

is just the matrix of the linear transformation and the determinant of a

linear transformation is the determinant of the matrix of the transformation
Perhaps the most interesting case to consider is that of an embedded -manifold in
. In this
context we must deal with both the coordinates of the ambient
and the local parametrizations

of the -manifold. In multivariate calculus we often consider vector elds which are dened on an
open subset of
3
and then we calculate the ux over a surfaces or the work along a curve. What
we have dened thus-far is in essence like denition how to integrate a vector eld on a surface
or a vector eld along a curve, no mention of the vector eld o the domain of integration was
made. We supposed the forms were already dened on the oriented manifold, but, what if we are
instead given a formula for a dierential form on
? How can we restrict that dierential form

to a surface or line or more generally a parametrized -dimensional submanifold of
? That is
the problem we concern ourselvew with for the remainder of this section.
Lets begin with a simple object. Consider a one-form =
=1
where the function
()
is smooth on some subset of
. Suppose is a curve parametrized by :
then
the natural chart on is provided by the parameter in particular we have
= {
}
where (
) = and
= {
} hence a vector eld along has the form ()

and a
dierential form has the form (). How can we use the one-form on
to naturally obtain a
one-form dened along C? I propose:
() =
=1
(())

It can be shown that
is a one-form on . If we change coordinates on the curve by reparametriz-

ing it then the component relative to vs. the component relative to are related:
=1
((()))
=1
(())
=1
(())
_
This is precisely the transformation rule we want for the components of a one-form.
Example 9.4.1. Suppose = + 3
2
+ and is the curve :
3
dened by
() = (1, ,
2
) we have = 1, = and =
2
hence = 0, = and = 2 on hence
= 0 + 3 +(2) = (3 + 2
2
).
Next, consider a two-form =
,=1
1
2
. Once more we consider a parametrized

submanifold of
. In particular use the notation :

2

where , serve as
coordinates on the surface . We can write an arbitrary two-form on in the form (, )
where : is a smooth function on . How should we construct (, ) given ? Again, I
think the following formula is quite natural, honestly, what else would you do
4
?
(, ) =
,=1
((, ))

4
include the
1
2
you say?, well see why not soon enough
The coecient function of is smooth because we assume
is smooth on
and the local

parametrization is also assumed smooth so the functions

and

are smooth. Moreover, the

component function has the desired coordinate change property with respect to a reparametrization
of . Suppose we reparametrize by , , then suppressing the point-dependence of
,=1
,=1
,=1
.
Therefore, the restriction of to is coordinate independent and we have thus constructed a
two-form on a surface from the two-form in the ambient space.
Example 9.4.2. Consider =
2
+ + ( + + + ) . Suppose
4
is
parametrized by
(, ) = (1,
2
2
, 3, )
In other words, we are given that
= 1, =
2
2
, = 3, =
Hence, = 0, = 2
2
+ 2
2
, = 3 and = . Computing
is just a matter of
substuting in all the formulas above, fortunately = 0 so only the term is nontrivial:
= (2
2
+ 2
2
) (3) = 6
2
2
= 6
2
2
.
It is fairly clear that we can restrict any -form on
to a -dimensional parametrized submanifold

by the procedure we explained above for = 1, 2. That is the underlying idea in the denitions
which follow. Beyond that, once we have restricted the -form on
to
then we pull-back the

restricted form to an open subset of
and reduce the problem to an ordinary multivariate integral.

Our goal in the remainder of the section is to make contact with the
5
integrals we study in calculus.
Note that an embedded manifold with a single patch is almost trivially oriented since there is no
overlap to consider. In particular, if :
is a local parametrization with
1
= (
1
,
2
, . . . ,
) then
1

2

is a volume form for . This is the natural

generalization of the normal-vector eld construction for surfaces in
3
.
Denition 9.4.3. integral of one-form along oriented curve:
Let =
be a one form and let be an oriented curve with parametrization () :

[, ] then we dene the integral of the one-form along the curve as follows,
(())
()
where () = (
1
(),
2
(), . . . ,
()) so we mean
to be the
component of ().
Moreover, the indices are understood to range over the dimension of the ambient space, if
we consider forms in
2
then = 1, 2 if in
3
then = 1, 2, 3 if in Minkowski
4
then
should be replaced with = 0, 1, 2, 3 and so on.
5
hopefully known to you already from multivariate calculus
Example 9.4.4. One form integrals vs. line integrals of vector elds: We begin with a
vector eld

and construct the corresponding one-form
. Next let be an oriented

curve with parametrization : [, ] , observe
(())
() =
You may note that the denition of a line integral of a vector eld is not special to three dimensions,
we can clearly construct the line integral in n-dimensions, likewise the correspondance can be
written between one-forms and vector elds in any dimension, provided we have a metric to lower
the index of the vector eld components. The same cannot be said of the ux-form correspondance,
it is special to three dimensions for reasons we have explored previously.
Denition 9.4.5. integral of two-form over an oriented surface:
Let =
1
2
be a two-form and let be an oriented piecewise smooth surface with

parametrization (, ) :
2

2

then we dene the integral of the two-form

over the surface as follows,
((, ))
(, )
(, )
where (, ) = (
1
(, ),
2
(, ), . . . ,
(, )) so we mean
to be the
component
of (, ). Moreover, the indices are understood to range over the dimension of the ambient
space, if we consider forms in
2
then , = 1, 2 if in
3
then , = 1, 2, 3 if in Minkowski
4
then , should be replaced with , = 0, 1, 2, 3 and so on.
Example 9.4.6. Two-form integrals vs. surface integrals of vector elds in
3
: We begin
with a vector eld

and construct the corresponding two-form
=
1
2
which is to
say
=
1
+
2
+
3
. Next let be an oriented piecewise smooth surface
with parametrization :
2

3
, then
Proof: Recall that the normal to the surface has the form,
(, ) =

at the point (, ). This gives us a vector which points along the outward normal to the surface
and it is nonvanishing throughout the whole surface by our assumption that is oriented. Moreover
the vector surface integral of

over was dened by the formula,
((, ))

(, ) .
now that the reader is reminded whats what, lets prove the proposition, dropping the (u,v) depence
to reduce clutter we nd,

=

(
notice that we have again used our convention that (
refers to the tensor components of

the 2-form
meaning we have
= (
whereas with the wedge product
=
1
2
(
, I mention this in case you are concerned there is a half in
yet we never found

a half in the integral. Well, we dont expect to because we dened the integral of the form with
respect to the tensor components of the form, again they dont contain the half.
Example 9.4.7. Consider the vector eld

= (0, 0, 3) then the corresponding two-form is simply
= 3 . Lets calculate the surface integral and two-form integrals over the square =
[0, 1][0, 1] in the -plane, in this case the parameters can be taken to be and so (, ) = (, )
and,
(, ) =

= (1, 0, 0) (0, 1, 0) = (0, 0, 1)

which is nice. Now calculate,

=

(0, 0, 3) (0, 0, 1)
=
1
0
1
0
3
= 3.
Consider that
= 3 = 3 3 therefore we may read directly that (
)
12
=
(
)
21
= 3 and all other components are zero,
=

(

=

_
3
_

=
1
0
1
0
_
3
_

= 3.
Denition 9.4.8. integral of a three-form over an oriented volume:
Let =
1
6
be a three-form and let be an oriented piecewise smooth

volume with parametrization (, , ) :
3

3

then we dene the integral

of the three-form in the volume as follows,
((, , ))

where (, , ) = (
1
(, , ),
2
(, , ), . . . ,
(, , )) so we mean
to be the
component of (, , ). Moreover, the indices are understood to range over the dimension
of the ambient space, if we consider forms in
3
then , , = 1, 2, 3 if in Minkowski
4
then
, , should be replaced with , , = 0, 1, 2, 3 and so on.
Finally we dene the integral of a -form over an -dimensional subspace of , we assume that
so that it is possible to embed such a subspace in ,
Denition 9.4.9. integral of a p-form over an oriented volume:
Let =
1
!
1
...
be a p-form and let be an oriented piecewise smooth

subspace with parametrization (
1
, . . . ,
) :
(for ) then we
dene the integral of the p-form in the subspace as follows,
1
...
((
1
, . . . ,
))
where (
1
, . . . ,
) = (
1
(
1
, . . . ,
),
2
(
1
, . . . ,
), . . . ,
(
1
, . . . ,
)) so we mean
to be the
component of (
1
, . . . ,
). Moreover, the indices are understood to range

over the dimension of the ambient space.
Integrals of forms play an important role in modern physics. I hope you can begin to appreciate
that forms recover all the formulas we learned in multivariate calculus and give us a way to extend
calculation into higher dimensions with ease. I include a toy example at the conclusion of this
chapter just to show you how electromagnetism is easily translated into higher dimensions.
9.5 Generalized Stokes Theorem
The generalized Stokes theorem contains within it most of the main theorems of integral calculus,
namely the fundamental theorem of calculus, the fundamental theorem of line integrals (a.k.a
the FTC in three dimensions), Greenes Theorem in the plane, Gauss Theorem and also Stokes
Theorem, not to mention a myriad of higher dimensional not so commonly named theorems. The
breadth of its application is hard to overstate, yet the statement of the theorem is simple,
Theorem 9.5.1. Generalized Stokes Theorem:
Let be an oriented, piecewise smooth (p+1)-dimensional subspace of
where +1
and let be it boundary which is consistently oriented then for a -form which behaves
reasonably on we have that
The proof of this theorem (and a more careful statement of it) can be found in a number of places,
Susan Colleys Vector Calculus or Steven H. Weintraubs Dierential Forms: A Complement to
Vector Calculus or Spivaks Calculus on Manifolds just to name a few. I believe the argument in
Edwards text is quite complete. In any event, you should already be familar with the idea from
the usual Stokes Theorem where we must insist the boundary curve to the surface is related to
the surfaces normal eld according to the right-hand-rule. Explaining how to orient the boundary
given an oriented is the problem of generalizing the right-hand-rule to many dimensions. I
leave it to your homework for the time being.
Lets work out how this theorem reproduces the main integral theorems of calculus.
Example 9.5.2. Fundamental Theorem of Calculus in : Let : be a zero-form
then consider the interval [, ] in . If we let = [, ] then = {, }. Further observe that
=
(). Notice by the denition of one-form integration
()
However on the other hand we nd ( the integral over a zero-form is taken to be the evaluation
map, perhaps we should have dened this earlier, oops., but its only going to come up here so Im
leaving it.)
= () ()
Hence in view of the denition above we nd that
() = () ()
9.5. GENERALIZED STOKES THEOREM 247

Example 9.5.3. Fundamental Theorem of Calculus in
3
: Let :
3
be a zero-form
then consider a curve from
3
to
3
parametrized by : [, ]
3
. Note that
= {() = , () = }. Next note that
=

Then consider that the exterior derivative of a function corresponds to the gradient of the function
thus we are not to surprised to nd that
()
On the other hand, we use the denition of the integral over a a two point set again to nd
= () ()
Hence if the Generalized Stokes Theorem is true then so is the FTC in three dimensions,
()
= () ()
another popular title for this theorem is the fundamental theorem for line integrals. As a nal
thought here we notice that this calculation easily generalizes to 2,4,5,6,... dimensions.
Example 9.5.4. Greenes Theorem: Let us recall the statement of Greenes Theorem as I have
not replicated it yet in the notes, let be a region in the -plane and let be its consistently
oriented boundary then if

= ((, ), (, ), 0) is well behaved on

+ =

_
_

We begin by nding the one-form corresponding to

namely
= + consider then that
= ( +) = + =

which simplies to,
=
_
_
=
(
Thus, using our discussion in the last section we recall

+
where we have reminded the reader that the notation in the rightmost expression is just another
way of denoting the line integral in question. Next observe,

And clearly, since
=

we have
)
Therefore,

+ =

_
Example 9.5.5. Gauss Theorem: Let us recall Gauss Theorem to begin, for suitably dened

and ,

First we recall our earlier result that
(
) = (

)
Now note that we may integrate the three form over a volume,
) =

(

)
whereas,
so there it is,

(

) =
) =
I have left a little detail out here, I may assign it for homework.
Example 9.5.6. Stokes Theorem: Let us recall Stokes Theorem to begin, for suitably dened

and ,

(

)
Next recall we have shown in the last chapter that,

(
) =
Hence,
) =

(

)
whereas,
which tells us that,

(

)
) =
9.6. POINCARES LEMMA AND CONVERSE 249

The Generalized Stokes Theorem is perhaps the most persausive argument for mathematicians to
be aware of dierential forms, it is clear they allow for more deep and sweeping statements of
the calculus. The generality of dierential forms is what drives modern physicists to work with
them, string theorists for example examine higher dimensional theories so they are forced to use
a language more general than that of the conventional vector calculus. See the end of the next
chapter for an example of such thinking.
9.6 poincares lemma and converse
This section is in large part inspired by M. Gockeler and T. Schuckers Dierential geometry, gauge
theories, and gravity page 20-22. The converse calculation is modelled on the argument found in
H. Flanders Dierential Forms with Applications to the Physical Sciences. The original work was
done around the dawn of the twentieth century and can be found in many texts besides the two I
mentioned here.
9.6.1 exact forms are closed
Proposition 9.6.1.
The exterior derivative of the exterior derivative is zero.
2
= 0
Proof: Let be an arbitrary -form then
=
1
!
(
2
...
)
(9.4)
then dierentiate again,
() =
_
1
!
(
2
...
)
_
=
1
!
(
2
...
)
= 0
(9.5)
since the partial derivatives commute whereas the wedge product anticommutes so we note that
the pair of indices (k,m) is symmetric for the derivatives but antisymmetric for the wedge, as we
know the sum of symmetric against antisymmetric vanishes ( see equation ?? part if you forgot.)
Denition 9.6.2.
A dierential form is closed i = 0. A dierential form is exact i there exists
such that = .
Proposition 9.6.3.
All exact forms are closed. However, there exist closed forms which are not exact.
Proof: Exact implies closed is easy, let be exact such that = then
= () = 0
using the theorem
2
= 0. To prove that there exists a closed form which is not exact it suces
to give an example. A popular example ( due to its physical signicance to magnetic monopoles,
Dirac Strings and the like..) is the following dierential form in
2
=
1
2
+
2
( ) (9.6)
You may verify that = 0 in homework. Observe that if were exact then there would exist
such that = meaning that
2
+
2
,

2
+
2
which are solved by = (/) + where is arbitrary. Observe that is ill-dened along
the -axis = 0 ( this is the Dirac String if we put things in context ), however the natural domain
of is

{(0, 0)}.
9.6.2 potentials for closed forms
Poincare suggested the following partial converse, he said closed implies exact provided we place
a topological restriction on the domain of the form. In particular, if the domain of a closed form
is smoothly deformable to a point then each closed form is exact. Well work out a proof of that
result for a subset of
. Be patient, we have to build some toys before we play.

Suppose
and = [0, 1] we denote a typical point in as (, ) where and
.
Dene maps,
1
: ,
0
:
by
1
() = (1, ) and
0
() = (0, ) for each . Flanders encourages us to view as a
cylinder and where the map
1
maps to the top and
0
maps to the base. We can pull-back
forms on the cylinder to the on the top ( = 1) or to the base ( = 0). For instance, if we consider
= ( +) + for the case = 1 then
0
=
1
= ( + 1).
Dene a smooth mapping of ( + 1) forms on to -forms on as follows:
(1.) ((, )
) = 0, (2.) ((, )
) =
_
1
0
(, )
_

for multi-indices of length ( + 1) and of length . The cases (1.) and (2.) simply divide
the possible monomial
6
inputs from
+1
( ) into forms which have and those which dont.
Then is dened for a general ( +1)-form on by linearly extending the formulas above to
multinomials of the basic monomials.
It turns out that the following identity holds for :
Lemma 9.6.4. the -lemma.
If is a dierential form on then
() +(()) =
0
.
Proof: Since the equation is given for linear operations it suces to check the formula for mono-
mials since we can extend the result linearly once those are armed. As in the denition of
there are two basic categories of forms on :
Case 1: If = (, )
then clearly () = 0 hence (()) = 0. Observe,

=
Hence () is calculated as follows:

() =
_
_
+
_
_
=
_
1
0
=
_
(, 1) (, 0)
0
(9.7)
where we used the FTC in the next to last step. The pull-backs in this case just amount to evalu-
ation at = 0 or = 1 as there is no -type term to squash in . The identity follows.
Case 2: Suppose = (, )
. Calculate,
=

. .
!
6
is a monomial whereas + is a binomial in this context
Thus, using
, we calculate:
() =
_
_
=
_
1
0
at which point we cannot procede further since is an arbitrary function which can include a
nontrivial time-dependence. We turn to the calculation of (()). Recall we dened
() =
_
1
0
(, )
_
.
We calculate the exterior derivative of ():
(()) =
_
1
0
(, )
_

=
_

_
1
0
(, )
. .

_
+
_
1
0
(, )
_
_
1
0
. (9.8)
Therefore, ()+(()) = 0 and clearly
0
=
1
= 0 in this case since the pull-backs squash
the to zero. The lemma follows. .
Denition 9.6.5.
A subset
is deformable to a point if there exists a smooth mapping :

such that (1, ) = and (0, ) = for all .
The map deforms smoothly into the point . Recall that
1
() = (1, ) and
0
() = (0, )
hence the conditions on the deformation can be expressed as:
(
1
()) = (
0
()) =
Denoting for the identity on and as the constant mapping with value on we have
1
=
0
=
Therefore, if is a ( + 1)-form on we calculate,
(
1
)
1
[
] =
whereas,
(
0
)
= 0
0
[
] = 0
Apply the -lemma to the form =
on and we nd:
((
)) +((
)) = .
However, recall that we proved that pull-backs and exterior derivatives commute thus
(
) =
()
and we nd an extremely interesting identity,
(
()) +((
)) = .
Proposition 9.6.6.
If is deformable to a point then a -form on is closed i is exact.
Proof: Suppose is exact then = for some ( 1)-form on hence = () = 0 by
Proposition 9.6.1 hence is closed. Conversely, suppose is closed. Apply the boxed consequence
of the -lemma, note that (
(0)) = 0 since we assume = 0. We nd,

((
)) =
identify that
is a -form on whereas (
) is a ( 1)-form on by the very con-

struction of . It follows that is exact since we have shown how it is obtained as the exterior
derivative of another dierential form of one degree less.
Where was deformability to a point used in the proof above? The key is the existence of the
mapping . In other words, if you have a space which is not deformable to a point then no
deformation map is available and the construction via breaks down. Basically, if the space
has a hole which you get stuck on as you deform loops to a point then it is not deformable to a
point. Often we call such spaces simply connected. Careful denition of these terms is too dicult
for calculus, deformation of loops and higher dimensional objects is properly covered in algebraic
topology. In any event, the connection of the deformation and exactness of closed forms allows
topologists to use dierential forms detect holes in spaces. In particular:
Denition 9.6.7. de Rham cohomology:
We dene several real vector spaces of dierential forms over some subset of ,
() {
closed}
the space of closed p-forms on . Then,
() {
exact}
the space of exact p-forms where by convention
0
() = {0} The de Rham cohomology
groups are dened by the quotient of closed/exact,
()
()/
().
the (
) =
Betti number of U.
We observe that simply connected regions have all the Betti numbers zero since
() =
()
implies that
() = {0}. Of course there is much more to say about Cohomology, I just wanted
to give you a taste and alert you to the fact that dierential forms can be used to reveal aspects of
topology. Not all algebraic topology uses dierential forms though, there are several other calcula-
tional schemes based on triangulation of the space, or studying singular simplexes. One important
event in 20-th century mathematics was the discovery that all these schemes described the same
homology groups. Steenrod reduced the problem to a few central axioms and it was shown that all
the calculational schemes adhere to that same set of axioms.
One interesting aspect of the proof we (copied from Flanders
7
) is that it is not a mere existence
proof. It actually lays out how to calculate the form which provides exactness. Lets call the
potential form of if = . Notice this is totally reasonable langauge since in the case of
classical mechanics we consider conservative forces

which as derivable from a scalar potential
by

= . Translated into dierential forms we have
= . Lets explore how the

-mapping and proof indicate the potential of a vector eld ought be calculated.
Suppose is deformable to a point and is a smooth conservative vector eld on . Perhaps you
recall that for conservative are irrotational hence = 0. Recall that
=
0
= 0
thus the one-form corresponding to a conservative vector eld is a closed form. Apply the identity:
let :
3
be the deformation of to a point ,
((
)) =
Hence, including the minus to make energy conservation natural,

= (
)
For convenience, lets suppose the space considered is the unit-ball and lets use a deformation to
the origin. Explicitly, (, ) = for all
3
such that 1. Note that clearly (0, ) = 0
7
I dont know the complete history of this calculation at the present. It would be nice to nd it since I doubt
Flanders is the originator.
whereas (1, ) = and has a nice formula so its smooth
8
. We wish to calculate the pull-back
of
= + + under , from the denition of pull-back we have

(
)() =
(())
for each smooth vector eld on . Dierential forms on are written as linear combi-
nations of , , , with smooth functions as coecients. We can calculate the coecents by
evalutaion on the corresponding vector elds
. Observe, since (, , , ) = (, , )
we have
(
) =

1
+

2
+

3
wheras,
(
) =

1
+

2
+

3
and similarly,
(
) =

) =

Furthermore,
((
)) =
) = ++
((
)) =
) = ,
((
)) =
) = ,
((
)) =
) =
Therefore,
= ( ++) + + + = ( ++) +
Now we can calculate (
) and hence . Note that only the coecient of gives a nontrivial

contribution so in retrospect we did a bit more calculation than necessary. That said, Ill just
keep it as a celebration of extreme youth for calculation. Also, Ive been a bit careless in omiting
the point up to this point, lets include the point dependence since it will be critical to properly
understand the formula.
(
)(, , , ) =
_
_
(, , ) +(, , ) +(, , )
_
_
Therefore,
(, , ) = (
) =
1
0
_
(, , ) +(, , ) +(, , )
_
Notice this is precisely the line-integral of =< , , >along the line with direction < , , >
from the origin to (, , ). In particular, if () =< , , > then

=< , , > hence we

identify
(, , ) =
1
0
_
()
_

8
there is of course a deeper meaning to the word, but, for brevity I gloss over this.
Perhaps you recall this is precisely how we calculate the potential function for a conservative vector
eld provided we take the origin as the zero for the potential.
Actually, this calculation is quite interesting. Suppose we used a dierent deformation
: . For xed point we travel to from the origin to the point by the path

(, ).
Of course this path need not be a line. The space considered might look like a snake where a line
cannot reach from the base point to the point . But, the same potential is derived. Why?
Path independence of the vector eld is one answer. The criteria = 0 suces for a sim-
ply connected region. However, we see something deeper. The criteria of a closed form paired
with a simply connected (deformable) domain suces to construct a potential for the given form.
This result reproduces the familar case of conservative vector elds derived from scalar potentials
and much more. In Flanders he calculates the potential for a closed two-form. This ought to
be the mathematics underlying the construction of the so-called vector potential of magnetism.
In junior-level electromagnetism
9
the magnetic eld satises = 0 and thus the two-form
has exterior derivative
= = 0. The magnetic eld corresponds to a

closed form. Poincares lemma shows that there exists a one-form
such that
. But
this means
hence in the langauge of vector elds we expect the vector potential

generated the magnetic eld throught the curl = . Indeed, this is precisely what a
typical junior level physics student learns in the magnetostatic case. Appreciate that it goes deeper
still, the Poincare lemma holds for -forms which correspond to objects which dont match up with
the quaint vector elds of 19-th century physics. We can be condent to nd potential for 3-form
uxes in a 10-dimensional space, or wherever our imagination takes us. I explain at the end of this
chapter how to translate electromagnetics into the langauge of dierential forms, it may well be
that in the future we think about forms the way we currently think about vectors. This is one of
the reasons I like Flanders text, he really sticks with the langauge of dierential forms throughout.
In contrast to these notes, he just does what is most interesting. I think undergraduates need to
see more detail and not just the most clever calculations, but, I can hardly blame Flanders! He
makes no claim to be an undergraduate work.
Finally, I should at least mention that though we can derive a potential for a given closed form
on a simply connected domain it need not be unique. In fact, it will not be unique unless we add
further criteria for the potential. This ambuity is called gauge freedomin physics. Mathematically
its really simple give form language. If = where is a ( 1)-form then we can take any
smooth ( 2) form and calculate that
( +) = +
2
= =
Therefore, if is a potential-form for then + is also a potential-form for .
9
just discussing magnetostatic case here to keep it simple
9.7. CLASSICAL DIFFERENTIAL GEOMETRY IN FORMS 257
9.7 classical dierential geometry in forms
XXX- already worked it out, copy from my notepad.
9.8 E & M in dierential form
Warning: I will use Einsteins implicit summation convention throughout this section.
I have made a point of abstaining from Einsteins convention in these notes up to this point.
However, I just cant bear the summations in this section. Theyre just too ugly.
9.8.1 dierential forms in Minkowski space
The logic here follows fairly close to the last section, however the wrinkle is that the metric here
demands more attention. We must take care to raise the indices on the forms when we Hodge dual
them. First we list the basis dierential forms, we have to add time to the mix ( again = 1 so
0
= = if you worried about it )
Name Degree Typical Element Basis for
(
4
)
function = 0 1
one-form = 1 =
, , ,
two-form = 2 =
1
2
, ,
, ,
three-form = 3 =
1
3!
,
,
four-form = 4
Greek indices are dened to range over 0, 1, 2, 3. Here the top form is degree four since in four
dimensions we can have four dierentials without a repeat. Wedge products work the same as they
have before, just now we have to play with. Hodge duality may oer some surprises though.
Denition 9.8.1. The antisymmetric symbol in at
4
is denoted
and it is dened by the

value
0123
= 1
plus the demand that it be completely antisymmetric.
We must not assume that this symbol is invariant under a cyclic exhange of indices. Consider,
0123
=
1023
ipped (01)
= +
1203
ipped (02)
=
1230
ipped (03).
(9.9)
Example 9.8.2. We now compute the Hodge dual of = with respect to the Minkowski metric
. First notice that has components
=
1
as is readily veried by the equation =

1
.
We raise the index using , as follows
=
1
=
1
.
9.8. E & M IN DIFFERENTIAL FORM 259
Beginning with the denition of the Hodge dual we calculate
() =
1
(41)!
= (1/6)
1
= (1/6)[
1023
+
1230
+
1302

+
1320
+
1203
+
1032
]
= (1/6)[
+ + + ]
= .
(9.10)
The dierence between the three and four dimensional Hodge dual arises from two sources, for
one we are using the Minkowski metric so indices up or down makes a dierence, and second the
antisymmetric symbol has more possibilities than before because the Greek indices take four values.
Example 9.8.3. We nd the Hodge dual of = with respect to the Minkowski metric
.
Notice that has components
=
0
as is easily seen using the equation =

0
. Raising the
index using as usual, we have
=
0
=
0
where the minus sign is due to the Minkowski metric. Starting with the denition of Hodge duality
we calculate
() = (1/6)
0
= (1/6)
0
= (1/6)
0
= (1/6)
= .
(9.11)
for the case here we are able to use some of our old three dimensional ideas. The Hodge dual of
cannot have a in it which means our answer will only have , , in it and that is why we
were able to shortcut some of the work, (compared to the previous example).
Example 9.8.4. Finally, we nd the Hodge dual of = with respect to the Minkowski metric
. Recall that

() =
1
(42)!
01
01
(
) and that
01
=
0
= (1)(1)
01
= 1.
Thus
( ) = (1/2)
01
= (1/2)[
0123
+
0132
]
= .
(9.12)
Notice also that since = we nd ( ) =
1 =

( ) = 1
( ) =

=
( ) =

=
( ) =

=
( ) =

=
( ) =

( ) =
( ) =

( ) =
( ) =

( ) =
The other Hodge duals of the basic two-forms follow from similar calculations. Here is a table of
all the basic Hodge dualities in Minkowski space, In the table the terms are grouped as they are to
emphasize the isomorphisms between the one-dimensional
0
() and
4
(), the four-dimensional
1
() and
3
(), the six-dimensional
2
() and itself. Notice that the dimension of () is
16 which just happens to be 2
4
.
Now that weve established how the Hodge dual works on the dierentials we can easily take the
Hodge dual of arbitrary dierential forms on Minkowski space. We begin with the example of the
4-current
Example 9.8.5. Four Current: often in relativistic physics we would even just call the four
current simply the current, however it actually includes the charge density and current density
. Consequently, we dene,
(
) (,

),
moreover if we lower the index we obtain,
(
) = (,

)
which are the components of the current one-form,
=
= +
This equation could be taken as the denition of the current as it is equivalent to the vector deni-
tion. Now we can rewrite the last equation using the vectors forms mapping as,
= +
.
Consider the Hodge dual of ,
=

( +
)
=
.
(9.13)
we will nd it useful to appeal to this calculation in a later section.
Example 9.8.6. Four Potential: often in relativistic physics we would call the four potential
simply the potential, however it actually includes the scalar potential and the vector potential

(discussed at the end of chapter 3). To be precise we dene,

(
) (,

)
we can lower the index to obtain,
(
) = (,

)
which are the components of the current one-form,
=
= +
Sometimes this equation is taken as the denition of the four potential. We can rewrite the four
potential vector eld using the vectors forms mapping as,
= +
.
The Hodge dual of is
. (9.14)
Several steps were omitted because they are identical to the calculation of the dual of the 4-current
above.
Denition 9.8.7. Faraday tensor.
Given an electric eld

= (
1
,
2
,
3
) and a magnetic eld

= (
1
,
2
,
3
) we dene a
2-form by
=
.
This 2-form is often called the electromagnetic eld tensor or the Faraday tensor. If
we write it in tensor components as =
1
2
and then consider its matrix (
)
of components then it is easy to see that
(
) =
0
1

2

3
1
0
3

2
2

3
0
1
3

2

1
0
(9.15)
Convention: Notice that when we write the matrix version of the tensor components we take the
rst index to be the row index and the second index to be the column index, that means
01
=
1
whereas
10
=
1
.
Example 9.8.8. In this example we demonstrate various conventions which show how one can
transform the eld tensor to other type tensors. Dene a type (1, 1) tensor by raising the rst index
by the inverse metric
as follows,

The zeroth row,
(
0
) = (
0
) = (0,
1
,
2
,
3
)
Then row one is unchanged since
1
=
1
,
(
1
) = (
1
) = (
1
, 0,
3
,
2
)
and likewise for rows two and three. In summary the (1,1) tensor
) has the
components below
(
) =
0
1

2

3
1
0
3

2
2

3
0
1
3

2

1
0
. (9.16)
At this point we raise the other index to create a (2, 0) tensor,
(9.17)
and we see that it takes one copy of the inverse metric to raise each index and
so
we can pick up where we left o in the (1, 1) case. We could proceed case by case like we did with
the (1, 1) case but it is better to use matrix multiplication. Notice that
is just
the (, ) component of the following matrix product,
(
) =
0
1

2

3
1
0
3

2
2

3
0
1
3

2

1
0
1 0 0 0
0 1 0 0
0 0 1 0
0 0 0 1
0
1

2

3
1
0
3

2
2

3
0
1
3

2

1
0
. (9.18)
So we nd a (2, 0) tensor
). Other books might even use the same symbol for
and
, it is in fact typically clear from the context which version of one is thinking about.
Pragmatically physicists just write the components so its not even an issue for them.
Example 9.8.9. Field tensors dual: We now calculate the Hodge dual of the eld tensor,
=

(
)
=
( ) +
( ) +
( )
+
( ) +
( ) +
( )
=
.
we can also write the components of

in matrix form:
(
) =
0
1

2

3
1
0
3

2
2

3
0
1
3

2

1
0.
(9.19)
Notice that the net-eect of Hodge duality on the eld tensor was to make the exchanges

and

.
9.8.2 exterior derivatives of charge forms, eld tensors, and their duals
In the last chapter we found that the single operation of the exterior dierentiation reproduces the
gradiant, curl and divergence of vector calculus provided we make the appropriate identications
under the work and ux form mappings. We now move on to some four dimensional examples.
Example 9.8.10. Charge conservation: Consider the 4-current we introduced in example 9.8.5.
Take the exterior derivative of the dual of the current to get,
(
) = (
)
= (
) [(
) ]
=

= (
+

) .
We work through the same calculation using index techniques,
(
) = (
)
= () [
1
2
)
= (
)
1
2

= (
)
1
2

= (
)
1
2

= (
)
1
2
2

= (
+

) .
Observe that we can now phrase charge conservation by the following equation
(
) = 0
+

= 0.
In the classical scheme of things this was a derived consequence of the equations of electromagnetism,
however it is possible to build the theory regarding this equation as fundamental. Rindler describes
that formal approach in a late chapter of Introduction to Special Relativity.
Proposition 9.8.11.
If (
) = (,

) is the vector potential (which gives the magnetic eld) and = +
, then =
= where is the electromagnetic eld tensor. Moreover,
.
Proof: The proof uses the denitions

=
and

=

and some vector identities:
= ( +
)
= +(
)
= + (
+ (
=
( )

= (
( )
) +
=
(

)
+
= =
1
2
.
Moreover we also have:
= (
=
1
2
(
+
1
2
(
=
1
2
(
.
Comparing the two identities we see that

Example 9.8.12. Exterior derivative of the eld tensor: We have just seen that the eld
tensor is the exterior derivative of the potential one-form. We now compute the exterior derivative
of the eld tensor expecting to nd Maxwells equations since the derivative of the elds are governed
by Maxwells equations,
= (
) +(
)
=
) + (

) +
1
2
)(
).
(9.20)
W pause here to explain our logic. In the above we dropped the
term because it was

wedged with another in the term so it vanished. Also we broke up the exterior derivative on the
ux form of

into the space and then time derivative terms and used our work in example 9.2.6.
Continuing the calculation,
= [
+
1
2
)]
+ (

)
= [
+
12
(
)]
+[
+
31
(
)]
+[
+
23
(
)]
+(

)
= (

+
+ (

)
=
+ (

)
(9.21)
where we used the fact that is an isomorphism of vector spaces (at a point) and
1
= ,
2
= , and
3
= . Behold, we can state two of Maxwells equations as
= 0

+
= 0,

= 0 (9.22)
Example 9.8.13. We now compute the exterior derivative of the dual to the eld tensor:
= (
) +(
)
=
+ (

)
(9.23)
This follows directly from the last example by replacing

and

. We obtain the two
inhomogeneous Maxwells equations by setting
equal to the Hodge dual of the 4-current,
,

= (9.24)
Here we have used example 9.8.5 to nd the RHS of the Maxwell equations.
We now know how to write Maxwells equations via dierential forms. The stage is set to prove that
Maxwells equations are Lorentz covariant, that is they have the same form in all inertial frames.
9.8.3 coderivatives and comparing to Griths relativitic E & M
Optional section, for those who wish to compare our tensorial E & M with that of
Griths, you may skip ahead to the next section if not interested
I should mention that this is not the only way to phrase Maxwells equations in terms of
dierential forms. If you try to see how what we have done here compares with the equations
presented in Griths text it is not immediately obvious. He works with
and
and
none
of which are the components of dierential forms. Nevertheless he recovers Maxwells equations
as
and
= 0. If we compare the components of

with equation 12.119 ( the
matrix form of
) in Griths text,
(
( = 1)) =
0
1

2

3
1
0
3

2
2

3
0
1
3

2

1
0
= (
). (9.25)
we nd that we obtain the negative of Griths dual tensor ( recall that raising the indices has
the net-eect of multiplying the zeroth row and column by 1). The equation
does not
follow directly from an exterior derivative, rather it is the component form of a coderivative. The
coderivative is dened =

, it takes a -form to an ()-form then makes it a (+1)-form

then nally the second Hodge dual takes it to an ( ( + 1))-form. That is takes a -form
to a 1-form. We stated Maxwells equations as
= 0

Now we can take the Hodge dual of the inhomogeneous equation to obtain,
= =

=
where I leave the sign for you to gure out. Then the other equation
= 0
can be understood as the component form of
= 0 but this is really = 0 in disguise,

0 =
= 0
so even though it looks like Griths is using the dual eld tensor for the homogeneous Maxwells
equations and the eld tensor for the inhomogeneous Maxwells equations it is in fact not the case.
The key point is that there are coderivatives implicit within Griths equations, so you have to
read between the lines a little to see how it matched up with what weve done here. I have not en-
tirely proved it here, to be complete we should look at the component form of = and explicitly
show that this gives us
, I dont think it is terribly dicult but Ill leave it to the reader.

Comparing with Griths is fairly straightforward because he uses the same metric as we have.
Other texts use the mostly negative metric, its just a convention. If you try to compare to such
a book youll nd that our equations are almost the same up to a sign. One good careful book
is Reinhold A. Bertlmanns Anomalies in Quantum Field Theory you will nd much of what we
have done here done there with respect to the other metric. Another good book which shares our
conventions is Sean M. Carrolls An Introduction to General Relativity: Spacetime and Geometry,
that text has a no-nonsense introduction to tensors forms and much more over a curved space (
in contrast to our approach which has been over a vector space which is at ). By now there are
probably thousands of texts on tensors; these are a few we have found useful here.
9.8.4 Maxwells equations are relativistically covariant
Let us begin with the denition of the eld tensor once more. We dene the components of the
eld tensor in terms of the 4-potentials as we take the view-point those are the basic objects (not
the elds). If
,
then the eld tensor =
is a tensor, or is it ? We should check that the components

transform as they ought according to the discussion in section ??. Let
then we observe,
(1.)

= (
1
)
(2.)

= (
1
)
(9.26)
where (2.) is simply the chain rule of multivariate calculus and (1.) is not at all obvious. We will
assume that (1.) holds, that is we assume that the 4-potential transforms in the appropriate way
for a one-form. In principle one could prove that from more base assumptions. After all electro-
magnetism is the study of the interaction of charged objects, we should hope that the potentials
are derivable from the source charge distribution. Indeed, there exist formulas to calculate the
potentials for moving distributions of charge. We could take those as denitions for the potentials,
then it would be possible to actually calculate if (1.) is true. Wed just change coordinates via a
Lorentz transformation and verify (1.). For the sake of brevity we will just assume that (1.) holds.
We should mention that alternatively one can show the electric and magnetic elds transform as to
make
a tensor. Those derivations assume that charge is an invariant quantity and just apply
Lorentz transformations to special physical situations to deduce the eld transformation rules. See
Griths chapter on special relativity or look in Resnick for example.
Let us nd how the eld tensor transforms assuming that (1.) and (2.) hold, again we consider

= (
1
)
((
1
)
) (
1
)
((
1
)
)
= (
1
)
(
1
)
)
= (
1
)
(
1
)
.
(9.27)
therefore the eld tensor really is a tensor over Minkowski space.
Proposition 9.8.14.
The dual to the eld tensor is a tensor over Minkowski space. For a given Lorentz trans-
formation
it follows that
= (
1
)
(
1
)
Proof: homework (just kidding in 2010), it follows quickly from the denition and the fact we
already know that the eld tensor is a tensor.
Proposition 9.8.15.
The four-current is a four-vector. That is under the Lorentz transformation
we
can show,
= (
1
)
Proof: follows from arguments involving the invariance of charge, time dilation and length con-
traction. See Griths for details, sorry we have no time.
Corollary 9.8.16.
The dual to the four current transforms as a 3-form. That is under the Lorentz transfor-
mation
we can show,

= (
1
)
(
1
)
(
1
)

Up to now the content of this section is simply an admission that we have been a little careless in
dening things upto this point. The main point is that if we say that something is a tensor then we
need to make sure that is in fact the case. With the knowledge that our tensors are indeed tensors
the proof of the covariance of Maxwells equations is trivial.
= 0
are coordinate invariant expressions which we have already proved give Maxwells equations in one
frame of reference, thus they must give Maxwells equations in all frames of reference.
The essential point is simply that
=
1
2
=
1
2
Again, we have no hope for the equation above to be true unless we know that
= (
1
)
(
1
)
. That transformation follows from the fact that the four-potential is a

four-vector. It should be mentioned that others prefer to prove the eld tensor is a tensor by
studying how the electric and magnetic elds transform under a Lorentz transformation. We in
contrast have derived the eld transforms based ultimately on the seemingly innocuous assumption
that the four-potential transforms according to

= (
1
)
. OK enough about that.

So the fact that Maxwells equations have the same form in all relativistically inertial frames
of reference simply stems from the fact that we found Maxwells equation were given by an arbitrary
frame, and the eld tensor looks the same in the new barred frame so we can again go through all
the same arguments with barred coordinates. Thus we nd that Maxwells equations are the same
in all relativistic frames of reference, that is if they hold in one inertial frame then they will hold
in any other frame which is related by a Lorentz transformation.
9.8.5 Electrostatics in Five dimensions
We will endeavor to determine the electric eld of a point charge in 5 dimensions where we are
thinking of adding an extra spatial dimension. Lets call the fourth spatial dimension the -direction
so that a typical point in space time will be (, , , , ). First we note that the electromagnetic
eld tensor can still be derived from a one-form potential,
= +
1
+
2
+
3
+
4
we will nd it convenient to make our convention for this section that , , ... = 0, 1, 2, 3, 4 whereas
, , ... = 1, 2, 3, 4 so we can rewrite the potential one-form as,
= +

This is derived from the vector potential
= (,
) under the assumption we use the natural

generalization of the Minkowski metric, namely the 5 by 5 matrix,
(
) =
1 0 0 0 0
0 1 0 0 0
0 0 1 0 0
0 0 0 1 0
0 0 0 0 1
= (
) (9.28)
we could study the linear isometries of this metric, they would form the group (1, 4). Now we
form the eld tensor by taking the exterior derivative of the one-form potential,
= =
1
2
(
now we would like to nd the electric and magnetic elds in 4 dimensions. Perhaps we should
say 4+1 dimensions, just understand that I take there to be 4 spatial directions throughout this
discussion if in doubt. Note that we are faced with a dilemma of interpretation. There are 10
independent components of a 5 by 5 antisymmetric tensor, naively we wold expect that the electric
and magnetic elds each would have 4 components, but that is not possible, wed be missing
two components. The solution is this, the time components of the eld tensor are understood to
correspond to the electric part of the elds whereas the remaining 6 components are said to be
magnetic. This aligns with what we found in 3 dimensions, its just in 3 dimensions we had the
fortunate quirk that the number of linearly independent one and two forms were equal at any point.
This denition means that the magnetic eld will in general not be a vector eld but rather a ux
encoded by a 2-form.
(
) =
0
3

1

2

3
0
(9.29)
Now we can write this compactly via the following equation,
= +
I admit there are subtle points about how exactly we should interpret the magnetic eld, however
Im going to leave that to your imagination and instead focus on the electric sector. What is the
generalized Maxwells equation that must satisfy?
( +) =
where = +
so the 5 dimensional Hodge dual will give us a 5 1 = 4 form, in

particular we will be interested in just the term stemming from the dual of ,
() =
the corresponding term in
is
( ) thus, using
=
1
( ) =
1
(9.30)
is the 4-dimensional Gausss equation. Now consider the case we have an isolated point charge
which has somehow always existed at the origin. Moreover consider a 3-sphere that surrounds the
charge. We wish to determine the generalized Coulomb eld due to the point charge. First we note
that the solid 3-sphere is a 4-dimensional object, it the set of all (, , , )
4
such that
2
+
2
+
2
+
2

2
We may parametrize a three-sphere of radius via generalized spherical coordinates,
= ()()()
= ()()()
= ()()
= ()
(9.31)
Now it can be shown that the volume and surface area of the radius three-sphere are as follows,
(
3
) =

2
2

4
(
3
) = 2
2
3
We may write the charge density of a smeared out point charge as,
=
_
2/
2
4
, 0
0, >
. (9.32)
Notice that if we integrate over any four-dimensional region which contains the solid three sphere
of radius will give the enclosed charge to be . Then integrate over the Gaussian 3-sphere
3
with radius call it ,
( ) =
1

now use the Generalized Stokes Theorem to deduce,
( ) =

but by the spherical symmetry of the problem we nd that must be independent of the direction
it points, this means that it can only have a radial component. Thus we may calculate the integral
with respect to generalized spherical coordinates and we will nd that it is the product of

and the surface volume of the four dimensional solid three sphere. That is,
( ) = 2
2
3
=


Thus,
=

2
2
3
the Coulomb eld is weaker if it were to propogate in 4 spatial dimensions. Qualitatively what has
happened is that the have taken the same net ux and spread it out over an additional dimension,
this means it thins out quicker. A very similar idea is used in some brane world scenarios. String
theorists posit that the gravitational eld spreads out in more than four dimensions while in con-
trast the standard model elds of electromagnetism, and the strong and weak forces are conned
to a four-dimensional brane. That sort of model attempts an explaination as to why gravity is so
weak in comparison to the other forces. Also it gives large scale corrections to gravity that some
hope will match observations which at present dont seem to t the standard gravitational models.
This example is but a taste of the theoretical discussion that dierential forms allow. As a
nal comment I remind the reader that we have done things for at space for the most part in
this course, when considering a curved space there are a few extra considerations that must enter.
Coordinate vector elds
must be thought of as derivations /
for one. Also the metric is not

a constant tensor like
or
rather is depends on position, this means Hodge duality aquires

a coordinate dependence as well. Doubtless I have forgotten something else in this brief warning.
One more advanced treatment of many of our discussions is Dr. Fulps Fiber Bundles 2001 notes
which I have posted on my webpage. He uses the other metric but it is rather elegantly argued, all
his arguments are coordinate independent. He also deals with the issue of the magnetic induction
and the dielectric, issues which we have entirely ignored since we always have worked in free space.
References and Acknowledgements:
I have drawn from many sources to assemble the content of the last couple chapters, the refer-
ences are listed approximately in the order of their use to the course, additionally we are indebted
to Dr. Fulp for his course notes from many courses (ma 430, ma 518, ma 555, ma 756, ...). Also
Manuela Kulaxizi helped me towards the correct (I hope) interpretation of 5-dimensional E&M in
the last example.
Vector Calculus, Susan Jane Colley
Introduction to Special Relativity, Robert Resnick
Dierential Forms and Connections, R.W.R. Darling
Dierential geometry, gauge theories, and gravity, M. Gockerler & T. Sch ucker
Anomalies in Quantum Field Theory, Reinhold A. Bertlmann
The Dierential Geometry and Physical Basis for the Applications of Feynman Diagrams, S.L.
Marateck, Notices of the AMS, Vol. 53, Number 7, pp. 744-752
Abstract Linear Algebra, Morton L. Curtis
Gravitation, Misner, Thorne and Wheeler
Introduction to Special Relativity, Wolfgang Rindler
Dierential Forms A Complement to Vector Calculus, Steven H. Weintraub
Dierential Forms with Applications to the Physical Sciences, Harley Flanders
Introduction to Electrodynamics, (3rd ed.) David J. Griths
The Geometry of Physics: An Introduction, Theodore Frankel
An Introduction to General Relativity: Spacetime and Geometry, Sean M. Carroll
Gauge Theory and Variational Principles, David Bleeker
Group Theory in Physics, Wu-Ki Tung
Chapter 10
supermath
273

Advanced Calculus Notes

Uploaded by

Copyright:

Available Formats

Advanced Calculus Notes

Uploaded by

Document Information

Original Description:

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

Advanced Calculus Notes

Uploaded by

Copyright:

Available Formats

Lecture Notes for Advanced Calculus

. In addition, I hope the exercise of abstracting calculus to supernumbers

= { there exists with

= { for all we have

is a set for = 1, 2, . . . then we denote the

for each = 1, 2, . . . . An element

for = 1, 2, . . . , }. In this course if we consider

) has -th component

. Equality of two vectors is only true if all

by the following rules: for each

and produces another vector +

. The operation is called scalar

and produces another

. Often we simply denote by juxtaposition .

. The size of the vector

the -th standard basis vector.

} is simply a nite weighted-sum of the objects from

are abstracted in linear algebra to form the denition of an abstract vector

also has a dot-product. The dot-product is a mapping from

. We can select the -th component by

has given us the length of the vector in the -th direction.

is a real number which is dened by =

we dene the angle between and by

. together with these dentions of length and angle forms a Euclidean

is harder to justify. The inequality that we need for it to be reasonable

1, otherwise we would not have a number in the (

and = [6, 7, 8, 9, 10]

nd the angle between these vectors

The angle is calculated from the denition of angle,

and suppose then

a normed linear space.

, matrix-valued functions and complex-valued functions of

be the distance function then

such that = 0 = where =

where we dened the

sin() and the direction can

. The directed line segment from to is

1 (, , ) {(1, 2, 3), (3, 1, 2), (2, 3, 1)}

standard basis expansion

: is the composite of with .

and suppose further that for each ,

) and for each

for each = 1, 2, . . . . This allows us to express each of the

be a mapping. If then we dene the restriction of

() = () for all . If and

be a mapping, if there exists a mapping : () such that

then is the inverse mapping of and we denote =

be a mapping. Furthermore, suppose that : is a

be a mapping. The restriction of to a cross section

be a mapping then we say a mapping is a local inverse of

. In the study of functions of one real variable

is the subset all points in

which are less

. We say is an interior point of i there exists some open ball

is an open set i for each there exists > 0 such that

is an closed set i its complement

be a mapping. We say that has limit

with 0 < < implies

be a mapping. If is a limit point of then we say that