Multivariable Calculus

Jerry Shurman
Reed College
Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix

1 Results from One-Variable Calculus . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 The Real Number System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Foundational and Basic Theorems . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Taylors Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

Part I Multivariable Differential Calculus

2 Euclidean Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.1 Algebra: Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 Geometry: Length and Angle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3 Analysis: Continuous Mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.4 Topology: Compact Sets and Continuity . . . . . . . . . . . . . . . . . . . . 51
2.5 Review of the One-Variable Derivative . . . . . . . . . . . . . . . . . . . . . 59
2.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

3 Linear Mappings and Their Matrices . . . . . . . . . . . . . . . . . . . . . . 63

3.1 Linear Mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.2 Operations on Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.3 The Inverse of a Linear Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.4 Inhomogeneous Linear Equations . . . . . . . . . . . . . . . . . . . . . . . . . . 91
3.5 The Determinant: Characterizing Properties and Their
Consequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.6 The Determinant: Uniqueness and Existence . . . . . . . . . . . . . . . . 101
3.7 An Explicit Formula for the Inverse . . . . . . . . . . . . . . . . . . . . . . . . 112
3.8 Geometry of the Determinant: Volume . . . . . . . . . . . . . . . . . . . . . 114
3.9 Geometry of the Determinant: Orientation . . . . . . . . . . . . . . . . . . 123
3.10 The Cross Product, Lines, and Planes in R3 . . . . . . . . . . . . . . . . 126
3.11 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
vi Contents

4 The Derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

4.1 Trying to Extend the Symbol-Pattern: Immediate, Irreparable
Catastrophe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
4.2 New Environment: the BachmannLandau Notation . . . . . . . . . 136
4.3 One-Variable Revisionism; the Derivative Redefined . . . . . . . . . . 143
4.4 Basic Results and the Chain Rule . . . . . . . . . . . . . . . . . . . . . . . . . 150
4.5 Calculating the Derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
4.6 Higher Order Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
4.7 Extreme Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
4.8 Directional Derivatives and the Gradient . . . . . . . . . . . . . . . . . . . 188
4.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200

5 Inverse and Implicit Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

5.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
5.2 The Inverse Function Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
5.3 The Implicit Function Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
5.4 Lagrange Multipliers: Geometric Motivation and Specific
Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
5.5 Lagrange Multipliers: Analytic Proof and General Examples . . 244
5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254

Part II Multivariable Integral Calculus

6 Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
6.1 Machinery: Boxes, Partitions, and Sums . . . . . . . . . . . . . . . . . . . . 257
6.2 Definition of the Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
6.3 Continuity and Integrability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
6.4 Integration of Functions of One Variable . . . . . . . . . . . . . . . . . . . . 281
6.5 Integration Over Nonboxes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
6.6 Fubinis Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
6.7 Change of Variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
6.8 Topological Preliminaries for the Change of Variable Theorem 331
6.9 Proof of the Change of Variable Theorem . . . . . . . . . . . . . . . . . . . 339
6.10 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352

7 Approximation by Smooth Functions . . . . . . . . . . . . . . . . . . . . . . 353

7.1 Spaces of Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
7.2 Pulse Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
7.3 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362
7.4 Test Approximate Identity and Convolution . . . . . . . . . . . . . . . . 368
7.5 Known-Integrable Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374
7.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378
 Contents vii

8 Parametrized Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379

8.1 Euclidean Constructions and Two Curves . . . . . . . . . . . . . . . . . . . 379
8.2 Parametrized Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388
8.3 Parametrization by Arc Length . . . . . . . . . . . . . . . . . . . . . . . . . . . 394
8.4 Plane Curves: Curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398
8.5 Space Curves: Curvature and Torsion . . . . . . . . . . . . . . . . . . . . . . 402
8.6 General Frenet Frames and Curvatures . . . . . . . . . . . . . . . . . . . . . 408
8.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412

9 Integration of Differential Forms . . . . . . . . . . . . . . . . . . . . . . . . . . . 413

9.1 Integration of Functions Over Surfaces . . . . . . . . . . . . . . . . . . . . . 414
9.2 Flow and Flux Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 422
9.3 Differential Forms Syntactically and Operationally . . . . . . . . . . . 428
9.4 Examples: 1-forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431
9.5 Examples: 2-forms on R3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
9.6 Algebra of Forms: Basic Properties . . . . . . . . . . . . . . . . . . . . . . . . 442
9.7 Algebra of Forms: Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . 443
9.8 Algebra of Forms: Differentiation . . . . . . . . . . . . . . . . . . . . . . . . . . 446
9.9 Algebra of Forms: the Pullback . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
9.10 Change of Variable for Differential Forms . . . . . . . . . . . . . . . . . . . 464
9.11 Closed Forms, Exact Forms, and Homotopy . . . . . . . . . . . . . . . . . 466
9.12 Cubes and Chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
9.13 Geometry of Chains: the Boundary Operator . . . . . . . . . . . . . . . . 474
9.14 The General Fundamental Theorem of Integral Calculus . . . . . . 481
9.15 Classical Change of Variable Revisited . . . . . . . . . . . . . . . . . . . . . 486
9.16 The Classical Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492
9.17 Divergence and Curl in Polar Coordinates . . . . . . . . . . . . . . . . . . 498
9.18 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507
Preface

This is the text for a two-semester multivariable calculus course. The setting is
n-dimensional Euclidean space, with the material on differentiation culminat-
ing in the Inverse Function Theorem and its consequences, and the material
on integration culminating in the Generalized Fundamental Theorem of Inte-
gral Calculus (often called Stokess Theorem) and some of its consequences in
turn. The prerequisite is a proof-based course in one-variable calculus. Some
familiarity with the complex number system and complex mappings is occa-
sionally assumed as well, but the reader can get by without it.
The books aim is to use multivariable calculus to teach mathematics as
a blend of reasoning, computing, and problem-solving, doing justice to the
structure, the details, and the scope of the ideas. To this end, I have tried
to write in a style that communicates intent early in the discussion of each
topic rather than proceeding coyly from opaque definitions. Also, I have tried
occasionally to speak to the pedagogy of mathematics and its effect on the
process of learning the subject. Most importantly, I have tried to spread the
weight of exposition among figures, formulas, and words. The premise is that
the reader is ready to do mathematics resourcefully by marshaling the skills
of
geometric intuition (the visual cortex being quickly instinctive),
algebraic manipulation (symbol-patterns being precise and robust),
and incisive use of natural language (slogans that encapsulate central ideas
enabling a large-scale grasp of the subject).
Thinking in these ways renders mathematics coherent, inevitable, and fluent.
In my own student days I learned this material from books by Apostol,
Buck, Rudin, and Spivak, books that thrilled me. My debt to those sources
pervades these pages. There are many other fine books on the subject as well,
such as the more recent one by Hubbard and Hubbard. Indeed, nothing in
this book is claimed as new, not even its neuroses. Whatever improvement
the exposition here has come to show over the years is due to innumerable
ideas and comments from my students in turn.
x Preface

By way of a warm-up, chapter 1 reviews some ideas from one-variable

calculus, and then covers the one-variable Taylors Theorem in detail.
Chapters 2 and 3 cover what might be called multivariable pre-calculus, in-
troducing the requisite algebra, geometry, analysis, and topology of Euclidean
space, and the requisite linear algebra, for the calculus to follow. A pedagogical
theme of these chapters is that mathematical objects can be better understood
from their characterizations than from their constructions. Vector geometry
follows from the intrinsic (coordinate-free) algebraic properties of the vector
inner product, with no reference to the inner product formula. The fact that
passing a closed and bounded subset of Euclidean space through a continuous
mapping gives another such set is clear once such sets are characterized in
terms of sequences. The multiplicativity of the determinant and the fact that
the determinant indicates whether a linear mapping is invertible are conse-
quences of the determinants characterizing properties. The geometry of the
cross product follows from its intrinsic algebraic characterization. Further-
more, the only possible formula for the (suitably normalized) inner product,
or for the determinant, or for the cross product, is dictated by the relevant
properties. As far as the theory is concerned, the only role of the formula is
to show that an object with the desired properties exists at all. The intent
here is that the student who is introduced to mathematical objects via their
characterizations will see quickly how the objects work, and that how they
work makes their constructions inevitable.
In the same vein, chapter 4 characterizes the multivariable derivative as a
well approximating linear mapping. The chapter then solves some multivari-
able problems that have one-variable counterparts. Specifically, the multivari-
able chain rule helps with change of variable in partial differential equations,
a multivariable analogue of the max/min test helps with optimization, and
the multivariable derivative of a scalar-valued function helps to find tangent
planes and trajectories.
Chapter 5 uses the results of the three chapters preceding it to prove the
Inverse Function Theorem, then the Implicit Function Theorem as a corollary,
and finally the Lagrange Multiplier Criterion as a consequence of the Implicit
Function Theorem. Lagrange multipliers help with a type of multivariable
optimization problem that has no one-variable analogue, optimization with
constraints. For example, given two curves in space, what pair of points
one on each curveis closest to each other? Not only does this problem have
six variables (the three coordinates of each point), but furthermore they are
not fully independent: the first three variables must specify a point on the
first curve, and similarly for the second three. In this problem, x1 through x6
vary though a subset of six-dimensional space, conceptually a two-dimensional
subset (one degree of freedom for each curve) that is bending around in the
ambient six dimensions, and we seek points of this subset where a certain
function of x1 through x6 is optimized. That is, optimization with constraints
can be viewed as a beginning example of calculus on curved spaces.
 Preface xi

For another example, let n be a positive integer, and let e1 , , en be

positive numbers with e1 + + en = 1. Maximize the function

f (x1 , , xn ) = xe11 xenn , xi 0 for all i,

subject to the constraint that

e1 x1 + + en xn = 1.

As in the previous paragraph, since this problem involves one condition on

the variables x1 through xn , it can be viewed as optimizing over an (n 1)-
dimensional space inside n dimensions. The problem may appear unmotivated,
but its solution leads
quickly to a generalization of the arithmetic-geometric
mean inequality ab (a + b)/2 for all nonnegative a and b,

ae11 aenn e1 a1 + + en an for all nonnegative a1 , , an .

Moving to integral calculus, chapter 6 introduces the integral of a scalar-

valued function of many variables, taken over a domain of its inputs. When the
domain is a box, the definitions and the basic results are essentially the same as
for one variable. However, in multivariable calculus we want to integrate over
regions other than boxes, and ensuring that we can do so takes a little work.
After this is done, the chapter proceeds to two main tools for multivariable
integration, Fubinis Theorem and the Change of Variable Theorem. Fubinis
Theorem reduces one n-dimensional integral to n one-dimensional integrals,
and the Change of Variable Theorem replaces one n-dimensional integral with
another that may be easier to evaluate. Using these techniques one can show,
for example, that the ball of radius r in n dimensions has volume

n/2 n
vol (Bn (r)) = r , n = 1, 2, 3, 4, .
(n/2)!
The meaning of the (n/2)! in the display when n is odd is explained by a
function called the gamma function. The sequence begins
4 3 1 2 4
2r, r2 , r , r , .
3 2
Chapter 7 discusses the fact that continuous functions, or differentiable
functions, or twice-differentiable functions, are well approximated by smooth
functions, meaning functions that can be differentiated endlessly. The approx-
imation technology is an integral called the convolution. One point here is that
the integral is useful in ways far beyond computing volumes. The second point
is that with approximation by convolution in hand, we feel free to assume in
the sequel that functions are smooth. The reader who is willing to grant this
assumption in any case can skip chapter 7.
Chapter 8 introduces parametrized curves as a warmup for chapter 9 to
follow. The subject of chapter 9 is integration over k-dimensional parametrized
xii Preface

surfaces in n-dimensional space, and parametrized curves are the special case
k = 1. Aside from being one-dimensional surfaces, parametrized curves are
interesting in their own right.
Chapter 9 presents the integration of differential forms. This subject poses
the pedagogical dilemma that fully describing its structure requires an in-
vestment in machinery untenable for students who are seeing it for the first
time, whereas describing it purely operationally is unmotivated. The approach
here begins with the integration of functions over k-dimensional surfaces in
n-dimensional space, a natural tool to want, with a natural definition suggest-
ing itself. For certain such integrals, called flow and flux integrals, the inte-
grand takes a particularly workable form consisting of sums of determinants
of derivatives. It is easy to see what other integrandsincluding integrands
suitable for n-dimensional integration in the sense of chapter 6, and includ-
ing functions in the usual sensehave similar features. These integrands can
be uniformly described in algebraic terms as objects called differential forms.
That is, differential forms assemble the smallest coherent algebraic structure
encompassing the various integrands of interest to us. The fact that differential
forms are algebraic makes them easy to study without thinking directly about
the analysis of integration. The algebra leads to a general version of the Fun-
damental Theorem of Integral Calculus that is rich in geometry. The theorem
subsumes the three classical vector integration theorems, Greens Theorem,
Stokess Theorem, and Gausss Theorem, also called the Divergence Theorem.
Comments and corrections should be sent to [email protected].

Exercises

0.0.1. (a) Consider two surfaces in space, each surface having at each of its
points a tangent plane and therefore a normal line, and consider pairs of
points, one on each surface. Conjecture a geometric condition, phrased in
terms of tangent planes and/or normal lines, about the closest pair of points.
(b) Consider a surface in space and a curve in space, the curve having at
each of its points a tangent line and therefore a normal plane, and consider
pairs of points, one on the surface and one on the curve. Make a conjecture
about the closest pair of points.
(c) Make a conjecture about the closest pair of points on two curves.
0.0.2. (a) Assume that the factorial of a half-integer makes sense, and grant
the general formula for the volume of a ball in n dimensions. Explain why
it follows that (1/2)! = /2. Further assume that the half-integral factorial
function satisfies the relation
x! = x (x 1)! for x = 3/2, 5/2, 7/2, .
Subject to these assumptions, verify that the volume of the ball of radius r
in three dimensions is 43 r3 as claimed. What is the volume of the ball of
radius r in five dimensions?
 Preface xiii

(b) The ball of radius r in n dimensions sits inside a circumscribing box of

sides 2r. Draw pictures of this configuration for n = 1, 2, 3. Determine what
portion of the box is filled by the ball in the limit as the dimension n gets
large. That is, find
vol (Bn (r))
lim .
n (2r)n
1
Results from One-Variable Calculus

As a warmup, these notes begin with a quick review of some ideas from one-
variable calculus. The material in the first two sections is assumed to be
familiar. Section 3 discusses Taylors Theorem at greater length, not assuming
that the reader has already seen it.

1.1 The Real Number System

We assume that there is a real number system, a set R that contains two
distinct elements 0 and 1 and is endowed with the algebraic operations of
addition,
+ : R R R,
and multiplication,
: R R R.
The sum +(a, b) is written a + b, and the product (a, b) is written a b or
simply ab.
Theorem 1.1.1 (Field Axioms for (R, +, )). The real number system, with
its distinct 0 and 1 and with its addition and multiplication, is assumed to
satisfy the following set of axioms.
(a1) Addition is associative: (x + y) + z = x + (y + z) for all x, y, z R.
(a2) 0 is an additive identity: 0 + x = x for all x R.
(a3) Existence of additive inverses: For each x R there exists y R such
that y + x = 0.
(a4) Addition is commutative: x + y = y + x for all x, y R.
(m1) Multiplication is associative: (xy)z = x(yz) for all x, y, z R.
(m2) 1 is a multiplicative identity: 1x = x for all x R.
(m3) Existence of multiplicative inverses: For each nonzero x R there exists
y R such that yx = 1.
(m4) Multiplication is commutative: xy = yx for all x, y R.
2 1 Results from One-Variable Calculus

(d1) Multiplication distributes over addition: (x+y)z = xz+yz for all x, y, z

R.

All of basic algebra follows from the field axioms. Additive and multi-
plicative inverses are unique, the cancellation law holds, 0 x = 0 for all real
numbers x, and so on.
Subtracting a real number from another is defined as adding the additive
inverse. In symbols,

: R R R, x y = x + (y) for all x, y R.

We also assume that R is an ordered field. That is, we assume that there
is a subset R+ of R (the positive elements) such that the following axioms
hold.

Theorem 1.1.2 (Order Axioms).

(o1) Trichotomy Axiom: For every real number x, exactly one of the following
conditions holds:

x R+ , x R+ , x = 0.

(o2) Closure of positive numbers under addition: For all real numbers x and y,
if x R+ and y R+ then also x + y R+ .
(o3) Closure of positive numbers under multiplication: For all real numbers x
and y, if x R+ and y R+ then also xy R+ .

For all real numbers x and y, define

x<y

to mean
y x R+ .
The usual rules for inequalities then follow from the axioms.
Finally, we assume that the real number system is complete. Complete-
ness can be phrased in various ways, all logically equivalent. A version of
completeness that is phrased in terms of binary search is as follows.

Theorem 1.1.3 (Completeness as a Binary Search Criterion). Every

binary search sequence in the real number system converges to a unique limit.

Convergence is a concept of analysis, and therefore so is completeness.

Another version of completeness is phrased in terms of set-bounds.

Theorem 1.1.4 (Completeness as a Set-Bound Criterion). Every non-

empty subset of R that is bounded above has a least upper bound.
 1.1 The Real Number System 3

Both statements of completeness are existence statements.

A subset S of R is inductive if
(i1) 0 S,
(i2) For all x R, if x S then x + 1 S.
Any intersection of inductive subsets of R is again inductive. The set of natu-
ral numbers, denoted N, is the intersection of all inductive subsets of R, i.e.,
N is the smallest inductive subset of R. There is no natural number between 0
and 1 (because if there were then deleting it from N would leave a smaller
inductive subset of R), and so

N = {0, 1, 2 , }.

Theorem 1.1.5 (Induction Theorem). Let P (n) be a proposition form de-

fined over N. Suppose that
P (0) holds.
For all n N, if P (n) holds then so does P (n + 1).
Then P (n) holds for all natural numbers n.

Indeed, the hypotheses of the theorem say that P (n) holds for a subset
of N that is inductive, and so the theorem follows from the definition of N as
the smallest inductive subset of R.
The Archimedean Principle states that the subset N of R is not
bounded above. Equivalently, the sequence {1, 1/2, 1/3, } converges to 0.
The Archimedean Principle follows from the assumption that R is complete
in the sense of binary search sequences or in the sense of set-bounds. A third
version of completeness is phrased in terms of monotonic sequences. Again it
is an existence statement.

Theorem 1.1.6 (Completeness as a Monotonic Sequence Criterion).

Every bounded monotonic sequence in R converges to a unique limit.

This version of completeness follows from either of the other two. How-
ever, it does not imply the other two unless we also assume the Archimedean
Principle.
The set of integers, denoted Z, is the union of the natural numbers and
their additive inverses,
Z = {0, 1, 2 , }.

Exercises

1.1.1. Referring only to the field axioms, show that 0x = 0 for all x R.

1.1.2. Prove that in any ordered field, 1 is positive. Prove that the complex
number field C can not be made an ordered field.
4 1 Results from One-Variable Calculus

1.1.3. Use a completeness property of the real number system to show that 2
has a positive square root.

1.1.4. (a) Prove by induction that

n
X n(n + 1)(2n + 1)
i2 = for all n Z+ .
i=1
6

(b) (Bernoullis Inequality) For any real number r 1, prove that

(1 + r)n 1 + rn for all n N.

(c) For what positive integers n is 2n > n3 ?

1.1.5. (a) Use the Induction Theorem to show that for any natural number m,
the sum m+n and the product mn are again natural for any natural number n.
Thus N is closed under addition and multiplication, and consequently so is Z.
(b) Which of the field axioms continue to hold for the natural numbers?
(c) Which of the field axioms continue to hold for the integers?

1.1.6. For any positive integer n, let Z/nZ denote the set {0, 1, . . . , n 1}
with the usual operations of addition and multiplication carried out taking
remainders. That is, add and multiply in the usual fashion but subject to the
additional condition that n = 0. For example, in Z/5Z we have 2 + 4 = 1 and
2 4 = 3. For what values of n does Z/nZ form a field?

1.2 Foundational and Basic Theorems

This section reviews the foundational theorems of one-variable calculus. The
first two theorems are not theorems of calculus at all, but rather they are
theorems about continuous functions and the real number system. The first
theorem says that under suitable conditions, an optimization problem is guar-
anteed to have a solution.

Theorem 1.2.1 (Extreme Value Theorem). Let I be a nonempty closed

and bounded interval in R, and let f : I R be a continuous function. Then
f takes a minimum value and a maximum value on I.

The second theorem says that under suitable conditions, any value trapped
between two output values of a function must itself be an output value.

Theorem 1.2.2 (Intermediate Value Theorem). Let I be a nonempty

interval in R, and let f : I R be a continuous function. Let y be a real
number, and suppose that

f (x) < y for some x I

 1.2 Foundational and Basic Theorems 5

and
f (x ) > y for some x I.
Then
f (c) = y for some c I.

The Mean Value Theorem relates the derivative of a function to values of

the function itself with no reference to the fact that the derivative is a limit,
but at the cost of introducing an unknown point.

Theorem 1.2.3 (Mean Value Theorem). Let a and b be real numbers

with a < b. Suppose that the function f : [a, b] R is continuous and that
f is differentiable on the open subinterval (a, b). Then

f (b) f (a)
= f (c) for some c (a, b).
ba
The Fundamental Theorem of Integral Calculus quantifies the idea that
integration and differentiation are inverse operations. In fact two different
results are both called the Fundamental Theorem, one a result about the
derivative of the integral and the other a result about the integral of the
derivative. Fundamental Theorem of Calculus, unmodified, usually refers
to the second of the next two results.

Theorem 1.2.4 (Fundamental Theorem of Integral Calculus I). Let I

be a nonempty interval in R, let a be a point of I, and let f : I R be a
continuous function. Define a second function,
Z x
F : I R, F (x) = f (t) dt.
a

Then F is differentiable on I with derivative F (x) = f (x) for all x I.

Theorem 1.2.5 (Fundamental Theorem of Integral Calculus II). Let

I be a nonempty interval in R, and let f : I R be a continuous function.
Suppose that the function F : I R has derivative f . Then for any closed
and bounded subinterval [a, b] of I,
Z b
f (x) dx = F (b) F (a).
a

Exercises

1.2.1. Use the Intermediate Value Theorem to show that 2 has a positive
square root.

1.2.2. Let f : [0, 1] [0, 1] be continuous. Use the Intermediate Value

Theorem to show that f (x) = x for some x [0, 1].
6 1 Results from One-Variable Calculus

1.2.3. Let a and b be real numbers with a < b. Suppose that f : [a, b] R
is continuous and that f is differentiable on the open subinterval (a, b). Use
the Mean Value Theorem to show that if f > 0 on (a, b) then f is strictly
increasing on [a, b]. (Note: The quantities called a and b in the Mean Value
Theorem when you cite it to solve this exercise will not be the a and b given
here. It may help to review the definition of strictly increasing.)

1.2.4. For the Extreme Value Theorem, the Intermediate Value Theorem,
and the Mean Value Theorem, give examples to show that weakening the
hypotheses of the theorem gives rise to examples where the conclusion of the
theorem fails.

1.3 Taylors Theorem

Let I R be a nonempty open interval, and let a I be any point. Let n be a
nonnegative integer. Suppose that the function f : I R has n continuous
derivatives,
f, f , f , . . . , f (n) : I R.
Suppose further that we know the values of f and its derivatives at a, the
n + 1 numbers

f (a), f (a), f (a), ..., f (n) (a).

(For instance, if f : R R is the cosine function, and a = 0, and n is even,

then the numbers are 1, 0, 1, 0, . . . , (1)n/2 .)
Question 1 (Existence and Uniqueness): Is there a polynomial p of
degree n that mimics the behavior of f at a in the sense that

p(a) = f (a), p (a) = f (a), p (a) = f (a), ..., p(n) (a) = f (n) (a)?

Is there only one such polynomial?

Question 2 (Accuracy of Approximation, Granting Existence and
Uniqueness): How well does p(x) approximate f (x) for x 6= a?
Question 1 is easy to answer. Consider a polynomial of degree n expanded
about x = a,

p(x) = a0 + a1 (x a) + a2 (x a)2 + a3 (x a)3 + + an (x a)n .

The goal is to choose the coefficients a0 , . . . , an to make p behave like the

original function f at a. Note that p(a) = a0 . We want p(a) to equal f (a), so
set
a0 = f (a).
Differentiate p to obtain
 1.3 Taylors Theorem 7

p (x) = a1 + 2a2 (x a) + 3a3 (x a)2 + + nan (x a)n1 ,

so that p (a) = a1 . We want p (a) to equal f (a), so set

a1 = f (a).

Differentiate again to obtain

p (x) = 2a2 + 3 2a3 (x a) + + n(n 1)an (x a)n2 ,

so that p (a) = 2a2 . We want p (a) to equal f (a), so set

f (a)
a2 = .
2
Differentiate again to obtain

p (x) = 3 2a3 + + n(n 1)(n 2)an (x a)n3 ,

so that p (a) = 3 2a3 . We want p (a) to equal f (a), so set

f (a)
a3 = .
32
Continue in this fashion to obtain a4 = f (4) (a)/4! and so on up to an =
f (n) (a)/n!. That is, the desired coefficients are

f (k) (a)
ak = for k = 0, . . . , n.
k!
Thus the answer to the existence part of Question 1 is Yes. Furthermore, since
the calculation offered us no choices en route, these are the only coefficients
that can work, and so the approximating polynomial is unique. It deserves a
name.

Definition 1.3.1 (nth degree Taylor Polynomial). Let I R be a

nonempty open interval, and let a be a point of I. Let n be a nonnegative
integer. Suppose that the function f : I R has n continuous derivatives.
Then the nth degree Taylor polynomial of f at a is

f (a) f (n) (a)

Tn (x) = f (a) + f (a)(x a) + (x a)2 + + (x a)n .
2 n!
In more concise notation,
n
X f (k) (a)
Tn (x) = (x a)k .
k!
k=0
8 1 Results from One-Variable Calculus

For example, if f (x) = ex and a = 0 then it is easy to generate the

following table:
f (k) (0)
k f (k) (x)
k!
0 ex 1
1 ex 1
1
2 ex
2
1
3 ex
3!
.. .. ..
. . .
1
n ex
n!
From the table we can read off the nth degree Taylor polynomial of f at 0,

x2 x3 xn
Tn (x) = 1 + x + + + +
2 3! n!
n
X xk
= .
k!
k=0

Recall that the second question is how well the polynomial Tn (x) approxi-
mates f (x) for x 6= a. Thus it is a question about the difference f (x) Tn (x).
Giving this quantity its own name is useful.

Definition 1.3.2 (nth degree Taylor Remainder). Let I R be a

nonempty open interval, and let a be a point of I. Let n be a nonnegative
integer. Suppose that the function f : I R has n continuous derivatives.
Then the nth degree Taylor remainder of f at a is

Rn (x) = f (x) Tn (x).

So the second question is to estimate the remainder Rn (x) for points x I.

The method to be presented here for doing so proceeds very naturally, but it
is perhaps a little surprising because although the Taylor polynomial Tn (x)
is expressed in terms of derivatives, as is the expression to be obtained for
the remainder Rn (x), we obtain the expression by using the Fundamental
Theorem of Integral Calculus repeatedly.
The method requires a calculation, and so, guided by hindsight, we first
carry it out so that then the ideas of the method itself will be uncluttered.
For any positive integer k and any x R, define a k-fold nested integral,
Z x Z x1 Z xk1
Ik (x) = dxk dx2 dx1 .
x1 =a x2 =a xk =a
 1.3 Taylors Theorem 9

This nested integral is a function only of x because a is a constant while x1

through xk are dummy variables of integration. That is, Ik depends only on
the upper limit of integration of the outermost integral. Although Ik may
appear daunting, it unwinds readily if we start from the simplest case. First,
Z x x

I1 (x) = dx1 = x1  = x a.
x1 =a x1 =a

Move one layer out and use this result to get

Z x Z x1 Z x
I2 (x) = dx2 dx1 = I1 (x1 ) dx1
x1 =a x2 =a x1 =a
Z x x
1  1
= (x1 a) dx1 = (x1 a)2  = (x a)2 .
x1 =a 2 x1 =a 2

Again move out and quote the previous calculation,

Z x Z x1 Z x2 Z x
I3 (x) = dx3 dx2 dx1 = I2 (x1 ) dx1
x1 =a x2 =a x3 =a x1 =a
Z x x
1 1  1
= (x1 a)2 dx1 = (x1 a)3  = (x a)3 .
x1 =a 2 3! x1 =a 3!

The method and pattern are clear, and the answer in general is
1
Ik (x) = (x a)k , k Z+ .
k!
f (k) (a)
Note that this is part of the kth term (xa)k of the Taylor polynomial,
k!
the part that makes no reference to the function f . That is, f (k) (a)Ik (x) is
the kth term of the Taylor polynomial for k = 1, 2, 3, . . .
With the formula for Ik (x) in hand, we return to using the Fundamental
Theorem of Integral Calculus to study the remainder Rn (x), the function f (x)
minus its nth degree Taylor polynomial Tn (x). According to the Fundamental
Theorem, Z x
f (x) = f (a) + f (x1 ) dx1 ,
a

That is, f (x) is equal to the constant term of the Taylor polynomial plus an
integral, Z x
f (x) = T0 (x) + f (x1 ) dx1 .
a
By the Fundamental Theorem again, the integral is in turn
Z x Z x Z x1 
f (x1 ) dx1 = f (a) + f (x2 ) dx2 dx1 .
a a a
10 1 Results from One-Variable Calculus

The first term of the outer integral is f (a)I1 (x), giving the first-order term
of the Taylor polynomial and leaving a doubly-nested integral,
Z x Z x Z x1

f (x1 ) dx1 = f (a)(x a) + f (x2 ) dx2 dx1 .
a a a

In other words, the calculation so far has shown that

Z x Z x1
f (x) = f (a) + f (a)(x a) + f (x2 ) dx2 dx1
a a
Z x Z x1
= T1 (x) + f (x2 ) dx2 dx1 .
a a

Once more by the Fundamental Theorem the doubly-nested integral is

Z x Z x1 Z x Z x1  Z x2 
f (x2 ) dx2 dx1 = f (a) + f (x3 ) dx3 dx2 dx1 ,
a a a a a

and the first term of the outer integral is f (a)I2 (x), giving the second-order
term of the Taylor polynomial and leaving a triply-nested integral,
Z x Z x1 Z x Z x1 Z x2
f (a)
f (x2 ) dx2 dx1 = 2
(x a) + f (x3 ) dx3 dx2 dx1 .
a a 2 a a a

So now the calculation so far has shown that

Z x Z x1 Z x2
f (x) = T2 (x) + f (x3 ) dx3 dx2 dx1 .
a a a

Continuing this process through n iterations shows that f (x) is Tn (x) plus an
(n + 1)-fold iterated integral,
Z x Z x1 Z xn
f (x) = Tn (x) + f (n+1) (xn+1 ) dxn+1 dx2 dx1 .
a a a

In other words, the remainder is the integral,

Z x Z x1 Z xn
Rn (x) = f (n+1) (xn+1 ) dxn+1 dx2 dx1 . (1.1)
a a a

Note that we now are assuming that f has n + 1 continuous derivatives.

For simplicity, assume that x > a. Since f (n+1) is continuous on the closed
and bounded interval [a, x], the Extreme Value Theorem says that it takes a
minimum value m and a maximum value M on the interval. That is,

m f (n+1) (xn+1 ) M, xn+1 [a, x].

Integrate these two inequalities n + 1 times to bound the remainder inte-

gral (1.1) on both sides by multiples of the integral that we have evaluated,
 1.3 Taylors Theorem 11

mIn+1 (x) Rn (x) M In+1 (x),

and therefore by the precalculated formula for In+1 (x),

(x a)n+1 (x a)n+1
m Rn (x) M . (1.2)
(n + 1)! (n + 1)!

Recall that m and M are particular values of f (n+1) . Define an auxiliary

function that will therefore assume the sandwiching values in (1.2),

(x a)n+1
g : [a, x] R, g(t) = f (n+1) (t) .
(n + 1)!

That is, since there exist values tm and tM in [a, x] such that f (n+1) (tm ) = m
and f (n+1) (tM ) = M , the result (1.2) of our calculation rephrases as

g(tm ) Rn (x) g(tM ).

The inequalities show that the remainder is an intermediate value of g. And

g is continuous, so by the Intermediate Value Theorem, there exists some
point c [a, x] such that g(c) = Rn (x). In other words, g(c) is the desired
remainder, the function minus its Taylor polynomial. We have proved

Theorem 1.3.3 (Taylors Theorem). Let I R be a nonempty open in-

terval, and let a I. Let n be a nonnegative integer. Suppose that the function
f : I R has n + 1 continuous derivatives. Then for each x I,

f (x) = Tn (x) + Rn (x)

where
f (n+1) (c)
Rn (x) = (x a)n+1 for some c between a and x.
(n + 1)!
We have proved Taylors Theorem only when x > a. It is trivial for x = a.
If x < a, then rather than repeat the proof while keeping closer track of signs,
with some of the inequalities switching direction, we may define

f : I R, f(x) = f (x).

Since f = f neg where neg is the negation function, a small exercise with
the chain rule shows that

f(k) (x) = (1)k f (k) (x), for k = 0, . . . , n + 1 and x I.

If x < a in I then x > a in I, and so we know by the version of Taylors

Theorem that we have already proved that

f(x) = Ten (x) + R

en (x)
12 1 Results from One-Variable Calculus

where
n (k)
X f (a)
Ten (x) = (x (a))k
k!
k=0

and
(n+1) (c)
en (x) = f
R (x (a))n+1 for some c between a and x.
(n + 1)!

But f(x) = f (x), and Ten (x) is precisely the desired Taylor polyno-
mial Tn (x),
n (k)
X f (a)
Ten (x) = (x (a))k
k!
k=0
Xn n
X f (k) (a)
(1)k f (k) (a)
= (1)k (x a)k = (x a)k = Tn (x),
k! k!
k=0 k=0

en (x) works out to the desired form of Rn (x),

and similarly R
(n+1)
en (x) = f
R
(c)
(x a)n+1 for some c between a and x.
(n + 1)!
Thus we obtain the statement of Taylors Theorem in the case x < a as well.
Whereas our proof of Taylors Theorem relies primarily on the Funda-
mental Theorem of Integral Calculus, and a similar proof relies on repeated
integration by parts (exercise 1.3.6), many proofs rely instead on the Mean
Value Theorem. Our proof neatly uses three different mathematical techniques
for the three different parts of the argument:
To find the Taylor polynomial Tn (x) we differentiated repeatedly, using a
substitution at each step to determine a coefficient.
To get a precise (if unwieldy) expression for the remainder Rn (x) =
f (x) Tn (x) we integrated repeatedly, using the Fundamental Theorem of
Integral Calculus at each step to produce a term of the Taylor polynomial.
To express the remainder in a more convenient form, we used the Extreme
Value Theorem and then the Intermediate Value Theorem once each. These
foundational theorems are not results from calculus but (as we will discuss
in section 2.4) from an area of mathematics called topology.
The expression for Rn (x) given in Theorem 1.3.3 is called the Lagrange
form of the remainder. Other expressions for Rn (x) exist as well. Whatever
form is used for the remainder, it should be something that we can estimate
by bounding its magnitude.
For example, we use Taylors Theorem to estimate ln(1.1) by hand to
within 1/500, 000. Let f (x) = ln(1 + x) on (1, ), and let a = 0. Compute
the following table:
 1.3 Taylors Theorem 13

f (k) (0)
k f (k) (x)
k!
0 ln(1 + x) 0
1
1 1
(1 + x)
1 1
2 2

(1 + x) 2
2 1
3
(1 + x)3 3
3! 1
4
(1 + x)4 4
.. .. ..
. . .
(1)n1 (n 1)! (1)n1
n
(1 + x)n n
(1)n n!
n+1
(1 + x)n+1

Next, read off from the table that for n 1, the nth degree Taylor polynomial
is
Xn
x2 x3 xn xk
Tn (x) = x + + (1)n1 = (1)k1 ,
2 3 n k
k=1

and the remainder is

(1)n xn+1
Rn (x) = for some c between 0 and x.
(1 + c)n+1 (n + 1)
This expression for the remainder may be a bit much to take in since it involves
three variables: the point x at which we are approximating the logarithm, the
degree n of the Taylor polynomial that is providing the approximation, and
the unknown value c in the error term. But in particular we are interested in
x = 0.1 (since we are approximating ln(1.1) using f (x) = ln(1 + x)), so that
the Taylor polynomial specializes to

(0.1)2 (0.1)3 (0.1)n

Tn (0.1) = (0.1) + + (1)n1 ,
2 3 n
and we want to bound the remainder in absolute value, so write

(0.1)n+1
|Rn (0.1)| = for some c between 0 and 0.1.
(1 + c)n+1 (n + 1)
Now the symbol x is gone. Next, note that although we dont know the value
of c, the smallest possible value of the quantity (1 + c)n+1 in the denominator
of the absolute remainder is 1 because c 0. And since this value occurs in
14 1 Results from One-Variable Calculus

the denominator it lets us write the greatest possible value of the absolute
remainder with no reference to c. That is,
(0.1)n+1
|Rn (0.1)| ,
(n + 1)
and the symbol c is gone as well. The only remaining variable is n, and the
goal is to approximate ln(1.1) to within 1/500, 000. Set n = 4 in the previous
display to get
1
|R4 (0.1)| .
500, 000
That is, the fourth degree Taylor polynomial
1 1 1 1
T4 (0.1) = +
10 200 3000 40000
= 0.10000000 0.00500000 + 0.00033333 0.00002500
= 0.09530833
agrees with ln(1.1) to within 0.00000200 , so that
0.09530633 ln(1.1) 0.09531033 .
Machine technology should confirm this.
Continuing to work with the function f (x) = ln(1 + x) for x > 1, set
x = 1 instead to get that for n 1,
1 1 1
Tn (1) = 1 + + (1)n1 ,
2 3 n
and  
 1 

|Rn (1)| =   for some c between 0 and 1.
(1 + c)n+1 (n + 1) 
Thus |Rn (1)| 1/(n + 1), and this goes to 0 as n . Therefore ln(2) is
expressible as an infinite series,
1 1 1
ln(2) = 1 + + .
2 3 4
This example illustrates an important general principle:
To check whether the Taylor polynomial Tn (x) converges to f (x) as n
grows, i.e., to check whether the infinite Taylor series

X f (k) (a)
T (x) = lim Tn (x) = (x a)k
n k!
k=0

reproduces f (x), check whether the remainder Rn (x) converges to 0.

To show that the remainder Rn (x) converges to 0, estimate |Rn (x)| in
a way that gets rid of the unknown c and then show that the estimate
goes to 0.
 1.3 Taylors Theorem 15

To repeat a formula from before, the nth degree Taylor polynomial of the
function ln(1 + x) is
X n
x2 x3 xn xk
Tn (x) = x + + (1)n1 = (1)k1 ,
2 3 n k
k=1

The graphs of the natural logarithm ln(x) and the first five Taylor polynomials
Tn (x 1) are plotted from 0 to 2 in figure 1.1. (The switch from ln(1 + x)
to ln(x) places the logarithm graph in its familiar position, and then the switch
from Tn (x) to Tn (x 1) is forced in consequence to fit the Taylor polynomials
through the repositioned function.) A good check of your understanding is to
see if you can determine which graph is which in the figure.

0.5 1 1.5 2

Figure 1.1. The natural logarithm and its Taylor polynomials

For another example, return to the exponential function f (x) = ex and

let a = 0. For any x, the difference between f (x) and the nth degree Taylor
polynomial Tn (x) satisfies
 
 c xn+1 

|Rn (x)| = e  for some c between 0 and x.
(n + 1)! 
If x 0 then ec could be as large as ex , while if x < 0 then ec could be as
large as e0 . The worst possible case is therefore
|x|n+1
|Rn (x)| max{1, ex } .
(n + 1)!
16 1 Results from One-Variable Calculus

As n (while x remains fixed, albeit arbitrary) the right side goes to 0

because the factorial growth of (n + 1)! dominates the polynomial growth
of |x|n+1 , and so we have in the limit that ex is expressible as a power series,

X xk
x2 x3 xn
ex = 1 + x + + + + + = .
2! 3! n! k!
k=0

The power series here can be used to define ex , but then obtaining the prop-
erties of ex depends on the technical fact that a power series can be differenti-
ated term by term in its open interval (or disk if we are working with complex
numbers) of convergence.
The power series in the previous display also allows a small illustration of
the utility of quantifiers. Since it is valid for every real number x, it is valid
with x2 in place of x,

X x2k
2 x4 x6 x2n
ex = 1 + x 2 + + + + + = for any x R.
2! 3! n! k!
k=0

2
There is no need here to introduce the function g(x) = ex , then work out its
Taylor polynomial and remainder, then analyze the remainder.
We end this chapter by sketching two cautionary examples. First, work
from earlier in the section shows that the Taylor series for the function ln(1+x)
at a = 0 is

X
x2 x3 xn xk
T (x) = x + + (1)n1 + = (1)k1 .
2 3 n k
k=1

The Ratio Test shows that this series converges absolutely when |x| < 1,
and the nth Term Test shows that the series diverges when x > 1. The se-
ries also converges at x = 1, as observed earlier. Thus, while the domain of
the function ln(1 + x) is (1, ), the Taylor series has no chance to match
the function outside of (1, 1]. As for whether the Taylor series matches the
function on (1, 1], recall the Lagrange form of the remainder,

(1)n xn+1
Rn (x) = for some c between 0 and x.
(1 + c)n+1 (n + 1)
Consequently, the absolute value of the Lagrange form of the remainder is
 n+1
1 |x|
|Rn (x)| = for some c between 0 and x.
n+1 1+c

From the previous display, noting that |x| is the distance from 0 to x while
1 + c is the distance from 1 to c, we see that
If 0 x 1 then |x| 1 1 + c, and so Rn (x) goes to 0 as n gets large.
 1.3 Taylors Theorem 17

If 1/2 x < 0 then |x| 1/2 1 + c, and so again Rn (x) goes to 0 as

n gets large.
But if 1 < x < 1/2 then possibly 1 + c < |x|, and so possibly Rn (x)
does not go to 0 as n gets large.
That is, we have shown that

ln(1 + x) = T (x) for x [1/2, 1],

but the Lagrange form does not readily show that the equality in the previous
display also holds for x (1, 1/2). Figure 1.1 suggests why: the Taylor
polynomials are converging more slowly to the original function the farther
left we go on the graph. However, a different form of the remainder, given in
exercise 1.3.6, proves that indeed the equality holds for all x (1, 1]. Also,
the geometric series relation
1
= 1 x + x2 x3 + , 1 < x < 1
1+x
gives the relation ln(1 + x) = T (x) for x (1, 1) upon integrating termwise
and then setting x = 0 to see that the resulting constant term is 0; but this
arguments invocation of the theorem that a power series can be integrated
termwise within its interval (or disk) of convergence is nontrivial.
For the last example, define f : R R by
( 2
e1/x if x 6= 0,
f (x) =
0 if x = 0.

It is possible to show that f is infinitely differentiable and that every derivative

of f at 0 is 0. That is, f (k) (0) = 0 for k = 0, 1, 2, . . . . Consequently, the Taylor
series for f at 0 is

T (x) = 0 + 0x + 0x2 + + 0xn + .

That is, the Taylor series is the zero function, which certainly converges for all
x R. But the only value of x for which it converges to the original function f
is x = 0. In other words, although this Taylor series converges everywhere,
it fails catastrophically to equal the function it is attempting to match. The
problem is that the function f decays exponentially, and since exponential
behavior dominates polynomial behavior, any attempt to discern f by using
polynomials will fail to see it. Figures 1.2 and 1.3 plot f to display its rapid
decay. The first plot is for x [25, 25] and the second is for x [1/2, 1/2].

Exercises

1.3.1. (a) Let n N. What is the (2n+1)st degree Taylor polynomial T2n+1 (x)
for the function f (x) = sin x at 0? (The reason for the strange indexing here
18 1 Results from One-Variable Calculus

20 10 10 20
Figure 1.2. Rapidly decaying function, wide view

0.4 0.2 0.2 0.4

Figure 1.3. Rapidly decaying function, zoom view

is that every second term of the Taylor polynomial is 0.) Prove that sin x is
equal to the limit of T2n+1 (x) as n , similarly to the argument in the text
for ex . Also find T2n (x) for f (x) = cos x at 0, and explain why the argument
for sin shows that cos x is the limit of its even-degree Taylor polynomials as
well.
(b) Many years ago, the authors high school physics textbook asserted,
bafflingly, that the approximation sin x x is good for x up to 8 . Decon-
struct.

1.3.2. What is the nth degree Taylor polynomial Tn (x) for the following func-
tions at 0?
(a) f (x) = arctan x. (This exercise is not just a matter of routine mechan-
ics. One way to proceed involves the geometric series, and another makes use
of the factorization 1 + x2 = (1 ix)(1 + ix).)
 1.3 Taylors Theorem 19

(b) f (x) = (1 + x) where R. (Although the answer can be written

in a uniform way for all , it behaves differently when N. Introduce the
generalized binomial coefficient symbol
 
( 1)( 2) ( k + 1)
= , kN
k k!

to help produce a tidy answer.)

1.3.3. (a) Further tighten the numerical estimate of ln(1.1) from the section
by reasoning as follows. As n increases, the Taylor polynomials Tn (0.1) add
terms of decreasing magnitude and alternating sign. Therefore T4 (0.1) un-
derestimates ln(1.1). Now that we know this, it is useful to find the smallest
possible value of the remainder (by setting c = 0.1 rather than c = 0 in the for-
mula). Then ln(1.1) lies between T4 (0.1) plus this smallest possible remainder
value and T4 (0.1) plus the largest possible remainder value, obtained in the
section. Supply the numbers, and verify by machine that the tighter estimate
of ln(1.1) is correct.
(b) In figure 1.1, identify the graphs of T1 through T5 and the graph of ln
near x = 0 and near x = 2.

1.3.4. Without a calculator, use the third degree Taylor polynomial for sin(x)
at 0 to approximate a decimal representation of sin(0.1). Also compute the
decimal representation of an upper bound for the error of the approximation.
Bound sin(0.1) between two decimal representations.

1.3.5. Use a second degree Taylor polynomial to approximate 4.2. Use Tay-
lors theorem to find a guaranteedaccuracy of the approximation and thus to
find upper and lower bounds for 4.2.

1.3.6. (a) Another proof of Taylors Theorem uses the Fundamental Theorem
of Integral Calculus once and then integrates by parts repeatedly. Begin with
the hypotheses of Theorem 1.3.3, and let x I. By the Fundamental Theorem,
Z x
f (x) = f (a) + f (t) dt.
a
Rx
Let u = f (t) and v = t x, so that the integral is a u dv, and integrating
by parts gives
Z x
f (x) = f (a) + f (a)(x a) f (t)(t x) dt.
a
Rx
Let u = f (t) and v = 12 (t x)2 , so that again the integral is a u dv, and
integrating by parts gives
Z x
(x a)2 (t x)2
f (x) = f (a) + f (a)(x a) + f (a) + f (t) dt.
2 a 2
20 1 Results from One-Variable Calculus

Show that after n steps the result is

Z x
n (t x)n
f (x) = Tn (x) + (1) f (n+1) (t) dt.
a n!

Whereas the expression for f (x) Tn (x) in Theorem 1.3.3 is called the La-
grange form of the remainder, this exercise has derived the integral form of
the remainder. Use the Extreme Value Theorem and the Intermediate Value
Theorem to derive the Lagrange form of the remainder from the integral form.
(b) Use the integral form of the remainder to show that

ln(1 + x) = T (x) for x (1, 1].

 Part I

Multivariable Differential Calculus

2
Euclidean Space

Euclidean space is a mathematical construct that encompasses the line, the

plane, and three-dimensional space as special cases. Its elements are called vec-
tors. Vectors can be understood in various ways: as arrows, as quantities with
magnitude and direction, as displacements, or as points. However, along with
a sense of what vectors are, we also need to emphasize how they interact. The
axioms in section 2.1 capture the idea that vectors can be added together and
can be multiplied by scalars, with both of these operations obeying familiar
laws of algebra. Section 2.2 expresses the geometric ideas of length and angle
in Euclidean space in terms of vector algebra. Section 2.3 discusses continu-
ity for functions (also called mappings) whose inputs and outputs are vectors
rather than scalars. Section 2.4 introduces a special class of sets in Euclidean
space, the compact sets, and shows that compact sets are preserved under
continuous mappings. Finally, section 2.5 reviews the one-variable derivative
in light of ideas from the two sections preceding it.

2.1 Algebra: Vectors

Let n be a positive integer. The set of all ordered n-tuples of real numbers,

Rn = {(x1 , , xn ) : xi R for i = 1, , n} ,

constitutes n-dimensional Euclidean space. When n = 1, the parentheses

and subscript in the notation (x1 ) are superfluous, so we simply view the
elements of R1 as real numbers x and write R for R1 . Elements of R2 and
of R3 are written (x, y) and (x, y, z) to avoid needless subscripts. These first
few Euclidean spaces, R, R2 and R3 , are conveniently visualized as the line,
the plane, and space itself. (See figure 2.1.)
Elements of R are called scalars, of Rn , vectors. The origin of Rn ,
denoted 0, is defined to be

0 = (0, , 0).
24 2 Euclidean Space

Figure 2.1. The first few Euclidean spaces

Sometimes the origin of Rn will be denoted 0n to distinguish it from other

origins that we will encounter later.
In the first few Euclidean spaces R, R2 , R3 , one can visualize a vector as
a point x or as an arrow. The arrow can have its tail at the origin and its
head at the point x, or its tail at any point p and its head correspondingly
translated to p + x. (See figure 2.2. Most illustrations will depict R or R2 .)

p+x
p
x x

Figure 2.2. Various ways to envision a vector

To a mathematician, the word space doesnt connote volume but instead

refers to a set endowed with some structure. Indeed, Euclidean space Rn comes
with two algebraic operations. The first is vector addition,

+ : Rn Rn Rn ,

defined by adding the scalars at each component of the vectors,

(x1 , , xn ) + (y1 , , yn ) = (x1 + y1 , , xn + yn ).

For example, (1, 2, 3) + (4, 5, 6) = (5, 7, 9). Note that the meaning of the
+ sign is now overloaded: on the left of the displayed equality, it denotes
the new operation of vector addition, whereas on the right side it denotes
the old addition of real numbers. The multiple meanings of the plus sign
shouldnt cause problems since which + is meant is clear from context, i.e.,
 2.1 Algebra: Vectors 25

the meaning of + is clear from whether it sits between vectors or scalars.

(An expression such as (1, 2, 3) + 4, with the plus sign between a vector and
a scalar, makes no sense according to our grammar.)
The interpretation of vectors as arrows gives a geometric description of
vector addition, at least in R2 . To add the vectors x and y, draw them as
arrows starting at 0 and then complete the parallelogram P that has x and y
as two of its sides. The diagonal of P starting at 0 is then the arrow depicting
the vector x + y. (See figure 2.3.) The proof of this is a small argument with
similar triangles, left to the reader as exercise 2.1.2.

x+y
y
P
x

Figure 2.3. The parallelogram law of vector addition

The second operation on Euclidean space is scalar multiplication,

: R Rn Rn ,

defined by
a (x1 , , xn ) = (ax1 , , axn ).
For example, 2(3, 4, 5) = (6, 8, 10). We will almost always omit the symbol
and write ax for a x. With this convention, juxtaposition is overloaded as
+ was overloaded above, but again this shouldnt cause problems.
Scalar multiplication of the vector x (viewed as an arrow) by a also has a
geometric interpretation: it simply stretches (i.e., scales) x by a factor of a.
When a is negative, ax turns x around and stretches it in the other direction
by |a|. (See figure 2.4.)

3x

x
2x

Figure 2.4. Scalar multiplication as stretching

26 2 Euclidean Space

With these two operations and distinguished element 0, Euclidean space

satisfies the following algebraic laws:

Theorem 2.1.1 (Vector Space Axioms).

(A1) Addition is associative: (x + y) + z = x + (y + z) for all x, y, z Rn .
(A2) 0 is an additive identity: 0 + x = x for all x Rn .
(A3) Existence of additive inverses: For each x Rn there exists y Rn such
that y + x = 0.
(A4) Addition is commutative: x + y = y + x for all x, y Rn .
(M1) Scalar multiplication is associative: a(bx) = (ab)x for all a, b R, x
Rn .
(M2) 1 is a multiplicative identity: 1x = x for all x Rn .
(D1) Scalar multiplication distributes over scalar addition: (a + b)x = ax + bx
for all a, b R, x Rn .
(D2) Scalar multiplication distributes over vector addition: a(x + y) = ax + ay
for all a R, x, y Rn .

All of these are consequences of how + and and 0 are defined for Rn
in conjunction with the fact that the real numbers, in turn endowed with +
and and containing 0 and 1, satisfy the field axioms (see section 1.1). For
example, to prove that Rn satisfies (M1), take any scalars a, b R and any
vector x = (x1 , , xn ) Rn . Then

a(bx) = a(b(x1 , , xn )) by definition of x

= a(bx1 , , bxn ) by definition of scalar multiplication
= (a(bx1 ), , a(bxn )) by definition of scalar multiplication
= ((ab)x1 , , (ab)xn ) by n applications of (m1) in R
= (ab)(x1 , , xn ) by definition of scalar multiplication
= (ab)x by definition of x.

The other vector space axioms for Rn can be shown similarly, by unwinding
vectors to their coordinates, quoting field axioms coordinatewise, and then
bundling the results back up into vectors (see exercise 2.1.3). Nonetheless, the
vector space axioms do not perfectly parallel the field axioms, and you are
encouraged to spend a little time comparing the two axiom sets to get a feel
for where they are similar and where they are different (see exercise 2.1.4).
Note in particular that
For n > 1, Rn is not endowed with vector-by-vector multiplication.
Although one can define vector multiplication on Rn componentwise, this mul-
tiplication does not combine with vector addition to satisfy the field axioms
except when n = 1. The multiplication of complex numbers makes R2 a field,
and in section 3.10 we will see an interesting noncommutative multiplication
of vectors for R3 , but these are special cases.
 2.1 Algebra: Vectors 27

One benefit of the vector space axioms for Rn is that they are phrased
intrinsically, meaning that they make no reference to the scalar coordinates
of the vectors involved. Thus, once you use coordinates to establish the vector
space axioms, your vector algebra can be intrinsic thereafter, making it lighter
and more conceptual. Also, in addition to being intrinsic, the vector space
axioms are general. While Rn is the prototypical set satisfying the vector space
axioms, it is by no means the only one. In coming sections we will encounter
other sets V (whose elements may be, for example, functions) endowed with
their own addition, multiplication by elements of a field F , and distinguished
element 0. If the vector space axioms are satisfied with V and F replacing Rn
and R then we say that V is a vector space over F .
The pedagogical point here is that although the similarity between vector
algebra and scalar algebra may initially make vector algebra seem uninspiring,
in fact the similarity is exciting. It makes mathematics easier because familiar
algebraic manipulations apply in a wide range of contexts. The same symbol-
patterns have more meaning. For example, we use intrinsic vector algebra to
show a result from Euclidean geometry, that the three medians of a triangle
intersect. (A median is a segment from a vertex to the midpoint of the opposite
edge.) Consider a triangle with vertices x, y, and z, and form the average of
the three vertices,
x+y+z
p= .
3
This algebraic average will be the geometric center of the triangle, where the
medians meet. (See figure 2.5.) Indeed, rewrite p as
 
2 y+z
p=x+ x .
3 2
The displayed expression for p shows that it is two thirds of the way from x
along the line segment from x to the average of y and z, i.e., that p lies on
the triangle median from vertex x to side yz. (Again see the figure. The idea
is that (y + z)/2 is being interpreted as the midpoint of y and z, each of these
viewed as a point, while on the other hand, the little mnemonic

head minus tail

helps us to remember quickly that (y + z)/2 x can be viewed as the arrow-

vector from x to (y + z)/2.) Since p is defined symmetrically in x, y, and z,
and it lies on one median, it therefore lies on the other two medians as well.
In fact, the vector algebra has shown that it lies two thirds of the way along
each median. (As for how a person might find this proof, it is a matter of
hoping that the center (x + y + z)/3 lies on the median by taking the form
x + c((y + z)/2 x) for some c and then seeing that indeed c = 2/3 works.)
The standard basis of Rn is the set of vectors

{e1 , e2 , , en }
28 2 Euclidean Space

y+z
2
p
x

Figure 2.5. Three medians of a triangle

where

e1 = (1, 0, , 0), e2 = (0, 1, , 0), , en = (0, 0, , 1).

(Thus each ei is itself a vector, not the ith scalar entry of a vector.) Any
vector x = (x1 , x2 , , xn ) (where the xi are scalar entries) decomposes as

x = (x1 , x2 , , xn )
= (x1 , 0, , 0) + (0, x2 , , 0) + + (0, 0, , xn )
= x1 (1, 0, , 0) + x2 (0, 1, , 0) + + xn (0, 0, , 1)
= x 1 e1 + x 2 e2 + + x n en ,

or, more succinctly,

n
X
x= x i ei . (2.1)
i=1

Note that in equation (2.1), x and the ei are vectors while the xi are scalars.
The equation shows that any x Rn is expressible as a linear combination
(sum of scalar multiples) Pn of the standard basis vectors. The expression is
unique, for if also x = i=1 xi ei for some scalars x1 , , xn then the equality
says that x = (x1 , x2 , , xn ), so that xi = xi for i = 1, , n.
(The reason that the geometric-sounding word linear is used here and
elsewhere in this chapter to describe properties having to do with the algebraic
operations of addition and scalar multiplication will be explained in chapter 3.)
The standard basis is handy in that it is a finite set of vectors from which
each of the infinitely many vectors of Rn can be obtained in exactly one way
as a linear combination. But it is not the only such set, nor is it always the
optimal one.

Definition 2.1.2 (Basis). A set of vectors {fi } is a basis of Rn if every

x Rn is uniquely expressible as a linear combination of the fi .
 2.1 Algebra: Vectors 29

For example, the set {f1 , f2 } = {(1, 1), (1, 1)} is a basis of R2 . To see
this, consider an arbitrary vector (x, y) R2 . This vector is expressible as a
linear combination of f1 and f2 if and only if there are scalars a and b such
that
(x, y) = af1 + bf2 .
Since f1 = (1, 1) and f2 = (1, 1), this vector equation is equivalent to a pair
of scalar equations,

x = a + b,
y = a b.

Add these equations and divide by 2 to get a = (x + y)/2, and similarly

b = (x y)/2. In other words, we have found that
x+y xy
(x, y) = (1, 1) + (1, 1),
2 2
and the coefficients a = (x + y)/2 and b = (x y)/2 on the right side of
the equation are the only possible coefficients a and b for the equation to
hold. That is, scalars a and b exist to express the vector (x, y) as a linear
combination of {f1 , f2 }, and the scalars are uniquely determined by the vector.
Thus {f1 , f2 } is a basis of R2 as claimed.
The set {g1 , g2 } = {(1, 3), (2, 6)} is not a basis of R2 , because any linear
combination ag1 + bg2 is (a + 2b, 3a + 6b), with the second entry equal to three
times the first. The vector (1, 0) is therefore not a linear combination of g1
and g2 .
Nor is the set {h1 , h2 , h3 } = {(1, 0), (1, 1), (1, 1)} a basis of R2 , because
h3 = 2h1 h2 , so that h3 is a nonunique linear combination of the hj .
See exercises 2.1.9 and 2.1.10 for practice with bases.

Exercises

2.1.1. Write down any three specific nonzero vectors u, v, w from R3 and any
two specific nonzero scalars a, b from R. Compute u+v, aw, b(v +w), (a+b)u,
u + v + w, abw, and the additive inverse to u.

2.1.2. Give a geometric proof that in R2 if we view the vectors x and y as

arrows from 0 and form the parallelogram P with these arrows as two of its
sides, then the diagonal z starting at 0 is the vector sum x + y viewed as an
arrow.

2.1.3. Verify that Rn satisfies vector space axioms (A2), (A3), (D1).

2.1.4. Are all the field axioms used in verifying that Euclidean space satisfies
the vector space axioms?
30 2 Euclidean Space

2.1.5. Show that 0 is the unique additive identity in Rn . Show that each vector
x Rn has a unique additive inverse, which can therefore be denoted x.
(And it follows that vector subtraction can now be defined,

: Rn Rn Rn , x y = x + (y) for all x, y Rn .)

Show that 0x = 0 for all x Rn .

2.1.6. Repeat the previous exercise, but with Rn replaced by an arbitrary

vector space V over a field F . (Work with the axioms.)

2.1.7. Show the uniqueness of additive identity and additive inverse using
only (A1), (A2), (A3). (This is tricky; the opening pages of some books on
group theory will help.)

2.1.8. Let x and y be non-collinear vectors in R3 . Give a geometric description

of the set of all linear combinations of x and y.

2.1.9. Which of the following sets are bases of R3 ?

S1 = {(1, 0, 0), (1, 1, 0), (1, 1, 1)},

S2 = {(1, 0, 0), (0, 1, 0), (0, 0, 1), (1, 1, 1)},
S3 = {(1, 1, 0), (0, 1, 1)},
S4 = {(1, 1, 0), (0, 1, 1), (1, 0, 1)}.

How many elements do you think a basis for Rn must have? Give (without
proof) geometric descriptions of all bases of R2 , of R3 .

2.1.10. Recall the field C of complex numbers. Define complex n-space Cn

analogously to Rn :

Cn = {(z1 , , zn ) : zi C for i = 1, , n} ,

and endow it with addition and scalar multiplication defined by the same
formulas as for Rn . You may take for granted that under these definitions, Cn
satisfies the vector space axioms with scalar multiplication by scalars from R,
and also Cn satisfies the vector space axioms with scalar multiplication by
scalars from C. That is, using language that was introduced briefly in the
section, Cn can be viewed as a vector space over R and also, separately, as a
vector space over C. Give a basis for each of these vector spaces.

Brief Pedagogical Interlude

Before continuing, a few comments about how to work with these notes may
be helpful.
 2.2 Geometry: Length and Angle 31

The material of chapters 2 through 5 is largely cumulative, with the main

theorem of chapter 5 being proved with main results of chapters 2, 3, and 4.
Each chapter is largely cumulative internally as well. To acquire detailed com-
mand of so much material and also a large-scale view of how it fits together,
the trick is to focus on each sections techniques while studying that section
and working its exercises, but thereafter to use the sections ideas freely by
reference. Specifically, after the scrutiny of vector algebra in the previous sec-
tion, ones vector manipulations should be fluent from now on, freeing one to
concentrate on vector geometry in the next section, after which the geometry
should also be light while concentrating on the analytical ideas of the following
section, and so forth.
Admittedly, the model that one has internalized all the prior material
before moving on is idealized. For that matter, so is the model that a body
of interplaying ideas is linearly cumulative. In practice, making ones work
focus entirely on the details of whichever topics are currently live while using
previous ideas by reference isnt always optimal. One might engage with the
details of previous ideas because one is coming to understand them better,
or because the current ideas showcase the older ones in a new way. Still,
the paradigm of technical emphasis on the current ideas and fluent use of
the earlier material does help a person who is navigating a large body of
mathematics to conserve energy and synthesize a larger picture.

2.2 Geometry: Length and Angle

The geometric notions of length and angle in Rn are readily described in terms
of the algebraic notion of inner product.

Definition 2.2.1 (Inner Product). The inner product is a function from

pairs of vectors to scalars,

h , i : Rn Rn R,

defined by the formula

n
X
h(x1 , , xn ), (y1 , , yn )i = x i yi .
i=1

For example,
n(n + 1)
h(1, 1, , 1), (1, 2, , n)i = ,
2
hx, ej i = xj where x = (x1 , , xn ) and j {1, , n},
hei , ej i = ij (this means 1 if i = j, 0 otherwise).

Proposition 2.2.2 (Inner Product Properties).

32 2 Euclidean Space

(IP1) The inner product is positive definite: hx, xi 0 for all x Rn , with
equality if and only if x = 0.
(IP2) The inner product is symmetric: hx, yi = hy, xi for all x, y Rn .
(IP3) The inner product is bilinear:

hx + x , yi = hx, yi + hx , yi, hax, yi = ahx, yi,

hx, y + y i = hx, yi + hx, y i, hx, byi = bhx, yi

for all a, b R, x, x , y, y Rn .

Proof. Exercise 2.2.4.

The reader should be aware that:

In general, hx + x , y + y i does not equal hx, yi + hx , y i.
Indeed, expanding hx + x , y + y i carefully with the inner product properties
shows that the cross-terms hx, y i and hx , yi are present in addition to hx, yi
and hx , y i.
Like the vector space axioms, the inner product properties are phrased
intrinsically, although they need to be proved using coordinates. As mentioned
in the previous section, intrinsic methods are neater and more conceptual than
using coordinates. More importantly:
The rest of the results of this section are proved by reference to the
inner product properties, with no further reference to the inner product
formula.
The notion of an inner product generalizes beyond Euclidean spacethis will
be demonstrated in exercise 2.3.4, for exampleand thanks to the displayed
sentence, once the properties (IP1) through (IP3) are established for any inner
product, all of the pending results in the section will follow automatically with
no further work. (But here a slight disclaimer is necessary. In the displayed
sentence, the word results does not refer to the pending graphic figures. The
fact that the length and angle to be defined in this section will agree with prior
notions of length and angle in the plane, or in three-dimensional space, does
depend on the specific inner product formula. In Euclidean space, the inner
product properties do not determine the inner product formula uniquely. This
point will be addressed in exercise 3.5.1.)

Definition 2.2.3 (Modulus). The modulus (or absolute value) of a vec-

tor x Rn is defined as p
|x| = hx, xi.

Thus the modulus is defined in terms of the inner product, rather than by
its own formula. The inner product formula shows that the modulus formula
is q
|(x1 , , xn )| = x21 + + x2n ,
 2.2 Geometry: Length and Angle 33

so that some particular examples are

r
n(n + 1)(2n + 1)
|(1, 2, , n)| = ,
6
|ei | = 1.

However, the definition of the modulus in terms of inner product combines

with the inner product properties to show, with no reference to the inner
product formula or the modulus formula, that the modulus satisfies (exer-
cise 2.2.5)

Proposition 2.2.4 (Modulus Properties).

(Mod1) The modulus is positive: |x| 0 for all x Rn , with equality if and
only if x = 0.
(Mod2) The modulus is absolute-homogeneous: |ax| = |a||x| for all a R and
x Rn .

Like other symbols, the absolute value signs are now overloaded, but their
meaning can be inferred from context, as in property (Mod2). When n is 1, 2,
or 3, the modulus |x| gives the distance from 0 to the point x, or the length
of x viewed as an arrow. (See figure 2.6.)

|x| |x|

x |x|
x

Figure 2.6. Modulus as length

The following relation between inner product and modulus will help to
show that distance in Rn behaves as it should, and that angle in Rn makes
sense. Since the relation is not obvious, its proof is a little subtle.

Theorem 2.2.5 (CauchySchwarz Inequality). For all x, y Rn ,

|hx, yi| |x| |y|,

with equality if and only if one of x, y is a scalar multiple of the other.

34 2 Euclidean Space

Note that the absolute value signs mean different things on each side of
the CauchySchwarz Inequality. On the left side, the quantities x and y are
vectors, their inner product hx, yi is a scalar, and |hx, yi| is its scalar absolute
value, while on the right side, |x| and |y| are the scalar absolute values of
vectors, and |x| |y| is their product. That is, the CauchySchwarz Inequality
says:
The size of the product is at most the product of the sizes.
The CauchySchwarz Inequality can be written out in coordinates, tem-
porarily abandoning the principle that we should avoid reference to formulas,

(x1 y1 + + xn yn )2 (x21 + + x2n )(y12 + + yn2 ).

And this inequality can be proved unconceptually as follows (the reader is

encouraged only to skim the following computation). The desired inequality
rewrites as X 2 X X
x i yi x2i yj2 ,
i i j

where the indices of summation run from 1 to n. Expand the square to get
X X X
x2i yi2 + x i yi x j yj x2i yj2 ,
i i,j i,j
i6=j

and canceling the terms common to both sides reduces it to

X X
x i yi x j yj x2i yj2 ,
i6=j i6=j

or X
(x2i yj2 xi yi xj yj ) 0.
i6=j

Rather than sum over all pairs (i, j) with i 6= j, sum over the pairs with
i < j, collecting the (i, j)-term and the (j, i)-term for each such pair, and the
previous inequality becomes
X
(x2i yj2 + x2j yi2 2xi yj xj yi ) 0.
i<j

Thus the desired inequality has reduced to a true inequality,

X
(xi yj xj yi )2 0.
i<j

So the main proof is done, although there is still the question of when equality
holds.
But surely the previous paragraph is not the graceful way to argue. The
computation draws on the minutiae of the formulas for the inner product and
 2.2 Geometry: Length and Angle 35

the modulus, rather than using their properties. It is uninformative, making

the CauchySchwarz Inequality look like a low-level accident. It suggests that
larger-scale mathematics is just a matter of bigger and bigger formulas. To
prove the inequality in a way that is enlightening and general, we should
work intrinsically, keeping the scalars hx, yi and |x| and |y| notated in their
concise forms, and we should use properties, not formulas. The idea is that the
calculation in coordinates reduces to the fact that squares are nonnegative.
That is, the CauchySchwarz Inequality is somehow quadratically hard, and its
verification amounted to completing many squares. The argument to be given
here is guided by this insight to prove the inequality by citing facts about
quadratic polynomials, facts established by completing one square back in
high school algebra at the moment that doing so was called for. Thus we
eliminate redundancy and clutter. So the argument to follow will involve an
auxiliary object, a judiciously-chosen quadratic polynomial, but in return it
will become coherent.

Proof. The result is clear when x = 0, so assume x 6= 0. For any a R,

0 hax y, ax yi by positive definiteness

= ahx, ax yi hy, ax yi by linearity in the first variable
2
= a hx, xi ahx, yi ahy, xi + hy, yi by linearity in the second variable
2 2 2
= a |x| 2ahx, yi + |y| by symmetry, definition of modulus.

View the right side as a quadratic polynomial in the scalar variable a, where
the scalar coefficients of the polynomial depend on the generic but fixed vec-
tors x and y,
f (a) = |x|2 a2 2hx, yia + |y|2 .
We have shown that f (a) is always nonnegative, so f has at most one root.
Thus by the quadratic formula its discriminant is nonpositive,

4hx, yi2 4|x|2 |y|2 0,

and the CauchySchwarz Inequality |hx, yi| |x| |y| follows. Equality holds
exactly when the quadratic polynomial f (a) = |ax y|2 has a root a, i.e.,
exactly when y = ax for some a R.

Geometrically, the condition for equality in CauchySchwarz is that the

vectors x and y, viewed as arrows at the origin, are parallel, though perhaps
pointing in opposite directions. A geometrically conceived proof of Cauchy
Schwarz is given in exercise 2.2.15 to complement the algebraic argument that
has been given here.
The CauchySchwarz Inequality shows that the modulus function satisfies

Theorem 2.2.6 (Triangle Inequality). For all x, y Rn ,

|x + y| |x| + |y|,
36 2 Euclidean Space

with equality if and only if one of x, y is a nonnegative scalar multiple of the

other.

Proof. To show this, compute

|x + y|2 = hx + y, x + yi
= |x|2 + 2hx, yi + |y|2 by bilinearity
2 2
|x| + 2|x||y| + |y| by CauchySchwarz
2
= (|x| + |y|) ,

proving the inequality. Equality holds exactly when hx, yi = |x||y|, or equiva-
lently when |hx, yi| = |x||y| and hx, yi 0. These hold when one of x, y is a
scalar multiple of the other and the scalar is nonnegative.

While the CauchySchwarz Inequality says that the size of the product is
at most the product of the sizes, the Triangle Inequality says:

The size of the sum is at most the sum of the sizes.

The Triangle Inequalitys name is explained by its geometric interpretation

in R2 . View x as an arrow at the origin, y as an arrow with tail at the head
of x, and x + y as an arrow at the origin. These three arrows form a triangle,
and the assertion is that the lengths of two sides sum to at least the length of
the third. (See figure 2.7.)

x+y
y

x
Figure 2.7. Sides of a triangle

The full Triangle Inequality says that for all x, y Rn ,

| |x| |y| | |x y| |x| + |y|.

The proof is exercise 2.2.7.

A small argument, which can be formalized as induction if one is painstak-
ing, shows that the basic Triangle Inequality extends from two vectors to any
finite number of vectors. For example,

|x + y + z| |x + y| + |z| |x| + |y| + |z|.

 2.2 Geometry: Length and Angle 37

The only obstacle to generalizing the basic Triangle Inequality in this fashion
is notation. The argument cant use the symbol n to denote the number of
vectors since n already denotes the dimension of the Euclidean space where
we are working; and furthermore, the vectors cant be denoted with subscripts
since a subscript denotes a component of an individual vector. Thus, for now
we are stuck writing something like

|x(1) + + x(k) | |x(1) | + + |x(k) | for all x(1) , , x(k) Rn ,

or  
Xk  X k
 
 x(i)  |x(i) |, x(1) , , x(k) Rn .
 
i=1 i=1

As our work with vectors becomes more intrinsic, vector entries will demand
less of our attention, and we will be able to denote vectors by subscripts. The
notation-change will be implemented in the next section.
For any vector x = (x1 , , xn ) Rn , useful bounds on the modulus |x|
in terms of the scalar absolute values |xi | are

Proposition 2.2.7 (Size Bounds). For any j {1, , n},

n
X
|xj | |x| |xi |.
i=1

The proof (by quick applications of the CauchySchwarz Inequality and

the Triangle Inequality) is exercise 2.2.8.
The modulus gives rise to a distance function on Rn that behaves as dis-
tance should. Define
d : Rn Rn R
by
d(x, y) = |y x|.

For example, d(ei , ej ) = 2(1 ij ).

Theorem 2.2.8 (Distance Properties).

(D1) Distance is positive: d(x, y) 0 for all x, y Rn , and d(x, y) = 0 if and
only if x = y.
(D2) Distance is symmetric: d(x, y) = d(y, x) for all x, y Rn .
(D3) Triangle Inequality: d(x, z) d(x, y) + d(y, z) for all x, y, z Rn .

(D1) and (D2) are clearly desirable as properties of a distance function.

Property (D3) says that you cant shorten your trip from x to z by making a
stop at y.

Proof. Exercise 2.2.9.

38 2 Euclidean Space

The CauchySchwarz Inequality also lets us define the angle between two
nonzero vectors in terms of the inner product. If x and y are nonzero vectors
in Rn , define their angle x,y by the condition

hx, yi
cos x,y = , 0 x,y . (2.2)
|x||y|
hx,yi
The condition is sensible because 1 1 by the CauchySchwarz
|x||y|

Inequality. For example, cos (1,0),(1,1) = 1/ 2, and so (1,0),(1,1) = /4. In
particular, two nonzero vectors x, y are orthogonal when hx, yi = 0. Natu-
rally, we would like x,y to correspond to the usual notion of angle, at least
in R2 , and indeed it doessee exercise 2.2.10. For convenience, define any two
vectors x and y to be orthogonal if hx, yi = 0, thus making 0 orthogonal to
all vectors.
Rephrasing geometry in terms of intrinsic vector algebra not only extends
the geometric notions of length and angle uniformly to any dimension, it also
makes some low-dimensional geometry easier. For example, vectors show in a
natural way that the three altitudes of any triangle must meet. Let x and y
denote two sides of the triangle, making the third side xy by the head minus
tail mnemonic. Let q be the point where the altitudes to x and y meet. (See
figure 2.8, which also shows the third altitude.) Thus

q y x and q x y.

We want to show that also q lies on the third altitude, i.e., that

q x y.

To rephrase matters in terms of inner products, we want to show that

( )
hq y, xi = 0
= hq, x yi = 0.
hq x, yi = 0

Since the inner product is linear in each of its arguments, a further rephrase
is that we want to show that
( )
hq, xi = hy, xi
= hq, xi = hq, yi.
hq, yi = hx, yi

And this is immediate since the inner product is symmetric: hq, xi and hq, yi
both equal hx, yi, and so they equal each other as desired. The point q where
the three altitudes meet is called the orthocenter of the triangle. In general
the orthocenter of a triangle is not the center, cf. the previous section.
 2.2 Geometry: Length and Angle 39

xy
q

x
Figure 2.8. Three altitudes of a triangle

Exercises

2.2.1. Let x = ( 23 , 12 , 0), y = ( 12 , 23 , 1), z = (1, 1, 1). Compute hx, xi,
hx, yi, hy, zi, |x|, |y|, |z|, x,y , y,e1 , z,e2 .

2.2.2. Show that the points x = (2, 1, 3, 1), y = (4, 2, 1, 4), z = (1, 3, 6, 1)
form the vertices of a triangle in R4 with two equal angles.
Pn
2.2.3. Explain why for all x Rn , x = j=1 hx, ej iej .

2.2.4. Prove the Inner Product Properties.

2.2.5. Use the Inner Product Properties and the definition of the modulus in
terms of the inner product to prove the Modulus Properties.

2.2.6. In the text, the modulus is defined in terms of the inner product. Prove
that this can be turned around by showing that for every x, y Rn ,

|x + y|2 |x y|2
hx, yi = .
4
2.2.7. Prove the full Triangle Inequality: for any x, y Rn ,

| |x| |y| | |x y| |x| + |y|.

Do not do this by writing three more variants of the proof of the Triangle
Inequality, but by substituting suitably into the basic Triangle Inequality,
which is already proved.

2.2.8. Let x = (x1 , , xn ) Rn . Prove the Size Bounds: For any j

{1, , n},
n
X
|xj | |x| |xi |.
i=1

(One approach is to start by noting that xj = hx, ej i and recalling equa-

tion (2.1).) When can each be an =?
40 2 Euclidean Space

2.2.9. Prove the Distance Properties.

2.2.10. In R2 , depict the nonzero vectors x and y as arrows from the origin
and depict x y as an arrow from the endpoint of y to the endpoint of x. Let
denote the angle (in the usual geometric sense) between x and y. Use the
Law of Cosines to show that
hx, yi
cos = ,
|x||y|

so that our notion of angle agrees with the geometric one, at least in R2 .
Pn
2.2.11. Prove that for any nonzero x Rn , i=1 cos2 x,ei = 1.

2.2.12. Prove that two nonzero vectors x, y are orthogonal if and only if
|x + y|2 = |x|2 + |y|2 .

2.2.13. Use vectors in R2 to show that the diagonals of a parallelogram are

perpendicular if and only if the parallelogram is a rhombus.

2.2.14. Use vectors to show that every angle inscribed in a semicircle is right.

2.2.15. Let x and y be vectors, with x nonzero. Define the parallel component
of y along x and the normal component of y to x to be

hx, yi
y(kx) = x and y(x) = y y(kx) .
|x|2

(a) Show that y = y(kx) + y(x) ; show that y(kx) is a scalar multiple of x; show
that y(x) is orthogonal to x. Show that the decomposition of y as a sum of
vectors parallel and perpendicular to x is unique. Draw an illustration.
(b) Show that
|y|2 = |y(kx) |2 + |y(x) |2 .
What theorem from classical geometry does this encompass?
(c) Explain why it follows from (b) that

|y(kx) | |y|,

with equality if and only y is a scalar multiple of x. Use this inequality to give
another proof of the CauchySchwarz Inequality. This argument gives the
geometric content of CauchySchwarz: The parallel component of one vector
along another is at most as long as the original vector.
(d) The proof of the CauchySchwarz Inequality in part (c) refers to parts
(a) and (b), part (a) refers to orthogonality, orthogonality refers to an angle,
and as explained in the text, the fact that angles make sense depends on the
CauchySchwarz Inequality. And so the proof in part (c) apparently relies on
circular logic. Explain why the logic is in fact not circular.
 2.3 Analysis: Continuous Mappings 41

2.2.16. Given nonzero vectors x1 , x2 , , xn in Rn , the GramSchmidt

process is to set

x1 = x1
x2 = x2 (x2 )(kx1 )
x3 = x3 (x3 )(kx2 ) (x3 )(kx1 )
..
.
xn = xn (xn )(kxn1 ) (xn )(kx1 ) .

(a) What is the result of applying the GramSchmidt process to the vectors
x1 = (1, 0, 0), x2 = (1, 1, 0), and x3 = (1, 1, 1)?
(b) Returning to the general case, show that x1 , , xn are pairwise or-
thogonal and that each xj has the form

xj = aj1 x1 + aj2 x2 + + aj,j1 xj1 + xj .

Thus any linear combination of the new {xj } is also a linear combination of the
original {xj }. The converse is also true and will be shown in exercise 3.3.13.

2.3 Analysis: Continuous Mappings

A mapping from Rn to Rm is some rule that assigns to each point x in Rn
a point in Rm . Generally mappings will be denoted by letters such as f , g, h.
When m = 1 we say function instead of mapping.
For example, the mapping

f : R2 R2

defined by
f (x, y) = (x2 y 2 , 2xy)
takes the real and imaginary parts of a complex number z = x+iy and returns
the real and imaginary parts of z 2 . By the nature of multiplication of complex
numbers, this means that each output point has modulus equal to the square
of the modulus of the input point and has angle equal to twice the angle of
the input point. Make sure that you see how this is shown in figure 2.9.
Mappings expressed by formulas may be undefined at certain points (e.g.,
f (x) = 1/|x| is undefined at 0), so we need to restrict their domains. For
a given dimension n, a given set A Rn , and a second dimension m,
let M(A, Rm ) denote the set of all mappings f : A Rm . This set forms a
vector space over R (whose points are functions) under the operations

+ : M(A, Rm ) M(A, Rm ) M(A, Rm ),

defined by
42 2 Euclidean Space

2
1

1 1 1

Figure 2.9. The complex square as a mapping from R2 to R2

(f + g)(x) = f (x) + g(x) for all x A,

and
: R M(A, Rm ) M(A, Rm ),
defined by
(a f )(x) = a f (x) for all x A.
As usual, + and are overloaded: on the left they denote operations
on M(A, Rm ), while on the right they denote the operations on Rm defined in
section 2.1. Also as usual, the is generally omitted. The origin in M(A, Rm )
is the zero mapping, 0 : A Rm , defined by
0(x) = 0m for all x A.
m
For example, to verify that M(A, R ) satisfies (A1), consider any mappings
f, g, h M(A, Rm ). For any x A,
((f + g) + h)(x) = (f + g)(x) + h(x) by definition of + in M(A, Rm )
= (f (x) + g(x)) + h(x) by definition of + in M(A, Rm )
= f (x) + (g(x) + h(x)) by associativity of + in Rm
= f (x) + (g + h)(x) by definition of + in M(A, Rm )
= (f + (g + h))(x) by definition of + in M(A, Rm ).
Since x is arbitrary, (f + g) + h = f + (g + h).
Let A be a subset of Rn . A sequence in A is an infinite list of vectors
{x1 , x2 , x3 , } in A, often written {x }. (The symbol n is already in use,
so its Greek counterpart pronounced nuis used as the index-counter.)
Since a vector has n entries, each vector x in the sequence takes the form
(x1, , , xn, ).
Definition 2.3.1 (Null Sequence). The sequence {x } in Rn is null if for
every > 0 there exists some 0 such that
if > 0 then |x | < .
That is, a sequence is null if for every > 0, all but finitely many terms of
the sequence lie within distance of 0n .
 2.3 Analysis: Continuous Mappings 43

Quickly from the definition, if {x } is a null sequence in Rn and {y } is a

sequence in Rn such that |y | |x | for all then also {y } is null.
Let {x } and {y } be null sequences in Rn , and let c be a scalar. Then the
sequence {x + y } is null because |x + y | |x | + |y | for each , and the
sequence {cx } is null because |cx | = |c||x | for each . These two results
show that the set of null sequences in Rn forms a vector space.
Since for any vector x Rn the absolute value |x| is a nonnegative scalar,
taking the scalar absolute value in turn has no further effect,

| |x| | = |x|, x Rn ,

and so a vector sequence {x } is null if and only if the scalar sequence {|x |}
is null.
Lemma 2.3.2 (Componentwise Nature of Nullness). The vector se-
quence {(x1, , , xn, )} is null if and only if each of its component scalar
sequences {xj, } (j {1, , n}) is null.
Proof. By the observation just before the lemma, it suffices to show that
{|(x1, , , xn, )|} is null if and only if each {|xj, |} is null. The Size Bounds
give for any j {1, , n} and any ,
n
X
|xj, | |(x1, , , xn, )| |xi, |.
i=1

If {|(x1, , , xn, )|} is null, then by the first inequality

Pnso is each {|xj, |}.
On the other hand, if each {|xj, |} is null then so is { i=1 |xi, |}, and thus
by the second inequality {|(x1, , , xn, )|} is null as well.

We define the convergence of vector sequences in terms of null sequences.
Definition 2.3.3 (Sequence Convergence, Sequence Limit). Let A be
a subset of Rn . Consider a sequence {x } in A and a point p Rn . The
sequence {x } converges to p (or has limit p), written {x } p, if the
sequence {x p} is null. When the limit p is a point of A, the sequence {x }
converges in A.
If a sequence {x } converges to p and also converges to p then the constant
sequence {p p} is the difference of the null sequences {x p} and {x p },
hence null, forcing p = p. Thus a sequence cannot converge to two distinct
values.
Many texts define convergence directly rather than by reference to nullness,
the key part of the definition being

if > 0 then |x p| < .

In particular, a null sequence is a sequence that converges to 0n . However, in

contrast to the situation for null sequences, for p 6= 0n it is emphatically false
44 2 Euclidean Space

that if {|x |} converges to |p| then necessarily {x } converges to p or even

converges at all. Also, for any nonzero p, the sequences that converge to p do
not form a vector space.
Vector versions of the Sum Rule and the Constant Multiple Rule for con-
vergent sequences follow immediately from the vector space properties of null
sequences:
Proposition 2.3.4 (Linearity of Convergence). Let {x } be a sequence
in Rn converging to p, let {y } be a sequence in Rn converging to q, and let c
be a scalar. Then the sequence {x + y } converges to p + q, and the sequence
{cx } converges to cp.
Similarly, since a sequence {x } converges to p if and only if {x p} is
null, we have in consequence of the componentwise nature of nullness (exer-
cise 2.3.5)
Proposition 2.3.5 (Componentwise Nature of Convergence). The vec-
tor sequence {(x1, , , xn, )} converges to the vector (p1 , , pn ) if and
only if each component scalar sequence {xj, } (j = 1, , n) converges to
the scalar pj .
Continuity, like convergence, is typographically indistinguishable in R
and Rn .
Definition 2.3.6 (Continuity). Let A be a subset of Rn , let f : A Rm
be a mapping, and let p be a point of A. Then f is continuous at p if for
every sequence {x } in A converging to p, the sequence {f (x )} converges
to f (p). The mapping f is continuous on A (or just continuous when A is
clearly established) if it is continuous at each point p A.
For example, the modulus function

| | : Rn R

is continuous on Rn . To see this, consider any point p Rn and consider any

sequence {x } in Rn that converges to p. We need to show that the sequence
{|x |} in R converges to |p|. But by the full Triangle Inequality,

| |x | |p| | |x p|.

Since the right side is the th term of a null sequence, so is the left, giving
the result.
For another example, let a Rn be any fixed vector and consider the
function defined by taking the inner product of this vector with other vectors,

T : Rn R, T (x) = ha, xi.

This function is also continuous on Rn . To see so, again consider any p Rn

and any sequence {x } in Rn converging to p. Then the definition of T , the
 2.3 Analysis: Continuous Mappings 45

bilinearity of the inner product, and the CauchySchwarz Inequality combine

to show that

|T (x ) T (p)| = |ha, x i ha, pi| = |ha, x pi| |a| |x p|.

Since |a| is a constant, the right side is the th term of a null sequence, hence
so is the left, and the proof is complete. We will refer to this example in
section 3.1. Also, note that as a special case of this example we may take
any j {1, , n} and set the fixed vector a to ej , showing that the jth
coordinate function map,

j : Rn R, j (x1 , , xn ) = xj ,

is continuous.

Proposition 2.3.7 (Vector Space Properties of Continuity). Let A be

a subset of Rn , let f, g : A Rm be continuous mappings, and let c R.
Then the sum and the scalar multiple mappings

f + g, cf : A Rm

are continuous. Thus the set of continuous mappings from A to Rm forms a

vector subspace of M(A, Rm ).

The vector space properties of continuity follows immediately from the

linearity of convergence and from the definition of continuity. Another conse-
quence of the definition of continuity is

Proposition 2.3.8 (Persistence of Continuity Under Composition).

Let A be a subset of Rn , and let f : A Rm be a continuous mapping. Let
B be a superset of f (A) in Rm , and let g : B R be a continuous mapping.
Then the composition mapping

g f : A R

is continuous.

The proof is exercise 2.3.7.

Let A be a subset of Rn . Any mapping f : A Rm decomposes as m
functions f1 , , fm with each fi : A R, by the formula

f (x) = (f1 (x), , fm (x)).

For example, if f (x, y) = (x2 y 2 , 2xy) then f1 (x, y) = x2 y 2 and f2 (x, y) =

2xy. The decomposition of f can also be written
m
X
f (x) = fi (x)ei ,
i=1
46 2 Euclidean Space

or equivalently, the functions fi are defined by the condition

fi (x) = f (x)i for i = 1, , m.

Conversely, given m functions f1 , , fm from A to R, any of the preced-

ing three displayed formulas assembles a mapping f : A Rm . Thus,
each mapping f determines and is determined by its component functions
f1 , , fm . Conveniently, to check continuity of the vector-valued mapping f
we only need to check its scalar-valued component functions.

Theorem 2.3.9 (Componentwise Nature of Continuity). Let A Rn ,

let f : A Rm have component functions f1 , , fm , and let p be a point
in A. Then

f is continuous at p each fi is continuous at p.

The componentwise nature of continuity follows from the componentwise

nature of convergence and is left as exercise 2.3.6.
Let A be a subset of Rn , let f and g be continuous functions from A
to R, and let c R. Then the familiar Sum Rule, Constant Multiple Rule,
Product Rule, and Quotient Rule for continuous functions hold. That is, the
sum f + g, the constant multiple cf , the product f g, and the quotient f /g
(at points p A such that g(p) 6= 0) are again continuous. The first two
of these facts are special cases of the vector space properties of continuity.
The proofs of the other two are typographically identical to their one-variable
counterparts. With the various continuity results so far in hand, it is clear
that a function such as
p
3 sin( x2 + y 2 + z 2 )
f : R R, f (x, y, z) =
exy+z
is continuous. The continuity of such functions, and of mappings with such
functions as their components, will go without comment from now on.
However, the continuity of functions of n variables also has new, subtle
features when n > 1. In R, a sequence {x } can approach the point p in only
two essential ways: from the left and from the right. But in Rn where n 2,
{x } can approach p from infinitely many directions, or not approach along a
line at all, so the convergence of {f (x )} can be trickier. For example, consider
the function f : R2 R defined by

2xy if (x, y) 6= 0,
f (x, y) = x2 + y 2
b if (x, y) = 0.

Can the constant b be specified to make f continuous at 0?

It cant. Take a sequence {(x , y )} approaching 0 along the line y = mx
of slope m. For any point (x , y ) of this sequence,
 2.3 Analysis: Continuous Mappings 47

2x mx 2mx2 2m
f (x , y ) = f (x , mx ) = = = .
x2 2
+ m x 2 (1 + m2 )x2 1 + m2

Thus, as the sequence of inputs {(x , y )} approaches 0 along the line of

slope m, the corresponding sequence of outputs {f (x , y )} holds steady
at 2m/(1 + m2 ), and so f (0) needs to take this value for continuity. Taking
input sequences {(x , y )} that approach 0 along lines of different slope shows
that f (0) needs to take different values for continuity, and hence f can not be
made continuous at 0. The graph of f away from 0 is a sort of spiral staircase,
and no height over 0 is compatible with all the stairs. (See figure 2.10. The
figure displays only the portion of the graph for slopes between 0 and 1 in the
input plane.) The reader who wants to work a virtually identical example can
replace the formula 2xy/(x2 + y 2 ) in f by (x2 y 2 )/(x2 + y 2 ) and run the
same procedure as in this paragraph.

Figure 2.10. A spiral staircase

The previous example was actually fairly simple in that we only needed to
study f (x, y) as (x, y) approached 0 along straight lines. Consider the function
g : R2 R defined by
2
x y if (x, y) 6= 0,
g(x, y) = x + y 2
4

b if (x, y) = 0.

For any nonzero slope m, take a sequence {(x , y )} approaching 0 along the
line y = mx. Compute that for each point of this sequence,

mx3 mx
g(x , y ) = g(x , mx ) = = 2 .
x4 + m2 x2 x + m2

This quantity tends to 0 as x goes to 0. That is, as the sequence of inputs

{(x , y )} approaches 0 along the line of slope m, the corresponding sequence
of outputs {g(x , y )} approaches 0, and so g(0) needs to take the value 0
48 2 Euclidean Space

for continuity. Since g is 0 at the nonzero points of either axis in the (x, y)-
plane, this requirement extends to the cases that {(x , y )} approaches 0 along
a horizontal or vertical line. However, next consider a sequence {(x , y )}
approaching 0 along the parabola y = x2 . For each point of this sequence,
x4 1
g(x , y ) = g(x , x2 ) = = .
x4 + x4 2
Thus, as the sequence of inputs {(x , y )} approaches 0 along the parabola,
the corresponding sequence of outputs {g(x , y )} holds steady at 1/2, and so
g(0) needs to be 1/2 for continuity as well. Thus g can not be made continuous
at 0, even though approaching 0 only along lines suggests that it can. The
reader who wants to work a virtually identical example can replace the formula
x2 y/(x4 + y 2 ) in g by x3 y/(x6 + y 2 ) and run the same procedure as in this
paragraph but using the curve y = x3 .
Thus, given a function f : R2 R, letting {(x , y )} approach 0 along
lines can disprove continuity at 0, but it can only suggest continuity at 0, not
prove it. To prove continuity, the Size Bounds may be helpful. For example,
let 3
x if (x, y) 6= 0,
h(x, y) = x2 + y 2

b if (x, y) = 0.
Can b be specified to make h continuous at 0? The estimate |x| |(x, y)| gives
for any (x, y) 6= 0,

|x3 | |x|3 |(x, y)|3

0 |h(x, y)| = = = |(x, y)|,
x2 + y 2 |(x, y)|2 |(x, y)|2
so as any sequence {(x , y )} of nonzero input vectors converges to 0, the
corresponding sequence of outputs {h(x , y )} is squeezed to 0 in absolute
value and hence converges to 0. Setting b = 0 makes h continuous at 0. The
reader who wants to work a virtually identical example can replace the formula
x3 /(x2 + y 2 ) in h by x2 y 2 /(x4 + y 2 ) and run the same procedure as in this
paragraph but applying the Size Bounds to vectors (x2 , y ).
Returning to the spiral staircase example,

2xy if (x, y) 6= 0,
f (x, y) = x2 + y 2
b if (x, y) = 0,

the Size Bounds show that that for any (x, y) 6= 0,

2|x| |y| 2|(x, y)|2

0 |f (x, y)| = 2
= 2.
|(x, y)| |(x, y)|2
The display tells us only that as a sequence of inputs {(x , y )} approaches 0,
the sequence of outputs {f (x , y )} might converge to some limit between 2
 2.3 Analysis: Continuous Mappings 49

and 2. The outputs neednt converge to 0 (or converge at all), but according
to this diagnostic they possibly could. Thus the Size Bounds tell us only that
f could be discontinuous at (0, 0), but they give no conclusive information.
In sum, these examples show that
The straight line test can prove that a limit does not exist, or it can
determine the only candidate for the value of the limit, but it can not
prove that the candidate value is the limit.
When the straight line test determines a candidate value of the limit,
approaching along a curve can further support the candidate, or it can
prove that the limit does not exist by determining a different candidate as
well.
The Size Bounds can prove that a limit does exist, but they can only
suggest that a limit does not exist.

The next proposition is a handy encoding of an intuitively plausible prop-

erty of continuous mappings. The result is so natural that it often is tacitly
taken for granted, but it is worth stating and proving carefully.

Proposition 2.3.10 (Persistence of Inequality). Let A be a subset of Rn

and let f : A Rm be a continuous mapping. Let p be a point of A, let b be
a point of Rm , and suppose that f (p) 6= b. Then there exists some > 0 such
that
for all x A such that |x p| < , f (x) 6= b.

Proof. Assume that the displayed statement in the proposition fails for ev-
ery > 0. Then in particular it fails for = 1/ for = 1, 2, 3, . So there
is a sequence {x } in A such that

|x p| < 1/ and f (x ) = b, = 1, 2, 3, .

Since f is continuous at p, this condition shows that f (p) = b. But in fact

f (p) 6= b, and so our assumption that the displayed statement in the propo-
sition fails for every > 0 leads to a contradiction. Therefore the statement
holds for some > 0, as desired.

Exercises

2.3.1. For A Rn , partially verify that M(A, Rm ) is a vector space over R

by showing that it satisfies vector space axioms (A4) and (D1).

2.3.2. Define multiplication : M(A, R)M(A, R) M(A, R). Is M(A, R)

a field with + from the section and this multiplication? Does it have a sub-
space that is a field?
50 2 Euclidean Space

2.3.3. For A Rn and m Z+ define a subspace of the space of mappings

from A to Rm ,

C(A, Rm ) = {f M(A, Rm ) : f is continuous on A}.

Briefly explain how the section has shown that C(A, Rm ) is a vector space.

2.3.4. Define an inner product and a modulus on C([0, 1], R) by

Z 1 p
hf, gi = f (t)g(t) dt, |f | = hf, f i.
0

Do the inner product properties (IP1), (IP2), and (IP3) (see Proposition 2.2.2)
hold for this inner product on C([0, 1], R)? How much of the material from
section 2.2 on the inner product and modulus in Rn carries over to hold for
C([0, 1], R)? Express the CauchySchwarz Inequality as a relation between
integrals.

2.3.5. Use the definition of convergence and the componentwise nature of

nullness to prove the componentwise nature of convergence. (The argument is
short.)

2.3.6. Use the definition of continuity and the componentwise nature of con-
vergence to prove the componentwise nature of continuity.

2.3.7. Prove the persistence of continuity under composition.

2.3.8. Define f : Q R by the rule

(
1 if x2 < 2,
f (x) =
0 if x2 > 2.

Is f continuous?

2.3.9. Which of the following functions on R2 can be defined continuously

at 0?
4 4 2 3
x y if (x, y) 6
= 0, x y if (x, y) 6= 0,
f (x, y) = (x2 + y 2 )2 g(x, y) = x2 + y 2

b if (x, y) = 0, b if (x, y) = 0,
3 3 2
x y if (x, y) 6= 0, xy if (x, y) 6= 0,
2
h(x, y) = x + y 2 k(x, y) = x + y 6
2

b if (x, y) = 0, b if (x, y) = 0.

2.3.10. Let f (x, y) = g(xy) where g : R R is continuous. Is f continuous?

2.3.11. Let f, g M(Rn , R) be such that f + g and f g are continuous. Are

f and g necessarily continuous?
 2.4 Topology: Compact Sets and Continuity 51

2.4 Topology: Compact Sets and Continuity

The Extreme Value Theorem from one-variable calculus states:
Let I be a nonempty closed and bounded interval in R, and let f :
I R be a continuous function. Then f takes a minimum value
and a maximum value on I.
This section generalizes the theorem from scalars to vectors. That is, we want
a result that if A is a set in Rn with certain properties, and if f : A Rm
is a continuous mapping, then the output set f (A) will also have certain
properties. The questions are, for what sorts of properties do such statements
hold, and when they hold, how do we prove them?
The one-variable theorem hypothesizes two data, the nonempty closed and
bounded interval I and the continuous function f . Each of these is described
in its own termsI takes the readily-recognizable but static form [a, b] where
a b, while the continuity of f is a dynamic assertion about convergence of
sequences. Because the two data have differently-phrased descriptions, a proof
of the Extreme Value Theorem doesnt suggest itself immediately: no ideas
at hand bear obviously on all the given information. Thus the work of this
section is not only to define the sets to appear in the pending theorem, but
also to describe them in terms of sequences, compatibly with the sequential
description of continuous mappings. The theorem itself will then be easy to
prove. Accordingly, most of the section will consist of describing sets in two
waysin terms that are easy to recognize, and in sequential language that
dovetails with continuity.
We begin with a little machinery to quantify the intuitive notion of near-
ness.
Definition 2.4.1 (-ball). For any point p Rn and any positive real num-
ber > 0, the -ball centered at p is the set
B(p, ) = {x Rn : |x p| < } .
(See figure 2.11.)

Figure 2.11. Balls in various dimensions

With -balls it is easy to describe the points that are approached by a

set A.
52 2 Euclidean Space

Definition 2.4.2 (Limit Point). Let A be a subset of Rn , and let p be a

point of Rn . The point p is a limit point of A if every -ball centered at p
contains some point x A such that x 6= p.
A limit point of A need not belong to A (exercise 2.4.2). On the other
hand, a point in A need not be a limit point of A (exercise 2.4.2 again); such
a point is called an isolated point of A. Equivalently, p is an isolated point
of A if p A and there exists some > 0 such that B(p, ) A = {p}. The
next lemma justifies the terminology of the previous definition: limit points
of A are precisely the (nontrivial) limits of sequences in A.
Lemma 2.4.3 (Sequential Characterization of Limit Points). Let A
be a subset of Rn , and let p be a point of Rn . Then p is the limit of a sequence
{x } in A with each x 6= p if and only if p is a limit point of A.
Proof. ( = ) If p is the limit of sequence {x } in A with each x 6= p then
any -ball about p contains an x (in fact, infinitely many), so p is a limit
point of A.
( = ) Conversely, if p is a limit point of A then B(p, 1/2) contains some
x1 A, x1 6= p. Let 2 = |x1 p|/2. The ball B(p, 2 ) contains some x2 A,
x2 6= p. Let 3 = |x2 p|/2 and continue defining a sequence {x } in this
fashion with |x p| < 1/2 for all . This sequence converges to p and each
x 6= p.

The lemma shows that Definition 2.4.2 is more powerful than it appears
every -ball centered at a limit point of A contains not only one but infinitely
many points of A.
Definition 2.4.4 (Closed Set). A subset A of Rn is closed if it contains
all of its limit points.
For example, the x1 -axis is closed as a subset of Rn since any point off
the axis is surrounded by a ball that misses the axisthat is, any point off
the axis is not a limit point of the axis, i.e., the axis is not missing any of
its limit points, i.e., the axis contains all of its limit points. The interval
(0, 1) is not closed because it does not contain the limit points at its ends.
These examples illustrate the fact that with a little practice it becomes easy
to recognize quickly whether a set is closed. Loosely speaking, a set is closed
when it contains all the points that it seems to want to contain.
Proposition 2.4.5 (Sequential Characterization of Closed Sets). Let
A be a subset of Rn . Then A is closed if and only if every sequence in A that
converges in Rn in fact converges in A.
Proof. ( = ) Suppose that A is closed, and let {x } be a sequence in A
converging in Rn to p. If x = p for some then p A since x A; and if
x 6= p for all then p is a limit point of A by = of Lemma 2.4.3, and
so p A since A is closed.
 2.4 Topology: Compact Sets and Continuity 53

( = ) Conversely, suppose that every convergent sequence in A has its

limit in A. Then all limit points of A are in A by = of Lemma 2.4.3,
and so A is closed.

The proposition equates an easily recognizable condition that we can un-

derstand intuitively (a set being closed) with a sequential characterization
that we can use in further arguments. Note that the sequential characteriza-
tion of a closed set A refers not only to A but also to the ambient space Rn
where A lies. We will return to this point later in this section.
Closed sets do not necessarily have good properties under continuous map-
pings. So next we describe another class of sets, the bounded sets. Bounded-
ness is again an easily recognizable condition that also has a characterization
in terms of sequences. The sequential characterization will turn out to be
complementary to the sequential characterization of closed sets, foreshadow-
ing that the properties of being closed and bounded will work well together.

Definition 2.4.6 (Bounded Set). The set A in Rn is bounded if A

B(0, R) for some R > 0.

Thus a bounded set is enclosed in some finite corral centered at the origin,
possibly a very big one. For example, any ball B(p, ), not necessarily centered
at the origin, is bounded, by a nice application of the Triangle Inequality
(exercise 2.4.5). On the other hand, the Archimedean property of the real
number system says that Z is an unbounded subset of R. The Size Bounds
show that any subset of Rn is bounded if and only if the jth coordinates of
its points form a bounded subset of R for each j {1, , n}. The geometric
content of this statement is that a set sits inside a ball centered at the origin
if and only if it sits inside a box centered at the origin.
Blurring the distinction between a sequence and the set of its elements
allows the definition of boundedness to apply to sequences. That is, a sequence
{x } is bounded if there is some R > 0 such that |x | < R for all Z+ . The
proof of the next fact in Rn is symbol-for-symbol the same as in R (or in C),
so it is only sketched.

Proposition 2.4.7 (Convergence Implies Boundedness). If the sequence

{x } converges in Rn then it is bounded.

Proof. Let {x } converge to p. Then there exists a starting index 0 such that
x B(p, 1) for all > 0 . Consider any real number R such that

R > max{|x1 |, , |x0 |, |p| + 1}.

Then clearly x B(0, R) for = 1, , 0 , and the Triangle Inequality

shows that also x B(0, R) for all > 0 . Thus {x } B(0, R) as a set.


54 2 Euclidean Space

Definition 2.4.8 (Subsequence). A subsequence of the sequence {x } is

a sequence consisting of some (possibly all) of the original terms, in ascending
order of indices.

Since a subsequence of {x } consists of terms x only for some values of ,

it is often written {xk }, where now k is the index variable. For example, given
the sequence
{x1 , x2 , x3 , x4 , x5 , } ,
a subsequence is
{x2 , x3 , x5 , x7 , x11 , },
with 1 = 2, 2 = 3, 3 = 5, and generally k = the kth prime.

Lemma 2.4.9 (Persistence of Convergence). Let {x } converge to p.

Then any subsequence {xk } also converges to p.

Proof. The hypothesis that {x } converges to p means that for any given
> 0, only finitely many sequence-terms x lie outside the ball B(p, ). Con-
sequently only finitely many subsequence-terms xk lie outside B(p, ), which
is to say that {xk } converges to p.

The sequence property that characterizes bounded sets is called the

BolzanoWeierstrass property. Once it is proved in R, the result follows
in Rn by arguing one component at a time.

Theorem 2.4.10 (BolzanoWeierstrass Property in R). Let A be a

bounded subset of R. Then every sequence in A has a convergent subsequence.

Proof. Let {x } be a sequence in A. Call a term x of the sequence a max-

point if it is at least as big as all later terms, i.e., x x for all > .
(For visual intuition, draw a graph plotting x as a function of , with line
segments connecting consecutive points. A max-point is a peak of the graph
at least as high as all points to its right.) If there are infinitely many max-
points in {x } then these form a decreasing sequence. If there are only finitely
many max-points then {x } has an increasing sequence starting after the
last max-pointthis follows almost immediately from the definition of max-
point. In either case, {x } has a monotonic subsequence which, being bounded,
converges because the real number system is complete.

Theorem 2.4.11 (BolzanoWeierstrass Property in Rn : Sequential

Characterization of Bounded Sets). Let A be a subset of Rn . Then A is
bounded if and only if every sequence in A has a subsequence that converges
in Rn .

Proof. ( = ) Suppose that A is bounded. Consider any sequence {x }

in A, written as {(x1, , , xn, )}. The real sequence {x1, } takes values in
a bounded subset of R and thus has a convergent subsequence, {x1,k }. The
 2.4 Topology: Compact Sets and Continuity 55

subscripts are getting out of hand, so keep only the k th terms of the orig-
inal sequence and relabel it. In other words, we may as well assume that
the sequence of first components, {x1, }, converges. The real sequence of
second components, {x2, }, in turn has a convergent subsequence, and by
Lemma 2.4.9 the corresponding subsequence of first components, {x1, }, con-
verges too. Relabeling again, we may assume that {x1, } and {x2, } both
converge. Continuing in this fashion n 2 more times exhibits a subsequence
of {x } that converges at each component.
( = ) Conversely, suppose that A is not bounded. Then there is a se-
quence {x } in A with |x | > for all . This sequence has no bounded subse-
quence, and hence it has no convergent subsequence by Proposition 2.4.7.

Note how the sequential characterizations in Proposition 2.4.5 and in the

BolzanoWeierstrass Property complement each other. The proposition char-
acterizes any closed set in Rn by the fact that if a sequence converges in the
ambient space then it converges in the set. The BolzanoWeierstrass Prop-
erty characterizes any bounded set in Rn by the fact that every sequence in
the set has a subsequence that converges, but not necessarily in the set. Also
note that like the sequential characterization of a closed set, the sequential
characterization of a bounded set A refers to the ambient space Rn where A
lies. We will return to this point once more in this section.

Definition 2.4.12 (Compact Set). A subset K of Rn is compact if it is

closed and bounded.

Since the static notions of closed and bounded are reasonably intuitive, we
can usually recognize compact sets on sight. But it is not obvious from how
compact sets look that they are related to continuity. So our program now
has two steps: first, combine Proposition 2.4.5 and the BolzanoWeierstrass
property to characterize compact sets in terms of sequences, and second, use
the characterization to prove that compactness is preserved by continuous
mappings.

Theorem 2.4.13 (Sequential Characterization of Compact Sets). Let

K be a subset of Rn . Then K is compact if and only if every sequence in K
has a subsequence that converges in K.

Proof. ( = ) We show that the sequential characterizations of closed and

bounded sets together imply the claimed sequential characterization of com-
pact sets. Suppose that K is compact and {x } is a sequence in K. Then K is
bounded, so by = of the BolzanoWeierstrass property {x } has a con-
vergent subsequence. But K is also closed, so by = of Proposition 2.4.5,
this subsequence converges in K.
( = ) Conversely, we show that the claimed sequential characterization of
compact sets subsumes the sequential characterizations of closed and bounded
sets. Thus, suppose that every sequence in K has a subsequence that converges
56 2 Euclidean Space

in K. Then in particular, every sequence in K that converges in Rn has a sub-

sequence that converges in K. By Lemma 2.4.9 the limit of the sequence is
the limit of the subsequence, so the sequence converges in K. That is, every
sequence in K that converges in Rn converges in K, and hence K is closed
by = of Proposition 2.4.5. Also in consequence of the claimed sequen-
tial property of compact sets, every sequence in K has a subsequence that
converges in Rn . Thus K is bounded by = of the BolzanoWeierstrass
Property.

By contrast to the sequential characterizations of a closed set and of a

bounded set, the sequential characterization of a compact set K makes no
reference to the ambient space Rn where K lies. A sets property of being
compact is innate in a way that a sets property of being closed or of being
bounded is not.
The next theorem is the main result of this section. Now that all of the
objects involved are described in the common language of sequences, its proof
is natural.

Theorem 2.4.14 (The Continuous Image of a Compact Set is Com-

pact). Let K be a compact subset of Rn and let f : K Rm be continuous.
Then f (K), the image set of K under f , is a compact subset of Rm .

Proof. Let {y } be any sequence in f (K); by = of Theorem 2.4.13, it

suffices to exhibit a subsequence converging in f (K). Each y has the form
f (x ), and this defines a sequence {x } in K. By = of Theorem 2.4.13,
since K is compact, {x } necessarily has a subsequence {xk } converging in K,
say to p. By the continuity of f at p, the sequence {f (xk )} converges in f (K)
to f (p). Since {f (xk )} is a subsequence of {y }, the proof is complete.

Again, the sets in Theorem 2.4.14 are defined with no direct reference to
sequences, but the theorem is proved entirely by using sequences. The point
is that with the theorem proved, we can easily see that it applies in particular
contexts without having to think any more about the sequences that were
used to prove it.
A corollary of Theorem 2.4.14 generalizes the theorem that was quoted to
begin the section:

Theorem 2.4.15 (Extreme Value Theorem). Let K be a nonempty com-

pact subset of Rn and let the function f : K R be continuous. Then f
takes a minimum and a maximum value on K.

Proof. By Theorem 2.4.14, f (K) is a compact subset of R. As a nonempty

bounded subset of R, f (K) has a greatest lower bound and a least upper
bound by the completeness of the real number system. Each of these bounds
is an isolated point or a limit point of f (K), since otherwise some -ball about
it would be disjoint from f (K), giving rise to greater lower bounds or lesser
upper bounds of f (K). Since f (K) is also closed it contains its limit points,
 2.4 Topology: Compact Sets and Continuity 57

so in particular it contains its greatest lower bound and its least upper bound.
This means precisely that f takes a minimum and a maximum value on K.

Even when n = 1, Theorem 2.4.15 generalizes the Extreme Value Theorem

from the beginning of the section. In the theorem here, K can be a finite union
of closed and bounded intervals in R rather than only one interval, or K can
be a more complicated set, provided only that it is compact.
A topological property of sets is a property that is preserved under continu-
ity. Theorem 2.4.14 says that compactness is a topological property. Neither
the property of being closed nor of being bounded is in itself topological. That
is, the continuous image of a closed set need not be closed, and the continuous
image of a bounded set need not be bounded; for that matter, the continuous
image of a closed set need not be bounded, and the continuous image of a
bounded set need not be closed (exercise 2.4.8).
The nomenclature continuous image in the slogan-title of Theorem 2.4.14
and in the previous paragraph is, strictly speaking, inaccurate: the image of
a mapping is a set, and the notion of a set being continuous doesnt even
make sense according to our grammar. As stated correctly in the body of the
theorem, continuous image is short for image under a continuous mapping.
The property that students often have in mind when they call a set contin-
uous is in fact called connectedness. Loosely, a set is connected if it has only
one piece, so that a better approximating word from everyday language is con-
tiguous. To define connectedness accurately, we would have to use methodol-
ogy exactly opposite that of the section: rather than relate sets to continuous
mappings by characterizing the sets in terms of sequences, the idea is to turn
the whole business around and characterize continuous mappings in terms of
sets, specifically in terms of open balls. However, the process of doing so, and
then characterizing compact sets in terms of open balls as well, is trickier
than characterizing sets in terms of sequences; and so we omit it since we do
not need connectedness. Indeed, the remark after Theorem 2.4.15 points out
that connectedness is unnecessary even for the one-variable Extreme Value
Theorem.
However, it deserves passing mention that connectedness is also a topologi-
cal property: again using language loosely, the continuous image of a connected
set is connected. This statement generalizes another theorem that underlies
one-variable calculus, the Intermediate Value Theorem. For a notion related
to connectedness that is easily shown to be a topological property, see exer-
cise 2.4.10.
The ideas of this section readily extend to broader environments. The first
generalization of Euclidean space is a metric space, a set with a well-behaved
distance function. Even more general is a topological space, a set with some
of its subsets designated as closed. Continuous functions, compact sets, and
connected sets can be defined meaningfully in these environments, and the
58 2 Euclidean Space

theorems remain the same: the continuous image of a compact set is compact,
and the continuous image of a connected set is connected.

Exercises

2.4.1. Are the following subsets of Rn closed, bounded, compact?

(a) B(0, 1),
(b) {(x, y) R2 : y x2 = 0},
(c) {(x, y, z) R3 : x2 + y 2 + z 2 1 = 0},
(d) {x : f (x) = 0m } where f M(Rn , Rm ) is continuous (this generalizes
(b) and (c)),
(e) Qn where Q denotes the rational numbers,
(f) {(x1 , , xn ) : x1 + + xn > 0}.

2.4.2. Give a set A Rn and limit point b of A such that b / A. Give a set
A Rn and a point a A such that a is not a limit point of A.

2.4.3. Let A be a closed subset of Rn and let f M(A, Rm ). Define the

graph of f to be
G(f ) = {(a, f (a)) : a A},
a subset of Rn+m . Show that if f is continuous then its graph is closed.

2.4.4. Prove the closed set properties: (1) The empty set and the full space
Rn are closed subsets of Rn , (2) any intersection of closed sets is closed, (3)
any finite union of closed sets is closed.

2.4.5. Prove that any ball B(p, ) is bounded in Rn .

2.4.6. Show that A is a bounded subset of Rn if and only if for each j

{1, , n}, the jth coordinates of its points form a bounded subset of R.

2.4.7. Show by example that a closed set need not satisfy the sequential char-
acterization of bounded sets, and that a bounded set need not satisfy the
sequential characterization of closed sets.

2.4.8. Show by example that the continuous image of a closed set need not
be closed, that the continuous image of a closed set need not be bounded,
that the continuous image of a bounded set need not be closed, and that the
continuous image of a bounded set need not be bounded.

2.4.9. A subset A of Rn is called discrete if each of its points is isolated.

(Recall that the term isolated was defined in the section.) Show or take for
granted the (perhaps surprising at first) fact that every mapping whose do-
main is discrete must be continuous. Is discreteness a topological property?
That is, need the continuous image of a discrete set be discrete?
 2.5 Review of the One-Variable Derivative 59

2.4.10. A subset A of Rn is called path-connected if for any two points

x, y A, there is a continuous mapping

: [0, 1] A

such that (0) = x and (1) = y. (This is the path that connects x and y.)
Draw a picture to illustrate the definition of a path-connected set. Prove that
path-connectedness is a topological property.

2.5 Review of the One-Variable Derivative

The limit of a sequence was introduced in Definition 2.3.3. The limit of a
mapping will now be defined as the common limit of all suitable sequences, if
this common sequence limit exists. Recall from Definition 2.4.2 that a point
a is a limit point of a set A if every -ball centered at a contains some point
x A such that x 6= a. A limit point of A may or may not be a point of A.
Also recall from Lemma 2.4.3 that a point a is a limit point of a set A if and
only if a is the limit of a sequence {x } in A with each x 6= a.
Definition 2.5.1 (Function Limit). Let A be a subset of Rn , let f : A
Rm be a mapping, and let a be a limit point of A. Let be a point of Rm .
Then f has limit as x approaches a, written

lim f (x) = ,
xa

if for every sequence {x } in A with each x 6= a such that {x } converges

to a, the corresponding output sequence {f (x )} converges to .
Thus the notion of limxa f (x) makes no reference to f (a) (which may
not even be defined), but only to values f (x) for x near a.
The Sum Rule and the Constant Multiple Rule for sequence limits give
rise to the same rules for mapping limits as well, but there is one technical
issue. The Sum Rule seems obvious,

lim (f (x) + g(x)) = lim f (x) + lim g(x),

xa xa xa

where f : A Rm and a is a limit point of A, and g : B Rm and a is a

limit point of B. But one needs to observe that the domain of f + g is A B,
and so the limit on the left can not exist unless the limit point a of A and
a limit point of the smaller set A B. For example, the functions
of B is also

x and x have the respective domains [0, ) and (, 0], and

lim x = lim x = 0,
x0 x0

but the function x + x has the domain [0, ) (, 0] = {0}, and 0 is
not a limit point of this set, so therefore
60 2 Euclidean Space

lim ( x + x) does not exist.
x0

This should be clear in light of the sentence immediately after Definition 2.5.1.
Returning to the general Sum Rule for mappings, other than this additional
detail to check, it follows from its counterpart for sequences. The Constant
Multiple Rule for mappings follows from its counterpart for sequences with-
out any additional technical considerations, since any constant multiple of a
mapping has the same domain as the original mapping.
Let A Rn be a set and let a Rn be a point. A mapping f : A Rm
is null at a if limxa f (x) = 0m . Thus if f is null at a then a must be a limit
point of A. Formulating the Sum Rule and the Constant Multiple Rule for
null mappings is left to you (exercise 2.5.1).
The notions of limit and continuity are closely related for mappings, but
again with a small technical issue present. The proof of the following propo-
sition is exercise 2.5.2.

Proposition 2.5.2 (Continuity in Terms of Function Limits). Let A be

a subset of Rn , and let a be a point of A, and let f : A Rm be a mapping.
Suppose that a is a limit point of A. Then f is continuous at a if and only
if limxa f (x) exists and is equal to f (a).
Suppose that a is not a limit point of A. Then f is continuous at a.

A careful discussion of the derivative is surprisingly technical even for

functions of one variable. The one-variable derivative is defined as a limit of a
difference quotient function. Usually the underlying assumptions, which can
easily get lost, are that f is a function from some interval I R to R and that
a is a point of I but not an endpoint. (Some authors allow differentiation at
endpoints, but then the derivative can exist and be nonzero at an extremum of
the function.) The difference quotient function is defined at all points except 0
of the interval J obtained by translating I by a, moving a to 0,

f (a + h) f (a)
g : J {0} R, g(h) = .
h
Thus 0 is a limit point of the domain of g (though not a point of the domain
of g), so that according to Definition 2.5.1, limh0 g(h) might exist. When it
does, the derivative of f at a is this function limit,

f (a + h) f (a)
f (a) = lim g(h) = lim .
h0 h0 h
In sum, the derivative of f at a is
a limit of a different function g, the difference quotient function whose
domain is obtained by translating and puncturing the domain of f ,
the limit being taken at the limit point 0 of the domain of g, which is not
in the domain of g,
 2.6 Summary 61

and the function limit being defined as the common value of output-
sequence limits over all input-sequences that approach but do not reach 0,
if this common value of output-sequence limits exists.
The full definition of the derivative in a first calculus course may have been dif-
ficult to digest, because all of these ideas had to be covered under intense time
pressure in the midst of everything else that was happening in that course,
and because the very process of getting all the ideas into play necessarily
rendered their presentation diffuse.
However, the author of these notes does not know any useful way to sim-
plify the setup without waving his hands. One can study an alternate differ-
ence quotient function g(x) = (f (x) f (a))/(x a) instead and thus avoid
translating the domain of f to place the puncture-point at 0, but this is not
not a good idea: in the definition of multivariable derivative to be introduced
in chapter 4, translating the situation to the origin will clarify rather than
complicate it. Also, one can define the limit of a function without reference
to sequence-limits, using the so-called epsilondelta definition rather than our
epsilonnu. For example, the formulation of the completeness of the real num-
ber system as a set-bound criterion in Theorem 1.1.4 makes no reference to
sequences, and if continuity of mappings is defined in epsilondelta language
then the Persistence of Inequality principle, which was a small nuisance to
prove, becomes true by definition. However, eschewing sequences and basing
all of the ideas in play here on an epsilondelta formulation of limit makes
other parts of the material harder. In particular, proving that compactness is a
topological property without using the sequential characterization of compact
sets requires considerable subtlety.

Exercises

2.5.1. Carefully state and prove the Sum Rule and the Constant Multiple
Rule for mappings and then for null mappings.

2.5.2. Prove Proposition 2.5.2.

2.6 Summary

Along with introducing Euclidean space and its properties, this chapter is
meant to provide a quick review of some ideas from one-variable calculus
while generalizing them to higher dimension. This chapter has also empha-
sized working with vectors intrinsically rather than using coordinates. The
multivariable Extreme Value Theorem will play a crucial role in our proof of
the Inverse Function Theorem in chapter 5.
3
Linear Mappings and Their Matrices

The basic idea of differential calculus is to approximate smooth-but-curved

objects in the small by straight ones. To prepare for doing so, this chapter
studies the multivariable analogues of lines. With one variable, lines are easily
manipulated by explicit formulas (e.g., the pointslope form is y = mx + b),
but with many variables we want to use the language of mappings. Section 3.1
gives an algebraic description of straight mappings, the linear mappings,
proceeding from an intrinsic definition to a description in coordinates. Each
linear mapping is described by a box of numbers called a matrix, so section 3.2
derives mechanical matrix manipulations corresponding to the natural ideas
of adding, scaling, and composing linear mappings. Section 3.3 discusses in
matrix terms the question of whether a linear mapping has an inverse, i.e.,
whether there is a second linear mapping such that each undoes the others
effect. Section 3.5 discusses the determinant, an elaborate matrix-to-scalar
function that extracts from a linear mapping a single number with remarkable
properties:
(Linear Invertibility Theorem) The mapping is invertible if and only if the
determinant is nonzero.
An explicit formula for the inverse of an invertible linear mapping can be
written using the determinant (section 3.7).
The factor by which the mapping magnifies volume is the absolute value
of the determinant (section 3.8).
The mapping preserves or reverses orientation according to the sign of the
determinant (section 3.9). Here orientation is an algebraic generalization of
clockwise versus counterclockwise in the plane and of right-handed versus
left-handed in space.
Finally, section 3.10 defines the cross product (a vector-by-vector multiplica-
tion special to three dimensions) and uses it to derive formulas for lines and
planes in space.
64 3 Linear Mappings and Their Matrices

3.1 Linear Mappings

The simplest interesting mappings from Rn to Rm are those whose output is
proportional to their input, the linear mappings. Proportionality means that
a linear mapping should take a sum of inputs to the corresponding sum of
outputs,
T (x + y) = T (x) + T (y) for all x, y Rn , (3.1)
and a linear mapping should take a scaled input to the correspondingly scaled
output,
T (x) = T (x) for all R, x Rn . (3.2)
(Here we use the symbol because a will be used heavily in other ways during
this chapter.) More formally,

Definition 3.1.1 (Linear Mapping). The mapping T : Rn Rm is

linear if !
Xk k
X
T i x i = i T (xi )
i=1 i=1

for all positive integers k, all real numbers 1 through k , and all vectors x1
through xk .

The reader may find this definition discomfiting. It does not say what form
a linear mapping takes, and this raises some immediate questions. How are we
to recognize linear mappings when we encounter them? Or are we supposed to
think about them without knowing what they look like? For that matter, are
there even any linear mappings to encounter? Another troublesome aspect of
Definition 3.1.1 is semantic: despite the geometric sound of the word linear,
the definition is in fact algebraic, describing how T behaves with respect to
the algebraic operations of vector addition and scalar multiplication. (Note
that on the left of the equality in the definition, the operations are set in Rn ,
while on the right they are in Rm .) So what is the connection between the
definition and actual lines? Finally, how exactly do conditions (3.1) and (3.2)
relate to the condition in the definition?
On the other hand, Definition 3.1.1 has the virtue of illustrating the prin-
ciple that to do mathematics effectively we should characterize our objects
rather than construct them. The characterizations are admittedly guided by
hindsight, but there is nothing wrong with that. Definition 3.1.1 says how
a linear mapping behaves. It says that whatever form linear mappings will
turn out to take, our reflex should be to think of them as mappings through
which we can pass sums and constants. (This idea explains why one of the
inner product properties is called bilinearity: the inner product is linear as a
function of either of its two vector variables when the other variable is held
fixed.) The definition of linearity tells us how to use linear mappings once we
know what they are, or even before we know what they are. Another virtue
of Definition 3.1.1 is that it is intrinsic, making no reference to coordinates.
 3.1 Linear Mappings 65

Some of the questions raised by Definition 3.1.1 have quick answers. The
connection between the definition and actual lines will quickly emerge from our
pending investigations. Also, an induction argument shows that (3.1) and (3.2)
are equivalent to the characterization in the definition, despite appearing
weaker (exercise 3.1.1). Thus, to verify that a mapping is linear, we only need
to show that it satisfies the easier-to-check conditions (3.1) and (3.2); but to
derive properties of mappings that are known to be linear, we may want to
use the more powerful condition in the definition. As for finding linear map-
pings, the definition suggests a two-step strategy: first, derive the form that
a linear mapping necessarily takes in consequence of satisfying the definition;
and second, verify that the mappings of that form are indeed linear, i.e., show
that the necessary form of a linear mapping is also sufficient for a mapping
to be linear. We now turn to this.
The easiest case to study is linear mappings from R to R. Following the
strategy, first we assume that we have such a mapping and determine its form,
obtaining the mappings that are candidates to be linear. Second we show
that all the candidates are indeed linear mappings. Thus suppose that some
mapping T : R R is linear. The mapping determines a scalar, a = T (1).
And then for any x R,

T (x) = T (x 1) since x 1 = x
= xT (1) by (3.2)
= xa by definition of a
= ax since multiplication in R commutes.

Thus, T is simply multiplication by a, where a = T (1). But to reiterate, this

calculation does not show that any mapping is linear, it shows only what form
a mapping must necessarily have once it is already known to be linear. We
dont yet know that any linear mappings exist at all.
So the next thing to do is show that conversely any mapping of the derived
form is indeed linearthe necessary condition is also sufficient. Fix a real
number a and define a mapping T : R R by T (x) = ax. Then the claim
is that T is linear and T (1) = a. Lets partially show this by verifying that T
satisfies (3.2). For any R and any x R,

T (x) = ax by definition of T
= ax since multiplication in R commutes
= T (x) by definition of T ,

as needed. You can check (3.1) similarly, and the calculation that T (1) = a is
immediate. These last two paragraphs combine to show
Proposition 3.1.2 (Description of Linear Mappings from Scalars to
Scalars). The linear mappings T : R R are precisely the mappings

T (x) = ax
66 3 Linear Mappings and Their Matrices

where a R. That is, each linear mapping T : R R is multiplication by a

unique a R and conversely.

The slogan encapsulating the proposition is:

For scalar input and scalar output, linear OF is scalar TIMES.

Also, the proposition explains the term linear: the graphs of linear mappings
from R to R are lines through the origin. (Mappings f (x) = ax + b with
b 6= 0 are not linear according to our definition even though their graphs are
also lines. However, see exercises 3.1.15 and 3.2.6.) For example, a typical
linear mapping from R to R is T (x) = (1/2)x. Figure 3.1 shows two ways
of visualizing this mapping. The left half of the figure plots the domain axis
and the codomain axis in one plane, orthogonally to each other, the familiar
way to graph a function. The right half of the figure plots the axes separately,
using the spacing of the dots to describe the mapping instead. The uniform
spacing along the rightmost axis depicts the fact that T (x) = xT (1) for all
x Z, and the spacing is half as big because the multiplying factor is 1/2.
Figures of this second sort can generalize up to three dimensions of input and
three dimensions of output, whereas figures of the first sort can display at
most three dimensions of input and output combined.

T (x)

T
x
0 1 0 T (1)

Figure 3.1. A linear mapping from R to R

Next consider a linear mapping T : Rn R. Recall the standard basis

vectors of Rn ,

e1 = (1, 0, , 0), , en = (0, 0, , 1).

Take the n real numbers

a1 = T (e1 ), , an = T (en ),

and define the vector a = (a1 , , an ) Rn . Any x Rn can be written

n
X
x = (x1 , , xn ) = x i ei , each xi R.
i=1
 3.1 Linear Mappings 67

(So here each xi is a scalar entry of the vector x, whereas in Definition 3.1.1,
each xi was itself a vector. The author does not know any graceful way to
avoid this notation collision, the systematic use of boldface or arrows to adorn
vector names being heavyhanded, and the systematic use of the Greek letter
rather than its Roman counterpart x to denote scalars being alien. Since
mathematics involves finitely many symbols and infinitely many ideas, the
reader will in any case eventually need the skill of discerning meaning from
context, a skill that may as well P start receiving practice now.) Returning to
n
the main discussion, since x = i=1 xi ei and T is linear, Definition 3.1.1
shows that
n
! n n
X X X
T (x) = T x i ei = xi T (ei ) = xi ai = hx, ai = ha, xi.
i=1 i=1 i=1

Again, the only possibility for the linear mapping is multiplication by an

element a, where now a = (T (e1 ), , T (en )) is a vector and the multiplication
is an inner product, but we dont yet know that such a mapping is linear.
However, fix a vector a = (a1 , , an ) and define the corresponding mapping
T : Rn R by T (x) = ha, xi. Then it is straightforward to show that indeed
T is linear and T (ej ) = aj for j = 1, , n (exercise 3.1.3). Thus we have

Proposition 3.1.3 (Description of Linear Mappings from Vectors to

Scalars). The linear mappings T : Rn R are precisely the mappings

T (x) = ha, xi

where a Rn . That is, each linear mapping T : Rn R is multiplication by

a unique a Rn and conversely.

The slogan encapsulating the proposition is:

For vector input and scalar output, linear OF is vector TIMES.

In light of the proposition, you should be able to recognize linear mappings

from Rn to R on sight. For 3
example, the mapping T : R R given by
T (x, y,
z) = x + ey + 2z is linear, being multiplication by the vector
(, e, 2).
In the previous chapter, the second example after Definition 2.3.6 showed
that every linear mapping T : Rn R is continuous. You are encouraged to
reread that example now before continuing.
A depiction of a linear mapping from R2 to R can again plot the domain
plane and the codomain axis orthogonally to each other or separately. See
figures 3.2 and 3.3 for examples of each type of plot. The first figure suggests
that the graph forms a plane in R3 and that a line of inputs is taken to
the output value 0. The second figure shows more clearly how the mapping
compresses the plane into the line. As in the right half of figure 3.1, the idea
is that T (x, y) = xT (1, 0) + yT (0, 1) for all x, y Z. The compression is that
68 3 Linear Mappings and Their Matrices

Figure 3.2. The graph of a linear mapping from R2 to R

Figure 3.3. Second depiction of a linear mapping from R2 to R

although (1, 0) and (0, 1) lie on separate input axes, T (1, 0) and T (0, 1) lie on
the same output axis.
The most general mapping is T : Rn Rm . Such a mapping decomposes
as T = (T1 , , Tm ) where each Ti : Rn R is the ith component function
of T . The next proposition reduces the linearity of such T to the linearity of
its components Ti , which we already understand.
Proposition 3.1.4 (Componentwise Nature of Linearity). The vector-
valued mapping T = (T1 , , Tm ) : Rn Rm is linear if and only if each
scalar-valued component function Ti : Rn R is linear.
Proof. For any x, y Rn ,


T (x + y) = T1 (x + y), , Tm (x + y)

and



T (x) + T (y) = T1 (x), , Tm (x) + T1 (y), , Tm (y)


= T1 (x) + T1 (y), , Tm (x) + Tm (y) .
But T satisfies (3.1) exactly when the left sides are equal, the left sides are
equal exactly when the right sides are equal, and the right sides are equal
 3.1 Linear Mappings 69

exactly when each Ti satisfies (3.1). A similar argument with (3.2), left as
exercise 3.1.5, completes the proof.

The componentwise nature of linearity combines with the fact that scalar-
valued linear mappings are continuous (as observed after Proposition 3.1.3)
and with the componentwise nature of continuity to show that all linear map-
pings are continuous. Despite being so easy to prove, this fact deserves a
prominent statement.
Theorem 3.1.5 (Linear Mappings are Continuous). Let the mapping
T : Rn Rm be linear. Then T is continuous.

By the previous proposition, a mapping T : Rn Rm is linear if and only

if each Ti determines n real numbers ai1 , , ain . Putting all mn numbers aij
into a box with m rows and n columns gives a matrix

a11 a12 a1n
a21 a22 a2n

A= . .. .. (3.3)
.. . .
am1 am2 amn

whose ith row is the vector determined by Ti , and whose (i, j)th entry (this
means ith row, jth column) is thus given by

aij = Ti (ej ). (3.4)

Sometimes one saves writing by abbreviating the right side of (3.3) to [aij ]mn ,
or even just [aij ] when m and n are firmly established.
The set of all m-by-n matrices (those with m rows and n columns) of real
numbers is denoted Mm,n (R). The n-by-n square matrices are denoted Mn (R).
Euclidean space Rn is often identified with Mn,1 (R) and vectors written as
columns,
x1
..
(x1 , , xn ) = . .
xn
This typographical convention may look odd, but it is useful. The idea is that
a vector in parentheses is merely an ordered list of entries, not inherently a
row or a column; but when a vectoror, more generally, a matrixis enclosed
by square brackets, the distinction between rows and columns is significant.
To make the linear mapping T : Rn Rm be multiplication by its matrix
A Mm,n (R), we need to define multiplication of an m-by-n matrix A by an
n-by-1 vector x appropriately. That is, the only sensible definition is
Definition 3.1.6 (Matrix-by-Vector Multiplication). Let A Mm,n (R)
and let x Rn . Then the product Ax Rm is defined the vector whose ith
entry is the inner product of As ith row and x,
70 3 Linear Mappings and Their Matrices

x1
a11 a12 a1n x2 a11 x1 + + a1n xn
a21 a22 a2n
..
a21 x1 + + a2n xn
Ax = . .. .. . = .. .
.. . . . .
..
am1 am2 amn am1 x1 + + amn xn
xn

For example,

  7    
123 17+28+39 50
8 = = .
456 47+58+69 122
9

Definition 3.1.6 is designed to give the following theorem, which encompasses

Propositions 3.1.2 and 3.1.3 as special cases.

Theorem 3.1.7 (Description of Linear Mappings from Vectors to

Vectors). The linear mappings T : Rn Rm are precisely the mappings

T (x) = Ax

where A Mm,n (R). That is, each linear mapping T : Rn Rm is multi-

plication by a unique A Mm,n (R) and conversely.

The slogan encapsulating the proposition is:

For vector input and vector output, linear OF is matrix TIMES.

Recall the meaning of the rows of a matrix A that describes a correspond-

ing linear mapping T :

The ith row of A describes Ti , the ith component function of T .

The columns of A also have a description in terms of T . Indeed, the jth column
is
a1j T1 (ej )
.. ..
. = . = T (ej ).
amj Tm (ej )
That is:

The jth column of A is T (ej ), i.e., is T of the jth standard basis vector.

For an example of using this last principle, let r : R2 R2 be the

mapping that rotates the plane counterclockwise through angle /6. It is
geometrically evident that r is linear: rotating the parallelogram P with sides
x1 and x2 (and thus with diagonal x1 + x2 ) by /6 yields the parallelogram
r(P ) with sides r(x1 ) and r(x2 ), so the diagonal of r(P ) is equal to both
 3.1 Linear Mappings 71

r(x1 + x2 ) = r(x1 ) + r(x2 )

x1 + x2
x2 r(x2 ) r(x1 )
x1

Figure 3.4. The rotation mapping is linear

r(x1 + x2 ) and r(x1 ) + r(x2 ). Thus r satisfies (3.1). The geometric verification
of (3.2) is similar. (See figure 3.4.)
To find the matrix A of r, simply compute that its columns are
   
3/2 1/2
r(e1 ) = r(1, 0) = , r(e2 ) = r(0, 1) = ,
1/2 3/2

and thus  
1/2
3/2
A= .
1/2 3/2
So now we know r because the rows of A describe its component functions,
    " 3 #  
1/2 x
3/2 x 21 y 3 1 1 3
r(x, y) = = 12 = x y, x + y .
1/2 3/2 y 2 x+ 2
3
y 2 2 2 2

Figures 3.5 through 3.8 show more depictions of linear mappings between
spaces of various dimensions. Note that although these mappings stretch and
torque their basic input grids, the grids still get taken to configurations of
straight lines. Contrast this to how the nonlinear mapping of figure 2.9 bent
the basic grid lines into curves.

Figure 3.5. A linear mapping from R to R2

We end this section by returning from calculations to intrinsic methods.

The following result could have come immediately after Definition 3.1.1, but it
72 3 Linear Mappings and Their Matrices

Figure 3.6. A linear mapping from R2 to R2

Figure 3.7. A linear mapping from R3 to R3

Figure 3.8. A linear mapping from R3 to R2

has been deferred to this point in order to present some of the objects involved
more explicitly first, to make them familiar. However, it is most easily proved
intrinsically.
Let L(Rn , Rm ) denote the set of all linear mappings from Rn to Rm . This
set not only sits inside the vector space M(Rn , Rm ), it is a vector space in its
own right:
 3.1 Linear Mappings 73

Proposition 3.1.8 (L(Rn , Rm ) Forms a Vector Space). Suppose that

S, T : Rn Rm are linear and that a R. Then the mappings

S + T, aS : Rn Rm

are also linear. Consequently, the set of linear mappings from Rn to Rm forms
a vector space.

Proof. The mappings S and T satisfy (3.1) and (3.2). We must show that
S + T and aS do the same. Compute for any x, y Rn ,

(S + T )(x + y)
= S(x + y) + T (x + y) by definition of + in M(Rn , Rm )
= S(x) + S(y) + T (x) + T (y) since S and T satisfy (3.1)
= S(x) + T (x) + S(y) + T (y) since addition in Rm commutes
= (S + T )(x) + (S + T )(y) by definition of + in M(Rn , Rm ).

Thus S + T satisfies (3.1). The other three statements about S + T and aS

satisfying (3.1) and (3.2) are similar and left as exercise 3.1.12. Once those
are established, the rest of the vector space axioms in L(Rn , Rm ) are readily
seen to be inherited from M(Rn , Rm ).

Also, linearity is preserved under composition. That is, if S : Rn Rm

and T : Rp Rn are linear then so is S T : Rp Rm (exercise 3.1.13).

Exercises

3.1.1. Prove that T : Rn Rm is linear if and only if it satisfies (3.1)

and (3.2). (It may help to rewrite (3.1) with the symbols x1 and x2 in place
of x and y. Then prove one direction by showing that (3.1) and (3.2) are
implied by the defining condition for linearity, and prove the other direction
by using induction to show that (3.1) and (3.2) imply the defining condition.
Note that as pointed out in the text, one direction of this argument has a bit
more substance than the other.)

3.1.2. Suppose that T : Rn Rm is linear. Show that T (0n ) = 0m . (An

intrinsic argument is nicer.)

3.1.3. Fix a vector a Rn . Show that the mapping T : Rn R given by

T (x) = ha, xi is linear, and that T (ej ) = aj for j = 1, , n.

3.1.4. Find the linear mapping T : R3 R such that T (0, 1, 1) = 1,

T (1, 0, 1) = 2, and T (1, 1, 0) = 3.

3.1.5. Complete the proof of the componentwise nature of linearity.

74 3 Linear Mappings and Their Matrices

3.1.6. Carry out the matrix-by-vector multiplications

100 1 ab   y1 1 1 0 1
 
1 1 0 2 , c d x ,
x1 xn ... , 0 1 1 1 .
y
111 3 ef yn 1 0 1 1

3.1.7. Prove that the identity mapping id : Rn Rn is linear. What is its

matrix? Explain.
3.1.8. Let denote a fixed but generic angle. Argue geometrically that the
mapping R : R2 R2 given by counterclockwise rotation by is linear, and
then find its matrix.
3.1.9. Show that the mapping Q : R2 R2 given by reflection through the
x-axis is linear. Find its matrix.
3.1.10. Show that the mapping P : R2 R2 given by orthogonal projection
onto the diagonal line x = y is linear. Find its matrix. (See exercise 2.2.15.)
3.1.11. Draw the graph of a generic linear mapping from R2 to R3 .
3.1.12. Continue the proof of Proposition 3.1.8 by proving the other three
statements about S + T and aS satisfying (3.1) and (3.2).
3.1.13. If S L(Rn , Rm ) and T L(Rp , Rn ), show that S T : Rp Rm
lies in L(Rp , Rm ).
3.1.14. (a) Let S L(Rn , Rm ). Its transpose is the mapping
S T : Rm Rn
defined by the characterizing condition
hx, S T (y)i = hS(x), yi for all x Rn and y Rm .
Granting that indeed a unique such S T exists, use the characterizing condition
to show that
S T (y + y ) = S T (y) + S T (y ) for all y, y Rm
by showing that
hx, S T (y + y )i = hx, S T (y) + S T (y )i for all x Rn and y, y Rm .
A similar argument (not requested here) shows that S T (y) = S T (y) for
all R and y Rm , and so the transpose of a linear mapping is linear.
(b) Keeping S from part (a), now further introduce T L(Rp , Rn ), so
that also S T L(Rp , Rm ). Show that the transpose of the composition is
the composition of the transposes in reverse order,
(S T )T = T T S T ,
by showing that
hx, (S T )T (z)i = hx, (T T S T )(z)i for all x Rp and z Rm .
 3.2 Operations on Matrices 75

3.1.15. A mapping f : Rn Rm is called affine if it has the form f (x) =

T (x) + b where T L(Rn , Rm ) and b Rm . State precisely and prove: the
composition of affine mappings is affine.

3.1.16. Let T : Rn Rm be a linear mapping. Note that since T is continu-

ous and since the absolute value function on Rm is continuous, the composite
function
|T | : Rn R
is continuous.
(a) Let S = {x Rn : |x| = 1}. Explain why S is a compact subset of Rn .
Explain why it follows that |T | takes a maximum value c on S.
(b) Show that |T (x)| c|x| for all x Rn . This result is the Linear
Magnification Boundedness Lemma. We will use it in chapter 4.

3.1.17. Let T : Rn Rm be a linear mapping.

(a) Explain why the set D = {x Rn : |x| = 1} is compact.
(b) Use part (a) of this exercise and part (b) of the preceding exercise
to explain why therefore the set {|T (x)| : x D} has a maximum. This
maximum is called the norm of T and is denoted kT k.
(c) Explain why kT k is the smallest value K that satisfies the condition
from part (b) of the preceding exercise, |T (x)| K|x| for all x Rn .
(d) Show that for any S, T L(Rn , Rm ) and any a R,

kS + T k kSk + kT k and kaT k = |a| kT k.

Define a distance function

d : L(Rn , Rm ) L(Rn , Rm ) R, d(S, T ) = kT Sk.

Show that this function satisfies the distance properties of Theorem 2.2.8.
(e) Show that for any S L(Rn , Rm ) and any T L(Rp , Rn ),

kST k kSkkT k.

3.2 Operations on Matrices

Having described abstract objects, the linear mappings T L(Rn , Rm ), with
explicit ones, the matrices A Mm,n (R) with (i, j)th entry aij = Ti (ej ), we
naturally want to study linear mappings via their matrices. The first step
is to develop rules for matrix manipulation corresponding to operations on
mappings. Thus if
S, T : Rn Rm
are linear mappings having matrices

A, B Mm,n (R),
76 3 Linear Mappings and Their Matrices

and if a is a real number, then the matrices for the linear mappings

S + T : Rn Rm and aS : Rn Rm

naturally should be denoted

A + B Mm,n (R) and aA Mm,n (R).

So + and (or juxtaposition) are about to acquire new meanings yet

again,
+ : Mm,n (R) Mm,n (R) Mm,n (R)
and
: R Mm,n (R) Mm,n (R).
To define the sum, fix j between 1 and n. Then

the jth column of A + B = (S + T )(ej )

= S(ej ) + T (ej )
= the sum of the jth columns of A and B.

And since vector addition is simply coordinatewise scalar addition, it follows

that for any i between 1 and m and any j between 1 and m, the (i, j)th entry
of A+B is the sum of the (i, j)th entries of A and B. (One can reach the same
conclusion in a different way by thinking about rows rather than columns.)
Thus the definition for matrix addition must be

Definition 3.2.1 (Matrix Addition).

If A = [aij ]mn and B = [bij ]mn then A + B = [aij + bij ]mn .

For example,      
12 1 0 02
+ = .
34 21 55
A similar argument shows that the appropriate definition to make for scalar
multiplication of matrices is
Definition 3.2.2 (Scalar-by-Matrix Multiplication).

If R and A = [aij ]mn then A = [aij ]mn .

For example,    
12 24
2 = .
34 68
The zero matrix 0m,n Mm,n (R), corresponding to the zero mapping in
L(Rn , Rm ), is the obvious one, with all entries 0. The operations in Mm,n (R)
precisely mirror those in L(Rn , Rm ), so
 3.2 Operations on Matrices 77

Proposition 3.2.3 (Mm,n (R) Forms a Vector Space). The set Mm,n (R)
of m-by-n matrices forms a vector space over R.
The remaining important operation on linear mappings is composition. As
shown in exercise 3.1.13, if

S : Rn Rm and T : Rp Rn

are linear then their composition

S T : Rp Rm

is linear as well. Suppose that S and T respectively have matrices

A Mm,n (R) and B Mn,p (R).

Then the composition S T has a matrix in Mm,p (R) that is naturally defined
as the matrix-by-matrix product

AB Mm,p (R),

the order of multiplication being chosen for consistency with the composition.
Under this specification,

(A times B)s jth column = (S T )(ej )

= S(T (ej ))
= A times (Bs jth column).

And A times (Bs jth column) is a matrix-by-vector multiplication, which

we know how to carry out: the result is a column vector whose ith entry for
i = 1, , m is the inner product of the ith row of A and the jth column
of B. In sum, the rule for matrix-by-matrix multiplication is as follows.
Definition 3.2.4 (Matrix Multiplication). Given two matrices,

A Mm,n (R) and B Mn,p (R),

such that A has as many columns as B has rows, their product,

AB Mm,p (R),

has for its (i, j)th entry (for any i {1, , m} and j {1, , p}) the inner
product of the ith row of A and the jth column of B. In symbols,

(AB)ij = hith row of A, jth column of Bi,

or, at the level of individual entries,

" n
#
X
If A = [aij ]mn and B = [bij ]np then AB = aik bkj .
k=1 mp
78 3 Linear Mappings and Their Matrices

Inevitably, matrix-by-matrix multiplication subsumes matrix-by-vector

multiplication, viewing vectors as one-column matrices. Also, once we have
the definition of matrix-by-matrix multiplication, we can observe that in com-
plement to the already-established rule that for any j {1, , n},

(A times B)s jth column equals A times (Bs jth column),

also, for any i {1, , m},

ith row of (A times B) equals (ith row of A) times B.

Indeed, both quantities in the previous display are the 1-by-p vector whose
jth entry is the inner product of the ith row of A and the jth column of B.
For example, consider the matrices

  1 2  
123 45
A= , B = 2 3 , C= ,
456 67
3 4

111   x
D = 0 1 1 , E= abc , F = y .
001 z

Some products among these (verify!) are

  8 9  
14 20 13 6
AB = , BC = 10 11 , AD = ,
32 47 4 9 15
12 13

6 9   ax bx cx
x + 2y + 3z
DB = 5 7 , AF = , F E = ay by cy ,
4x + 5y + 6z
3 4 az bz cz
EF = ax + by + cz.

Matrix multiplication is not commutative. Indeed, when the product AB

is defined, the product BA may not be, or it may be but have different dimen-
sions from AB; cf. EF and F E above. Even when A and B are both n-by-n,
so that AB and BA are likewise n-by-n, the products need not agree. For
example,
         
01 00 10 00 01 00
= , = .
00 10 00 10 00 01

Of particular interest is the matrix associated to the identity mapping,

id : Rn Rn , id(x) = x.

Naturally, this matrix is denoted the identity matrix; it is written In . Since

idi (ej ) = ij ,
 3.2 Operations on Matrices 79

1 0 0
0 1 0

In = [ij ]nn = . .. .. .
.. . .
0 0 1
Although matrix multiplication fails to commute, it does have the following
properties.

Proposition 3.2.5 (Properties of Matrix Multiplication). Matrix mul-

tiplication is associative,

A(BC) = (AB)C for A Mm,n (R), B Mn,p (R), C Mp,q (R).

Matrix multiplication distributes over matrix addition,

A(B + C) = AB + AC for A Mm,n (R), B, C Mn,p (R),

(A + B)C = AC + BC for A, B Mm,n (R), C Mn,p (R).

Scalar multiplication passes through matrix multiplication,

(AB) = (A)B = A(B) for R, A Mm,n (R), B Mn,p (R).

The identity matrix is a multiplicative identity,

Im A = A = AIn for A Mm,n (R).

Proof. The right way to show these is intrinsic, by remembering that addition,
scalar multiplication, and multiplication of matrices precisely mirror addition,
scalar multiplication, and composition of mappings. For example, if A, B, C
are the matrices of the linear mappings S L(Rn , Rm ), T L(Rp , Rn ), and
U L(Rq , Rp ), then (AB)C and A(BC) are the matrices of (S T ) U and
S (T U ). But these two mappings are the same since the composition of
mappings (mappings in general, not only linear mappings) is associative. To
verify the associativity, we cite the definition of four different binary compo-
sitions to show that the ternary composition is independent of parentheses,
as follows. For any x Rq ,

((S T ) U )(x) = (S T )(U (x)) by definition of R U where R = S T

= S(T (U (x))) by definition of S T
= S((T U )(x)) by definition of T U
= (S (T U ))(x) by definition of S V where V = T U .

So indeed ((S T ) U ) = (S (T U )), and consequently (AB)C = A(BC).

Alternatively, one can verify the equalities elementwise by manipulating
sums. Adopting the notation Mij for the (i, j)th entry of a matrix M ,
80 3 Linear Mappings and Their Matrices
n
X n
X p
X p
n X
X
(A(BC))ij = Aik (BC)kj = Aik Bk Cj = Aik Bk Cj
k=1 k=1 =1 k=1 =1
Xp Xn p
X
= Aik Bk Cj = (AB)i Cj = ((AB)C)ij .
=1 k=1 =1

The steps here are not explained in detail because the author finds this method
as grim as it is gratuitous: the coordinates work because they must, but their
presence only clutters the argument. The other equalities are similar.

Composing mappings is most interesting when all the mappings in ques-
tion take a set S back to the same set S, for the set of such mappings is
closed under composition. In particular, L(Rn , Rn ) is closed under compo-
sition. The corresponding statement about matrices is that Mn (R) is closed
under multiplication.

Exercises

3.2.1. Justify Definition 3.2.2 of scalar multiplication of matrices.

3.2.2. Carry out the matrix multiplications
e
0 10 0
  
ab d b   a 1 b1 0 01 0
, x 1 x 2 x 3 a 2 b2 , (e = 2, 3, 4),
cd c a 0 00 1
a 3 b3
0 00 0

11 1 10 0 100 111
0 1 1 1 1 0 , 1 1 0 0 1 1 .
00 1 11 1 111 001

3.2.3. Prove more of Proposition 3.2.5, that A(B + C) = AB + AC, (A)B =

A(B) and Im A = A for suitable matrices A, B, C and any scalar .
3.2.4. (If you have not yet worked exercise 3.1.14 then do so before working
this exercise.) Let A = [aij ] Mm,n (R) be the matrix of S L(Rn , Rm ). Its
transpose AT Mn,m (R) is the matrix of the transpose mapping S T . Since
S and S T act respectively as multiplication by A and AT , the characterizing
property of S T from exercise 3.1.14 gives

hx, AT yi = hAx, yi for all x Rn and y Rm .

Make specific choices of x and y to show that the transpose AT Mn,m (R) is
obtained by flipping A about its NorthwestSoutheast diagonal; that is, show
that the (i, j)th entry of AT is aji . It follows that the rows of AT are the
columns of A and the columns of AT are the rows of A.
(Similarly, let B Mn,p (R) be the matrix of T L(Rp , Rn ), so that B T
is the matrix of T T . Since matrix multiplication is compatible with linear
 3.2 Operations on Matrices 81

mapping composition, we know immediately from exercise 3.1.14(b), with no

reference to the concrete description of the matrix transposes AT and B T in
terms of the original matrices A and B, that the transpose of the product is
the product of the transposes in reverse order,

(AB)T = B T AT for all A Mm,n (R) and B Mn,p (R).

That is, by characterizing the transpose mapping in exercise 3.1.14, we eas-

ily derived the construction of the transpose matrix here and obtained the
formula for the product of transpose matrices with no reference to their con-
struction.)

3.2.5. The trace of a square matrix A Mn (R) is the sum of its diagonal
elements,
n
X
tr(A) = aii .
i=1

Show that
tr(AB) = tr(BA), A, B Mn (R).
(This exercise may entail double subscripts.)

3.2.6. For any matrix A Mm,n (R) and column vector a Rm define the
affine mapping (cf. exercise 3.1.15)

AffA,a : Rn Rm

by the rule AffA,a (x) = Ax + a for all x Rn , viewing x as a column vector.

(a) Explain why every affine mapping from Rn to Rm takes this form.
(b) Given such A and a, define the matrix A Mm+1,n+1 (R) to be
 
A a
A = .
0n 1

Show that for all x Rn ,

   
x AffA,a (x)
A = .
1 1

Thus, affine mappings, like linear mappings, behave as matrix-by-vector mul-

tiplications but where the vectors are the usual input and output vectors
augmented with an extra 1 at the bottom.
(c) If the affine mapping AffB,b : Rp Rn determined by B Mn,p (R)
and b Rn has matrix  
B b
B =
0p 1
show that AffA,a AffB,b : Rp Rm has matrix A B . Thus, matrix multi-
plication is compatible with composition of affine mappings.
82 3 Linear Mappings and Their Matrices

3.2.7. The exponential of any square matrix A is the infinite matrix sum
1 2 1
eA = I + A + A + A3 + .
2! 3!
Compute the exponentials of the following matrices:

1 0 0
  10
1 0 1 0
A = [], A= , A = 0 1 , A=
0
.
0 0 1
00
0 0 0

What is the general pattern?

3.2.8. Let a, b, d be real numbers with ad = 1. Show that

    
ab 1 ab a0
= .
0d 0 1 0d

Let a, b, c, d be real numbers with c 6= 0 and ad bc = 1. Show that

     1     1 
ab 1 ac1 c 0 0 1 1c d
= .
cd 0 1 0 c 1 0 0 1
 
Thus this exercise has shown that all matrices ac db with ad bc = 1 can
1   0 
be expressed in terms of matrices 0 1 and matrices 0 1 and the matrix
 0 1 
1 0 .

3.3 The Inverse of a Linear Mapping

Given a linear mapping S : Rn Rm , does it have an inverse? That is, is
there a mapping T : Rm Rn such that

S T = idm and T S = idn ?

If so, what is T ?
The symmetry of the previous display shows that if T is an inverse of S
then S is an inverse of T in turn. Also, the inverse T , if it exists, must be
unique, for if T : Rm Rn also inverts S then

T = T idm = T (S T ) = (T S) T = idn T = T.

Thus T can unambiguously be denoted S 1 . In fact, this argument has shown

a little bit more than claimed: If T inverts S from the left and T inverts S
from the right then T = T . On the other hand, the argument does not show
that if T inverts S from the left then T also inverts S from the rightthis is
not true.
 3.3 The Inverse of a Linear Mapping 83

If the inverse T exists then it too is linear. To see this, note that the
elementwise description of S and T being inverses of one another is that every
y Rm takes the form y = S(x) for some x Rn , every x Rn takes the
form x = T (y) for some y Rm , and

for all x Rn and y Rm , y = S(x) x = T (y).

Now compute that for any y1 , y2 Rm ,

T (y1 + y2 ) = T (S(x1 ) + S(x2 )) for some x1 , x2 Rn

= T (S(x1 + x2 )) since S is linear
= x1 + x2 since T inverts S
= T (y1 ) + T (y2 ) since y1 = S(x1 ) and y2 = S(x2 ).

Thus T satisfies (3.1). The argument that T satisfies (3.2) is similar.

Since matrices are more explicit than linear mappings, we replace the
question at the beginning of this section with its matrix counterpart: Given a
matrix A Mm,n (R), does it have an inverse matrix, a matrix B Mn,m (R)
such that
AB = Im and BA = In ?
As above, if the inverse exists then it is unique, and so it can be denoted A1 .
The first observation to make is that if the equation Ax = 0m has a
nonzero solution x Rn then A has no inverse. Indeed, also A0n = 0m , so an
inverse A1 would have to take 0m both to x and to 0n , which is impossible.
And so we are led to a subordinate question: When does the matrix equation

Ax = 0m

have nonzero solutions x Rn ?

For example, let A be the 5-by-6 matrix

5 1 17 26 1 55
3 1 13 20 0 28

A= 2 1 3 5 0 3.
2 0 4 6 0 10
5 0 10 15 1 42

If there is a nonzero x R6 such that Ax = 05 then A is not invertible.

Left multiplication by certain special matrices will simplify the matrix A.

Definition 3.3.1 (Elementary Matrices). There are three kinds of ele-

mentary matrices. For any i, j {1, , m} (i 6= j) and any a R, the
m-by-m (i; j, a) recombine matrix is
84 3 Linear Mappings and Their Matrices

1
..
.

1 a

. ..
Ri;j,a = .

1

..
.
1

(Here the a sits in the (i, j)th position, the diagonal entries are 1 and all other
entries are 0. The a is above the diagonal as shown only when i < j, otherwise
it is below.)
For any i {1, , m} and any nonzero a R, the m-by-m (i, a) scale
matrix is
1
..
.

1


Si,a = a .

1

.
. .
1
(Here the a sits in the ith diagonal position, all other diagonal entries are 1
and all other entries are 0.)
For any i, j {1, , m} (i 6= j), the m-by-m (i; j) transposition ma-
trix is
1
..
.

1

0 1

1

. ..
Ti;j = .

1

1 0

1

..
.
1
(Here the diagonal entries are 1 except the ith and jth, the (i, j)th and (j, i)th
entries are 1, all other entries are 0.)
The plan is to study the equation Ax = 0m by using these elementary
matrices to reduce A to a nicer matrix E and then solve the equation Ex = 0m
instead. Thus we are developing an algorithm rather than a formula. The next
 3.3 The Inverse of a Linear Mapping 85

proposition describes the effect that the elementary matrices produce by left
multiplication.

Proposition 3.3.2 (Effects of the Elementary Matrices). Let M be an

m-by-n matrix; call its rows rk . Then
(1) The m-by-n matrix Ri;j,a M has the same rows as M except that its ith
row is ri + arj ;
(2) The m-by-n matrix Si,a M has the same rows as M except that its ith row
is ari ;
(3) The m-by-n matrix Ti;j M has the same rows as M except that its ith row
is rj and its jth row is ri .

Proof. (1) As observed immediately after Definition 3.2.4, each row of Ri;j,a M
equals the corresponding row of Ri;j,a times M . For any row index k 6= i, the
only nonzero entry of the row is a 1 in the kth position, so the product of the
row and M simply picks of the kth row of M . Similarly, the ith row of Ri;j,a
has a 1 in the ith position and an a in the jth, so the row times M equals the
ith row of M plus a times the jth row of M .
(2) and (3) are similar, left as exercise 3.3.2.

To get a better sense of why the statements in the proposition are true, it
may be helpful to do the calculations explicitly with some moderately sized
matrices. But then, the point of the proposition is that once one believes it, left
multiplication by elementary matrices no longer requires actual calculation.
Instead, one simply carries out the appropriate row operations. For example,
   
123 13 17 21
R1;2,3 = ,
456 4 5 6

because R1;2,3 adds 3 times the second row to the first. The slogan here is:

Elementary matrix TIMES is row operation ON.

Thus we use the elementary matrices to reason about this material, but for
hand calculation we simply carry out the row operations.
The next result is that performing row operations on A doesnt change the
set of solutions x to the equation Ax = 0m .

Lemma 3.3.3 (Invertibility of Products of the Elementary Matri-

ces). Products of elementary matrices are invertible. More specifically:
(1) The elementary matrices are invertible by other elementary matrices.
Specifically,

(Ri;j,a )1 = Ri;j,a , (Si,a )1 = Si,a1 , (Ti;j )1 = Ti;j .

(2) If the m-by-m matrices M and N are invertible by M 1 and N 1 , then the
product matrix M N is invertible by N 1 M 1 . (Note the order reversal.)
86 3 Linear Mappings and Their Matrices

(3) Any product of elementary matrices is invertible by another such product,

specifically the product of the inverses of the original matrices, but taken
in reverse order.

Proof. (1) To prove that Ri;j,a Ri;j,a = Im , note that Ri;j,a is the identity
matrix Im with a times its jth row added to its ith row, and multiplying this
from the left by Ri;j,a subtracts back off a times the jth row, restoring Im .
The proof that Ri;j,a Ri;j,a = Im is either done similarly or by citing the
proof just given with a replaced by a. The rest of (1) is similar.
(2) Compute:

(M N )(N 1 M 1 ) = M (N N 1 )M 1 = M Im M 1 = M M 1 = Im .

Similarly for (N 1 M 1 )(M N ) = Im .

(3) This is immediate from (1) and (2).

Proposition 3.3.4 (Persistence of Solution). Let A be an m-by-n matrix

and let P be a product of m-by-m elementary matrices. Then the equations

Ax = 0m and (P A)x = 0m

are satisfied by the same vectors x in Rn .

Proof. Suppose that the vector x Rn satisfies the left equation, Ax = 0m .

Then
(P A)x = P (Ax) = P 0m = 0m .
Conversely, suppose that x satisfies (P A)x = 0m . Lemma 3.3.3 says that P
has an inverse P 1 , so

Ax = Im Ax = (P 1 P )Ax = P 1 (P A)x = P 1 0m = 0m .

The machinery is in place to solve the equation Ax = 05 where as before,

5 1 17 26 1 55
3 1 13 20 0 28

A= 2 1 3 50 3
.
2 0 4 6 0 10
5 0 10 15 1 42

Scale As fourth row by 1/2; transpose As first and fourth rows:

1 0 2 30 5
3 1 13 20 0 28
call
T1;4 S4,1/2 A =
2 1 3 50 3 = B.
5 1 17 26 1 55
5 0 10 15 1 42
 3.3 The Inverse of a Linear Mapping 87

Note that B has a 1 as the leftmost entry of its first row. Recombine various
multiples of the first row with the other rows to put 0s beneath the leading 1
of the first row:

1 0 2 30 5
0 1 7 11 0 13
call
R5;1,5 R4;1,5 R3;1,2 R2;1,3 B =
0 1 7 11 0 13 = C.
0 1 7 11 1 30
0 0 0 0 1 17

Recombine various multiples of the second row with the others to put 0s
above and below its leftmost nonzero entry; scale the second row to make its
leading nonzero entry a 1:

102 30 5
0 1 7 11 0 13
call
S2,1 R4;2,1 R3;2,1 C =
0 0 0 0 0 0 = D.
0 0 0 0 1 17
0 0 0 0 1 17

Transpose the third and fifth rows; put 0s above and below the leading 1 in
the third row:
102 30 5
0 1 7 11 0 13
call
R4;3,1 T3;5 D =
0 0 0 0 1 17 = E.
0 0 0 0 0 0
000 00 0
Matrix E is a prime example of a so-called echelon matrix. (The term will be
defined precisely in a moment.) Its virtue is that the equation Ex = 05 is now
easy to solve. This equation expands out to

x1
102 30 5 x1 + 2x3 + 3x4 + 5x6 0
x
0 1 7 11 0 13 2 x2 + 7x3 + 11x4 + 13x6 0
x3
Ex =
0 0 0 0 1 17 x4 = x5 + 17x6
= 0.
0 0 0 0 0 0 0 0
x5
000 00 0 0 0
x6

Matching the components in the last equality gives

x1 = 2x3 3x4 5x6

x2 = 7x3 11x4 13x6
x5 = 17x6 .

Thus, x3 , x4 and x6 are free variables that may take any values we wish, but
then x1 , x2 and x5 are determined from these equations. For example, setting
x3 = 5, x4 = 3, x6 = 2 gives the solution x = (9, 24, 5, 3, 34, 2).
88 3 Linear Mappings and Their Matrices

Definition 3.3.5 (Echelon Matrix). A matrix E is called echelon if it has

the form
0 0 1 0 0
1 0

E= 1
.

Here the s are arbitrary entries and all entries below the stairway are 0.
Thus each rows first nonzero entry is a 1, each rows leading 1 is farther
right than that of the row above it, each leading 1 has a column of 0s above
it, and any rows of 0s are at the bottom.

Note that the identity matrix I is a special case of an echelon matrix.

The algorithm for reducing any matrix A to echelon form by row operations
should be fairly clear from the previous example. The interested reader may
want to codify it more formally, perhaps in the form of a computer program.
Although different sequences of row operations may reduce A to echelon form,
the resulting echelon matrix E will always be the same. This result can be
proved by induction on the number of columns of A, and its proof is in many
linear algebra books.

Theorem 3.3.6 (Matrices Reduce to Echelon Form). Every matrix A

row reduces to a unique echelon matrix E.

In an echelon matrix E, the columns with leading 1s are called new

columns, and all others are old columns. The recipe for solving the equation
Ex = 0m is then
1. Freely choose the entries in x that correspond to the old columns of E.
2. Then each nonzero row of E will determine the entry of x corresponding
to its leading 1 (which sits in a new column). This entry will be a linear
combination of the free entries to its right.
Lets return to the problem of determining whether A Mm,n (R) is in-
vertible. The idea was to see if the equation Ax = 0m has any nonzero solu-
tions x, in which case A is not invertible. Equivalently, we may check whether
Ex = 0m has nonzero solutions, where E is the echelon matrix to which A
row reduces. The recipe for solving Ex = 0m shows that there are nonzero
solutions unless all of the columns are new.
If A Mm,n (R) has more columns than rows then its echelon matrix E
must have old columns. Indeed, each new column comes from the leading 1 in
a distinct row, so

new columns of E rows of E < columns of E,

showing that not all the columns are new. Thus A is not invertible when
m < n. On the other hand, if A Mm,n (R) has more rows than columns and
 3.3 The Inverse of a Linear Mapping 89

it has an inverse matrix A1 Mn,m (R), then A1 in turn has inverse A, but
this is impossible since A1 has more columns than rows. Thus A is also not
invertible when m > n.
The remaining case is that A is square. The only square echelon matrix
with all new columns is I, the identity matrix (exercise 3.3.10). Thus, unless
As echelon matrix is I, A is not invertible. On the other hand, if As echelon
matrix is I, then P A = I for some product P of elementary matrices. Multiply
from the left by P 1 to get A = P 1 ; this is invertible by P , giving A1 = P .
Summarizing,
Theorem 3.3.7 (Invertibility and Echelon Form for Matrices). A non-
square matrix A is never invertible. A square matrix A is invertible if and only
if its echelon form is the identity matrix.
When A is square, the discussion above gives an algorithm that simulta-
neously checks whether it is invertible and finds its inverse when it is.
Proposition 3.3.8 (Matrix Inversion Algorithm). Given A Mn (R),
set up the matrix  
B = A | In
in Mn,2n (R). Carry out row operations on this matrix to reduce the left side
to echelon form. If the left side reduces to In then A is invertible and the right
side is A1 . If the left side doesnt reduce to In then A is not invertible.
The algorithm works because if B is left multiplied by a product P of
elementary matrices, the result is
 
PB = PA | P .
As discussed, P A = In exactly when P = A1 .
For example, the calculation

1 1 0 100 100 111
R1;2,1 R2;3,1 0 1 1 0 1 0 = 0 1 0 0 1 1
0 0 1 001 001 001
shows that 1
1 1 0 111
0 1 1 = 0 1 1 ,
0 0 1 001
and one readily checks that the claimed inverse really works. Since arithmetic
by hand is so error-prone a process, one always should confirm ones answer
from the matrix inversion algorithm.
We now have an algorithmic answer to the question at the beginning of
the section.
Theorem 3.3.9 (Echelon Criterion for Invertibility). The linear map-
ping S : Rn Rm is invertible only when m = n and its matrix A has
echelon matrix In , in which case its inverse S 1 is the linear mapping with
matrix A1 .
90 3 Linear Mappings and Their Matrices

Exercises

3.3.1. Write down the following 3-by-3 elementary matrices and their inverses:
R3;2, , S3,3 , T3;2 , T2;3 .
3.3.2. Finish the proof of Proposition 3.3.2.
h1 2i
3.3.3. Let A = 3 4 . Evaluate the following products without actually mul-
56
tiplying matrices: R3;2, A, S3,3 A, T3;2 A, T2;3 A.
3.3.4. Finish the proof of Lemma 3.3.3, part (1).
3.3.5. What is the effect of right multiplying the m-by-n matrix M by an
n-by-n matrix Ri;j,a ? By Si,a ? By T i; j?
3.3.6. Recall the transpose of a matrix M (cf. exercise 3.2.4), denoted M T .
T T T
Prove: Ri;j,a = Rj;i,a ; Si,a = Si,a ; Ti;j = Ti;j . Use these results and the
T T T
formula (AB) = B A to redo the previous problem.
3.3.7. Are the following matrices echelon? For each matrix M , solve the equa-
tion M x = 0.

00
103     1000 011
0001 1100
1 0
0 1 1 , , ,
0 1 ,
0 1 1 0 , 1 0 3 .
0000 0011
001 0010 000
00

3.3.8. For each matrix A solve the equation Ax = 0.

1 1 4 2 1 3 2 3 1 2
1 3 8, 1 4 0 1, 2 1 1.
125 2 6 1 5 1 3 0

3.3.9. Balance the chemical equation

Ca + H3 PO4 Ca3 P2 O8 + H2 .

3.3.10. Prove by induction that the only square echelon matrix with all new
columns is the identity matrix.
3.3.11. Are the following matrices invertible? Find the inverse when possible,
and then check your answer.
1 1
1 1 1 2 5 1 1 2 3
2 0 1 , 4 1 2 , 1 1 1 .
2 3 4
1 1 1
3 01 6 4 1 3 4 5

3.3.12. The matrix A is called lower triangular if aij = 0 whenever i < j.

If A is a lower triangular square matrix with all diagonal entries equal to 1,
show that A is invertible and A1 takes the same form.
 3.4 Inhomogeneous Linear Equations 91

3.3.13. This exercise refers back to the GramSchmidt exercise in chapter 2.

That exercise expresses the relation between the vectors {xj } and the vectors
{xj } formally as x = Ax where x is a column vector whose entries are the
vectors x1 , , xn , x is the corresponding column vector of xj s, and A is an
n-by-n lower triangular matrix.
Show that each xj has the form

xj = aj1 x1 + aj2 x2 + + aj,j1 xj1 + xj ,

and thus any linear combination of the original {xj } is also a linear combina-
tion of the new {xj }.

3.4 Inhomogeneous Linear Equations

The question of whether a linear mapping T is invertible led to solving the

linear equation Ax = 0, where A was the matrix of T . Such a linear equa-
tion, with right side 0, is called homogeneous. An inhomogeneous linear
equation has nonzero right side,

Ax = b, A Mm,n (R), x Rn , b Rm , b 6= 0.

The methods of the homogeneous case apply here too. If P is a product of m-

by-m elementary matrices such that P A is echelon (call it E), then multiplying
the inhomogeneous equation from the left by P gives

Ex = P b,

and since P b is just a vector, the solutions to this can be read off as in the
homogeneous case. There may not always be solutions, however.

Exercises

3.4.1. Solve the inhomogeneous equations

1 1 2 1 1 2 1 2 1
2 0 2 x = 1 , 1 1 1 1 x = 2 .
1 3 4 2 1 7 5 1 3

3.4.2. For what values b1 , b2 , b3 does the equation

3 1 2 b1
2 1 1 x = b2
1 3 0 b3

have a solution?
92 3 Linear Mappings and Their Matrices

3.4.3. A parent has a son and a daughter. The parent is four times as old as
the daughter, the daughter is four years older than the son. In three years the
parent will be five times as old as the son. How old are the parent, daughter
and son?

3.4.4. Show that to solve an inhomogeneous linear equation, one may solve a
homogeneous system in one more variable and then restrict to solutions where
the last variable is equal to 1.

3.5 The Determinant: Characterizing Properties and

Their Consequences

In this section all matrices are square, n-by-n. The goal is to define a function
that takes such a matrix, with its n2 entries, and returns a single number.
The putative function is called the determinant,

det : Mn (R) R.

For any square matrix A Mn (R), the scalar det(A) should contain as much
algebraic and geometric information about the matrix as possible. Not sur-
prisingly, so informative a function is complicated to encode.
This context nicely demonstrates a pedagogical principle already men-
tioned in section 3.1: characterizing a mathematical object illuminates its
construction and its use. Rather than beginning with a definition of the de-
terminant, we will stipulate a few natural behaviors for it, and then we will
eventually see that
there is a function with these behaviors (existence),
there is only one such function (uniqueness), and, most importantly,
these behaviors, rather than the definition, further show how the function
works (consequences).
We could start at the first bullet (existence) and proceed from the construction
of the determinant to its properties, but when a construction is complicated
(as the determinants construction is) it fails to communicate intent, and
pulling it out of thin air as the starting point of a long discussion is an obstacle
to understanding. A few naturally gifted readers will see what the unexplained
idea really is, enabling them to skim the ensuing technicalities and go on to
start using the determinant effectively; some other tough-minded readers can
work through the machinery and then see its operational consequences; but
it is all too easy for the rest of us to be defeated by disorienting detail-fatigue
before the presentation gets to the consequential points and provides any
energizing clarity.
Another option would be to start at the second bullet (uniqueness), letting
the desired properties of the determinant guide our construction of it. This
 3.5 The Determinant: Characterizing Properties and Their Consequences 93

process wouldnt be as alienating as starting with existence, but deriving the

determinants necessary construction has only limited benefit since we intend
to use the construction as little as possible. Working through the derivation
would still squander our energy on the internal mechanisms of the determi-
nant before getting to its behavior, when its behavior is what truly lets us
understand it. We first want to learn to use the determinant easily and art-
fully. Doing so will make its internals feel of secondary importance, as they
should.
The upshot is that in this section we will pursue the third bullet (conse-
quences), and then the next section will proceed to the second bullet (unique-
ness) and finally the first one (existence).
Instead of viewing the determinant only as a function of a matrix A
Mn (R) with n2 scalar entries, view it also as a function of As n rows, each
of which is an n-vector. If A has rows r1 , , rn , write det(r1 , , rn ) for
det(A). Thus, det is now being interpreted as a function of n vectors, i.e., the
domain of det is n copies of Rn ,

det : Rn Rn R.

The advantage of this view is that now we can impose conditions on the deter-
minant, using language already at our disposal in a natural way. Specifically,
we make three requirements:
(1) The determinant is multilinear, meaning that it is linear as a function
of each of its vector variables when the rest are held fixed. That is, for
any vectors r1 , , rk , rk , , rn and any scalar ,

det(r1 , , rk + rk , , rn ) = det(r1 , , rk , , rn )
+ det(r1 , , rk , , rn ).

(2) The determinant is skew-symmetric as a function of its vector variables,

meaning that exchanging any two inputs negates the determinant,

det(r1 , , rj , , ri , , rn ) = det(r1 , , ri , , rj , , rn ).

(Here i 6= j.) Consequently, the determinant is also alternating, meaning

that if two inputs ri and rj are equal then det(r1 , , rn ) = 0.
(3) The determinant is normalized, meaning that the standard basis has
determinant 1,
det(e1 , , en ) = 1.
Condition (1) does not say that det(A+A ) = det(A)+det(A ) for scalars
and square matrices A, A . Especially, the determinant is not additive,

det(A + B) is in general not det(A) + det(B). (3.5)

What the condition does say is that if all rows but one of a square matrix are
held fixed, then the determinant of the matrix varies linearly as a function
94 3 Linear Mappings and Their Matrices

of the one row. By induction, an equivalent statement of multilinearity is the

more cluttered
X X
det(r1 , , i rk,i , , rn ) = i det(r1 , , rk,i , , rn ),
i i

but to keep the notation manageable we work with the simpler version.
We will prove the following theorem in the next section.

Theorem 3.5.1 (Existence and Uniqueness of the Determinant). One,

and only one, multilinear skew-symmetric normalized function from the n-fold
product of Rn to R exists. This function is the determinant,

det : Rn Rn R.

Furthermore, all multilinear skew-symmetric functions from the n-fold product

of Rn to R are scalar multiples of of the determinant. That is, any multilinear
skew-symmetric function : Rn Rn R is

= c det where c = (e1 , , en ).

In more structural language, Theorem 3.5.1 says that the multilinear skew-
symmetric functions from the n-fold product of Rn to R form a 1-dimensional
vector space over R, and {det} is a basis.
The reader may object that even if the conditions of multilinearity, skew-
symmetry, and normalization are grammatically natural, they are concep-
tually opaque. Indeed they reflect considerable hindsight, since the idea of
a determinant originally emerged from explicit calculations. But again, the
payoff is that characterizing the determinant rather than constructing it illu-
minates its many useful properties. The rest of the section can be viewed as
an amplification of this idea.
For one quick application of the existence of the determinant, consider the
standard basis of Rn taken in order,

(e1 , , en ).

Suppose that some succession of m pair-exchanges of the vectors in this or-

dered n-tuple has no net effect, i.e., after the m pair-exchanges, the vectors are
back in their original order. By skew-symmetry each pair-exchange negates
the determinant, and so after all m pair-exchanges the net result is

(1)m det(e1 , , en ) = det(e1 , , en ).

Since det is normalized, it follows that (1)m = 1, i.e., m is even. That is,
no odd number of pair-exchanges can leave an ordered n-tuple in its initial
order. Consequently, if two different sequences of pair-exchanges have the
same net effect then their lengths are both odd or both eventhis is because
 3.5 The Determinant: Characterizing Properties and Their Consequences 95

running one sequence forwards and then the other back has no net effect and
hence comes to an even number of moves. In other words, although a net
rearrangement of an n-tuple does not determine a unique succession of pair-
exchanges to bring it about, or even a unique number of such exchanges, it does
determine the parity of any such number: the net rearrangement requires an
odd number of exchanges, or it requires an even number. (For reasons related
to this, an old puzzle involving fifteen squares that slide in a 4-by-4 grid can
be made unsolvable by popping two pieces out and exchanging them.)
The fact that the parity of a rearrangement is well defined may be easy to
believe, perhaps so easy that the need for a proof is hard to see, but a proof
really is required. The determinants skewness and normalization are so pow-
erful that they give the result essentially as an afterthought. See exercise 3.5.2
for an elementary proof that does not invoke the existence of the determinant.
To summarize clearly, with reference to the exercise:
Independently of the determinant, every rearrangement of n objects
has a well-defined parity, meaning that for any rearrangement of the
objects, either all sequences of pairwise exchanges that put the objects
back in order have even length or all such sequences have odd length.
Easy though it is to use the determinant to show that parity is well defined,
in the next section we will need the fact that parity is well defined to show
that the determinant is unique. Thus exercise 3.5.2 keeps us from arguing in
a circle.
The next result is a crucial property of the determinant.

Theorem 3.5.2 (The Determinant is Multiplicative). For all matrices

A, B Mn (R),
det(AB) = det(A) det(B).
In particular, if A is invertible then the determinant of the matrix inverse is
the scalar inverse of the determinant,

det(A1 ) = (det(A))1 .

Multilinearity says that the determinant behaves well additively and

scalar-multiplicatively as a function of each of n vectors, while (3.5) says that
the determinant does not behave well additively as a function of one matrix.
Theorem 3.5.2 says that the determinant behaves perfectly well multiplica-
tively as a function of one matrix. Also, the theorem tacitly says that if A is
invertible then det(A) is nonzero. Soon we will establish the converse as well.

Proof. Let B Mn (R) be fixed. Consider the function

: Mn (R) R, (A) = det(AB).

As a function of the rows of A, is the determinant of the rows of AB,

96 3 Linear Mappings and Their Matrices

: Rn Rn R, (r1 , , rn ) = det(r1 B, , rn B).

The function is multilinear and skew-symmetric. To show multilinearity,

compute (using the definition of in terms of det, properties of vectormatrix
algebra, the multilinearity of det, and the definition of again),

(r1 , , rk + rk , , rn ) = det(r1 B, , (rk + rk )B, , rn B)

= det(r1 B, , rk B + rk B, , rn B)
= det(r1 B, , rk B, , rn B)
+ det(r1 B, , rk B, , rn B)
= (r1 , , rk , , rn )
+ (r1 , , rk , , rn ).

To show skew-symmetry, take two distinct indices i, j {1, , n} and com-

pute similarly,

(r1 , , rj , , ri , , rn ) = det(r1 B, , rj B, , ri B, , rn B)
= det(r1 B, , ri B, , rj B, , rn B)
= (r1 , , ri , , rj , , rn ).

Also compute that

(e1 , , en ) = det(e1 B, , en B) = det(B).

It follows from Theorem 3.5.1 that (A) = det(B) det(A), and this is the
desired main result det(AB) = det(A) det(B) of the theorem. Finally, if A is
invertible then

det(A) det(A1 ) = det(AA1 ) = det(I) = 1.

That is, det(A1 ) = (det(A))1 . The proof is complete.

One consequence of the theorem is

det(A1 BA) = det(B), A, B Mn (R), A invertible.

And we note that the same result holds for the trace, introduced in exer-
cise 3.2.5, in consequence of that exercise,

tr(A1 BA) = tr(B), A, B Mn (R), A invertible.

More facts about the determinant are immediate consequences of its char-
acterizing properties.

Proposition 3.5.3 (Determinants of Elementary and Echelon Matri-

ces).
 3.5 The Determinant: Characterizing Properties and Their Consequences 97

(1) det(Ri;j,a ) = 1 for all i, j {1, , n} (i 6= j) and a R.

(2) det(Si,a ) = a for all i {1, , n} and nonzero a R.
(3) det(Ti;j ) = 1 for all i, j {1, , n} (i 6= j).
(4) If E is n-by-n echelon then
(
1 if E = I,
det(E) =
0 otherwise.

Proof. (1) Compute,

det(Ri;j,a ) = det(e1 , , ei + aej , , ej , , en )

= det(e1 , , ei , , ej , , en ) + a det(e1 , , ej , , ej , , en )
= 1 + a 0 = 1.

(2) and (3) are similar. For (4), if E = I then det(E) = 1 since the determinant
is normalized. Otherwise the bottom row of E is 0, and since a linear function
takes 0 to 0, it follows that det(E) = 0.

For one consequence of Theorem 3.5.2 and Proposition 3.5.3, recall that
every matrix A Mn (R) has a transpose matrix AT , obtained by flipping A
about its NorthwestSoutheast diagonal. The next theorem (whose proof is
exercise 3.5.4) says that all statements about the determinant as a function
of the rows of A also apply to the columns. This fact will be used without
comment from now on. In particular, det(A) is the unique multilinear skew-
symmetric normalized function of the columns of A.

Theorem 3.5.4 (Determinant and Transpose). For all A Mn (R),

det(AT ) = det(A).

We also give another useful consequence of the determinants characteriz-

ing properties. A type of matrix that has an easily calculable determinant is a
triangular matrix, meaning a matrix all of whose subdiagonal entries are 0
or all of whose superdiagonal entries are 0. (Lower triangular matrices have
already been introduced in exercise 3.3.12.) For example, the matrices

a11 a12 a13 a11 0 0
0 a22 a23 and a21 a22 0
0 0 a33 a31 a32 a33

are triangular.

Proposition 3.5.5 (Determinant of a Triangular Matrix). The deter-

minant of a triangular matrix is the product of its diagonal entries.

Proof. We may consider only upper triangular matrices since a lower trian-
gular matrix has an upper triangular matrix for its transpose. The 3-by-3
98 3 Linear Mappings and Their Matrices

case makes the general argument clear. The determinant of a 3-by-3 upper
triangular matrix A is
3
X 3
X 3
X
det A = det( a1i1 ei1 , a2i2 ei2 , a3i3 ei3 )
i1 =1 i2 =2 i3 =3

which, since the determinant is multilinear, is

X 3
3 X
3 X
det A = a1i1 a2i2 a3i3 det(ei1 , ei2 , ei3 ).
i1 =1 i2 =2 i3 =3

Since the summation-index i3 takes only the value 3, this is

3 X
X 3
det A = a1i1 a2i2 a33 det(ei1 , ei2 , e3 ),
i1 =1 i2 =2

and the terms with i1 = 3 or i2 = 3 vanish since the determinant is alternating,

so the determinant further simplifies to

2
X
det A = a1i1 a22 a33 det(ei1 , e2 , e3 ).
i1 =1

Now the term with i1 = 2 vanishes similarly, leaving

det A = a11 a22 a33 det(e1 , e2 , e3 ).

Finally, since the determinant is normalized we have

det A = a11 a22 a33 .

A far more important consequence of Theorem 3.5.2 and Proposition 3.5.3

is one of the main results of this chapter. Recall that any matrix A row reduces
as
R1 RN A = E
where the Rk are elementary, E is echelon, and A is invertible if and only if
E = I. Since the determinant is multiplicative,

det(R1 ) det(RN ) det(A) = det(E). (3.6)

But each det(Rk ) is nonzero, and det(E) is 1 if E = I and 0 otherwise, so

this gives the algebraic significance of the determinant:
 3.5 The Determinant: Characterizing Properties and Their Consequences 99

Theorem 3.5.6 (Linear Invertibility Theorem). The matrix A Mn (R)

is invertible if and only if det(A) 6= 0.

That is, the zeroness or nonzeroness of the determinant says whether the
matrix is invertible. Once the existence and uniqueness of the determinant
are established in the next section, we will continue to use the determinant
properties to interpret the magnitude and the sign of the determinant as well.
Not only does equation (3.6) prove the Linear Invertibility Theorem, but
furthermore it describes an algorithm for computing the determinant of any
square matrix A: reduce A to echelon form by recombines, scales, and trans-
positions; if the echelon form is I then det(A) is the reciprocal product of the
scaling factors times 1 raised to the number of transpositions, and if the
echelon form is not I then det(A) = 0.

Exercises

3.5.1. Consider a scalar-valued function of pairs of vectors,

ip : Rn Rn R,

satisfying the following three properties.

(1) The function is bilinear,

ip(x + x , y) = ip(x, y) + ip(x , y),

ip(x, y + y ) = ip(x, y) + ip(x, y )

for all , , , R and x, x , y, y Rn .

(2) The function is symmetric,

ip(x, y) = ip(y, x) for all x, y Rn .

(3) The function is normalized,

ip(ei , ej ) = ij for all i, j {1, , n}.

(The Kronecker delta ij was defined in section 2.2.)

Compute that this function, if it exists at all, must be the inner product.
On the other hand, we already know that the inner product has these three
properties, so this exercise has shown that it is characterized by them.

3.5.2. Let n 2. This exercise proves, without invoking the determinant,

that any succession of pair-exchanges of the ordered set

(1, 2, , n)

that has no net effect consists of an even number of exchanges.

100 3 Linear Mappings and Their Matrices

To see this, consider a shortest-possible succession of an odd number of

pair-exchanges having in total no net effect. Certainly it must involve at least
three exchanges. We want to show that it cant exist at all.
Let the notation
(i j) (where i 6= j)
stand for exchanging the elements in positions i and j. Then in particular,
the first two exchanges in the succession take the form
(i j)( ),
meaning to exchange the elements in positions i and j and then to exchange
the elements in another pair of positions. There are four cases,
(i j)(i j),
(i j)(i k), k
/ {i, j},
(i j)(j k), k / {i, j},
(i j)(k ), k,
/ {i, j}, k 6= .
The first case gives a shorter succession of an odd number of pair-exchanges
having in total no net-effect, contradiction. Show that the other three cases
can be rewritten in the form
( )(i )
where the first exchange does not involve the ith slot. Next we may apply
the same argument to the second and third exchanges, then to the third and
fourth, and so on. Eventually, either a contradiction arises from the first of
the four cases, or only the last pair-exchange involves the ith slot. Explain
why the second possibility is untenable, completing the argument.
3.5.3. Let f : Rn Rn R be a multilinear skew-symmetric function,
and let c be any real number. Show that the function cf is again multilinear
and skew-symmetric.
3.5.4. This exercise shows that for any square matrix A, det(AT ) = det(A).
(a) Show that det(RT ) = det(R) for any elementary matrix R. (That is,
R can be a recombine matrix, a scale matrix, or a transposition matrix.)
(b) If E is a square echelon matrix then either E = I or the bottom row
of E is 0. In either case, show that det(E T ) = det(E). (For the case E 6= I, we
know that E is not invertible. What is E T en , and what does this say about
the invertibility of E T ?)
(c) Use the formula (M N )T = N T M T , Theorem 3.5.2, and Proposi-
tion 3.5.3 to show that det(AT ) = det(A) for all A Mn (R).
3.5.5. The square matrix A is orthogonal if AT A = I. Show that if A is
orthogonal then det(A) = 1. Give an example with determinant 1.
3.5.6. The matrix A is skew-symmetric if AT = A. Show that if A is
n-by-n skew-symmetric with n odd then det(A) = 0.
 3.6 The Determinant: Uniqueness and Existence 101

3.6 The Determinant: Uniqueness and Existence

Recall that Theorem 3.5.1 asserts that exactly one multilinear skew-symmetric
normalized function from the n-fold product of Rn to R exists. That is, a
unique determinant exists.
We warm up for the proof of the theorem by using the three defining
conditions of the determinant to show that only one formula is possible for
the determinant of a general 2-by-2 matrix,
 
ab
A= .
cd

The first row of this matrix is

r1 = (a, b) = a(1, 0) + b(0, 1) = ae1 + be2 ,

and similarly its second row is r2 = ce1 + de2 . Thus, since we view the deter-
minant as a function of rows, its determinant must be

det(A) = det(r1 , r2 ) = det(ae1 + be2 , ce1 + de2 ).

Since det is linear in its first vector variable, this expands to

det(ae1 + be2 , ce1 + de2 ) = a det(e1 , ce1 + de2 ) + b det(e2 , ce1 + de2 ),

and since det is also linear in its second vector variable, this expands further,

a det(e1 , ce1 + de2 )+b det(e2 , ce1 + de2 )

= ac det(e1 , e1 ) + ad det(e1 , e2 )
+ bc det(e2 , e1 ) + bd det(e2 , e2 ).

But since det is skew-symmetric and alternating, the expanded expression

simplifies considerably,

ac det(e1 , e1 ) + ad det(e1 , e2 ) + bc det(e2 , e1 )+bd det(e2 , e2 )

= (ad bc) det(e1 , e2 ).

And finally, since det is normalized, we have found the only possible formula
for the determinant of a 2-by-2 matrix,

det(A) = ad bc.

All three characterizing properties of the determinant were required to derive

this formula. More subtly (and in this context trivially), the fact that this
is the only possible formula tacitly relies on the fact that any sequence of
exchanges of e1 and e2 that leaves them in order has even length, and any
such sequence that exchanges their order has odd length.
102 3 Linear Mappings and Their Matrices

(As a brief digression, the reader can use the matrix inversion algorithm
from section 3.3 to verify that the 2-by-2 matrix A is invertible if and only
if ad bc is nonzero, showing that the formula for the 2-by-2 determinant
arises from considerations of invertibility as well as from our three conditions.
However, the argument requires cases, e.g., a 6= 0 or a = 0, making this
approach uninviting for larger matrices.)
Returning to the main line of exposition, nothing here has yet shown that
a determinant function exists at all for 2-by-2 matrices. What it has shown is
that there is only one possibility,

det((a, b), (c, d)) = ad bc.

But now that we have the only possible formula, checking that indeed it
satisfies the desired properties is purely mechanical. For example, to verify
linearity in the first vector variable, compute

det((a, b) + (a , b ), (c, d)) = det((a + a , b + b ), (c, d))

= (a + a )d (b + b )c
= (ad bc) + (a d b c)
= det((a, b), (c, d)) + det((a , b ), (c, d)).

For skew-symmetry,

det((c, d), (a, b)) = cb da = (ad bc) = det((a, b), (c, d)).

And for normalization,

det(1, 0), (0, 1)) = 1 1 0 0 = 1.

We should also verify linearity in the second vector variable, but this no longer
requires the defining formula. Instead, since the formula is skew-symmetric
and is linear in the first variable,

det(r1 , r2 + r2 ) = det(r2 + r2 , r1 )


= det(r2 , r1 ) + det(r2 , r1 )


= det(r1 , r2 ) det(r1 , r2 )
= det(r1 , r2 ) + det(r1 , r2 ).

This little trick illustrates the value of thinking in general terms: a slight
modification, inserting a few occurrences of and replacing the subscripts
1 and 2 by i and j, shows that for any n, the three required conditions for the
determinant are redundantlinearity in one vector variable combines with
skew-symmetry to ensure linearity in all the vector variables.
One can similarly show that for a 1-by-1 matrix,

A = [a],
 3.6 The Determinant: Uniqueness and Existence 103

the only possible formula for its determinant is

det(A) = a,

and that indeed this works. The result is perhaps silly, but the exercise of
working through a piece of language and logic in the simplest instance can
help one to understand its more elaborate cases. As another exercise, the same
techniques show that the only possible formula for a 3-by-3 determinant is

ab c
det d e f = aek + bf g + cdh af h bdk ceg.
ghk
And again, this is the only possible formula because parity is well defined for
all rearrangements of e1 , e2 , and e3 . This formula is complicated enough that
we should rethink it in a more systematic way before verifying that it has
the desired properties. And we may as well generalize it to arbitrary n in the
process. Here are some observations about the 3-by-3 formula:
It is a sum of 3-fold products of matrix entries.
Every 3-fold product contains one element from each row of the matrix.
Every 3-fold product also contains one element from each column of the
matrix. So every 3-fold product arises from the positions of three rooks
that dont threaten each other on a 3-by-3 chessboard.
Every 3-fold product comes weighted by a + or a .
Similar observations apply to the 1-by-1 and 2-by-2 formulas. Our general
formula should encode them. Making it do so is partly a matter of notation,
but also an idea is needed to describe the appropriate distribution of plus
signs and minus signs among the terms. The following language provides all
of this.
Definition 3.6.1 (Permutation). A permutation of {1, 2, , n} is a vec-
tor
= ((1), (2), , (n))
whose entries are {1, 2, , n}, each appearing once, in any order. An inver-
sion in the permutation is a pair of entries with the larger one to the left.
The sign of the permutation , written (1) , is 1 raised to the number of
inversions in . The set of permutations of {1, , n} is denoted Sn .
Examples are the permutations = (1, 2, 3, , n), = (2, 1, 3, , n),
and = (5, 4, 3, 2, 1) (here n = 5). In these examples has no inversions,
has one, and has ten. Thus (1) = 1, (1) = 1, and (1) = 1. In
general, the sign of a permutation with an even number of inversions is 1 and
the sign of a permutation with an odd number of inversions is 1. There are
n! permutations of {1, 2, , n}; that is, the set Sn contains n! elements.
As advertised, permutations and their signs provide the notation for the
only possible n-by-n determinant formula. Consider any n vectors
104 3 Linear Mappings and Their Matrices
n
X n
X n
X
r1 = a1i ei , r2 = a2j ej , , rn = anp ep .
i=1 j=1 p=1

By multilinearity, any multilinear function (if it exists at all) must satisfy

Xn n
X n
X
(r1 , r2 , , rn ) = a1i ei , a2j ej , , anp ep
i=1 j=1 p=1

X n
n X n
X
= a1i a2j anp (ei , ej , , ep ).
i=1 j=1 p=1

If is also alternating then for any i, j, , p {1, , n},

(ei , ej , , ep ) = 0 if any two subscripts agree.
Thus we may sum only over permutations,
X
(r1 , r2 , , rn ) = a1i a2j anp det(ei , ej , , ep ).
(i,j, ,p)Sn

Consider any permutation = (i, j, , p). Suppose that contains an in-

version, i.e., two elements are out of order. Then necessarily two elements in
adjacent slots are out of order. (For example, if i > p then either i > j, giving
adjacent elements out of order as desired; or j > i > p, so that j and p are
an out of order pair in closer slots than i and p, and so on.) If a permutation
contains any inversions, then exchanging a suitable adjacent pair decreases
the number of inversions by one, negating the sign of the permutation, while
exchanging the corresponding two input vectors negates the determinant. Re-
peating this process until the permutation has no remaining inversions shows
that
(ei , ej , , ep ) = (1) (e1 , e2 , , en ).
That is, a possible formula for a multilinear skew-symmetric function is
X
(r1 , r2 , , rn ) = (1) a1i a2j anp c
=(i,j, ,p)

where
c = (e1 , , en ).
Especially, a possible formula for a multilinear skew-symmetric normalized
function is
X
det(r1 , r2 , , rn ) = (1) a1i a2j anp .
=(i,j, ,p)

And as we have discussed twice already in this section, the previous display
gives the unique possible formula for a multilinear skew-symmetric normalized
function because any method of rearranging (ei , ej , , ep ) into order must
produce the same factor (1) .
 3.6 The Determinant: Uniqueness and Existence 105

Definition 3.6.2 (Determinant). The determinant function,

det : Mn (R) R,

is defined as follows. For any A Mn (R) with entries (aij ),

X
det(A) = (1) a1(1) a2(2) an(n) .
Sn

The formula in the definition is is indeed the formula computed a moment

ago, since for any permutation = (i, j, , p) Sn we have (1) = i,
(2) = j, , (n) = p.
As an exercise to clarify the formula, we use it to reproduce the 3-by-3
determinant. Each permutation in S3 determines a rook-placement, and the
sign of the permutation is the parity of the number of NortheastSouthwest
segments joining any two of its rooks. For example, the permutation (2, 3, 1)
specifies that the rooks in the top, middle, and bottom rows are respectively
in columns 2, 3, and 1, and the sign is positive since there are two Northeast
Southwest segments. (See figure 3.9.) The following table lists each permuta-
tion in S3 followed by the corresponding term in the determinant formula. For
each permutation, the term is its sign times the product of the three matrix
entries where its rooks are placed.

(1) a1(1) a2(2) a3(3)

(1, 2, 3) aek
(1, 3, 2) af h
(2, 1, 3) bdk
(2, 3, 1) bf g
(3, 1, 2) cdh
(3, 2, 1) ceg

The sum of the right column entries is the anticipated formula from before,

ab c
det d e f = aek + bf g + cdh af h bdk ceg.
ghk

The same procedure reproduces the 2-by-2 determinant as well,

 
ab
det = ad bc,
cd

and even the silly 1-by-1 formula det[a] = a. The 2-by-2 and 3-by-3 cases
are worth memorizing. They can be visualized as adding the products along
NorthwestSoutheast diagonals of the matrix and then subtracting the prod-
ucts along SouthwestNortheast diagonals, where the word diagonal connotes
106 3 Linear Mappings and Their Matrices

Figure 3.9. The rook placement for (2, 3, 1), showing the two inversions

+ + +

Figure 3.10. The 3-by-3 determinant

wraparound in the 3-by-3 case. (See figure 3.10.) But be aware that this
pattern of the determinant as the NorthwestSoutheast diagonals minus the
SouthwestNortheast diagonals is valid only for n = 2 and n = 3.
We have completed the program of the second bullet at the beginning of the
previous section, finding the only possible formula (the one in Definition 3.6.2)
that could satisfy the three desired determinant properties. We dont yet know
that it does, just that it is the only formula that could. That is, we have
now proved the uniqueness but not yet the existence of the determinant in
Theorem 3.5.1.
The first bullet tells us to prove the existence by verifying that the com-
puted determinant formula indeed does satisfy the three stipulated determi-
nant properties. Similarly to the 2-by-2 case, this is a mechanical exercise.
The impediments are purely notational, but the notation is admittedly cum-
bersome, and so the reader is encouraged to skim the next proof.

Proposition 3.6.3 (Properties of the Determinant).

(1) The determinant is linear as a function of each row of A.
(2) The determinant is skew-symmetric as a function of the rows of A.
 3.6 The Determinant: Uniqueness and Existence 107

(3) The determinant is normalized.

Proof. (1) If A has rows ri = (ai1 , , ain ) except that its kth row is the linear
combination rk +rk where rk = (ak1 , , akn ) and rk = (ak1 , , akn ), then
its (i, j)th entry is (
aij if i 6= k,
akj + akj if i = k.
Thus

det(r1 , , rk + rk , , rn )
X
= (1) a1(1) (ak(k) + ak(k) ) an(n)
Sn
X
= (1) a1(1) ak(k) an(n)
Sn
X
+ (1) a1(1) ak(k) an(n)
Sn

= det(r1 , , rk , , rn ) + det(r1 , , rk , , rn ),

as desired.
(2) Let A have rows r1 , , rn where ri = (ai1 , , ain ). Suppose that
rows k and k + 1 are exchanged. The resulting matrix has (i, j)th entry


aij if i
/ {k, k + 1},
ak+1,j if i = k,


akj if i = k + 1.

For each permutation Sn , define a companion permutation by exchang-

ing the kth and (k + 1)st entries,

= ((1), , (k + 1), (k), , (n)).

Thus (k) = (k + 1), (k + 1) = (k), and (i) = (i) for all other i.
As varies through Sn , so does , and for each we have the relation

(1) = (1) (exercise 3.6.6). The defining formula of the determinant
gives

det(r1 , , rk+1 , rk , , rn )
X
= (1) a1(1) ak+1,(k) ak(k+1) an(n)

X
= (1) a1 (1) ak+1, (k+1) ak (k) an (n)

= det(r1 , , rk , rk+1 , , rn ).
108 3 Linear Mappings and Their Matrices

The previous calculation establishes the result when adjacent rows of A are
exchanged. To exchange rows k and in A where > k, carry out the following
adjacent row exchanges:
rows k and k + 1, k and k + 1.
k + 1 and k + 2, k + 1 and k + 2,
, ,
2 and 1, 2 and 1,
1 and ,

This process trickles the kth row down to the th and then bubbles the th
row back up to the kth, bobbing each row in between them up one position
and then back down. And the display shows that the process carries out an
odd number of exchanges (all but the bottom one come in pairs), each of
which negates the determinant.
(3) This is left to the reader (exercise 3.6.7).

So a unique determinant function with the stipulated behavior exists. And
we have seen that any multilinear skew-symmetric function must be a scalar
multiple of the determinant. The last comment necessary to complete the
proof of Theorem 3.5.1 is that since the determinant is multilinear and skew-
symmetric, so are its scalar multiples. This fact was shown in exercise 3.5.3.
The reader is invited to contemplate how tortuous it would have been to
prove the various facts about the determinant in the previous section by using
the unwieldy determinant formula, with its n! terms.
The previous section has already established that the determinant of a
triangular matrix is the product of the diagonal entries, but the result also
follows immediately from the determinant formula (exercise 3.6.8). This fact
should be cited freely to save time.
An algorithm for computing det(A) for any A Mn (R) is now at hand.
Algebraically, the idea is that if
P1 AP2 =
where P1 and P2 are products of elementary matrices and is a triangular
matrix, then since the determinant is multiplicative,
det(A) = det(P1 )1 det() det(P2 )1 .
Multiplying A by P2 on the right carries out a sequence of column operations
on A, just as multiplying A by P1 on the left carries out row operations. Recall
that the determinants of the elementary matrices are
det(Ri;j,a ) = 1,
det(Si,a ) = a,
det(Ti;j ) = 1.
 3.6 The Determinant: Uniqueness and Existence 109

Procedurally, this all plays out as follows.

Proposition 3.6.4 (Determinant Algorithm). Given A Mn (R), use
row and column operationsrecombines, scales, transpositionsto reduce A
to a triangular matrix . Then det(A) is det() times the reciprocal of each
scale factor and times 1 for each transposition.
The only role that the determinant formula (as compared to the determi-
nant properties) played in obtaining this algorithm is that it gave the deter-
minant of a triangular matrix easily.
For example, the matrix

1/0! 1/1! 1/2! 1/3!
1/1! 1/2! 1/3! 1/4!
A= 1/2! 1/3! 1/4! 1/5!

1/3! 1/4! 1/5! 1/6!

becomes, after scaling the first row by 3!, the second row by 4!, the third row
by 5!, and the fourth row by 6!,

6 6 31
24 12 4 1
B=
60 20 5 1 .
120 30 6 1

Subtract the first row from each of the others to get

6 6 31
18 6 1 0
C=
54 14 2 0 ,
114 24 3 0

and then scale the third row by 1/2 and the fourth row by 1/3, yielding

6 631
18 6 1 0
D=
27 7 1 0 .
38 8 1 0

Next subtract the second row from the third row and the fourth rows, and
scale the fourth row by 1/2 to get

6 631
18 6 1 0
E=
9 1 0 0 .
10 1 0 0

Subtract the third row from the fourth, transpose the first and fourth columns,
and transpose the second and third columns, leading to
110 3 Linear Mappings and Their Matrices

1 36 6
0 16 18

= .
0 01 9
0 00 1

This triangular matrix has determinant 1, and so according to the algorithm,

232 1
det(A) = = .
6! 5! 4! 3! 1036800
In the following exercises, feel free to use the determinant properties and
the determinant formula in whatever combined way gives you the least work.

Exercises

3.6.1. For this exercise, let n and m be positive integers, not necessarily equal,
and let Rn Rn denote m copies of Rn . Consider any multilinear function

f : Rn Rn R.

For any m vectors in Rn ,

a1 = (a11 , , a1n ),
a2 = (a21 , , a2n ),
..
.
am = (am1 , , amn ),

explain why
X n
n X n
X
f (a1 , a2 , , am ) = a1i a2j amp f (ei , ej , , ep ).
i=1 j=1 p=1

Since each f (ei , ej , , ep ) is a constant (it depends on f , but not on the

vectors a1 , , am ) the multilinear function f is thus a polynomial in the
entries of its vector-variables. Therefore this exercise has shown that every
multilinear function is continuous.

3.6.2. Use the three desired determinant properties to derive the formulas in
the section for 1-by-1 and 3-by-3 determinant. Verify that the 1-by-1 formula
satisfies the properties.

3.6.3. For each permutation, count the inversions and compute the sign:
(2, 3, 4, 1), (3, 4, 1, 2), (5, 1, 4, 2, 3).

3.6.4. Explain why there are n! permutations of {1, , n}.

 3.6 The Determinant: Uniqueness and Existence 111

3.6.5. Define the permutation = (n, n 1, n 2, , 1) Sn . Show that

has (n 1)n/2 inversions and that
(
1 if n has the form 4k or 4k + 1 (k Z),
(1) =
1 otherwise.

3.6.6. Explain why (1) = (1) in the proof of part (2) of Proposi-
tion 3.6.3.
3.6.7. Use the defining formula of the determinant to reproduce the result
that det(In ) = 1.

3.6.8. ExplainP Pn (1) a1(1) a2(2) an(n) from the determi-
why in any term
n
nant formula, i=1 (i) = i=1 i. Use this to re-explain why the determinant
of a triangular matrix is the product of its diagonal entries.
3.6.9. Calculate the determinants of the following matrices:

4 3 1 2 1 1 2 3
0 1 2 3 2 2 0 2

1 0 4 1, 4 1 1 1 .
20 30 1 2 3 0
3.6.10. Show that the Vandermonde matrix,

1 a a2
1 b b2 ,
1 c c2

has determinant (b a)(c a)(c b). For what values of a, b, c is the Vander-
monde matrix invertible? (The idea is to do the problem conceptually rather
than writing out the determinant and then factoring it, so that the same ideas
would work for larger matrices. The determinant formula shows that the de-
terminant in the problem is a polynomial in a, b, and c. What is its degree in
each variable? Why must it vanish if any two variables are equal? Once you
have argued that that the determinant is as claimed, dont forget to finish the
problem.)
3.6.11. Consider the following n-by-n matrix based on Pascals triangle:

11 1 1 1
1 2 3 4 n

1 3
6 10 n(n+1)
2


A = 1 4 10 20 .

.. .. .. .. ..
. . . . .
n(n+1)
1n 2

Find det(A). (Hint: Row and column reduce.)

112 3 Linear Mappings and Their Matrices

3.7 An Explicit Formula for the Inverse

Consider an invertible linear mapping

T : Rn Rn

having matrix
A Mn (R).
In section 3.3 we discussed a process to invert A and thereby invert T . Now,
with the determinant in hand, we can also write the inverse of A explicitly in
closed form. Because the formula giving the inverse involves many determi-
nants, it is hopelessly inefficient for computation. Nonetheless, it is of interest
to us for a theoretical reason (the pending Corollary 3.7.3) that we will need
in chapter 5.

Definition 3.7.1 (Classical Adjoint). Let n 2 be an integer, and let

A Mn (R) be an n-by-n matrix. For any i, j {1, , n}, let

Ai,j Mn1 (R)

be the (n 1)-by-(n 1) matrix obtained by deleting the ith row and the jth
column of A. The classical adjoint of A is the n-by-n matrix whose (i, j)th
entry is (1)i+j times the determinant of Aj,i ,

Aadj = [(1)i+j det(Aj,i )] Mn (R).

The factor (1)i+j in the formula produces an alternating checkerboard

pattern of plus and minus signs, starting with a plus sign in the upper left cor-
ner of Aadj . Note that the (i, j)th entry of Aadj involves Aj,i rather than Ai,j .
For instance, in the 2-by-2 case,
 adj  
ab d b
= .
cd c a

Already for a 3-by-3 matrix the formula for the classical adjoint is daunting,
     
ef b c b c
adj det det det
ab c h k  h k   e f 
ac
d e f = det d f det
ac
det

ghk g k  g k d f 
de ab ab
det det det
gh gh de

ek f h ch bk bf ce
= f g dk ak cg cd af .
dh eg bg ah ae bd

Returning to the 2-by-2 case, where

 3.7 An Explicit Formula for the Inverse 113
   
ab d b
A= and Aadj = ,
cd c a

compute that
   
adj ad bc 0 10
AA = = (ad bc) = det(A)I2 .
0 ad bc 01

The same result holds in general:

Proposition 3.7.2 (Classical Adjoint Identity). Let n 2 be an integer,

let A Mn (R) be an n-by-n matrix, and let Aadj be its classical adjoint. Then

A Aadj = det(A)In .

Especially, if A is invertible then

1
A1 = Aadj .
det(A)

The idea of the proof is that the inner product of the ith row of A and
the ith column of Aadj gives precisely the formula for det(A), while for i 6= j
the inner product of the ith row of A and the jth column of Aadj gives the
formula for the determinant of a matrix having the ith row of A as two of its
rows. The argument is purely formal but notationally tedious, and so we omit
it.
In the 2-by-2 case the proposition gives us a slogan:
To invert a 2-by-2 matrix, exchange the diagonal elements, negate the
off-diagonal elements, and divide by the determinant.
Again, for n > 2 the explicit formula for the inverse is rarely of calculational
use. We care about it for the following reason.

Corollary 3.7.3. Let A Mn (R) be an invertible n-by-n matrix. Then each

entry of the inverse matrix A1 is a continuous function of the entries of A.

Proof. Specifically, the (i, j)th entry of A1 is

(A1 )i,j = (1)i+j det(Aj,i )/ det(A),

a rational function (ratio of polynomials) of the entries of A. As such it varies

continuously in the entries of A so long as A remains invertible.

Exercise

3.7.1. Verify at least one diagonal entry and at least one off-diagonal entry
in the formula A Aadj = det(A)In for n = 3.
114 3 Linear Mappings and Their Matrices

3.8 Geometry of the Determinant: Volume

Consider a linear mapping from n-space back to n-space,

T : Rn Rn .

This section discusses two ideas:

The mapping T magnifies volume by a constant factor. (Here volume is a
pan-dimensional term that in particular means length when n = 1, area
when n = 2, and the usual notion of volume when n = 3.) That is, there
is some number t 0 such that if one takes a set,

E Rn ,

and passes it through the mapping to get another set,

T E Rn ,

then the sets volume is multiplied by t,

vol T E = t vol E.

The magnification factor t depends on T but is independent of the set E.

Furthermore, if the matrix of T is A then the magnification factor associ-
ated to T is
t = | det A|.
That is, the absolute value of det A has a geometric interpretation as the
factor by which T magnifies volume.
(The geometric interpretation of the sign of det A will be discussed in the next
section.)
An obstacle to pursuing these ideas is that we dont have a theory of
volume in Rn readily at hand. In fact, volume presents real difficulties. For
instance, any notion of volume that has sensible properties can not apply
to all sets; so either volume behaves unreasonably or some sets dont have
well defined volumes at all. Here we have been tacitly assuming that volume
does behave well and that the sets E under consideration do have volumes.
This section will investigate volume informally by considering how it ought
to behave, stating assumptions as they arise and arriving only at a partial
description. The resulting arguments will be heuristic, and the skeptical reader
will see gaps in the reasoning. Volume will be discussed further in chapter 6,
but a full treatment of the subject (properly called measure) is beyond the
range of this text.
The standard basis vectors e1 , , en in Rn span the unit box,

B = {1 e1 + + n en : 0 1 1, , 0 n 1}.
 3.8 Geometry of the Determinant: Volume 115

Thus box means interval when n = 1, rectangle when n = 2, and the usual
notion of box when n = 3. Let p be a point in Rn , let a1 , , an be positive
real numbers, and let B denote the box spanned by the vectors a1 e1 , , an en
and translated by p,

B = {1 a1 e1 + + n an en + p : 0 1 1, , 0 n 1}.

(See figure 3.11. The figures of this section are set in two dimensions, but the
ideas are general and hence so are the figure captions.) A face of a box is the
set of its points such that some particular i is held fixed at 0 or at 1 while
the others vary. A box in Rn has 2n faces.
A natural definition is that the unit box has unit volume,

vol B = 1.

We assume that volume is unchanged by translation. Also, we assume that box

volume is finitely additive, meaning that given finitely many boxes B1 , , BM
that are disjoint except possibly for shared faces or shared subsets of faces,
the volume of their union is the sum of their volumes,
M
[ M
X
vol Bi = vol Bi . (3.7)
i=1 i=1

And we assume that scaling any spanning vector of a box affects the boxs
volume continuously in the scaling factor. It follows that scaling any spanning
vector of a box by a real number a magnifies the volume by |a|. To see this,
first note that scaling a spanning vector by an integer creates || abutting
translated copies of the original box, and so the desired result follows in this
case from finite additivity. A similar argument applies to scaling a spanning
vector by a reciprocal integer 1/m (m 6= 0), since the original box is now |m|
copies of the scaled one. These two special cases show that the result holds
for scaling a spanning vector by any rational number r = /m. Finally, the
continuity assumption extends the result from the rational numbers to the
real numbers, since every real number is approached by a sequence of rational
numbers. Since the volume of the unit box is normalized to 1, since volume
is unchanged by translation, and since scaling any spanning vector of a box
by a magnifies its volume by |a|, the volume of the general box is (recalling
that a1 , , an are assumed to be positive)

vol B = a1 an .

A subset of Rn that is well approximated by boxes plausibly has a volume.

To be more specific, a subset E of Rn is well approximated by boxes if for any
> 0 there exist boxes B1 , , BN , BN +1 , , BM that are disjoint except
possibly for shared faces, such that E is contained between a partial union of
the boxes and the full union,
116 3 Linear Mappings and Their Matrices

p + a 2 e2
B e2
p p + a 1 e1 B

e1
Figure 3.11. Scaling and translating the unit box

N
[ M
[
Bi E Bi , (3.8)
i=1 i=1

and such that the boxes that complete the partial union to the full union have
a small sum of volumes,
M
X
vol Bi < . (3.9)
i=N +1

(See figure 3.12, where E is an elliptical region, the boxes B1 through BN that
it contains are dark, and the remaining boxes BN +1 through BM are light.)
To see that E should have a volume, note that the first containment of (3.8)
says that a number at most big enough to serve as vol E (a lower bound) is
SN
L = vol i=1 Bi , and the second containment says that a number at least
SM
big enough (an upper bound) is U = vol i=1 Bi . By the finite additivity
PN
condition (3.7), the lower and upper bounds are L = i=1 vol Bi and U =
PM
i=1 vol Bi . Thus they are close to each other by (3.9),

M
X
U L= vol Bi < .
i=N +1

Since is arbitrarily small, the bounds should be squeezing down on a unique

value that is the actual volume of E, and so indeed E should have a volume.
For now this is only a plausibility argument, but it is essentially the idea of
integration and it will be quantified in chapter 6.
Any n vectors v1 , , vn in Rn span an n-dimensional parallelepiped

P(v1 , , vn ) = {1 v1 + + n vn : 0 1 1, , 0 n 1},

abbreviated to P when the vectors are firmly fixed. Again the terminology
is pan-dimensional, meaning in particular interval, parallelogram, and paral-
lelepiped in the usual sense for n = 1, 2, 3. We will also consider translations
of parallelepipeds away from the origin by offset vectors p,

P = P + p = {v + p : v P}.
 3.8 Geometry of the Determinant: Volume 117

Figure 3.12. Inner and outer approximation of E by boxes

(See figure 3.13.) A face of a parallelepiped is the set of its points such that
some particular i is held fixed at 0 or at 1 while the others vary. A paral-
lelepiped in Rn has 2n faces. Boxes are special cases of parallelepipeds. The
methods of chapter 6 will show that parallelepipeds are well approximated by
boxes, and so they have well defined volumes. We assume that parallelepiped
volume is finitely additive, and we assume that any finite union of paral-
lelepipeds each having volume zero again has volume zero.

p + v2
P
v2
p + v1
P
v1
p

Figure 3.13. Parallelepipeds

To begin the argument that the linear mapping T : Rn Rn magnifies

volume by a constant factor, we pass the unit box B and the scaled translated
box B from earlier in the section through T . The image of B under T is
a parallelepiped T B spanned by T (e1 ), , T (en ), and the image of B is a
parallelepiped T B spanned by T (a1 e1 ), , T (an en ) and translated by T (p).
(See figure 3.14.) Since T (a1 e1 ) = a1 T (e1 ), , T (an en ) = an T (en ), it follows
that scaling the sides of T B by a1 , , an and then translating the scaled
parallelepiped by T (p) gives T B . As for boxes, scaling any spanning vector
of a parallelepiped by a real number a magnifies the volume by |a|, and so we
have
vol T B = a1 an vol T B.
But recall that also
118 3 Linear Mappings and Their Matrices

vol B = a1 an .
The two displays combine to give
vol T B
= vol T B.
vol B
That is, the volume of the T -image of a box divided by the volume of the
box is constant, regardless of the boxs location or side lengths, the constant
being the volume of T B, the T -image of the unit box B. Call this constant
magnification factor t. Thus,

vol T B = t vol B for all boxes B . (3.10)

T (p) T B

B
p
B
TB

Figure 3.14. Linear image of the unit box and of a scaled translated box

We need one last preliminary result about volume. Again let E be a subset
of Rn that is well approximated by boxes. Fix a linear mapping T : Rn
Rn . Very similarly to the argument for E, the set T E also should have a
volume, because it is well approximated by parallelepipeds. Indeed, the set
containments (3.8) are preserved under the linear mapping T ,
N
[ M
[
T Bi T E T Bi .
i=1 i=1

In general, the image of a union is the union of the images, so this rewrites as
N
[ M
[
T Bi T E T Bi .
i=1 i=1

(See figure 3.15.) As before, numbers at most big enough and at least big
enough for the volume of T E are
N
[ N
X M
[ M
X
L = vol T Bi = vol T Bi , U = vol T Bi = vol T Bi .
i=1 i=1 i=1 i=1
 3.8 Geometry of the Determinant: Volume 119

The only new wrinkle is that citing the finite additivity of parallelepiped
volume here assumes that the parallelepipeds T Bi either inherit from the
original boxes Bi the property of being disjoint except possibly for shared
faces, or they all have volume zero. The assumption is valid because if T is
invertible then the inheritance holds, while if T is not invertible then we will
see later in this section that the T Bi have volume zero as desired. With this
point established, let t be the factor by which T magnifies box-volume. The
previous display and (3.10) combine to show that the difference of the bounds
is
XM M
X M
X
U L= vol T Bi = t vol Bi = t vol Bi t.
i=N +1 i=N +1 i=N +1

The inequality is strict if t > 0, and it collapses to U L = 0 if t = 0. In

either case, since is arbitrarily small, the argument that T E should have a
volume is the same as for E.

Figure 3.15. Inner and outer approximation of T E by parallelepipeds

To complete the argument that the linear mapping T : Rn Rn mag-

nifies volume by a constant factor, we argue that for any subset E of Rn
SN approximated by boxes, vol T E is t times the volume of E. Let
that is well
V = vol i=1 Bi . Then E is contained between a set of volume V and a set of
volume less than V + (again see figure 3.12, where V is the shaded area and
V + is the total area), and T E is contained between a set of volume tV and
a set of volume at most t(V + ) (again see figure 3.15, where tV is the shaded
area and t(V +) is the total area). Thus the volumes vol E and vol T E satisfy
the condition
tV vol T E t(V + )
.
V + vol E V
Since can be arbitrarily small, the left and right quantities in the display
can be arbitrarily close to t, and so the only possible value for the quantity in
the middle (which is independent of ) is t. Thus we have the desired equality
announced at the beginning of the section,
120 3 Linear Mappings and Their Matrices

vol T E = t vol E.
In sum, subject to various assumptions about volume, T magnifies the volumes
of all boxes and of all figures that are well approximated by boxes by the same
factor, which we have denoted t.
Now we investigate the magnification factor t associated to the linear map-
ping T , with the goal of showing that it is | det A| where A is the matrix of T .
As a first observation, if the linear mappings S, T : Rn Rn magnify volume
by s and t respectively, then their composition S T magnifies volume by st.
In other words, the magnification of linear mappings is multiplicative. Also,
recall that the mapping T is simply multiplication by the matrix A. Since any
matrix is a product of elementary matrices times an echelon matrix, we only
need to study the magnification of multiplying by such matrices. Temporarily
let n = 2.
The 2-by-2 recombine matrices take the form R = [ 10 a1 ] and R = [ a1 10 ]
with a R. The standard basis vectors e1 and e2 are taken by R to its
columns, e1 and ae1 + e2 . Thus R acts geometrically as a shear by a in the
e1 -direction, magnifying volume by 1. (See figure 3.16.) Note that 1 = | det R|
as desired. The geometry of R is left as an exercise.

Figure 3.16. Shear

The scale matrices are S = [ a0 10 ] and S = [ 10 a0 ]. The standard basis gets

taken by S to ae1 and e2 , so S acts geometrically as a scale in the e1 -direction,
magnifying volume by |a|; this is | det S|, again as desired. (See figure 3.17.)
The situation for S is similar.
The transposition matrix is T = [ 01 10 ]. It exchanges e1 and e2 , acting as
a reflection through the diagonal, magnifying volume by 1. (See figure 3.18.)
Since det T = 1, the magnification factor is the absolute value of the deter-
minant.
Finally, the identity matrix E = I has no effect, magnifying volume by 1,
and any other echelon matrix E has bottom row (0, 0) and hence squashes e1
and e2 to vectors whose last component is 0, magnifying volume by 0. (See
figure 3.19.) The magnification factor is | det E| in both cases.
The discussion for scale matrices, transposition matrices, and echelon ma-
trices generalizes effortlessly from 2 to n dimensions, but generalizing the
 3.8 Geometry of the Determinant: Volume 121

Figure 3.17. Scale

Figure 3.18. Reflection

Figure 3.19. Squash

discussion for recombine matrices Ri;j,a takes a small argument. Since trans-
position matrices have no effect on volume, we may multiply Ri;j,a from the
left and from the right by various transposition matrices to obtain R1;2,a and
study it instead. Multiplication by R1;2,a preserves all of the standard basis
vectors except e2 , which is taken to ae1 + e2 as before. The resulting paral-
lelepiped P(e1 , ae1 + e2 , e3 , , en ) consists of the parallelogram shown in the
right side of figure 3.16, extended one unit in each of the remaining orthogonal
n 2 directions of Rn . The n-dimensional volume of the parallelepiped is its
base (the area of the parallelogram, 1) times its height (the (n2)-dimensional
volume of the unit box over each point of the parallelogram, again 1). That is,
the n-by-n recombine matrix still magnifies volume by 1, the absolute value
of its determinant, as desired. The base times height property of volume is yet
122 3 Linear Mappings and Their Matrices

another invocation here, but it is a consequence of a theorem to be proved in

chapter 6, Fubinis Theorem. Summarizing,

Theorem 3.8.1 (Geometry of Linear Mappings). Any linear mapping

T : Rn Rn is the composition of a possible squash followed by shears, scales
and reflections. If the matrix of T is A then T magnifies volume by | det A|.

Proof. The matrix A of T is a product of elementary matrices and an echelon

matrix. The elementary matrices act as shears, scales and reflections, and if
the echelon matrix is not the identity then it acts as a squash. This proves
the first statement. Each elementary or echelon matrix magnifies volume by
the absolute value of its determinant. The second statement follows since
magnification and | det | are both multiplicative.

The work of this section has given a geometric interpretation of the mag-
nitude of det A: it is the magnification factor of multiplication by A. If the
columns of A are denoted c1 , , cn then Aej = cj for j = 1, , n, so that
even more explicitly | det A| is the volume of the parallelepiped spanned by
the columns of A. For instance, to find the volume of the three-dimensional
parallelepiped spanned by the vectors (1, 2, 3), (2, 3, 4), and (3, 5, 8), compute
that
123
| det 2 3 5 | = 1.
348

Exercises

3.8.1. (a) The section states that the image of a union is the union of the
images. More specifically, let A and B be any sets, let f : A B be any
mapping, and let A1 , , AN be any subsets of A. Show that
N
! N
[ [
f Ai = f (Ai ).
i=1 i=1

(This exercise is purely set-theoretic, making no reference to our working

environment of Rn .)
(b) Consider a two-point set A = {a1 , a2 } where a1 6= a2 , a one-point
set B = {b}, and the only possible mapping f : A B, given by
f (a1 ) = f (a2 ) = b. Let A1 = {a1 } and A2 = {a2 }, subsets of A. What is
the intersection A1 A2 ? What is the image of the intersection, f (A1 A2 )?
What are the images f (A1 ) and f (A2 )? What is the intersection of the images,
f (A1 ) f (A2 )? Is the image of an intersection in general the intersection of
the images?

3.8.2. Describe the geometric effect of multiplying by the matrices R and S

in the section. Describe the effect of multiplying by R and S if a < 0.
 3.9 Geometry of the Determinant: Orientation 123

3.8.3. Describe the geometric effect of multiplying by the 3-by-3 elementary

matrices R2;3,1 , R3;1,2 , and S2,3 .
 
3.8.4. (a) Express the matrix 01 10 as a product of recombine and scale
matrices (you may not need both types).
(b) Use part (a) to describe counterclockwise rotation of the plane through
angle /2 as a composition of shears and scales.

3.8.5. Describe counterclockwise rotation of the plane through angle (where

cos 6= 0 and sin 6= 0) as a composition of shears and scales.

3.8.6. In R3 , describe the linear mapping that takes e1 to e2 , e2 to e3 , and e3

to e1 as a composition of shears, scales, and transpositions.

3.8.7. Let P be the parallelogram a b in R2 spanned by (a, c) and (b, d). Cal-
culate directly that | det c d | = area P. (Hint: area = base height
= |(a, c)| |(b, d)| | sin (a,c),(b,d) |. It may be cleaner to find the square of the
area.)

3.8.8. This exercise shows directly that | det | = volume in R3 . Let P be the
parallelepiped in R3 spanned by v1 , v2 , v3 , let P be spanned by the vectors
v1 , v2 , v3 obtained from performing the GramSchmidt process on the vj s,
let A M3 (R) have rows v1 , v2 , v3 and let A M3 (R) have rows v1 , v2 , v3 .
(a) Explain why det A = det A.
(b) Give a plausible geometric argument that vol P = vol P.
(c) Show that 2
|v1 | 0 0
t
A A = 0 |v2 |2 0 .
0 0 |v3 |2
Explain why therefore | det A | = vol P . It follows from parts (a) and (b) that
that | det A| = vol P.

3.9 Geometry of the Determinant: Orientation

Recall from section 2.1 that a basis of Rn is a set of vectors {f1 , , fp } such
that any vector in Rn is a unique linear combination of the {fj }. Though
strictly speaking a basis is only a set, we adopt here the convention that the
basis vectors are given in the specified order indicated. Given such a basis,
view the vectors as columns and let F denote the matrix in Mn,p (R) with
columns f1 , , fp . Thus the order of the basis vectors is now relevant. For
a standard basis vector ej of Rp , the matrix-by-vector product F ej gives the
jth column fj of F . Therefore, for any vector x = (x1 , , xp ) Rp (viewed
as a column),
124 3 Linear Mappings and Their Matrices

Xp p
X p
X
Fx = F
x j ej = x j F ej = x j fj .
j=1 j=1 j=1

Thus, multiplying all column vectors x Rp by the matrix F gives precisely

the linear combinations of f1 , , fp , and so we have the equivalences

{f1 , , fp } is a basis of Rn
!
each y Rn is uniquely expressible

as a linear combination of the {fj }
!
each y Rn takes the form

y = F x for a unique x Rp
F is invertible
F is square (i.e., p = n) and det F 6= 0.

These considerations have proved

Theorem 3.9.1 (Characterization of Bases). Any basis of Rn has n el-

ements. The vectors {f1 , , fn } form a basis exactly when the matrix F
having them as its columns has nonzero determinant.

Let {f1 , , fn } be a basis of Rn , and let F be the matrix formed by their

columns. Abuse terminology and call det F the determinant of the basis,
written det{f1 , , fn }. Again, this depends on the order of the {fj }. There
are then two kinds of basis of Rn , positive and negative bases, according
to the sign of their determinants. The standard basis {e1 , , en } forms the
columns of the identity matrix I and is therefore positive.
The multilinear function det F is continuous in the n2 entries of f1 , , fn
(see exercise 3.6.1). If a basis {f1 , , fn } can be smoothly deformed via other
bases to the standard basis then the corresponding determinants must change
continuously to 1 without passing through 0. Such a basis must therefore be
positive. Similarly, a negative basis can not be smoothly deformed via other
bases to the standard basis. It is also true but less clear (and not proved here)
that every positive basis deforms smoothly to the standard basis.
The plane R2 is by convention drawn with {e1 , e2 } forming a counterclock-
wise angle of /2. Two vectors {f1 , f2 } form a basis if they are not collinear.
Therefore the basis {f1 , f2 } can be deformed via bases to {e1 , e2 } exactly
when the angle f1 ,f2 goes counterclockwise from f1 to f2 . (Recall from equa-
tion (2.2) that the angle between two nonzero vectors is between 0 and .)
That is, in R2 , the basis {f1 , f2 } is positive exactly when the angle from f1
to f2 is counterclockwise. (See figure 3.20.)
Three-space R3 is by convention drawn with {e1 , e2 , e3 } forming a right-
handed triple, meaning that when the fingers of your right hand curl from
e1 to e2 your thumb forms an acute angle with e3 . Three vectors {f1 , f2 , f3 }
 3.9 Geometry of the Determinant: Orientation 125
f2
f1

f1 f2

Figure 3.20. Positive and negative bases of R2

form a basis if they are not coplanar. In other words they must form a right-
or left-handed triple. Only right-handed triples deform via other nonplanar
triples to {e1 , e2 , e3 }. Therefore in R3 , the basis {f1 , f2 , f3 } is positive exactly
when it forms a right-handed triple. (See figure 3.21.)

f3
f2
f2
f1
f1 f3

Figure 3.21. Positive and negative bases of R3

The geometric generalization to Rn of a counterclockwise angle in the plane

and a right-handed triple in space is not so clear, but the algebraic notion of
positive basis is the same for all n.
Consider any invertible mapping T : Rn Rn with matrix A Mn (R),
and any basis {f1 , , fn } of Rn . If F again denotes the matrix with columns
f1 , , fn then AF has columns {Af1 , , Afn } = {T (f1 ), , T (fn )}. These
form a new basis of Rn with determinant

det{T (f1 ), , T (fn )} = det AF = det A det F = det A det{f1 , , fn }.

The calculation lets us interpret the sign of det A geometrically: If det A > 0
then T preserves the orientation of bases, and if det A < 0 then T reverses
orientation. For example, the mapping with matrix

0001
1 0 0 0

0 1 0 0
0010
126 3 Linear Mappings and Their Matrices

reverses orientation in R4 .
To summarize: Let A be an n-by-n matrix. Whether det A is nonzero says
whether A is invertible, the magnitude of det A is the factor by which A
magnifies volume, and (assuming that det A 6= 0) the sign of det A determines
how A affects orientation.

Exercises

3.9.1. Any invertible mapping T : Rn Rn is a composition of scales,

shears and transpositions. Give conditions on such a composition to make the
mapping orientation-preserving, orientation-reversing.

3.9.2. Does the linear mapping T : Rn Rn that takes e1 to e2 , e2 to e3 ,

, en to e1 preserve or reverse orientation? (The answer depends on n.)
More generally, if is a permutation in Sn , does the linear mapping taking e1
to e(1) , , en to e(n) preserve or reverse orientation? (This depends on .)

3.9.3. Argue geometrically in R2 that any basis can be smoothly deformed

via other bases to the standard basis or to {e1 , e2 }. Do the same for R3
and {e1 , e2 , e3 }.

3.10 The Cross Product, Lines, and Planes in R3

Generally in Rn there is no natural way to associate to a pair of vectors u
and v a third vector. In R3 , however, the plane specified by u and v has only
one orthogonal direction, i.e., dimension 3 is special because 3 2 = 1. In R3
a normal vector to u and v can be specified by making suitable conventions on
its orientation viz a viz the other two vectors, and on its length. This will give
a vector-valued product of two vectors that is special to three-dimensional
space, called the cross product. The first part of this section develops these
ideas.
Given any two vectors u, v R3 , we want their cross product u v R3
to be orthogonal to u and v,

uv u and u v v. (3.11)

There is the question of which way uv should point along the line orthogonal
to the plane spanned by u and v. The natural answer is that the direction
should be chosen to make the ordered triple of vectors {u, v, u v} positive
unless it is degenerate,
det(u, v, u v) 0. (3.12)
Also there is the question of how long u v should be. With hindsight, we
assert that specifying the length to be the area of the parallelogram spanned
by u and v will work well. That is,
 3.10 The Cross Product, Lines, and Planes in R3 127

|u v| = area P(u, v). (3.13)

The three desired geometric properties (3.11) through (3.13) seem to describe
the cross product completely. (See figure 3.22.)

u
Figure 3.22. The cross product of u and v

The three geometric properties also seem disparate. However, they combine
into a uniform algebraic property, as follows. Since the determinant in (3.12) is
nonnegative, it is the volume of the parallelepiped spanned by u, v, and u v.
The volume is the base times the height, and since u v is normal to u and v
the base is the area of P(u, v) and the height is |u v|. Thus

det(u, v, u v) = area P(u, v) |u v|.

It follows from the previous display and (3.13) that

|u v|2 = det(u, v, u v).

Since orthogonal vectors have inner product 0, since the determinant is 0 when
two rows agree, and since the square of the absolute value is the vectors inner
product with itself, we can rewrite (3.11) and this last display (obtained from
(3.12) and (3.13)) uniformly as equalities of the form hu v, wi = det(u, v, w)
for various w,
hu v, ui = det(u, v, u),
hu v, vi = det(u, v, v), (3.14)
hu v, u vi = det(u, v, u v).
Instead of saying what the cross product is, as an equality of the form u v =
f (u, v) would, the three equalities of (3.14) say how the cross product interacts
with certain vectorsincluding itselfvia the inner product. Again, the idea
is to characterize rather than construct.
(The reader may object to the argument just given that det(u, v, u v) =
area P(u, v) |u v|, on the grounds that we dont really understand the area
of a 2-dimensional parallelogram in 3-dimensional space to start with, that
128 3 Linear Mappings and Their Matrices

in R3 we measure volume rather than area, and the parallelogram surely has
volume zero. In fact, the argument can be viewed as motivating the formula
as the definition of the area. This idea will be discussed more generally in
section 9.1.)
Based on (3.14), we leap boldly to an intrinsic algebraic characterization
of the cross product.
Definition 3.10.1 (Cross Product). Let u and v be any two vectors in R3 .
Their cross product u v is defined by the property

hu v, wi = det(u, v, w) for all w R3 .

That is, u v is the unique vector x R3 such that hx, wi = det(u, v, w) for
all w R3 .
As with the determinant earlier, we do not yet know that the characterizing
property determines the cross product uniquely, or even that a cross product
that satisfies the characterizing property exists at all. But also as with the
determinant, we defer those issues and first reap the consequences of the
characterizing property with no reference to an unpleasant formula for the
cross product. Of course the cross product will exist and be unique, but for
now the point is that graceful arguments with its characterizing property show
that it has all the further properties that we want it to have.
Proposition 3.10.2 (Properties of the Cross Product).
(CP1) The cross product is skew-symmetric: v u = u v for all u, v R3 .
(CP2) The cross product is bilinear: For all scalars a, a , b, b R and all
vectors u, u , v, v R3 ,

(au + a u ) v = a(u v) + a (u v),

u (bv + b v ) = b(u v) + b (u v ).

(CP3) The cross product u v is orthogonal to u and v.

(CP4) u v = 0 if and only if u and v are collinear (meaning that u = av or
v = au for some a R).
(CP5) If u and v are not collinear then the triple {u, v, u v} is right-handed.
(CP6) The magnitude |u v| is the area of the parallelogram spanned by u
and v.
Proof. (1) This follows from the skew-symmetry of the determinant. For any
w R3 ,

hv u, wi = det(v, u, w) = det(u, v, w) = hu v, wi = hu v, wi.

Since w is arbitrary, v u = u v.
(2) For the first variable, this follows from the linearity of the determinant
in its first row-vector variable and the linearity of the inner product in its first
vector variable. Fix a, a R, u, u , v R3 . For any w R3 ,
 3.10 The Cross Product, Lines, and Planes in R3 129

h(au + a u ) v, wi = det(au + a u , v, w)
= a det(u, v, w) + a det(u , v, w)
= ahu v, wi + a hu v, wi
= ha(u v) + a (u v), wi.

Since w is arbitrary, (au + a u ) v = a(u v) + a (u v). The proof for the

second variable follows from the result for the first variable and from (1).
(3) hu v, ui = det(u, v, u) = 0 since the determinant of a matrix with
two equal rows vanishes. Similarly, hu v, vi = 0.
(4) If u = av then for any w R3 ,

hu v, wi = hav v, wi = det(av, v, w) = a det(v, v, w) = 0.

Since w is arbitrary, u v = 0. And similarly if v = au.

Conversely, suppose that u and v are not collinear. Then they are linearly
independent, and so no element of R3 can be written as a linear combination
of u and v in more than one way. The set {u, v} is not a basis of R3 , because
every basis consists of three elements. Since no elements of R3 can be written
as a linear combination of u and v in more than one way, and since {u, v}
is not a basis, the only possibility is that some w R3 can not be written
as a linear combination of u and v at all. Thus the set {u, v, w} is a linearly
independent set of three elements, making it a basis of R3 . Compute that
since {u, v, w} is a basis,

hu v, wi = det(u, v, w) 6= 0.

Therefore u v 6= 0.
(5) By (4), u v 6= 0, so 0 < hu v, u vi = det(u, v, u v). By the results
on determinants and orientation, {u, v, u v} is right-handed.
(6) By definition, |u v|2 = hu v, u vi = det(u, v, u v). As discussed
earlier in the section, det(u, v, u v) = areaP(u, v) |u v|. The result follows
from dividing by |u v| if it is nonzero, and from (4) otherwise.

Now we show that the characterizing property determines the cross prod-
uct uniquely. The idea is that a vectors inner products with all other vectors
completely describe the vector itself. The observation to make is that for any
vector x Rn (n need not be 3 in this paragraph),

if hx, wi = 0 for all w Rn then x = 0n .

To justify the observation, specialize w to x to show that hx, xi = 0, giving the

result since 0n is the only vector whose inner product with itself is 0. (Here
we use the nontrivial direction of the degeneracy condition in the positive
definiteness property of the inner product.) In consequence of the observation,
for any two vectors x, x Rn ,

if hx, wi = hx , wi for all w Rn then x = x .

130 3 Linear Mappings and Their Matrices

That is, the inner product values hx, wi for all w Rn specify x, as anticipated.
To prove that the cross product exists, it suffices to write a formula for it
that satisfies the characterizing property in Definition 3.10.1. Since we need

hu v, e1 i = det(u, v, e1 ),
hu v, e2 i = det(u, v, e2 ),
hu v, e3 i = det(u, v, e3 ),

the only possible formula is

u v = (det(u, v, e1 ), det(u, v, e2 ), det(u, v, e3 )).

This formula indeed satisfies the definition because by definition of the inner
product and then by the linearity of the determinant in its third argument we
have for any w = (w1 , w2 , w3 ) R3 ,

hu v, wi = det(u, v, e1 ) w1 + det(u, v, e2 ) w2 + det(u, v, e3 ) w3

= det(u, v, w1 e1 + w2 e2 + w3 e3 )
= det(u, v, w).

In coordinates, the formula for the cross product is

u1 u2 u3 u1 u2 u3 u1 u2 u3
u v = (det v1 v2 v3 , det v1 v2 v3 , det v1 v2 v3 )
1 0 0 0 1 0 0 0 1
= (u2 v3 u3 v2 , u3 v1 u1 v3 , u1 v2 u2 v1 ).

A bit more conceptually, the cross product formula in coordinates is

u1 u2 u3
u v = det v1 v2 v3 .
e1 e2 e3

The previous display is only a mnemonic device: strictly speaking, it doesnt

lie within our grammar because the entries of the bottom row are vectors
rather than scalars. But even so, its two terms u1 v2 e3 u2 v1 e3 do give the
third entry of the cross product, and similarly for the others. In chapter 9,
where we will have to compromise our philosophy of working intrinsically
rather than in coordinates, this formula will be cited and generalized. In the
meantime its details are not important except for mechanical calculations,
and we want to use it as little as possible, as with the determinant earlier.
Indeed, the display shows that the cross product is essentially a special case
of the determinant.
It is worth knowing the cross products of the standard basis pairs,
 3.10 The Cross Product, Lines, and Planes in R3 131

e1 e1 = 03 , e1 e2 = e3 , e1 e3 = e2 ,
e2 e1 = e3 , e2 e2 = 03 , e2 e3 = e1 ,
e3 e1 = e2 , e3 e2 = e1 , e3 e3 = 03 .

Here ei ej is 03 if i = j, and ei ej is the remaining standard basis vector

if i 6= j and i and j are in order in the diagram
/2

1Y


3
and ei ej is minus the remaining standard basis vector if i 6= j and i and j
are out of order in the diagram.

The remainder of this section describes lines and planes in R3 .

A line in R3 is determined by a point p and a direction vector d. (See
figure 3.23.) A point q lies on the line exactly when it is a translation from p
by some multiple of d. Therefore,

(p, d) = {p + td : t R}.

In coordinates, a point (x, y, z) lies on ((xp , yp , zp ), (xd , yd , zd )) exactly when

x = xp + txd , y = yp + tyd , z = zp + tzd for some t R.

If the components of d are all nonzero then the relation between the coordi-
nates can be expressed without the parameter t,
x xp y yp z zp
= = .
xd yd zd
For example, the line through (1, 1, 1) in the direction (1, 2, 3) consists of all
points (x, y, z) satisfying x = 1 + t, y = 1 + 2t, z = 1 + 3t for t R, or,
equivalently, satisfying x 1 = (y 1)/2 = (z 1)/3.
A plane in R3 is determined by a point p and a normal (orthogonal)
vector n. (See figure 3.24.) A point x lies on the plane exactly when the
vector from p to x is orthogonal to n. Therefore,

P (p, n) = {x R3 : hx p, ni = 0}.

In coordinates, a point (x, y, z) lies on P ((xp , yp , zp ), (xn , yn , zn )) exactly when

(x xp )xn + (y yp )yn + (z zp )zn = 0.

132 3 Linear Mappings and Their Matrices

p+d

Figure 3.23. Line in R3

Figure 3.24. Plane in R3

Exercises

3.10.1. Evaluate (2, 0, 1) (1, 3, 2).

3.10.2. Suppose that a vector v R3 takes the form v = u1 e1 = u2 e2
for some u1 and u2 . Describe v.
3.10.3. True or false: For all u, v, w in R3 , (u v) w = u (v w).
3.10.4. Express (u + v) (u v) as a scalar multiple of u v.
3.10.5. (a) Let U, V Mn (R) be skew-symmetric, meaning that U T = U
and similarly for V , where U T is the transpose of U (exercise 3.2.4). Show that
aU is skew-symmetric for any a R, and that U + V is skew-symmetric. Thus
the skew-symmetric matrices form a vector space. Show furthermore that the
commutant product (or Lie bracket) [U, V ] = U V V U is skew-symmetric.
One can optionally check that although the commutant product is not in
general associative, it instead satisfies the Jacobi identity,
[[U, [V, W ]] + [V, [W, U ]] + [W, [U, V ]] = 0.
(b) Encode the vectors u = (u1 , u2 , u3 ) and v = (v1 , v2 , v3 ) as 3-by-3
skew-symmetric matrices,
 3.10 The Cross Product, Lines, and Planes in R3 133

0 u1 u2 0 v1 v2
U = u1 0 u3 , V = v1 0 v3 .
u2 u3 0 v2 v3 0

Show that the commutant product [U, V ] encodes the cross product u v.

3.10.6. Investigate the extent to which a cancellation law holds for the cross
product, as follows: for fixed u, v in R3 with u 6= 0, describe the vectors w
satisfying the condition u v = u w.

3.10.7. What is the line specified by two points p and p ?

3.10.8. Give conditions on the points p, p and the directions d, d so that

(p, d) = (p , d ).

3.10.9. Express the relation between the coordinates of a point on (p, d) if

the x-component of d is 0.

3.10.10. What can you conclude about the lines

x xp y yp z zp x xp y yp z zp
= = and = =
xd yd zd xD yD zD

given that xd xD + yd yD + zd zD = 0? What can you conclude if xd /xD =

yd /yD = zd /zD ?

3.10.11. Show that (p, d) and (p , d ) intersect if and only if the linear equa-
tion Dt = p is solvable,
 where D M3,2 (R) has columns d and d , t is the
t1
column vector t2 , and p = p p. For what points p and p do (p, (1, 2, 2))
and (p , (2, 1, 4)) intersect?

3.10.12. Use vector geometry to show that the distance from the point q to
the line (p, d) is
|(q p) d|
.
|d|
(Hint: what is the area of the parallelogram spanned by q p and d?) Find
the distance from the point (3, 4, 5) to the line ((1, 1, 1), (1, 2, 3)).

3.10.13. Show that the time of nearest approach of two particles whose po-
sitions are s(t) = p + tv, s(t) = p + tv is t = hp, vi/|v|2 . (You may
assume that the particles are at their nearest approach when the difference of
their velocities is orthogonal to the difference of their positions.)

3.10.14. Write the equation of the plane through (1, 2, 3) with normal direc-
tion (1, 1, 1).

3.10.15. Where does the plane x/a + y/b + z/c = 1 intersect each axis?
134 3 Linear Mappings and Their Matrices

3.10.16. Specify the plane containing the point p and spanned by directions
d and d . Specify the plane containing the three points p, q, and r.

3.10.17. Use vector geometry to show that the distance from the point q to
the plane P (p, n) is
|hq p, ni|
.
|n|
(Hint: Resolve q p into components parallel and normal to n.) Find the
distance from the point (3, 4, 5) to the plane P ((1, 1, 1), (1, 2, 3)).

3.11 Summary

Linear algebra is easy in the abstract since the vector space operations pass
through linear mappings, and it is easy in the concrete since mechanical ma-
trix manipulations are straightforward. While the matrix methods from this
chapter are handy computational tools, it is also crucial to understand the
intrinsic notion of a linear mapping: this is the idea that we will use to define
multivariable differentiation in the next chapter.
4
The Derivative

In one-variable calculus the derivative is a limit of difference quotients, but

this idea does not generalize to many variables. The multivariable definition
of the derivative to be given in this chapter has three noteworthy features:
The derivative is defined as a linear mapping.
The derivative is characterized intrinsically rather than constructed in co-
ordinates.
The derivative is characterized by the property of closely approximating
the original mapping near the point of approximation.
Section 4.1 shows that the familiar definition of the one-variable derivative
can not scale up to many variables. Section 4.2 introduces a pan-dimensional
notation scheme that describes various closenesses of approximation. The no-
tation packages a range of ideas that arise in calculus, handling them uni-
formly. Section 4.3 revisits the one-variable derivative, rephrasing it in the
new scheme, and then scales it up to many variables. Handy basic properties
of the derivative follow immediately. Section 4.4 obtains some basic results
about the derivative intrinsically, notably the Chain Rule. Section 4.5 com-
putes with coordinates to calculate the derivative by considering one variable
at a time and using the techniques of one-variable calculus. This section also
obtains a coordinate-based version of the Chain Rule. Section 4.6 studies the
multivariable counterparts of higher order derivatives from one-variable calcu-
lus. Section 4.7 discusses optimization of functions of many variables. Finally,
section 4.8 discusses the rate of change of a function of many variables as its
input moves in any fixed direction, not necessarily parallel to a coordinate
axis.
136 4 The Derivative

4.1 Trying to Extend the Symbol-Pattern: Immediate,

Irreparable Catastrophe
In one variable calculus, the derivative of a function f : R R at a point
a R is defined as a limit,
f (a + h) f (a)
f (a) = lim .
h0 h
But for any integer n > 1, the corresponding expression makes no sense for a
mapping f : Rn Rm and for a point a of Rn . Indeed, the expression is
f (a + h) f (a)
lim ,
h0n h
but this is not even grammatically admissiblethere is no notion of division by
the vector h. That is, the standard definition of derivative does not generalize
to more than one input variable.
The breakdown here can not repaired by any easy patch. We must rethink
the derivative altogether in order to extend it to many variables.
Fortunately, the reconceptualization is richly rewarding.

Exercise

4.1.1. For a mapping f : Rn Rm and for a point a of Rn , the repair-

attempt of defining f (a) as
f (a + h) f (a)
lim
h0n |h|
is grammatically sensible. Does it reproduce the usual derivative if n = m = 1?

4.2 New Environment: the BachmannLandau Notation

The notation to be introduced in this section, originally due to Bachmann
late in the 19th century, was also employed by Landau. It was significantly
repopularized in the 1960s by Knuth in his famous computer science books,
and it is now integral to mathematics, computer science, and mathematical
statistics.
Definition 4.2.1 (o(1)-mapping, O(h)-mapping, o(h)-mapping). Con-
sider a mapping from some ball about the origin in one Euclidean space to a
second Euclidean space,
: B(0n , ) Rm
where n and m are positive integers and > 0 is a positive real number. The
mapping is smaller than order 1 if
 4.2 New Environment: the BachmannLandau Notation 137

for any c > 0, |(h)| c for all small enough h.

The mapping is of order h if

for some c > 0, |(h)| c|h| for all small enough h.

The mapping is smaller than order h if

for any c > 0, |(h)| c|h| for all small enough h.

A mapping smaller than order 1 is denoted o(1), and a mapping of order h is

denoted O(h), and a mapping smaller than order h is denoted o(h). Also o(1)
can denote the collection of o(1)-mappings, and similarly for O(h) and o(h).

The definition says that in terms of magnitudes, an o(1)-mapping is smaller

than any constant as h gets small, and an O(h)-mapping is at most some
constant multiple of h as h gets small, and an o(h)-mapping is smaller than
any constant multiple of h as h gets small. That is,


|o(1)| 0





|O(h)|

is bounded
|h| as h 0,




|o(h)|


0

|h|

but the definitions of O(h) and o(h) avoid the divisions in the previous display,
and the definitions further stipulate that any o(1)-mapping or O(h)-mapping
or o(h)-mapping takes the value 0 at h = 0. That is, beyond avoiding division,
the definitions are strictly speaking slightly stronger than the previous display.
Also, the definitions quickly give the containments

o(h) O(h) o(1),

meaning that every o(h)-mapping is an O(h)-mapping, and every O(h)-

mapping is an o(1)-mapping.
Visually, the idea is that:
For any c > 0, however small, close enough to the origin the graph of an
o(1)-mapping lies between the horizontal lines at height c, although the
requisite closeness of h to 0 can change as c gets smaller.
For some particular c > 0, close enough to the origin the graph of an
O(h)-mapping lies inside the bow-tie-shaped envelope determined by the
lines y = cx.
For any c > 0, however small, close enough to the origin the graph of
an o(h)-mapping lies inside the y = cx bow-tie, although the requisite
closeness of h to 0 can change as c gets smaller.
138 4 The Derivative

These images are oversimplified, representing a mappings n-dimensional

domain-ball and m-dimensional codomain-space as axes, but still the im-
ages correctly suggest that the o(1) condition describes continuity in local
coordinates, and the O(h) condition describes at-most-linear growth in local
coordinates, and the o(h) condition describes smaller-than-linear growth in
local coordinates. (A local coordinate system has its origin placed at some
particular point of interest, allowing us to assume that the point is simply the
origin.)
The next proposition gives the important basic example to have at hand.
Proposition 4.2.2 (Basic Family of Landau Functions). Consider the
function
e : Rn R, e (x) = |x|e (where e 0 is a real number).
Then
e is o(1) if e > 0,
e is O(h) if e 1,
e is o(h) if e > 1.
The proof is exercise 4.2.2. Examples are shown in figure 4.1.

1/2
1
3

Figure 4.1. Basic o(1), O(h), and o(h) functions

Since Definition 4.2.1 stipulates growth-bounds, the following result is im-

mediate.
Proposition 4.2.3 (Dominance Principle for the Landau Spaces). Let
be o(1), and suppose that |(h)| |(h)| for all small enough h. Then also
is o(1). And similarly for O(h) and for o(h).
For example, the function
(
h2 sin(1/h) if h 6= 0,
: R R, (h) =
0 if h = 0
 4.2 New Environment: the BachmannLandau Notation 139

is o(h) despite oscillating ever-faster as h approaches 0, because || |2 |

where 2 (h) = h2 is o(h) by Proposition 4.2.2. The reader should draw a
sketch of this situation.
Similarly, the functions , : R2 R given by

(h, k) = h, (h, k) = k

are O((h, k)) since the Size Bounds say that they are bounded absolutely
by the O(h)-mapping 1 (h, k) = |(h, k)|, i.e., |(h, k)| = |h| |(h, k)| and
similarly for . For general n and for any i {1, , n}, now letting h denote
a vector again as usual rather than the first component of a vector as it did
a moment ago, the ith component function

: Rn R, (h) = hi

is O(h) by the same argument. We will use this observation freely in the
sequel.
The o(1) and O(h) and o(h) conditions give rise to predictable closure
properties.
Proposition 4.2.4 (Vector Space Properties of the Landau Spaces).
For any fixed domain-ball B(0n , ) and codomain-space Rm , the o(1)-mappings
form a vector space, and O(h) forms a subspace, of which o(h) forms a sub-
space in turn. Symbolically,

o(1) + o(1) = o(1), R o(1) = o(1),

O(h) + O(h) = O(h), R O(h) = O(h),
o(h) + o(h) = o(h), R o(h) = o(h),

i.e., o(1) and O(h) and o(h) absorb addition and scalar multiplication.
The fact that o(1) forms a vector space encodes the rules that sums and
constant multiples of continuous mappings are again continuous.
Proof (Sketch). Consider any , o(1). For any c > 0,

|(h)| c/2 and |(h)| c/2 for all small enough h,

and so by the Triangle Inequality,

|( + )(h)| c for all small enough h.

(A fully-quantified version of the argument is as follows. Let c > 0 be given.

There exists > 0 such that |(h)| c/2 if |h| , and there exists > 0
such that |(h)| c/2 if |h| . Let = min{ , }. Then |( + )(h)| c
if |h| .) Similarly, for any nonzero R,

|(h)| c/|| for all small enough h,

140 4 The Derivative

so that since the modulus is absolute-homogeneous,

|()(h)| c for all small enough h.
If instead , O(h) then for all small enough h,
|(h)| c|h| and |(h)| c |h| for some c, c > 0,
so that for all small enough h,
|( + )(h)| (c + c )|h|.
Similarly, for any nonzero R, for all small enough h,
|()(h)| (||c)|h|.
The argument for o(h) is similar to the argument for o(1) (exercise 4.2.3).

For example, the function
: Rn R, (x) = 12|x|1/2 7|x| + 5|x|3/2
is an o(1)-function because all three of its terms are. It is not an O(h)-function
even though its second and third terms are, and it is not an o(h)-function even
though its third term is.
Another handy fact is the componentwise nature of the conditions o(1)
and O(h) and o(h). To see this, first note that any : B(0n , ) Rm is
o(1) if and only if the corresponding absolute value || : B(0n , ) R is.
Now let have component functions 1 , , m . For any h B(0n , ) and
for each j {1, , m}, the Size Bounds give
m
X
|j (h)| |(h)| |i (h)|.
i=1

Using the left side of the Size Bounds and then the vector space properties of
o(1) and then the right side of the Size Bounds,
m
X
|| is o(1) = each |j | is o(1) = |i | is o(1) = || is o(1).
i=1

Thus || is o(1) if and only if each |i | is. As explained just above, we may
drop the absolute values, and so in fact is o(1) if and only if each i is,
as desired. The arguments for the O(h) and o(h) conditions are the same
(exercise 4.2.4). The componentwise nature of the o(1) condition encodes the
componentwise nature of continuity.
The place of linear mappings in the Landau notation scheme is straight-
forward, affirming the previously-mentioned intuition that the O(h) condition
describes at-most-linear growth and the o(h) condition describes smaller-than-
linear growth.
 4.2 New Environment: the BachmannLandau Notation 141

Proposition 4.2.5. Every linear mapping is O(h). The only o(h) linear map-
ping is the zero mapping.

Proof. Let T : Rn Rm be a linear mapping. The unit sphere in Rn is

compact and T is continuous, so the image of the unit sphere under T is
again compact, hence bounded. That is, some positive c R exists such that
|T (ho )| c whenever |ho | = 1. The homogeneity of T shows that |T (h)| c|h|
for all nonzero h: letting ho = h/|h|,

|T (h)| = |T (|h|ho )| = | |h|T (ho ) | = |h| |T (ho )| c|h|.

And the inequality holds for h = 0 as well. Thus T is O(h).

Now assume that T is not the zero mapping. Thus T (ho ) is nonzero for
some nonzero ho , and we may take |ho | = 1. Let c = |T (ho )|/2, a positive
real number. For any scalar multiple h = ho of ho , however small, compute
(noting for the last step that |h| = ||)

|T (h)| = |T (ho )| = |T (ho )| = || |T (ho )| = 2c|| = 2c|h|.

That is, |T (h)| > c|h| for some arbitrarily small h-values, i.e., it is not the
case that |T (h)| c|h| for all small enough h. Thus T fails the o(h) definition
for the particular constant c = |T (ho )|/2.

For scalar-valued functions, a product property is useful to have at hand.

Proposition 4.2.6 (Product Property for Landau Functions). Con-

sider two scalar-valued functions and their product function,

, , : B(0n , ) R.

If is o(1) and is O(h) then is o(h). Especially, the product of two

linear functions is o(h).

Proof. Let c > 0 be given. For some d > 0, for all h close enough to 0n ,

|(h)| c/d and |(h)| d|h|,

and so
|()(h)| c|h|.
The second statement of the proposition follows from its first statement and
the previous proposition.

For two particular examples, consider the linear functions

1 , 2 : R2 R, 1 (h, k) = h, 2 (h, k) = k.

The proposition combines with the vector space properties of o(h, k) to say
that the functions
142 4 The Derivative

, : R2 R, (h, k) = h2 k 2 , (h, k) = hk

are both o(h, k).

Beyond their vector space properties, the Landau spaces carry composition
properties. If : B(0n , ) Rm and : B(0m , ) R are both o(1),
then after shrinking if necessary, the composition : B(0n , ) R is
also defined. That is, composition of o(1)-mappings is defined after suitably
shrinking a domain-ball. From now on we shrink domain-balls as necessary
without further comment.

Proposition 4.2.7 (Composition Properties of the Landau Spaces).

The composition of o(1)-mappings is again an o(1)-mapping. Also, the com-
position of O(h)-mappings is again an O(h)-mapping. Furthermore, the com-
position of an O(h)-mapping and an o(h)-mapping, in either order, is again
an o(h)-mapping. Symbolically,

o(o(1)) = o(1),
O(O(h)) = O(h),
o(O(h)) = o(h),
O(o(h)) = o(h).

That is, o(1) and O(h) absorb themselves, and o(h) absorbs O(h) from either
side.

The rule o(o(1)) = o(1) encodes the persistence of continuity under com-
position.

Proof. For example, to verify the third rule, suppose that : B(0n , ) Rm
is O(h) and that : B(0m , ) R is o(k). Then

for some c > 0, |(h)| c|h| for all small enough h.

Thus if h is small then so is (h), so that

for any d > 0, |((h))| d|(h)| for all small enough h.

Since c is some particular positive number and d can be any positive number,
cd again can be any positive number. That is, letting e = cd and combining
the previous two displays,

for any e > 0, |( )(h)| e|h| for all small enough h.

Hence is o(h) as desired. (A fully quantified version of the argument is

as follows. The hypotheses are that

there exist c > 0 and > 0 such that |(h)| c|h| if |h|

and that
 4.3 One-Variable Revisionism; the Derivative Redefined 143

for any d > 0 there exists d > 0 such that |(k)| d|k| if |k| d .

Now let e > 0 be given. Define d = e/c and e = min{, d /c}. Suppose that
|h| e . Then

|(h)| c|h| d since |h| and |h| d /c,

and so

|((h))| d|(h)| cd|h| since |(h)| d and |(h)| c|h|.

That is,
|((h))| e|h| since cd = e.
This shows that is o(h) since for any e > 0 there exists e > 0 such that
|( )(h)| e|h| if |h| e .)
The other rules are proved similarly (exercise 4.2.5).

Exercises

4.2.1. By analogy to Definition 4.2.1, give the appropriate definition of an

O(1)-mapping. What is the geometric interpretation of the definition? Need
an O(1)-mapping take 0 to 0?

4.2.2. Let e be a nonnegative real number. Consider the function

e : Rn R, (x) = |x|e .

(a) Suppose that e > 0. Let c > 0 be given. If |h| c1/e then what do we
know about |e (h)| in comparison to c? What does this tell us about e ?
(b) Prove that 1 is O(h).
(c) Suppose that e > 1. Combine parts (a) and (b) with the product
property for Landau functions (Proposition 4.2.6) to show that e is o(h).
(d) Explain how parts (a), (b), and (c) have proved Proposition 4.2.2.

4.2.3. Complete the proof of Proposition 4.2.4.

4.2.4. Establish the componentwise nature of the O(h) condition, and estab-
lish the componentwise nature of the o(h) condition.

4.2.5. Complete the proof of Proposition 4.2.7.

4.3 One-Variable Revisionism; the Derivative Redefined

The one-variable derivative as recalled at the beginning of the chapter,
f (a + h) f (a)
f (a) = lim ,
h0 h
144 4 The Derivative

is a construction. To rethink the derivative, we should characterize it instead.

To think clearly about what it means for the graph of a function to have
tangent slope t at a point (a, f (a)), we should work in local coordinates and
normalize to the case of a horizontal tangent. That is, given a function f of
x-values near some point a, and given a candidate tangent-slope t at (a, f (a)),
define a related function g of h-values near 0,

g(h) = f (a + h) f (a) th.

Thus g takes 0 to 0, and the graph of g near the origin is like the graph of f
near (a, f (a)) but with the line of slope t subtracted. To reiterate, the idea
that f has tangent slope t at (a, f (a)) has been normalized to the tidier idea
that g has slope 0 at the origin. Here the idea is:
To say that the graph of g is horizontal at the origin is to say that
for any positive real number c, however small, the region between the
lines of slope c contains the graph of g close enough to the origin.
That is:
The intuitive condition for the graph of g to be horizontal at the origin
is precisely that g is o(h). The horizontal nature of the graph of g at
the origin connotes that the graph of f has tangent slope t at (a, f (a)).
The symbolic connection between this characterization of the derivative
and the constructive definition is immediate. As always, the definition of f
having derivative f (a) at a is

f (a + h) f (a)
lim = f (a),
h0 h
which is to say,
f (a + h) f (a) f (a)h
lim = 0,
h0 h
and indeed this is precisely the o(h) condition on g. Figure 4.2 illustrates the
idea that when h is small, not only is the vertical distance f (a + h) f (a)
f (a)h from the tangent line to the curve small as well, but it is small even
relative to the horizontal distance h.
We need to scale these ideas up to many dimensions. Instead of viewing
the one-variable derivative as the scalar f (a), think of it as the corresponding
linear mapping Ta : R R, multiplication by f (a). That is, think of it as
the mapping
Ta (h) = f (a)h for all h R.
Figure 4.3 incorporates this idea. The figure is similar to figure 4.2, but it
shows the close approximation in the local coordinate system centered at the
point of tangency, and in the local coordinate system the tangent line is indeed
the graph of the linear mapping Ta . The shaded axis-portions in the figure
 4.3 One-Variable Revisionism; the Derivative Redefined 145

f (x)

f (a + h)
f (a) + f (a)h

f (a)
x
a a+h

Figure 4.2. Vertical distance from tangent line to curve

Ta (h)

f (a + h) f (a)
Ta (h)

h
h

Figure 4.3. Vertical distance in local coordinates

are h horizontally and g(h) = f (a + h) f (a) f (a)h vertically, and the fact
that the vertical portion is so much smaller illustrates that g(h) is o(h).
We are nearly ready to rewrite the derivative definition pan-dimensionally.
The small remaining preliminary matter is to take into account the local
nature of the characterizing condition: it depends on the behavior of f only
on an -ball about a, but on the other hand, it does require an entire -ball.
Thus the following definition is appropriate for our purposes.

Definition 4.3.1 (Interior Point). Let A be a subset of Rn , and let a be

a point of A. Then a is an interior point of A if some -ball about a is a
subset of A. That is, a is an interior point of A if B(a, ) A for some > 0.
146 4 The Derivative

Now we can define the derivative in a way that encompasses many variables
and is suitably local.
Definition 4.3.2 (Derivative). Let A be a subset of Rn , let f : A Rm
be a mapping, and let a be an interior point of A. Then f is differentiable
at a if there exists a linear mapping Ta : Rn Rm satisfying the condition
f (a + h) f (a) Ta (h) is o(h). (4.1)
This Ta is called the derivative of f at a, written Dfa or (Df )a . When f
is differentiable at a, the matrix of the linear mapping Dfa is written f (a)
and is called the Jacobian matrix of f at a.
Here are two points to note about Definition 4.3.2:
Again, any assertion that a mapping is differentiable at a point has the
connotation that the point is an interior point of the mappings domain.
That is, if f is differentiable at a then B(a, ) A for some > 0. In the
special case n = 1 we are disallowing the derivative at an endpoint of the
domain.
The domain of the linear mapping Ta is unrestricted even if f itself is
defined only locally about a. Indeed, the definition of linearity requires
that the linear mapping have all of Rn as its domain. Any linear mapping
is so uniform that in any case its behavior on all of Rn is determined by its
behavior on any -ball about 0n (exercise 4.3.1). In geometric terms, the
graph of T , the tangent object approximating the graph of f at (a, f (a)),
extends without bound, even if the graph of f itself is restricted to points
near (a, f (a)). But the approximation of the graph by the tangent object
needs to be close only near the point of tangency.
Returning to the idea of the derivative as a linear mapping, when n = 2
and m = 1 a function f : A R is differentiable at an interior point
(a, b) of A if for small scalar values h and k, f (a + h, b + k) f (a, b) is well
approximated by a linear function
T (h, k) = h + k
where and are scalars. Since the equation z = f (a, b) + h + k describes
a plane in (x, y, z)-space (where h = x a and k = y b), f is differentiable
at (a, b) if its graph has a well-fitting tangent plane through (a, b, f (a, b)).
(See figure 4.4.) Here the derivative of f at (a, b) is the linear mapping tak-
ing (h, k) to h + k and the Jacobian matrix of f at a is therefore [, ].
The tangent plane in the figure is not the graph of the derivative Df(a,b) ,
but rather a translation of the graph. Another way to say this is that the
(h, k, Df(a,b) (h, k))-coordinate system has its origin at the point (a, b, f (a, b))
in the figure.
When n = 1 and m = 3, a mapping f : A R3 is differentiable at an
interior point a of A if f (a + h) f (a) is closely approximated for small real h
by a linear mapping
 4.3 One-Variable Revisionism; the Derivative Redefined 147

f (x, y)
T (h, k)

k
h

x (a, b) y

Figure 4.4. Graph and tangent plane

T (h) = h

for some scalars , , and . As h varies through R, f (a) + T (h) traverses the
line = (f (a), (, , )) in R3 that ish tangent at f (a) to the output curve
i
of f . (See figure 4.5.) Here Dfa (h) = h and the corresponding Jacobian
hi
matrix is . Note that the figure does not show the domain of f , so it may
help to think of f as a time-dependent traversal of the curve rather than as
the curve itself. The figure does not have room for the (h, Dfa (h))-coordinate
system (which is 4-dimensional), but the Dfa (h)-coordinate system has its
origin at the point f (a).
For an example, let A = B((0, 0), 1) be the unit disk in R2 , and consider
the function
f : A R, f (x, y) = x2 y 2 .
We show that for any point (a, b) A, f is differentiable at (a, b) and its
derivative is the linear mapping

T(a,b) : R2 R, T(a,b) (h, k) = 2ah 2bk.

To verify this, we need to check Definition 4.3.2. The point that is written
in the definition intrinsically as a (where a is a vector) is written here in
coordinates as (a, b) (where a and b are scalars), and similarly the vector h in
the definition is written (h, k) here, because the definition is intrinsic whereas
here we are going to compute. To check the definition, first note that every
point (a, b) of A is an interior point; the fact that every point of A is interior
148 4 The Derivative

f (a)

Figure 4.5. Tangent to a parametrized curve

doesnt deserve a detailed proof right now, only a quick comment. Second,
confirm the derivatives characterizing property (4.1) by calculating that

f (a + h, b + k) f (a, b) T(a,b) (h, k)

= (a + h)2 (b + k)2 a2 + b2 2ah + 2bk
= h2 k 2 .

We saw immediately after the product property for Landau functions (Propo-
sition 4.2.6) that h2 k 2 is o(h, k). This is the desired result. Also, the calcula-
tion tacitly shows how the derivative was found for us to verify: the difference
f (a + h, b + k) f (a, b) is 2ah 2bk + h2 k 2 , which as a function of h and k
has a linear part 2ah 2bk and a quadratic part h2 k 2 that is much smaller
when h and k are small. The linear approximation of the difference is the
derivative.
Before continuing, we need to settle a grammatical issue. Definition 4.3.2
refers to any linear mapping that satisfies condition (4.1) as the derivative of f
at a. Fortunately, the derivative, if it exists, is unique, justifying the definite
article. The uniqueness is geometrically plausible: if two straight objects (e.g.,
lines or planes) approximate the graph of f well near (a, f (a)), then they
should also approximate each other well enough that straightness forces them
to coincide. The quantitative argument amounts to recalling that the only
linear o(h)-mapping is zero.

Proposition 4.3.3 (Uniqueness of the Derivative). Let f : A Rm

(where A Rn ) be differentiable at a. Then there is only one linear mapping
satisfying the definition of Dfa .

Proof. Suppose that the linear mappings Ta , Ta : Rn Rm are both deriva-

tives of f at a. Then the two mappings

f (a + h) f (a) Ta (h) and f (a + h) f (a) Ta (h)

 4.3 One-Variable Revisionism; the Derivative Redefined 149

are both o(h). By the vector space properties of o(h), so is their difference
(Ta Ta )(h). Since the linear mappings from Rn to Rm form a vector space
as well, the difference Ta Ta is linear. But the only o(h) linear mapping is
the zero mapping, so Ta = Ta as desired.

Finally, another result is immediate in our setup.

Proposition 4.3.4. If f is differentiable at a then f is continuous at a.

Proof. Compute, using the differentiability of f at a and the fact that linear
mappings are O(h), then the containment o(h) O(h) and the closure of O(h)
under addition, and finally the containment O(h) o(1), that

f (a+h)f (a) = f (a+h)f (a)Ta (h)+Ta (h) = o(h)+O(h) = O(h) = o(1).

Since the o(1) condition describes continuity, the argument is complete.

We will study the derivative via two routes. On the one hand, the linear
mapping Dfa : Rn Rm is specified by mn scalar entries of its matrix f (a),
and so calculating the derivative is tantamount to determining these scalars
by using coordinates. On the other hand, developing conceptual theorems
without getting lost in coefficients and indices requires the intrinsic idea of
the derivative as a well-approximating linear mapping.

Exercises

4.3.1. Let T : Rn Rm be a linear mapping. Show that for any > 0, the
behavior of T on B(0n , ) determines the behavior of T everywhere.

4.3.2. Give a geometric interpretation of the derivative when n = m = 2.

Give a geometric interpretation of the derivative when n = 1 and m = 2.

4.3.3. Let f : A Rm (where A Rn ) have component functions

f1 , , fm , and let a be an interior point of A. Let T : Rn Rm be a lin-
ear mapping with component functions T1 , , Tm . Using the componentwise
nature of the o(h) condition, established in section 4.2, show the componen-
twise nature of differentiability: f is differentiable at a with derivative T
if and only if each component fi is differentiable at a with derivative Ti .

4.3.4. Let f (x, y) = (x2 y 2 , 2xy). Show that Df(a,b) (h, k) = (2ah2bk, 2bh+
2ak) for all (a, b) R2 . (By the previous problem, you may work componen-
twise.)

4.3.5. Let g(x, y) = xey . Show that Dg(a,b) (h, k) = heb + kaeb for all (a, b)
R2 . (Note that because e0 = 1 and because the derivative of the exponential
function at 0 is 1, the one-variable characterizing property says that ek 1 =
k + o(k).)
150 4 The Derivative

4.3.6. Show that if f : Rn Rm satisfies |f (x)| |x|2 for all x Rn then

f is differentiable at 0n .
p
4.3.7. Show that the function f (x, y) = |xy| for all (x, y) R2 is not
differentiable at (0, 0). (First see what Df(0,0) (h, 0) and Df(0,0) (0, k) need to
be.)

4.4 Basic Results and the Chain Rule

Before constructing the derivative coordinatewise via the Jacobian matrix, we
derive some results intrinsically from its characterizing property. We begin by
computing two explicit derivatives.

Proposition 4.4.1 (Derivatives of Constant and Linear Mappings).

(1) Let C : A Rm (where A Rn ) be the constant mapping C(x) = c for
all x A, where c is some fixed value in Rm . Then the derivative of C at
any interior point a of A is the zero mapping.
(2) The derivative of a linear mapping T : Rn Rm at any point a Rn is
again T .

Proof. Both of these results hold essentially by grammar. In general the

derivative of a mapping f at a is the linear mapping that well approximates
f (a + h) f (a) for h near 0n . But C(a + h) C(a) is the zero mapping for
all h A, so it is well approximated near 0n by the zero mapping on Rn .
Similarly, T (a + h) T (a) is T (h) for all h Rn , and this linear mapping is
well approximated by itself near 0n .
To prove (1) more symbolically, let Z : Rn Rm denote the zero map-
ping, Z(h) = 0m for all h Rn . Then

C(a + h) C(a) Z(h) = c c 0 = 0 for all h Rn .

Being the zero mapping, C(a + h) C(a) Z(h) is crushingly o(h), showing
that Z meets the condition to be DCa . And (2) is similar (exercise 4.4.1).

Of course, differentiation passes through addition and scalar multiplication

of mappings.

Proposition 4.4.2 (Linearity of the Derivative). Let f : A Rm

(where A Rn ) and g : B Rm (where B Rn ) be mappings, and let a be
a point of A B. Suppose that f and g are differentiable at a with derivatives
Dfa and Dga . Then
(1) The sum f + g : A B Rm is differentiable at a with derivative
D(f + g)a = Dfa + Dga .
(2) For any R, the scalar multiple f : A Rm is differentiable at a
with derivative D(f )a = Dfa .
 4.4 Basic Results and the Chain Rule 151

The proof is a matter of seeing that the vector space properties of o(h)
encode the Sum Rule and Constant Multiple Rule for derivatives.
Proof. Since f and g are differentiable at a, some ball about a lies in A and
some ball about a lies in B. The smaller of these two balls lies in A B. That
is, a is an interior point of the domain of f + g. With this topological issue
settled, proving the proposition reduces to direct calculation. For (1),

(f + g)(a + h) (f + g)(a) (Dfa + Dga )(h)

= f (a + h) f (a) Dfa (h) + g(a + h) g(a) Dga (h)
= o(h) + o(h) = o(h).

And (2) is similar (exercise 4.4.2).

You may want to contrast how nicely our topological setup worked at
the beginning of this proof to the irritating example that we encountered in
connection with the Sum Rule for mappings back in section 2.5.
Elaborate mappings are built by composing simpler ones. The next theo-
rem is the important result that the derivative of a composition is the composi-
tion of the derivatives. That is, the best linear approximation of a composition
is the composition of the best linear approximations.
Theorem 4.4.3 (Chain Rule). Let f : A Rm (where A Rn ) be a
mapping, let B Rm be a set containing f (A), and let g : B R be a
mapping. Thus the composition gf : A R is defined. If f is differentiable
at the point a A, and g is differentiable at the point f (a) B, then the
composition g f is differentiable at the point a, and its derivative there is

D(g f )a = Dgf (a) Dfa .

In terms of Jacobian matrices, since the matrix of a composition is the product

of the matrices, the Chain Rule is

(g f ) (a) = g (f (a)) f (a).

The fact that we can prove that the derivative of a composition is the
composition of the derivatives without an explicit formula for the derivative
is akin to the fact in the previous chapter that we could prove that the deter-
minant of the product is the product of the determinants without an explicit
formula for the determinant.
Proof. To showcase the true issues of the argument clearly, we reduce the
problem to a normalized situation. For simplicity, we first take a = 0n and
f (a) = 0m . So we are given that

f (h) = S(h) + o(h),

g(k) = T (k) + o(k),
152 4 The Derivative

and we need to show that

(g f )(h) = (T S)(h) + o(h).

Compute that

g(f (h)) = g(Sh + o(h)) by the first given

= T Sh + T (o(h)) + o(Sh + o(h)) by the second.

We know that T k = O(k) and Sh = O(h), so the previous display gives

(g f )(h) = (T S)(h) + O(o(h)) + o O(h) + o(h) .

Since o(h) O(h) and O(h) is closed under addition, since o(h) absorbs O(h)
from either side, and since o(h) is closed under addition, the error (the last
two terms on the right side of the previous display) is


O(o(h)) + o O(h) + o(h) = O(o(h)) + o(O(h)) = o(h) + o(h) = o(h).

Therefore we have shown that

(g f )(h) = (T S)(h) + o(h),

exactly as desired. The crux of the matter is that o(h) absorbs O(h) from
either side.
For the general case, now longer assuming that a = 0n and f (a) = 0m , we
are given that

f (a + h) = f (a) + S(h) + o(h),

g(f (a) + k) = g(f (a)) + T (k) + o(k),

and we need to show that

(g f )(a + h) = (g f )(a) + (T S)(h) + o(h).

Compute that

g(f (a + h)) = g(f (a) + Sh + o(h)) by the first given

= g(f (a)) + T Sh + T (o(h)) + o(Sh + o(h)) by the second,

and from here the proof that the remainder term is o(h) is precisely as it is
in the normalized case.

Two quick applications of the Chain Rule arise naturally for scalar-valued
functions. Given two such functions, not only is their sum defined, but since R
is a field (unlike Rm for m > 1), so is their product and so is their quotient at
points where g is nonzero. With some help from the Chain Rule, the derivative
laws for product and quotient follow easily from elementary calculations.
 4.4 Basic Results and the Chain Rule 153

Lemma 4.4.4 (Derivatives of the Product and Reciprocal Func-

tions). Define the product function,

p : R2 R, p(x, y) = xy,

and define the reciprocal function

r : R {0} R, r(x) = 1/x.

Then
(1) The derivative of p at any point (a, b) R2 exists and is

Dp(a,b) (h, k) = ak + bh.

(2) The derivative of r at any nonzero real number a exists and is

Dra (h) = h/a2 .

Proof. (1) Compute,

p(a + h, b + k) p(a, b) ak bh = (a + h)(b + k) ab ak bh = hk.

By the Size Bounds |h| |(h, k)| and |k| |(h, k)|, so |hk| = |h| |k| |(h, k)|2 .
Since |(h, k)|2 is 2 (h, k) (where e is the example from Proposition 4.2.2), it
is o(h, k).
(2) is left as exercise 4.4.3.

Proposition 4.4.5 (Multivariable Product and Quotient Rules). Let

f : A R (where A Rn ) and g : B R (where B Rn ) be functions,
and let f and g differentiable at a. Then
(1) f g is differentiable at a with derivative

D(f g)a = f (a)Dga + g(a)Dfa .

(2) If g(a) 6= 0 then f /g is differentiable at a with derivative

 
f g(a)Dfa f (a)Dga
D = .
g a g(a)2

Proof. (1) As explained in the proof of Proposition 4.4.2, a is an interior point

of the domain A B of f g, so we need only to compute. The product function
f g is the composition p (f, g), where (f, g) : A B R2 is the mapping
with component functions f and g. For any h Rn , the Chain Rule and the
componentwise nature of differentiation (this was exercise 4.3.3) give


D(f g)a (h) = D(p (f, g))a (h) = Dp(f,g)(a) D(f, g)a (h)
= Dp(f (a),g(a)) (Dfa (h), Dga (h)),
154 4 The Derivative

and by the previous lemma,

Dp(f (a),g(a)) (Dfa (h), Dga (h)) = f (a)Dga (h) + g(a)Dfa (h)
= (f (a)Dga + g(a)Dfa )(h).

This proves (1) since h is arbitrary. (2) is similar (exercise 4.4.4) but with the
wrinkle that one needs to show that since g(a) 6= 0 and since Dga exists, it
follows that a is an interior point of the domain of f /g. Here it is relevant that
g must be continuous at a, and so by the Persistence of Inequality principle
(Proposition 2.3.10), g is nonzero on some -ball at a as desired.

With the results accumulated so far, we can compute the derivative of

any mapping whose component functions are given by rational expressions in
its component input scalars. By the componentwise nature of differentiabil-
ity, it suffices to find the derivatives of the component functions. Since these
are compositions of sums, products, and reciprocals of constants and linear
functions, their derivatives are calculable with the existing machinery.
Suppose, for instance, that f (x, y) = (x2 y)/(y + 1) for all (x, y) R2
such that y 6= 1. Note that every point of the domain of f is an interior
point. Rewrite f as
X2 Y
f=
Y +1
where X is the linear function X(x, y) = x on R2 and similarly Y (x, y) = y.
Applications of the Chain Rule and virtually every other result on derivatives
so far shows that at any point (a, b) in the domain of f , the derivative Df(a,b)
is given by (justify the steps)

Df(a,b) (h, k)
(Y + 1)(a, b)D(X 2 Y )(a,b) (X 2 Y )(a, b)D(Y + 1)(a,b)
= (h, k)
((Y + 1)(a, b))2
(b + 1)(D(X 2 )(a,b) DY(a,b) ) (a2 b)(DY(a,b) + D1(a,b) )
= (h, k)
(b + 1)2
(b + 1)(2X(a, b)DX(a,b) Y ) (a2 b)Y
= (h, k)
(b + 1)2
(b + 1)(2aX Y ) (a2 b)Y
= (h, k)
(b + 1)2
(b + 1)(2ah k) (a2 b)k
=
(b + 1)2
2a a2 + 1
= h k.
b+1 (b + 1)2

In practice this method is too unwieldy for any functions beyond the simplest,
and in any case it applies only to mappings with rational component functions.
 4.4 Basic Results and the Chain Rule 155

But on the other hand, there is no reason to expect much in the way of
computational results from our methods so far, since we have been studying
the derivative based on its intrinsic characterization. In the next section we
will construct the derivative in coordinates, enabling us to compute easily by
drawing on the results of one-variable calculus.
For another application of the Chain Rule, let A and B be subsets of Rn ,
and suppose that f : A B is invertible with inverse g : B A. Suppose
further that f is differentiable at a A and that g is differentiable at f (a).
The composition g f is the identity mapping idA : A A which, being the
restriction of a linear mapping, has the linear mapping as its derivative at a.
Therefore,
id = D(idA )a = D(g f )a = Dgf (a) Dfa .
This argument partly shows that for invertible f as described, the linear map-
ping Dfa is also invertible. (A symmetric argument completes the proof by
showing that also id = Dfa Dgf (a) .) Since we have methods available to
check the invertibility of a linear map, we can apply this criterion once we
know how to compute derivatives.
Not too much should be made of this result, however; its hypotheses are
too strong. Even in the one-variable case the function f (x) = x3 from R
to R is invertible
and yet has the noninvertible derivative 0 at x = 0. (The
inverse, g(x) = 3 x is not differentiable at 0, so the conditions above are not
met.) Besides, we would prefer a converse statement, that if the derivative is
invertible then so is the mapping. The converse statement is not true, but we
will see in chapter 5 that it is locally true, i.e., it is true in the small.

Exercises

4.4.1. Prove part (2) of Proposition 4.4.1.

4.4.2. Prove part (2) of Proposition 4.4.2.

4.4.3. Prove part (2) of Lemma 4.4.4.

4.4.4. Prove the Quotient Rule.

4.4.5. Let f (x, y, z) = xyz. Find Df(a,b,c) for arbitrary (a, b, c) R3 . (Hint:
f is the product XY Z where X is the linear function X(x, y, z) = x and
similarly for Y and Z.)

4.4.6. Define f (x, y) = xy 2 /(y 1) on {(x, y) R2 : y 6= 1}. Find Df(a,b)

where (a, b) is a point in the domain of f .

4.4.7. (A generalization of the product rule.) Recall that a function

f : Rn Rn R

is called bilinear if for all x, x , y, y Rn and all R,

156 4 The Derivative

f (x + x , y) = f (x, y) + f (x , y),
f (x, y + y ) = f (x, y) + f (x, y ),
f (x, y) = f (x, y) = f (x, y).

(a) Show that if f is bilinear then f (h, k) is o(h, k).

(b) Show that if f is bilinear then f is differentiable with Df(a,b) (h, k) =
f (a, k) + f (h, b).
(c) What does this exercise say about the inner product?
4.4.8. (A bigger generalization of the product rule.) A function

f : Rn Rn R

(there are k copies of Rn ) is called multilinear if for each j {1, , k}, for
all x1 , , xj , xj , , xk Rn and all R,

f (x1 , , xj + xj , , xk ) = f (x1 , , xj , , xk ) + f (x1 , , xj , , xk )

f (x1 , , xj , , xk ) = f (x1 , , xj , , xk ).

(a) Show that if f is multilinear and a1 , , ak , h1 , , hk Rn then

for i, j {1, , k} (distinct), f (a1 , , hi , , hj , , ak ) is o(h1 , , hk ).
(Use the previous problem.)
(b) Show that if f is multilinear then f is differentiable with
k
X
Df(a1 , ,ak ) (h1 , , hk ) = f (a1 , , aj1 , hj , aj+1 , , ak ).
j=1

(c) When k = n, what does this exercise say about the determinant?

4.5 Calculating the Derivative

Working directly from Definition 4.3.2 of the multivariable derivative with-
out using coordinates has yielded some easy results and one harder onethe
Chain Rulebut no explicit description of the derivative except in the sim-
plest cases. We dont even know that any multivariable derivatives exist except
for mappings with rational coefficient functions.
Following the general principle that necessary conditions are more easily
obtained than sufficient ones, we assume that the derivative exists and de-
termine what it then must be. Geometry provides the insight. By the usual
componentwise argument, there is no loss in studying a function f with scalar
output, i.e., we may take m = 1. Setting n = 2 fits the graph of f in R3 where
we can see it. Thus take f : A R where A R2 .
Suppose that f is differentiable at the point (a, b). Then the graph of f
has a well-fitting tangent plane P at the point (a, b, f (a, b)), as shown ear-
lier in figure 4.4. To determine this plane, we need two of its lines through
 4.5 Calculating the Derivative 157

f (x, y)

x y
(a, b)

Figure 4.6. Cross-sectional lines

(a, b, f (a, b)). The natural lines to consider are those whose (x, y)-shadows run
in the x and y directions. Call them x and y . (See figure 4.6.)
The line x is tangent to a cross section of the graph of f . To see this cross
section, freeze the variable y at the value b and look at the resulting function
of one variable, (x) = f (x, b). The slope of x in the vertical (x, b, z)-plane
is precisely (a). A small technicality here is that since (a, b) is an interior
point of A, also a is an interior point of the domain of .
Similarly, y has slope (b) where (y) = f (a, y). The linear function
approximating f (a + h, b + k) f (a, b) for small (h, k) is now specified as
T (h, k) = (a)h + (b)k. Thus Df(a,b) has matrix [ (a) (b)]. Since the
entries of this matrix are simply one variable derivatives, this is something
that we can compute.

Definition 4.5.1 (Partial Derivative). Let A be a subset of Rn , let f :

A R be a function, and let a = (a1 , , an ) be an interior point of A. Fix
j {1, , n}. Define

(t) = f (a1 , , aj1 , t, aj+1 , , an ) for t near aj .

Then the jth partial derivative of f at a is defined as

Dj f (a) = (aj )

if (aj ) exists. Here the prime signifies ordinary one variable differentiation.
Equivalently,
f (a + tej ) f (a)
Dj f (a) = lim
t0 t
if the limit exists and it is not being taken at an endpoint of the domain of
the difference quotient.
158 4 The Derivative

Partial derivatives are easy to compute: fix all but one of the variables,
and then take the one-variable derivative with respect to the variable that
remains. For example if

f (x, y, z) = ey cos x + z

then
d b 
D1 f (a, b, c) = (e cos x + c)x=a = eb sin a,
dx
D2 f (a, b, c) = eb cos a,
D3 f (a, b, c) = 1.

Theorem 4.5.2 (The Derivative in Coordinates: Necessity). Let the

mapping f : A Rm (where A Rn ) be differentiable at the point a
A. Then for each i {1, , m} and j {1, , n}, the partial derivative
Dj fi (a) exists. Furthermore, each Dj fi (a) is the (i, j)th entry of the Jacobian
matrix of f at a. Thus the Jacobian matrix is

D1 f1 (a) Dn f1 (a)
D1 f2 (a) Dn f2 (a)

f (a) = .. . .. = [Dj fi (a)]i=1, ,m .
. . . . j=1, ,n
D1 fm (a) Dn fm (a)

Proof. The idea is to read off the (i, j)th entry of f (a) by studying the ith
component function of f and letting h 0n along the jth coordinate direction
in the defining property (4.1) of the derivative. The ensuing calculation will
repeat the quick argument in section 4.3 that the characterization of the
derivative subsumes the construction in the one-variable case.
The derivative of the component function fi at a is described by the ith
row of f (a). Call the row entries di1 , di2 , , din . Since linear of is matrix
times, it follows that

(Dfi )a (tej ) = dij t for all t R.

Let h = tej with t a variable real number, so that h 0n as t 0R . Since

(Dfi )a exists, we have as a particular instance of the characterizing property
that fi (a + h) fi a) (Dfi )a (h) is o(h),

|fi (a + tej ) fi (a) (Dfi )a (tej )|

0 = lim
t0 |tej |
 
 fi (a + tej ) fi (a) dij t 
= lim  

t0 t
 
 fi (a + tej ) fi (a) 
= lim  dij  .
t0 t
 4.5 Calculating the Derivative 159

That is,
fi (a + tej ) fi (a)
lim = dij .
t0 t
The previous display says precisely that Dj fi (a) exists and equals dij .

So the existence of the derivative Dfa makes necessary the existence of all
partial derivatives of all component functions of f at a. The natural question
is whether their existence is also sufficient for the existence of Dfa . It is
not. The proof of Theorem 4.5.2 was akin to the Straight Line Test from
section 2.3: the general condition h 0n was specialized to h = tej , i.e.,
to letting h approach 0n only along the axes. The specialization let us show
that the derivative matrix entries are the partial derivatives of the component
functions of f . But the price for this specific information was loss of generality,
enough loss that the derived necessary conditions are not sufficient.
For example, the function
(
2xy
2 2 2 if (x, y) 6= (0, 0),
f : R R, f (x, y) = x +y
0 if (x, y) = (0, 0)
has for its first partial derivative at the origin
f (t, 0) f (0, 0) 00
D1 f (0, 0) = lim = lim = 0,
t0 t t0 t
and similarly D2 f (0, 0) = 0; but as discussed in chapter 2, f is not continuous
at the origin, much less differentiable there. However, this example is con-
trived, the sort of function that one sees only in a mathematics class, and in
fact with slightly stronger hypotheses a statement in the spirit of the converse
to Theorem 4.5.2 does hold.
Theorem 4.5.3 (The Derivative in Coordinates: Sufficiency). Let f :
A Rm (where A Rn ) be a mapping, and let a be an interior point
of A. Suppose that for each i {1, , m} and j {1, , n}, the partial
derivative Dj fi exists not only at a but at all points in some -ball about a,
and the partial derivative Dj fi is continuous at a. Then f is differentiable
at a.
Note that if f meets the conditions of Theorem 4.5.3 (all partial derivatives
of all component functions of f exist at and about a, and they are continuous
at a) then the theorems conclusion (f is differentiable at a) is the condition
of Theorem 4.5.2, so that the latter theorem tells us the derivative of f (the
entries of its matrix are the partial derivatives). But the example given just be-
fore Theorem 4.5.3 shows that the converse fails: even if all partial derivatives
of all component functions of f exist at a then f need not be differentiable
at a.
The difference between the necessary conditions in Theorem 4.5.2 and the
sufficient conditions in Theorem 4.5.3 has a geometric interpretation when
n = 2 and m = 1. The necessary conditions in Theorem 4.5.2 are:
160 4 The Derivative

If a graph has a well-fitting plane at some point, then at that point

we see well-fitting lines in the cross sections parallel to the coordinate
axes.
The sufficient conditions in Theorem 4.5.3 are:
If a graph has well-fitting lines in the cross sections at and near the
point, and if those lines dont change much as we move among cross
sections at and near the point, then the graph has a well-fitting plane.
But well-fitting cross sectional lines at the point are not enough to guarantee
a well-fitting plane at the point. The multivariable derivative is truly a pan-
dimensional construct, not just an amalgamation of cross sectional data.
Proof. It suffices to show the differentiability of each component function fi ,
so we may assume that m = 1, i.e., that f is scalar-valued. To thin out the
notation, the proof will be done for n = 3 (so for example a = (a1 , a2 , a3 )),
but its generality should be clear.
Theorem 4.5.2 says that if the derivative Dfa exists then it is defined by
the matrix of partial derivatives Dj f (a). The goal therefore is to show that
the linear mapping
Ta (h1 , h2 , h3 ) = D1 f (a)h1 + D2 f (a)h2 + D3 f (a)h3
satisfies the defining property of the derivative. That is, we need to show that
f (a + h) f (a) = D1 f (a)h1 + D2 f (a)h2 + D3 f (a)h3 + o(h).
We may take h small enough that the partial derivatives Dj f exist at all
points within distance |h| of a. Here we use the hypothesis that the partial
derivatives exist everywhere near a.
The first idea is to move from a to a + h in steps, changing one coordinate
at a time,
f (a + h) f (a) = f (a1 + h1 , a2 + h2 , a3 + h3 ) f (a1 , a2 + h2 , a3 + h3 )
+ f (a1 , a2 + h2 , a3 + h3 ) f (a1 , a2 , a3 + h3 )
+ f (a1 , a2 , a3 + h3 ) f (a1 , a2 , a3 ).
Since the partial derivatives exist we may apply the Mean Value Theorem in
two directions and the one-variable derivatives characterizing property in the
third,
f (a + h) f (a) = D1 f (a1 + c1 , a2 + h2 , a3 + h3 )h1
+ D2 f (a1 , a2 + c2 , a3 + h3 )h2
+ D3 f (a1 , a2 , a3 )h3 + o(h3 ),
where |ci | |hi | for i = 1, 2. Since D1 f and D2 f are continuous at the
point a = (a1 , a2 , a3 ), and since the condition h 03 squeezes each hi and ci
to 0,
 4.5 Calculating the Derivative 161

D1 f (a1 + c1 , a2 + h2 , a3 + h3 ) = D1 f (a) + o(1),

D2 f (a1 , a2 + c2 , a3 + h3 ) = D2 f (a) + o(1).

Also, o(1)hi = o(h) for i = 1, 2 and o(h3 ) = o(h), and so altogether we have

f (a + h) f (a) = D1 f (a)h1 + D2 f (a)h2 + D3 f (a)h3 + o(h).

This is the desired result.

Thus, to reiterate some earlier discussion and to amplify slightly,

the differentiability of f at a implies the existence of all the partial deriva-
tives at a, and the partial derivatives are the entries of the derivative
matrix,
while the existence of all the partial derivatives at and about a, and their
continuity at a, combine to imply the differentiability of f at a,
but the existence of all partial derivatives at a need not imply the differ-
entiability of f at a.
And in fact the previous proof shows that we need to check the scope and
continuity only of all but one of the partial derivatives. The proof used
the existence of D3 f at a but not its existence near a or its continuity
at a, and a variant argument or a reindexing shows that nothing is special
about the last variable. This observation is a bit of a relief, telling us that
in the case of one input variable our methods do not need to assume that
the derivative exists at and about a point and is continuous at the point
in order to confirm merely that it exists at the point. We codify this bullet
as a variant sufficiency theorem:

Theorem 4.5.4 (The Derivative in Coordinates: Sufficiency). Let f :

A Rm (where A Rn ) be a mapping, and let a be an interior point of A.
Suppose that for each i {1, , m},
for all j {1, , n} the partial derivative Dj fi (a) exists,
while for all but at most one j {1, , m} the partial derivatives Dj fi
exist in some -ball about a and are continuous at a.
Then f is differentiable at a.

Note how all this compares to the discussion of the determinant in the
previous chapter. There we wanted the determinant to satisfy characterizing
properties, we found the only function that could possibly satisfy them, and
then we verified that it did. Here we wanted the derivative to satisfy a char-
acterizing property, and we found the only possibility for the derivativethe
linear mapping whose matrix consists of the partial derivatives, which must
exist if the derivative does. But analysis is more subtle than algebra: this linear
mapping need not satisfy the characterizing property of the derivative unless
we add further assumptions. The derivative-existence theorem, Theorem 4.5.3
or the slightly stronger Theorem 4.5.4, is the most substantial result so far
162 4 The Derivative

in this chapter. We have already seen a counterexample to the converse of

Theorem 4.5.3, in which the function had partial derivatives but wasnt differ-
entiable because it wasnt even continuous (page 159). For a one-dimensional
counterexample to the converse of Theorem 4.5.3, in which the derivative ex-
ists but is not continuous, see exercise 4.5.3. The example in the exercise does
not contradict the weaker converse of the stronger Theorem 4.5.4.
To demonstrate the ideas of this section so far, consider the function
( 2
x y
2 2 if (x, y) 6= (0, 0),
f (x, y) = x +y
0 if (x, y) = (0, 0).

The top formula in the definition describes a rational function of x and y

on the punctured plane R2 {(0, 0)}. Any rational function and all of its
partial derivatives are continuous on its domain (feel free to invoke this result),
and furthermore every point (a, b) away from (0, 0) lies in some -ball that
is also away from (0, 0). That is, for any point (a, b) 6= (0, 0) the partial
derivatives of f exist at and about (a, b) and they are continuous at (a, b).
Thus the conditions for Theorem 4.5.3 are met, and so its conclusion follows:
f is differentiable at (a, b). Now Theorem 4.5.2 says that the derivative matrix
at (a, b) is the matrix of partial derivatives,
 
  2ab3 a2 (a2 b2 )
f (a, b) = D1 f (a, b) D2 f (a, b) = .
(a2 + b2 )2 (a2 + b2 )2

Consequently the derivative of f at any nonzero (a, b) is the corresponding

linear map
2ab3 a2 (a2 b2 )
Df(a,b) (h, k) = 2 2 2
h+ k.
(a + b ) (a2 + b2 )2
However, this analysis breaks down at the point (a, b) = (0, 0). Here our only
recourse is to figure out whether a candidate derivative exists and then test
whether it works. The first partial derivative of f at (0, 0) is

f (t, 0) f (0, 0) 00
D1 f (0, 0) = lim = lim = 0,
t0 t t0 t
and similarly D2 f (0, 0) = 0. So by Theorem 4.5.2, the only possibility for the
derivative of f at (0, 0) is the zero mapping. Now the question is:

Is f (h, k) f (0, 0) 0 o(h, k)?

Since the denominator h2 + k 2 of f away from the origin is |(h, k)|2 ,

|h|2 |k|
|f (h, k) f (0, 0) 0| = |f (h, k)| = .
|(h, k)|2

Let (h, k) approach 02 along the line h = k. Since |h| = |(h, h)|/ 2,
 4.5 Calculating the Derivative 163

|h|3 |(h, h)|

|f (h, h) f (0, 0) 0| = = .
|(h, h)|2 2 2
Thus along this line the condition |f (h, k)f (0, 0)0| c|(h, k)| fails for (say)
c = 1/4, and so f (h, k)f (0, 0)0 is not o(h, k). That is, the function f is not
differentiable at (0, 0). And indeed, the graph of f near (0, 0) shows a surface
that isnt well approximated by any plane through its center, no matter how
closely we zoom in. (See figure 4.7. The figure shows that the cross-sectional
slopes over the axes are 0 while the cross-sectional slopes over the diagonals
are not, confirming our symbolic calculations.) Here we have used the straight
line test to get a negative answer; but recall that the straight line test alone
can not give a positive answer, so the method here would need modification
to show that a function is differentiable.

Figure 4.7. The crimped sheet is differentiable everywhere except at the origin

For another example, exercise 4.3.4 used the characterizing property to

confirm the derivative of the function f (x, y) = (x2 y 2 , 2xy). Now we can
obtain the derivative and know that it works by using the theorems of this
section. The function f has domain R2 , so every domain point is interior.
Since each component of f is a polynomial, so are all partial derivatives of the
components, making them continuous everywhere. Thus f is differentiable at
any point (a, b) R2 . The matrix of partial derivatives at (a, b) is
   
D1 f1 (a, b) D2 f1 (a, b) 2a 2b
= ,
D1 f2 (a, b) D2 f2 (a, b) 2b 2a

and so the derivative of f at (a, b) is, as before,

164 4 The Derivative

Df(a,b) (h, k) = (2ah 2bk, 2bh + 2ak).

Similarly, the function g(x, y) = xey from exercise 4.3.5 has domain R2 ,
all of whose points are interior, and its partial derivatives D1 g(x, y) = ey and
D2 g(x, y) = xey are continuous everywhere. Thus it is differentiable every-
where. Its matrix of partial derivatives at any point (a, b) is

[D1 g(a, b) D2 g(a, b)] = [eb aeb ],

and so its derivative at (a, b) is

Dg(a,b) (h, k) = eb h + aeb k.

The reader is encouraged to reproduce the derivative of the product func-

tion (Lemma 4.4.4, part (1)) similarly.
Returning to the discussion (at the end of the previous section) of invert-
ibility of a mapping and invertibility of its derivative, consider the mapping

f : R2 {(0, 0)} R2 {(0, 0)}, f (x, y) = (x2 y 2 , 2xy).

At any (x, y) where f is defined, the partial derivatives are D1 f1 (x, y) =

2x, D2 f1 (x, y) = 2y, D1 f2 (x, y) = 2y, and D2 f2 (x, y) = 2x. These are
continuous functions of (x, y), so for any (a, b) 6= (0, 0), Df(a,b) exists and its
matrix is    
D1 f1 (a, b) D2 f1 (a, b) 2a 2b
f (a, b) = = .
D1 f2 (a, b) D2 f2 (a, b) 2b 2a
The matrix has determinant 4(a2 + b2 ) > 0, and hence it is always invertible.
On the other hand, the mapping f takes the same value at points (x, y)
and (x, y), so it is definitely not invertible.
With the Jacobian matrix described explicitly, a more calculational version
of the Chain Rule is available.

Theorem 4.5.5 (Chain Rule in Coordinates). Let f : A Rm (where

A Rn ) be differentiable at the point a of A, and let g : f (A) R be
differentiable at the point b = f (a). Then the composition g f : A R is
differentiable at a, and its partial derivatives are
m
X
Dj (g f )i (a) = Dk gi (b)Dj fk (a) for i = 1, , , j = 1, , n.
k=1

Proof. The composition is differentiable by the intrinsic Chain Rule. The Ja-
cobian matrix of g at b is

g (b) = [Dk gi (b)]m (row index i, column index k),

and the Jacobian matrix of f at a is

 4.5 Calculating the Derivative 165

f (a) = [Dj fk (a)]mn (row index k, column index j),

and the Jacobian matrix of g f at a is

(g f ) (a) = [Dj (g f )i (a)]n (row index i, column index j).

By the intrinsic Chain Rule,

(g f ) (a) = g (b)f (a).

Equate the (i, j)th entries to obtain the result.

Notations for the partial derivative vary. A function is often described by

a formula such as w = f (x, y, z). Other notations for D1 f are

f w
f1 , fx , , wx , .
x x
If x, y, z are in turn functions of s and t then a classical formulation of the
Chain Rule would be
w w x w y w z
= + + . (4.2)
t x t y t z t
The formula is easily visualized as chasing back along all dependency chains
from t to w in a diagram where an arrow means contributes to:

7 E x




s




8
&
y /w
qqqq @
q
qqqq
qqqq
t



& 
z

Unfortunately, for all its mnemonic advantages, the classical notation is a

veritable minefield of misinterpretation. Formula (4.2) doesnt indicate where
the various partial derivatives are to be evaluated, for one thing. Specifying the
variable of differentiation by name rather than by position also becomes con-
fusing when different symbols are substituted for the same variable, especially
since the symbols themselves may denote specific values or other variables.
For example one can construe many different meanings for the expression
166 4 The Derivative

f
(y, x, z).
x
Blurring the distinction between functions and the variables denoting their
outputs is even more problematic. If one has, say, z = f (x, t, u), x = g(t, u),

t




'
x 7/ z
>



u

then chasing all paths from z back to t gives

z z x z
= +
t x t t
with z/t meaning something different on each side of the equality. While
the classical formulas are useful and perhaps simpler to apply in elementary
situations, they are not particularly robust until one has a solid understand-
ing of the Chain Rule. On the other hand, the classical formulas work fine
in straightforward applications, so several exercises are phrased in the older
language to give you practice with it.
For example, let

(x, y) = f (r, ) = (r cos , r sin ),

(z, w) = g(x, y) = (x2 y 2 , 2xy).

We compute (z/r)(2, /3). The Chain Rule in coordinates gives

     
z/r z/ z/x z/y x/r x/
= ,
w/r w/ w/x w/y y/r y/

and the upper left entry is

z z x z y
= + = 2x cos 2y sin .
r x r y r

We are given (r, ) = (2, /3), and it follows that (x, y) = (1, 3). So the
answer is
z 1 3
(2, /3) = 2 1 2 3 = 2 .
r 2 2
To confirm the result without using the Chain Rule, note that f is the polar-
to-cartesian change of coordinates, and g is the complex squaring function in
cartesian coordinates, so that the composition g f is the squaring function
in polar coordinates. That is, the composition is
 4.5 Calculating the Derivative 167

(z, w) = (g f )(r, ) = (r2 cos 2, r2 sin 2).

Consequently z/r = 2r cos 2, and substituting (r, ) = (2, /3) gives in

particular (z/r)(2, /3) = 2 2 cos 2/3 = 2 2 (1/2) = 2, as we know
it must.

Exercises

4.5.1. Explain why in the discussion beginning this section the tangent
plane P consists of all points (a, b, f (a, b)) + (h, k, T (h, k)) where T (h, k) =
(a)h + (b)k.
4.5.2. This exercise shows that all partial derivatives of a function can exist at
and about a point without being continuous at the point. Define f : R2 R
by (
2xy
2 2 if (x, y) 6= (0, 0),
f (x, y) = x +y
0 if (x, y) = (0, 0).
(a) Show that D1 f (0, 0) = D2 f (0, 0) = 0.
(b) Show that D1 f (a, b) and D2 f (a, b) exist and are continuous at all other
(a, b) R2 .
(c) Show that D1 f and D2 f are discontinuous at (0, 0).
4.5.3. Define f : R R by
(
x2 sin x1 if x 6= 0,
f (x) =
0 if x = 0.

Show that f (x) exists for all x but that f is discontinuous at 0. Explain how
this disproves the converse of Theorem 4.5.3.
4.5.4. Discuss the derivatives of the following mappings at the following
points.
2
y
(a) f (x, y) = xy+1 on {(x, y) R2 : y 6= 1} at generic (a, b) with b 6= 1.
(After you are done, compare the effort of doing the problem now to the effort
of doing it as we did at the end of section 4.4.)
xy 2
(b) f (x, y) = y1 on {(x, y) R2 : y 6= 1} at generic (a, b) with b 6= 1.
( xy
2 2 if (x, y) 6= (0, 0)
(c) f (x, y) = x +y at generic (a, b) 6= (0, 0) and at
0 if (x, y) = (0, 0)
(0, 0).
For the rest of these exercises, assume as much differentiability as necessary.
4.5.5. For what differentiable mappings f : A Rm is f (a) a diagonal
matrix for all a A? (A diagonal matrix is a matrix whose (i, j)th entries for
all i 6= j are 0.)
168 4 The Derivative

4.5.6. Show that if z = f (xy) then x, y, and z satisfy the differential equation
x zx y zy = 0.
4.5.7. Let w = F (xz, yz). Show that x wx + y wy = z wz .
4.5.8. If z = f (ax + by), show that bzx = azy .
4.5.9. The function f : R2 R is called homogeneous of degree k if
f (tx, ty) = tk f (x, y) for all scalars t and vectors (x, y). Letting f1 and f2
denote the first and second partial derivatives of f , show that such f satisfies
the differential equation

xf1 (x, y) + yf2 (x, y) = kf (x, y).

(Hint: First differentiate the homogeneity condition with respect to t, viewing

x and y as fixed but generic; the derivative of one side will require the chain
rule. Second, since the resulting condition holds for all scalars t, it holds for
any particular t of your choosing.)
4.5.10. Let
f : R2 R
be a function such that for all (x, y) R2 , the integral
Z y
F : R2 R, F (x, y) = f (x, v) dv
v=0

exists and is differentiable with respect to x, its partial derivative with respect
to x being obtained by passing the x-derivative through the v-integral,
Z y
F (x, y)
= f (x, v) dv
x x v=0
Ry Ry
v=0
f (x + h, v) dv v=0 f (x, v) dv
= lim
h0 h
Z y
f (x + h, v) f (x, v)
= lim dv
h0 v=0 h
Z y
! f (x + h, v) f (x, v)
= lim dv
h0 h
Zv=0
y
f
= (x, v) dv.
v=0 x

(The ! step requires justification, but under reasonable circumstances it can

be carried out.) Define a function
Z x
G : R R, G(x) = f (x, v) dv.
v=0

Thus x affects G in two ways: as a parameter for the integrand, and as the
upper limit of integration. What is dG(x)/dx?
 4.6 Higher Order Derivatives 169

4.6 Higher Order Derivatives

Partial differentiation can be carried out more than once on nice enough func-
tions. For example if
f (x, y) = ex sin y
then
D1 f (x, y) = sin yex sin y , D2 f (x, y) = x cos yex sin y .
Taking partial derivatives again yields
D1 D1 f (x, y) = sin2 yex sin y ,
D1 D2 f (x, y) = cos yex sin y + x sin y cos yex sin y ,
D2 D1 f (x, y) = cos yex sin y + x sin y cos yex sin y = D1 D2 f (x, y),
D2 D2 f (x, y) = x sin yex sin y + x2 cos2 yex sin y ,
and some partial derivatives of these in turn are,
D1 D1 D2 f (x, y) = 2 sin y cos yex sin y + x sin2 y cos yex sin y ,
D1 D2 D1 f (x, y) = D1 D1 D2 f (x, y),
D2 D1 D2 f (x, y) = sin yex sin y + 2x cos2 yex sin y x sin2 yex sin y
+ x2 sin y cos2 yex sin y ,
D2 D2 D1 f (x, y) = D2 D1 D2 f (x, y),
D1 D2 D2 f (x, y) = sin yex sin y + 2x cos2 yex sin y x sin2 yex sin y
+ x2 sin y cos2 yex sin y
= D2 D1 D2 f (x, y),
D2 D1 D1 f (x, y) = 2 sin y cos yex sin y + x sin2 y cos yex sin y
= D1 D1 D2 f (x, y).

Suspiciously many of these match. The result of two or three partial differen-
tiations seems to depend only on how many were taken with respect to x and
how many with respect to y, not on the order in which they were taken.
To analyze the situation, it suffices to consider only two differentiations.
Streamline the notation by writing D2 D1 f as D12 f . (The subscripts may look
reversed, but reading D12 from left to right as D-one-two suggests the appro-
priate order of differentiating.) The definitions for D11 f , D21 f , and D22 f are
similar. These four functions are called the second order partial derivatives
of f , and in particular D12 f and D21 f are the second order mixed partial
derivatives. More generally, the kth order partial derivatives of a function f
are those that come from k partial differentiations. A C k -function is a func-
tion for which all the kth order partial derivatives exist and are continuous.
The theorem is that with enough continuity the order of differentiation doesnt
matter. That is, the mixed partial derivatives agree.
170 4 The Derivative

Theorem 4.6.1 (Equality of Mixed Partial Derivatives). Suppose that

f : A R (where A R2 ) is a C 2 -function. Then at any point (a, b) of A,
D12 f (a, b) = D21 f (a, b).
We might try to prove the theorem as follows:
D1 f (a, b + k) D1 f (a, b)
D12 f (a, b) = lim
k0 k
f (a+h,b+k)f (a,b+k)
limh0 h limh0 f (a+h,b)f
h
(a,b)
= lim
k0 k
f (a + h, b + k) f (a, b + k) f (a + h, b) + f (a, b)
= lim lim ,
k0 h0 hk
and similarly
f (a + h, b + k) f (a + h, b) f (a, b + k) + f (a, b)
D21 f (a, b) = lim lim .
h0 k0 hk
So, letting (h, k) = f (a + h, b + k) f (a, b + k) f (a + h, b) + f (a, b), we
want to show that
(h, k) (h, k)
lim lim = lim lim .
h0 k0 hk k0 h0 hk
If the order of taking the limits doesnt matter then we have the desired re-
sult. However, if f is not a C 2 -function then the order of taking the limits can
in fact matter, i.e., the two mixed partial derivatives can both exist but not
be equal (see exercise 4.6.1 for an example). Thus a correct proof of Theo-
rem 4.6.1 requires a little care. The theorem is similar to Taylors Theorem
from section 1.3 in that both are stated entirely in terms of derivatives, but
they are most easily proved using integrals. The following proof uses integra-
tion to show that (h, k)/(hk) is an average value of both D12 f and D21 f
near (a, b), and then letting h and k shrink to 0 forces D12 f and D21 f to
agree at (a, b) as desired. That is, the proof shows that the two quantities in
the previous display are equal by showing that each of them equals a common
third quantity.
Proof. Since f is a C 2 -function on A, every point of A is interior. Take any
point (a, b) A. Then some box B = [a, a + h] [b, b + k] lies in A. Compute
the nested integral
Z a+h Z b+k Z a+h
dy dx = k dx = hk.
a b a

Also, by the Fundamental Theorem of Integral Calculus twice,

Z a+h Z b+k Z a+h
D12 f (x, y) dy dx = (D1 f (x, b + k) D1 f (x, b)) dx
a b a
= f (a + h, b + k) f (a, b + k) f (a + h, b) + f (a, b) = (h, k).
 4.6 Higher Order Derivatives 171

(Thus the integral has reproduced the quantity that arose in the discussion
leading into this proof.) Let mh,k be the minimum value of D12 f on the box B,
and let Mh,k be the maximum value. These exist by Theorem 2.4.15 since B
is nonempty compact and D12 f : B R is continuous. Thus

mh,k D12 f (x, y) Mh,k for all (x, y) B.

Integrate this inequality, using the two previous calculations, to get

mh,k hk (h, k) Mh,k hk,

or
(h, k)
mh,k Mh,k .
hk
As (h, k) (0+ , 0+ ), the continuity of D12 f at (a, b) forces mh,k and Mh,k
to D12 f (a, b), and hence
(h, k)
D12 f (a, b) as (h, k) (0+ , 0+ ).
hk
But also, reversing the order of the integrations and of the partial derivatives
gives the symmetric calculations
Z b+k Z a+h
dx dy = hk,
b a

and Z Z
b+k a+h
D21 f (x, y) dx dy = (h, k),
b a
and so the same argument shows that
(h, k)
D21 f (a, b) as (h, k) (0+ , 0+ ).
hk
Since both D12 f (a, b) and D21 f (a, b) are the limit of (h, k)/(hk), they are
equal.

Extending Theorem 4.6.1 to more variables and to higher derivatives is
straightforward provided that one supplies enough continuity. The hypotheses
of the theorem can be weakened a bit, in which case a subtler proof is required,
but such technicalities are more distracting than useful.
Higher order derivatives are written in many ways. If a function is de-
scribed by the equation w = f (x, y, z) then D233 f is also denoted
  
f 3f
f233 , fyzz , , ,
z z y z 2 y
  
w 3w
wyzz , , .
z z y z 2 y
172 4 The Derivative

As with one derivative, these combine mnemonic advantages with conceptual

dangers.
A calculation using higher order derivatives and the Chain Rule transforms
the heat equation of Laplace from cartesian to polar coordinates. The C 2
quantity u = f (x, y) depending on the cartesian variables x and y satisfies
Laplaces equation if (blurring the distinction between u and f )

2u 2u
+ 2 = 0.
x2 y

If instead u is viewed as a function g(r, ) of the polar variables r and then

how is Laplaces equation expressed?
The cartesian coordinates in terms of the polar coordinates are

x = r cos , y = r sin .

Thus u = f (x, y) = f (r cos , r sin ) = g(r, ), showing that u depends on r

and via x and y:

r /x

@


'


qqq8 u
qq

qqqqq
 qq
/yq

The Chain Rule begins a hieroglyphic calculation,

u r = u x x r + u y yr ,

so that by the product rule,

urr = (ux xr + uy yr )r
= uxr xr + ux xrr + uyr yr + uy yrr .

Since ux and uy depend on r and via x and y just as u does, each of them can
take the place of u in the diagram above, and the Chain Rule gives expansions
of uxr and uyr as it did for ur ,

urr = uxr xr + ux xrr + uyr yr + uy yrr

= (uxx xr + uxy yr ) xr + ux xrr + (uyx xr + uyy yr ) yr + uy yrr
= uxx x2r + uxy yr xr + ux xrr + uyx xr yr + uyy yr2 + uy yrr
= uxx x2r + 2uxy xr yr + uyy yr2 + ux xrr + uy yrr .
 4.6 Higher Order Derivatives 173

Note the use of equality of mixed partial derivatives. The same calculation
with instead of r gives

u = uxx x2 + 2uxy x y + uyy y2 + ux x + uy y .

Since x = r cos and y = r sin we have the relations

xr = x/r, yr = y/r, x = y, y = x,
xrr = 0, yrr = 0, x = x, y = y.

It follows that

r2 urr = uxx x2 + 2uxy xy + uyy y 2 ,

rur = ux x + uy y,
u = uxx y 2 2uxy xy + uyy x2 ux x uy y,

and so

r2 urr + rur + u = uxx x2 + uyy y 2 + uxx y 2 + uyy x2

= (uxx + uyy )(x2 + y 2 ).

Recall that the cartesian form of Laplaces equation is uxx + uyy = 0. Now
the polar form follows:

r2 urr + rur + u = 0.

That is,
2u u 2 u
r22
+r + 2 = 0.
r r
The point of this involved calculation is that having done it once, and only
once, we now can check directly whether any function g of the polar variables
r and satisfies Laplaces equation. We no longer need to transform each
u = g(r, ) into cartesian terms u = f (x, y) before checking.
An n-by-n matrix A is orthogonal if AT A = I. (This concept was in-
troduced in exercise 3.5.5.) Let A be orthogonal and consider its associated
linear map,
TA : Rn Rn , TA (x) = Ax.
We show that prepending TA to a twice-differentiable function on Rn is inde-
pendent from applying the Laplacian operator to the function. That is, letting
denote the Laplacian operator on Rn ,

= D11 + D22 + + Dnn ,

and taking any twice-differentiable function on Rn ,

f : Rn R,
174 4 The Derivative

we show that
(f TA ) = f TA .
To see this, start by noting that for any x Rn , the chain rule and then the
fact that the derivative of any linear map is itself give two equalities of linear
mappings,

D(f TA )x = DfTA (x) D(TA )x = DfTA (x) TA .

In terms of matrices, the equality of the first and last quantities in the previous
display is an equality of row vector valued functions of x,
   
D1 (f TA ) Dn (f TA ) (x) = ( D1 f Dn f TA )(x) A.

Because we view vectors as columns, transpose the quantities in the previ-

ous display, using the fact that A is orthogonal to write A1 for AT , and
universalize over x to get an equality of column valued functions,

D1 (f TA ) D1 f
.. .
. = TA1 .. TA .
Dn (f TA ) Dn f

The derivative matrix of the left side has as its rows the row vector derivative
matrices of its entries, while the derivative matrix of the right side is computed
by the chain rule and the fact that the derivative of any linear map is itself,
   
Dij (f TA ) nn = A1 Dij f TA nn A.

The trace of a square matrix was introduced in exercise 3.2.5 as the sum of its
diagonal entries, and the fact that tr(A1 BA) = tr(B) if A is invertible was
noted just after the proof of Theorem 3.5.2. Equate the traces of the matrices
in the previous display to get the desired result,

(f TA ) = f TA .

(To complement the proof just given in functional notation, here is a more
elementary second proof. Let the matrix A have entries aij . For any x Rn ,
compute that for i = 1, , n,
n
X n
X
Di (f TA )(x) = Dj f (Ax)Di (Ax)j = aji Dj f (Ax),
j=1 j=1

and thus
n
X n
X n
X
Dii (f TA )(x) = aji Djk f (Ax)Di (Ax)k = aji aki Djk f (Ax),
j=1 k=1 j,k=1
 4.6 Higher Order Derivatives 175

and thus, because A is orthogonal, so that (AAT )jk is 1 when j = k and 0

otherwise,
n X
X n
(f TA )(x) = aji aki Djk f (Ax)
j,k=1 i=1
Xn
= (AAT )jk Djk f (Ax)
j,k=1
n
X
= Dii f (Ax) = (f TA )(x),
i=1

as desired.)

Exercises

4.6.1. This exercise shows that continuity is necessary for the equality of
mixed partial derivatives. Let
(
xy(y 2 x2 )
x2 +y 2 if (x, y) 6= (0, 0)
f (x, y) =
0 if (x, y) = (0, 0).

Away from (0, 0), f is rational and so it is continuous and and all its partial
derivatives of all orders exist and are continuous. Show: (a) f is continous
at (0, 0), (b) D1 f , and D2 f exist and are continuous at (0, 0), (c) D12 f (0, 0) =
1 6= 1 = D21 f (0, 0).

For the rest of these exercises, assume as much differentiability as necessary.

4.6.2. Suppose u, as a function of x and y, satisfies the differential equation

uxx uyy = 0. Make the change of variables x = s + t, y = s t. What
corresponding differential equation does u satisfy when viewed as a function
of s and t? (That is, find a nontrivial relation involving at least one of u, us ,
ut , uss , utt , and ust .)

4.6.3. (The wave equation) (a) Let c be a constant, tacitly understood to

denote the speed of light. Let x and t denote a space variable and a time
variable, and introduce variables

p = x + ct, q = x ct.

Show that a quantity w, viewed as a function of x and t, satisfies the wave

equation,
c2 wxx = wtt ,
if and only if it satisfies the equation
176 4 The Derivative

wpq = 0.

(b) Using part (a), show that in particular if w = F (x + ct) + G(x ct)
(where F and G are arbitrary C 2 -functions of one variable) then w satisfies
the wave equation.
(c) Now let 0 < v < c (both v and c are constant), and define new space
and time variables in terms of the original ones by a Lorentz transformation,

y = (x vt), u = (t (v/c2 )x) where = (1 v 2 /c2 )1/2 .

Show that

y + cu = (1 v/c)(x + ct), y cu = (1 + v/c)(x ct),

so that consequently (y, u) has the same spacetime norm as (x, t),

y 2 c 2 u 2 = x 2 c 2 t2 .

(d) Recall the variables p = x + ct and q = x ct from part (a). Similarly,

let r = y + cu and s = y cu. Suppose that a quantity w, viewed as a function
of p and q satisfies the wave equation wpq = 0. Use the results r = (1v/c)p,
s = (1 + v/c)q from part (c) to show that also it satisfies the wave equation
in the (r, s)-coordinate system, wrs = 0. Consequently, if w satisfies the wave
equation c2 wxx = wtt in the original space and time variables then also it
satisfies the wave equation c2 wyy = wuu in the new space and time variables.

4.6.4. Show that the substitution x = es , y = et converts the equation

x2 uxx + y 2 uyy + xux + yuy = 0

into Laplaces equation uss + utt = 0.

4.6.5. (a) Show that the substitution x = s2 t2 , y = 2st converts Laplaces

equation uxx + uyy = 0 back into Laplaces equation uss + utt = 0.
(b) Let k be a nonzero real number. Show that the substitution r = k ,
= k converts the polar Laplaces equation r2 urr + rur + u = 0 back
into the polar Laplaces equation 2 u + u + u = 0. (When k = 2 this
subsumes part (a) since the substitution here encodes the complex kth power
function in polar coordinates while the substitution in part (a) encodes the
complex squaring function in Cartesian coordinates.)

4.6.6. Let u be a function of x and y, and suppose that x and y in turn depend
linearly on s and t,
    
x ab s
= , ad bc = 1.
y cd t

What is the relation between uss utt u2st and uxx uyy u2xy ?
 4.6 Higher Order Derivatives 177

4.6.7. (a) Let H denote the set of points (x, y) R2 such that y > 0. Associate
to each point (x, y) H another point,
 
x y
(z, w) = , .
x2 + y 2 x2 + y 2
You may take for granted or verify that
zx = z 2 w 2 , zy = 2zw, zxx = 2z(z 2 3w2 ), zyy = 2z(z 2 3w2 )
and
wx = 2zw, wy = z 2 w 2 , wxx = 2w(3z 2 w2 ), wyy = 2w(3z 2 w2 ).

Consider a quantity u = f (z, w), so that also u = f(x, y) for a different

function f. As usual, we have
uxx = uzz zx2 + 2uzw zx wx + uww wx2 + uz zxx + uw wxx ,
uyy = uzz zy2 + 2uzw zy wy + uww wy2 + uz zyy + uw wyy .
Show that
y 2 (uxx + uyy ) = w2 (uzz + uww ).
The operator y 2 ( 2 /x2 + 2 /y 2 ) on H is the hyperbolic Laplacian, de-
noted H . We have just established the invariance of H under the hyperbolic
transformation that takes (x, y) to (z, w) = (x/(x2 + y 2 ), y/(x2 + y 2 )).
(b) Show that the invariance relation y 2 (uxx + uyy ) = w2 (uzz + uww ) also
holds when (z, w) = (x+b, y) for any fixed real number b, and that the relation
also holds when (z, w) = (rx, ry) for any fixed positive real number r. It is
known that any hyperbolic transformation of H takes the form (z, w) = (x, y)
where is a finite succession of transformations of the type in part (a) or of the
two types just addressed here. Note that consequently this exercise has shown
that the invariance relation holds for any hyperbolic transformation of H.
That is, for any hyperbolic transformation and for any twice-differential
function f : H R we have, analogously to the result at the very end of
the section,
H (f ) = H f .
4.6.8. Consider three matrices,
    
01 00 1 0
X= , Y = , H= .
00 10 0 1
Establish the relations
XY Y X = H, HX XH = 2X, HY Y H = 2Y.
Now consider three operators on smooth functions from Rn to R, reusing the
names X, Y , and H, and letting = D11 + D22 + + Dnn denote the
Laplacian operator,
178 4 The Derivative

1 2
(Xf )(x) = 2 |x| f (x),
(Y f )(x) = 12 f (x),
X n
n
(Hf )(x) = 2 f (x) + xi Di f (x).
i=1

Establish the same relations as a moment ago,

XY Y X = H, HX XH = 2X, HY Y H = 2Y.

The three matrices generate a small instance of a Lie algebra, and this exercise
shows that the space of smooth functions on Rn can be made a representation
of the Lie algebra. Further show, partly by citing the work at the end of
the section, that the action of any orthogonal matrix A on smooth functions
commutes with the representation,

X(f TA ) = (Xf ) TA ,
Y (f TA ) = (Y f ) TA ,
H(f TA ) = (Hf ) TA .

4.7 Extreme Values

In one variable calculus the derivative is used to find maximum and minimum
values (extrema) of differentiable functions. Recall the following useful facts.
(Extreme Value Theorem.) If f : [, ] R is continuous then it assumes
a maximum and a minimum on the interval [, ].
(Critical Point Theorem.) Suppose that f : [, ] R is differentiable
on (, ) and that f assumes a maximum or minimum at an interior point a
of [, ]. Then f (a) = 0.
(Second Derivative Test.) Suppose that f : [, ] R is C 2 on (, ) and
that f (a) = 0 at an interior point a of [, ]. If f (a) > 0 then f (a) is a
local minimum of f , and if f (a) < 0 then f (a) is a local maximum.
Geometrically the idea is that just as the affine function

A(a + h) = f (a) + f (a)h

specifies the tangent line to the graph of f at (a, f (a)), the quadratic function
1
P (a + h) = f (a) + f (a)h + f (a)h2
2
determines the best fitting parabola. When f (a) = 0 the tangent line is
horizontal and the sign of f (a) specifies whether the parabola opens upward
or downward. When f (a) = 0 and f (a) = 0, the parabola degenerates to
 4.7 Extreme Values 179

Figure 4.8. Approximating parabolas

the horizontal tangent line, and the second derivative provides no information.
(See figure 4.8.)
This section generalizes these facts to functions f of n variables. The Ex-
treme Value Theorem has already generalized as Theorem 2.4.15: a continuous
function f on a compact subset of Rn takes maximum and minimum values.
The Critical Point Theorem also generalizes easily to say that each extreme
value of the function f : A R that occurs at a point where f is differen-
tiable occurs at critical point of f , meaning a point a where Dfa is the zero
function.

Theorem 4.7.1 (Multivariable Critical Point Theorem). Suppose that

the function f : A R (where A Rn ) takes an extreme value at the point a
of A, and suppose that f is differentiable at a. Then all partial derivatives of f
at a are zero.

Proof. For each j {1, , n}, the value f (a) is an extreme value for the
one-variable function from definition 4.5.1 of the partial derivative Dj f (a).
By the one-variable Critical Point Theorem, (aj ) = 0. That is, Dj f (a) = 0.

The generalization of the second derivative test is more elaborate. From

now on, all functions are assumed to be of type C 2 on the interiors of their
domains, meaning that all their second order partial derivatives exist and are
continuous.

Definition 4.7.2 (Second Derivative Matrix). Let f : A R (where

A Rn ) be a function and let a be an interior point of A. The second
derivative matrix of f at a is the n-by-n matrix whose (i, j)th entry is the
second order partial derivative Dij f (a). Thus

D11 f (a) D1n f (a)
.. .. ..
f (a) = . . . .
Dn1 f (a) Dnn f (a)

By the equality of mixed partial derivatives, the second derivative matrix

is symmetric, i.e., f (a)T = f (a). Beware of confusing the second derivative
matrix and the Jacobian matrix: the second derivative matrix is a square
180 4 The Derivative

matrix defined only for scalar valued functions and its entries are second order
partial derivatives, while for scalar valued functions the Jacobian matrix is the
row vector of first partial derivatives.
Irksomely, f is not (f ) . The problem is that the seemingly-reasonable
map f : A Rn taking each point a A to f (a) does not fit correctly
into our conventions: each f (a) is specifically a row vector, but we view the
elements of Rn either as ordered lists with no shape at all or as column vectors.
T
Thus (f ) does not even exist. The correction is that the map f : A Rn
T
taking each point a A to f (a) does fit our scheme, and indeed
T
(f ) (a) = f (a) for interior points a of A.

The pettiness of this issue makes clear that eventually we should loosen our
conventions and freely transpose vectors and matrices as called for by context,
as indeed many software packages do. But since the conventions can be helpful
for a student who is seeing the material for the first time, we retain them for
now and grudgingly accept the transpose here.
As an example, if

f (x, y) = sin2 x + x2 y + y 2 ,

then for any (a, b) R2 ,

f (a, b) = [sin 2a + 2ab a2 + 2b]

and  
2 cos 2a + 2b 2a
f (a, b) = .
2a 2
Any n-by-n matrix M determines a quadratic function

QM : Rn R, QM (h) = hT M h.

Here h is viewed as a column vector. If M has entries mij and h = (h1 , , hn )

then the rules of matrix multiplication show that

m11 m1n h1 n X n
  . . X
QM (h) = h1 hn .. . . .. ...
.
= mij hi hj .
mn1 mnn hn i=1 j=1

The function QM is homogeneous of degree 2, meaning that each of its terms

has degree 2 in the entries of h and therefore QM (th) = t2 QM (h) for all t R
and h Rn .
When M is the second derivative matrix of a function f at a point a, the
corresponding quadratic function is denoted Qfa rather than Qf (a) . Just as
f (a) + Dfa (h) gives the best affine approximation of f (a + h) for small h,
f (a) + Dfa (h) + 21 Qfa (h) gives the best quadratic approximation.
 4.7 Extreme Values 181

In the example f (x, y) = sin2 x + x2 y + y 2 , the second derivative matrix

at a point (a, b) defines the quadratic function
  
  2 cos 2a + 2b 2a h
Qf(a,b) (h, k) = h k
2a 2 k
= 2((cos 2a + b) h2 + 2a hk + k 2 ) for (h, k) R2 ,

and so the best quadratic approximation of f near, for instance, the point
(/2, 1) is
1
f (/2 + h, 1 + k) f (/2, 1) + Df(/2,1) (h, k) + Qf(/2,1) (h, k)
2
= 2 /4 + 2 + h + ( 2 /4 + 2)k + hk + k 2 .

Suppose that f : A R (where A R2 ) has a critical point at (a, b),

i.e., f (a, b) = [0 0]. Working in local coordinates, we will approximate f
by a quadratic function on R2 having a critical point at (0, 0). The graphs of
nine such quadratic functions are shown in figure 4.9. If the best quadratic
approximation of f at (a, b) is a bowl then f should have a minimum at (a, b).
Similarly for an inverted bowl and a maximum. If the best quadratic ap-
proximation is a saddle then there should be points (x, y) near (a, b) where
f (x, y) > f (a, b) and points (x , y ) near (a, b) where f (x , y ) < f (a, b). In
this case (a, b) is called for obvious reasons a saddle point of f .

Figure 4.9. Two bowls, two saddles, four halfpipes, and a plane
182 4 The Derivative

Returning to the example f (x, y) = sin2 x + x2 y + y 2 , note that (0, 0)

is a critical point of f since f (0, 0) = [0 0]. The second derivative matrix
f (0, 0) is [ 20 02 ], and so the quadratic function 12 Qf(0,0) is given by
  
1 1  20 h
Qf(0,0) (h, k) = hk = h2 + k 2 .
2 2 02 k

Thus the graph of f looks like a bowl near (0, 0) and f (0, 0) should be a local
minimum.
This discussion is not yet rigorous. Justifying the ideas and proving the
appropriate theorems will occupy the rest of this section. The first task is to
study quadratic approximation of C 2 -functions.

Proposition 4.7.3 (Special Case of Taylors Theorem). Let I be an open

interval in R containing [0, 1]. Let : I R be a C 2 -function. Then
1
(1) = (0) + (0) + (c) for some c [0, 1].
2
The proposition follows from the general Taylors Theorem in section 1.3
since the first degree Taylor polynomial of at 0 is T1 (t) = (0) + (0)t, so
that in particular, T1 (1) = (0) + (0).

Theorem 4.7.4 (Quadratic Taylor Approximation). Let f : A R

(where A Rn ) be a C 2 -function on the interior points of A. Let a be an
interior point of A. Then for all small enough h Rn ,
1
f (a + h) = f (a) + Dfa (h) + Qfa+ch (h) for some c [0, 1],
2
or, in matrices, viewing h as a column vector,
1
f (a + h) = f (a) + f (a)h + hT f (a + ch)h for some c [0, 1].
2
Proof. Let I = (, 1 + ) be a small superinterval of [0, 1] in R. Define

: I A, (t) = a + th.

Thus (0) = a, (1) = a + h, and (t) = h for all t I. Further define

= f : I R.

That is, (t) = f (a + th) is the restriction of f to the line segment from a
to a + h. By the Chain Rule and the fact that = h,

(t) = (f ) (t) = f ((t))h = Dfa+th (h).

T
The previous display rephrases as (t) = hf ((t)), hi, so the Chain Rule,
T
the irksome formula (f ) = f , and the symmetry of f give
 4.7 Extreme Values 183

(t) = hf ((t))h, hi = hT f (a + th)h = Qfa+th (h).

Since f (a + h) = (1), the special case of Taylors Theorem says that for some
c [0, 1],
1 1
f (a + h) = (0) + (0) + (c) = f (a) + Dfa (h) + Qfa+ch (h),
2 2
giving the result.

Thus, to study f near a critical point a Rn where Dfa is zero, we need to
look at the sign of Qfa+ch (h) for small vectors h. The next order of business
is therefore to discuss the values taken by a homogeneous quadratic function.
Definition 4.7.5 (Positive Definite, Negative Definite, Indefinite Ma-
trix). The symmetric square n-by-n matrix M is called
positive definite if QM (h) > 0 for every nonzero h Rn ,
negative definite if QM (h) < 0 for every nonzero h Rn ,
indefinite if QM (h) is positive for some h and negative for others.
The identity
 1 0 matrix
 I is positive definite since hT Ih = |h|2 for all h.
The matrix 0 1 is indefinite. The general question of whether a symmetric
n-by-n matrix is positive definite leads to an excursion into linear algebra
too lengthy for this course. (See exercise 4.7.10 for the result without proof.)
However, in the special case of n = 2, basic methods give the answer. Recall
that the quadratic polynomial h2 +2h+ takes positive and negative values
if and only if it has distinct real roots, i.e., 2 < 0.
Proposition
  4.7.6 (Two-by-two Definiteness Test). Consider a matrix

M= M2 (R). Then

(1) M is positive definite if and only if > 0 and 2 > 0.
(2) M is negative definite if and only if < 0 and 2 > 0.
(3) M is indefinite if and only if 2 < 0.
Proof. Since QM (t(h, k)) = t2 QM (h, k) for all real t, scaling the input vector
(h, k) by nonzero real numbers doesnt affect the sign of the output. The
second entry k can therefore be scaled to 0 or 1, and if k = 0 then the first
entry h can be scaled to 1. Therefore, to show (1), reason that
M is positive definite QM (1, 0) > 0 and QM (h, 1) > 0 for all h R
> 0 and h2 + 2h + > 0 for all h R
> 0 and 2 > 0.
(2) is similar. As for (3),
M is indefinite h2 + 2h + takes positive and negative values
2 < 0.


184 4 The Derivative

The proposition provides no information if 2 = 0. Geometrically,

the proposition gives conditions on M to determine that the graph of QM is
a bowl, an inverted bowl, or a saddle. The condition 2 = 0 indicates a
degenerate graph: a halfpipe, an inverted halfpipe, or a plane.
For nonzero the matrix calculation
     
1 0 0 1 1
=
1 1 0 1 ( 2 ) 0 1

gives a corresponding equality of quadratic functions,

x2 + 1 ( 2 )y 2 ,
x2 + 2xy + y 2 = = x + 1 y.
x

That is, a change of variables eliminates the cross term, and the variant
quadratic function makes the results of the definiteness test clear.
The positive definite, negative definite, or indefinite character of a matrix
is preserved if the matrix entries vary by small enough amounts. Again we
restrict our discussion to the 2-by-2 case. Here the result is plausible geomet-
rically, since it says that if the matrix M (a, b) defines a function whose graph
is (for example) a bowl, then matrices close to M (a, b) should define functions
with similar graphs, which thus should still be bowl-shaped. The same persis-
tence holds for a saddle, but a halfpipe can deform immediately into either a
bowl or a saddle, and so can a plane.

Proposition 4.7.7 (Persistence of Definiteness). Let A be a subset of R2 ,

and let the matrix-valued mapping
 
(x, y) (x, y)
M : A M2 (R), M (x, y) =
(x, y) (x, y)

be continuous. Let (a, b) be an interior point of A. Suppose that the matrix

M (a, b) is positive definite. Then for all (x, y) in some -ball about (a, b), the
matrix M (x, y) is also positive definite. Similar statements hold for negative
definite and indefinite matrices.

Proof. By the Persistence of Inequality principle (Proposition 2.3.10), the

criteria > 0 and 2 > 0 remain valid if x and y vary by a small enough
amount. The other statements follow similarly.

When a function f has continuous second order partial derivatives, the

entries of the second derivative matrix f (a) vary continuously with a. The
upshot of the last proposition is therefore that we may replace the nebulous
notion of Qfa+ch for some c with the explicit function Qfa .

Proposition 4.7.8 (Two-variable Max/min Test). Let f : A R

(where A R2 ) be C 2 on its interior points. Let (a, b)h be an
i interior point

of A, and suppose that f (a, b) = [0 0]. Let f (a, b) = . Then
 4.7 Extreme Values 185

(1) If > 0 and 2 > 0 then f (a, b) is a local minimum.

(2) If < 0 and 2 > 0 then f (a, b) is a local maximum.
(3) If 2 < 0 then f (a, b) is a saddle point.
Proof. This follows from Theorem 4.7.4, Proposition 4.7.6 and Proposi-
tion 4.7.7.

Again, the test gives no information if 2 = 0.
Returning once again to the example f (x, y) = sin2 x + x2 y + y 2 with its
critical point (0, 0) and second derivative matrix f (0, 0) = [ 20 02 ], the max/min
test shows that f has a local minimum at (0, 0).
Another example is to find the extrema of the function
f (x, y) = xy(x + y 3)
on the triangle
T = {(x, y) R2 : x 0, y 0, x + y 3}.
To solve this, first note that T is compact. Therefore f is guaranteed to take
a maximum and a minimum value on T . These are assumed either at interior
points of T or along the edge. Examining the signs of x, y and x + y 3 shows
that f is zero at all points on the edge of T and negative on the interior of T .
Thus f assumes its maximum valuezeroalong the boundary of T and must
assume its minimum somewhere inside. (See figure 4.10.) To find the extrema
of f inside T , we first find the critical points. The partial derivatives of f (now
viewed as a function only on the interior of T ) are
fx (x, y) = y(2x + y 3) fy (x, y) = x(x + 2y 3),
and since x and y are nonzero on the interior of G, these are both zero only at
the unique solution (x, y) = (1, 1) of the simultaneous equations 2x + y = 3,
x + 2y = 3. Therefore f (1, 1) = 1 must be the minimum value of f . A quick
calculation shows that f (1, 1) = [ 21 12 ], and the max/min test confirms the
minimum at (1, 1).
Another example is to find the extreme values of the function
1 2 1
f : R2 R, f (x, y) = x + xy 2x y 2 .
2 2
Since R2 is not compact, there is no guarantee that f has any extrema. In fact,
for large x, f (x, 0) gets arbitrarily large, and for large y, f (0, y) gets arbitrarily
small (where small means negative, not epsilontic). So f has no global extrema.
Nonetheless there may be local ones. Every point of R2 is interior, so it suffices
to examine the critical points of f . The partial derivatives are
fx (x, y) = x + y 2 fy (x, y) = x y,
and the only point where both of them  vanish is (x, y) = (1, 1). The second
derivative matrix is f (1, 1) = 11 11 , so the critical point (1, 1) is a saddle
point. The function f has no extrema, local or global.
186 4 The Derivative

Figure 4.10. Zero on the boundary, negative on the interior

Exercises

4.7.1. Compute the best quadratic approximation of f (x, y) = ex cos y at the

point (0, 0), f (h, k) f (0, 0) + Df(0,0) (h, k) + 12 Qf(0,0) (h, k).

4.7.2. Compute the best quadratic approximation of f (x, y) = ex+2y at the

point (0, 0).

4.7.3. Explain, making whatever reasonable assumptions seem to be helpful,

of why the n-dimensional conceptual analogue of figure 4.9 should have 3n
pictures. How does this relate to figure 4.8?

4.7.4. Find the extreme values taken by f (x, y) = xy(4x2 + y 2 16) on the
quarter ellipse

E = {(x, y) R2 : x 0, y 0, 4x2 + y 2 16}.

4.7.5. Find the local extrema of the function f (x, y) = x2 +xy 4x+ 23 y 2 7y
on R2 .

4.7.6. Determine the nature of f (x, y) = 31 x3 + 13 y 3 + (x 23 )2 (y + 4)2 at

each of its critical points. Are there global extrema?

4.7.7. Find the critical points. Are they maxima, minima, or saddle points?
(The max/min test will not always help.)

f (x, y) = x2 y + xy 2 g(x, y) = ex+y h(x, y) = x5 y + xy 5 + xy.

1 1
4.7.8. Discuss local and global extrema of f (x, y) = x1 y1 on the open
ball B((0, 0); 1) in R2 .
 4.7 Extreme Values 187

4.7.9. The graph of the function m(x, y) = 6xy 2 2x3 3y 4 is called a

monkey saddle. Find the three critical points of m and classify each as a
maximum, minimum or saddle. (The max/min test will work on two. Study
m(x, 0) and m(0, y) to classify the third.) Explain the name monkey saddle
computer graphing software may help.
4.7.10. Linear algebra readily addresses the question of whether an n-by-n
matrix is positive definite, negative definite, or indefinite.
Definition 4.7.9 (Characteristic Polynomial). Let M be an n-by-n ma-
trix. Its characteristic polynomial is
pM () = det(M I).
The characteristic polynomial of M is a polynomial of degree n in the scalar
variable .
While the roots of a polynomial with real coefficients are in general com-
plex, the roots of the characteristic polynomial of a symmetric matrix in
Mn (R) are guaranteed to be real. The characterization we want is
Theorem 4.7.10 (Description of Definite/Indefinite Matrices). Let M
be a symmetric matrix in Mn (R). Then
(1) M is positive definite if and only if all the roots of pM () are positive.
(2) M is negative definite if and only if all the roots of pM () are negative.
(3) M is indefinite if and only if pM () has positive roots and negative roots.
With this result one can extend the methods in the section to functions of
more than two variables. h i
(a) Let M be the symmetric matrix
M2 (R). Show that

pM () = 2 ( + ) + ( 2 ).
(b) Show that Theorem 4.7.10 is equivalent to Proposition 4.7.6 when
n = 2.
(c) Classify the 3-by-3 matrices

1 1 0 010
1 2 0 1 0 1 .
0 01 010
A generalization of Proposition 4.7.7 also holds, since the roots of a poly-
nomial vary continuously with the polynomials coefficients. The generalized
proposition leads to
Proposition 4.7.11 (General Max/min Test). Let f : A R (where
A Rn ) be C 2 on its interior points. Let a be an interior point of A, and
suppose that f (a) = 0n . Let the second derivative matrix f (a) have charac-
teristic polynomial p().
188 4 The Derivative

(1) If all roots of p() are positive then f (a) is a local minimum.
(2) If all roots of p() are negative then f (a) is a local maximum.
(3) If p() has positive and negative roots then f (a) is a saddle point.
4.7.11. This exercise eliminate the cross terms from a quadratic function of n
variables, generalizing the calculation for n = 2 in the section. Throughout, we
abbreviate positive definite to positive. Let M be a positive n-by-n symmetric
matrix where n > 1. This exercise shows how to diagonalize M as a quadratic
function. (This is different from diagonalizing M as a transformation, as is
done in every linear algebra course.) Decompose M as
 T
ac
M= ,
cN

with a > 0 and c Rn1 a column vector and N positive (n 1)-by-(n 1)

symmetric. Define
M2 = N a1 ccT ,
again (n 1)-by-(n 1) symmetric, though we dont yet know whether it is
positive. Check that
 T   
1 a1 cT a 0T 1 a1 cT
M= .
0 In1 0 M2 0 In1

Show that in terms of quadratic functions this says (letting v = (x1 , , xn )

and v2 = (x2 , , xn ) with these vectors viewed as columns, and letting
1 = x1 + a1 cT v2 )
x
x21 + v2T M2 v2 .
v T M v = a
Consequently M2 is positive: indeed, if the last term of the previous display
is nonpositive then setting x1 = a1 cT v2 makes x 1 zero and thus makes
the entire right side nonpositive, so that v = 0n because M is positive and
consequently v2 = 0n1 . Repeating the process on M2 and so on eventually
gives
21 + + an x
v T M v = a1 x 2n ,
x1 , , x
with all ai > 0 and with the vector v = ( n ) of modified variables the
image of the vector v of original variables by a linear transformation whose
matrix is upper triangular with 1s on the diagonal.

4.8 Directional Derivatives and the Gradient

Let f be a scalar-valued function, f : A R where A Rn , and as-
sume that f is differentiable at a point a of A. While the derivative Dfa is a
rather abstract objectthe linear mapping that gives the best approximation
of f (a + h) f (a) for small hthe partial derivatives Dj f (a) are easy to
understand. The jth partial derivative of f at a,
 4.8 Directional Derivatives and the Gradient 189

f (a + tej ) f (a)
Dj f (a) = lim ,
t0 t
measures the rate of change of f at a as its input varies in the jth direction.
Visually, Dj f (a) gives the slope of the jth cross section through a of the graph
of f .
Analogous formulas measure the rate of change of f at a as its input varies
in a direction that doesnt necessarily parallel a coordinate axis. A direction
in Rn is specified by a unit vector d, i.e., a vector d such that |d| = 1. As the
input to f moves distance t in the d direction, f changes by f (a + td) f (a).
Thus the following definition is natural.
Definition 4.8.1 (Directional Derivative). Let f : A R (where A
Rn ) be a function, let a be an interior point of A, and let d Rn be a unit
vector. The directional derivative of f at a in the d direction is
f (a + td) f (a)
Dd f (a) = lim ,
t0 t
if this limit exists.
The directional derivatives of f in the standard basis vector directions are
simply the partial derivatives.
When n = 2 and f is differentiable at (a, b) R2 , its graph has a well-
fitting tangent plane through (a, b, f (a, b)). The plane is determined by the
two slopes D1 f (a, b) and D2 f (a, b), and it geometrically determines the rate
of increase of f in all other directions. (See figure 4.11.) The geometry suggests
that if f : A R (where A Rn ) is differentiable at a then all directional
derivatives are expressible in terms of the partial derivatives. This is true and
easy to show. A special case of the differentiability property (4.1) is

f (a + td) f (a) Dfa (td) is o(td) = o(t),

or, since the constant t passes through the linear map Dfa ,
f (a + td) f (a)
lim = Dfa (d),
t0 t
or, since the linear map Dfa has matrix [D1 f (a), , Dn f (a)],
n
X
Dd f (a) = Dj f (a)dj ,
j=1

as desired.
The derivative matrix f (a) of a scalar-valued function f at a is often
called the gradient of f at a and written f (a). That is,

f (a) = f (a) = [D1 f (a), , Dn f (a)].

The previous calculation and this definition lead to

190 4 The Derivative

Figure 4.11. General directional slope determined by axis-directional slopes

Theorem 4.8.2 (Directional Derivative and Gradient). Let the func-

tion f : A R (where A Rn ) be differentiable at a, and let d Rn be a
unit vector. Then the directional derivative of f at a in the d direction exists,
and it is equal to
n
X
Dd f (a) = Dj f (a)dj
j=1

= hf (a), di
= |f (a)| cos f (a),d .

Therefore:
The rate of increase of f at a in the d direction varies with d, from
|f (a)| when d points in the direction opposite to f (a), to |f (a)|
when d points in the same direction as f (a).
In particular, the vector f (a) points in the direction of greatest increase
of f at a, and its modulus |f (a)| is precisely this greatest rate.
Also, the directions orthogonal to f (a) are the directions in which f
neither increases nor decreases at a.

This theorem gives necessary conditions that arise in consequence of the

derivative of f existing at a point a. As in section 4.5, the converse statement,
that these conditions are sufficient to make the derivative of f exist at a,
is false. Each directional derivative Dd f (a) can exist without the derivative
Dfa existing (exercise 4.8.10). Furthermore, each directional derivative can
exist at a and satisfy the formula Dd f (a) = hf (a), di in the theorem, but
still without the derivative Dfa existing (exercise 4.8.11). The existence of
the multivariable derivative Dfa is a stronger condition than any amount of
one-variable cross-sectional derivative data at a.
 4.8 Directional Derivatives and the Gradient 191

For an example of using the theorem, if you are skiing on the quadratic
mountain f (x, y) = 9 x2 2y 2 at the point (a, f (a)) = (1, 1, 6), then your
gradient meter shows

f (1, 1) = (D1 f (1, 1), D2 f (1, 1)) = (2x, 4y)(x,y)=(1,1) = (2, 4).

Therefore the direction of steepest descent down the hillside is the (2, 4)-
direction (this could be divided by its modulus 20 to make it a unit vector),

and the slope of steepest descent is the absolute value |f (1, 1)| = 20.
On the other hand, cross-country skiing in the (2, 1)-direction, which is
orthogonal to f (1, 1), neither gains nor loses elevation immediately. (See
figure 4.12.) The cross-country skiing trail that neither climbs nor descends
has a mathematical name.

Figure 4.12. Gradient and its orthogonal vector for the parabolic mountain

Definition 4.8.3 (Level Set). Let f : A R (where A Rn ) be a func-

tion. A level set of f is the set of points in A that map under f to some fixed
value b in R,
L = {x A : f (x) = b}.

The curves on a topographical map are level sets of the altitude function.
The isotherms on a weather map are level sets of the temperature function,
and the isobars on a weather map are level sets of the pressure function.
Indifference curves in economics are level sets of the utility function, and iso-
quants are level sets of the production function. Surfaces of constant potential
in physics are level sets of the potential function.
For example, on the mountain

f : R2 R, f (x, y) = 9 x2 2y 2 ,
192 4 The Derivative

the level set for b = 5 is an ellipse in the plane,

L = {(x, y) R2 : x2 + 2y 2 = 4}.

And similarly the level set is an ellipse for any real number b up to 9. As
just mentioned, plotting the level sets of a function f of two variables gives a
topographical map description of f . The geometry is different for a function
of one variable: each level set is a subset of the line. For example, consider a
restriction of the sine function,

f : (0, ) R, f (x) = sin(x).

The level set taken by f to 1/2 consists of two points,

L = {/6, 5/6}.

For a function of three variables, each level set is a subset of space. For ex-
ample, if a, b, and c are positive numbers, and the function is

f : R3 R, f (x, y, z) = (x/a)2 + (y/b)2 + (z/c)2 ,

then its level sets are ellipsoids. Specifically, for every positive
r, the level set
of points taken
by f to r is the ellipsoid of x-radius a r, y-radius b r, and
z-radius c r,
(  2  2  2 )
3 x y z
L = (x, y, z) R : + + =1 .
a r b r c r

The third bullet in Theorem 4.8.2 says that the gradient is normal to the
level set. This fact may seem surprising, since the gradient is a version of the
derivative, and we think of the derivative as describing a tangent object to a
graph. The reason that the derivative has become a normal object is that

a level set is different from a graph.

A level set of f is a subset of the domain of f , whereas the graph of f ,

which simultaneously shows the domain and the range of f , is a subset of a
space that is one dimension larger. For instance, if we think of f as measuring
elevation, then the graph of f is terrain in three-dimensional space, while a
level set of f is set of points in the plane that lie beneath the terrain at some
constant altitude; the level set is typically a curve. Figure 4.12 illustrates the
difference in the case of the mountain function. Note that in the left part of the
figure, the gradient is orthogonal to the ellipse on which it starts. Similarly,
figure 4.13 illustrates the difference in the case of the restricted sine function
from the previous paragraph. In the figure, the x-axis shows the two-point
level set from the previous paragraph, and the gradient of f at each of the
two points. The fact that one gradient points rightward indicates that to climb
 4.8 Directional Derivatives and the Gradient 193

the graph of f over that point, one should move to the right, and the slope
to be encountered on the graph will be the length of the gradient on the axis.
Similarly, the other gradient points leftward because to climb the graph over
the other point, one should move to the left. Here each gradient is trivially
orthogonal to the level set, because the level set consists of isolated points.
For the three-variable function from the previous paragraph, we still can see
the level setsthey are concentric ellipsoidsbut not the graph, which would
require four dimensions. Instead, we can conceive of the function as measuring
temperature in space, and of the gradient as pointing in the direction to move
for greatest rate of temperature-increase, with the length of the gradient being
that rate. Figure 4.14 shows a level set for the temperature function, and
several gradients, visibly orthogonal to the level set.

Figure 4.13. Level set and gradients for the sine function

Figure 4.14. Level set and gradients for the temperature function
194 4 The Derivative

Although Theorem 4.8.2 has already stated that the gradient is orthogonal
to the level set, we now amplify the argument. Let f : A R (where
A Rn ) be given, and assume that it is differentiable. Let a be a point of A,
and let b = f (a). Consider the level set of f containing a,

L = {x A : f (x) = b} Rn ,

and consider any smooth curve from some interval into the level set, passing
through a,
: (, ) L, (0) = a.
The composite function
f : (, ) R
is the constant function b, so that its derivative at 0 is 0. By the Chain Rule
this relation is
f (a) (0) = 0.
Every tangent vector to L at a takes the form (0) for some of the sort that
we are considering. Therefore, f (a) is orthogonal to every tangent vector
to L at a, i.e., f (a) is normal to L at a.
Before continuing to work with the gradient, we pause to remark that level
sets and graphs are related. For one thing:
The graph of any function is also the level set of a different function.
To see this, let n > 1, let A0 be a subset of Rn1 , and let f : A0 R be any
function. Given this information, let A = A0 R and define a second function
g : A R,

g(x1 , , xn1 , xn ) = f (x1 , , xn1 ) xn .

Then the graph of f is a level of g, specifically the set of inputs that g takes
to 0,

graph(f ) = {x A0 R : xn = f (x1 , , xn1 )}

= {x A : g(x) = 0}.

For example, the graph of the mountain function f (x, y) = 9 x2 2y 2 is also

a level set of the function g(x, y, z) = 9 x2 2y 2 z. But in contrast to this
quick method defining g explicitly in terms of f to show that every graph is
a level set, the converse question is much more subtle:
To what extent is some given level set also a graph?
For example, the level sets of the mountain function f are ellipses (as shown in
figure 4.12), but an ellipse is not the graph of y as a function of x or vice versa.
The converse question will be addressed by the Implicit Function Theorem in
the next chapter.
 4.8 Directional Derivatives and the Gradient 195

Returning to the gradient, the geometrical fact that it is normal to the

level set makes it easy to find the tangent plane to a two-dimensional surface
in R3 . For example, consider the surface

H = {(x, y, z) R3 : x2 + y 2 z 2 = 1}.

(This surface is a hyperboloid of one sheet.) The point (2 2, 3, 4) belongs
to H. Note that H as a level set of the function f (x, y, z) = x2 + y 2 z 2 , and
compute the gradient

f (2 2, 3, 4) = (4 2, 6, 8).

Since this is the normal vector to H at (2 2, 3, 4), the tangent plane equation
at the
end of section 3.10 shows that the equation of the tangent plane to H
at (2 2, 3, 4) is

4 2(x 2 2) + 6(y 3) 8(z 4) = 0.

If a function f : Rn R has a continuous gradient, then from any

starting point a Rn where the gradient f (a) is nonzero, there is a path of
steepest ascent of f (called an integral curve of f ) starting at a. If n = 2
and the graph of f is seen as a surface in 3-space, then the integral curve from
the point (a, b) R2 is the shadow of the path followed by a particle climbing
the graph, starting at (a, b, f (a, b)). If n = 2 or n = 3 and f is viewed as
temperature, then the integral curve is the path followed by a heat-seeking
bug.
To find the integral curve, we set up an equation that describes it. The
idea is to treat the gradient vector as a divining rod and follow it starting
at a. Doing so produces a path in Rn that describes time-dependent motion,
always in the direction of the gradient, and always with speed equal to the
modulus of the gradient. Computing the path amounts to finding an interval
I R containing 0 and a mapping

: I Rn

that satisfies the differential equation with initial conditions

(t) = f ((t)), (0) = a. (4.3)

Whether (and how) one can solve this for depends on the data f and a.
In the case of the mountain function f (x, y) = 9 x2 2y 2 , with gradient
f (x, y) = (2x, 4y), the path has two components 1 and 2 , and the
differential equation and initial conditions (4.3) become

(1 (t), 2 (t)) = (21 (t), 42 (t)), (1 (0), 2 (0)) = (a, b),

to which the unique solution is

196 4 The Derivative

(1 (t), 2 (t)) = (ae2t , be4t ).

Let x = 1 (t) and y = 2 (t). Then the previous display shows that

a2 y = bx2 ,

and so the integral curve lies on a parabola. (The parabola is degenerate if the
starting point (a, b) lies on either axis.) Every parabola that forms an integral
curve for the mountain function meets orthogonally with every ellipse that
forms a level set. (See figure 4.15.)

Figure 4.15. Level sets and integral curves for the parabolic mountain

For another example, let f (x, y) = x2 y 2 . The level sets for this function
are hyperbolas having the 45 degree lines x = y and x = y as asymptotes.
The gradient of the function is f (x, y) = (2x, 2y), so to find the integral
curve starting at (a, b), we need to solve the equations

(1 (t), 2 (t)) = (21 (t), 22 (t)), (1 (0), 2 (0)) = (a, b).

Thus (1 (t), 2 (t)) = (ae2t , be2t ), so that the integral curve lies on the hy-
perbola xy = ab having the axes x = 0 and y = 0 as asymptotes. The integral
curve hyperbola is orthogonal to the level set hyperbolas. (See figure 4.16.)
For another example, let f (x, y) = ex y. The level sets for this function
are the familiar exponential curve y = ex and all of its vertical translates. The
gradient of the function is f (x, y) = (ex , 1), so to find the integral curve
starting at (0, 1), we need to solve the equations

(1 (t), 2 (t)) = (e1 (t) , 1), (1 (0), 2 (0)) = (0, 1).

To find 1 , reason that

e1 (t) 1 (t) = 1 for all t 0 where the system is sensible,

and so for all t 0 where the system is sensible,

 4.8 Directional Derivatives and the Gradient 197

Figure 4.16. Hyperbolic level sets and integral curves

Z t
e1 ( ) 1 ( ) d = t.
=0

Integration gives
e1 (t) + e1 (0) = t,
and so, recalling that 1 (0) = 0,

1 (t) = ln(1 t), 0 t < 1.

Also, 2 (t) = 1 t. Thus the integral curve,

(1 (t), 2 (t)) = ( ln(1 t), 1 t), 0t<1

is the portion of the curve y = ex where x 0. (See figure 4.17.) The entire
integral curve is traversed in one unit of time.

Figure 4.17. Negative exponential integral curve for exponential level sets

For another example, let f (x, y) = x2 + xy + y 2 . The level sets for this
function are tilted ellipses. The gradient of f is f (x, y) = (2x + y, x + 2y),
so to find the integral curve starting at (a, b), we need to solve the equations
198 4 The Derivative

1 (t) = 21 (t) + 2 (t) 1 (0) = a

2 (t) = 1 (t) + 22 (t) 2 (0) = b.

Here the two differential equations are coupled, meaning that the derivative
of 1 depends on both 1 and 2 , and similarly for the derivative of 2 . How-
ever, the system regroups conveniently,

(1 + 2 ) (t) = 3(1 + 2 )(t) (1 + 2 )(0) = a + b

(1 2 ) (t) = (1 2 )(t) (1 2 )(0) = a b.

Thus

(1 + 2 )(t) = (a + b)e3t
(1 2 )(t) = (a b)et ,

from which

1 (t) = 12 (a + b)e3t + 12 (a b)et

2 (t) = 21 (a + b)e3t 12 (a b)et .

These call to be checked, and indeed

1 (t) = 23 (a + b)e3t + 12 (a b)et = 21 (t) + 2 (t)

2 (t) = 23 (a + b)e3t 12 (a b)et = 1 (t) + 22 (t).

The motion takes place along the cubic curve having equation

x+y (x y)3
= .
a+b (a b)3

(See figure 4.18.) The integral curves in the first two examples were quadratic
only by happenstance, in consequence of the functions 9x2 2y 2 and x2 y 2
having such simple coefficients. Changing the mountain function to 9x2 3y 2
would produce cubic integral curves, and changing x2 y 2 to x2 5y 2 in the
second example would produce integral curves x5 y = a5 b.
For another example, suppose the temperature in space is given by
T (x, y, z) = 1/(x2 + y 2 + z 2 ). (This function blows up at the origin, so we
dont work there.) The level sets of this function are spheres and the integral
curves are rays going toward the origin. The level set passing through the
point (a, b, c) in space is again orthogonal to the integral curve through the
same point. In general, solving the vector differential equation (4.3) to find
the integral curves of a function f can be difficult.

Exercises

4.8.1. Let f (x, y, z) = xy 2 + yz. Find D( 32 , 31 , 23 ) f (1, 1, 2).

 4.8 Directional Derivatives and the Gradient 199

Figure 4.18. Cubic integral curve for elliptic level sets

4.8.2. Let g(x, y, z) = xyz, let d be the unit vector in the direction from
(1, 2, 3) to (3, 1, 5). Find Dd g(1, 2, 3).
4.8.3. Let f be differentiable at a point a, and let d = e1 , a unit vector. Are
the directional derivative Dd f (a) and the partial derivative D1 f (a) equal?
Explain.
4.8.4. Formulate and prove a version of Rolles theorem for functions of n
variables.
4.8.5. Show that if f : Rn R and g : Rn R are differentiable then so
is their product f g : Rn R and (f g) = f g + gf .
4.8.6. Find the tangent plane to the surface {(x, y, z) : x2 +2y 2 +3zx10 = 0}
in R3 at the point (1, 2, 13 ).
4.8.7. (a) Consider the surface S = {(x, y, z) R3 : xy = z}. Let p = (a, b, c)
be a generic point of S. Find the tangent plane Tp to S at p.
(b) Show that the intersection S Tp consists of two lines.
4.8.8. (a) Let A and be nonzero constants. Solve the one-variable differential
equation
z (t) = Aez(t) , z(0) = 0.
(b) The pheromone concentration in the plane is given by f (x, y) = e2x +
y
4e . What path does a bug take, starting from the origin?
4.8.9. (a) Sketch some level sets and integral curves for the function f (x, y) =
x2 + y. Find the integral curves analytically if you can.
(b) Sketch some level sets and integral curves for the function f (x, y) = xy.
Find the integral curves analytically if you can.
4.8.10. Recall the function f : R2 R whose graph is the crimped sheet,
( 2
x y
2 2 if (x, y) 6= (0, 0),
f (x, y) = x +y
0 if (x, y) = (0, 0).
200 4 The Derivative

(a) Show that f is continuous at (0, 0).

(b) Find the partial derivatives D1 f (0, 0) and D2 f (0, 0).
(c) Let d be any unit vector in R2 (thus d takes the form d = (cos , sin )
for some R). Show that Dd f (0, 0) exists by finding it.
(d) Show that in spite of (c), f is not differentiable at (0, 0). (Use your re-
sults from parts (b) and (c) to contradict Theorem 4.8.2.) Thus, the existence
of every directional derivative at a point is not sufficient for differentiability
at the point.
4.8.11. Define f : R2 R by
(
1 if y = x2 and (x, y) 6= (0, 0)
f (x, y) =
0 otherwise.

(a) Show that f is discontinuous at (0, 0). It follows that f is not differen-
tiable at (0, 0).
(b) Let d be any unit vector in R2 . Show that Dd f (0, 0) = 0. Show that
consequently the formula Dd f (0, 0) = hf (0, 0), di holds for every unit vec-
tor d. Thus, the existence of every directional derivative at a point, and the
fact that each directional derivative satisfies the formula are still not sufficient
for differentiability at the point.
4.8.12. Fix two real numbers a and b satisfying 0 < a < b. Define a mapping
T = (T1 , T2 , T3 ) : R2 R3 by
T (s, t) = ((b + a cos s) cos t, (b + a cos s) sin t, a sin s).
(a) Describe the shape of the set in R3 mapped to by T . (The answer will
explain the name T .)
(b) Find the points (s, t) R2 such that T1 (s, t) = 02 . The points map
to only four image points p under T . Show that one such p is a maximum
of T1 , another is a minimum, and the remaining two are saddle points.
(c) Find the points(s, t) R2 such that T3 (s, t) = 02 . To what points q
do these (s, t) map under T ? Which such q are maxima of T3 ? Minima? Saddle
points?

4.9 Summary
The multivariable derivative is defined as a linear mapping by an intrinsic
characterization. The characterization shows that the derivative is unique and
that it satisfies the Chain Rule. Looking at cross sections shows that if the
derivative exists then the entries of its matrix are the partial derivatives of the
component functions. Conversely, if the partial derivatives exist and behave
well then the derivative exists. The derivative in coordinates gives the Chain
Rule in coordinates. The multivariable derivative figures in solving various
kinds of problems:
 4.9 Summary 201

changing variables in a partial differential equation,

optimizing a scalar-valued function of many variables,
finding the tangent plane to a surface at a point,
determining integral curves.
The next chapter will consider another kind of problem, optimization with
constraints.
5
Inverse and Implicit Functions

The question of whether a mapping f : A Rn (where A Rn ) is globally

invertible is beyond the local techniques of differential calculus. However, a lo-
cal theorem is finally in reach. The idea sounds plausible: if the derivative of f
is invertible at the point a then f itself, being well approximated near a by its
derivative, should also be invertible in the small. However, it is by no means
a general principle that an approximated object must inherit the properties
of the object approximating it. On the contrary, mathematics often approx-
imates complicated objects by simpler ones. For instance, Taylors Theorem
approximates any function that has many derivatives by a polynomial, but
this does not make the function itself a polynomial as well.
To further illustrate the issue via an example, consider an argument in
support of the one-variable Critical Point Theorem. Let f : A R (where
A R) be differentiable, and let f (a) be positive at an interior point a of A.
We might reason as follows:
f can not have a maximum at a because the tangent line to the graph
of f at (a, f (a)) has a positive slope, so that as we move our input
rightward from a, we climb.
But this reasoning is vague. What do we climb, the tangent line or the graph?
The argument linearizes the question by fitting the tangent line through the
graph, and then it solves the linearized problem instead by checking whether
we climb the tangent line rather than whether we climb the graph. The calcu-
lus is light and graceful. But strictly speaking, part of the argument is tacit:
Since the tangent line closely approximates the graph near the point of
tangency, the fact that we climb the tangent line means that we climb
the graph as well for a while.
And the tacit part of the argument is not fully quantitative. How does the
climbing property of the tangent line transfer to the graph? The Mean Value
Theorem, and a stronger hypothesis that f is positive about a as well as at a,
resolve the question, since for x slightly larger than a,
204 5 Inverse and Implicit Functions

f (x) f (a) = f (c)(x a) for some c (a, x),

and the right side is the product of two positive numbers, hence positive. But
the Mean Value Theorem is an abstract existence theorem (for some c)
whose proof relies on foundational properties of the real number system. Thus,
moving from the linearized problem to the actual problem is far more sophis-
ticated technically than linearizing the problem or solving the linearized prob-
lem. In sum, this one-variable example is meant to amplify the point of the
preceding paragraph, that (now returning to n dimensions) if f : A Rn
has an invertible derivative at a then the Inverse Function Theoremthat f
itself is invertible in the small near ais surely inevitable, but its proof will
be technical and require strengthening our hypotheses.
Already in the one-variable case, the Inverse Function Theorem relies on
foundational theorems about the real number system, on a property of con-
tinuous functions, and on a foundational theorem of differential calculus. We
quickly review the ideas. Let f : A R (where A R) be a function, let
a be an interior point of A, and let f be continuously differentiable on some
interval about a, meaning that f exists and is continuous on the interval.
Suppose that f (a) > 0. Since f is continuous about a, the Persistence of In-
equality principle (Proposition 2.3.10) says that f is positive on some closed
interval [a , a + ] about a. By an application of the Mean Value Theorem
as in the previous paragraph, f is therefore strictly increasing on the interval,
and so its restriction to the interval does not take any value twice. By the
Intermediate Value Theorem, f takes every value from f (a ) to f (a + )
on the interval. Therefore f takes every such value exactly once, making it
locally invertible. A slightly subtle point is that the inverse function f 1 is
continuous at f (a), but then a purely formal calculation with difference quo-
tients will verify that the derivative of f 1 exists at f (a) and is 1/f (a). Note
how heavily this proof relies on the fact that R is an ordered field. A proof of
the multivariable Inverse Function Theorem must use other methods.
Although the proof to be given in this chapter is technical, its core idea
is simple common sense. Let a mapping f be given that takes x-values to y-
values and in particular takes a to b. Then the local inverse function must take
y-values near b to x-values near a, taking each such y back to the unique x
that f took to y in the first place. We need to determine conditions on f
that make us believe that a local inverse exists. As explained above, the basic
condition is that the derivative of f at agiving a good approximation of f
near a, but easier to understand than f itselfshould be invertible, and the
derivative should be continuous as well. With these conditions in hand, an
argument similar to the one-variable case (though more painstaking) shows
that f is locally injective:
Given y near b, there is at most one x near a that f takes to y.
So the remaining problem is to show that f is locally surjective:
Given y near b, show that there is some x near a that f takes to y.
 5.1 Preliminaries 205

This problem decomposes to two subproblems. First:

Given y near b, show that there is some x near a that f takes closest to y.
Then:
Show that f takes this particular x exactly to y.
And once the appropriate environment is established, solving each subprob-
lem is just a matter of applying the main theorems from the previous three
chapters.
Not only does the Inverse Function Theorem have a proof that uses so
much previous work from this course so nicely, it also has useful consequences.
It leads easily to the Implicit Function Theorem, which answers a different
question: When does a set of constraining relations among a set of variables
make some of the variables dependent on the others? The Implicit Function
Theorem in turn fully justifies (rather than linearizing) the Lagrange multi-
plier method, a technique for solving optimization problems with constraints.
As discussed in the preface to these notes, optimization with constraints has
no one-variable counterpart, and it can be viewed as the beginning of calculus
on curved spaces.

5.1 Preliminaries

The basic elements of topology in Rn -balls; limit points; closed, bounded,

and compact setswere introduced in section 2.4 to provide the environment
for the Extreme Value Theorem. A little more topology is now needed before
we proceed to the Inverse Function Theorem. Recall that for any point a Rn
and any radius > 0, the -ball at a is the set

B(a, ) = {x Rn : |x a| < }.

Recall also that a subset of Rn is called closed if it contains all of its limit
points. Not unnaturally, a subset S of Rn is called open if its complement
S c = Rn S is closed. A set, however, is not a door: it can be neither open
or closed, and it can be both open and closed. (Examples?)

Proposition 5.1.1 (-balls Are Open). For any a Rn and any > 0,
the ball B(a, ) is open.

Proof. Let x be any point in B(a, ), and set = |xa|, a positive number.
The Triangle Inequality shows that B(x, ) B(a, ) (exercise 5.1.1), and
therefore x is not a limit point of the complement B(a, )c . Consequently all
limit points of B(a, )c are in fact elements of B(a, )c , which is thus closed,
making B(a, ) itself open.

206 5 Inverse and Implicit Functions

This proof shows that any point x B(a, ) is an interior point. In fact,
an equivalent definition of open is that a subset of Rn is open if each of its
points is interior (exercise 5.1.2).
The closed -ball at a, denoted B(a, ), consists of the corresponding
open ball with its edge added in,

B(a, ) = {x Rn : |x a| }.

The boundary of the closed ball B(a, ), denoted B(a, ), is the points on
the edge,
B(a, ) = {x Rn : |x a| = }.
(See figure 5.1.) Any closed ball B and its boundary B are compact sets
(exercise 5.1.3).

Figure 5.1. Open ball, closed ball, and boundary

Let f : A Rm (where A Rn ) be continuous, let W be an open subset

of Rm , and let V be the set of all points in A that f maps into W ,

V = {x A : f (x) W }.

The set V is called the inverse image of W under f ; it is often denoted

f 1 (W ), but this is a little misleading since f need not actually have an
inverse mapping f 1 . For example, if f : R R is the squaring function
f (x) = x2 , then the inverse image of [4, 9] is [3, 2] [2, 3], and this set is
denoted f 1 ([4, 9]) even though f has no inverse. (See figure 5.2, in which
f is not the squaring function, but the inverse image f 1 (W ) also has two
components.) The inverse image concept generalizes an idea that we saw in
section 4.8: the inverse image of a one-point set under a mapping f is a level
set of f , as in Definition 4.8.3.
Although the forward image under a continuous function of an open set
need not be open (exercise 5.1.4), inverse images behave more nicely. The
connection between continuous mappings and open sets is

Theorem 5.1.2 (Inverse Image Characterization of Continuity). Let

f : A Rm (where A is an open subset of Rn ) be continuous. Let W Rm
be open. Then f 1 (W ), the inverse image of W under f , is open.
 5.1 Preliminaries 207

f 1 (W ) f 1 (W )

Figure 5.2. Inverse image with two components

Proof. Let a be a point of f 1 (W ). We want to show that it is an interior

point. Let w = f (a), a point of W . Since W is open, some ball B(w, ) is
contained in W . Consider the function

g : A R, g(x) = |f (x) w|.

This function is continuous and it satisfies g(a) = > 0, and so by a slight

variant of the Persistence of Inequality principle (Proposition 2.3.10) there
exists a ball B(a, ) A on which g remains positive. That is,

f (x) B(w, ) for all x B(a, ).

Since B(w, ) W , this shows that B(a, ) f 1 (W ), making a an interior

point of f 1 (W ) as desired.

The converse to Theorem 5.1.2 is also true and is exercise 5.1.8. We need
one last technical result for the proof of the Inverse Function Theorem.

Lemma 5.1.3 (Difference Magnification Lemma). Let B be a closed ball

in Rn and let g be a differentiable mapping from an open superset of B in Rn
back to Rn . Suppose that there is a number c such that |Dj gi (x)| c for all
i, j {1, , n} and all x B. Then

x) g(x)| n2 c|
|g( x x| B.
for all x, x

A comment about the lemmas environment might be helpful before we go

into the details of the proof. We know that continuous mappings behave well
on compact sets. On the other hand, since differentiability is sensible only at
interior points, differentiable mappings behave well on open sets. And so, to
work effectively with differentiability, we want a mapping on a domain that is
open, allowing differentiability everywhere, but then we restrict our attention
to a compact subset of the domain so that continuity (which follows from
differentiability) will behave well too. The closed ball and its open superset
in the lemma arise from these considerations.
208 5 Inverse and Implicit Functions

B. The Size Bounds give

Proof. Consider any two points x, x
n
X
|g(x) g(
x)| |gi (
x) gi (x)|,
i=1

and so to prove the lemma it suffices to prove that

|gi (
x) gi (x)| nc|
x x| for i = 1, , n.

Thus we have reduced the problem from vector output to scalar output. To
create an environment of scalar input as well, make the line segment from x
to x
the image of a function of one variable,

: [0, 1] Rn , x x).
(t) = x + t(

Note that (0) = x, (1) = x , and (t) = x x for all t (0, 1). Fix any
i {1, , n} and consider the restriction of gi to the segment, a scalar-valued
function of scalar input,

: [0, 1] R, (t) = (gi )(t).

Thus (0) = gi (x) and (1) = gi (

x). By the Mean Value Theorem,

x) gi (x) = (1) (0) = (t)

gi ( for some t (0, 1),

and so since = gi the Chain Rule gives

x) gi (x) = (gi ) (t) = gi ((t)) (t) = gi ((t))(

gi ( x x).

Since gi ((t)) is a row vector and x

x is a column vector, the last quantity
in the previous display is their inner product. Hence the display and the
CauchySchwarz Inequality give

x) gi (x)| |gi ((t))| |

|gi ( x x|.

For each j, the jth entry of the vector gi ((t)) is the partial derivative
Dj gi ((t)). And we are given that |Dj gi ((t))| c, so the Size Bounds show
that |gi ((t))| nc and therefore

|gi (
x) gi (x)| nc|
x x|.

As explained at the beginning of the proof, the result follows.

Exercises

5.1.1. Let x B(a; ) and let = |x a|. Explain why > 0 and why
B(x; ) B(a; ).
 5.1 Preliminaries 209

5.1.2. Show that a subset of Rn is open if and only if each of its points is
interior.
5.1.3. Prove that any closed ball B is indeed a closed set, as is its boundary
B. Show that any closed ball and its boundary are also bounded, hence
compact.
5.1.4. Find a continuous function f : Rn Rm and an open set A Rn
such that the image f (A) Rm of A under f is not open. Feel free to choose
n and m.
5.1.5. Define f : R R by f (x) = x3 3x. Compute f (1/2). Find
f 1 ((0, 11/8)), f 1 ((0, 2)), f 1 ((, 11/8) (11/8, )). Does f 1 exist?
5.1.6. Show that for f : Rn Rm and B Rm , the inverse image of the
complement is the complement of the inverse image,

f 1 (B c ) = f 1 (B)c .

Does the analogous formula hold for forward images?

5.1.7. If f : Rn Rm is continuous and B Rm is closed, show that f 1 (B)
is closed. What does this say about the level sets of continuous functions?
5.1.8. Prove the converse to Theorem 5.1.2: If f : A Rm (where A Rn
is open) is such that for any open W Rm also f 1 (W ) A is open, then f
is continuous.
5.1.9. Let a and b be real numbers with a < b. Let n > 1, and suppose that
the mapping g : [a, b] Rn is continuous and that g is differentiable on
the open interval (a, b). It is tempting to generalize the Mean Value Theorem
(Theorem 1.2.3) to the assertion

g(b) g(a) = g (c)(b a) for some c (a, b). (5.1)

The assertion is grammatically meaningful, since it posits an equality between

two n-vectors. The assertion would lead to a slight streamlining of the proof
of Lemma 5.1.3, since there would be no need to reduce to scalar output.
However, the assertion is false.
(a) Let g : [0, 2] R2 be g(t) = (cos t, sin t). Show that (5.1) fails for
this g. Describe the situation geometrically.
(b) Let g : [0, 2] R3 be g(t) = (cos t, sin t, t). Show that (5.1) fails for
this g. Describe the situation geometrically.
(c) Here is an attempt to prove (5.1): Let g = (g1 , , gn ). Since each gi
is scalar-valued, we have for i = 1, , n by the Mean Value Theorem,

gi (b) gi (a) = gi (c)(b a) for some c (a, b).

Assembling the scalar results gives the desired vector result.

What is the error here?
210 5 Inverse and Implicit Functions

5.2 The Inverse Function Theorem

Theorem 5.2.1 (Inverse Function Theorem). Let A be an open subset
of Rn , and let f : A Rn have continuous partial derivatives at every point
of A. Let a be a point of A. Suppose that det f (a) 6= 0. Then there is an
open set V A containing a and an open set W Rn containing f (a) such
that f : V W has a continuously differentiable inverse f 1 : W V .
For each y = f (x) W , the derivative of the inverse is the inverse of the
derivative,
D(f 1 )y = (Dfx )1 .

Before the proof, it is worth remarking that the formula for the derivative
of the local inverse, and the fact that the derivative of the local inverse is
continuous, are easy to establish once everything else is in place. If the local
inverse f 1 of f is known to exist and to be differentiable, then for any x V
the fact that the identity mapping is its own derivative combines with the
chain rule to say that

idn = D(idn )x = D(f 1 f )x = D(f 1 )y Dfx where y = f (x),

and similarly idn = Dfx (Df 1 )y , where this time idn is the identity mapping
on y-space. The last formula in the theorem follows. In terms of matrices, the
formula is
(f 1 ) (y) = f (x)1 where y = f (x).
This formula combines with Corollary 3.7.3 (the entries of the inverse matrix
are continuous functions of the entries of the matrix) to show that since the
mapping is continuously differentiable and the local inverse is differentiable,
the local inverse is continuously differentiable. Thus we need to show only
that the local inverse exists and is differentiable.

Proof. The proof begins with a simplification. Let T = Dfa , a linear map-
ping from Rn to Rn that is invertible because its matrix f (a) has nonzero
determinant. Let
f = T 1 f.
By the chain rule, the derivative of f at a is

Dfa = D(T 1 f )a = D(T 1 )f (a) Dfa = T 1 T = idn .

Also, suppose we have a local inverse g of f, so that

g f = idn near a

and
f g = idn near f(a).
 5.2 The Inverse Function Theorem 211

The situation is shown in the following diagram, in which V is an open set

f is an open set con-
containing a, W is an open set containing f (a), and W
1
taining T (f (a)) = f (a).

f T 1
V^ /W o f
/W
T

The diagram shows that the way to invert f locally, going from W back to V ,
f : g = g T 1 . Indeed, since f = T f,
is to proceed through W

g T 1 ) (T f) = idn near a,
g f = (

and, since T 1 (f (a)) = f(a),

f g = (T f) (
g T 1 ) = idn near f (a).

That is, to invert f , it suffices to invert f. And if g is differentiable then so

is g = g T 1 . The upshot is that we may prove the theorem for f rather
than f . Equivalently, we may assume with no loss of generality that Dfa =
idn and therefore that f (a) = In . This normalization will let us carry out
a clean, explicit computation in the following paragraph. (Note: With the
normalization complete, our use of the symbol g to denote a local inverse of f
now ends. The mapping to be called g in the following paragraph is unrelated
to the local inverse g in this paragraph.)
Next we find a closed ball B around a where the behavior of f is somewhat
controlled by the fact that f (a) = In . More specifically, we will quantify
the idea that since f (x) In for x near a, also f ( x) f (x) x x for
x, x
near a. Recall that the (i, j)th entry of In is ij and that det(In ) = 1.
As x varies continuously near a, the (i, j)th entry Dj fi (x) of f (x) varies
continuously near ij , and so the scalar det f (x) varies continuously near 1.
Since Dj fi (a) ij = 0 and since det f (a) = 1, applying the Persistence of
Inequality Principle (Proposition 2.3.10) n2 + 1 times shows that there exists
a closed ball B about a small enough that
1
|Dj fi (x) ij | < for all i, j {1, , n} and x B (5.2)
2n2
and
det f (x) 6= 0 for all x B. (5.3)
Let g = f idn , a differentiable mapping near a, whose Jacobian matrix at x,
g (x) = f (x) In , has (i, j)th entry Dj gi (x) = Dj fi (x) ij . Equation (5.2)
and Lemma 5.1.3 (with c = 1/(2n2 )) show that for any two points x and x
in B,
212 5 Inverse and Implicit Functions

x) g(x)| 21 |
|g( x x|,
and therefore, since f = idn + g,
|f (
x) f (x)| = |(
x x) + (g(
x) g(x))|
|
x x| |g(
x) g(x)|
x x| 12 |
| x x| (by the previous display)
1
= 2 |
x x|.

The previous display shows that f is injective on B, i.e., any two distinct
points of B are taken by f to distinct points of Rn . For future reference, we
note that the result of the previous calculation rearranges as
|
x x| 2|f (
x) f (x)| B.
for all x, x (5.4)
The boundary B of B is compact, and so is the image set f (B) since f
is continuous. Also, f (a)
/ f (B) since f is injective on B. And f (a) is not
a limit point of f (B) since f (B), being compact, is closed. Consequently,
some open ball B(f (a), 2) contains no point from f (B). (See figure 5.3.)

2
f
a f (a)

B f (B)

Figure 5.3. Ball about f (a) away from f (B)

Let W = B(f (a), ), the open ball with radius less than half the distance
from f (a) to f (B). Thus
|y f (a)| < |y f (x)| for all y W and x B. (5.5)
That is, every point y of W is closer to f (a) than it is to any point of f (B).
(See figure 5.4.)
The goal now is to exhibit a mapping on W that inverts f near a. In
other words, the goal is to show that for each y W , there exists a unique x
interior to B such that f (x) = y. So fix an arbitrary y W . Define a function
: B R that measures for each x the square of the distance from f (x)
to y,
Xn
(x) = |y f (x)|2 = (yi fi (x))2 .
i=1
 5.2 The Inverse Function Theorem 213

f y
W

x f (x)

Figure 5.4. Ball closer to f (a) than to f (B)

The idea is to show that for one and only one x near a, (x) = 0. Since
modulus is always nonnegative, the x we seek must minimize . As mentioned
at the beginning of the chapter, this simple idea inside all the technicalities
is the heart of the proof: the x to be taken to y by f must be the x that is
taken closest to y by f .
The function is continuous and B is compact, so the Extreme Value The-
orem guarantees that does indeed take a minimum on B. Condition (5.5)
guarantees that takes no minimum on the boundary B. Therefore the
minimum of must occur at an interior point x of B; this interior point x
must be a critical point of , so all partial derivatives of vanish at x. Thus
by the Chain Rule,
n
X
0 = Dj (x) = 2 (yi fi (x))Dj fi (x) for j = 1, , n.
i=1

This condition is equivalent to the matrix equation

D1 f1 (x) D1 fn (x) y1 f1 (x) 0
.. . . .. .. ..
. . . . =
.
Dn f1 (x) Dn fn (x) yn fn (x) 0
or
f (x)T (y f (x)) = 0n .
But det f (x)T = det f (x) 6= 0 by condition (5.3), so f (x)T is invertible and
the only solution of the equation is y f (x) = 0n . Thus our x is the desired x
interior to B such that y = f (x). And there is only one such x because f is
injective on B. We no longer need the boundary B, whose role was to make
a set compact. In sum, we now know that f is injective on B and that f (B)
contains W .
Let V = f 1 (W ) B, the set of all points x B such that f (x) W .
(See figure 5.5.) By the inverse image characterization of continuity (Theo-
rem 5.1.2), V is open. We have established that f : V W is inverted by
f 1 : W V .
214 5 Inverse and Implicit Functions

f
V f 1 W

Figure 5.5. The sets V and W of the Inverse Function Theorem

The last thing to prove is that f 1 is differentiable on W . Again, reducing

the problem makes it easier. By (5.3), the condition det f (x) 6= 0 is in effect
at each x V . Therefore a is no longer a distinguished point of V , and it
suffices to prove that the local inverse f 1 is differentiable at f (a). To reduce
the problem to working at the origin, consider the mapping f defined by the
formula f(x) = f (x + a) b. Since f (a) = b it follows that f(0n ) = 0n ,
and since f is f up to prepended and postpended translations, f is locally
invertible at 0n and its derivative there is Df0 = Dfa = idn . The upshot is
that in proving that f 1 is differentiable at f (a), there is no loss of generality
in normalizing to a = 0n and f (a) = 0n while also retaining the normalization
that Dfa is the identity mapping.
So now we have that f (0n ) = 0n = f 1 (0n ) and

f (h) h = o(h),

and we want to show that

f 1 (k) k = o(k).

For any point k W , let h = f 1 (k). Note that |h| 2|k| by condition (5.4)
with x
= h and x = 0n so that f ( x) = k and f (x) = 0n , and thus h = O(k).
So now we have

f 1 (k) k = (f (h) h) = o(h) = o(h) = o(O(k)) = o(k),

exactly as desired. That is, f 1 is indeed differentiable at 0n with the identity

mapping for its derivative. For a non-normalized proof that f 1 is differen-
tiable on W , see exercise 5.2.9.

Note the range of mathematical skills that this proof of the Inverse Func-
tion Theorem required. The ideas were motivated and guided by pictures,
but the actual argument was symbolic. At the level of fine detail, we nor-
malized the derivative to the identity in order to reduce clutter, we made an
adroit choice of quantifier in choosing a small enough B to apply the Differ-
ence Magnification Lemma with c = 1/(2n2 ), and we used the full Triangle
 5.2 The Inverse Function Theorem 215

Inequality to obtain (5.4). This technique sufficed to prove that f is locally

injective. Since the proof of the Difference Magnification Lemma used the
Mean Value Theorem many times, the role of the Mean Value Theorem in
the multivariable Inverse Function Theorem is thus similar to its role in the
one-variable proof reviewed at the beginning of the chapter. However, while
the one-variable proof that f is locally surjective relied on the Intermediate
Value Theorem, the multivariable argument was far more elaborate. The idea
was that the putative x taken by f to a given y must be the actual x taken
by f closest to y. We exploited this idea by working in broad strokes:
The Extreme Value Theorem from chapter 2 guaranteed that there is such
an actual x.
The Critical Point Theorem and then the Chain Rule from chapter 4
described necessary conditions associated to x.
And finally, the Linear Invertibility Theorem from chapter 3 showed that
f (x) = y as desired. Very satisfyingly, the hypothesis that the derivative is
invertible sealed the argument that the mapping itself is locally invertible.
Indeed, the proof of local surjectivity used nearly every significant result from
chapters 2 through 4 of these notes.
For an example, define f : R2 R2 by f (x, y) = (x3 2xy 2 , x + y). Is
f locally invertible at (1, 1)? If so, what is the best affine approximation to
the inverse near f (1, 1)? To answer the first question, calculate the Jacobian
 2    
3x 2y 2 4xy  14
f (1, 1) =  = .
1 1  11
(x,y)=(1,1)
 
This matrix is invertible with inverse f (1, 1)1 = 31 11 14 . Therefore
f is locally invertible at (1, 1) and the affine approximation to f 1 near
f (1, 1) = (1, 0) is
    
1 1 1 4 h 1 4 1 1
f 1 (1 + h, 0 + k) + = (1 h + k, 1 + h k).
1 3 1 1 k 3 3 3 3

The actual inverse function f 1 about (1, 0) may not be clear, but the In-
verse Function Theorem guarantees its existence, and its affine approximation
is easy to find.

Exercises

5.2.1. Define f : R2 R2 by f (x, y) = (x3 + 2xy + y 2 , x2 + y). Is f locally

invertible at (1, 1)? If so, what is the best affine approximation to the inverse
near f (1, 1)?

5.2.2. Same question for f (x, y) = (x2 y 2 , 2xy) at (2, 1).

216 5 Inverse and Implicit Functions

5.2.3. Same question for C(r, ) = (r cos , r sin ) at (1, 0).

5.2.4. Same question for C(, , ) = ( cos sin , sin sin , cos ) at
(1, 0, /2).
5.2.5. At what points (a, b) R2 is each of the following mappings guaranteed
to be locally invertible by the Inverse Function Theorem? In each case, find
the best affine approximation to the inverse near f (a, b).
(a) f (x, y) = (x + y, 2xy 2 ).
(b) f (x, y) = (sin x cos y + cos x sin y, cos x cos y sin x sin y).
5.2.6. Define f : R2 R2 by f (x, y) = (ex cos y, ex sin y). Show that f is
locally invertible at each point (a, b) R2 , but that f is not globally invertible.
Let (a, b) = (0, 3 ); let (c, d) = f (a, b); let g be the local inverse to f near (a, b).
Find an explicit formula for g, compute g (c, d) and verify that it agrees with
f (a, b)1 .
5.2.7. If f and g are functions from R3 to R, show that the mapping F =
(f, g, f + g) : R3 R3 , does not have a differentiable local inverse anywhere.
5.2.8. Define f : R R by
(
x + 2x2 sin x1 if x 6= 0,
f (x) =
0 if x = 0.

(a) Show that f is differentiable at x = 0 and that f (0) 6= 0. (Since this

is a one-dimensional problem you may verify the old definition of derivative
rather than the new one.)
(b) Despite the result from (a), show that f is not locally invertible at
x = 0. Why doesnt this contradict the Inverse Function Theorem?
5.2.9. The proof of the Inverse Function Theorem ended with a normalized
argument that the inverse function on W is again differentiable. Supply expla-
nation as necessary to the non-normalized version of the argument, as follows.
Let y be a fixed point of W , and let y + k lie in W as well. Take x = f 1 (y)
in V , and let f 1 (y + k) = x + h, thus defining h = f 1 (y + k)) f 1 (y). We
know that f (x) is invertible and that
f (x + h) f (x) f (x)h = o(h).
We want to show that
f 1 (y + k) f 1 (y) f (x)1 k = o(k).
Compute,
f 1 (y + k) f 1 (y) f (x)1 k = h f (x)1 (f (x + h) f (x))
= h f (x)1 (f (x)h + o(h))
= f (x)1 o(h).
 5.3 The Implicit Function Theorem 217

Using (5.4), |h| = |x + h x| 2|f (x + h) f (x)| = 2|k| = O(k), so we have

f 1 (y + k) f 1 (y) f (x)1 k = f (x)1 o(O(k)) = f (x)1 o(k).

Multiplication by the fixed matrix f (x)1 is a linear mapping, and any

linear mapping is O of its input. Altogether,

f 1 (y + k) f 1 (y) f (x)1 k = f (x)1 o(k) = O(o(k)) = o(k),

as desired.

5.3 The Implicit Function Theorem

Let n and c be positive integers with c n, and let r = n c. This section
addresses the following question:
When do c conditions on n variables locally specify c of the variables
in terms of the remaining r variables?
The symbols in this question will remain in play throughout the section. That
is,
n = r + c is the total number of variables,
c is the number of conditions, i.e., the number of constraints on the vari-
ables, and therefore the number of variables that might be dependent on
the others,
and r is the number of remaining variables and therefore the number of
variables that might be free.
The word conditions (or constraints) provides a mnemonic for the symbol c,
and similarly remaining (or free) provides a mnemonic for r.
The question can be rephrased:
When is a level set locally a graph?
To understand the rephrasing, we begin by reviewing the idea of a level set,
given here in a slightly more general form than in Definition 4.8.3.

Definition 5.3.1 (Level Set). Let g : A Rm (where A Rn ) be a

mapping. A level set of g is the set of points in A that map under g to some
fixed vector w in Rm ,
L = {v A : g(v) = w}.
That is, L is the inverse image under g of the one-point set {w}.

Also we review the argument in section 4.8 that every graph is a level
set. Let A0 be a subset of Rr , and let f : A0 Rc be any mapping. Let
A = A0 Rc (a subset of Rn ) and define a second mapping g : A Rc ,
218 5 Inverse and Implicit Functions

g(x, y) = f (x) y, (x, y) A0 Rc .

Then the graph of f is

graph(f ) = {(x, y) A0 Rc : y = f (x)}

= {(x, y) A : g(x, y) = 0c },

and this is the set of inputs to g that g takes to 0c , a level set of g as desired.
Now we return to rephrasing the question at the beginning of this sec-
tion. Let A be an open subset of Rn , and let a mapping g : A Rc have
continuous partial derivatives at every point of A. Points of A can be written

(x, y), x Rr , y Rc .

(Throughout this section, we routinely will view an n-vector as the concate-

nation of an r-vector and c-vector in this fashion.) Consider the level set

L = {(x, y) A : g(x, y) = 0c }.

The question was whether the c scalar conditions g(x, y) = 0c on the n = c+r
scalar entries of (x, y) define the c scalars of y in terms of the r scalars of x
near (a, b). That is, the question is whether the vector relation g(x, y) = 0c
for (x, y) near (a, b) is equivalent to a vector relation y = (x) for some
mapping that takes r-vectors near a to c-vectors near b. This is precisely
the question of whether the level set L is locally the graph of such a mapping .
If the answer is yes, then we would like to understand as well as possible
by using the techniques of differential calculus. In this context we view the
mapping as implicit in the condition g = 0c , explaining the name of the
pending Implicit Function Theorem.
The first phrasing of the question, whether c conditions on n variables
specify c of the variables in terms of the remaining r variables, is easy to
answer when the conditions are affine. Affine conditions take the matrix form
P v = w where P Mc,n (R), v Rn , and w Rc , and P and w are fixed
while v is the vector of variables. Partition the matrix P into a left c-by-r
block M and a right square c-by-c block N , and partition the vector v into
its first r entries x and its last c entries y. Then the relation P v = w is
 
  x
M N = w,
y

that is,
M x + N y = w.
Assume that N is invertible. Then subtracting M x from both sides and then
left multiplying by N 1 shows that the relation is

y = N 1 (w M x).
 5.3 The Implicit Function Theorem 219

Thus, when the right c-by-c submatrix of P is invertible, the relation P v = w

explicitly specifies the last c variables of v in terms of the first r variables.
A similar statement applies to any invertible c-by-c submatrix of P and the
corresponding variables. A special case of this calculation, the linear case,
will be used throughout the section: for any M Mc,r (R), any invertible
N Mc (R), any h Rr , and any k Rc ,
 
  h
M N = 0c k = N 1 M h. (5.6)
k

When the conditions are nonaffine the situation is not so easy to analyze.
However:
The problem is easy to linearize. That is, given a point (a, b) (where a Rr
and b Rc ) on the level set {(x, y) : g(x, y) = w}, differential calculus
tells us how to describe the tangent object to the level set at the point.
Depending on the value of r, the tangent object will be a line, or a plane,
or higher-dimensional. But regardless of its dimension, it is described by
the linear conditions g (a, b)v = 0c , and these conditions take the form
that we have just considered,
 
  h
M N = 0c , M Mc,r (R), N Mc (R), h Rr , k Rc .
k

Thus if N is invertible then we can solve the linearized problem as in (5.6).

The Inverse Function Theorem says:
If the linearized inversion problem is solvable then the actual inver-
sion problem is locally solvable.
With a little work, we can use the Inverse Function Theorem to establish
the Implicit Function Theorem:
If the linearized level set is a graph then the actual level set is locally
a graph.
And in fact, the Implicit Function Theorem will imply the Inverse Function
Theorem as well.
For example, the unit circle C is described by one constraint on two vari-
ables (n = 2 and c = 1, so r = 1),

x2 + y 2 = 1.

Globally (in the large), this relation neither specifies x as a function of y nor
y as a function of x. It cant: the circle is visibly not the graph of a function
of either sortrecall the Vertical Line Test to check whether a curve is the
graph of a function y = (x), and analogously for the Horizontal Line Test.
The situation does give a function, however, if one works locally (in the small)
by looking only at part of the circle at a time. Any arc in the bottom half of
the circle is described by the function
220 5 Inverse and Implicit Functions
p
y = (x) = 1 x2 .

Similarly, any arc in the right half is described by

p
x = (y) = 1 y 2 .

Any arc in the bottom right quarter is described by both functions. (See
figure 5.6.) On the other hand, no arc of the circle about the point (a, b) =
(1, 0) is described by a function y = (x), and no arc about (a, b) = (0, 1) is
described by a function x = (y). (See figure 5.7.) Thus, about some points
(a, b), the circle relation x2 + y 2 = 1 contains the information to specify each
variable as a function of the other. These functions are implicit in the relation.
About other points, the relation implicitly defines one variable as a function
of the other, but not the second as a function of the first.

y = (x)
x = (y)

Figure 5.6. Arc of a circle

x 6= (y)

y 6= (x)

Figure 5.7. Trickier arcs of a circle

To bring differential calculus to bear on the situation, think of the circle

as a level set. Specifically, it is a level set of the function g(x, y) = x2 + y 2 ,
 5.3 The Implicit Function Theorem 221

C = {(x, y) : g(x, y) = 1}.

Let (a, b) be a point on the circle. The derivative of g at the point is

 
g (a, b) = 2a 2b .

The tangent line to the circle at (a, b) consists of the points (a + h, b + k) such
that (h, k) is orthogonal to g (a, b),
 
  h
2a 2b = 0.
k

That is,
2ah + 2bk = 0.
Thus whenever b 6= 0 we have

k = (a/b)h,

showing that on the tangent line, the second coordinate is a linear function
of the first, and the function has derivative a/b. And so on the circle it-
self near (a, b), plausibly the second coordinate is a function of the first as
well, provided that b 6= 0. Note that indeed this argument excludes the two
points (1, 0) and (1, 0) about which y is not an implicit function of x. But
about points (a, b) C where D2 g(a, b) 6= 0, the circle relation should im-
plicitly define y as a function of x. And at such points (say, on the lower
half-circle), the function is explicitly
p
(x) = 1 x2 ,

so that (x) = x/ 1 x2 = x/y (the last minus sign is present because
the square root is positive but y is negative) and in particular

(a) = a/b.

Thus (a) is exactly the slope that we found a moment earlier by solving
the linear problem g (a, b)v = 0 where v = (h, k) is a column vector. That
is, using the constraint g(x, y) = 0 to set up and solve the linear problem,
making no reference in the process to the function implicitly defined by the
constraint, found the derivative (a) nonetheless. The procedure illustrates
the general idea of the pending Implicit Function Theorem:
Constraining conditions do locally define some variables implicitly in
terms of others, and the implicitly defined function can be differenti-
ated without being found explicitly.
(And returning to the circle example, yet another way to find the derivative
is to differentiate the relation x2 + y 2 = 1 at a point (a, b) about which we
assume that y = (x),
222 5 Inverse and Implicit Functions

2a + 2b (a) = 0,
so that again (a) = a/b. The reader may recall from elementary calculus
that this technique is called implicit differentiation.)
It may help the reader visualize the situation if we revisit the idea of
the previous paragraph more geometrically. Since C is a level set of g, the
gradient g (a, b) is orthogonal to C at the point (a, b). When g (a, b) has a
nonzero y-component, C should locally have a big shadow on the x-axis, from
which there is a function back to C. (See figure 5.8, in which the arrow
drawn is quite a bit shorter than the true gradient, for graphical reasons.)

Figure 5.8. Nonhorizontal gradient and x-shadow

Another set defined by a constraining relation is the unit sphere, also

specified as a level set. Let

g(x, y, z) = x2 + y 2 + z 2 .

Then the sphere is

S = {(x, y, z) : g(x, y, z) = 1}.
Imposing one condition on three variables should generally leave two of them
free (say, the first two) and define the remaining one in terms of the free ones.
That is, n = 3 and c = 1, so that r = 2. And indeed, the sphere implicitly
describes z as a function (x, y) about any point p = (a, b, c) S off the
equator, where c = 0. (So for this example we have just overridden the general
use of c as the number of constraints; here c is the third coordinate of a point on
the level set.) The equator is precisely the points where D3 g(p) = 2c vanishes.
Again geometry makes this plausible. The gradient g (p) is orthogonal to S
at p. When g (p) has a nonzero z-component, S should locally have a big
shadow in the (x, y)-plane from which there is a function back to S and then
to the z-axis. (See figure 5.9.)
 5.3 The Implicit Function Theorem 223
z

y
x

Figure 5.9. Function from the (x, y)-plane to the z-axis via the sphere

The argument based on calculus and linear algebra to suggest that near
points (a, b, c) S such that D3 g(a, b, c) 6= 0, z is implicitly a function (x, y)
on S is similar to the case of the circle. The derivative of g at the point is
 
g (a, b, c) = 2a 2b 2c .

The tangent plane to the sphere at (a, b, c) consists of the points (a + h, b +

k, c + ) such that (h, k, ) is orthogonal to g (a, b, c),

  h
2a 2b 2c k = 0.

That is,
2ah + 2bk + 2c = 0.
Thus whenever c 6= 0 we have

= (a/c)h (b/c)k,

showing that on the tangent plane, the third coordinate is a linear function
of the first two, and the function has partial derivatives a/c and b/c.
And so on the sphere itself near (a, b, c), plausibly the third coordinate is a
function of the first two as well, provided that c 6= 0. This argument excludes
points on the equator, about which z is not an implicit function of (x, y). But
about points (a, b, c) S where D3 g(a, b, c) 6= 0, the sphere relation should
implicitly define z as a function of (x, y). And at such points (say, on the
upper hemisphere), the function is explicitly
p
(x, y) = 1 x2 y 2 ,
224 5 Inverse and Implicit Functions
p p
so that (x, y) = [x/ 1 x2 y 2 y/ 1 x2 y 2 ] = [x/z y/z] and in
particular  
(a, b) = a/c b/c .
The partial derivatives are exactly as predicted by solving the linear problem
g (a, b, c)v = 0, where v = (h, k, ) is a column vector, with no reference to .
(As with the circle, a third way to find the derivative is to differentiate the
sphere relation x2 + y 2 + z 2 = 1 at a point (a, b, c) about which we assume
that z = (x, y), differentiating with respect to x and then with respect to y,

2a + 2cD1 (a, b) = 0, 2b + 2cD2 (a, b) = 0.

Again we obtain (a, b) = [a/c b/c].)

Next consider the intersection of the unit sphere and the 45-degree plane
z = y. The intersection is a great circle, again naturally described as a level
set. That is, if we consider the mapping

g : R3 R2 , g(x, y, z) = (x2 + y 2 + z 2 , y + z),

then the great circle is a level set of g,

GC = {(x, y, z) : g(x, y, z) = (1, 0)}.

The two conditions on the three variables should generally leave one variable
(say, the first one) free and define the other two variables in terms of it. That
is, n = 3 and c = 2, so that r = 1. Indeed, GC is a circle that is orthogonal
to the plane of the page, and away from its two points (1, 0, 0) that are
farthest in and out of the page, it does define (y, z) locally as functions of x.
(See figure 5.10.) This time we first proceed by linearizing the problem to
obtain the derivatives of the implicit function without finding the implicit
function = (1 , 2 ) itself. The derivative matrix of g at p is
 
2a 2b 2c
g (a, b, c) = .
0 1 1

The level set GC is defined by the condition that g(x, y, z) hold constant
at (1, 0) as (x, y, z) varies. Thus the tangent line to GC at a point (a, b, c)
consists of points (a + h, b + k, c + ) such that neither component function
of g is instantaneously changing in the (h, k, )-direction,

  h  
2a 2b 2c 0
k = .
0 1 1 0

The right 2-by-2 submatrix of g (a, b, c) has nonzero determinant whenever

b 6= c, that is, at all points of GC except the two aforementioned ex-
treme points (1, 0, 0). Assuming that b 6= c, let M denote the first column
 5.3 The Implicit Function Theorem 225

of g (a, b, c) and let N denote the right 2-by-2 submatrix. Then by (5.6), the
linearized problem has solution
      a
k 1 1 1 2c 2a 2b
= N M h = h= a h
2(c b) 1 2b 0 2c

(the condition c = b was used in the last step), or

a a
k= h, = h. (5.7)
2b 2c
And so for all points (a + h, b + k, c + ) on the tangent line to GC at (a, b, c),
the last two coordinate-offsets k and are specified in terms of the first co-
ordinate offset h via (5.7), and the component functions have partial deriva-
tives a/(2b) and a/(2c). (And as with the circle and the sphere, the two
partial derivatives can be obtained by implicit differentiation as well.)

Figure 5.10. y and z locally as functions of x on a great circle

To make the implicit function in the great circle relations explicit, note
that near the point p = (a, b, c) in the figure,
r r !
1 x2 1 x2
(y, z) = (1 (x), 2 (x)) = , .
2 2

At p the component functions have derivatives

a a
1 (a) = q and 2 (a) = q .
2 2
2 1a2 2 1a
2
226 5 Inverse and Implicit Functions

But 1 a2 = 2b2 = 2c2 , and b2 = b since b < 0 while c2 = c since c > 0,
so the derivatives are
a a
1 (a) = and 2 (a) = .
2b 2c
Predictably enough, these component derivatives of the true mapping defin-
ing the last two coordinates y and z in terms of the first coordinate x for points
on GC itself near p, match the implicitly calculated values shown in (5.7).
In the examples of the circle, the sphere, and the great circle, the functions
implicit in the defining relations could in fact be found explicitly. But in
general, relations may snarl the variables so badly that expressing some as
functions of the others is beyond our algebraic capacity. For instance, do the
simultaneous conditions

y 2 = ez cos(y + x2 ) and y 2 + z 2 = x2 (5.8)

define y and z implicitly in terms of x near the point (1, 1, 0)? (This point
meets both conditions.) Answering this directly by solving for y and z is
manifestly unappealing. But linearizing the problem is easy. At our point
(1, 1, 0), the mapping

g(x, y, z) = (y 2 ez cos(y + x2 ), y 2 + z 2 x2 )

has derivative matrix

  
2xez sin(y + x2 ) 2y + ez sin(y + x2 ) ez cos(y + x2 ) 
g (1, 1, 0) = 
2x 2y 2z 
(1,1,0)
 
0 2 1
= .
2 2 0

Since the right 2-by-2 determinant is nonzero, we expect that indeed y and z
are implicit functions 1 (x) and 2 (x) near (1, 1, 0). Furthermore, solving
the linearized problem as in the previous example with M and N similarly
defined suggests that if (y, z) = (x) = (1 (x), 2 (x)) then
 1       
1 2 1 0 1 0 1 0 1
(1) = N M = = = .
2 0 2 2 2 2 2 2

Thus for a point (x, y, z) = (1 + h, 1 + k, 0 + ) near (1, 1, 0) satisfying

conditions (5.8), we expect that (k, ) (h, 2h), i.e.,

for x = 1 + h. (y, z) (1, 0) + (h, 2h).

The Implicit Function Theorem fulfills these expectations.

 5.3 The Implicit Function Theorem 227

Theorem 5.3.2 (Implicit Function Theorem). Let c and n be positive

integers with n > c, and let r = n c. Let A be an open subset of Rn , and let
g : A Rc have continuous partial derivatives at every point of A. Consider
the level set
L = {v A : g(v) = 0c }.
r
c
 g(p) = 0c . Let p = (a, b) where a R and
Let p be a point of L, i.e., let
b R , and let g (p) = M N where M is the left c-by-r submatrix and N is
the remaining right square c-by-c submatrix.
If det N 6= 0 then the level set L is locally a graph near p. That is, the
condition g(x, y) = 0c for (x, y) near (a, b) implicitly defines y as a function
y = (x) where takes r-vectors near a to c-vectors near b, and in particular
(a) = b. The function is differentiable at a with derivative matrix

(a) = N 1 M.

Hence is well approximated near a by its affine approximation,

(a + h) b N 1 M h.

We make three remarks before the proof.

The condition g(x, y) = 0c could just as easily be g(x, y) = w for any fixed
point w Rc , as in our earlier examples. Normalizing to w = 0c amounts
to replacing g by g w (with no effect on g ), which we do to tidy up the
statement of the theorem.
The Implicit Function Theorem gives no information when det N = 0. In
this case, the condition g(x, y) = 0c may or may not define y in terms of x.
While the theorem strictly addresses only whether the last c of n variables
subject to c conditions depend on the first r variables, it can be suitably
modified to address whether any c variables depend on the remaining ones
by checking the determinant of a suitable c-by-c submatrix of g (p). The
modification is merely a matter of reindexing or permuting the variables,
not worth writing down formally in cumbersome notation, but the reader
should feel free to use the modified version.

Proof. Examining the derivative has already shown the theorems plausibility
in specific instances. Shoring up these considerations into a proof is easy with
a well-chosen change of variables and the Inverse Function Theorem. For the
change of variables, define
G : A Rn
as follows: for all x Rr and y Rc such that (x, y) A,

G(x, y) = (x, g(x, y)).

Note that G incorporates g, but unlike g it is a map between spaces of the

same dimension n. Note also that the augmentation that changes g into G is
228 5 Inverse and Implicit Functions

highly reversible, being the identity mapping on the x-coordinates. That is, it
is easy to recover g from G. The mapping G affects only y-coordinates, and it
is designed to take the level set L = {(x, y) A : g(x, y) = 0c } to the x-axis.
(See figure 5.11, in which the inputs and the outputs of G are shown in the
same copy of Rn .)

Rc

A
b p

G(x, y) = (x, g(x, y))

G(A) x
R n a Rr

Figure 5.11. Mapping A to Rn and the constrained set to x-space

The mapping G is differentiable at the point p = (a, b) with derivative

matrix  
Ir 0rc
G (a, b) = Mn (R).
M N
This matrix has determinant det G (a, b) = det N 6= 0, and so by the Inverse
Function Theorem G has a local inverse mapping defined near the point
G(a, b) = (a, 0c ). (See figure 5.12.) Since the first r components of G are the
identity mapping, the same holds for the inverse. That is, the inverse takes
the form
(x, y) = (x, (x, y)),
where maps n-vectors near (a, 0c ) to c-vectors near b. The inversion criterion
is that for all (x, y) near (a, b) and all (x, y) near (a, 0c ),
G(x, y) = (x, y) (x, y) = (x, y).
Equivalently, since neither G nor affects x-coordinates, for all x near a, y
near b, and y near 0c ,
g(x, y) = y y = (x, y). (5.9)
Also by the Inverse Function Theorem and a short calculation.
 
1 Ir 0rc
(a, 0c ) = G (a, b) =
N 1 M N 1
 5.3 The Implicit Function Theorem 229
y

Rc

b p

(x, y) = (x, (x, y))

x
Rn a Rr

Figure 5.12. Local inverse of G

Now we can exhibit the desired mapping implicit in the original g. Define
a mapping
(x) = (x, 0c ) for x near a. (5.10)
The idea is that locally this lifts the x-axis to the level set L where g(x, y) = 0c
and then projects horizontally to the y-axis. (See figure 5.13.) For any (x, y)
near (a, b), a specialization of condition (5.9) combines with the definition
(5.10) of to give

g(x, y) = 0c y = (x).

This equivalence exhibits y as a local function of x on the level set of g, as

desired. And since by definition (5.10), is the last c component functions of
restricted to the first r inputs to , the derivative (a) is exactly the lower
left c-by-r block of (a, 0c ), which is N 1 M . This completes the proof.

Thus the Implicit Function Theorem follows easily from the Inverse Func-
tion Theorem. The converse implication is even easier. Imagine a scenario
where somehow we know the Implicit Function Theorem but not the Inverse
Function Theorem. Let f : A Rn (where A Rn ) be a mapping that
satisfies the hypotheses for the Inverse Function Theorem at a point a A.
That is, f is continuously differentiable in an open set containing a, and
det f (a) 6= 0. Define a mapping

g : A Rn Rn , g(x, y) = f (x) y.

(This mapping should look familiar from the beginning of the section.) Let
b = f (a). Then g(a, b) = 0, and the derivative matrix of g at (a, b) is
 
g (a, b) = f (a) In .
230 5 Inverse and Implicit Functions
y

Rc

(x, 0c ) p
b

(x) (x, 0c ) = (x, (x, 0c ))

x
Rn a Rr

Figure 5.13. The implicit mapping from x-space to y-space via the level set

Since f (a) is invertible, we may apply the Implicit Function Theorem, with
the roles of c, r, and n in theorem taken by the values n, n, and 2n here, and
with the theorem modified as in the third remark before its proof so that we
are checking whether the first n variables depend on the last n values. The
theorem supplies us with a differentiable mapping defined for values of y
near b such that for all (x, y) near (a, b),

g(x, y) = 0 x = (y).

But by the definition of g, this equivalence is

y = f (x) x = (y).

That is, inverts f . Also by the Implicit Function Theorem, is differentiable

at b with derivative

(b) = f (a)1 (In ) = f (a)1

(as it must be), and we have recovered the Inverse Function Theorem. In a
nutshell, the argument converts the graph y = f (x) into a level set g(x, y) = 0,
and then the Implicit Function Theorem says that locally the level set is also
the graph of x = (y). (See figure 5.14.)
Rederiving the Inverse Function Theorem so easily from the Implicit Func-
tion Theorem is not particularly impressive, since proving the Implicit Func-
tion Theorem without citing the Inverse Function Theorem would be just as
hard as the route we took of proving the Inverse Function Theorem first. The
point is that the two theorems have essentially the same content.
We end this section with one more example. Consider the function

g : R2 R, g(x, y) = (x2 + y 2 )2 x2 + y 2
 5.3 The Implicit Function Theorem 231
y

Rn

((y), y)
b

(y)
f (x) (x, f (x))

x
R2n a Rn

Figure 5.14. The Inverse Function Theorem from the Implicit Function Theorem

and the corresponding level set, a curve in the plane,

L = {(x, y) R2 : g(x, y) = 0}.

The Implicit Function Theorem lets us analyze L qualitatively. The derivative

matrix of g is

g (x, y) = 4[x((x2 + y 2 ) 1/2) y((x2 + y 2 ) + 1/2)].

By the theorem, L is locally the graph of a function y = (x) except possibly

at its points where y((x2 + y 2 ) + 1/2) = 0, which is to say y = 0. To find all
such points is to find all x such that g(x, 0) = 0. This condition is x4 x2 = 0,
or x2 (x2 1) = 0, and so the points of L where locally it might not be the
graph of y = (x) are (0, 0) and (1, 0). Provisionally we imagine L to have
vertical tangents at these points.
Similarly, L is locally the graph of a function x = (y) except possibly at
its points where x((x2 + y 2 ) 1/2) = 0, which is to say x = 0 or x2 + y 2 = 1/2.
The condition g(0, y) = 0 is y 4 + y 2 = 0, whose only solution is y = 0. And
if x2 + y 2 p
= 1/2 then g(x, y) = 1/4 x2p + y 2 = 3/4 2x2 , which vanishes
for x = 3/8, also determining y = 1/8. Thus the points p of L p where
locally it might not be the graph of x = (y) are (0, 0) and ( 3/8, 1/8)
with the two signs independent. Provisionally we imagine L to have horizontal
tangents at these points.
However, since also we imagined a vertical tangent at (0, 0), this point
requires further analysis. Keeping only the lowest-order terms of the relation
g(x, y) = 0 gives y 2 x2 , or y x, and so L looks like two crossing lines of
slopes 1 near (0, 0). This analysis suffices to sketch L, as shown in figure 5.15.
The level set L is called a lemniscate. The lemniscate originated in astronomy,
and the study of its arc length led to profound mathematical ideas by Gauss,
Abel, and many others.
232 5 Inverse and Implicit Functions

Figure 5.15. Lemniscate

Exercises
5.3.1. Does the relation x2 + y + sin(xy) = 0 implicitly define y as a function
of x near the origin? If so, what is its best affine approximation? How about
x as a function of y and its affine approximation?
5.3.2. Does the relation xy z ln y + exz = 1 implicitly define z as a function
of (x, y) near (0, 1, 1)? How about y as a function of (x, z)? When possible,
give the affine approximation to the function.
5.3.3. Do the simultaneous conditions x2 (y 2 + z 2 ) = 5 and (x z)2 + y 2 = 2
implicitly define (y, z) as a function of x near (1, 1, 2)? If so, then what is
the functions affine approximation?
5.3.4. Same question for the conditions x2 + y 2 = 4 and 2x2 + y 2 + 8z 2 = 8
near (2, 0, 0).
5.3.5. Do the simultaneous conditions xy + 2yz = 3xz and xyz + x y = 1
implicitly define (x, y) as a function of z near (1, 1, 1)? How about (x, z) as a
function of y? How about (y, z) as a function of x? Give affine approximations
when possible.
5.3.6. Do the conditions xy 2 + xzu + yv 2 = 3 and u3 yz + 2xv u2 v 2 = 2
implicitly define (u, v) in terms of (x, y, z) near the point (1, 1, 1, 1, 1)? If so,
what is the derivative matrix of the implicitly defined mapping at (1, 1, 1)?
5.3.7. Do the conditions x2 +yu+xv +w = 0 and x+y +uvw = 1 implicitly
define (x, y) in terms of (u, v, w) near (x, y, u, v, w) = (1, 1, 1, 1, 1)? If so,
what is the best affine approximation to the implicitly defined mapping?
5.3.8. Do the conditions
2x + y + 2z + u v = 1
xy + z u + 2v = 1
yz + xz + u2 + v = 0
define the first three variables (x, y, z) as a function (u, v) near the point
(x, y, z, u, v) = (1, 1, 1, 1, 1)? If so, find the derivative matrix (1, 1).
 5.4 Lagrange Multipliers: Geometric Motivation and Specific Examples 233

5.3.9. Define g : R2 R by g(x, y) = 2x3 3x2 + 2y 3 + 3y 2 and let L be

the level set {(x, y) : g(x, y) = 0}. Find those points of L about which y need
not be defined implicitly as a function of x, and find the points about which
x need not be defined implicitly as a function of y. Describe L preciselythe
result should explain the points you found.

5.4 Lagrange Multipliers: Geometric Motivation and

Specific Examples
How close does the intersection of the planes x+y +z = 1 and xy +2z = 1
in R3 come to the origin? This question is an example of an optimization prob-
lem with constraints. The goal in such problems is to maximize or minimize
some function, but with relations imposed on its variables. Equivalently, the
problem is to optimize some function whose domain is a level set.
A geometric solution of the sample problem just given is that the planes
intersect in a line through the point p = (0, 1, 0) in direction d = (1, 1, 1)
(1, 1, 2), so the point-to-line distance formula from exercise 3.10.12 answers
the question. This method is easy and efficient.
A more generic method of solution is via substitution. The equations of
the constraining planes are x + y = 1 z and x y = 1 2z; adding gives
x = 3z/2, and subtracting gives y = 1+z/2. To finish the problem, minimize
the function d2 (z) = (3z/2)2 + (1 + z/2)2 + z 2 , where d2 denotes distance
squared from the origin. Minimizing d2 rather than d avoids square roots.
Not all constrained problems yield readily to either of these methods. The
more irregular the conditions, the less amenable they are to geometry, and the
more tangled the variables, the less readily they distill. Merely adding more
variables to the previous problem produces a nuisance: How close does the
intersection of the planes v + w + x + y + z = 1 and v w + 2x y + z = 1
in R5 come to the origin? Now no geometric procedure lies conveniently at
hand. As for substitution, linear algebra shows that

v
  w  
1 11 11
x = 1
1 1 2 1 1
y 1
z
implies

   1     x  
v 1 1 1 1 1 1 3x/2 z
= y = .
w 1 1 1 2 1 1 1 + x/2 y
z
2 2
Since the resulting function d2 (x, y, z) = (3x/2 z) +(1 + x/2 y) +x2 +
y 2 + z 2 is quadratic, partial differentiation and more linear algebra will find
its critical points. But the process is getting tedious.
234 5 Inverse and Implicit Functions

Lets step back from specifics (but we will return to the currently unre-
solved example soon) and consider in general the necessary nature of a critical
point in a constrained problem. The discussion will take place in two stages:
first we consider the domain of the problem, and then we consider the critical
point.
The domain of the problem is the points in n-space that satisfy a set of c
constraints. To satisfy the constraints is to meet a condition

g(x) = 0c

where g : A Rc is a C 1 -mapping, with A Rn an open set. That is, the

constrained set forming the domain in the problem is a level set L, the inter-
section of the level sets of the component functions gi of g. (See figures 5.16
and 5.17. The first figure shows two individual level sets for scalar-valued
functions on R3 , and the second figure shows them together and then shows
their intersection, the level set for a vector-valued mapping.)

Figure 5.16. Level sets for two scalar-valued functions on R3

At any point p L, the set L must be locally orthogonal to each gradient

gi (p). (See figures 5.18 and 5.19. The first figure shows the level sets for
the component functions of the constraint mapping, and the gradients of the
component functions at p, while the second figure shows the tangent line and
the normal plane to the level set at p. In the first figure, neither gradient is
tangent to the other surface, and so in the second figure the two gradients are
not normal to one another.) Therefore:
L is orthogonal at p to every linear combination of the gradients,
c
X
i gi (p) where 1 , , c are scalars.
i=1

Equivalently:
 5.4 Lagrange Multipliers: Geometric Motivation and Specific Examples 235

Figure 5.17. The intersection is a level set for a vector-valued mapping on R3

Every such linear combination of gradients is orthogonal to L at p.

But we want to turn this idea around and assert the converse, that:
Every vector that is orthogonal to L at p is such a linear combination.
However, the converse does not always follow. Intuitively, the argument is that
if the gradients g1 (p), , gc (p) are linearly independent (i.e., they point
in c nonredundant directions) then the Implicit Function Theorem should say
that the level set L therefore looks (n c)-dimensional near p, so the space of
vectors orthogonal to L at p is c-dimensional, and so any such vector is indeed
a linear combination of the gradients. This intuitive argument is not a proof,
but for now it is a good heuristic.

Figure 5.18. Gradients to the level sets at a point of intersection

Proceeding to the second stage of the discussion, now suppose that p is

a critical point of the restriction to L of some C 1 -function f : A R.
236 5 Inverse and Implicit Functions

Figure 5.19. Tangent line and normal plane to the intersection

(Thus f has the same domain A Rn as g.) Then for any unit vector d
describing a direction in L at p, the directional derivative Dd f (p) must be 0.
But Dd f (p) = hf (p), di, so this means that:
f (p) must be orthogonal to L at p.
This observation combines with our description of the most general vector
orthogonal to L at p, in the third bullet above, to give Lagranges condition:
Suppose that p is a critical point of the function f restricted to the
level set L = {x : g(x) = 0c } of g. If the gradients gi (p) are linearly
independent, then
c
X
f (p) = i gi (p) for some scalars 1 , , c ,
i=1

and since p is in the level set, also

g(p) = 0c .

Approaching a constrained problem by setting up these conditions and then

working with the new variables 1 , , c is sometimes easier than the other
methods. The i are useful but irrelevant constants.
This discussion has derived the Lagrange multiplier criterion for the lin-
earized version of the constrained problem. The next section will use the
Implicit Function Theorem to derive the criterion for the actual constrained
problem, and then it will give some general examples. The remainder of this
section is dedicated to specific examples,
Returning to the unresolved second example at the beginning of the sec-
tion, the functions in question are
 5.4 Lagrange Multipliers: Geometric Motivation and Specific Examples 237

f (v, w, x, y, z) = v 2 + w2 + x2 + y 2 + z 2
g1 (v, w, x, y, z) = v + w + x + y + z 1
g2 (v, w, x, y, z) = v w + 2x y + z + 1

and the corresponding Lagrange condition and constraints are (after absorbing
a 2 into the s, whose particular values are irrelevant anyway)

(v, w, x, y, z) = 1 (1, 1, 1, 1, 1) + 2 (1, 1, 2, 1, 1)

= (1 + 2 , 1 2 , 1 + 22 , 1 2 , 1 + 2 )
v+w+x+y+z = 1
v w + 2x y + z = 1.

Substitute the expressions from the Lagrange condition into the constraints
to get 51 + 22 = 1 and 21 + 82 = 1. That is,
    
5 2 1 1
= .
2 8 2 1

and so, inverting the matrix to solve the system,

      
1 1 8 2 1 10/36
= = .
2 36 2 5 1 7/36

Note how much more convenient the two s are to work with than the five
original variables. Their values are auxiliary to the original problem, but sub-
stituting back now gives the nearest point to the origin,
1
(v, w, x, y, z) = (3, 17, 4, 17, 3),
36

and its distance from the origin is 612/36. This example is just one instance
of a general problem of finding the nearest point to the origin in Rn subject
to c affine constraints. We will solve the general problem in the next section.
An example from geometry is Euclids Least Area Problem. Given an angle
ABC and a point P interior to the angle as shown in figure 5.20, what line
through P cuts off from the angle the triangle of least area?
Draw the line L through P parallel to AB and let D be its intersection
with AC. Let a denote the distance AD and let h denote the altitude from
AC to P . Both a and h are constants. Given any other line L through P ,
let x denote its intersection with AC and H denote the altitude from AC to
the intersection of L with AB. (See figure 5.21.) The shaded triangle and its
subtriangle in the figure are similar, giving the relation x/H = (x a)/h.
The problem is now to minimize the function f (x, H) = 12 xH subject to
the constraint g(x, H) = 0 where g(x, H) = (x a)H xh = 0. Lagranges
condition f (x, H) = g(x, H) and the constraint g(x, H) = 0 become,
after absorbing a 2 into ,
238 5 Inverse and Implicit Functions

A C

Figure 5.20. Setup for Euclids Least Area Problem

P
D h
x
a xa

Figure 5.21. Construction for Euclids Least Area Problem

(H, x) = (H h, x a),
(x a)H = xh.
The first relation quickly yields (x a)H = x(H h). Combining this with
the second shows that H h = h, that is, H = 2h. The solution of Euclids
problem is, therefore, to take the segment that is bisected by P between the
two sides of the angle. (See figure 5.22.)
Euclids problem has the interpretation of finding the point of tangency
between the level set g(x, H) = 0, a hyperbola having asymptotes x = a and
H = h, and the level sets of f (x, H) = (1/2)xH, a family of hyperbolas having
asymptotes x = 0 and H = 0. (See figure 5.23, where the dashed asymptotes
meet at (a, h) and the point of tangency is visibly (x, H) = (2a, 2h).)
An example from optics is Snells Law. A particle travels through medium 1
at speed v, and through medium 2 at speed w. If the particle travels from
 5.4 Lagrange Multipliers: Geometric Motivation and Specific Examples 239

Figure 5.22. Solution of Euclids Least Area Problem

Figure 5.23. Level sets for Euclids Least Area Problem

point A to point B as shown in the least possible amount of time, what is the
relation between angles and ? (See figure 5.24.)
Since time is distance over speed, a little trigonometry shows that this
problem is equivalent to minimizing f (, ) = a sec /v + b sec /w subject
to the constraint g(, ) = a tan + b tan = d. (g measures lateral distance
traveled.) The Lagrange condition f (, ) = g(, ) is
 
a b
sin sec , sin sec = (a sec2 , b sec2 ).
2 2
v w

Therefore = sin /v = sin /w, giving Snells famous relation,

sin v
= .
sin w
240 5 Inverse and Implicit Functions

A a tan()
medium 1
a sec() a

d

medium 2 b sec()
b

b tan() B

Figure 5.24. Geometry of Snells Law

Figure 5.25 depicts the situation using the variables x = tan and y = tan .
The level set of possible configurations becomes the portion of the line
ax
+ by = d in p the first quadrant, and the function to be optimized becomes
a 1 + x2 /v + b 1 + y 2 /w. A level set for a large value of the function passes
through the point (0, d/b), the configuration with = 0 where the particle
travels vertically in medium 1 and then travels a long path in medium 2,
and a level set for a smaller value of the function passes through the point
(d/a, 0), the configuration with = 0 where the particle travels a long path
in medium 1 and then travels vertically in medium 2, while a level set for an
even smaller value of the function is tangent to the line segment at its point
that describes the optimal configuration specified by Snells Law.

Figure 5.25. Level sets for the optics problem

For an example from analytic geometry, let the function f measure the
square of the distance between the points x = (x1 , x2 ) and y = (y1 , y2 ) in the
 5.4 Lagrange Multipliers: Geometric Motivation and Specific Examples 241

plane,
f (x1 , x2 , y1 , y2 ) = (x1 y1 )2 + (x2 y2 )2 .
Fix points a = (a1 , a2 ) and b = (b1 , b2 ) in the plane, and fix positive numbers
r and s. Define

g1 (x1 , x2 ) = (x1 a1 )2 + (x2 a2 )2 r2 ,

g2 (y1 , y2 ) = (y1 b1 )2 + (y2 b2 )2 s2
g(x1 , x2 , y1 , y2 ) = (g1 (x1 , x2 ), g2 (y1 , y2 )).

Then the set of four-tuples (x1 , x2 , y1 , y2 ) such that

g(x1 , x2 , y1 , y2 ) = (0, 0)

can be viewed as the set of pairs of points x and y that lie respectively on the
circles centered at a and b with radii r and s. Thus, to optimize the function f
subject to the constraint g = 0 is to optimize the distance between pairs of
points on the circles. The rows of the 2-by-4 matrix
 
x a1 x 2 a2 0 0
g (x, y) = 2 1
0 0 y 1 b1 y 2 b2

are linearly independent because x 6= a and y 6= b. The Lagrange condition

works out to

(x1 y1 , x2 y2 , y1 x1 , y2 x2 ) = 1 (x1 a1 , x2 a2 , 0, 0)
2 (0, 0, y1 b1 , y2 b2 ),

or
(x y, y x) = 1 (x a, 02 ) 2 (02 , y b).
The second half of the vector on the left is the additive inverse of the first, so
the condition rewrites as

x y = 1 (x a) = 2 (y b).

If 1 = 0 or 2 = 0 then x = y and both i are 0. Otherwise 1 and 2 are

nonzero, forcing x and y to be distinct points such that

x y k x a k y b,

and so the points x, y, a, and b are collinear. Granted, these results are obvious
geometrically, but it is pleasing to see them follow so easily from the Lagrange
multiplier condition. On the other hand, not all points x and y such that x,
y, a, and b are collinear are solutions to the problem. For example, if both
circles are bisected by the x-axis and neither circle sits inside the other, then
x and y could be the leftmost points of the circles, neither the closest nor the
farthest pair.
242 5 Inverse and Implicit Functions

The last example of this section begins by maximizing the geometric mean
of n nonnegative numbers,

f (x1 , , xn ) = (x1 xn )1/n , each xi 0,

subject to the constraint that their arithmetic mean is 1,

x1 + + xn
= 1, each xi 0.
n
The set of such (x1 , , xn )-vectors is compact, being a closed subset of [0, n]n .
Since f is continuous on its domain [0, )n , it is continuous on the constrained
set, and so it takes minimum and maximum values on the constrained set. At
any constrained point set having some xi = 0, the function-value f = 0 is the
minimum. All other constrained points, having each xi > 0, lie in the interior
of the domain of f . The upshot is that we may assume that all xi are positive
and expect the Lagrange multiplier method to produce the maximum value
of f among the values that it produces. Especially, if the Lagrange multi-
plier method produces only one value (as it will) then that value must be the
maximum.
The constraining function is g(x1 , , xn ) = (x1 + + xn )/n, and the
gradients of f and g are
 
f (x1 , , xn ) 1 1
f (x1 , , xn ) = , ,
n x1 xn
1
g(x1 , , xn ) = (1, , 1).
n
The Lagrange condition f = g shows that all xi are equal, and the
constraint g = 1 forces their value to be 1. Therefore, the maximum value of
the geometric mean when the arithmetic mean is 1 is the value

f (1, , 1) = (1 1)1/n = 1.

This Lagrange multiplier argument provides most of the proof of

Theorem 5.4.1 (ArithmeticGeometric Mean Inequality). The geo-

metric mean of n positive numbers is at most their arithmetic mean:
a1 + + an
(a1 an )1/n for all nonnegative a1 , , an .
n
Proof. If any ai = 0 then the inequality is clear. Given positive numbers
a1 , , an , let a = (a1 + + an )/n and let xi = ai /a for i = 1, , n. Then
(x1 + + xn )/n = 1 and therefore
a1 + + an
(a1 an )1/n = a(x1 xn )1/n a = .
n


 5.4 Lagrange Multipliers: Geometric Motivation and Specific Examples 243

Despite these pleasing examples, Lagrange multipliers are in general no

computational panacea. Some problems of optimization with constraint are
solved at least as easily by geometry or substitution. Nonetheless, Lagranges
method provides a unifying idea that addresses many different types of op-
timization problem without reference to geometry or physical considerations.
In the following exercises, use whatever methods you find convenient.

Exercises

5.4.1. Find the nearest point to the origin on the intersection of the hyper-
planes x + y + z 2w = 1 and x y + z + w = 2 in R4 .

5.4.2. Find the nearest point on the ellipse x2 + 2y 2 = 1 to the line x + y = 4.

5.4.3. Minimize f (x, y, z) = z subject to the constraints 2x+4y = 5, x2 +z 2 =

2y.

5.4.4. Maximize f (x, y, z) = xy + yz subject to the constraints x2 + y 2 = 2,

yz = 2.

5.4.5. Find the extrema of f (x, y, z) = xy+z subject to the constraints x 0,

y 0, xz + y = 4, yz + x = 5.

5.4.6. Find the rectangular box of greatest volume, having sides parallel to the

2
2
2
coordinate axes, that can be inscribed in the ellipsoid xa + yb + zc = 1.

5.4.7. The lengths of the twelve edges of a rectangular block sum to 4, and
the areas of the six faces sum to 4. Find the lengths of the edges when the
excess of the blocks volume over that of a cube with edge equal to the least
edge of the block is greatest.

5.4.8. A cylindrical can (with top and bottom) has volume V . Subject to this
constraint, what dimensions give it the least surface area?

5.4.9. Find the distance in the plane from the point (0, 1) to the parabola
y = ax2 where a > 0. Note: The answer depends on whether a > 1/2 or 0 <
a 1/2.

5.4.10. This exercise extends the ArithmeticGeometric

Pn Mean Inequality. Let
e1 , , en be positive numbers with i=1 e i = 1. Maximize the function
fP(x1 , , xn ) = xe11 xenn (where each xi 0) subject to the constraint
n
i=1 ei xi = 1. Use your result to derive the weighted ArithmeticGeometric
Mean Inequality,

ae11 aenn e1 a1 + + en an for all nonnegative a1 , , an .

What values of the weights, e1 , , en reduce this to the basic Arithmetic

Geometric Mean Inequality?
244 5 Inverse and Implicit Functions

5.4.11. Let p and q be positive numbers satisfying the equation p1 + 1q = 1.

Pn
Maximize the function of 2n
Pnvariables f (x1 , P , xn , y1 , , yn ) = i=1 xi yi
p n q
subject to the constraints i=1 xi = 1 and i=1 yi = 1. Derive H olders
Inequality: For all nonnegative a1 , , an , b1 , , bn ,

n n
!1/p n
!1/q
X X X
a i bi api bqi .
i=1 i=1 i=1

5.5 Lagrange Multipliers: Analytic Proof and General

Examples
Recall that the environment for optimization with constraints consists of
an open set A Rn ,
a constraining C 1 -mapping g : A Rc ,
the corresponding level set L = {v A : g(v) = 0c },
and a C 1 -function f : A R to optimize on L.
We have argued geometrically, and not fully rigorously, that if f on L is
optimized at a point p L then the gradient f (p) is orthogonal to L at p.
Also, every linear combination of the gradients of the component functions
of g is orthogonal to L at p. We want to assert the converse, that every
vector that is orthogonal to L at p is such a linear combination. The desired
converse assertion does not always hold, but when it does it gives the Lagrange
condition,
X c
f (p) = i gi (p).
i=1

Here is the rigorous analytic justification that the Lagrange multiplier method
usually works. The Implicit Function Theorem will do the heavy lifting, and
it will reaffirm that the method is guaranteed only where the gradients of the
component functions of g are linearly independent. The theorem makes the
rigorous proof of the Lagrange criterion easier and more persuasiveat least
in the authors opinionthan the heuristic argument given earlier.

Theorem 5.5.1 (Lagrange Multiplier Condition). Let n and c be posi-

tive integers with n > c. Let g : A Rc (where A Rn ) be a mapping that
is continuously differentiable at each interior point of A. Consider the level
set
L = {x A : g(x) = 0c }.
Let f : A R be a function. Suppose that the restriction of f to L has an
extreme value at a point p L that is an interior point of A. Suppose that
f is differentiable at p, and suppose that the c-by-n derivative matrix g (p)
contains a c-by-c block that is invertible. Then the following conditions hold:
 5.5 Lagrange Multipliers: Analytic Proof and General Examples 245

f (p) = g (p) for some row vector Rc ,

g(p) = 0c .

The proof will culminate the ideas in this chapter as follows. The Inverse
Function Theorem says:
If the linearized inversion problem is solvable then the actual inversion
problem is locally solvable.
The Inverse Function Theorem is equivalent to the Implicit Function Theorem:
If the linearized level set is a graph then the actual level set is locally
a graph.
And finally, the idea for proving the Lagrange Condition is:
Although the graph is a curved space, where the techniques of chapter 4
do not apply, its domain is a straight space, where they do.
That is, the Implicit Function Theorem lets us reduce optimization on the
graph to optimization on the domain, which we know how to do.

Proof. The second condition holds since p is a point in L. The first condition
needs to be proved. Let r = n c, the number of variables that should remain
free under the constraint g(x) = 0c , and notate the point p as p = (a, b),
where a  Rr and c
 b R . Using this notation, we have g(a, b) = 0c and

g (a, b) = M N where M is c-by-r and N is c-by-c and invertible. (We may
assume that N is the invertible block in the hypotheses to the theorem because
we may freely permute the variables.) The Implicit Function Theorem gives
a mapping : A0 Rc (where A0 Rr and a is an interior point of A0 )
with (a) = b, (a) = N 1 M , and for all points (x, y) A near (a, b),
g(x, y) = 0c if and only if y = (x).
Make f depend only on the free variables by defining

f0 = f (idr , ) : A0 R, f0 (x) = f (x, (x)).

(See figure 5.26.) Since the domain of f0 doesnt curve around in some larger
space, f0 is optimized by the techniques from chapter 4. That is, the Implicit
Function Theorem has reduced optimization on the curved set to optimization
in Euclidean space. Specifically, the multivariable Critical Point Theorem says
that f0 has a critical point at a,

f0 (a) = 0r .

Our task is to express the previous display in terms of the given data f and g.
Doing so will produce the Lagrange condition.
Since f0 = f (idr , ) is a composition, the Chain Rule says that the
condition f0 (a) = 0r is f (a, (a)) (idr , ) (a) = 0r , or
246 5 Inverse and Implicit Functions
 
Ir
f (a, b) = 0r .
(a)

Let f (a, b) = (u, v) where u Rr and v Rc are row vectors, and recall
that (a) = N 1 M . The previous display becomes
 
Ir
[u v] = 0r ,
N 1 M

giving u = vN 1 M . This expression for u and the trivial identity v = vN 1 N

combine to give in turn
   
u v = vN 1 M N .
   
But u v = f (a, b) and M N = g (a, b) and (a, b) = p. So set = vN 1 (a
row vector in Rc ), and the previous display is precisely Lagranges condition,

f (p) = g (p).

y
Rc

p
f

(idr , ) f0 R

x
Rn A0 a Rr
Figure 5.26. The Lagrange Multiplier Criterion from the Implicit Function Theo-
rem

We have seen that the Lagrange Multiplier Condition is necessary but not
sufficient for an extreme value. That is, it can report a false positive, as in the
two-circle problem in the previous section. False positives are not a serious
problem since inspecting all the points that meet the Lagrange condition will
determine which of them give the true extrema of f . A false negative would
be a worse situation, giving us no indication that an extreme value might
exist, much less how to find it. The following example shows that the false
negative scenario can arise without the invertible c-by-c block required in
Theorem 5.5.1.
 5.5 Lagrange Multipliers: Analytic Proof and General Examples 247

Let the temperature in the plane be given by

f (x, y) = x,

and consider a plane set defined by one constraint on two variables,

L = {(x, y) R2 : y 2 = x3 }.

(See figure 5.27.) Since temperature increases as we move to the right, the
coldest point of L is its leftmost point, the cusp at (0, 0). However, the La-
grange condition does not find this point. Indeed, the constraining function
is g(x, y) = x3 y 2 (which does have continuous derivatives, notwithstanding
that its level set has a cusp: the graph of a smooth function is smooth, but
the level set of a smooth function need not be smooththis is exactly the
issue addressed by the Implicit Function Theorem). Therefore the Lagrange
condition and the constraint are

(1, 0) = (3x2 , 2y),

x3 = y 2 .

These equations have no solution. The problem is that the gradient at the cusp
is g(0, 0) = (0, 0), and neither of its 1-by-1 subblocks is invertible. In general,
the Lagrange Multiplier Condition will not report a false negative so long as
we remember that it only claims to check for extrema at the nonsingular
points of L, the points p such that g (p) has an invertible c-by-c subblock.

Figure 5.27. Curve with cusp

The previous section gave specific examples of the Lagrange multiplier

method. This section now gives some general families of examples.
Recall that the previous section discussed the problem of optimizing the
distance between two points in the plane, each point lying on an associated
circle. Now, as the first general example of the Lagrange multiplier method,
let (x, y) Rn Rn denote a pair of points each from Rn , and let the function
f measure the square of distance between such a pair,

f : Rn Rn R, f (x, y) = |x y|2 .
248 5 Inverse and Implicit Functions

Note that f (x, y) = [x y y x], viewing x and y as row vectors. Given two
mappings g1 : Rn Rc1 and g2 : Rn Rc2 , define

g : Rn Rn Rc1 +c2 , g(x, y) = (g1 (x), g2 (y)).

To optimize the function f subject to the constraint g(x, y) = (0c1 , 0c2 ) is to

optimize the distance between pairs of points x and y on the respective level
sets cut out of Rn by the c1 conditions g1 (x) = 0c1 and the c2 conditions
g2 (y) = 0c2 . Assuming that the Lagrange condition holds for the optimizing
pair, it is
 
 
  g1 (x) 0c2 n
x y y x = g (x, y) = 1 2
0c1 n g2 (y)
= 1 (g1 (x), 0c2 n ) 2 (0c1 n , g2 (y)),

where 1 Rc1 and 2 Rc2 are row vectors. The symmetry of f reduces
this equality of 2n-vectors to an equality of n-vectors,

x y = 1 g1 (x) = 2 g2 (y).

That is, either x = y or the line through x and y is normal to the first level
set at x and normal to the second level set at y, generalizing the result from
the two-circle problem. With this result in mind, you may want to revisit
exercise 0.0.1 from the preface to these notes.
The remaining general Lagrange multiplier methods optimize a linear func-
tion or a quadratic function subject to affine constraints or a quadratic con-
straint. We gather the results in one theorem.

Theorem 5.5.2 (Low Degree Optimization With Constraints).

(1) Let f (x) = aT x (where a Rn ) subject to the constraint M x = b (where
M Mc,n (R) has linearly independent rows, with c < n, and b Rc ).
Check whether aT M T (M M T )1 M = aT . If so then f subject to the con-
traint is identically aT M T (M M T )1 b; otherwise f subject to the constraint
has no optima.
(2) Let f (x) = xT Ax (where A Mn (R) is symmetric and invertible) subject
to the constraint M x = b (where M Mc,n (R) has linearly independent
rows, with c < n, and b Rc ). The x that optimizes f subject to the
constraint and the optimal value are

x = A1 M T (M A1 M T )1 b and f (x) = bT (M A1 M T )1 b.

Especially when A = I, the point x such that M x = b closest to the origin

and its square distance from the origin are

x = M T (M M T )1 b and |x|2 = bT (M M T )1 b.
 5.5 Lagrange Multipliers: Analytic Proof and General Examples 249

(3) Let f (x) = aT x (where a Rn ) subject to the constraint xT M x = b (where

M Mn (R) is symmetric and invertible, and b R is nonzero). Check
whether aT M 1 ab > 0. If so then the optimizing inputs and the optimal
values are

x = M 1 ab/ aT M 1 ab and f (x) = aT M 1 ab .

Otherwise f subject to the constraint has no optima.

(4) Let f (x) = xT Ax (where A Mn (R) is symmetric) subject to the con-
straint xT M x = b (where M Mn (R) is symmetric and invertible,
and b R is nonzero). The possible optimal values of f subject to the
constraint are

f (x) = b where is an eigenvalue of M 1 A.

(The term eigenvalue will be explained in the proof.) Especially when

A = I, the nearest square-distances from the origin on the quadratic sur-
face xT M x = b take the form b where is an eigenvalue of M 1 .
Proof. (1) The data are (viewing vectors as columns)

f : Rn R, f (x) = aT x where a Rn ,
g : Rn Rc , g(x) = M x b where M Mc,n (R) and b Rc .

Here we assume that c < n, i.e., there are fewer constraints than variables.
Also, we assume that the c rows of M are linearly independent in Rn , or equiv-
alently (invoking a result from linear algebra), that some c columns of M are
a basis of Rc , or equivalently, that some c-by-c subblock of M (not necessarily
contiguous columns) has nonzero determinant. The Lagrange condition and
the constraints are

aT = T M where Rc ,
M x = b.

Before solving the problem, we need to consider the two relations in the pre-
vious display.
The Lagrange condition aT = T M is solvable for exactly when aT is a
linear combination of the rows of M . Since M has c rows, each of which is
a vector in Rn , and since c < n, generally aT is not a linear combination
of the rows of M , so the Lagrange conditions can not be satisfied. That is:
Generally the constrained function has no optimum.
However, we will study the exceptional case, when aT is a linear combi-
nation of the rows of M . In this case, the linear combination of the rows
that gives aT is unique because the rows are linearly independent. That
is, when exists it is uniquely determined.
To find the only candidate , note that the Lagrange condition aT =
T M gives aT M T = T M M T , and thus T = aT M T (M M T )1 . This
250 5 Inverse and Implicit Functions

calculations first step is not reversible, and so the calculation does not
show that exists to be found in all cases. But it does show that to check
whether aT is a linear combination of the rows of M , one checks whether
aT M T (M M T )1 M = aT , in which case T = aT M T (M M T )1 .
Note that furthermore, the Lagrange condition aT = T M makes no ref-
erence to x
The constraining condition M x = b has solutions x only if b is a linear
combination of the columns of M . Our assumptions about M guarantee
that this is the case.
With aT being a linear combination of the rows of M and with b being a linear
combination of the columns of M , the Lagrange condition and the constraints
immediately show that for any x in the constrained set,

f (x) = aT x = T M x = T b = aT M T (M M T )1 b.

That is, f subject to the constraint g = b is the constant = aT M T (M M T )1 b.

For geometric insight into the calculation, envision the space of linear
combinations of the rows of M (a c-dimensional subspace of Rn ) as a plane,
and envision the space of vectors x such that M x = 0c (an (nc)-dimensional
subspace of Rn ) as an axis orthogonal to the plane. The condition aT = T M
says that a lies in the plane, and the condition M x = b says that x lies on an
axis parallel to the x-axis. (From linear algebra, the solutions of M x = b are
the vectors
x = x0 + x,
where x0 is the unique linear combination of the rows of M such that M x0 = b,
and x is any vector such that M x = 0c .) The constant value of f is aT x for
any x on the axis. In particular, the value is aT x0 where x0 is the point where
the axis meets the plane.
(2) Now we optimize a quadratic function subject to affine constraints.
Here the data are

f : Rn R, f (x) = xT Ax where A Mn (R) is symmetric,

n c
g : R R , g(x) = M x b where M Mc,n (R) and b Rc .

As in (1), we assume that c < n, and we assume that the c rows of M are
linearly independent in Rn , i.e., some c columns of M are a basis of Rc , i.e.,
some c-by-c subblock of M has nonzero determinant. Thus the constraints
M x = b have solutions x for any b Rc .
To set up the Lagrange condition, we need to differentiate the quadratic
function f . Compute that

f (x + h) f (x) = (x + h)T A(x + h) xT Ax = 2xT Ah + hT Ah,

and so the best linear approximation of this difference is T (h) = 2xT Ah. It
follows that
 5.5 Lagrange Multipliers: Analytic Proof and General Examples 251

f (x) = 2xT A.
Returning to the optimization problem, the Lagrange condition and the
constraints are

x T A = T M where Rc ,
M x = b.

Having solved a particular problem of this sort in section 5.4, we use its
particular solution to guide our solution of the general problem. The first step
was to express x in terms of , so here we transpose the Lagrange condition to
get Ax = M T , then assume that A is invertible and thus get x = A1 M T .
The second step was to write the constraint in terms of and then solve
for , so here we have b = M x = M A1 M T , so that = (M A1 M T )1 b,
assuming that the c-by-c matrix M A1 M T is invertible. Now the optimizing
input x = A1 M T is

x = A1 M T (M A1 M T )1 b,

and the optimal function value f (x) = xT Ax = T M x = T b is

f (x) = bT (M A1 M T )1 b.

In particular, letting A = I, the closest point x to the origin such that M x = b

and the square of its distance from the origin are

x = M T (M M T )1 b, |x|2 = bT (M M T )1 b.

(3) Next we optimize a linear function subject to a quadratic constraint.

The data are

f : Rn R, f (x) = aT x where a Rn ,
(
M Mn (R) is symmetric
g : Rn R, T
g(x) = x M x b where
b R is nonzero.

The Lagrange condition and the constraint are

aT = xT M where R,
T
x M x = b.

Therefore the possible optimized values of f are

f (x) = aT x = xT M x = b,

and so to find these values it suffices to find the possible values of . Assuming
that M is invertible, the Lagrange condition is aT M 1 = xT , and hence
252 5 Inverse and Implicit Functions

aT M 1 ab = xT ab = 2 b2 = f (x)2 .

Thus (assuming that aT M 1 ab > 0) the optimal values are

f (x) = aT M 1 ab .

The penultimate display also shows that = aT M 1 ab/b, so that the
Lagrange condition gives the optimizing x-values,

x = M 1 ab/ aT M 1 ab .

One readily confirms that indeed xT M x = b for these x.

As a small geometric illustration of the sign-issues in this context, suppose
that n = 2 and M = [ 01 10 ], so that the quadratic constraint is 2x1 x2 = b. For
b > 0 the optimizing problem is thus set on a hyperbola in the first and third
quadrants of the plane. The function to be optimized is f (x, y) = a1 x1 + a2 x2
for some a1 , a2 R. Since M is its own inverse, the quantity aT M 1 ab under
the square root is 2a1 a2 b, and thus the constrained optimization problem has
solutions only when a1 a2 > 0. Meanwhile, the level sets of f are lines of
slope a1 /a2 , meaning that the problem has solutions only when the level
sets have negative slope. In that case, the solutions will be at the two points
where the hyperbola is tangent to a level set: a pair of opposite points, one in
the first quadrant and one in the third. For b < 0 the constraining hyperbola
moves to the second and fourth quadrants, and the problem has solutions
when the level sets of f have a positive slope.
(4) Finally, we optimize a quadratic function subject to a quadratic con-
straint. The data are

f : Rn R, f (x) = xT Ax where A Mn (R) is symmetric,

(
M Mn (R) is symmetric
g : Rn R, T
g(x) = x M x b where
b R is nonzero.

The Lagrange condition and the constraint are

xT A = xT M where R,
T
x M x = b.

By the Lagrange condition and the constraint, the possible optimal values
of f take the form
f (x) = xT Ax = xT M x = b,
which we will know as soon as we find the possible values of , without needing
to find x. Assuming that M is invertible, the Lagrange condition gives

M 1 Ax = x.
 5.5 Lagrange Multipliers: Analytic Proof and General Examples 253

In other words, x must satisfy the condition that multiplying x by M 1 A

gives a scalar multiple of x. Any nonzero vector x that satisfies this condi-
tion is called an eigenvector of M 1 A. The scalar multiple factor is the
corresponding eigenvalue. We will end the section with a brief discussion of
eigenvalues.

The eigenvalues of any square matrix B are found by a systematic proce-
dure. The first step is to observe that the condition Bx = x is
(B I)x = 0.
Since any eigenvector x is nonzero by definition, B I is not invertible, i.e.,
det(B I) = 0.
Conversely, for every R satisfying this equation there is at least one
eigenvector x of B because the equation (B I)x = 0 has nonzero solutions.
And so the eigenvalues are the real roots of the polynomial
pB () = det(B I).
This polynomial is the characteristic polynomial of B, already discussed in
exercise 4.7.10. For example,
  part (a) of that exercise covered the case n = 2,
showing that if B = ab db then
pB () = 2 (a + d) + (ad b2 ).
The discriminant of this quadratic polynomial is
= (a + d)2 4(ad b2 ) = (a d)2 + 4b2 .
Since is nonnegative, all roots of characteristic polynomial are real. And a
result of linear algebra says that for any positive n, all roots of the character-
istic polynomial of a symmetric n-by-n matrix B are real as well. However,
returning to our example, even though the square matrices A and M are
assumed to be symmetric, the product M 1 A need not be.
As a particular case of Theorem 5.5.2, part (4), if A = I then finding the
eigenvectors of M encompasses finding the points of a quadric surface that
are closest
 to the origin or farthest from the origin. For instance, if n = 2 and
M = ab db then we are optimizing on the set of points (x1 , x2 ) R2 such
that, say,
ax21 + 2bx1 x2 + dx22 = 1.
The displayed equation is the equation of a conic section. When b = 0 we
have an unrotated ellipse or hyperbola, and the only possible optimal points
will be the scalar multiples of e1 and e2 that lie on the curve. For an ellipse,
a pair of points on one axis is closest to the origin, and a pair on the other
axis is farthest; for a hyperbola, a pair on one axis is closest and there are
no points on the other axis. In the case of a circle, the matrix M is a scalar
multiple of the identity matrix, and so all vectors are eigenvectors compatibly
with the geometry that all points are equidistant from the origin. Similarly if
n = 3 then L is a surface such as an ellipsoid or a hyperboloid.
254 5 Inverse and Implicit Functions

Exercises

5.5.1. Let f (x, y) = y and let g(x, y) = y 3 x4 . Graph the level set L =
{(x, y) : g(x, y) = 0}. Show that the Lagrange multiplier criterion does not find
any candidate points where f is optimized on L. Optimize f on L nonetheless.

5.5.2. Consider the linear mapping

g(x, y, z) = (x + 2y + 3z, 4x + 5y + 6z).

(a) Use Theorem 5.5.2, part (1) to optimize the linear function f (x, y, z) =
6x + 9y + 12z subject to the affine constraints g(x, y, z) = (7, 8).
(b) Verify without using the Lagrange multiplier method that the function
f subject to the constraints g = (7, 8) (with f and g from part (a)) is constant,
always taking the value that you found in part (a).
(c) Show that the function f (x, y, z) = 5x + 7y + z can not be optimized
subject to any constraint g(x, y, z) = b.

5.5.3. (a) Use Theorem 5.5.2, part (2) to minimize the quadratic function
f (x, y) = x2 + y 2 subject to the affine constraint 3x + 5y = 8.
(b) Use the same result to find the extrema of f (x, y, z) = 2xy + z 2 subject
to the constraints x + y + z = 1, x + y z = 0.
(c) Use the same result to find the nearest point to the origin on the
intersection of the hyperplanes x + y + z 2w = 1 and x y + z + w = 2
in R4 , reproducing your answer to exercise 5.4.1.

5.5.4. (a) Use Theorem 5.5.2, part (3) to optimize f (x, y, z) = x 2y + 2z on

the sphere of radius 3.
(b) Use the same result to optimize the function f (x, y, z, w) = x+yzw
subject to the constraint g(x, y, z, w) = 1, g(x, y, z, w) = x2 /2 y 2 + z 2 w2 .

5.5.5. (a) Use Theorem 5.5.2, part (4) to optimize the function f (x, y) = 2xy
subject to the constraint g(x, y) = 1 where g(x, y) = x2 + 2y 2 .
(b) Use the same result to optimize the function f (x, y, z) = 2(xy+yz+zx)
subject to the constraint g(x, y, z) = 1 where g(x, y, z) = x2 + y 2 z 2 .

5.6 Summary
The main arguments in this chapter manipulate larger idea-units than those
in the previous chapters, and they are standard mathematical set-pieces. The
proof of the Inverse Function Theorem uses the main results from chapters 2,
3, and 4. The Inverse Function Theorem further implies the Implicit Function
Theorem, and conversely. The Implicit Function Theorem gives the Lagrange
multiplier condition, a systematic approach to problems of optimization with
constraints. The Lagrange multiplier method can also be viewed as optimiza-
tion on a curved set inside a larger-dimensional space.
 Part II

Multivariable Integral Calculus

6
Integration

The integral of a scalar-valued function of many variables, taken over a box

of its inputs, is defined in sections 6.1 and 6.2. Intuitively, the integral can be
understood as representing mass or volume, but the definition is purely math-
ematical: the integral is a limit of sums, as in one-variable calculus. Multivari-
able integration has many familiar propertiesfor example, the integral of a
sum is the sum of the integrals. Section 6.3 shows that continuous functions
can be integrated over boxes. However, we want to carry out multivariable
integration over more generally-shaped regions. That is, the theory has geo-
metric aspects not present in the one-dimensional case, where integration is
carried out over intervals. After a quick review of the one-variable theory in
section 6.4, section 6.5 shows that continuous functions can also be integrated
over nonboxes that have manageable shapes. The main tools for evaluating
multivariable integrals are Fubinis Theorem (section 6.6), which reduces an
n-dimensional integral to an n-fold nesting of one-dimensional integrals,and
the Change of Variable Theorem (section 6.7), which replaces one multivari-
able integral by another that may be easier to evaluate. Section 6.8 provides
some preliminaries for the proof of the Change of Variable Theorem, and then
section 6.9 gives the proof.

6.1 Machinery: Boxes, Partitions, and Sums

The integral represents physical ideas such as volume or mass or work, but
defining it properly in purely mathematical terms requires some care. Here is
some terminology that is standard from the calculus of one variable, perhaps
other than compact (meaning closed and bounded ) from section 2.4 of these
notes. The language describes a domain of integration and the machinery to
subdivide it.

Definition 6.1.1 (Compact Interval, Length, Partition, Subinterval).

A nonempty compact interval in R is a set
258 6 Integration

I = [a, b] = {x R : a x b} ,

where a and b are real numbers with a b. The length of the interval is

length(I) = b a.

A partition of I is a set of real numbers

P = {t0 , t1 , . . . , tk }

satisfying
a = t0 < t1 < < tk = b.
Such a partition divides I into k subintervals J1 , . . . , Jk where

Jj = [tj1 , tj ], j = 1, . . . , k.

A generic nonempty compact subinterval of I is denoted J. (See figure 6.1.)

Since the only intervals that we are interested in are nonempty and compact,
either or both of these properties will often be tacit from now on, rather
than stated again and again. As a special case, Definition 6.1.1 says that any
length-zero interval [a, a] has only one partition, P = {a}, which divides it
into no subintervals.

J1 J Jk

a = t0 t1 t2 t3 tk1 tk = b

Figure 6.1. Interval and subintervals

The next definition puts an initial loose stipulation on functions to be

integrated.

Definition 6.1.2 (Bounded Function). Let A be a subset of R, and let

f : A R be a function. Then f is bounded if its range, {f (x) : x A},
is bounded as a set in R, as in Definition 2.4.6. That is, f is bounded if there
exists some R > 0 such that |f (x)| < R for all x A.

Visually, a function is bounded if its graph is contained inside a horizontal

strip. On the other hand, the graph of a bounded function neednt be contained
in a vertical strip, because the domain (and therefore the graph) need not be
bounded. For example, these functions are bounded:

f (x) = sin x, f (x) = 1/(1 + x2 ), f (x) = arctan x,

 6.1 Machinery: Boxes, Partitions, and Sums 259

and these functions are not:

f (x) = ex , f (x) = 1/x for x 6= 0.

However, since we want to integrate a bounded function over a compact in-

terval, the entire process is set inside a rectangle in the plane.
The next definition describes approximations of the integral by finite sums,
the integral to be a limit of such sums if it exists at all. The summands involve
limits, so already these sums are analytic constructs despite being finite.

Definition 6.1.3 (One-dimensional Lower Sum and Upper Sum). Let

I be a nonempty compact interval in R, and let f : I R be a bounded
function. For any subinterval J of I, the greatest lower bound of the values
taken by f on J is denoted mJ (f ),

mJ (f ) = inf {f (x) : x J} ,

and similarly the least upper bound is denoted MJ (f ),

MJ (f ) = sup {f (x) : x J} .

The lower sum of f over P is

X
L(f, P ) = mJ (f ) length(J),
J

and the upper sum of f over P is

X
U (f, P ) = MJ (f ) length(J).
J

If the interval I in Definition 6.1.3 has length zero, then the lower and
upper sums are empty and so they are assigned the value 0 by convention.
The function f in Definition 6.1.3 is not required to be differentiable or
even continuous, only bounded. Even so, the values mJ (f ) and MJ (f ) in the
previous definition exist by the set-bound phrasing of the principle that the
real number system is complete. To review this idea, see Theorem 1.1.4. When
f is in fact continuous, the Extreme Value Theorem (Theorem 2.4.15) justi-
fies substituting min and max for inf and sup in the definitions of mJ (f )
and MJ (f ), since each subinterval J is nonempty and compact. It may be
easiest at first to understand mJ (f ) and MJ (f ) by imagining f to be contin-
uous and mentally substituting appropriately. But we will need to integrate
discontinuous functions f . Such functions may take no minimum or maximum
on J, and so we may run into a situation like the one pictured in figure 6.2,
in which the values mJ (f ) and MJ (f ) are not actual outputs of f . Thus the
definition must be as given to make sense.
The technical properties of inf and sup will figure in Lemmas 6.1.6, 6.1.8,
and 6.2.2. To see them in isolation first, we rehearse them now. So, let S
260 6 Integration

MJ (f )

mJ (f )

Figure 6.2. Sup and inf but no max or min

and T be nonempty sets of real numbers, both bounded. In the context of

integration, S and T will be sets of outputs of a bounded function f . This
particular of S and T is irrelevant for the moment, but it may help you to
see later how these ideas are used in context if you now imagine S and T on
a vertical axis, as in figure 6.2, rather than on a horizontal one. In any case,
the necessary results are as follows.
inf(S) sup(S). In fact any lower bound of S is at most as big as any
upper bound, because any element of S lies between them. In particular,
this argument applies to the greatest lower bound inf(S) and the least
upper bound sup(S), giving the stated inequality.
If S T then inf(T ) inf(S) sup(S) sup(T ). We already have
the middle inequality. To establish the others, the idea is that since S is
a subset, the bounds on S are innately at least as tight as those on T .
More specifically, since inf(T ) is a lower bound of T , it is a lower bound
of the subset S, and since inf(S) is the greatest lower bound of S the first
inequality follows. The third inequality is similar.
In particular, let I be a compact interval, let f : I R be a bounded
function, let J be a subinterval of I, let J be a subinterval of J in turn,
and then take S and T to be sets of output-values of f ,

S = {f (x) : x J }, T = {f (x) : x J}.

Then S T since S is a set of fewer outputs than T , and so this bullet

has shown that

mJ (f ) mJ (f ) MJ (f ) MJ (f ).

If s t for all s S and t T then sup(S) inf(T ). Imprecisely, the

idea is that S is entirely below T on the vertical axis, and so the smallest
number that traps S from above is still below the largest number that
traps T from below. A more careful proof is in the next section.
Graphing f over I in the usual fashion and interpreting the lower and upper
sum as sums of rectangle-areas shows that they are respectively too small and
too big to be the area under the graph. (See figure 6.3.) Alternatively, thinking
 6.1 Machinery: Boxes, Partitions, and Sums 261

of f as the density function of a wire stretched over the interval I shows that
the lower and upper sum are too small and too big to be the mass of the
wire. The hope is that the lower and upper sums are trapping a yet-unknown
quantity (possibly to be imagined as area or mass) from each side, and that as
the partition P becomes finer, the lower and upper sums will actually converge
to this value.

Figure 6.3. Too small and too big

All the terminology so far generalizes easily from one dimension to many,
i.e., from R to Rn . Recall that if S1 , S2 , . . . , Sn are subsets of R then their
cartesian product is a subset of Rn ,

S1 S2 Sn = {(s1 , s2 , . . . , sn ) : s1 S1 , s2 S2 , . . . , sn Sn } .

(See figure 6.4, in which n = 2, and S1 has two components, and S2 has one
component, so that the cartesian product S1 S2 has two components.)

Figure 6.4. Cartesian product

262 6 Integration

Definition 6.1.4 (Compact Box, Box Volume, Partition, Subbox). A

nonempty compact box in Rn is a cartesian product

B = I1 I2 In

of nonempty compact intervals Ij for j = 1, . . . , n. The volume of the box is

the product of the lengths of its sides,
n
Y
vol(B) = length(Ij ).
j=1

A partition of B is a cartesian product of partitions Pj of Ij for j = 1, . . . , n,

P = P1 P2 Pn .

Such a partition divides B into subboxes J, each such subbox being a carte-
sian product of subintervals. By a slight abuse of language, these are called
the subboxes of P .

(See figure 6.5, and imagine its three-dimensional Rubiks cube counterpart.)
Every nonempty compact box in Rn has partitions, even such boxes with
some length-zero sides. This point will arise at the very beginning of the next
section.

Figure 6.5. Box and subboxes

The definition of a bounded function f : A R, where now A is a subset

of Rn , is virtually the same as earlier in the section: again the criterion is that
its range must be bounded as a set. (In fact the definition extends just as
easily to mappings f : A Rm , but we need only scalar-valued functions
here.)
 6.1 Machinery: Boxes, Partitions, and Sums 263

Definition 6.1.5 (n-dimensional Lower Sum and Upper Sum). Let B

be a nonempty compact box in Rn , and let f : B R be a bounded function.
For any subbox J of B, define mJ (f ) and MJ (f ) analogously to before,

mJ (f ) = inf {f (x) : x J} and MJ (f ) = sup {f (x) : x J} .

The lower sum and upper sum of f over P are similarly

X X
L(f, P ) = mJ (f ) vol(J) and U (f, P ) = MJ (f ) vol(J).
J J

With minor grammatical modifications, this terminology includes the pre-

vious definition as a special case when n = 1 (e.g., volume reverts to length,
as it should revert to area when n = 2), so from now on we work in Rn .
However, keeping the cases n = 1 and n = 2 in mind should help to make the
pan-dimensional ideas of multivariable integration geometrically intuitive. If
the box B in Definition 6.1.5 has any sides of length zero then the upper and
lower sums are 0.
Graphing f over B in the usual fashion when n = 2 and interpreting the
lower and upper sum as sums of box-volumes shows that they are respectively
too small and too big to be the volume under the graph. (See figure 6.6.)
Alternatively, if n = 2 or n = 3, then thinking of f as the density of a plate
or a block occupying the box B shows that the lower and upper sum are too
small and too big to be the objects mass. Again, the hope is that as the
partitions become finer, the lower and upper sums will converge to a common
value that they are trapping from either side.

Figure 6.6. Too small and too big

The first result supports this intuition.

Lemma 6.1.6. For any box B, any partition P of B, and any bounded func-
tion f : B R,
L(f, P ) U (f, P ).
264 6 Integration

Proof. For any subbox J of P , mJ (f ) MJ (f ) by the first bullet from earlier

in this section with S = {f (x) : x J }, while also vol(J) 0, and therefore
mJ (f ) vol(J) MJ (f ) vol(J). Sum this relation over all subboxes J to get
the result.

The next thing to do is express the notion of taking a finer partition.

Definition 6.1.7 (Refinement). Let P and P be partitions of B. Then P

is a refinement of P if P P .

Figure 6.7 illustrates the fact that if P refines P then every subbox of P
is contained in a subbox of P . The literal manifestation in the figure of the
containment P P is that the set of points where a horizontal line segment
and a vertical line segment meet in the right side of the figure subsumes the
set of such points in the left side.
Refining a partition brings the lower and upper sums nearer each other:

Figure 6.7. Refinement

Lemma 6.1.8. Suppose that P refines P as a partition of the box B. Then

L(f, P ) L(f, P ) and U (f, P ) U (f, P ).

See figure 6.8 for a picture-proof for lower sums when n = 1, thinking of
the sums in terms of area. The formal proof is just a symbolic rendition of
the figures features.

Proof. Every subbox J of P divides further under the refinement P into

subboxes J . Since each J J, we have mJ (f ) mJ (f ) by the second
bullet from earlier in this section, but even without reference to the bullet the
idea is that
 6.1 Machinery: Boxes, Partitions, and Sums 265

Figure 6.8. Lower sum increasing under refinement

mJ (f ) mJ (f ) because f has less opportunity to be small on the

subbox J of J.
Thus
X X
mJ (f ) vol(J ) mJ (f ) vol(J )
J J J J
X
= mJ (f ) vol(J ) = mJ (f )vol(J).
J J
P
Sum the relation J J mJ (f ) vol(J ) mJ (f ) vol(J) over all subboxes J
of P to get L(f, P ) L(f, P ). The argument is similar for upper sums.

The proof uncritically assumes that the volumes of a boxs subboxes sum
to the volume of the box. This assumption is true, and left as an exercise.
The emphasis here isnt on boxes (which are straightforward), but on defining
the integral of a function f whose domain is a box. The next result helps
investigate whether the lower and upper sums indeed trap some value from
both sides. First we need a definition.

Definition 6.1.9 (Common Refinement). Given two partitions of B,

P = P1 P2 Pn and P = P1 P2 Pn ,

their common refinement is the partition

P = (P1 P1 ) (P2 P2 ) (Pn Pn ).

(See figure 6.9.) The common refinement of two partitions P and P is

certainly a partition that refines both P and P , and it is the smallest such
partition. The union P P is not taken as the definition of the common
refinement because it need not be a partition at all. The common refinement
does all the work for the next result.
266 6 Integration

Figure 6.9. Common refinement

Proposition 6.1.10 (Lower Sums Are At Most Upper Sums). Let P

and P be partitions of the box B, and let f : B R be any bounded function.
Then
L(f, P ) U (f, P ).

Proof. Let P be the common refinement of P and P . By the two lemmas,

L(f, P ) L(f, P ) U (f, P ) U (f, P ),

proving the result.

Exercises

6.1.1. (a) Let I = [0, 1], let P = {0, 1/2, 1}, let P = {0, 3/8, 5/8, 1}, and let
P be the common refinement of P and P . What are the subintervals of P ,
and what are their lengths? Same question for P . Same question for P .
(b) Let B = I I, let Q = P {0, 1/2, 1}, let Q = P {0, 1/2, 1}, and
let Q be the common refinement of Q and Q . What are the subboxes of Q
and what are their areas? Same question for Q . Same question for Q .

6.1.2. Show that the lengths of the subintervals of any partition of [a, b] sum
to the length of [a, b]. Same for the areas of the subboxes of [a, b] [c, d].
Generalize to Rn .

6.1.3. Let J = [0, 1]. Compute mJ (f ) and MJ (f ) for each of the following
functions f : J R.
( x),
(a) f (x) = x(1
1 if x is irrational
(b) f (x) =
1/m if x = n/m in lowest terms, n, m Z and m > 0,
(
(1 x) sin(1/x) if x 6= 0
(c) f (x) =
0 if x = 0.
 6.2 Definition of the Integral 267

6.1.4. (a) Let I, P , P and P be as in exercise 6.1.1(a), and let f (x) = x2

on I. Compute the lower sums L(f, P ), L(f, P ), L(f, P ) and the correspond-
ing upper sums, and check that they conform to Lemma 6.1.6, Lemma 6.1.8,
and Proposition 6.1.10.
(b) Let B, Q, Q and Q be as in exercise 6.1.1(b), and define f : B R
by (
0 if 0 x < 1/2
f (x, y) =
1 if 1/2 x 1.
Compute L(f, Q), L(f, Q ), L(f, Q ) and the corresponding upper sums,
and check that they conform to Lemma 6.1.6, Lemma 6.1.8, and Proposi-
tion 6.1.10.
6.1.5. Draw the cartesian product ([a1 , b1 ] [c1 , d1 ]) ([a2 , b2 ] [c2 , d2 ]) R2
where a1 < b1 < c1 < d1 and similarly for the other subscript.
6.1.6. When is a cartesian product empty?
6.1.7. Show that the union of partitions of a box B need not be a partition
of B.
6.1.8. Draw a picture illustrating the proof of Proposition 6.1.10 when n = 1.

6.2 Definition of the Integral

Fix a nonempty compact box B and a bounded function f : B R. The
set of lower sums of f over all partitions P of B,
{L(f, P ) : P is a partition of B} ,
is nonempty because such partitions exist (as observed in the previous sec-
tion), and similarly for the set of upper sums. Proposition 6.1.10 shows that
the set of lower sums is bounded above by any upper sum, and similarly the
set of upper sums is bounded below. Thus the next definition is natural.
Definition 6.2.1 (Lower Integral, Upper Integral, Integrability, Inte-
gral). The lower integral of f over B is the least upper bound of the lower
sums of f over all partitions P ,
Z
L f = sup {L(f, P ) : P is a partition of B} .
B

Similarly, the upper integral of f over B is the greatest lower bound of the
upper sums of f over all partitions P ,
Z
U f = inf {U (f, P ) : P is a partition of B} .
B

The function f is called

R integrable
R over B if the lower and upper integrals
are equal, i.e., if L B f = U B f . InR this case, their shared value is called the
integral of f over B and written B f .
268 6 Integration

So we have a quantitative definition that seems appropriate. The integral,

if it exists, is at least as big as any lower sum and at least as small as any upper
sum; and it is specified as the common value that is approached from below
by lower sums and from above by upper sums. Less formally, if quantities that
we view as respectively too small and too big approach a common value, then
that value must be what were after. R R
The following lemma shows that L B f U B f . Its proof provides an
example of how to work with lower and upper bounds. Note that the argument
does not require a contradiction or an , but rather it goes directly to the point.
Lemma 6.2.2 (Persistence of Order). Let L and U be nonempty sets of
real numbers such that

u for all L and u U . (6.1)

Then sup(L) and inf(U ) exist, and they satisfy

sup(L) inf(U ).

Proof. The given condition (6.1) rephrases as

for each L, u for all u U,

meaning precisely that

each L is a lower bound of U .

Since U is nonempty and has lower bounds, it has a greatest lower bound
inf(U ). Since each L is a lower bound and inf(U ) is the greatest lower
bound,
inf(U ) for each L,
meaning precisely that

inf(U ) is an upper bound of L.

Since L is nonempty and has an upper bound, it has a least upper bound
sup(L). Since sup(L) is the least upper bound and inf(U ) is an upper bound,

sup(L) inf(U ).

This is the desired result.

Again let a nonempty compact box R B and a Rbounded function f : B R
be given. The
R lemma shows that L B
f U B f (exercise). Therefore, to
show
R that R
B
f exists it suffices to show only that the reverse inequality holds,
L B f U B f.
Not all bounded functions f : B R are integrable. The standard
counterexample is the interval B = [0, 1] and the function
 6.2 Definition of the Integral 269
(
1 if x is rational,
f : B R, f (x) =
0 if x is irrational.
Chasing through the definitions shows that forR this B and f , any lower sum
is RL(f, P ) = 0, so the lower integral is L B f = sup {0} =R 0. Similarly,
U B f = 1. Since the upper and lower integrals dont agree, B f does not
exist.
So the questions are: what functions are integrable, or at least, what are
some general classes of integrable functions, and how does one evaluate their
integrals? Working from the definitions, as in the last example, is a good
exercise in simple cases to get familiar with the machinery, but as a general
procedure it is hopelessly unwieldy. Here is one result that will help us in the
next section to show that continuous functions are integrable.
Proposition 6.2.3 (Integrability Criterion). Let B be a box, and let f :
B R be a bounded function. Then f is integrable over B if and only if for
every > 0, there exists a partition P of B such that U (f, P ) L(f, P ) < .

RProof. ( = ) Let f be integrable over B and let > 0 be given. Since

B
f /2 is less than the least upper boundRof the lower sums, it is not an
upper bound of the lower sums, and similarly B f + /2 is not a lower bound
of the upper sums. Thus there exist partitions P and P of B such that
Z Z
L(f, P ) > f /2 and U (f, P ) < f + /2.
B B

Let P be the common refinement of P and P . Then since refining increases
lower sums and decreases upper sums, also
Z Z
L(f, P ) > f /2 and U (f, P ) < f + /2.
B B

This shows that U (f, P ) L(f, P R) < , as Rrequired.
( = ) We need to show that U B f L B f = 0. To do so, use the little
principle that to prove that a nonnegative number is zero, it suffices to show
that it is less than any positive number. Let > 0 be given. By assumption
there exists a partition P such that
U (f, P ) L(f, P ) < ,
and by the definition of upper and lower integral, also
Z Z
L(f, P ) L f U f U (f, P ).
B B

The last two displays combine to give

Z Z
U f L f < .
B B
R R
Since the positive number is arbitrary, U B
f L B
f = 0 as desired.

270 6 Integration

Here is an example of using the Integrability Criterion. It subsumes the

Rb Rc Rb
result from one-variable calculus that if a f exists then also a f and c f
Rb
exist for any c between a and b, and they they sum to a f .
Proposition 6.2.4. Let B be a box, let f : B R be a bounded function,
and let P be a partition of B. If f is integrable over B then f is integrable
over each subbox J of P , in which case
XZ Z
f= f.
J J B

Proof. Consider any partition P of B that refines P . For each subbox J of P ,

let PJ = P J, a partition of J. Let the symbol J denote subboxes of P ,
and compute that
X X X X
L(f, P ) = mJ (f ) vol(J ) = mJ (f ) vol(J ) = L(f, PJ ).
J J J J J

P
Similarly, U (f, P ) = J U (f, PJ ).
Suppose that f is integrable over B. Let an arbitrary > 0 be given. By
= of the Integrability Criterion, there exists a partition P of B such
that
U (f, P ) L(f, P ) < .
Since refining a partition cannot increase the difference between the upper and
lower sums, we may replace P by its common refinement with P and thus
assume that P refines P . Therefore the formulas from the previous paragraph
show that X
(U (f, PJ ) L(f, PJ )) < ,
J
and so
U (f, PJ ) L(f, PJ ) < for each subbox J of B.
Therefore f is integrable over each subbox J of B by = of the Integra-
bility Criterion.
Now assume that f is integrable over B and hence over each subbox J.
Still letting P be any partition of B that refines P , the integral over each
subbox J lies between the corresponding lower and upper sums, and so
X XZ X
L(f, P ) = L(f, PJ ) f U (f, PJ ) = U (f, P ).
J J J J
P R
Thus J J f is an upper bound of all lower sums L(f, P ) and a lower bound
of all upper sums U (f, P ), giving
Z XZ Z
L f f U f.
B J J B
R R R
But L B f = U B f = B f because f is integrable over B, and so the
inequalities in the previous display collapse to give the desired result.

 6.2 Definition of the Integral 271

Similar techniques show that the converse of the proposition holds as well,
so that given B, f , and P , f is integrable over B if and only if f is integrable
over each subbox J, but we do not need this full result. Each of the proposition
and its converse requires both implications of the Integrability Criterion.
The symbol B denotes a box in the next set of exercises.

Exercises

6.2.1. Let f R: B R
R be a bounded function. Explain how Lemma 6.2.2
shows that L B f U B f .

6.2.2. Let U and L be real numbers satisfying U L. Show that U = L if

and only if for all > 0, U L < .

6.2.3. Let f : B R be the constant

R function f (x) = k for all x B. Show
that f is integrable over B and B f = k vol(B).

6.2.4. Granting that any interval of positive length contains both rational
and irrational numbers, fill in the details in the argument that the function
f : [0, 1] R with f (x) = 1 for rational x and f (x) = 0 for irrational x is
not integrable over [0, 1].

6.2.5. Let B = [0, 1] [0, 1] R2 . Define a function f : B R by

(
0 if 0 x < 1/2,
f (x, y) =
1 if 1/2 x 1.
R
Show that f is integrable and B
f = 1/2.

6.2.6. This exercise shows that integration is linear. Let f : B R and

g : B R be integrable.
(a) Let P be a partition of B and let J be some subbox of P . Show that

mJ (f ) + mJ (g) mJ (f + g) MJ (f + g) MJ (f ) + MJ (g).

Show that consequently,

L(f, P ) + L(g, P ) L(f + g, P ) U (f + g, P ) U (f, P ) + U (g, P ).

(b) Part (a) of this exercise obtained comparisons between lower and upper
sums, analogously to the first paragraph of the proof R of Proposition 6.2.4.
Argue
R analogously
R to the rest of the proof to show B
(f + g) exists and equals
B
f + B
g. (One way to begin is by using the Integrability Criterion twice
and then a common refinement to show that there exists a partition P of B
such that U (f, P ) L(f, P ) < /2 and U (g, P ) L(g, P ) < /2.)
(c) Let c 0 be any constant. Let P be any partition of B. Show that for
any subbox J of P ,
272 6 Integration

mJ (cf ) = c mJ (f ) and MJ (cf ) = c MJ (f ).

Explain why consequently

L(cf, P ) = c L(f, P ) and U (cf, P ) = c U (f, P ).

Explain why consequently

Z Z Z Z
L cf = c L f and U cf = c U f.
B B B B
R
Explain why consequently B
cf exists and
Z Z
cf = c f.
B B

(d) Let P be any partition of B. Show that for any subbox J of P ,

mJ (f ) = MJ (f ) and MJ (f ) = mJ (f ).

Explain why consequently

L(f, P ) = U (f, P ) and U (f, P ) = L(f, P ).

Explain why consequently

Z Z Z Z
L (f ) = U f and U (f ) = L f.
B B B B
R
Explain why consequently B
(f ) exists and
Z Z
(f ) = f.
B B

Explain why the work so far here in part (d) combines

R with part (c) to show
that for any c R (positive, zero, or negative), B cf exists and
Z Z
cf = c f.
B B

6.2.7. This exercise shows that integration preserves order. Let f : B R

and g : B R both be integrable, and R suppose
R that f g, meaning that
f (x) g(x) for all x B. Show that B f B g. (Comment: Even though
f (x) g(x) for all x, upper sums for f can be bigger than upper sums for g (!),
so the argument requires a little finesse. Perhaps
R begin by explaining why the
previous exercise lets us show instead that B (g f ) 0. That is, introducing
Rthe function h = g f , we have h(x) 0 for all x and we need to show that
B
h 0. This is precisely the original problem with g = h and f = 0, so once
one has assimilated this idea one often says in similar contexts, We may take
f = 0.)
 6.3 Continuity and Integrability 273

6.2.8.
R  Suppose
R that f : B R is integrable, and that so is |f |. Show that
 f |f |.
B B

6.2.9. Prove the converse to Proposition 6.2.4: Let B be a box, let f : B R

be a bounded function, and let P be a partition of B. If f is integrable over
each subbox J of P then f is integrable over B. (You may quote the formulas
from the first paragraph of the proof in the text, since that paragraph makes
no assumptions of integrability. It may help to let b denote the number of
subboxes J, so that this quantity has a handy name.)

6.3 Continuity and Integrability

Although the Integrability Criterion gives a test for the integrability of any
specific function f , it is cumbersome to apply case by case. But handily, it
will provide the punchline of the proof of the next theorem, which says that
a natural class of functions is integrable.

Theorem 6.3.1 (Continuity Implies Integrability). Let B be a box, and

let f : B R be a continuous function. Then f is integrable over B.

To prove this, as we will at the end of the section, we first need to sharpen
our understanding of continuity on boxes. The version of continuity that were
familiar with isnt strong enough to prove certain theorems, this one in par-
ticular. Formulating the stronger version of continuity requires first revising
the grammar of the familiar brand.

Definition 6.3.2 (Sequential Continuity). Let S Rn be a set, and let

f : S Rm be a mapping. For any x S, f is sequentially continuous
at x if for every sequence {x } in S converging to x, the sequence {f (x )}
converges to f (x). The mapping f is sequentially continuous on S if f is
sequentially continuous at each point x in S.

Definition 6.3.3 (- Continuity). Let S Rn be a set, and let f : S

Rm be a mapping. For any x S, f is - continuous at x if for every > 0
there exists some > 0 such that

S and |
if x x x| < then |f (
x) f (x)| < .

The mapping f is - continuous on S if f is - continuous at each point

x in S.

Both definitions of continuity at a point x capture the idea that as inputs

to f approach x, the corresponding outputs from f should approach f (x).
This idea is exactly the substance of sequential continuity. (See figure 6.10.)
For - continuity at x, imagine that someone has drawn a ball of radius
(over which you have no control, and its probably quite small) about the
274 6 Integration

f
x f (x)

Figure 6.10. Sequential continuity

point f (x) in Rm . The idea is that in response, you can draw a ball of some
radiusthis is the in the definitionabout the point x in S, so that every
point in the -ball about x gets taken by f into the -ball about f (x). (See
figure 6.11.)

f
x f (x)

Figure 6.11. continuity

For example, the function f : Rn R given by f (x) = 2|x| is -

continuous on Rn . To show this, consider any point x Rn , and let > 0 be
given. Set = /2. Then whenever | x x| < , a calculation that uses the
generalized Triangle Inequality at the third step shows that

|f (
x) f (x)| = |2|
x| 2|x|| = 2||
x| |x|| 2|
x x| < 2 = ,

as needed. Thus f is - continuous at x, and since x is arbitrary, f is -

continuous on Rn .
For another example, to prove that the function f : R R given by
f (x) = x2 is - continuous on R, consider any x R and let > 0 be given.
This time set
= min{1, /(1 + 2|x|)}.
This choice of may look strange, but its first virtue is that since 1, for any
R with |
x x x| < , we have | x +x| = |
x x+2x| | x x|+2|x| < 1+2|x|;
and its second virtue is that also /(1+2|x|). These conditions fit perfectly
into the following calculation,
 6.3 Continuity and Integrability 275

|f ( x2 x2 |
x) f (x)| = |
= |
x + x| |
x x|

< (1 + 2|x|) by the two virtues of
1 + 2|x|
= .

And this is exactly what we needed to show that f is - continuous at x.

Since x is arbitrary, f is - continuous on R.
The tricky part of writing this sort of proof is finding the right . Doing
so generally requires some preliminary fiddling around on scratch paper. For
the proof just given, the initial scratch calculation would be

|f ( x2 x2 | = |(
x) f (x)| = | x x)| = |
x + x)( x + x| |
x x|,

exhibiting the quantity that we need to bound by as a product of two terms,

the second bounded directly by whatever we choose. The idea is initially
to make the first term reasonably small by stipulating that be at most 1,
giving as in the previous paragraph

|
x + x| = |
x x + 2x| |
x x| + 2|x| < 1 + 2|x|.

Now |f (x) f (x)| (1 + 2|x|)|

x x|. Next we constrain further to make this
estimate less than when | x x| < . Stipulating that be at most /(1+2|x|)
does so. Hence the choice of in the proof.
To prove instead that the function f : R R given by f (x) = x2 is
sequentially continuous on R, again take any x R. Consider any sequence
{x } in R converging to x. To show that the sequence {f (x )} in R converges
to f (x), compute

{f (x ) f (x)} = {x2 x2 } = {(x + x)(x x)}.

That is, the sequence {f (x ) f (x)} is the product of the sequences {x + x}

and {x x}. Because the sequence {x } converges it is bounded, and the
constant sequence {x} is bounded as well, so the sum {x + x} is bounded.
Also, because the sequence {x } converges to x, the sequence {x x} is
null. A bounded sequence times a null sequence is again null, and hence alto-
gether the sequence in the previous display is null. That is, {f (x )} converges
to f (x), showing that f is sequentially continuous at x. Since x is arbitrary,
f is sequentially continuous on R.
In fact there is no need to continue distinguishing between sequential con-
tinuity and - continuity, because each type of continuity implies the other.

Proposition 6.3.4 (Sequential Continuity and Continuity are

Equivalent). For any set S Rn and any mapping f : S Rm , f is
sequentially continuous on S if and only if f is - continuous on S.
276 6 Integration

Proof. Let x be any point of S.

( = ) Suppose that f is - continuous at x. We need to show that f is
sequentially continuous x. So, let {x } be a sequence in S converging to x. To
show that {f (x )} converges to f (x) means that given an arbitrary > 0, we
need to exhibit a starting index N such that

for all > N , |f (x ) f (x)| < .

The definition of - continuity gives a such that

S and |
if x x x| < then |f (
x) f (x)| < .

And since {x } converges in S to x, there is some starting index N such that

for all > N , |x x| < .

The last two displays combine to imply the first display, showing that f is
sequentially continuous at x.
( = ) Now suppose that f is not - continuous at x. Then for some > 0,
no > 0 satisfies the relevant conditions. In particular, = 1/ fails the
conditions for = 1, 2, 3, . So there is a sequence {x } in S such that

|x x| < 1/ and |f (x ) f (x)| , = 1, 2, 3, .

The display shows that f is not sequentially continuous at x.

Since the two types on continuity imply one another at each point x of S,
they imply one another on S.

The fact that the second half of this proof has to proceed by contrapo-
sition, whereas the first half is straightforward, shows that - continuity is
a little more powerful than sequential continuity on the face of it, until we
do the work of showing that they are equivalent. Also, the very definition
of - continuity seems harder for students than the definition of sequential
continuity, which is why these notes have used sequential continuity up to
now. However, the exceptionally alert reader may have recognized that the
second half of this proof is essentially identical to the proof of the Persistence
of Inequality Principle (Proposition 2.3.10). Thus, the occasional arguments
in these notes that cited Persistence of Inequality were tacitly using - con-
tinuity already, because sequential continuity was not transparently strong
enough for their purposes. The reader who dislikes redundancy is encouraged
to rewrite the second half of this proof to quote Persistence of Inequality
rather than re-prove it.
The reason that we bother with this new - type of continuity, despite
its equivalence to sequential continuity meaning that it is nothing new, is
that its grammar generalizes to describe the more powerful continuity that
we need. The two examples above of - continuity differed: in the example
 6.3 Continuity and Integrability 277

f (x) = x2 , the choice of = min{1, /(2|x| + 1)} for any given x and to
satisfy the definition of - continuity at x depended not only on but on x
as well. In the example f (x) = 2|x|, the choice of = /2 for any given x
and depended only on , i.e., it was independent of x. Here, one value of
works simultaneously at all values of x once is specified. This technicality
has enormous consequences.

Definition 6.3.5 (Uniform Continuity). Let S Rn be a set, and let

f : S Rm be a mapping. Then f is uniformly continuous on S if for
every > 0 there exists some > 0 such that

S and |
if x, x x x| < then |f (
x) f (x)| < .

The nomenclature uniformly continuous on S is meant to emphasize that

given > 0, a single, uniform value of works in the definition of - continuity
simultaneously for all points x S. The scope of its effectiveness is large-scale.
Uniform continuity depends on both the mapping f and the set S.
A visual image may help distinguish between the old notion of continuity
(henceforth called pointwise continuity) and the new, stronger notion of
uniform continuity. Imagine the graph of a function f : S R (where
S R), and take some input point x. Then f is pointwise continuous at x
if for any > 0, one can draw a rectangle of height 2 centered at the point
(x, f (x)) that is narrow enough that the graph of f protrudes only from the
sides of the rectangle, not the top or bottom. The base of the rectangle is 2,
where comes from - continuity. Note that for a given , one may need
rectangles of various widths at different points. A rectangle that works at
x may not be narrow enough to work again at some other point x . (See
figure 6.12, where ever-narrower rectangles are required as we move to the left
on the graph.) On the other hand, the function f is uniformly continuous if
given > 0, there is a single 2-by-2 rectangle that can slide along the entire
graph of f with its centerpoint on the graph, and the graph never protruding
from the top or bottom. (See figure 6.13. A tacit assumption here is that the
graph of f either doesnt extend beyond the picture frame, or it continues to
rise and fall tamely if it does.) By contrast, no single rectangle will work in
figure 6.12.
The domain of the nonuniformly continuous function f (x) = sin(1/x) in
figure 6.12 is not compact, not being closed at its left endpoint. We are about
to prove that on a compact domain, uniform continuity follows for free from
pointwise continuity. In conjunction with the compactness of the boxes B over
which we integrate, this is is the crucial ingredient for proving Theorem 6.3.1
(continuous functions are integrable over boxes), the goal of this section.

Theorem 6.3.6 (Continuity on Compact Sets is Uniform). Let K Rn

be compact, and let f : K Rm be pointwise continuous on K. Then f is
uniformly continuous on K.
278 6 Integration

Figure 6.12. One can require different values at different points x

Figure 6.13. Or one can work uniformly for at all x

As with the proof that sequential continuity implies - continuity, we

proceed by contraposition. That is, we show that in the case of a compact
domain, if f is not uniformly continuous then f can not be continuous either.

Proof. Suppose that f is not uniformly continuous. Then for some > 0
there exists no suitable uniform , and so in particular no reciprocal positive
integer 1/ will serve as in the definition of uniform continuity. Thus for
each Z+ there exist points x and y in K such that

|x y | < 1/ and |f (x ) f (y )| . (6.2)

Consider the sequences {x } and {y } in K. By the sequential characteri-

zation of compactness (Theorem 2.4.13), {x } has a convergent subsequence
converging in K; call it {xk }. Throw away the rest of the x s and throw
away the y s of corresponding index, reindex the remaining terms of the two
sequences, and now {x } converges to some p K. Since |x y | < 1/ for
each (this remains true after the reindexing), {y } converges to p as well.
So
lim x = p = lim y ,
and thus
 6.3 Continuity and Integrability 279

f (lim x ) = f (lim y ).
But the second condition in (6.2) shows that

lim f (x ) 6= lim f (y ),

i.e., even if both limits exist then still they can not be equal. (If they both
exist and they agree then lim(f (x ) f (y )) = 0, but this is incompatible
with the second condition in (6.2), |f (x ) f (y )| for all .) The previous
two displays combine to show that

lim f (x ) 6= f (lim x ) or lim f (y ) 6= f (lim y ),

i.e., at least one of the left sides in the previous display doesnt match the
corresponding right side or doesnt exist at all. Thus f is not continuous at p.

Recall the main Rresult that we want: If B is a box in Rn and f : B R

is continuous then B f exists. The result is easy to prove now. The crucial
line of the proof is the opener.
Proof (of Theorem 6.3.1). The continuity of f on B is uniform. Thus, given
> 0, there exists > 0 such that

B and |
if x, x x x| < then |f (
x) f (x)| < .
vol(B)
(We may take vol(B) > 0, making the volume safe to divide by, since otherwise
all lower sums and upper sums are 0, making the integral 0 as well, and there
is nothing to prove.) Take a partition P of B whose subboxes J have sides
of length less than /n. By the Size Bounds (Proposition 2.2.7), any points x
and x in any subbox J satisfy |
x x| < , so

J then |f (
if x, x x) f (x)| < .
vol(B)
Let x and x vary over J, and cite the Extreme Value Theorem (Theorem
2.4.15) to show that

MJ (f ) mJ (f ) < .
vol(B)
Multiply by vol(J) to get
vol(J)
MJ (f )vol(J) mJ (f )vol(J) < ,
vol(B)
and sum this relation over subboxes J to get

U (f, P ) L(f, P ) < .

R
The Integrability Criterion now shows that B f exists.

280 6 Integration

Integration synthesizes local data at each point of a domain into one whole.
The idea of this section is that integrating a continuous function over a box
is more than a purely local process: it requires the uniform continuity of the
function all through the box, a large-scale simultaneous estimate that holds
in consequence of the box being compact.

Exercises

6.3.1. Reread the proof that sequential and - continuity are equivalent, then
redo the proof with the book closed.

6.3.2. Let f : R R be the cubing function f (x) = x3 . Give a direct proof

that f is - continuous on R. (Hint: A3 B 3 = (A B)(A2 + AB + B 2 ).)

6.3.3. Here is a proof that the squaring function f (x) = x2 is not uniformly
continuous on R. Suppose that some > 0 satisfies the definition of uniform
continuity for = 1. Set x = 1/ and x = 1/+/2. Then certainly |
x x| < ,
but
   
 1 2 1   1 2 1 2

|f (
x) f (x)| =  + 2 =  2 + 1 + 2  = 1 + > .
 2  4 4

This contradicts uniform continuity.

Is the cubing function of the previous exercise uniformly continuous on R?
On [0, 500]?

6.3.4. (a) Show: If I R is an interval (possibly all of R), f : I R is

differentiable, and there exists a positive constant R such that |f (x)| R for
all x I then f is uniformly continuous on I.
(b) Prove that sine and cosine are uniformly continuous on R.

6.3.5. Let f : [0, +) R be the square root function f (x) = x. You
may take for granted that f is - continuous on [0, +).
(a) What does part (a) of the previous problem say about the uniform
continuity of f ?
(b) Is f uniformly continuous?

6.3.6. Let J be a box in Rn with sides of length less than /n. Show that any
points x and x in J satisfy |
x x| < .
R
6.3.7. For B f to exist, it is sufficient that f : B R be continuous, but it
is not necessary. What preceding exercise provides an example of this? Here is
another example. Let B = [0, 1] and let f : B R be monotonic increasing,
meaning that if x1 < x2 in B then f (x1 ) f (x2 ). Show that such a function
is bounded, though
R it need not be continuous. Use the Integrability Criterion
to show that B f exists.
 6.4 Integration of Functions of One Variable 281

6.3.8. The natural logarithm is defined as an integral. Let r : R+ R be

the reciprocal function, r(x) = 1/x for x > 0. The natural logarithm is
( R
r if x 1,
ln : R+ R, ln(x) = R[1,x]
[x,1] r if 0 < x < 1.

We know that the integrals in the previous display exist because the reciprocal
function is continuous.
(a) Show that limx ln x/x = 0 as follows. Let some small > 0 be given.
For x > 2/, let u(x, ) denote the sum of the areas of the boxes [1, 2/][0, 1]
and [2/, x] [0, /2]. Show that u(x, ) ln x. (Draw a figure showing the
boxes and the graph of r, and use the words upper sum in your answer.)
Compute limx u(, x)/x (here remains fixed), and use your result to show
that u(, x)/x < for all large enough x. This shows that limx ln x/x = 0.
(b) Let a > 0 and b > 1 be fixed real numbers. Part (a) shows that

ln x/x < ln b/(a + 1) for all large x.

Explain why consequently

xa /bx < 1/x for all large x.

This proves that exponential growth dominates polynomial growth,

xa
lim = 0, a > 0, b > 1.
x bx

Thus, for example,

x1000000
lim = 0,
x 1.0000001x

even though the values of x1000000 /1.0000001x are enormous as x begins to

grow.

6.4 Integration of Functions of One Variable

In a first calculus course one learns to do computations such as: to evaluate
Z e
(ln x)2
dx,
x=1 x

let u = ln x; then du = dx/x, and as x goes from 1 to e, u goes from 0 to 1,

so the integral equals 1
Z 1
2 1 3  1
u du = u  = .
u=0 3 0 3
Or such as: to evaluate
282 6 Integration
Z 9
dx
p .
0 1+ x
p
let u = 1 + x. Then some algebra shows that x = (u2 1)2 , and so
dx = 4(u2 1)u du. Also, when x = 0, u = 1, and when x = 9, u = 2.
Therefore the integral is
Z 9 Z 2 2 Z 2
dx (u 1)u
p =4 du = 4 (u2 1) du
0 1+ x 1 u 1
  2
1 3  16
=4 u u  = .
3 3 1

Although both of these examples use substitution, they differ from one
another in a way that a first calculus course may not explain. The first substi-
tution involved picking an x-dependent u (i.e., u = ln x) where u (x) (i.e., 1/x)
was present in the integral and got absorbed by the substitution. The second
substitution took an opposite form to the first: this time the x-dependent u
was inverted to produce a u-dependent x, and the factor u (x) was introduced
into the integral rather than eliminated from it. Somehow, two different things
are going on under the guise of u-substitution.
In this section we specialize our theory of multivariable integration to n = 1
and review two tools for evaluating one-dimensional integrals, the Fundamen-
tal Theorem of Integral Calculus (FTIC) and the Change of Variable Theorem.
Writing these down precisely will clarify the examples we just worked. More
importantly, generalizing these results appropriately to n dimensions is the
subject of the remainder of these notes.
The multivariable
R integral notation of this chapter, specialized to one di-
mension, is [a,b] f . For familiarity, replace this by the usual notation,
Z b Z
f= f for a b.
a [a,b]
Rb
As matters stand, the redefined notation a f makes sense only when a b,
so extend its definition to
Z b Z a
f = f for a > b.
a b

Once this is done, the same relation between signed integrals holds regardless
of which (if either) of a and b is larger,
Z b Z a
f = f for all a and b.
a b

Something nontrivial is happening here: when the multivariable integration of

this chapter is specialized to one dimension, it can be extended to incorporate
 6.4 Integration of Functions of One Variable 283
Rb
a sign convention to represent the order on R. If a < b then a describes
Ra
positive traversal along the real line from a up to b, while b describes negative
traversal from b down to a. This sort of thing does not obviously generalize
to higher dimensions since Rn is not ordered.
Casewise inspection shows that for any three points a, b, c R in any
order, and for any integrable function f : [min{a, b, c}, max{a, b, c}] R,
Z c Z b Z c
f= f+ f.
a a b

Also, if f : [min{a, b}, max{a, b}] R takes the constant value k then
Z b
f = k(b a),
a

again regardless of which of a and b is larger. These facts generalize Proposi-

tion 6.2.4 and exercise 6.2.3 to signed one-variable integration.
Each of the next two theorems describes a sense in which one-variable dif-
ferentiation and integration are inverse operations. Both are called the Fun-
damental Theorem of Integral Calculus, but the second is more deserving of
the title because of how far it generalizes.

Theorem 6.4.1. Let the function f : [a, b] R be continuous. Define a

function Z x
F : [a, b] R, F (x) = f.
a

Then F is differentiable on [a, b] and F = f .

Proof. Let x and x + h lie in [a, b] with h 6= 0. Study the difference quotient
R x+h Rx R x+h
F (x + h) F (x) a
f a
f x
f
= = .
h h h
R x+h
If h > 0 then m[x,x+h] (f ) h x f M[x,x+h] (f ) h, and dividing
through by h shows that the difference quotient lies between m[x,x+h] (f ) and
M[x,x+h] (f ). Thus the difference quotient is forced to f (x) as h goes to 0,
since f is continuous. Similar analysis applies when h < 0.
(Alternatively, an argument using the characterizing property of the
derivative and the LandauBachmann notation does not require separate cases
depending on the sign of h. Compute that
Z x+h Z x+h
F (x + h) F (x) f (x)h = (f f (x)) = o(1) = o(h),
x x

But here the reader needs to believe, or check, the last equality.)

284 6 Integration

The alert reader will recall the convention in these notes that a mapping
can be differentiable only at an interior point of its domain. In particular,
the derivative of a function F : [a, b] R is undefined at a and b. Hence
the statement of Theorem 6.4.1 is inconsistent with our usage, and strictly
speaking the theorem should conclude that F is continuous on [a, b] and dif-
ferentiable on (a, b) with derivative F = f . The given proof does show this,
since the existence of the one-sided derivative of F at each endpoint makes F
continuous there.
However, we prohibited derivatives at endpoints only to tidy up our state-
ments. An alternative would have been to make the definition that for any
compact, connected set K Rn (both of these terms were discussed in sec-
tion 2.4), a mapping f : K Rm is differentiable on K if there exists an
open set A Rn containing K, and an extension of f to a differentiable map-
ping f : A Rm . Here the word extension means that the new function f
on A has the same behavior on K as the old f . One reason that we avoided
this slightly more general definition is that it is tortuous to track through the
material in chapter 4, especially for the student who is seeing the ideas for
the first time. Also, this definition requires that the Critical Point Theorem
(stating that the extrema a function occur at points where its derivative is 0)
be fussily rephrased to say that this criterion applies only to the extrema that
occur at the interior points of the domain. From the same preference for tidy
statements over fussy ones, now we do allow the more general definition of
the derivative.
Proving the FTIC from Theorem 6.4.1 requires the observation that if two
functions F1 , F2 : [a, b] R are differentiable, and F1 = F2 , then F1 = F2 +c
for some constant c. The observation follows from the Mean Value Theorem
and is an exercise.

Theorem 6.4.2 (Fundamental Theorem of Integral Calculus). Sup-

pose that the function F : [a, b] R is differentiable and F is continuous.
Then Z b
F = F (b) F (a).
a
Rx
Proof. Define F2 : [a, b] R by F2 (x) = a F . Then F2 = F by the
preceding theorem, so (exercise 6.4.3) there exists a constant c such that for
all x [a, b],
F2 (x) = F (x) + c. (6.3)
Plug x = a into (6.3) to get 0 = F (a) + c, so c = F (a). Next plug in x = b
Rb
to get F2 (b) = F (b) F (a). Since F2 (b) = a F by definition, the proof is
complete.

(One can also prove the Fundamental Theorem with no reference to Theo-
rem 6.4.1, letting the Mean Value Theorem do all the work instead. Compute
that for any partition P of [a, b], whose points are a = t0 < t1 < < tk = b,
 6.4 Integration of Functions of One Variable 285
k
X
F (b) F (a) = F (ti ) F (ti1 ) (telescoping sum)
i=1
k
X
= F (ci )(ti ti1 ) with each ci (ti1 , ti ), by the MVT
i=1
U (F , P ).

Since P is arbitrary, F (b)F (a) is a lower bound of the upper sums and hence
Rb
is at most the upper integral U a F . Since F is continuous, its integral exists
and the upper integral is the integral. That is,
Z b
F (b) F (a) F .
a

A similar argument with lower sums gives the opposite inequality.)

In one-variable calculus one learns various techniques to find antideriva-
tives; i.e., given continuous f , one finds F such that F = f . Once this is
Rb
done, evaluating a f is mere plug-in to the FTIC. But since not all continu-
ous functions have antiderivatives that are readily found, or even possible to
2
write in an elementary form (for example, try f (x) = ex or f (x) = sin(x2 )),
the FTIC has its limitations.
Another tool for evaluating one-dimensional integrals is the Change of
Variable Theorem. The idea is to transform one integral to another that may
be better suited to the FTIC.

Theorem 6.4.3 (Change of Variable Theorem; Forward Substitution

Formula). Let : [a, b] R be differentiable with continuous derivative
and let f : [a, b] R be continuous. Then
Z b Z (b)

(f ) = f. (6.4)
a (a)

Proof. Use Theorem 6.4.1 to define F : [a, b] R such that F = f . By

the chain rule, F has derivative (F ) = (F ) = (f ) , which
is continuous on [a, b]. Thus by the FTIC twice,
Z b Z b
(f ) = (F ) = (F )(b) (F )(a)
a a
Z (b) Z (b)
= F ((b)) F ((a)) = F = f.
(a) (a)

Rb
One way to apply the Change of Variable Theorem to an integral a g is
to recognize that the integrand takes the form g = (f ) , giving the
286 6 Integration
R (b)
left side of (6.4) for suitable f and such that the right side (a) f is easier
to evaluate. This method is calledR integration by forward substitution.
e
For instance, for the first integral x=1 ((ln x)2 )/x) dx at the beginning of the
section, take
g : R+ R, g(x) = (ln x)2 /x.
Re
To evaluate 1 g, define

: R+ R, (x) = ln x

and
f : R R, f (u) = u2 .
Then g = (f ) , and (1) = 0, (e) = 1, so by the Change of Variable
Theorem,
Z e Z e Z (e) Z 1
g= (f ) = f= f.
1 1 (1) 0

Since f has antiderivative F where F (u) = u3 /3, the last integral equals
F (1) F (0) = 1/3 by the FTIC.
The second integral at the beginning of the section was evaluated not by
the Change of Variable Theorem as given, but by a consequence of it:
Corollary 6.4.4 (Inverse Substitution Formula). Let : [a, b] R be
continuous and let f : [a, b] R be continuous. Suppose further that is
invertible and that 1 is differentiable with continuous derivative. Then
Z b Z (b)
(f ) = f (1 ) .
a (a)

The formula in the corollary is the formula for integration by inverse

substitution. To obtain it from (6.4), consider the diagrams for forward and
inverse substitution:
1
[a, b] / [(a), (b)] [(a), (b)] / [a, b]



f
f f

f
! z
$
}
R R
Noting where the various elements of the left diagram occur in the forward
Rb R (b)
substitution formula a (f ) = (a) f shows that applying the forward
R (b) Rb
substitution suitably to the right diagram gives (a) f (1 ) = a (f ),
the inverse substitution formula as claimed. Rb
To apply the formula in Corollary 6.4.4 to an integral a g, write the
integrand as g = f , giving the left side, and then invert and differentiate
the inverse to see
R 9 if the
prightside is easier to evaluate. For instance, for the
second integral 0 dx/ 1 + x at the beginning of the section, define
 6.4 Integration of Functions of One Variable 287
q

: R0 R1 , (x) = 1+ x

and
f : R1 R, f (u) = 1/u.
Then the integral is Z Z
9 9
dx
p = (f ).
0 1+ x 0

Let q

u = (x) = 1 + x.
Then a little algebra gives

x = (u2 1)2 = 1 (u),

so that
(1 ) (u) = 4u(u2 1).
Since (0) = 1 and (9) = 2, the integral becomes
Z 9 Z 9 Z 2 Z 2
dx u(u2 1) du
p = (f ) = f (1 ) = 4 ,
0 1+ x 0 1 1 u

and as before this evaluates easily to 16/3.

The variable-based notation used to work the two integrals at the begin-
ning of the section, with x and u and dx and du, is much easier mnemonically
than the function-based notation used to rework them with the Change of
Variable Theorem and its corollary. But a purist would object to it on two
counts. First, expressions such as (ln x)2 /x and u2 are not functions, they are
the outputs of functions, so strictly speaking we cant integrate them. The
problem is not serious, it is mere pedantry: we simply need to loosen our
Rb Rb
notation to let x=a f (x) be synonymous with a f , at the cost of an unnec-
essary new symbol x. This x is called a dummy variable since another symbol
Rb Rb Rb
would do just as well: y=a f (y) and =a f () also denote a f . At the the-
oretical level, where we deal with functions qua functions, this extra notation
is useless and cumbersome, but in any down-to-earth example it is in fact a
convenience since describing functions by formulas is easier and more direct
than introducing new symbols to name them.
The second, more serious objection to the variable-based notation is to
the dx, the du, and mysterious relations such as du = dx/x between them.
What kind of objects are dx and du? In a first calculus course they are typ-
ically described as infinitesimally small changes in x and u, but our theory
of integration is not based on such hazy notions; in fact it was created in
the 19th century to answer objections to their validity. (Though infinitesimals
were revived and put on a firm footing in the 1960s, we have no business with
them here.) An alternative is to view dx and du as formal symbols that serve,
288 6 Integration
R
along with the integral sign , as bookends around the expression for the
function being integrated. This viewpoint leaves notation such as du = dx/x
still meaningless in its own right. In a first calculus course it may be taught
as a procedure with no real justification, whereas by contrast, the revisited
versions of the two integral-calculations of this section are visibly applications
of results that have been proved. However, the classical method is probably
easier for most of us, its notational conventions dovetailing with the Change
of Variable Theorem and its corollary so well. So feel free to continue using
it. (And remember to switch the limits of integration when you do.)
However, to underscore that dx is an unnecessary, meaningless symbol, it
will generally not be used in these notes until it is defined in chapter 9, as
something called a differential form.

Exercises

6.4.1. (a) Show that for three points a, b, c R in any order, and any inte-
Rc Rb Rc
grable function f : [min{a, b, c}, max{a, b, c}] R, a f = a f + b f .
(b) Show that if f : [min{a, b}, max{a, b}] R takes the constant value k
Rb
then a f = k(b a), regardless of which of a and b is larger.
6.4.2. Complete the proof of Theorem 6.4.1 by analyzing the case h < 0.
6.4.3. Show that if F1 , F2 : [a, b] R are differentiable and F1 = F2 , then
F1 = F2 + C for some constant C. This result was used in the section to prove
the Fundamental Theorem of Calculus (Theorem 6.4.2), so do not use that
theorem to address this exercise. However, this exercise does require a theo-
rem. Reducing to the case where F2 = 0, as in the comment in exercise 6.2.7,
will make this exercise a bit tidier.
6.4.4. (a) Suppose that 0 a b and f : [a2 , b2 ] R is continuous. Define
R x2
F : [a, b] R by F (x) = a2 f . Does F exist, and if so then what is it?
(b) More generally, suppose f : R R is continuous, and , : R R
R (x)
are differentiable. Define F : R R by F (x) = (x) f . Does F exist, and
if so then what is it?
6.4.5. Let f : [0, 1] R be continuous and suppose that for all x [0, 1],
Rx R1
0
f = x f . What is f ?
6.4.6. Find all differentiable
Rx functions f : R0 R such that for all x
R0 , (f (x))2 = 0 f .
1
R6.4.7. Define f : R+ R by f (u) = e(u+ u ) /u and F : R+ R by F (x) =
x
1
f . Show that F behaves somewhat like a logarithm in that F (1/x) = F (x)
for all x R+ . Interpret this property of F as a statement about area under
the graph of f . (Hint: define : R+ R+ by (u) = 1/u, and show that
(f ) = f .)
 6.5 Integration Over Nonboxes 289

6.5 Integration Over Nonboxes

R
So far we know that B f exists if B is a box and f : B R is continuous
(Theorem 6.3.1). With some more work, the R theorem can be refined to relax
these requirements. The basic idea is that B f still exists if f is discontinuous
on a small enough subset of B. The idea isnt hard conceptually, but its
justification requires some bookkeeping. Once it is established, integration
over compact sets K other than boxes is easy to define provided that their
boundaries are suitably small.
To quantify the notion of small, and more generally the notion of set size,
let a set S Rn be given. The characteristic function of S is
(
n 1 if x S
S : R R, S (x) =
0 otherwise.

Suppose that S is bounded, meaning that S sits in some box B.

Definition 6.5.1 (Volume of a Set). The volume of a bounded set S Rn

is Z
vol(S) = S where B is any box containing S,
B
if this integral exists.

This definition
R requires several comments. At first glance it seems ill-posed.
Conceivably, B S could exist for some boxes B containing S but not others,
and it could take different
R values for the various B where it exists. In fact, some
technique shows that if B S exists for some box B containing S then it exists
for any such box and always takes the same value, so the definition makes sense
after all. See the exercises. Also, an exercise shows that the volume of a box B
is the same under Definition 6.5.1 as under Definition 6.1.4, as it must be
for grammatical consistency. Finally, note that not all sets have volume, only
those whose characteristic functions are integrable.
Sets of volume zero are small enough that they dont interfere with inte-
gration. To prove such a result explicitly, we first translate the definition of
volume zero into statements about the machinery of the integral. Let S Rn
sit in a box B, and let P be a partition of B. The subboxes J of P consist of
two types:
type I : J such that J S 6=
and
type II : J such that J S = .
Thus S sits in the unionRof subboxes J of type I and the sum of their volumes
gives an upper sum for B S .
For example, figure 6.14 shows a circle S inside a box B, and a partition P
of B, where the type I subboxes of the partition are shaded. The shaded boxes
290 6 Integration

visibly have a small total area. Similarly, figure 6.15 shows a smooth piece of
surface in R3 , then shows it inside a partitioned box, and figure 6.16 shows
some of the type I subboxes of the partition. Figure 6.16 also shows a smooth
arc in R3 and some of the type I rectangles that cover it, with the ambient
box and the rest of the partition now tacit. Figure 6.16 is meant to show that
all the type I boxes, which combine to cover the surface or the arc, have a
small total volume.

Figure 6.14. Circle, box, partition, and type I subboxes

Figure 6.15. A two dimensional set in R3 ; the set inside a partitioned box

The following fact is convenient.

 6.5 Integration Over Nonboxes 291

Figure 6.16. Some type I subboxes of the partition, and for an arc in R3

Proposition 6.5.2 (Volume Zero Criterion). A set S contained in the

box B has volume zero if and only if for every > 0 there exists a partition
P of B such that X
vol(J) < .
J:type I

The proof is an exercise. This criterion makes it plausible that any bounded
smooth arc in R2 has volume zero, and similarly for a bounded smooth arc or
smooth piece of surface in R3 . The next result uses the criterion to provide
a general class of volume zero sets. Recall that for any set S Rk and any
mapping : S R , the graph of is a subset of Rk+ ,

graph() = {(x, (x)) : x S}.

Proposition 6.5.3 (Graphs Have Volume Zero). Let B be a box in Rm ,

and let : B R be continuous. Then graph() has volume zero.

The idea is that the graph of the function in the proposition will describe
some of the points of discontinuity of a different function f that we want to
integrate. Thus the dimension m in the proposition is typically n 1, where
the function f that we want to integrate has n-dimensional input.

Proof. The continuity of on B is uniform, and the image of , being com-

pact, sits in some interval I.
Let > 0 be given. Set equal to any positive number less than
/(2vol(B)) such that length(I)/ is an integer. There exists a partition Q of I
whose subintervals K have length , and a > 0 such that for all x, x B,

|
x x| < = x) (x)| < .
|( (6.5)
292 6 Integration

Now take a partition P of B whose subboxes J have sides of length less

than /m, so that if two points are in a common subbox J then the distance
between them is less than . Consider the partition P Q of B I. For each
subbox J of P there exist at most two subboxes J K of P Q over J
that intersect the graph of , i.e., subboxes of type I. To see this, note that
if we have three or more such subboxes, then some pair J K and J
K are not vertical neighbors, and so any hypothetical pair of points of the
graph, one in each subbox, are less than distance apart horizontally but
at least distance apart vertically. But by (6.5), this is impossible. (See
figure 6.17. The horizontal direction in the figure is only a schematic of the
m-dimensional box B, but the vertical direction accurately depicts the one-
dimensional codomain of .)

Figure 6.17. The graph meets at most two boxes over each base

Now, working with subboxes J K of P Q, compute that

X X
vol(J K) = vol(J) since length(K) =
type I type I
X
2 vol(J) by the preceding paragraph
J
= 2vol(B) < since < /(2vol(B)),

and the proof is complete by the Volume Zero Criterion.

An exercise shows that any finite union of sets of volume zero also has
volume zero, and another exercise shows that any subset of a set of volume
zero also has volume zero. These facts and the preceding proposition are
enough to demonstrate that many regions have boundaries of volume zero. The
boundary of a set consists of all points simultaneously near the set and near
its complementroughly speaking, its edge. (Unfortunately, the mathematical
 6.5 Integration Over Nonboxes 293

terms bounded and boundary need have nothing to do with each other. A set
with a boundary need not be bounded, and a bounded set need not have any
boundary points nor contain any of its boundary points if it does have them.)
For example, the set in figure 6.18 has a boundary consisting of four graphs
of functions on one-dimensional boxes, i.e., on intervals. Two of the boundary
pieces are graphs of functions y = (x), and the other two are graphs of
functions x = (y). Two of the four functions are constant functions.

y = 2 x2

x = sin(y)

x=2

y=0

Figure 6.18. Boundary with area zero

The main result of this section is that discontinuity on a set of volume

zero does not interfere with integration.

Theorem 6.5.4 (Near-continuity Implies Integrability). Let B Rn

be a box. Let f : B
R R be bounded, and continuous except on a set S B
of volume zero. Then B f exists.

Proof. Let > 0 be given.

The proof involves two partitions. Since f is bounded there exists a pos-
itive number R such that |f (x)| < R for all x B. Take a partition P of B
whose subboxes J of type I (those intersecting the set S where f is discon-
tinuous) have volumes adding to less than /(4R). (See figure 6.19, in which
the function f is the dome shape over the unit disk but is 0 outside the unit
disk, making it discontinuous on the unit circle.)
Consider some yet unspecified refinement P of P , dividing each subbox J
of P into further subboxes J . (See figure 6.20, in which the original boxes J
of type I remain shaded, but each box J of either type has been further
partitioned.) On any J , MJ (f ) mJ (f ) 2R, and so a short calculation
shows that regardless how the refinement P is to be specified, its subboxes J
that sit inside type I subboxes J of P make only a small contribution to the
difference between the lower sum and the upper sum of f over P ,
294 6 Integration

Figure 6.19. Type I subboxes of small total area

X X
(MJ (f ) mJ (f )) vol(J )
J : type I J J
X X X (6.6)
2R vol(J ) = 2R vol(J) < 2R = .
4R 2
J : type I J J J : type I

Figure 6.20. Refinement of the partition

To specify the refinement P of P that we need, consider the type II

subboxes J of P , i.e., the union of the unshaded boxes in figure 6.19. The
function f is continuous on each such subbox and hence integrable over it by
Theorem 6.3.1. Let the number of these subboxes be denoted N . By ( = )
of the Integrability Criterion, each type II subbox J has a partition PJ such
that
 6.5 Integration Over Nonboxes 295

U (f, PJ ) L(f, PJ ) <
.
2N
Let P be a partition of the full box B that refines the original partition P
and also incorporates all the partitions PJ of the type II subboxes J. Thus
the intersection of P with any particular type II subbox J refines PJ . Since
refinement can not increase the distance between lower and upper sums, an-
other short calculation shows that the subboxes J of P that sit inside type II
subboxes J of P also make only a small contribution to the difference between
the lower sum and the upper sum of f over P ,
X X
(MJ (f ) mJ (f )) vol(J )
J : type II J J
X (6.7)
U (f, PJ ) L(f, PJ ) < N = .
2N 2
J : type II

Finally, combining (6.6) and (6.7) shows Rthat U (f, P ) L(f, P ) < , and so
by ( = ) of the Integrability Criterion, B f exists.

To recapitulate the argument: The fact that f is bounded means that its
small set of discontinuities cant cause much difference between lower and up-
per sums, and the continuity of f on the rest of its domain poses no obstacle
to integrability either. The only difficulty was making the ideas fit into our
box-counting definition of the integral. The reader could well object that prov-
ing Theorem 6.5.4 shouldnt have to be this complicated. Indeed, the theory
of integration being presented here, Riemann integration, involves laborious
proofs precisely because it uses such crude technology: finite sums over boxes.
More powerful theories of integration exist, with stronger theorems and more
graceful arguments. However, these theories also entail the startup cost of as-
similating a larger, more abstract set of working ideas, making them difficult
to present as quickly as Riemann integration.
Now we can discuss integration over nonboxes.
Definition 6.5.5 (Known-Integrable Function). A function
f : K R
is known-integrable if K is a compact subset of Rn having boundary of
volume zero, and if f is bounded on K and is continuous on all of K except
possibly a subset of volume zero.
For example, let K = {(x, y) : |(x, y)| 1} be the closed unit disk in R2 ,
and define (
1 if x 0,
f : K R, f (x, y) =
1 if x < 0.
To see that this function is known-integrable, note that the boundary of K
is the union of the upper and lower unit semicircles, which are graphs of
continuous functions on the same 1-dimensional box,
296 6 Integration
p
: [1, 1] R, (x) = 1 x2 .

Thus the boundary of K has area zero. Furthermore, f is bounded on K, and

f is continuous on all of K except the vertical interval {0} [1, 1], which
has area zero by the 2-dimensional box area formula.

Definition 6.5.6 (Integral Over a Nonbox). Let

f : K R

be a known-integrable function. Extend its domain to Rn by defining a new

function (
n f (x) if x K,
f : R R, f (x) =
0 if x
/ K.
Then the integral of f over K is
Z Z
f= f where B is any box containing K.
K B

For the example just before the definition, the extended function is


1 if |(x, y)| 1 and x 0,
2
f : R R, f (x, y) = 1 if |(x, y)| 1 and x < 0,


0 if |(x, y)| > 1,

and to integrate the original function over the disk we integrate the extended
function over the box B = [0, 1] [0, 1].
Returning to generality, the integral on the right side of the equality in the
definition exists because f is bounded and discontinuous on a set of volume
zero, as required for Theorem 6.5.4. In particular, the definition of volume is
now, sensibly enough, Z
vol(K) = 1.
K
Naturally, the result of Proposition 6.2.4, that the integral over the whole
is the sum of the integrals over the pieces, is not particular to boxes and
subboxes.

Proposition 6.5.7. Let K Rn be a compact set whose boundary has volume

zero. Let f : K R be continuous. Further, let K = K1 K2 where each Kj
is compact and the intersection K1 K2 has volume zero. Then f is integrable
over K1 and K2 , and Z Z Z
f+ f= f.
K1 K2 K
 6.5 Integration Over Nonboxes 297

Proof. Define
(
f (x) if x K1 ,
f1 : K R, f1 (x) =
0 otherwise.
R R
Then f1 is known-integrable on K, and so K f1 exists and equals K1 f1 .
Define a corresponding function f2 : K R, for which the corresponding
conclusions hold. It follows that
Z Z Z Z Z
f1 + f2 = f1 + f2 = (f1 + f2 ).
K1 K2 K K K

But f1 +f2 equals f except on the volume-zero set K1 K2 , which contributes

nothing to the integral. The result follows.

Exercises

6.5.1. (a) Suppose that I1 = [a1 , b1 ], I2 = [a2 , b2 ], . . . are intervals in R. Show

that their intersection I1 I2 is another interval (possibly empty).
(b) Suppose that S = S1 Sn , T = T1 Tn , U = U1 Un ,
. . . are cartesian products of sets. Show that their intersection is

S T U = (S1 T1 U1 ) (Sn Tn Un ).

(c) Show that any intersection of boxes in Rn is another box (possibly

empty).
(d) If S is a set and T1 , T2 , T3 , . . . are all sets that contain S, show that
T1 T2 T3 contains S.

6.5.2. Let S be a nonempty bounded subset of Rn , let B be any box con-

taining S, and let B be the intersection of all boxes containing S. By the
preceding problem, B Ris also a box
R containing S. Use Proposition 6.2.4 to
show that if either of B S and B S exist then both exist and they are
equal. It follows, as remarked in the text, that the definition of the volume
of S is independent of the containing box B.

6.5.3. Let B Rn be a box. Show that its volume under Definition 6.5.1
equals its volume under Definition 6.1.4. (Hint: Exercise 6.2.3.)

6.5.4. Let S be the set of rational numbers in [0, 1]. Show that under Defini-
tion 6.5.1, the volume (i.e., length) of S does not exist.

6.5.5. Prove the Volume Zero Criterion.

6.5.6. If S Rn has volume zero and R is a subset of S, show R has volume

zero. (Hint: 0 R S .)
298 6 Integration

6.5.7. Prove that if S1 and S2 have volume zero, then so does S1 S2 . (Hint:
S1 S2 S1 + S2 .)

6.5.8. Find an unbounded set with a nonempty boundary, and a bounded set
with empty boundary.

6.5.9. Review figure 6.18 and its discussion in the section. Also review the ex-
ample that begins after Definition 6.5.5 and continues after Definition 6.5.6.
Similarly, use results from the section such as Theorem 6.5.4 and Proposi-
tion 6.5.3R to explain why for each set K and function f : K R below, the
integral K f exists. Draw a picture each time, taking n = 3 for the picture
in part (f).
(a) K = {(x, y) : 2 y 3, 0 x 1 + ln y/y}, f (x, y) = exy .
2
(b) K = {(x, y) : 1 x 4, 1 y x}, f (x, y) = ex/y /y 5 .
(c) K = the region between the curves y = 2x2 and x = 4y 2 , f (x, y) = 1.
(d) K = {(x, y) : 1 x2 + y 2 2}, f (x, y) = x2 .
(e) K = the pyramid with vertices (0, 0, 0), (3, 0, 0), (0, 3, 0), (0, 0, 3/2),
f (x, y, z) = x.
(f) K = {x Rn : |x| 1} (the solid unit ball in Rn ), f (x1 , . . . , xn ) =
x1 xn .

6.6 Fubinis Theorem

With existence theorems for the integral now in hand, this section and the
next one present tools to compute integrals.
An n-fold iterated integral is n one-dimensional integrals nested inside
each other, such as
Z b1 Z b2 Z bn
f (x1 , x2 , . . . , xn ),
x1 =a1 x2 =a2 xn =an

for some function f : [a1 , b1 ] [an , bn ] R. An iterated integral is

definitely not the same sort of object as an n-dimensional integral. We can
evaluate an iterated integral by working from the inside out. For the innermost
integral, f is to be viewed as a function of the variable xn with its other inputs
treated as constants, and so on outwards. For example,
Z 1 Z 2 Z 1 2 Z 1 1
2 1 3  8 4 2  4
xy = xy  = x= x  = .
x=0 y=0 x=0 3 y=0 x=0 3 3 x=0 3

There are n! different orders in which one can

R 2 iterate
R 1 n integrals, e.g., the ex-
ample just worked is not the same object as y=0 x=0 xy 2 . Regardless of order,
each one-dimensional integral requires varying its particular input to f while
holding the other inputs fixed. The upshot of all this variable-dependence is
 6.6 Fubinis Theorem 299

that there is no reasonable alternative to naming and writing the variables in

an iterated integral.
In an inner integral, outermore variables may figure not only as inputs
to the integrand, but also in the limits of integration. For example, in the
calculation
Z Z x Z x Z

cos(x + y) = sin(x + y) = sin(2x) sin(x) = 2,
x=0 y=0 x=0 y=0 x=0

each inner integral over y is being taken over a segment of x-dependent length
as the outer variable x varies from 0 to . (See figure 6.21.)

y=x

Figure 6.21. Variable Range of Inner Integration

Fubinis Theorem says that under suitable conditions, the n-dimensional

integral is equal to the n-fold iterated integral. The theorem thus provides an
essential calculational tool for multivariable integration.
Theorem 6.6.1 (Fubinis Theorem). Let B = [a, b] [c, d] R2 , and let
f : B R R be bounded, and continuous except on a subset S B of area
zero, so B f exists. Suppose that for each x [a, b], S contains only finitely
many points (possibly none) with first coordinate x. Then the iterated integral
Rb Rd
x=a y=c
f (x, y) also exists, and
Z Z b Z d
f= f (x, y).
B x=a y=c

For notational convenience, the theorem is stated only in two dimensions.

Replacing [a, b] and [c, d] by boxes gives a more general version with a virtually
identical proof. Thinking geometrically in terms of area and volume makes the
theorem plausible in two dimensions since each inner integral is the area of a
cross-section of the volume under the graph of f . (See figure 6.22.)
However, since the multiple integral and the iterated integral are defined
analytically as limits of sums, our only available method for proving the the-
orem is analytic: we must compare approximating sums for the two integrals.
300 6 Integration

y
x

Figure 6.22. Inner integral as cross-sectional area

We now Rdiscuss the ideas before giving the actual proof. A lower sum for the
integral B f is shown geometrically in the left side of figure 6.23. A partition
P Q divides the box B = [a, b] [c, d] into subboxes I J, and the volume
of each solid region in the figure is the area of a subbox times the minimum
Rd
height of the graph over the subbox. By contrast, letting g(x) = y=c f (x, y)
be the area of the cross-section at x, the right side of figure 6.23 shows a lower
Rb
sum for the integral x=a g(x) The partition P divides the interval [a, b] into
subintervals I, and the volume of each bread-slice in the figure is the length
of a subinterval times the minimum area of the cross-sections orthogonal to I.
The proof will show that because integrating in the y-direction is a finer di-
agnostic than summing minimal box-areas in the y-direction, the bread-slices
in the right side of the figure are a superset of the boxes in the left side. Con-
sequently, the volume beneath the bread-slices is at least the volume of the
boxes,
L(f, P Q) L(g, P ).
By similar reasoning for upper sums, in fact we expect that

L(f, P Q) L(g, P ) U (g, P ) U (f, P Q). (6.8)

R
Since L(f, P Q) and U (f, P Q) converge to B f under suitable refinement
of P Q, so do L(g, P ) and U (g, P ). Thus the iterated integral exists and
equals the double integral as desired. The details of turning the geometric
intuition of this paragraph into a proof of Fubinis Theorem work out fine
provided that we carefully tend to matters in just the right order. However,
the need for care is genuine. A subtle point not illustrated by figure 6.23 is
that
although the boxes lie entirely beneath the bread-slices (this is a relation
between two sets),
 6.6 Fubinis Theorem 301

and although the boxes lie entirely beneath the graph (so is this),
and although the volume of the bread-slices is at most the volume beneath
the graph (but this is a relation between two numbers),
the bread-slices need not lie entirely beneath the graph.
Since the bread-slices need not lie entirely beneath
R the graph, the fact that
their volume L(g, P ) estimates the integral B f from below does not follow
from pointwise considerations. The proof finesses this point by establishing
the inequalities (6.8) without reference to the integral, only then bringing the
integral into play as the limit of the extremal sums in (6.8).

y y
x x

Figure 6.23. Geometry of two lower sums

Proof. For each x [a, b], define the cross-sectional function

x : [c, d] R, x (y) = f (x, y).
The hypotheses of Fubinis Theorem ensure that as x varies from a to b, each
cross-sectional function x is continuous except at finitely many points and
hence it is integrable on [c, d]. Give the cross-sectional integral a name,
Z d
g : [a, b] R, g(x) = x .
c
Rb Rd Rb
The iterated integral x=a y=c f (x, y) is precisely the integral g. We need
R a
to show that this exists and equals B f .
302 6 Integration

Consider any partition P Q of B into subboxes J K. Thus P partitions

[a, b] into subintervals J, and Q partitions [c, d] into subintervals K. Take any
subinterval J of P , and take any point x of J. Note that x on each K
samples f only on a cross-section of J K, and so f has more opportunity
to be small on J K than x has on K. That is,

mJK (f ) mK (x ).

The lower sum of the cross-sectional function x over the y-partition Q is a

lower bound for the cross-sectional integral g(x),
X Z d
mK (x ) length(K) = L(x , Q) x = g(x).
K c

The previous two displays combine to give a lower bound for the cross-
sectional integral g(x), the lower bound making reference to the interval J
where x lies but independent of the particular point x of J,
X
mJK (f ) length(K) g(x) for all x J.
K

That is, the left side of this last display is a lower bound of all values g(x)
as x varies through J. So it is at most the greatest lower bound,
X
mJK (f ) length(K) mJ (g).
K

Multiply through by the length of J to get

X
mJK (f ) area(J K) mJ (g) length(J).
K

(This inequality says that each y-directional row of boxes in the left half of
figure 6.23 has at most the volume of the corresponding bread-slice in the
right half of the figure.) As noted at the end of the preceding paragraph, the
iterated integral is the integral of g. The estimate just obtained puts us in
a position to compare lower sums for the double integral and the iterated
integral,
X X
L(f, P Q) = mJK (f ) area(J K) mJ (g) length(J) = L(g, P ).
J,K J

Concatenating a virtually identical argument with upper sums gives the an-
ticipated chain of inequalities,

L(f, P Q) L(g, P ) U (g, P ) U (f, P Q).

R
The outer terms converge to B f under suitable refinement of P Q, and
Rb R
hence so do the inner terms, showing that a g exists and equals B f .

 6.6 Fubinis Theorem 303

Since we will use Fubinis Theorem to evaluate actual examples, all the
notational issues discussed in section 6.4 arise here again. A typical notation
for examples is
Z Z b Z d
f (x, y) = f (x, y),
B x=a y=c

where the left side is a 2-dimensional integral, the right side is an iterated
integral, and f (x, y) is an expression defining f . For example, by Fubinis
Theorem and the calculation at the beginning of this section,
Z Z 1 Z 2
2 4
xy = xy 2 = .
[0,1][0,2] x=0 y=0 3
R Rd Rb
Of course, an analogous theorem asserts that B f (x, y) = y=c x=a f (x, y)
provided that the set S of discontinuity meets horizontal segments at only
finitely many points too. In other words, the double integral also equals the
other iterated
R 2 integral,
R1 and consequently the two iterated integrals agree. For
example, y=0 x=0 xy 2 also works out easily to 4/3.
In many applications, the integral over B is really an integral over a non-
rectangular compact set K, as defined at the end of the previous section. If
K is the area between the graphs of continuous functions 1 , 2 : [a, b] R,
i.e., if
K = {(x, y) : a x b, 1 (x) y 2 (x)},
R b R 2 (x)
then one iterated integral takes the form x=a y= 1 (x)
f (x, y). Similarly, if

K = {(x, y) : c y d, 1 (y) x 2 (y)},

Rd R 2 (y)
then the other iterated integral is y=c x=1 (y)
f (x, y). (See figure 6.24.)

y = 2 (x)

x = 2 (y)

x = 1 (y)
y = 1 (x)

Figure 6.24. Setting up nonrectangular double integrals

304 6 Integration

The interchangeability of the order of integration leads to a fiendish class

of iterated integral problems where one switches order to get a workable inte-
grand. For example, the iterated integral
Z 2 Z 1
2
ex
y=0 x=y/2

2
looks daunting because the integrand ex has no convenient antiderivative,
but after exchanging the order of the integrations and then carrying out a
change of variable, it becomes
Z 1 Z 2x Z 1 Z 1
2 2
ex = 2xex = eu = 1 e1 .
x=0 y=0 x=0 u=0

Interchanging the order of integration can be tricky in such cases; often one
has to break K up into several pieces first, e.g.,
Z 2 Z 2 Z 1 Z 2 Z 2 Z 2
= + .
x=1 y=1/x y=1/2 x=1/y y=1 x=1

A carefully labeled diagram facilitates this process. For example, figure 6.25
shows the sketch that arises from the integral on the left side, and then the
resulting sketch that leads to the sum of two integrals on the right side.

y y
y=2
2
x=1
x=2
1
y = 1/x 1/2 x = 1/y

x x
1 2

Figure 6.25. Sketches for iterated integrals

Interchanging the outer two integrals in a triply iterated integral is no dif-

ferent from the double case, but interchanging the inner two is tricky because
of the constant-but-unknown value taken by the outer variable. Sketching a
generic two-dimensional cross-section usually makes the substitutions clear.
For example, consider the iterated integral
 6.6 Fubinis Theorem 305
Z 1 Z x2 Z x2
. (6.9)
x=0 y=x3 z=y

(The function being integrated is irrelevant to this discussion of how to ex-

change the order of integration, so it is omitted from the notation.) Exchanging
the outer two integrals is carried out via the first diagram in figure 6.26. The
diagram leads to the iterated integral
Z 1 Z
3 y Z x2

.
y=0 x= y z=y

On the other hand, to exchange the inner integrals of (6.9), think of x as fixed
but generic between 0 and 1 and consider the second diagram in figure 6.26.
This diagram shows that (6.9) is also the iterated integral
Z 1 Z x2 Z z
. (6.10)
x=0 z=x3 y=x3

y z

y = x2
z = x2
x= y x2
y = x3

y = x3
x3 z=y
x= 3y y=z
x y
1 x3 x2

Figure 6.26. Sketches for a triply-iterated integral

Switching the outermost and innermost integrals of (6.9) while leaving the
middle one in place requires three successive switches of adjacent integrals.
For instance, switching the inner integrals as we just did and then doing an
outer exchange on (6.10) virtually identical to the outer exchange of a moment
earlier (substitute z for y in the first diagram of figure 6.26) shows that (6.9)
is also Z Z 3 Z
1 z z

.
z=0 x= z y=x3
306 6 Integration

Finally, the first diagram of figure 6.27 shows how to exchange the inner
integrals once more. The result is
Z 1 Z z Z
3 y

.
z=0 y=z 3/2 x= z

The second diagram of figure 6.27 shows the three-dimensional figure that our
iterated integral has traversed in various fashions. It is satisfying to see how
this picture is compatible with the cross-sectional sketches, and to determine
which axis is which. However, the three-dimensional figure is unnecessary
for exchanging the order of integration. The author of these notes finds using
two-dimensional cross-sections easier and more reliable than trying to envision
an entire volume at once. Also, the two-dimensional cross-section technique
will work in an n-fold iterated integral for any n 3, even when the whole
situation is hopelessly beyond visualizing.

z y = x3
3/2
z x= 3y
x
z 3z

Figure 6.27. Another cross-section and the three-dimensional region

The unit simplex in R3 is the set

S = {(x, y, z) : x 0, y 0, z 0, x + y + z 1}

(see figure 6.28). Its centroid is (x, y, z), where

R R R
S
x S
y z
x= , y= , z= S .
vol(S) vol(S) vol(S)

Fubinis Theorem lets us treat the integrals as iterated, giving

 6.6 Fubinis Theorem 307
Z Z 1 Z 1x Z 1xy
x= x
S x=0 y=0 z=0
Z 1 Z 1x
= x(1 x y)
x=0 y=0
Z 1
1
= 1
2 x(1 x)2 = ,
x=0 24

where theRroutine one-variable calculations are not shown in detail. Similarly,

vol(S) = S 1 works out to 1/6, so x = 1/4. By symmetry, y = z = 1/4 also.
See the exercises for an n-dimensional generalization of this result.

Figure 6.28. Unit simplex

To find the volume between the two paraboloids z = 8 x2 y 2 and

z = x2 + 3y 2 , first set 8 x2 y 2 = x2 + 3y 2 to find that the graphs intersect
over the ellipse {(x, y) : (x/2)2 + (y/ 2)2 = 1}. (See figure 6.29.) By Fubinis
Theorem the volume is
Z 2 Z 42y2 Z 8x2 y2
V = 1 = 8 2
y= 2 x= 42y 2 z=x2 +3y 2

where again the one-dimensional calculations are omitted.

Another example is to find the volume of the region K common to the
cylinders x2 + y 2 = 1 and x2 + z 2 = 1. For
each x-value between 1 and 1,
y and z vary independently between 1 x2 and 1 x2 . That is, the
intersection of the two cylinders is a union of squares, whose corners form two
tilted ellipses. (See figure 6.30.) By the methods of this section, the integral
has the same value as the iterated integral, which is
Z Z Z Z
1 1x2 1x2 1
16
1=4 (1 x2 ) = .
x=1 y= 1x2 z= 1x2 x=1 3

Finally, we end the section with a more theoretical example.

308 6 Integration

Figure 6.29. Volume between two graphs

Figure 6.30. Volume common to two cylinders

Proposition 6.6.2 (Differentiation under the Integral Sign). Consider

a function
f : [a, b] [c, d] R.
Suppose that f and D1 f are continuous. Also consider the cross-sectional
integral function,
Z d
g : [a, b] R, g(x) = f (x, y).
y=c

Rd
Then g is differentiable, and g (x) = y=c
D1 f (x, y). That is,
 6.6 Fubinis Theorem 309
Z d Z d
d
f (x, y) = f (x, y).
dx y=c y=c x

Proof. Compute for any x [a, b], using the Fundamental Theorem of Integral
Calculus (Theorem 6.4.2) for the second equality and then Fubinis Theorem
for the fourth,
Z d
g(x) = f (x, y)
y=c
Z d Z x 
= D1 f (t, y) + f (a, y)
y=c t=a
Z d Z x Z d
= D1 f (t, y) + C (where C = f (a, y))
y=c t=a y=c
Z x Z d
= D1 f (t, y) + C.
t=a y=c

It follows from Theorem 6.4.1 that the derivative equals the integrand evalu-
ated at x,
Z d
g (x) = D1 f (x, y),
y=c

as desired.

See exercise 6.6.10 for another example in this spirit.

Exercises

6.6.1. Let S be the set of points (x, y) R2 between the x-axis and the
sine curve as x varies between 0 and 2. Since the sine curve has two arches
between 0 and 2, and since the area of an arch of sine is 2,
Z
1 = 4.
S

On the other hand,

Z 2 Z sin x Z 2
1= sin x = 0.
x=0 y=0 x=0

Why doesnt this contradict Fubinis Theorem?

Rb Rx
6.6.2. Exchange the order of integration in x=a y=a f (x, y).
R1 R 1x R x+y
6.6.3. Exchange the inner order of integration in x=0 y=0 z=0
f.
310 6 Integration
R 1 R 1 R x2 +y2
6.6.4. Exchange the inner order of integration in x=0 y=0 z=0 f . Sketch
the region of integration.
R
6.6.5. Evaluate K f from parts (a), (b), (c), (f) of exercise 6.5.9, except
change K to [0, 1]n for part (f).
6.6.6. Find the volume of the region K bounded by the coordinate planes,
x + y = 1, and z = x2 + y 2 . Sketch K.
R
6.6.7. Evaluate K (1 + x + y + z)3 where K is the unit simplex.
6.6.8. Find the volume of the region K in the first octant bounded by x = 0,
z = 0, z = y, and x = 4 y 2 . Sketch K.
6.6.9. Find the volume of the region K between z = x2 + 9y 2 and z =
18 x2 9y 2 . Sketch K.
6.6.10. Let f : R2 R have continuous mixed second order partial deriva-
tives, i.e., let D12 f and D21 f exist and be continuous. Rederive the familiar
fact that D12 f = D21 f as follows. If D12 f (p, q)D21 f (p, q) > 0 at some point
(p, q) thenR D12 f D21 f > 0 on some rectangle B = [a, b] [c, d] containing
(p, q), so B (D12 f D21 f ) > 0. Obtain a contradiction by evaluating this
integral.
6.6.11. Let K and L be compact subsets of Rn with boundaries of volume
zero. Suppose that for each x1 R, the cross sectional sets

Kx1 = {(x2 , . . . , xn ) : (x1 , x2 , . . . , xn ) K}

Lx1 = {(x2 , . . . , xn ) : (x1 , x2 , . . . , xn ) L}

have equal (n 1)-dimensional volumes. Show that K and L have the same
volume. (Hint: Use Fubinis Theorem to decompose the n-dimensional volume-
integral as the iteration of a 1-dimensional integral of (n 1)-dimensional
integrals.) Illustrate for n = 2.
6.6.12. Let x0 be a positive real, and let f : [0, x0 ] R be continuous. Show
that
Z x0 Z x1 Z xn1 Z x0
1
f (xn ) = (x0 t)n1 f (t).
x1 =0 x2 =0 xn =0 (n 1)! t=0

(Use induction. The base case n = 1 is easy, then the induction hypothesis
applies to the inner (n 1)-fold integral.)
6.6.13. Let n Z+ and r R0 . The n-dimensional simplex of side r is

Sn (r) = {(x1 , . . . , xn ) : 0 x1 , . . . , 0 xn , x1 + + xn r}.

(a) Make sketches for n = 2 and n = 3 showing that Sn (r) is a disjoint

union of cross-sectional (n 1)-dimensional simplices of side r xn at height
 6.7 Change of Variable 311

xn as xn varies from 0 to r. Explain this symbolically for general n > 1. That

is, explain why G
Sn (r) = Sn1 (r xn ) {xn }.
xn [0,r]

(b) Prove that vol(S1 (r)) = r. Use part (a) and Fubinis Theorem (cf. the
hint to exercise 6.6.11) to prove that
Z r
vol(Sn (r)) = vol(Sn1 (r xn )) for n > 1,
xn =0

and show by induction that vol(Sn (r)) = rnZ/n!. Z r

(r xn )n1
(c) Use Fubinis Theorem to show that xn = xn .
Sn (r) xn =0 (n 1)!
R
Work this integral by substitution or by parts to get Sn (r) xn = rn+1 /(n+1)!.
R 
(d) The centroid of Sn (r) is (x1 , . . . , xn ), where xj = Sn (r) xj vol(Sn (r))
for each j. What are these coordinates explicitly? (Make sure your answer
agrees with the case in the text.)

6.7 Change of Variable

Any point p R2 with cartesian coordinates (x, y) is also specified by its
polar coordinates (r, ), where r is the distance from the origin to p and
is the angle from the positive x-axis to p. (See figure 6.31.)

y p
r

x

Figure 6.31. Polar coordinates

The angle is defined only up to multiples of 2, and it isnt defined at

all when p = (0, 0). Trigonometry expresses (x, y) in terms of (r, ),

x = r cos , y = r sin . (6.11)

But expressing (r, ) in terms of (x, y) is a little more subtle. Certainly

312 6 Integration
p
r= x2 + y 2 .

Also, tan = y/x provided that x 6= 0, but this doesnt mean that =
arctan(y/x). Indeed, arctan isnt even a well-defined function until its range
is specified, e.g., as (/2, /2). With this particular restriction, the actual
formula for , even given that not both x and y are 0, is not arctan(y/x), but

if x > 0 and y 0 (this lies in [0, /2)),
arctan(y/x)




/2 if x = 0 and y > 0,
= arctan(y/x) + if x < 0 (this lies in (/2, 3/2)),



3/2 if x = 0 and y < 0,


arctan(y/x) + 2 if x > 0 and y < 0 (this lies in (3/2, 2)).

The formula is unwieldy, to say the least. (The author probably would not
read through the whole thing if he were instead a reader. In any case, see
figure 6.32.) A better approach is that given (x, y), the polar radius r is the
unique nonnegative number such that

r 2 = x2 + y 2 ,

and then, if r 6= 0, the polar angle is the unique number in [0, 2) such
that (6.11) holds. But still, going from polar coordinates (r, ) to cartesian
coordinates (x, y) as in (6.11) is considerably more convenient than conversely.
This is good since, as we will see, doing so is also more natural.

3/2

/2

Figure 6.32. The angle between 0 and 2

The change of variable mapping from polar to cartesian coordinates is

: R0 [0, 2] R2 , (r, ) = (r cos , r sin ).

 6.7 Change of Variable 313

The mapping is injective except that the half-lines R0 {0} and R0 {2}
both map to the nonnegative x-axis, and the vertical segment {0} [0, 2] is
squashed to the point (0, 0). Each horizontal half-line R0 {} maps to the
ray of angle with the positive x-axis, and each vertical segment {r} [0, 2]
maps to the circle of radius r. (See figure 6.33.)

y

2

x
r

Figure 6.33. The polar coordinate mapping

It follows that regions in the (x, y)-plane defined by radial or angular con-
straints are images under of (r, )-regions defined by rectangular constraints.
For example, the cartesian disk
Db = {(x, y) : x2 + y 2 b2 }
is the -image of the polar rectangle
Rb = {(r, ) : 0 r b, 0 2}.
(See figure 6.34.) Similarly the cartesian annulus and quarter disk,
Aa,b = {(x, y) : a2 x2 + y 2 b2 },
Qb = {(x, y) : x 0, y 0, x2 + y 2 b2 },
are the images of rectangles. (See figures 6.35 and 6.36.)
Iterated integrals over rectangles are especially convenient to evaluate, be-
cause the limits of integration for the two one-variable integrals are constants
rather than variables that interact. For example,
Z b Z 2 Z 2 Z b
= .
r=a =0 =0 r=a

These tidy (r, ) limits describe the (x, y) annulus Aa,b indirectly via , while
the more direct approach of an (x, y)-iterated integral over Aa,b requires four
messy pieces,
Z a Z b2 x2 Z a "Z a2 x2 Z b2 x2 # Z b Z b2 x2
+ + + .
x=b y= b2 x2 x=a y= b2 x2 y= a2 x2 x=a y= b2 x2
314 6 Integration

y

2

x
b
r
b

Figure 6.34. Rectangle to disk under the polar coordinate mapping

y

2

x
a b
r
a b

Figure 6.35. Rectangle to annulus under the polar coordinate mapping

x
/2 b
r
b

Figure 6.36. Rectangle to quarter disk under the polar coordinate mapping

Thus, since Fubinis Theorem equates integrals over two-dimensional regions

to twofold iterated integrals, it would be a real convenience to reduce integrat-
ing over the (x, y)-annulus to integrating over the (r, ) rectangle that maps
to it under . The Change of Variable Theorem will do so. This is the sense
in which it is natural to map from polar to cartesian coordinates rather than
in the other direction.
The Change of Variable Theorem says in some generality how to transform
an integral from one coordinate system to another. Recall that given a set
A Rn and a differentiable mapping : A Rn , the n-by-n matrix of
partial derivatives of is denoted ,
 6.7 Change of Variable 315

= [Dj i ]i,j=1...n .

A differentiable mapping whose partial derivatives are all continuous is called

a C 1 -mapping. Also, for any set K Rn , an interior point of K is a point
of K that is not a boundary point, and the interior of K is the set of all such
points,
K = {interior points of K}.
We will discuss boundary points and interior points more carefully in the next
section. In the specific sorts of examples that arise in calculus, they are easy
enough to recognize.

Theorem 6.7.1 (Change of Variable Theorem for Multiple Inte-

grals). Let K Rn be a compact and connected set having boundary of
volume zero. Let A be an open superset of K, and let

: A Rn

be a C 1 -mapping such that

is injective on K and det 6= 0 on K .

Let
f : (K) R
be a continuous function. Then
Z Z
f= (f ) | det |.
(K) K

This section will end with a heuristic argument to support Theorem 6.7.1,
and then section 6.9 will prove the theorem after some preliminaries in sec-
tion 6.8. In particular, section 6.8 will explain why the left side integral in
the theorem exists. (The right side integral exists because the integrand is
continuous on K, which is compact and has boundary of volume zero, but the
fact that (K) is nice enough for the left side integral to exist requires some
discussion.) From now to the end of this section, the focus is on how the the-
orem is used. Generally, the idea is to carry out substitutions of the sort that
were called inverse substitutions in the one-variable
R discussion of section 6.4.
That is, to apply the theorem to an integral D fR, find a suitable set K and
mapping such that D = (K) and the integral K (f ) | det | is easier
to evaluate instead. The new integral most likely will be easier because K has
a nicer shape than D (this wasnt an issue in the one-variable case), but also
possibly because the new integrand is more convenient.
For example, to integrate the function f (x, y) = x2 + y 2 over the annulus
Aa,b , recall the polar coordinate mapping (r, ) = (r cos , r sin ), and recall
that under this mapping, the annulus is the image of a box,
316 6 Integration

Aa,b = ([a, b] [0, 2]).

The composition of the integrand with is

(f )(r, ) = r2 ,

and the polar coordinate has derivative matrix

 
cos r sin
= ,
sin r cos

with absolute determinant

| det | = r.
So by the Change of Variable Theorem, the desired integral is instead an
integral over a box in polar coordinate space,
Z Z
f= r2 r
Aa,b [a,b][0,2]

By Fubinis Theorem, the latter integral can be evaluated as an iterated in-

tegral,
Z Z 2 Z b
3
r = r3 = (b4 a4 ).
[a,b][0,2] =0 r=a 2
Similarly, the quarter disk Qb = ([0, b] [0, /2]) has centroid (x, y) where
R R /2 R b
Q
x r cos r b3 /3 4b
x= = =0 r=02 = 2 = ,
area(Q) b /4 b /4 3
and y takes the same value by symmetry. Indeed 4/(3) is somewhat less
than 1/2, in conformance with our physical intuition of the centroid of a
region as its balancing point.
The sharp-eyed reader has noticed that a subtle aspect of Theorem 6.7.1
was in play for this example. Although the polar change of coordinate map-
ping (r, ) is defined on all of R2 , it fails to be injective on all of the box
K = [a, b] [0, 2]: the 2-periodic behavior of as a function of maps the
top and bottom edges of the box to the same segment of the x-axis. Further-
more, if the annulus has inner radius a = 0, i.e., if the annulus is a disk, then
not only collapses the left edge of the box to the origin in the (x, y)-plane,
but also det = 0 on the left edge of the box. Thus we need the theorems
hypotheses that need be injective only on the interior of K, and that the
condition det 6= 0 need hold only on the interior of K.
Just as polar coordinates are convenient for radial symmetry in R2 , cylin-
drical coordinates in R3 conveniently describe regions with symmetry about
the z-axis. A point p R3 with cartesian coordinates (x, y, z) has cylindrical
coordinates (r, , z) where (r, ) are the polar coordinates for the point (x, y).
(See figure 6.37.)
 6.7 Change of Variable 317

z z
y

x
r

Figure 6.37. Cylindrical coordinates

The cylindrical change of variable mapping is thus

: R0 [0, 2] R R3

given by
(r, , z) = (r cos , r sin , z).
That is, is just the polar coordinate mapping on z cross-sections, so like the
polar map, it is mostly injective. Its derivative matrix is

cos r sin 0
= sin r cos 0 ,

0 0 1

and again
| det | = r.
So, for example, to integrate f (x, y, z) = y 2 z over the cylinder C : x2 +y 2 1,
0 z 2, note that C = ([0, 1][0, 2][0, 2]), and therefore by the Change
of Variable Theorem and then Fubinis Theorem,
Z Z 2 Z 1 Z 2 Z 2 1 2
2 2 2 r4  z 2 
f= r sin z r = sin   = .
C =0 r=0 z=0 =0 4 r=0 2 z=0 2

From now on, Fubinis Theorem no longer necessarilyR pwarrants comment.

For another example, we evaluate the integral S x2 + y 2 where S is the
region bounded by z 2 = x2 + y 2 , z = 0, and z = 1. (This region looks like an
ice cream cone with the ice cream licked down flat.) The Change of Variable
Theorem transforms the integral into (r, , z)-coordinates,
318 6 Integration
Z p Z 1 Z 2 Z 1

x2 + y 2 = r2 1= .
S r=0 =0 z=r 6

Spherical coordinates in R3 are designed to exploit symmetry about

the origin. A point p = (x, y, z) R3 has spherical coordinates (, , ) where
the spherical radius is the distance from the origin to p, the longitude
is the angle from the positive x-axis to the (x, y)-projection of p, and the
colatitude is the angle from the positive z-axis to p. By some geometry, the
spherical coordinate mapping is

: R0 [0, 2] [0, ] R3

given by
(, , ) = ( cos sin , sin sin , cos ).
The spherical coordinate mapping has derivative matrix

cos sin sin sin cos cos
= sin sin cos sin sin cos ,
cos 0 sin

with determinant (using column-linearity)

cos sin sin cos cos
det = 2 sin det sin sin cos sin cos
cos 0 sin
2 2
!
2
cos sin + sin cos2
2
= sin
+ cos2 cos2 + sin2 sin2
= 2 sin ,

so that, since 0 ,

| det | = 2 sin .

That is, the spherical coordinate mapping reverses orientation. It can be re-
defined to preserve orientation by changing to the latitude angle, varying
from /2 to /2, rather than the colatitude.
Figure 6.38 shows the image under the spherical coordinate mapping of
some (, )-rectangles, each having a fixed value of , and similarly for fig-
ure 6.39 for some fixed values of , and figure 6.40 for some fixed values of .
Thus the spherical coordinate mapping takes boxes to regions with these sorts
of walls, such as the half ice cream cone with a bite taken out of its bottom
in figure 6.41.
For an example of the Change of Variable Theorem using spherical coor-
dinates, the solid ball of radius r in R3 is
 6.7 Change of Variable 319

z
y

Figure 6.38. Spherical coordinates for some fixed spherical radii

z
y

Figure 6.39. Spherical coordinates for some fixed longitudes

B3 (r) = ([0, r] [0, 2] [0, ]),

and therefore its volume is
Z Z 2 Z r Z
1 4
vol(B3 (r)) = 1= 2 sin = 2 r3 2 = r3 .
B3 (r) =0 =0 =0 3 3

It follows that the cylindrical shell B3 (b) B3 (a) has volume 4(b3 a3 )/3.
See exercises 6.7.12 through 6.7.14 for the lovely formula giving the volume
of the n-ball for arbitrary n.
The Change of Variable Theorem and spherical coordinates work together
to integrate over the solid ellipsoid of (positive) axes a, b, c,
320 6 Integration

z
y

Figure 6.40. Spherical coordinates for some fixed colatitudes

z
y

Figure 6.41. The spherical coordinate mapping on a box

Ea,b,c = {(x, y, z) : (x/a)2 + (y/b)2 + (z/c)2 1}.

For example, to compute the integral

Z
(Ax2 + By 2 + Cz 2 ),
Ea,b,c

first define a change of variable mapping that stretches the unit sphere into
the ellipsoid,

: B3 (1) Ea,b,c , (u, v, w) = (au, bv, cw).

 6.7 Change of Variable 321

Thus
a00
= 0 b 0 , | det | = abc.
00c
Let f (x, y, z) = Cz 2 . Then since Ea,b,c = (B3 (1)) and (f )(u, v, w) =
Cc2 w2 , part of the integral is
Z Z Z
f= (f ) | det | = abc3 C w2 .
(B3 (1)) B3 (1) B3 (1)

Apply the Change of Variable Theorem again, this time using the spherical
coordinate mapping into (u, v, w)-space,
Z Z 1 Z 2 Z
3 2 3 4
abc C w = abc C 2 cos2 2 sin = abc3 C.
B3 (1) =0 =0 =0 15

By the symmetry of the symbols in the original integral, its overall value is
therefore
Z
4
(Ax2 + By 2 + Cz 2 ) = abc(a2 A + b2 B + c2 C).
Ea,b,c 15

Another example is to find the centroid of upper hemispherical shell

S = (B3 (b) B3 (a)) {z 0}.

By symmetry, x = y = 0. As for z, compute using spherical coordinates that

Z Z b Z 2 Z /2
4
z= cos 2 sin = (b a4 ).
S =a =0 =0 4

This integral needs to be divided by the volume 2(b3 a3 )/3 of S to give

3(b4 a4 )
z= .
8(b3 a3 )

In particular, the centroid of the solid hemisphere is 3/8 of the way up. It
is perhaps surprising that does not figure in this formula, as it did in the
two-dimensional case.
Here is a heuristic argument to support the Change of Variable Theorem.
Suppose that K is a box. Recall the assertion: under certain conditions,
Z Z
f= (f ) | det |.
(K) K

Take a partition P dividing K into subboxes J, and in each subbox choose a

point xJ . If the partition is fine enough, then each J maps under to a small
322 6 Integration

patch A of volume vol(A) | det (xJ )|vol(J) (cf. section 3.8), and each xJ
maps to a point yA A. (See figure 6.42.) Since the integral is a limit of
weighted sums,
Z X
f f (yA )vol(A)
(K) A
X
f ((xJ ))| det (xJ )|vol(J)
ZJ
(f ) | det |,
K

and these should become equalities in the limit as P becomes finer. What
makes this reasoning incomplete is that the patches A are not boxes, as are
required for our theory of integration.

J A

Figure 6.42. Change of variable

Recall from sections 3.8 and 3.9 that the absolute value of det (x) de-
scribes how the mapping scales volume at x, while the sign of det (x)
says whether the mapping locally preserves or reverses orientation. The fac-
tor | det | in the n-dimensional Change of Variable Theorem (rather than
the signed det ) reflects the fact that n-dimensional integration does not
take orientation into account. This unsigned result is less satisfying than the
one-variable theory, which does consider orientation and therefore comes with
R (b) Rb
a signed change of variable theorem, (a) f = a (f ) . An orientation-
sensitive n-dimensional integration theory will be developed in chapter 9.

Exercises
R
6.7.1. Evaluate S x2 + y 2 where S is the region bounded by x2 + y 2 = 2z
and z = 2. Sketch S.
6.7.2. Find the volume of the region S above x2 + y 2 = 4z and below x2 +
y 2 + z 2 = 5. Sketch S.
 6.7 Change of Variable 323

6.7.3. Find the volume of the region between the graphs of z = x2 + y 2 and
z = (x2 + y 2 + 1)/2.

6.7.4. Derive the spherical coordinate mapping.

6.7.5. Let be the spherical coordinate mapping. Describe (K) where

K = {(, , ) : 0 2, 0 /2, 0 cos }.

(Hint: Along with visualizing the geometry, set = 0 and consider the condi-
tion 2 = cos in cartesian coordinates.) Same question for

K = {(, , ) : 0 2, 0 , 0 sin }.
R
6.7.6. Evaluate S xyz where S is the first octant of B3 (1).

6.7.7. Find the mass of a solid figure filling the spherical shell

S = B3 (b) B3 (a)

with density (x, y, z) = x2 + y 2 + z 2 .

2
+y 2 +z 2 )3/2
6.7.8. A solid
R sphere of radius b has density (x, y, z) = e(x . Find
its mass, B3 (b) .

6.7.9. Find the centroid of the region S = B3 (a) {x2 + y 2 z 2 } {z 0}.

Sketch S.

6.7.10. (a) Prove Pappuss Theorem: Let K be a compact set in the (x, z)-
plane lying to the right of the z-axis and with boundary of area zero. Let S
be the solid obtained by rotating K about the z-axis in R3 . Then

vol(S) = 2x area(K),
R
where as always, x = K x/area(K). (Use cylindrical coordinates.)
(b) What is the volume of the torus Ta,b of cross-sectional radius a and
major radius b from the center of rotation to the center of the cross-sectional
disk? (See figure 6.43.)

6.7.11. Prove the change of scale principle: If the set K Rn has volume
v then for any r 0, the set rK = {rx : x K} has volume rn v. (Change
variables by (x) = rx.)

6.7.12. (Volume of the n-ball, first version.) Let n Z+ and r R0 . The

n-dimensional ball of radius r is

Bn (r) = {x : x Rn |x| r} = {(x1 , . . . , xn ) : x21 + + x2n r2 }.

Let
324 6 Integration

Figure 6.43. Torus

vn = vol(Bn (1)).
(a) Explain how exercise 6.7.11 reduces computing the volume of Bn (r) to
computing vn .
(b) Explain why v1 = 2 and v2 = .
(c) Let D denote the unit disk B2 (1). Explain why for n > 2,
G q
Bn (1) = {(x1 , x2 )} Bn2 ( 1 x21 x22 ).
(x1 ,x2 )D

p unit n-ball is a union of cross-sectional (n 2)-dimensional balls

That is, the
of radius 1 x21 x22 as (x1 , x2 ) varies through the unit disk. Make a sketch
for n = 3, the only value of n for which we can see this.
(d) Explain why for n > 2,
Z
n
vn = vn2 (1 x21 x22 ) 2 1
(x1 ,x2 )D
Z 2 Z 1
n
= vn2 (1 r2 ) 2 1 r
=0 r=0
= vn2 /(n/2).

(Use the definition of volume at the end of section 6.5, Fubinis Theorem, the
definition of volume again, the change of scale principle from the previous
exercise, and the Change of Variable Theorem.)
(e) Show by induction only the for n even case of the formula


n/2
for n even,
vn = (n/2)!
(n1)/2 n
2 ((n 1)/2)!

for n odd.
n!
(The for n odd case can be shown by induction as well, but the next two
exercises provide a better, more conceptual approach to the volumes of odd-
dimensional balls.)
 6.7 Change of Variable 325
R 2
6.7.13. This exercise computes the improper integral I = x=0 ex , defined
RR 2 RR 2
as the limit limR x=0 ex . Let I(R) = x=0 ex for any R 0.
R 2 2
(a) Use Fubinis Theorem to show that I(R)2 = S(R) ex y , where S(R)
is the square
S(R) = {(x, y) : 0 x R, 0 y R}.
(b) Let Q(R) be the quarter disk

Q(R) = {(x, y) : 0 x, 0 y, x2 + y 2 R2 },

and similarly for Q( 2 R). Explain why
Z Z Z
2 2 2 2 2 2
ex y ex y ex y .
Q(R) S(R) Q( 2 R)
R 2 2 R 2 2
(c) Change variables, and evaluate Q(R) ex y and Q(2 R) ex y .
What are the limits of these two quantities as R ?
(d) What is I?

6.7.14. (Volume of the n-ball, improved version) Define the gamma function
as an integral, Z
(s) = xs1 ex dx, s > 0.
x=0
(This improper integral is well-behaved, even though it is not being carried
out over a bounded region and even though the integrand is unbounded near
x = 0 when 0 < s < 1. We use dx here because this exercise is computational.)
(a) Show: (1) = 1, (1/2) = , (s + 1) = s (s). (Substitute and see
the previous exercise for the second identity, integrate by parts for the third.)
(b) Use part (a) to show that n! = (n+1) for n = 0, 1, 2, . Accordingly,
define x! = (x+1) for all real numbers x > 1, not only nonnegative integers.
(c) Use exercise 6.7.12(b), exercise 6.7.12(d), and the extended definition
of the factorial in part (b) of this exercise to obtain a uniform formula for the
volume of the unit n-ball,

n/2
vn = , n = 1, 2, 3, .
(n/2)!

(We already have this formula for n even. For n odd, the argument is essen-
tially identical to exercise 6.7.12(e) but starting at the base case n = 1.) Thus
the n-ball of radius r has volume
n/2 n
vol(Bn (r)) = r , n = 1, 2, 3, .
(n/2)!

(d) The Legendre duplication formula for the gamma function is

(2s) = 22s1 1/2 (s) (s + 1/2).

326 6 Integration

For odd n, what value of s shows that the values of vn from part (c) of this
exercise and from part (e) of exercise 6.7.12 are equal?
(e) (Read-only. While the calculation of vn in these exercises shows the
effectiveness of our integration toolkit, the following heuristic argument illus-
trates that we would profit from an even more effective theory of integration.)
Decompose Euclidean space Rn into concentric n-spheres (the n-sphere is
the boundary of the n-ball), each having a radius r and a differential radial
thickness dr. Since each such n-sphere is obtained by removing the n-ball of
radius r from the n-ball of radius r + dr, its differential volume is

vn (r + dr)n vn rn vn nrn1 dr.

Here we ignore the higher powers of dr on the grounds that they are so much
smaller than the dr-term. Thus, re-using some ideas from a moment ago, and
using informal notation,
Z n
2
n/2 = ex dx since the integral equals
Z R
2
= e|x| dV by Fubinis Theorem
Rn
Z
2
= vn n rn1 er dr integrating over spherical shells
r=0
Z
= vn n/2 tn/21 et dt substituting t = r2
t=0
= vn n/2 (n/2)
= vn (n/2)!.

The formula vn = n/2 /(n/2)! follows immediately. The reason that this
induction-free argument lies outside our theoretical framework is that it in-
tegrates directly (rather than by the Change of Variable Theorem) even as it
decomposes Rn into small pieces that arent boxes. Although we would prefer
a more flexible theory of integration that allows such procedures, developing
it takes correspondingly more time.

6.7.15. This exercise evaluates the improper integral

Z
dx
Is = (for any real number s > 1/2).
x= (1 + x 2 )s

(a) For any > 0, make a substitution in the integral

Z
dt
ts et
t=0 t

to show that it equals (s)s .

(b) Explain how a particular choice of in (a) leads to
 6.7 Change of Variable 327
Z Z
1 2
)t dt
Is = ts e(1+x dx.
(s) x= t=0 t

(c) Explain how after exchanging the order of integration, a few other steps
lead to Z Z
1 s1/2 s 2 dt
Is = t e ex dx .
(s) t=0 x= t
(d) Use earlier exercises to conclude that

(s 1/2)
Is = .
(s)
Can you check this formula for s = 1?

6.7.16. Let A and B be positive real numbers. This exercise evaluates the
improper integral
Z
dx
Is = (for any real number s > 0).
x= (Ae
2x + Be2x )s

(a) Recall from exercise 6.7.15(a) that for any > 0,

Z
dt
ts et = (s)s .
t=0 t
Explain how a particular choice of leads to
Z Z
1 2x 2x dt
Is = ts e(Ae +Be )t dx.
(s) x= t=0 t
1
(b) Let x = 2log u (natural logarithm) and show that
Z Z
1 1 dt du
Is = ts e(Au+Bu )t .
2 (s) u=0 t=0 t u
Replace t by ut to get
Z Z
1 2 dt du
Is = ts us e(Au +B)t
.
2 (s) u=0 t=0 t u

Replace u by u to get
Z Z
1 dt du
Is = ts us/2 e(Au+B)t .
4 (s) u=0 t=0 t u

(c) Exchange the order of integration and replace u by u/t to get

Z Z
1 du dt
Is = ts/2 us/2 e(Au+Bt) .
4 (s) t=0 u=0 u t
328 6 Integration

Replace u by u/A and t by t/B to get

Z Z
s/2 s/2 1 s/2 t dt du
Is = A B t e us/2 eu .
4 (s) t=0 t u=0 u

Thus the integral is

Z
dx (s/2) (s/2) s/2 s/2
= A B , s > 0.
x= (Ae2x + Be2x )s 4 (s)

6.7.17. (Read-only. This exercise makes use not only of the gamma function
but of some results beyond our scope, in the hope of interesting the reader in
those ideas.)
(a) Consider any x R>0 , R, and s R>1 . We show that
Z (
2 x s1
eiy (s) e if > 0,
s
dy =
y= (x + iy) 0 if 0.

Indeed, replacing by x in the gamma function integral gives a variant

expression of gamma that incorporates x,
Z Z
s d d
(s) = e =x s
ex s .
=0 =0

A result from complex analysis says that this formula extends from the open
half-line of positive x-values to the open half-plane of complex numbers x + iy
with x positive. That is, for any y R,
Z
d
(s) = (x + iy)s e(x+iy) s .
=0

This is
Z (

(s) iy ex s1 if > 0,
= e x () d where x () =
(x + iy)s =0 0 if 0.

The integral here is a Fourier transform. That is, letting F denote the Fourier
transform operator, the previous display says

(s)
= (Fx )(y), y R.
(x + iy)s
R
The integral (s) y= eiy (x + iy)s dy is consequently the inverse Fourier
transform at of the Fourier transform of x . Fourier inversion says that
the inverse Fourier transform of the Fourier transform is the original function
multiplied by 2. Putting all of this together gives the value of the integral at
the beginning of the exercise.
 6.7 Change of Variable 329

(b) We introduce an n-dimensional gamma function for any positive inte-

ger n. Let

Cn = {n-by-n symmetric positive definite matrices}.

The set Cn is so denoted because it forms a structure called a cone: it is

closed under addition and under dilation by positive real numbers. For n > 1,
exercise 4.7.11 gives a decomposition
 
a cT
Rn1 R>0 Cn1 Cn , c a 2 .
c a1 ccT + 2
The nth gamma function is
Z
d
n (s) = etr (det )s ,
Cn (det )(n+1)/2
Q
in which d = ij dij is the product of the differentials of the diagonal and
superdiagonal elements of , remembering that because is symmetric the
subdiagonal entries are redundant. The decomposition of Cn combines with
some other facts (which the reader is encouraged to identify, if not to prove)
to show that
1 2

Z Z Z ea |c| atr 2 as (det 2 )s

n (s) = d2 da dc .
cRn1 aR>0 2 Cn1 (n+1)/2 (n+1)/2
a (det 2 )

Replacing c by a1/2 c (and thus dc by a(n1)/2 dc) lets the integral be separated,
Z Z
2 da
n (s) = e|c| dc ea as
cR n1 aR>0 a
Z
d2
etr 2 (det 2 )s1/2
2 Cn1 (det 2 )n/2
= (n1)/2 (s)n1 (s 21 ).

And iterating the argument gives the value of the nth gamma function in
terms of the basic gamma function,

n (s) = (n1)n/4 (s) (s 21 ) (s 22 ) (s n2

2 ) (s n1
2 ).

Similarly to part (a), one now can evaluate an integral over the vector space Vn
of n-by-n symmetric matrices for a given Vn ,
Z (
(2)n (n1)n/2 tr(x)
ei tr(y) n (s) e (det )s(n+1)/2 if Cn ,
s
dy =
yVn det(x + iy) 0 else.

using the fact that the constant for Fourier inversion over the space of n-by-n
symmetric matrices is (2)n (n1)n/2 .
330 6 Integration

6.7.18. Figure 6.44 shows a geodesic dome with 5-fold vertices and 6-fold
vertices. (A geodesic of the sphere is a great circle.) Figure 6.45 shows a
birds-eye view of the dome. The thinner edges emanate from the 5-vertices,
while four of the six edges emanating from each 6-vertex are thicker. The five
triangles that meet at a 5-vertex are isoceles, while two of the six triangles
that meet at a 6-vertex are equilateral. This exercise uses vector algebra and
the spherical coordinate system to work out the lengths and angles of the
dome. Integration and the Change of Variable Theorem play no role in this
exercise.
(a) Take all vertices to lie on a sphere of radius 1. The ten thick edges
around the equator form a regular 10-gon. Show that consequently the thick
edges have length
a = 2 sin(/10) = 2 cos(2/5).
This famous number from geometry goes back to Euclid. Note that a = 5 +
51 where 5 = e2i/5 = cos(2/5) + i sin(2/5) is the fifth root of unity one
fifth of the way clockwise around the complex unit circle. Thus a2 + a 1 =
52 + 2 + 53 + 5 + 54 1, and the right side is 0 by the finite geometric sum
formula. That is,
a2 + a 1 = 0, a > 0,
and so the length of the thick edges is

1 + 5
a= = 0.618033999 .
2
This number is a variant of the so-called Golden Ratio.
(b) The dome has a point at the north pole (0, 0, 1), then a layer of five
points p0 through p4 around the north pole at some colatitude , then a
layer of ten points q0 through q9 , five at colatitude 2 and the other five at
some second colatitude , and finally the layer of ten equatorial points r0
through r9 . The colatitude must be such that the triangle with vertices n,
q0 , and q2 is equilateral. These vertices may be taken to be

n = (0, 0, 1),
q0 , q2 = (cos(/5) sin(2), sin(/5) sin(2), cos(2)).

Show that the equilateral condition |n q0 |2 = |q2 q0 |2 gives the condition

cos2 () = 1/(2 a), then sin2 () = (1 a)/(2 a), then tan2 () = a2 , so
that the colatitude of the five points about the north pole is

= arctan(a) = 31.7175 .

Use the cross-sectional triangle having vertices 0, n, p0 and the law of cosines
to show that the shorter segments have length
q
b = 2(1 1/ 2 a) = 0.546533058 .
 6.8 Topological Preliminaries for the Change of Variable Theorem 331

(Alternatively, one can find and b by using the triangle with vertices n, p0 ,
p1 .) For reference in part (e), show that
2 = arctan(2) = 63.4350 .
(c) Show that the angle of an isoceles triangle where its equal sides meet
at a 5-vertex is
= 2 arcsin(a/(2b)) = 68.862 ,
and the angles where its unequal sides meet at 6-vertices are
= arccos(a/(2b)) = 55.569 .
(d) Show that the angle where two a-segments meet along a geodesic
is 180 36 . Show that the angle where two b-segments meet along a geodesic
(this happens at the 6-vertices but not at the 5-vertices) is 180 .
(e) To find the colatitude of q1 , q3 , , q9 , take q9 and q1 to be
q9 , q1 = (cos(/5) sin(2), sin(/5) sin(2), cos(2)),
and consider the geodesic containing them. Their cross product is normal to
the plane of the geodesic. Show that this cross product is
q9 q1 = 2 sin(/5) sin(2)( cos(2), 0, cos(/5) sin(2)).
Show by illustration that the latitude
of this cross product is the colatitude
of q9 and q1 . Show that = arctan( 2 + a). Show further that (a+1)2 = 2+a,
so that in fact
= arctan(a + 1) = 58.2825 .

6.8 Topological Preliminaries for the Change of Variable

Theorem
This section establishes some topological results to prepare for proving the
Change of Variable Theorem (Theorem 6.7.1), and then the next section gives
the proof. Both sections are technical, and so the reader is invited to skim
as feels appropriate. For instance, one might focus on the discussion, the
statements, and the figures, but go light on the proofs.
In preparation for proving the Change of Variable Theorem, we review its
statement. The statement includes the terms boundary and interior, which
we have considered only informally so far, but we soon will discuss them
more carefully. The statement also includes the term open, and the reader
is reminded that a set is called open if its complement is closed; we soon
will review the definition of a closed set. The statement includes the term
C 1 -mapping, meaning a mapping such that all partial derivatives of all of its
component functions exist and are continuous. And the statement includes
the notation K for the interior of a set K. The theorem says:
332 6 Integration

Figure 6.44. Geodesic dome

Let K Rn be a compact and connected set having boundary of volume

zero. Let A be an open superset of K, and let

: A Rn

be a C 1 -mapping such that

is injective on K and det 6= 0 on K .

Let
f : (K) R
be a continuous function. Then
Z Z
f= (f ) | det |.
(K) K

Thus the obvious data for the theorem are K, , and f . (The description of
subsumes A, and in any case the role of A is auxiliary.) But also, although the
dimension n is conceptually generic-but-fixed, in fact the proof of the theorem
will entail induction on n, so that we should view n as a variable part of the
setup as well. Here are some comments about the data.
 6.8 Topological Preliminaries for the Change of Variable Theorem 333

Figure 6.45. Geodesic dome, birds-eye view

The continuous image of a compact set is compact (Theorem 2.4.14), so

that (K) is again compact. Similarly, by an invocation in section 2.4, the
continuous image of a connected set is connected, so that (K) is again
connected. The reader who wants to minimize invocation may instead as-
sume that that K is path-connected, so that (K) is again path-connected
(see exercise 2.4.10 for the definition of path-connectedness and the fact
that path-connectedness is a topological property); the distinction between
connectedness and path-connectedness is immaterial for any example that
will arise in calculus. We soon will see that also the image (K) again has
boundary of volume zero, so that in fact (K) inherits all of the assumed
properties of K.
Thus both integrals in the Change of Variable Theorem exist, because in
each case the integrand is continuous on the domain of integration and the
domain of integration is compact and has boundary of volume zero.
The hypotheses of the theorem can be weakened or strengthened in vari-
ous ways with no effect on the outcome. Indeed, the proof of the theorem
proceeds partly by strengthening the hypotheses. The hypotheses in The-
orem 6.7.1 were chosen to make the theorem fit the applications that arise
in calculus. Especially, parametrizations by polar, cylindrical, or spheri-
cal coordinates often degenerate on the boundary of the parameter-box,
hence the conditions that is injective and det 6= 0 being required
only on the interior K . (See figure 6.46. In the figure, the polar coor-
dinate mapping collapses the left side of the parametrizing rectangle to
the origin in the parametrized disk, and it takes the top and bottom sides
of the parametrizing rectangle to the same portion of the x-axis in the
parametrized disk. Furthermore, neither the origin nor the portion of the
334 6 Integration

x-axis is on the boundary of the parametrized disk even though they both
come from the boundary of the parametrizing rectangle. On the other
hand, every nonboundary point of the parametrizing rectangle is taken
to a nonboundary point of the parametrized disk, so that every bound-
ary point of the parametrized disk comes from a boundary point of the
parametrizing rectangle.)
While the hypotheses about are weaker than necessary in order to make
the theorem easier to use, the hypothesis that f is continuous is stronger
than necessary in order to make the theorem easier to prove. The theorem
continues to hold if f is assumed only to be integrable, but then the proof
requires more work. In calculus examples, f is virtually always continuous.
This subject will be revisited at the end of chapter 7.

Figure 6.46. The change of variable mapping need not behave well on the boundary

This section places a few more topological ideas into play to set up the
proof of the Change of Variable Theorem in the next section. The symbols K,
A, , and f denoting the set, the open superset, the change of variable, and the
function in the theorem will retain their meanings throughout the discussion.
Symbols such as S will denote other sets, symbols such as will denote other
transformations, and symbols such as g will denote other functions.
Recall some topological ideas that we have already discussed.
For any point a Rn and any positive real number r > 0, the open ball
centered at a of radius r is the set
B(a, r) = {x Rn : |x a| < r} .
A point a Rn is called a limit point of a set S Rn if every open ball
centered at a contains some point x S such that x 6= a. The subset A
of Rn is called closed if it contains all of its limit points.
Definition 6.8.1. Let S be a subset of Rn . Its closure S is the smallest closed
superset of S.
 6.8 Topological Preliminaries for the Change of Variable Theorem 335

Here smallest is taken in the sense of set-containment. The intersection of

closed sets is closed (exercise 6.8.1(a)), and so S is the intersection of all closed
supersets of S, including Rn . Thus S exists and is unique. The special-case
definition
B(a, r) = {x Rn : |x a| r}
from section 5.1 is consistent with Definition 6.8.1.
Closed sets can also be described in terms of boundary points rather than
limit points.

Definition 6.8.2. Let S be a subset of Rn . A point p Rn is called a bound-

ary point of S if for every r > 0 the open ball B(p, r) contains a point from S
and a point from the complement S c . The boundary of S, denoted S, is the
set of boundary points of S.

A boundary point of a set need not be a limit point of the set, and a limit
point of a set need not be a boundary point of the set (exercise 6.8.1(b)).
Nonetheless, similarly to the definition of closed set in the second bullet be-
fore Definition 6.8.1, a set is closed if and only if it contains all of its boundary
points (exercise 6.8.1(c)). The boundary of any set is closed (exercise 6.8.1(d)).
Since the definition of boundary point is symmetric in the set and its comple-
ment, the boundary of the set is also the boundary of the complement,

S = (S c ).

In general, the closure of a set is the union of the set and its boundary (exer-
cise 6.8.2(a)),
S = S S.
If S is bounded then so is its closure S (exercise 6.8.2(b)), and therefore the
closure of a bounded set is compact. The special-case definition

B(a, r) = {x Rn : |x a| = r}

from section 6.1 is consistent with Definition 6.8.2.

Definition 6.8.3. An open box in Rn is a set of the form

J = (a1 , b1 ) (a2 , b2 ) (an , bn ).

The word box, unmodified, continues to mean a closed box.

Proposition 6.8.4 (Finiteness Property of Compact Sets: Special

Case of the HeineBorel Theorem). Consider a compact set K Rn .
Suppose that some collection of open boxes Ji covers K. Then a finite collec-
tion of the open boxes Ji covers K.
336 6 Integration

Proof (Sketch). Suppose that no finite collection of the open boxes Ji cov-
ers K. Let B1 be a box that contains K. Partition B1 into 2n subboxes B e by
bisecting it in each direction. If for each subbox B,e some finite collection of
the open boxes Ji covers K B, e then the 2n -fold collection of these finite col-
lections in fact covers all of K. Thus no finite collection of the open boxes Ji
covers K B e for at least one subbox B e of B1 . Name some such subbox B2 ,
repeat the argument with B2 in place of B1 , and continue in this fashion,
obtaining nested boxes
B1 B2 B3
whose sides are half as long at each succeeding generation, and such that no
K Bj is covered by a finite collection of the open boxes Ji . The intersection
K B1 B2 contains at most one point because the boxes Bj eventually
shrink smaller than the distance between any two given distinct points. On
the other hand, since each K Bj is nonempty (else the empty subcollection
of the open boxes Ji covers it), there is a sequence {cj } with each cj K Bj ;
and since K is compact and each Bj is compact and the Bj are nested, the
sequence {cj } has a subsequence that converges in K and in each Bj , hence
converging in the intersection K B1 B2 . Thus the intersection is a
single point c. Some open box Ji covers c because c K, and so because the
boxes Bj shrink to c, also Ji covers Bj for all high enough indices j. This
contradicts the fact that no K Bj is covered by finitely many Ji . Thus the
initial supposition that no finite collection of the open boxes Ji covers K is
untenable.

Although the finiteness property of compact sets plays only a small role in
these notes, the idea is important and far-reaching. For example, it lies at the
heart of sequence-free proofs that the continuous image of a compact set is
compact, the continuous image of a connected set is connected, and continuity
on compact sets is uniform.
The following lemma is similar to the Difference Magnification Lemma
(Lemma 5.1.3). Its content is that although passing a box through a mapping
neednt give another box, if the box is somewhat uniform in its dimensions
and if the mapping has bounded derivatives then the mapping takes the box
into a second box that isnt too much bigger than the original.

Lemma 6.8.5 (Box-Volume Magnification Lemma). Let B be a box

in Rn whose longest side is at most twice its shortest side. Let g be a dif-
ferentiable mapping from an open superset of B in Rn back to Rn . Suppose
that there is a number c such that |Dj gi (x)| c for all i, j {1, , n} and
all x B. Then g(B) sits in a box B such that vol(B ) (2nc)n vol(B).

Proof. Let x be the centerpoint of B and let x

be any point of B. Make the
line segment connecting x to x
the image of a function of one variable,

: [0, 1] Rn , x x).
(t) = x + t(
 6.8 Topological Preliminaries for the Change of Variable Theorem 337

Fix any i {1, , n}. Identically to the proof of the Difference Magnification
Lemma, we have for some t (0, 1),

x) gi (x) = hgi ((t)), x

gi ( xi.

For each j, the jth entry of the vector gi ((t)) is Dj gi ((t)), and we are
given that |Dj gi ((t))| c. Also, the jth entry of the vector x
x satisfies
|
xj xj | /2 where is the longest side of B. Thus

|gi (
x) gi (x)| nc/2,

and so
gi (B) [gi (x) nc/2, gi (x) + nc/2].
Apply this argument for each i {1, , n} to show that g(B) lies in the
box B centered at g(x) having sides nc and therefore having volume

vol(B ) = (nc)n .

On the other hand, since the shortest side of B is at least /2,

vol(B) (/2)n .

The result follows.

Using the previous two results, we can show that the property of having
volume zero is preserved under mappings that are well enough behaved. How-
ever, we need to assume more than just continuity. The property of having
volume zero is not a topological property.

Proposition 6.8.6 (Volume Zero Preservation Under C 1 -Mappings).

Let S Rn be a compact set having volume zero. Let A be an open superset
of S, and let
: A Rn
be a C 1 -mapping. Then (S) again has volume zero.

Proof. For each s S there exists an rs > 0 such that the copy of the
box [rs , rs ]n centered at s lies in A (exercise 6.8.5(a)). Let Js denote the cor-
responding open box, i.e., a copy of (rs , rs )n centered at s. By the Finiteness
Property of compact sets, a collection of finitely many of the open boxes Js
covers S, so certainly the corresponding collection U of the closed boxes does
so as well. As a finite union of compact sets, U is compact (exercise 6.8.1(f)).
Therefore the partial derivatives Dj i for i, j = 1, , n are uniformly con-
tinuous on U , and so some constant c bounds all Dj i on U .
Let > 0 be given. Cover S by finitely many boxes Bi having total volume
less than /(2nc)n . After replacing each box by its intersections with the boxes
of U , we may assume that the boxes all lie in U . (Here it is relevant that the
intersection of two boxes is a box.) And after further subdividing the boxes if
338 6 Integration

necessary, we may assume that the longest side of each box is at most twice the
shortest side (exercise 6.8.6(b)). By the Box-Volume Magnification Lemma,
the -images of the boxes lie in a union of boxes Bi having volume
X X
vol(Bi ) (2nc)n vol(Bi ) < .
i i

The last topological preliminary that we need is the formal definition of

interior.

Definition 6.8.7 (Interior Point, Interior of a Set). Let S Rn be a

set. Any nonboundary point of S is an interior point of S. Thus x is an
interior point of S if some open ball B(x, r) lies entirely in S. The interior
of S is
S = {interior points of S}.

The interior of any set S is open (exercise 6.8.6(a)). Any set decomposes
as the disjoint union of its interior points and its boundary points (exer-
cise 6.8.6(b)),
S = S (S S), S S = .
As anticipated at the beginning of the section, we now can complete the
argument that the properties of the set K in the Change of Variable Theorem
are preserved by the mapping in the theorem.

Proposition 6.8.8. Let K Rn be a compact and connected set having

boundary of volume zero. Let A be an open superset of K, and let : A Rn
be a C 1 -mapping such that det 6= 0 everywhere on K . Then (K) is again
a compact and connected set having boundary of volume zero.

Proof. We have discussed the fact that (K) is again compact and connected.
Restrict to K. The Inverse Function Theorem says that maps interior
points of K to interior points of (K), and thus ((K)) (K). By the
Volume-Zero Preservation proposition vol((K)) = 0. So vol(((K))) = 0
as well.

Exercises

6.8.1. (a) Show that every intersectionnot just twofold intersections and
not even just finite-fold intersectionsof closed sets is closed. (Recall from
Proposition 2.4.5 that a set S is closed if and only if every sequence in S that
converges in Rn in fact converges in S.)
(b) Show by example that a boundary point of a set need not be a limit
point of the set. Show by example that a limit point of a set need not be a
boundary point of the set.
 6.9 Proof of the Change of Variable Theorem 339

(c) Show that a set is closed if and only if it contains each of its boundary
points. (Again recall the characterization of closed sets mentioned in part (a).)
(d) Show that the boundary of any set is closed.
(e) Show that every union of two closed sets is closed. It follows that any
union of finitely many closed sets is closed. Recall that by definition a set is
open if its complement is closed. Explain why consequently every intersection
of finitely many open sets is open.
(f) Explain why any union of finitely many compact sets is compact.

6.8.2. Let S be any subset of Rn .

(a) Show that its closure is its union with its boundary, S = S S.
(b) Show that if S is bounded then so is S.

6.8.3. (a) Which points of the proof of Proposition 6.8.4 are sketchy? Fill in
the details.
(b) Let S be an unbounded subset of Rn , meaning that S is not contained
in any ball. Find a collection of open boxes Ji that covers S but such that no
finite subcollection of the open boxes Ji covers S.
(c) Let S be an bounded but non-closed subset of Rn , meaning that S is
bounded but missing a limit point. Find a collection of open boxes Ji that
covers S but such that no finite subcollection of the open boxes Ji covers S.

6.8.4. Let > 0. Consider the box B = [0, 1] [0, ] R2 , and consider
the mapping g : R2 R2 given by g(x, y) = (x, x). What is the small-
est box B containing g(B)? What is the ratio vol(B )/vol(B)? Discuss the
relation between this example and Lemma 6.8.5.

6.8.5. The following questions are about the proof of Proposition 6.8.6.
(a) Explain why for each s S there exists an rs > 0 such that the copy
of the box [rs , rs ]n centered at s lies in A.
(b) Explain why any box (with all sides assumed to be positive) can be
subdivided into boxes whose longest side is at most twice the shortest side.

6.8.6. Let S Rn be any set.

(a) Show that the interior S is open.
(b) Show that S decomposes as the disjoint union of its interior points and
its boundary points.

6.9 Proof of the Change of Variable Theorem

Again recall the statement of the Change of Variable Theorem:
Let K Rn be a compact and connected set having boundary of volume
zero. Let A be an open superset of K, and let : A Rn be a C 1 -
mapping such that is injective on K and det 6= 0 on K . Let
f : (K) R be a continuous function. Then
340 6 Integration
Z Z
f= (f ) | det |.
(K) K

We begin chipping away at the theorem by strengthening its hypotheses.

Proposition 6.9.1 (Optional Hypothesis-Strengthening). To prove the

Change of Variable Theorem, it suffices to prove the theorem subject to any
combination of the following additional hypotheses:
K is a box.
is injective on all of A,
det 6= 0 on all of A.

Before proceeding to the proof of the proposition, it deserves comment

that we will not always want K to be a box. But once the proposition is
proved, we may take K to be a box or not as convenient.

Proof. Let > 0 be given.

Let B be a box containing K, and let P be a partition of B into subboxes J.
Define three types of subbox,

type I : J such that J K ,

type II : J such that J K 6= (and thus J (B\K) 6= ),
type III : J such that J B (K K).

(In the left side of figure 6.47, the type I subboxes are shaded and the type II
subboxes are white. There are no type III subboxes in the figure, but type III
subboxes play no role in the pending argument anyway.) The three types of
box are exclusive and exhaustive (exercise 6.9.2(a)).

Figure 6.47. Type I and type II subboxes, image of the type I subboxes

Also define a function

 6.9 Proof of the Change of Variable Theorem 341
(
(f )(x) | det (x)| if x K,
g : B R, g(x) =
0 if x
/ K.

The continuous function f is necessarily bounded on (K), say by R. The

partial derivatives Dj i of the component functions of are continuous on K,
and so the continuous function | det | is bounded on the compact set K, say
e Thus RR
by R. e bounds g on B.
As in the proof of the Volume Zero Preservation proposition (Proposi-
tion 6.8.6), we can cover the subset K of A by a collection U of finitely
many boxes that is again a subset of A, and so the continuous partial deriva-
tives Dj i of the component functions of are bounded on the compact
set U , say by c. We may assume that the partition P is fine enough that all
subboxes J of type I and type II lie in U (exercise 6.9.2(b)). And we may
assume that the longest side of each subbox J is at most twice the shortest
side. Recall that > 0 has been given. Because the boundary of K has volume
zero, we may further assume that the partition P is fine enough that
X  

vol(J) < min ,
R(2nc)n RR e
J:type II

(exercise 6.9.2(c)).
Let [
(K)I = (J), (K)II = (K)\(K)I .
J:type I

(Thus (K)I is shaded in the right side of figure 6.47 while (K)II is white.)
Then the integral on the left side of the equality in the Change of Variable
Theorem decomposes into two parts,
Z Z Z
f= f+ f,
(K) (K)I (K)II

and because is injective on K , the previous display rewrites as

Z X Z Z
f= f+ f. (6.12)
(K) J : type I (J) (K)II

Also, [
(K)II (J),
J : type II

so that Z  Z Z
  X
 
 f |f | |f |.
 (K)II  (K)II (J)
J : type II

By the Box-Volume Magnification Lemma (Lemma 6.8.5), for each box J of

type II, vol((J)) (2nc)n vol(J). Thus, by the bounds on f and on the sum
of the type II box-volumes, it follows that
342 6 Integration
Z 
 
 
 f  < .
 (K)II 

That is, the second term on the right side of (6.12) contributes as negligibly
as desired to the integral on the left side, which is the integral on the left side
of the Change of Variable Theorem. In terms of figure 6.47, the idea is that if
the boxes in the left half of the figure are refined until the sum of the white
box-areas is small enough then the integral of f over the corresponding small
white region in the right half of the figure becomes negligible
Meanwhile, the integral on the right side of the equality in the Change of
Variable Theorem also decomposes into two parts,
Z X Z X Z
(f ) | det | = g+ g. (6.13)
K J : type I J J : type II J

By the bounds on g and on the sum of the type II box-volumes,

 
 Z 
 X  X Z
 g  |g| < .
 
J : type II J  J : type II J

That is, the second term on the right side of (6.13) contributes as negligibly
as desired to the integral on the left side, which is the integral on the right
side of the Change of Variable Theorem. In terms of figure 6.47, the idea is
that if the boxes in the left half of the figure are refined until the sum of the
white box-areas is small enough then the integral of (f ) | det | over the
white boxes becomes negligible. That is, it suffices to prove the Change of
Variable Theorem for boxes like the shaded boxes in the left half of the figure.
The type I subboxes J of the partition of the box B containing the orig-
inal K (which is not assumed to be a box) satisfy all of the additional hy-
potheses in the statement of the proposition: each J is a box, and we may
shrink the domain of to the open superset K of each J, where is injective
and where det 6= 0. Thus, knowing the Change of Variable Theorem sub-
ject to any of the additional hypotheses says that the first terms on the right
sides of (6.12) and (6.13) are equal, making the integrals on the left sides lie
within of each other. Since is arbitrary, the integrals are in fact equal. In
sum, it suffices to prove the Change of Variable Theorem assuming any of the
additional hypotheses as desired.

Proposition 6.9.2 (Alternative Optional Hypothesis-Strengthening).

To prove the Change of Variable Theorem, it suffices to prove the theorem sub-
ject to the following additional hypotheses:
(K) is a box (but now we may not assume that K is a box).
is injective on all of A.
det =
6 0 on all of A.
 6.9 Proof of the Change of Variable Theorem 343

Similarly to the remark after Proposition 6.9.1, we will not always want
the additional hypotheses.

Proof. With the previous proposition in play, the idea now is to run through its
proof in reverse, starting from the strengthened hypotheses that it grants us.
Thus we freely assume that K is a box, that the change of variable mapping
is injective on all of A, and that det 6= 0 on all of A. By the Inverse Function
Theorem, the superset (A) of (K) is open and : A (A) has a C 1
inverse
1 : (A) A.
Let > 0 be given.
Let B be a box containing (K), and let P be a partition of B into
subboxes J. Define three types of subbox,

type I : J such that J (K) ,

type II : J such that J (K) 6= (and thus J (B\(K)) 6= ),
type III : J such that J B ((K) (K)).

These three types of box are exclusive and exhaustive. Also, define as before
(
(f )(x) | det (x)| if x K,
g : B R, g(x) =
0 if x
/ K.

Again, f is bounded on (K), say by R, and | det | is bounded on K, say

e so that RR
by R, e bounds g on B. (See figure 6.48, in which the type I subboxes
cover nearly all of (K) and their inverse images cover nearly all of K.)

Figure 6.48. Type I, II, and III subboxes, inverse image of the type I subboxes

Cover the subset (K) of (A) by a collection U of finitely many boxes

that is again a subset of (A). Then the continuous partial derivatives Dj 1
i
344 6 Integration

of the component functions of 1 are bounded on the compact set U , say

by c. We may assume that the partition P is fine enough that all subboxes J
of type I and type II lie in U . And we may assume that the longest side of each
subbox J is at most twice the shortest side. Recall that > 0 has been given.
Because the boundary of (K) has volume zero, we may further assume that
the partition P is fine enough that
( )
X
vol(J) < min , .
e
R RR(2nc) n
J:type II

Let [
KI = 1 (J), KII = K\KI .
J:type I

Then the integral on the left side of the equality in the Change of Variable
Theorem decomposes into two parts,
Z X Z X Z
f= f+ f. (6.14)
(K) J : type I J J : type II J

By the bounds on f and on the sum of the type II box-volumes,

 
 X Z  X Z
 
 f  |f | < .
 
J : type II J  J : type II J

That is, the second term on the right side of (6.14) contributes as negligibly
as desired to the integral on the left side, which is the integral on the left side
of the Change of Variable Theorem.
Meanwhile, the integral on the right side of the equality in the Change of
Variable Theorem also decomposes into two parts,
Z Z Z
(f ) | det | = g+ g,
K KI KII

and because 1 is injective, the previous display rewrites as

Z X Z Z
(f ) | det | = g+ g. (6.15)
K J : type I 1 (J) KII

Also, [
KII 1 (J),
J : type II

so that Z  Z Z
  X
 g  |g| |g|.

KII KII J : type II 1 (J)
 6.9 Proof of the Change of Variable Theorem 345

For each box J of type II, vol(1 (J)) (2nc)n vol(J). Thus, by the bounds
on g and on the sum of the type II box-volumes, it follows that
Z 
 
 g  < .
 
KII

That is, the second term on the right side of (6.15) contributes as negligibly
as desired to the integral on the left side, which is the integral on the right
side of the Change of Variable Theorem.
The type I subboxes J of the partition of the box B containing the orig-
inal (K) (which is not assumed to be a box) satisfy the new additional
hypothesis in the statement of the proposition. The other two additional hy-
pothesis in the statement of the proposition are already assumed. Thus, know-
ing the Change of Variable Theorem subject to the additional hypotheses says
that the first terms on the right sides of (6.14) and (6.15) are equal, making
the integrals on the left sides lie within of each other. Since is arbitrary, the
integrals are in fact equal. In sum, it suffices to prove the Change of Variable
Theorem assuming the additional hypotheses as desired.

Proposition 6.9.3 (Further Optional Hypothesis-Strengthening). To
prove the Change of Variable Theorem, it suffices to prove the theorem subject
to the additional hypothesis that f is identically 1.
As with the other hypothesis-strengthenings, we will not always want f to
be identically 1, but we may take it to be so when convenient.
Proof. We assume the strengthened hypotheses given us by Proposition 6.9.2.
Let P be a partition of the box (K) into subboxes J. For each subbox J,
view the quantity MJ (f ) = sup {f (x) : x J} both as a number and as a
constant function. Assume that the Change of Variable Theorem holds for
the constant function 1 and therefore for any constant function, and compute
Z XZ
(f ) | det | = (f ) | det |
K J 1 (J)
XZ
(MJ (f ) ) | det |
J 1 (J)
XZ
= MJ (f ) by the assumption
J J
X
= MJ (f ) vol(J)
J
= U (f, P ).
R
As a lower bound of the upper sums, K (f )| det | is at most the integral,
Z Z

(f ) | det | f.
K (K)
346 6 Integration

A similar argument gives the opposite inequality, making the integrals equal
as desired.

The next result will allow the proof of the Change of Variable Theorem to
decompose the change of variable mapping.

Proposition 6.9.4 (Persistence Under Composition). In the Change of

Variable Theorem, suppose that the change of variable mapping is a composi-
tion
=
where the mappings
: A Rn
and
e Rn
:A e is an open superset of (K))
(where A
satisfy the hypotheses of the Change of Variable Theorem. If
Z Z
g= (g ) | det | for continuous functions g : (K) R
(K) K

and Z Z
1= | det |
( (K)) (K)

then also Z Z
1= | det |.
(K) K

Proof. The argument is a straightforward calculation using the definition of ,

the second given equality, the first given equality, the multiplicativity of the
determinant, the Chain Rule, and again the definition of ,
Z Z Z
1= 1= | det |
(K) ( (K)) (K)
Z
= | det( )| | det |
K
Z


= | det ( ) |
ZK Z
= | det( ) | = | det |.
K K

Proposition 6.9.5 (Linear Change of Variable). The Change of Variable

Theorem holds for invertible linear mappings.
 6.9 Proof of the Change of Variable Theorem 347

Proof. Let
T : Rn Rn
be an invertible linear mapping having matrix M . Thus T (x) = M for all x.
Also, T is a composition of recombines, scales, and transpositions, and so
by the persistence of the Change of Variable Theorem under composition it
suffices to prove the theorem assuming that T is a recombine or a scale or a
transposition. In each case, Propositions 6.9.1 and 6.9.3 allow us to assume
that K is a box B and f = 1. Thus the desired result is simply

vol(T (B)) = | det M | vol(B),

and we established this formula back in section 3.8.

The Change of Variable Theorem is proved partially by induction on the

dimension n.

Proposition 6.9.6 (Base Case for the Induction). The Change of Vari-
able Theorem holds if n = 1.

Proof. Because n = 1, K is an interval [a, b] R where a b. Here is

where we use the hypothesis that K is connected. Since we have not studied
connected sets closely, the reader is being asked to take for granted that any
compact and connected subset of R is a closed and bounded interval. (Or see
exercise 6.9.1 for a proof that any compact and path-connected subset of R is
a closed and bounded interval.) The continuous function

: [a, b] R

can take the value 0 only at a and b. Thus by the Intermediate Value Theorem,
never changes sign on [a, b]. If 0 on [a, b] then is increasing, and so
(using Theorem 6.4.3 for the second equality),
Z Z (b) Z b Z
f= f= (f ) = (f ) | |.
([a,b]) (a) a [a,b]

If 0 on [a, b] then is decreasing, and so

Z Z (a) Z (b) Z b Z
f= f = f = (f ) = (f ) | |.
([a,b]) (b) (a) a [a,b]

Thus in either case the desired result holds.

Proposition 6.9.7 (Bootstrap Induction Step). For any n > 1, if the

Change of Variable Theorem holds in dimension n 1 then it holds in dimen-
sion n subject to the additional hypothesis that the transformation fixes at
least one coordinate.
348 6 Integration

Figure 6.49. A transformation that fixes the third coordinate

A 3-dimensional transformation that fixes the third coordinate is shown

in figure 6.49. The figure makes the proof of Proposition 6.9.7 inevitable: the
desired result holds for each slice since we are assuming the Change of Variable
Theorem in dimension n 1, and so Fubinis Theorem gives the result for the
entire figure.

Proof. Propositions 6.9.1 and 6.9.3 allow us to assume that K is a box B,

that is injective on B, that det 6= 0 on B, and that f = 1. Also, we may
assume that the coordinate fixed by is the last coordinate. There is a box
Bn1 Rn1 and an interval I = [a, b] R such that
[
B= Bn1 {t}.
tI

By assumption, is a C 1 -mapping on an open superset A of B. For each t I

let At denote the cross-section of A with last coordinate t,

At = {x Rn1 : (x, t) A}.

Then At is an open superset of Bn1 in Rn1 . For each t I define a mapping

t : At Rn1 , t (x) = (1 (x, t), , n1 (x, t)).

Each t is a C 1 -mapping on an open superset of Bn1 , and

[
(B) = t (Bn1 ) {t}.
tI

Since is injective on B and det 6= 0 on B, it follows that each t is

injective on Bn1 , and the formula

| det t (x)| = | det (x, t)|, (x, t) B (6.16)

 6.9 Proof of the Change of Variable Theorem 349

(exercise 6.9.3) shows that det t 6= 0 on Bn1 . Thus for each t, the set Bn1
and the transformation t satisfy the Change of Variable Theorem hypotheses
in dimension n 1. Compute, using Fubinis Theorem, quoting the Change
of Variable Theorem in dimension n 1, and citing formula (6.16) and again
using Fubinis Theorem, that
Z Z Z Z Z Z

1= 1= | det t | = | det |.
(B) tI t (Bn1 ) tI Bn1 B

At long last we can prove the Change of Variable Theorem for n > 1.

Proof. We may assume the result for dimension n 1, and we may assume
that K is a box B, that A is an open superset of B, and that : A Rn is
a C 1 -mapping such that is injective on A and det 6= 0 on A. We need to
show that Z Z
1= | det |. (6.17)
(B) B

To prove the theorem, we will partition B into subboxes J, each J having an

open superset AJ on which is a composition

= T ,

where and are C 1 -mappings that fix at least one coordinate and T is a
linear transformation. Note that , , and T inherit injectivity and nonzero
determinant-derivatives from , so that in particular T is invertible. Since the
theorem holds for each of , , and T , it holds for their composition. In more
detail,
Z Z
1= | det T | by Proposition 6.9.5
T ( ( (J))) ( (J))
Z
= | det(T )| | det | by Proposition 6.9.7
(J)
Z
= | det(T ) | by the Chain Rule
(J)
Z


= | det (T ) | | det | by Proposition 6.9.7
ZJ
= | det(T ) | by the Chain Rule.
J

That is, for each J, Z Z

1= | det |,
(J) J

and so summing over all subboxes J finally gives (6.17).

350 6 Integration

To obtain the subboxes J, proceed as follows for each point x B. Let

T = Dx

and define
e = T 1 ,

so that D ex = idn is the n-dimensional identity map. Introduce the nth
projection function, n (x1 , , xn ) = xn , and further define

: A Rn , e1 , ,
= ( en1 , n ),

so that Dx = idn as well. By the Inverse Function Theorem, is locally

invertible. Let Jx be a subbox of B containing x having an open superset Ax
such that 1 exists on (Ax ). Now define

: (Ax ) Rn , en 1 ).
= (1 , , n1 ,

Then = e = T 1 on Ax , so that T = on Ax , and

thus , , and T have the desired properties. (Figure 6.50 illustrates the
decomposition for the polar coordinate mapping. In the figure, changes
only the first coordinate, changes only the second, and then the linear
mapping T completes the polar coordinate change of variable.)

Figure 6.50. Local decomposition of the polar coordinate mapping

Cover B by the collection of open interiors of the boxes Jx . By the finite-

ness property of B, some finite collection of the interiors covers B, and so
certainly the corresponding finite collection of the boxes Jx themselves cov-
ers B. Partition B into subboxes J so that each J lies in one of the finitely
many Jx , and the process is complete.

In contrast to all of this, recall the much-easier proof of the one-dimensional

Change of Variable Theorem, using the construction of an antiderivative by
integrating up to variable endpoint (Theorem 6.4.1, sometimes called the first
Fundamental Theorem of Integral Calculus) and using the (second) Funda-
mental Theorem of Integral Calculus twice,
 6.9 Proof of the Change of Variable Theorem 351
Z b Z b Z x
(f ) = (F ) where F (x) = f , so F = f
a a a
Z b
= (F ) by the chain rule
a
= (F )(b) (F )(a) by the FTIC
= F ((b)) F ((a)) by definition of composition
Z (b)
= F by the FTIC again
(a)
Z (b)
= f since F = f .
(a)

We see that the one-variable Fundamental Theorem of Integral Calculus is

doing the bulk of the work. Chapter 9 will give us a multivariable Fundamental
Theorem, after which we can sketch a proof of the multivariable Change of
Variable Theorem in the spirit of the one-variable argument just given. A fully
realized version of that proof still has to handle topological issues, but even
so it is more efficient than the long, elementary method of this section.

Exercises
6.9.1. Let K be a nonempty compact subset of R. Explain why the quantities
a = min{x : x K} and b = max{x : x K} exist. Now further assume that
K is path-connected, so that in particular there is a continuous function
: [0, 1] R
such that (0) = a and (1) = b. Explain why consequently K = [a, b].
6.9.2. (a) Explain to yourself as necessary why the three types of rectangle
in the proof of Proposition 6.9.1 are exclusive. Now suppose that the three
types are not exhaustive, i.e., some rectangle J lies partly in K and partly
in (B\K) without meeting the set K = (B\K). Supply details as necessary
for the following argument. Let x J lie in K and let x J lie in (B\K) .
Define a function from the unit interval to R by mapping the interval to the
line segment from x to x , and then mapping each point of the segment to 1 if
it lies in K and to 1 if it lies in B\K. The resulting function is continuous
on the interval, and it changes sign on the interval, but it does not take the
value 0. This is impossible, so the rectangle J can not exist.
(b) In the proof of Proposition 6.9.1, show that we may assume that the
partition P is fine enough that all subboxes J of type I and type II lie in U .
(c) In the proof of Proposition 6.9.1, show that given > 0, we may assume
that the partition P is fine enough that
X  

vol(J) < min , .
R(2nc)n RR e
J:type II
352 6 Integration

6.9.3. In the proof of Proposition 6.9.7, establish formula (6.16).

6.9.4. Here is a sketched variant of the endgame of the Change of Variable

proof: A slightly easier variant of Proposition 6.9.7 assumes that the transfor-
mation changes at most one coordinate, and then the process of factoring
locally as a composition can be iterated until each factor is either linear or
changes at most one coordinate. Fill in the details.

6.10 Summary

Integration is a synthesis of many small elements into a whole. The integral

of a continuous function on a reasonably-shaped region exists, as is shown
by careful management of details and the technical point that continuity is
uniform on compact sets. Fubinis Theorem, which is so apparent intuitively,
is also easy to prove essentially from the definitions. However, the Change
of Variable Theorem, which also seems plausible, requires a more elaborate
proof.
7
Approximation by Smooth Functions

Let k be a nonnegative integer. Recall that a C k -function on Rn is a function

all of whose partial derivatives up through order k exist and are continuous.
That is, to say that a function

f : Rn R

is C k is to say that f , and Dj f for j = 1, , n, and Djj f for j, j = 1, , n,

and so on up to all Dj1 jk f exist and are continuous on Rn . Various ideas
that we have discussed so far have required different values of k:
If f is C 1 then f is differentiable in the multivariable sense of derivative
(Theorem 4.5.3).
If f is C 2 then its mixed second order derivatives D12 f and D21 f are equal
(Theorem 4.6.1).
The multivariable max/min test (Proposition 4.7.8) assumes a C 2 -function.
The Inverse Function Theorem says that if A Rn is open and f : A
Rn is componentwise C 1 and its derivative Dfa is invertible at a point a
then f is locally invertible about a, and the local inverse is again C 1 (The-
orem 5.2.1).
A C 0 -mapping from the unit interval can fill the square, but a C 1 -mapping
cannot.
If f (again scalar-valued now) is C 0 then it is integrable over any compact
set having boundary of volume zero R (sectionR6.5).
In the Change of Variable formula (K) f = K (f )| det | for multiple
integrals (Theorem 6.7.1) the change of variable mapping is assumed to
be C 1 and for now the integrand f is assumed to be C 0 . We will return to
this example at the very end of the chapter.
Meanwhile, a smooth function is a function on Rn all of whose partial deriva-
tives of all orders exist. Smooth functions are also called C -functions, an
appropriate notation because the derivatives of each order are continuous
in consequence of the derivatives of one-higher order existing. This chapter
354 7 Approximation by Smooth Functions

briefly touches on the fact that for functions that vanish off a compact set,
C 0 -functions and C 1 -functions and C 2 -functions are well approximated by C -
functions.
The approximation technology is an integral called the convolution. The
idea is as follows. Suppose that we had a function
: Rn R
with the following properties:
(1) (x)
R = 0 for all x 6= 0,
(2) xRn (x) = 1.
So conceptually the graph of is an infinitely high, infinitely narrow spike
above 0 having total volume 1. No such function exists, at least not in the
usual sense of function. (The function , known as the Dirac delta function,
is an example of a distribution, distributions being objects that generalize
functions.) Nonetheless, if were sensible then for any function f : Rn R
and for any x Rn we would have in consequence that the graph of the
product of f and the x-translate of is an infinitely high, infinitely narrow
spike above x having total volume f (x),
(1) Rf (y)(x y) = 0 for all y 6= x,
(2) yRn f (y)(x y) = f (x).
That is, granting the Dirac delta function, any function f can be expressed
as an integral. The idea motivating convolution is that if we replace the ideal-
ized delta function by a smooth pulse , big-but-finitely high and small-but-
positively wide and having total volume 1, and if f is well enough behaved
(e.g., f is continuous and vanishes off a compact set) then we should still
recover a close approximation of f from the resulting integral,
Z
f (y)(x y) f (x).
yRn

The approximating integral on the left side of the previous display is the
convolution of f and evaluated at x. Although f is assumed only to be con-
tinuous, the convolution is smooth. Indeed, any xi -derivative passes through
the y-integral and is smooth, so that
Z Z

f (y)(x y) = f (y) (x y),
xi y y x i

and similarly for higher derivatives.

One can see convolution in action visually by comparing graphs of convolu-
tions to the graph of the original function. And the conceptual framework for
establishing the properties of convolution analytically is not difficult. Having
discussed approximation by convolution, we will freely assume in the remain-
ing chapters of these notes that our functions are C , i.e., that they are
smooth.
 7.1 Spaces of Functions 355

7.1 Spaces of Functions

To begin, we quantify the phrase functions that vanish off a compact set from
the chapter introduction.

Definition 7.1.1 (Support). Consider a function

f : Rn R

The support of f is the closure of the set of its inputs that produce nonzero
outputs,
supp(f ) = {x Rn : f (x) 6= 0}.
The function f is compactly supported if its support is compact. The class
of compactly supported C k -functions is denoted Cck (Rn ). Especially, Cc0 (Rn )
denotes the class of compactly supported continuous functions.

Each class Cck (Rn ) of functions forms a vector space over R (exercise 7.1.1).
Figure 7.1 shows a compactly supported C 0 -function on R and its support.
The graph has some corners, so the function is not C 1 .

Figure 7.1. Compactly supported continuous function on R and its support

The spaces of compactly supported functions shrink as their member-

functions are required to have more derivatives,

Cc0 (Rn ) Cc1 (Rn ) Cc2 (Rn ) ,

and we will see that all of the containments are proper.

Definition 7.1.2 (Test Function). A test function is a compactly sup-

ported smooth function. The class of test functions is denoted Cc (Rn ).

The class of test functions sits at the end of the chain of containments of
function-spaces from a moment ago,
356 7 Approximation by Smooth Functions
\
Cc (Rn ) = Cck (Rn ),
k0

and as an intersection of vector spaces over R, the test functions Cc (Rn )

again form a vector space over R. In the chain of containments

Cc0 (Rn ) Cc1 (Rn ) Cc2 (Rn ) Cc (Rn ),

all of the containments are proper. Indeed, for the first containment, Weier-
strass showed how to construct a function f of one variable, having sup-
port [0, 1], that is continuous everywhere but differentiable nowhere on its
support. The function of n variables

f0 (x1 , x2 , , xn ) = f (|x1 , x2 , , xn |)

thus lies in Cc0 (Rn ) but not in Cc1 (Rn ). Next, the function
Z x1
f1 (x1 , x2 , , xn ) = f0 (t1 , x2 , , xn )
t1 =0

lies in Cc1 (Rn ) but not Cc2 (Rn ) because its first partial derivative is f0 , which
does not have a first partial derivative. Defining f2 as a similar integral of f1
gives a function that lies in Cc2 (Rn ) but not Cc3 (Rn ), and so on. Finally, none
of the functions fk just described lies in Cc (Rn ).
For any k > 0 and any f Cck (Rn ), the supports of the partial derivatives
are contained in the support of the original function,

supp(Dj f ) supp(f ), j = 1, , n.

Thus the partial derivative operators Dj take Cck (Rn ) to Cck1 (Rn ) as sets.
The operators are linear because

Dj (f + f) = Dj f + Dj f, f, f Cck (Rn )

and
Dj (cf ) = c Dj f, f Cck (Rn ), c R.
In addition, more can be said about the Dj operators. Each space Cck (Rn ) of
functions carries an absolute value function having properties similar to the
absolute value on Euclidean space Rn . With these absolute values in place,
the partial differentiation operators are continuous.

Definition 7.1.3 (Cck (Rn ) Absolute Value). The absolute value function
on Cc0 (Rn ) is

| | : Cc0 (Rn ) R, |f | = sup{|f (x)| : x Rn }.

Let k be a nonnegative integer. The absolute value function on Cck (Rn ) is

 7.1 Spaces of Functions 357

| |k : Cck (Rn ) R
given by


|f |,





|D j f | for j = 1, , n,





|f |k = max |D jj f | for j, j = 1, , n,
.

.


.


.




|Dj1 jk f | for j1 , , jk = 1, , n
That is, |f |k is the largest absolute value of f or of any derivative of f up to
order k. In particular, | |0 = | |.
The largest absolute values mentioned in the definition exist by the Ex-
treme Value Theorem since the relevant partial derivatives are compactly sup-
ported and continuous. By contrast, we have not defined an absolute value
on the space of test functions Cc (Rn ) because the obvious attempt to extend
Definition 7.1.3 to test function would involve the maximum of an infinite set,
a maximum that certainly need not exist.
Proposition 7.1.4 (Cck (Rn ) Absolute Value Properties).
(A1) Absolute value is positive: |f |k 0 for all f Cck (Rn ), and |f |k = 0 if
and only if f is the zero function.
(A2) Scaling Property: |cf |k = |c| |f |k for all c R and f Cck (Rn ).
(A3) Triangle Inequality: |f + g|k |f |k + |g|k for all f, g Cck (Rn ).
Proof. The first two properties are straightforward to check. For the third
property, note that for any f, g Cc0 (Rn ) and any x Rn ,
|(f + g)(x)| |f (x)| + |g(x)| |f | + |g|.
Thus |f | + |g| is an upper bound of all values |(f + g)(x)|, so that
|f + g| |f | + |g|.
That is, |f + g|0 |f |0 + |g|0 . If f, g Cc1 (Rn ) then the same argument shows
that also |Dj (f + g)| |Dj f | + |Dj g| for j = 1, , n, so that
( )
|f + g|,
|f + g|1 = max
|Dj f + Dj g| for j = 1, , n
( )
|f | + |g|,
max
|Dj f | + |Dj g| for j = 1, , n
( ) ( )
|f |, |g|,
max + max
|Dj f | for j = 1, , n |Dj g| for j = 1, , n
= |f |1 + |g|1 .
358 7 Approximation by Smooth Functions

(For the second inequality, note for example that

max(|f | + |g|) = (|f | + |g|)(
x) for some x

= |f |(
x) + |g|(
x) max |f | + max |g|,
and similarly for each partial derivative.) The proof that |f + g|k |f |k + |g|k
for higher values of k is more of the same.

Now we can verify the anticipated continuity of the linear operators Dj
from Cck (Rn ) to Cck1 (Rn ).
Proposition 7.1.5 (Continuity of Differentiation). For any k 1, the
partial differentiation mappings
Dj : Cck (Rn ) Cck1 (Rn ), j = 1, , n
are continuous.
Proof. Consider any function f Cck (Rn ) and any sequence {fm } in Cck (Rn ).
Suppose that
lim |fm f |k = 0.
m
Then
lim |fm f | = 0,
m
lim |Dj fm Dj f | = 0 for j = 1, , n,
m
lim |Djj fm Djj f | = 0 for j, j = 1, , n,
m
..
.
lim |Dj1 j2 jk fm Dj1 j2 jk f | = 0 for j1 , j2 , jk = 1, , n.
m

Fix any j {1, , n}. As a subset of the information in the previous display,
lim |Dj fm Dj f | = 0,
m
lim |Djj fm Djj f | = 0 for j = 1, , n,
m
..
.
lim |Djj2 jk fm Djj2 jk f | = 0 for j2 , jk = 1, , n.
m

That is,
lim |Dj fm Dj f |k1 = 0.
m
The implication that we have just shown,
lim |fm f |k = 0 = lim |Dj fm Dj f |k1 = 0,
m m

is exactly the assertion that Dj : Cck (Rn ) Cck1 (Rn ) is continuous, and the
proof is complete.

 7.1 Spaces of Functions 359

Again let k 1. The fact that |f |k1 |f |k for any f Cck (Rn ) (exer-
cise 7.1.2) shows that for any f Cck (Rn ) and any sequence {fm } in Cck (Rn ), if
limm |fm f |k = 0 then limm |fm f |k1 = 0. That is, the inclusion mapping

i : Cck (Rn ) Cck1 (Rn ), i(f ) = f

is continuous.
The space Cc (Rn ) of test functions is closed under partial differentiation,
meaning that the partial derivatives of a test function are again test functions
(exercise 7.1.3).
In this chapter we will show that just as any real number x R is ap-
proximated as closely as desired by rational numbers q Q, any compactly
supported continuous function f Cck (Rn ) is approximated as closely as de-
sired by test functions g Cc (Rn ). More precisely, we will show that:
For any f Cck (Rn ), there exists a sequence {fm } in Cc (Rn ) such
that limm |fm f |k = 0.
The fact that limm |fm f |k = 0 means that given any > 0, there exists
a starting index m0 such that fm for all m m0 uniformly approximates f
to within up to kth order. That is, for all m m0 , simultaneously for
all x Rn ,

|fm (x) f (x)| < ,

|Dj fm (x) Dj f (x)| < for j = 1, , n,
|Djj fm (x) Djj f (x)| < for j, j = 1, , n,
..
.
|Dj1 jk fm (x) Dj1 jk f (x)| < for j1 , , jk = 1, , n.

The use of uniform here to connote that a condition holds simultaneously

over a set of values is similar to its use in uniform continuity.

Exercises

7.1.1. Show that each class Cck (Rn ) of functions forms a vector space over R.

7.1.2. Verify that |f |k1 |f |k for any f Cck (Rn ).

7.1.3. Explain why each partial derivative of a test function is again a test
function.

7.1.4. Let {fn } be a sequence of functions in Cc0 (Rn ), and suppose that the
sequence converges, meaning that there exists a function f : Rn R such
that limn fn (x) = f (x) for all x Rn . Must f have compact support? Must
f be continuous?
360 7 Approximation by Smooth Functions

7.2 Pulse Functions

A pulse function is a useful type of test function. To construct pulse functions,
first consider the function
(
0 if x 0,
s : R R, s(x) = 1/x
e if x > 0.

(See figure 7.2.) Each x < 0 lies in an open interval on which s is the constant
function 0, and each x > 0 lies in an open interval on which s is a composition
of smooth functions, so in either case all derivatives s(k) (x) exist. More specif-
ically, for any nonnegative integer k, there exists a polynomial pk (x) such that
the kth derivative of s takes the form


0 if x < 0,
s(k) (x) = pk (x)x2k e1/x if x > 0,


? if x = 0.

Only s(k) (0) is in question. However, s(0) (0) = 0, and if we assume that
s(k) (0) = 0 for some k 0 then it follows (because exponential behavior
dominates polynomial behavior) that

s(k) (h) s(k) (0)

lim+ = lim+ pk (h)h2k1 e1/h = 0.
h0 h h0

That is, s(k+1) (0) exists and equals 0 as well. By induction, s(k) (0) = 0 for
all k 0. Thus s is smooth: each derivative exists, and each derivative is
continuous because the next derivative exists as well. But s is not a test
function because its support is not compact: supp(s) = [0, ).

Figure 7.2. Smooth function

 7.2 Pulse Functions 361

Now the pulse function is defined in terms of the smooth function,

s(x + 1)s(x + 1)
p : R R, p(x) = R 1 .
x=1
s(x + 1)s(x + 1)

The graph of p (figure 7.3) explains the name pulse function. As a product of
compositions of smooth functions, p is smooth. The support of p is [1, 1], so
p is a test function. Also, p is normalized so that
Z
p = 1.
[1,1]

The maximum pulse value p(0) is therefore close to 1 because the pulse graph
is roughly a triangle of base 2, but p(0) is not exactly 1. The pulse function
p2 (x, y) = p(x)p(y) from R2 to R, having support [1, 1]2 , is shown in fig-
ure 7.4. A similar pulse function p3 on R3 can be imagined as a concentration
of density in a box about the origin.

-1 1

Figure 7.3. Pulse function

Exercises

7.2.1. Since the function s in the section is smooth, it has nth degree Taylor
polynomials Tn (x) at a = 0 for all nonnegative integers n. (Here n does not
denote the dimension of Euclidean space.) For what x does s(x) = Tn (x)?

7.2.2. Let p be the pulse function defined in the section. Explain why
supp(p) = [1, 1].

7.2.3. Let p : R R be the one-dimensional pulse function from the section.

362 7 Approximation by Smooth Functions

0
-1

-1
1

Figure 7.4. Two-dimensional pulse function

R x b + x(b a))), where a < b.

(a) Graph the function q(x) = p(2a
(b) Graph the function r(x) = t=1 p(t).
(c) Use the function r from part (b) to give a formula for a test function
that is 0 for x < a, climbs from 0 to 1 for a x b, is 1 for b < x < c, drops
from 1 to 0 for c x d, and is 0 for d < x.

7.3 Convolution

This section shows how to construct test functions from Cc0 (Rn )-functions. In
preparation, we introduce a handy piece of notation.

Definition 7.3.1 (Sum, Difference of Two Sets). Let S and T be subsets

of Rn . Their sum is the set consisting of all sums of a point of S plus a point
of T ,
S + T = {s + t : s S, t T }.
Their difference is similarly

S T = {s t : s S, t T }.

Visually, S + T can be imagined as many copies of T , one based at each

point of S, or vice versa. For example, if K is a three-dimensional box and B
is a small ball about 03 then K +B is slightly larger than K, again shaped like
a box except that the edges and corners are rounded. Similarly, {0} T is the
reflection of T through the origin. The sum or difference of two compact sets
is compact (exercise 7.3.1(a)). The sum of of the open balls B(a, r) and B(b, s)
is B(a+b, r+s) (exercise 7.3.1(b)). The reader is alerted that the set difference
here is different from another, more common notion of set difference, that
being the elements of one set that are not elements of another,
 7.3 Convolution 363

S\T = {s S : s
/ T }.

Returning to Cc0 (Rn )-functions, any such function can be integrated over
all of Rn .

Definition 7.3.2 (Integral of a Cc0 (Rn )-Function). Let f Cc0 (Rn ). The
integral of f is the integral of f over any box that contains its support,
Z Z
f= f where supp(f ) B.
B

In Definition 7.3.2 the integral on the right side exists by Theorem 6.3.1.
Also, the integral on the right side is independent of the suitable box B, always
being the integral over the intersection of all such boxes, the smallest suitable
box. Thus the integral
R on the leftR side exists and is unambiguous. We do not
bother writing Rn f rather than f , because it is understood that by default
we are integrating f over Rn .

Definition 7.3.3 (Mollifying Kernel). Let f Cc0 (Rn ) be a compactly sup-

ported continuous function, and let Cc (Rn ) be a test function. The mol-
lifying kernel associated to f and is the function

: Rn Rn R, (x, y) = f (y)(x y).

For any fixed x Rn , the corresponding cross section of the mollifying kernel
is denoted x ,
x : Rn R, x (y) = (x, y).

For each x Rn , the mollifying kernel x (y) can be nonzero only if y

supp(f ) and x y supp(). It follows that

supp(x ) supp(f ) ({x} supp()).

Therefore x is compactly supported. (Figure 7.5 shows an example of the

multiplicands f (y) and (xy) of x (y), and figure 7.6 shows their compactly
supported product.) Also, since f and are continuous, x is continuous.
That is, for each x, the mollifying kernel x viewed as a function of y again
lies in Cc0 (Rn ), making it integrable by Theorem 6.3.1.
The mollifying kernel is so named for good reason. First, it is a kernel in
the sense that we integrate it to get a new function.

Definition 7.3.4 (Convolution). Let f Cc0 (Rn ) and let Cc (Rn ). The
convolution of f and is the function defined by integrating the mollifying
kernel,
Z Z
f : Rn R, (f )(x) = x (y) = f (y)(x y).
y y
364 7 Approximation by Smooth Functions

f (y) (x y)

Figure 7.5. Multiplicands of the mollifying kernel

x (y)

Figure 7.6. The mollifying kernel is compactly supported

Second, although the mollifying kernel is only as well-behaved as f , inte-

grating it indeed mollifies f in the sense that the integral is as well behaved
as , i.e., the integral is a test function. Even if f is nowhere differentiable,
f has all partial derivatives of all orders while remaining compactly sup-
ported. Furthermore, the derivatives have the natural formula obtained by
passing them through the integral.
Proposition 7.3.5 (Derivatives of the Convolution). Let f Cc0 (Rn )
and let Cc (Rn ). Then also f Cc (Rn ). Specifically, the partial
derivatives of the convolution are the convolutions with the partial derivatives,

Dj (f ) = f Dj , j = 1, , n,

and similarly for the higher order partial derivatives.

The following result helps to prove Proposition 7.3.5. In its statement, the
symbol , which usually denotes a test function, instead denotes a Cc1 (Rn )-
function. The reason for the weaker hypothesis will appear soon in the proof
of Corollary 7.3.7.
Lemma 7.3.6 (Uniformity Lemma for C 1 -Functions). Let Cc1 (Rn ).
Given any > 0, there exists a corresponding > 0 such that for all a Rn
 7.3 Convolution 365

and all nonzero h R, and for any j {1, , n},

 
 (a + hej ) (a) 
|h| < =  Dj (a) < .
h

Proof. Supposing first that for each j {1, , n} a corresponding j exists,

define = min{1 , , n }. Then for all nonzero h R and for any j
{1, , n},
|h| < = |h| < j .
Thus works simultaneously for all j. So it suffices to find j for one fixed-
but-arbitrary j, and from now on in this proof j is fixed
The Mean Value Theorem at the jth coordinate gives for all a Rn and
all nonzero h R,
 
 (a + hej ) (a) 
 D (a)  = |Dj (a + tej ) Dj (a)| where |t| < |h|.
 h
j 

Since Dj is continuous on Rn and is compactly supported, it is uniformly

continuous on Rn , and so given any > 0 there exists a corresponding j > 0
such that for all a Rn and t R,

|Dj (a + tej ) Dj (a)| < if |t| < j .

The desired result follows from the two displays.

Now we can establish the derivative formula for the convolution.

Proof (of Proposition 7.3.5). To see that f is compactly supported, recall

the observation that for a given x, the mollifying kernel x (y) = f (y)(x y)
can be nonzero only at y-values such that

y supp(f ) ({x} supp()).

Such y can exist only if x takes the form

x = y + z, y supp(f ), z supp().

That is, the integrand is always zero if x

/ supp(f ) + supp() (see figure 7.7).
Hence,
supp(f ) supp(f ) + supp().
To show that Dj (f ) exists and equals f Dj for j = 1, , n is
precisely to show that each x-derivative passes through the y-integral,
Z Z

f (y)(x y) = f (y) (x y), j = 1, , n.
xj y y xj

Since the integral is being taken over some box B, the equality follows from
Proposition 6.6.2. But we prove it using other methods, for reasons that will
366 7 Approximation by Smooth Functions

(x y)
f (y)

Figure 7.7. The mollifying kernel is zero for x outside supp(f ) + supp()

emerge later in the chapter. The function f is bounded, say by R, so we can

estimate that for any x Rn and any nonzero h R and any j,
 
 (f )(x + hej ) (f )(x) 
 (f Dj )(x)
 h
R R 
 f (y)(x + hej y) f (y)(x y) Z 
 y y 
= f (y)Dj (x y)
 h y 
Z  
 (x y + hej ) (x y) 
=  f (y) Dj (x y) 
y h
Z  
 (x y + hej ) (x y) 
R  Dj (x y) .
y h

Assuming that |h| < 1, the support of the integrand as a function of y lies in
the bounded set
{x + tej : 1 < t < 1} supp(),
and therefore the integral can be taken over some box B. By the Unifor-
mity Lemma, given any > 0, for all small enough h the integrand is less
than /(R vol(B)) uniformly in y. Consequently the integral is less than /R.
In sum, given any > 0, for all small enough h we have
 
 (f )(x + hej ) (f )(x) 
 (f Dj )(x) < .
 h
Since x is arbitrary, this gives the desired result for first-order partial deriva-
tives,
Dj (f ) = f Dj , j = 1, , n.
As for higher-order partial derivatives, note that Dj Cc (Rn ) for each j.
So the same result for second-order partial derivatives follows,

Djj (f ) = Dj (f Dj ) = f Djj , j, j = 1, , n,

and so on.

 7.3 Convolution 367

The proof of Proposition 7.3.5 required only that each x be integrable,

that f be bounded, and that lie in Cc1 (Rn ). We will make use of this obser-
vation in section 7.5.
If the function f lies in the subspace Cc1 (Rn ) of Cc0 (Rn ) then the partial
derivatives of the convolution pass through the integral to f as well as to .
That is, for differentiable functions, the derivative of the convolution is the
convolution of the derivative.

Corollary 7.3.7. Let k 1, let f Cck (Rn ), and let Cc (Rn ). Then

Dj1 jk (f ) = Dj1 jk f , j1 , , jk = 1, , n.

Proof. Since Z
(f )(x) = f (y)(x y),
y

it follows by the Change of Variable Theorem (replace y by x y) that also

Z
(f )(x) = f (x y)(y).
y

Now the proof of the proposition works with the roles of f and exchanged
to show that Dj (f ) = Dj f for j = 1, , n. (Here is where it is
relevant that the Uniformity Lemma requires only a Cc1 (Rn )-function rather
than a test function.) Similarly, if f Cc2 (Rn ) then because Dj f Cc1 (Rn ) for
j = 1, , n it follows that.

Djj (f ) = Djj f , j, j = 1, , n.

The argument for higher derivatives is the same.

Consider a function f Cc0 (Rn ). Now that we know that any convolution
f (where Cc (Rn )) lies in Cc (Rn ), the next question is to what extent
the test function f resembles the original compactly supported continuous
function f . As already noted, for any x the integral
Z
(f )(x) = f (y)(x y)
y

refers to values of f only on {x}supp(). Especially, if supp() is a small set

about the origin then the convolution value (f )(x), depends only on the be-
havior of the original function f near x. The next section will construct useful
test functions having small support, the idea being that convolutions f
with such test functions will approximate the functions f being convolved.
For example, in figure 7.5 f (x) is small and positive, while the integral of the
mollifying kernel shown in figure 7.6 is plausibly small and positive as well.
368 7 Approximation by Smooth Functions

Exercises

7.3.1. (a) Show that the sum of two compact sets is compact.
(b) Let B(a, r) and B(b, s) be open balls. Show that their sum is B(a +
b, r + s).
(c) Recall that there are four standard axioms for addition, either in the
context of a field or a vector space. Which of the four axioms are satisfied by
set addition, and which are not?
(d) Let 0 < a < b. Let A be the circle of radius b in the (x, y)-plane,
centered at the origin. Let B be the closed disk of radius a in the (x, z)-plane,
centered at (b, 0, 0). Describe the sum A + B.

7.3.2. Let f Cc0 (Rn ), and let Cc (Rn ). Assume

R that 0, i.e., all
output values of are nonnegative, and assume that = 1. Suppose that R
bounds f , meaning that |f (x)| < R for all x. Show that also R bounds f .

7.4 Test Approximate Identity and Convolution

Our next technical tool is a sequence of test functions whose graphs are ever
taller and more narrow, each enclosing volume 1.

Definition 7.4.1 (Test Approximate Identity). A test approximate

identity is a sequence of test functions

{m } = {1 , 2 , 3 , }

such that:
(1) Each m is nonnegative, i.e.,Reach m maps Rn to R0 .
(2) Each m has integral 1, i.e., m = 1 for each m.
(3) The supports of the m shrink to {0}, i.e.,

\
supp(1 ) supp(2 ) , supp(m ) = {0}.
m=1

We can construct a test approximate identity using the pulse function p

from section 7.2. Define for m = 1, 2, 3,

m : Rn R, m (x) = mn p(mx1 ) p(mx2 ) p(mxn ).

n n
Then supp(
R m ) = [1/m, 1/m] for each m. Here the coefficient m is chosen
so that m = 1 (exercise 7.4.1). Figure 7.8 shows the graphs of 2 , 4 , 8 ,
and 15 when n = 1. The first three graphs have the same vertical scale, but
not the fourth. Figure 7.9 shows the graphs of 1 through 4 when n = 2, all
having the same vertical scale.
 7.4 Test Approximate Identity and Convolution 369

8 8

-0.5 0.5 -0.25 0.25

8 15

-0.125 0.125
Figure 7.8. The functions 2 , 4 , 8 , and 15 from an approximate identity

16 16

-1
1
-
2
1
1 2

16 16

1 1
- -
3 1 4 1
3 4

Figure 7.9. The functions 1 through 4 from a two-dimensional approximate

identity

The identity being approximated by the sequence of test functions {m } is

the Dirac delta function from the chapter introduction, denoted . To repeat
ideas from the introduction, is conceptually a unit point mass at the origin,
and so its properties should be
370 7 Approximation by Smooth Functions
Z
supp() = {0}, = 1.

No such function exists in the orthodox sense of the word function. But re-
gardless of sense, for any function f : Rn R and any x Rn , the mollifying
kernel associated to f and ,

x (y) = f (y)(x y),

is conceptually a point of mass f (x) at each x. That is, its properties should
be Z
supp(x ) = {x}, (f )(x) = x (y) = f (x).
y

Under a generalized notion of function, the Dirac delta makes perfect sense as
an object called a distribution, defined by the integral in the previous display
but only for a limited class of functions:

For all x, (f )(x) = f (x) for test functions f .

Yes, now it is f that is restricted to be a test function. The reason for this is
that is not a test function, not being a function at all, and to get a good
theory of distributions such as , we need to restrict the functions that they
convolve with. In sum, the Dirac delta function is an identity in the sense that

f = f for test functions f .

Distribution theory is beyond the scope of these notes, but we may conceive of
the identity property of the Dirac delta function as the expected limiting be-
havior of any test approximate identity. That is, returning to the environment
of f Cc0 (Rn ) and taking any test approximate identity {m }, we expect that

lim(f m ) = f for Cc0 (Rn )-functions f .

m

As explained in section 7.1, this limit will be uniform, meaning that the values
(f m )(x) will converge to f (x) at one rate simultaneously for all x in Rn .
See exercise 7.4.3 for an example of nonuniform convergence.
For an example of convolution with elements of a test approximate identity,
consider the sawtooth function


|x| if |x| 1/4,
f : R R, f (x) = 1/2 |x| if 1/4 < |x| 1/2,


0 if 1/2 < |x|.

Recall the test approximate identity {m } from after Definition 7.4.1. Fig-
ure 7.10 shows f and its convolutions with 2 , 4 , 8 , and 15 . The convo-
lutions approach the original function while smoothing its corners, and the
 7.4 Test Approximate Identity and Convolution 371

Figure 7.10. The sawtooth function convolved with various m

convolutions are bounded by the bound on the original function as shown in

exercise 7.3.2. Also, the convolutions have larger supports than the original
function, but the supports shrink toward the original support as m grows.
The following lemma says that if compact sets shrink to a point, then
eventually they lie inside any given ball about the point. Specifically, the sets
that we have in mind are the supports of a test approximate identity.

Lemma 7.4.2 (Shrinking Sets Lemma). Let

{Sm } = {S1 , S2 , S3 , }

be a sequence of compact subsets of Rn such that

\
S1 S2 S3 , Sm = {0}.
m=1

Then for any > 0 there exists some positive integer m0 such that

for all m m0 , Sm B(0, ).

Proof. Let > 0 be given. If no Sm lies in B(0, ) then there exist points

x1 S1 \B(0, ),
x2 S2 \B(0, ),
x3 S3 \B(0, ),
372 7 Approximation by Smooth Functions

and so on. The sequence {xm } lies in S1 , so it has a convergent subsequence.

The containments S1 S2 show that replacing the sequence by the sub-
sequence preserves the displayed conditions, so we may assume that the origi-
nal sequence converges. Let x denote its limit. For any m 1, theT terms of the
sequence from index m onward lie in Sm , so x Sm . Thus x m Sm = {0},
i.e., x = 0. But also, |xm | for each m, so |x| . This is a contradiction,
so we are done.

The hypothesis of compactness is necessary in the Shrinking Sets Lemma
(exercise 7.4.2).
Theorem 7.4.3 (Cc0 (Rn )-Approximation by Convolutions). Consider a
function f Cc0 (Rn ) and let {m } : Rn R be a test approximate identity.
Given > 0, there exists a positive integer m0 such that for all integers m,

m m0 = |f m f | < .

That is, the convolutions f m converge uniformly to the original function f .

Proof. Let > 0 be given. Since the support of f is compact, f is uniformly
continuous on its support, and hence f is uniformly continuous on all of Rn .
So there exists some > 0 such that for all x, y Rn ,

|y x| < = |f (y) f (x)| < .

Since the supports of the approximate identity functions shrink to {0}, the
Shrinking Sets Lemma says that there exists some positive integer m0 such
that for all integers m m0 , supp(m ) B(0, ). Note that m0 depends only
on , which in turn depends only on , all of this with no reference to any
particular x Rn . Now, for all x, y Rn , and all m m0 ,

y x supp(m ) = y x B(0, ) = x + B(0, )

= |y x| <
= |f (y) f (x)| < .

Because the approximate identity functions m have integral 1, we have for

all x Rn and all positive integers m,
Z
f (x) = f (x)m (x y).
y

Use the fact that the approximate identity functions m are nonnegative to
estimate that for all x Rn and all positive integers m,
Z 
 
|(f m )(x) f (x)| =  (f (y) f (x))m (x y)
y
Z
|f (y) f (x)|m (x y).
y
 7.4 Test Approximate Identity and Convolution 373

We may integrate only over y-values in x supp(m ), so that if m m0 then

the integrand is less than m (x y). That is, since the approximate identity
functions have integral 1 we have for all x Rn and all positive integers m,
Z
m m0 = |(f m )(x) f (x)| < m (x y) = .
y

This is the desired result. Note how the argument has used all three defining
properties of the approximate identity.

Corollary 7.4.4 (Cck (Rn )-Approximation by Convolutions). Let k be a
positive integer. Consider a function f Cck (Rn ) and let {m } : Rn R
be a test approximate identity. Given > 0, there exists a positive integer m0
such that for all integers m,

m m0 = |f m f |k < .

That is, the convolutions and their derivatives converge uniformly to the orig-
inal function and its derivatives up to order k.
Proof. Recall from Corollary 7.3.7 that if f Cc1 (Rn ) then for any test func-
tion the derivative of the convolution is the convolution of the derivative,

Dj (f ) = Dj f , j = 1, , n.

Since the derivatives Dj f lie in Cc0 (Rn ), the theorem says that their convo-
lutions Dj f m converge uniformly to the derivatives Dj f as desired. The
argument for higher derivatives is the same.

Exercises
R
7.4.1. Recall that p = 1 where p : Rn R is the pulse function from
section 7.2. Let m be any positive integer and recall the definition in the
section,
m (x) = mn p(mx1 ) p(mx2 ) p(mxn )..
R
Explain why consequently m = 1.
7.4.2. Find a sequence {Sm } of subsets of R satisfying all of the hypotheses
of the Shrinking Sets Lemma except for compactness, and such that no Sm is
a subset of the interval B(0, 1) = (1, 1).
7.4.3. This exercise illustrates a nonuniform limit. For each positive integer m,
define
fm : [0, 1] R, fm (x) = xm .
Also define (
0 if 0 x < 1,
f : [0, 1] R, f (x) =
1 if x = 1.
374 7 Approximation by Smooth Functions

(a) Using one set of axes, graph f1 , f2 , f3 , f10 , and f .

(b) Show that for any x [0, 1], limm fm (x) = f (x). That is, given > 0,
there exists some positive integer m0 such that for all positive integers m,

m m0 = |fm (x) f (x)| < .

Thus the function f is the limit of the sequence of functions {fm }. That is:

For each x, f (x) = lim{fm (x)}.

m

(c) Now let = 1/2. Show that for any positive integer m, no matter how
large, there exists some corresponding x [0, 1] such that |fm (x) f (x)| .
That is:

For each m, |fm (x) f (x)| fails to be small for some x.

Thus the convergence of {fm } to f is not uniform, i.e., the functions do not
converge to the limit-function at one rate simultaneously for all x [0, 1].

7.5 Known-Integrable Functions

Recall that the slogan-title of Theorem 6.5.4 is near-continuity implies in-
tegrability. The largest space of functions that we have considered so far in
this chapter is Cc0 (Rn ), so we have not yet discussed the entire class of func-
tions that we know to be integrable. This section gives some results about
convolution and approximation for such functions.
Recall also that a function is called bounded if its outputs form a bounded
set.
Definition 7.5.1 (Known-Integrable Function). A function

f : Rn R.

is known-integrable if it is bounded, compactly supported, and continuous

except on a set of volume zero. The class of known-integrable functions is
denoted Ic (Rn ).
Unsurprisingly, the class Ic (Rn ) forms a vector space over R.
Let f Ic (Rn ). The integral of f is the integral of f over any box that
contains its support,
Z Z
f= f where supp(f ) B.
B

Similarly to the remarks after Definition 7.3.2, the integral on the right side
exists, but this time by Theorem 6.5.4. The integral on the right side is inde-
pendent of the box B, and so the integral on the left side exists, is unambigu-
ous, and is understood to be the integral of f over all of Rn .
 7.5 Known-Integrable Functions 375

The convolution remains sensible when f is known-integrable. That is, if

f Ic (Rn ) and Cc (Rn ) then for each x Rn the mollifying kernel

x : Rn R, x (y) = f (y)(x y)

again lies in Ic (Rn ). And so we may continue to define the convolution of f

and as Z
f : Rn R, (f )(x) = x (y).
y

The formulas for convolution derivatives remain valid as well. That is, if f
Ic (Rn ) and Cc (Rn ) then also f Cc (Rn ), and

Dj (f ) = f j , j = 1, , n,
Djj (f ) = f Djj j , j, j = 1, , n,

and so on. Here is where it is relevant that our proof of Proposition 7.3.5
required only that each x be integrable, that f be bounded, and that lie
in Cc1 (Rn ).
Given a known-integrable function f Ic (Rn ) and a test approximate
identity {m }, we would like the convolutions {f m } to approximate f uni-
formly as m grows. But the following proposition shows that this is impossible
when f has discontinuities.

Proposition 7.5.2 (The Uniform Limit of Continuous Functions is

Continuous). Let
{fm } : Rn R
be a sequence of continuous functions that converges uniformly to a limit func-
tion
f : Rn R.
Then f is continuous as well.

Rn and for any positive integer m we have

Proof. For any two points x, x

|f (
x) f (x)| = |f (
x) fm ( x) fm (x) + fm (x) f (x)|
x) + fm (
|f (
x) fm (
x)| + |fm (
x) fm (x)| + |fm (x) f (x)|.

Let > 0 be given. For all m large enough, the first and third terms are less
than /3 regardless of the values of x and x
. Fix such a value of m, and fix x.
Then since fm is continuous, the middle term is less than /3 if x is close
enough to x. It follows that

|f (
x) f (x)| < for all x
close enough to x.

That is, f is continuous.

376 7 Approximation by Smooth Functions

Thus the convergence property of convolutions must become more tech-

nical for known-integrable functions rather than compactly supported con-
tinuous functions. In preparation for proving the convergence property, the
following lemma says that if K is a compact subset of an open set then so is
the sum of K and some closed ball.
Lemma 7.5.3 (Thickening Lemma). Let K and A be subsets of Rn such
that
K A, K is compact, A is open.
Then
for some r > 0, K + B(0, r) A.
Proof. Since K is compact, it lies in some ball B(0, R). Solving the problem
with the open set A B(0, R) in place of A also solves the original problem.
Having replaced A by A B(0, R), define a function on K that takes
positive real values,
d : K R>0 , d(a) = sup{r : B(a, r) A}.
The fact that we have shrunk A (if necessary) to lie inside the ball has ensured
that d is finite because specifically d(a) R for all a. Fix some a K and let
r = d(a). Let {rm } be a strictly increasing sequence of positive real numbers
such that limm {rm } = r. Then B(a, rm ) A for each m, and so

[
B(a, r) = B(a, rm ) A.
m=1

This argument shows that in fact

d(a) = max{r : B(a, r) A}.
The function d is continuous. To see this, fix some point a K and let
r = d(a). Consider also a second point a K such that |
a a| < r, and let
r = d(
a). Then
a, r |
B( a a|) B(a, r) A,
showing that r r | a a|. Either r r + |
a a|, or r > r + |
a a| r so
that also |a a| < r and the same argument shows that r r | a a|, i.e.,
r r + |
a a| after all. That is, we have shown that for any a K,
( )
K
a
= |d( a) d(a)| |a a|.
|
a a| < r(a)

Thus d is continuous at a (given > 0, let = min{r(a), /2}), and since

a K is arbitrary, d is continuous on K as claimed.
Since K is compact and d is continuous, d takes a minimum value r > 0.
Thus K + B(0, r) A. Finally, let r = r/2. Then K + B(0, r) A as desired.


 7.5 Known-Integrable Functions 377

Now we can establish the convergence property of convolutions for known-

integrable functions.
Theorem 7.5.4 (Ic (Rn )-Approximation by Convolutions). Consider a
function f Ic (Rn ) and let {m } : Rn R be a test approximate identity.
Let K be a compact subset of Rn such that f is continuous on an open su-
perset of K. Given > 0, there exists a positive integer m0 such that for all
integers m,
m m0 = |(f m )(x) f (x)| < for all x K.
That is, the convolutions converge uniformly to the original function on com-
pact subsets of open sets where the function is continuous.
Proof. Let > 0 be given. By the Thickening Lemma, there exists some r > 0
such that f is continuous on K + B(0, r). Hence f is uniformly continuous, on
K + B(0, r). That is, there exists > 0 (with < r) such that for all x K
and all y Rn ,
|y x| < = |f (y) f (x)| < .
There exists some positive integer m0 such that for all integers m m0 ,
supp(m ) B(0, ). For all x K, all y Rn , and all m m0 ,
y x supp(m ) = y x B(0, ) = x + B(0, )
= |y x| <
= |f (y) f (x)| < .
From here, the proof is virtually identical to the proof of Theorem 7.4.3.

For example, consider the truncated squaring function
(
x2 if |x| 1/2,
f : R R, f (x) =
0 if 1/2 < |x|.

Note that f lies in Ic (Rn ) rather than in Cc0 (Rn ) because of its discontinuities
at x = 1/2. Figure 7.11 shows f and its convolutions with 2 , 4 , 8 ,
and 15 . The convolutions converge uniformly to the truncated parabola on
compact sets away from the two points of discontinuity. But the convergence
is not well behaved at or near those two points. Indeed, the function value
f (1/2) = 1/4 rather than f (1/2) = 0 is arbitrary and has no effect on
the convolution in any case. And again the convolutions are bounded by the
bound on the original function and their supports shrink toward the original
support as m grows.
In consequence of Ic (Rn )-approximation by convolutions, any integral of
a known-integrable function is approximated as closely as desired by the in-
tegral of a test function. Thus the hypothesis of a continuous integrand f in
the Change of Variable Theorem for multiple integrals (Theorem 6.7.1), men-
tioned in the last bullet of the chapter introduction, can now be weakened to
a known-integrable integrand.
378 7 Approximation by Smooth Functions

Figure 7.11. The truncated squaring function convolved with various m

7.6 Summary

Given a continuous, compactly supported function f , the convolutions of f and

the elements of any approximate identity are a sequence of smooth functions
that approximate f uniformly. If f is further assumed to have derivatives then
the corresponding derivatives of the convolutions approximate the derivatives
of f uniformly as well. On the other hand, if f is allowed to have a small
amount of discontinuity then the convolutions still approximate f uniformly
on compact sets where f is continuous.
8
Parametrized Curves

This chapter introduces parametrized curves as a warmup for chapter 9 to

follow. The subject of chapter 9 is integration over k-dimensional parametrized
surfaces in n-dimensional space, and the parametrized curves of this chapter
are the special case k = 1. Multivariable integration plays no role in this
chapter. Aside from being one-dimensional surfaces, parametrized curves are
interesting in their own right.
Section 8.1 leads into the subject of curves by introducing two specific
curves that solve problems beyond the capabilities of classical straightedge
and compass constructions. One striking idea here is the fact that by using
algebra to study geometry, we can describe precisely how the classical con-
structions are limited. Section 8.2 begins the study of parametrized curves,
meaning curves that we view not only as sets but as specified traversals of
the sets. Section 8.3 discusses the canonical parametrization of a curve by arc
length, the traversal at unit speed. Section 8.4 specializes the discussion to
curves in the plane. In this case, a local parameter called the curvature gives
a fairly complete description of curves in the large. Similarly, section 8.5 dis-
cusses curves in three-dimensional space. Here a second local parameter called
torsion is needed along with curvature to describe curves. Finally, section 8.6
generalizes the idea of describing a curve in optimal local coordinates to n
dimensions.

8.1 Euclidean Constructions and Two Curves

The straightedge constructs the line that passes through any two given points
in the Euclidean plane. The compass constructs the circle that is centered at
any given point and has any given distance as its radius. Any finite succession
of straightedge and compass constructions is called a Euclidean construction.
Physical straightedge and compass constructions are imprecise. Further-
more, there is really no such thing as a straightedge: aside from having to be
380 8 Parametrized Curves

infinite, the line-constructor somehow requires a prior line for its own con-
struction. But we dont concern ourselves with the details of actual tools for
drawing lines and circles. Instead we imagine the constructions to be ideal,
and we focus on the theoretical question of what Euclidean constructions can
or can not accomplish.
With computer graphics being a matter of course to us today, the techno-
logical power of Euclidean constructions, however idealized, is underwhelming,
and so one might reasonably wonder why they deserve study. One point of this
section is to use the study of Euclidean constructions to demonstrate the idea
of investigating the limitations of a technology. That is, mathematical reason-
ing of one sort (in this case, algebra) can determine the capacities of some some
other sort of mathematical technique (in this case, Euclidean constructions).
In a similar spirit, a subject called Galois theory uses the mathematics of fi-
nite group theory to determine the capacities of solving polynomial equations
by radicals.
In a high school geometry course one should learn that Euclidean con-
structions have the capacity to
bisect an angle,
bisect a segment,
draw the line through a given point and perpendicular to a given line,
and draw the line through a given point and parallel to a given line.
These constructions (exercise 8.1.1) will be taken for granted here.
Two classical problems of antiquity are trisecting the angle and doubling
the cube. This section will argue algebraically that neither of these problems
can be solved by Euclidean constructions, and then the second point of this
section is to introduce particular curvesand methods to generate them
that solve the classical problems where Euclidean constructions fail to do so.
Take any two points in the plane and denote them 0 and 1. Use the
straightedge to draw the line through them. We may as well take the line
to be horizontal with 1 appearing to the right of 0. Now define any real num-
ber r as Euclidean if we can locate it on our number line with a Euclidean
construction. For instance, it is clear how the compass constructs the integers
from 0 to any specified n, positive or negative, in finitely many steps. Thus
the integers are Euclidean. Further, we can add an orthogonal line through
any integer. Repeating the process on such orthogonal lines gives us as much
of the integer-coordinate grid as we want.
Proposition 8.1.1. The Euclidean numbers form a subfield of R. That is, 0
and 1 are Euclidean, and if r and s are Euclidean, then so are r s, rs, and
(if s 6= 0) r/s.
Proof. We have already constructed 0 and 1, and given any r and s it is
easy to construct r s. If s 6= 0 then the construction shown in figure 8.1
produces r/s. Finally, to construct rs when s 6= 0, first construct 1/s and
then rs = r/(1/s) is Euclidean as well.

 8.1 Euclidean Constructions and Two Curves 381
y

x
r/s r
Figure 8.1. Constructing r/s

Let E denote the field of Euclidean numbers. Since Q is the smallest sub-
field of R, it follows that Q E R. The questions are whether E is no more
than Q, whether E is all of R, andassuming that in fact E lies properly be-
tween Q and Rhow we can describe the elements of E. The next proposition
shows that E is a proper superfield of Q.

Proposition 8.1.2. If c 0 is constructible, i.e., if c E, then so is c.

Proof. In the construction shown in figure 8.2 we have a semicircle of radius

(c + 1)/2 centered at ((c + 1)/2, 0). This semicircle contains the point (1, y),
where p
y = ((c + 1)/2)2 ((c 1)/2)2 = c.
(Due to a tacit assumption in the figure, this proof isnt quite complete, but
see exercise 8.1.2.)

x
c+1
2 1

Figure 8.2. Constructing c
382 8 Parametrized Curves

Thus, any real numberq expressible in terms of finitely many square roots
p
starting from Q, such as 1 + 2 + 3, lies in E. Next we show that the
converse holds as well. That is, any number in E is expressible in finitely
square roots starting from Q.
Definition 8.1.3. Let F be any subfield of R. A point in F is a point (x, y)
in the plane whose coordinates x and y belong to F. A line in F is a line
through two points in F. A circle in F is a circle whose center is a point in F
and whose radius is a number in F.
Exercise 8.1.3 shows that any line in F has equation ax + by + c = 0 where
a, b, c F, and any circle in F has equation x2 + y 2 + ax + by + c = 0 with
a, b, c F.
Proposition 8.1.4. Let F be any subfield of R. Let L1 , L2 be any nonparallel
lines in F, and let C1 , C2 be distinct circles in F. Then
(1) L1 L2 is a point in F.
(2) C1 C2 is either empty or it is one or two points whose coordinates are
expressible in terms of F and a square root of a value in F.
(3) C1 L1 is either empty or it is one or two points whose coordinates are
expressible in terms of F and a square root of a value in F.
Proof. (1) is exercise 8.1.4(a).
(2) reduces to (3), for if the circles
C1 : x2 + y 2 + a1 x + b1 y + c1 = 0,
C 2 : x 2 + y 2 + a 2 x + b2 y + c 2 = 0
intersect, then C1 C2 = C1 L where L is the line
L : (a1 a2 )x + (b1 b2 )y + (c1 c2 ) = 0
(exercise 8.1.4(b)). Since C1 is a circle in F, the equations for C2 and L show
that C2 is a circle in F if and only if L is a line in F.
To prove (3), keep the equation for the circle C1 and suppose the line L1
has equation dx + ey + f = 0. The case d = 0 is exercise 8.1.4(c). Otherwise
we may take d = 1 after dividing through by d, an operation that keeps the
other coefficients in F. Thus x = ey f . Now, for (x, y) to lie in C1 L1 ,
we need
(ey f )2 + y 2 + a1 (ey f ) + b1 y + c1 = 0,
a condition of the form Ay 2 + By + C = 0 with A, B, C F. Solving for y
involves at most a square root over F, and then x = ey f involves only
further operations in F.

This result characterizes the field E of constructible numbers. Points in E
are obtained by intersecting lines and circles, starting with lines and circles
in Q. By the proposition, this means taking a succession of square roots. Thus:
 8.1 Euclidean Constructions and Two Curves 383

The field E is the set of numbers expressible in finitely many field and
square root operations starting from Q.
Now we can dispense with the two classical problems mentioned earlier.

Theorem 8.1.5. An angle of 60 degrees can not be trisected by straightedge

and compass.

Proof. If we could construct a 20-degree angle then we could construct the

number cos(20 ) (exercise 8.1.5(a)). From trigonometry,

cos(3) = 4 cos3 () 3 cos()

(exercise 8.1.5(b)), so in particular, cos(20 ) satisfies the cubic polynomial

relation
4x3 3x 1/2 = 0.
This cubic relation has no quadratic factors, so its root cos(20 ) is not con-
structible. (Strictly speaking, this last statement requires some algebraic jus-
tification, but at least it should feel plausible.)

Theorem 8.1.6. The side of a cube having volume 2 is not constructible.

Proof. Indeed, the side satisfies the relation x3 2 = 0, which again has no
quadratic factors.

Thinking algebraically had made certain complicated-seeming geometric

questions easy.
The second half of this section introduces curves to trisect the angle and
to double the cube. The first curve, the conchoid of Nicomedes, is defined
as follows. Fix a point O and a line L in the plane. For convenience, take
O = (0, 0) and L : {y = b} where b > 0 is constant. Fix a positive real
number d. For each point P R2 with y-coordinate bigger than b, let (O, P )
denote the line through O and P . The conchoid is then the set

{P R2 : (O, P ) meets L at a point distance d from P }.

(See figure 8.3.)

The conchoid can be organically generated, as shown in figure 8.4. The
lighter piece of the device in the figure swivels at the origin, and as it swivels
the tack at the lower end of the darker piece tracks the horizontal groove
at y = b. Thus the lighter piece slides along the length of the darker one, and
the pen at its upper end draws the conchoid.
The conchoid trisects angles, as shown in figure 8.5. Given an angle AOB
with AB OA, construct the conchoid with d = 2OB. (The conchoid in this
figure is rotated 90 degrees clockwise from those in the previous two figures.)
Then proceed as follows.
384 8 Parametrized Curves

x
O
Figure 8.3. A Conchoid

Figure 8.4. Organic generation of the conchoid

Let E be the point on the conchoid with the same y-coordinate as B.

Let C be the intersection point of AB and OE. Thus CE = 2 OB by our
choice of conchoid.
Let D be the midpoint of CE. Thus CD = DE = OB, and also BD is
the same length.
Let = AOC. Then also = BED and = EBD.
Let = BOD. Then also = BDO. The angle AOB that we want
to trisect equals + .
 8.1 Euclidean Constructions and Two Curves 385

So the other angle at D equals 2 since it is the remaining angle in

triangle BDE, but also it is visibly . Thus = 2.
The angle AOB = + that we want to trisect now equals 3, and so
is the desired trisection.

B E

D
C

x
O
A

Figure 8.5. Trisecting the angle with the conchoid

The cissoid of Diocles is defined as follows. Take a circle of radius a > 0

centered at (a, 0). Each ray emanating from the origin into the right half-
plane intersects the circle at a point C and intersects the circles right vertical
tangent at a point B. Let P denote the the point on the ray such that OP =
CB. The cissoid is the set of all such p (see figure 8.6).
Newton showed how to generate the cissoid organically (see figure 8.7).
As the tack at the end of the shorter piece of the device in the figure tracks
the vertical groove at x = a, the longer piece of the device slides past the
bumper at (a, 0). Consequently, the pen in the center of the shorter piece
draws the cissoid. Verifying that this construction indeed gives the cissoid is
exercise 8.1.6.
The cissoid doubles the cube. In the left half of figure 8.8, M is the mid-
point of the vertical segment from (1, 0) to (1, 1), so that the smaller right
triangle has height-to-base ratio 1/2. The line through (2, 0) and M meets the
point P on the cissoid, and the larger right triangle also has height-to-base
ratio 1/2. In the right side of the figure, the line through (0, 0) and P meets
386 8 Parametrized Curves
y

P
C

O x
a 2a

C
B

Figure 8.6. The cissoid

the circle, and the two horizontal distances labeled x are equal by the nature
of the cissoid. Continuing to work in the right half of the figure, the right
triangle with base x and height y is similar to the two other right triangles,
and the analysis of the left half of the figure has shown that the unlabeled
vertical segment in the right half has height (2 x)/2. Thus the similar right
triangles give the relations
y 2x y x
= and = .
x y x (2 x)/2

It follows that
y2 y 2
= 2 x and 2
= .
x x 2x
Multiply the two equalities to get
 y 3
= 2.
x
That is, multiplying the sides of a cube by y/x doubles the volume of the
cube, as desired.
 8.1 Euclidean Constructions and Two Curves 387

Figure 8.7. Organic generation of the cissoid

P
M y

1 x 2 2x x

Figure 8.8. Doubling the cube with the cissoid

Exercises

8.1.1. Show how straightedge and compass constructions bisect an angle, bi-
sect a segment, draw the line through point P perpendicular to line L, and
draw the line through point P parallel to line L.

8.1.2. What tacit assumption does the proof of Proposition 8.1.2 make
about c? Complete the proof for constructible c 0 not satisfying the as-
sumption.
388 8 Parametrized Curves

8.1.3. For any subfield F of R, show that any line in F has equation ax + by +
c = 0 with a, b, c F; show that any circle in F has equation x2 + y 2 + ax +
by + c = 0 with a, b, c F. Are the converses to these statements true? If the
line passes through the pointp in direction d, what are the relations between
p, d and a, b, c? If the circle has center p and radius r, what are the relations
between p, r and a, b, c?
8.1.4. (a) If L1 and L2 are nonparallel lines in F, show that L1 L2 is a point
with coordinates in F.
(b) If C1 and C2 are distinct intersecting circles in F with equations x2 +
2
y +a1 x+b1 y+c1 = 0 for C1 and similarly for C2 , show that C1 C2 is equal to
C1 L where L is the line with equation (a1 a2 )x + (b1 b2 )y + (c1 c2 ) = 0.
(c) Prove Proposition 8.1.4 part (3) when C1 is as in part (b) here and L1
has equation ey + f = 0 with e 6= 0.
8.1.5. (a) Suppose that the angle is constructible. Show that the number
cos is constructible as well.
3
(b) Equate the real parts of the equality ei3 = ei to establish the
trigonometry identity cos 3 = 4 cos3 3 cos .
8.1.6. Show that Newtons organic construction really does generate the cis-
soid.

8.2 Parametrized Curves

For our purposes a curve is not specified as a subset of Rn , but instead as a
traversal of such a set.
Definition 8.2.1 (Parametrized Curve). A parametrized curve is a
smooth mapping
: I Rn
where I R is a nonempty interval and n 1.
Here smooth means that the mapping has derivatives of all orders. A
small technicality is that the definition should, strictly speaking, insist that if
the interval I is not open then extends smoothly to some open superinterval
of I. We wont bother checking this in our examples.
The interval I in the definition is the parameter interval of . Any
point t I is a parameter value, and the corresponding point (t) Rn is
a point on the curve. Collectively, the set of points on the curve,
b = {(t) : t I},

is the trace of the curve. So the nomenclature point on the curve is a slight
abuse of language: a curve is a mapping and its trace is a set, and really
(t) is a point on the trace of the curve. But maintaining this distinction is
pointlessly pedantic. Also, since all of our curves will be parametrized, we will
refer to them simply as curves.
 8.2 Parametrized Curves 389

Definition 8.2.2 (Tangent Vector, Regular Curve). Let : I Rn be

a curve, and let t I. The tangent vector of at t is (t). The curve
is regular if its tangent vector (t) is nonzero for all t I.

It is often helpful to think of I as an interval of time, so that describes a

time-dependent motion through space. Thinking this way suggests some more
terminology.
The tangent vector (t) is also called the velocity vector of at t.
The scalar magnitude | (t)| of the velocity vector is the speed of at t.
Thus we may visualize (t) as a point and the velocity (t) as an arrow
emanating from the point in the direction of motion, the length of the arrow
being the speed of the motion. The definition of a regular curve rephrases as
the criterion that its time-dependent traversal never comes to a halt.

Definition 8.2.3 (Arc Length of a Curve). Let : I Rn be a curve,

and let t, t be points of I with t < t . The arc length of from t to t is
Z t
L(t, t ) = | ( )| d.
=t

In physical terms, this definition is a curvy version of the familiar idea that
distance equals speed times time. For a more purely mathematical definition
of a curves arc length, we should take the limit of the lengths of inscribed
polygonal paths. Take a partition t0 < t1 < < tn of the parameter interval
[t, t ], where t0 = t and tn = t . The partition determines the corresponding
points on the curve, (t0 ), (t1 ), . . . , (tn ). The arc length should be the
limit of the sums of the lengths of the line segments joining the points,
n
X
L(t, t ) = lim |(tk ) (tk1 )|.
n
k=1

It is possible to show that for smooth curvesin fact, for C 1 -curvesthe

limit exists and is equal to the integral definition of arc length. (The details
of the argument are too technical to deserve full explication here, but very
briefly: since the integral is conceptually the length of an inscribed polygon
with infinitely many sides, each infinitesimally long, and since the length of
an inscribed polygon increases when any of its segments is replaced by more
segments by adding more points of inscription, the definition of L(t, t ) as an
integral should be at least as big as the definition of L(t, t ) as a limit of sums,
and in fact this is easy to show. For the other inequality we need to argue
that the limit of sums gets as close to the integral as we wish. Since the sums
arent quite Riemann sums for the integral, this is where things get slightly
tricky.) Using the limit of sums as the definition of arc length is more general,
since it makes no reference to the smoothness of . A continuous curve for
which the arc length (defined as the limit of inscribed polygon lengths) is
390 8 Parametrized Curves

finite is called rectifiable. Perhaps surprisingly, not all continuous curves are
rectifiable. For that matter, the image of a continuous curve need not match
our intuition of a curve. For instance, there is a continuous mapping from the
closed interval [0, 1] to all of the square [0, 1] [0, 1], a so-called area-filling
curve. In any case, we will continue to assume that our curves are smooth,
and we will use the integral definition of arc length.
For example, the helix is the curve : R R3 where

(t) = (a cos t, a sin t, bt).

Here a > 0 and b > 0 are constants. (See figure 8.9.)

Figure 8.9. The helix

The velocity vector of the helix is

(t) = (a sin t, a cos t, b),

and so the speed is p

| (t)| = a 2 + b2 .

For another example, the cycloid is the curve made by a point on a rolling
wheel of radius 1. (See figure 8.10.) Its parametrization, in terms of the angle
through which the wheel has rolled, is

C() = ( sin , 1 cos ), 0 2.

 8.2 Parametrized Curves 391

Its velocity vector is

C () = (1 cos , sin ), 0 2,

and so its speed is

p
|C ()| = (1 cos )2 + (sin )2 = 2 2 cos
r q
1
= 4 (1 cos ) = 4 sin2 (/2)
2
= 2| sin(/2)|
= 2 sin(/2) since sin 0 on [0, ].

So the speed of the cycloid is greatestequal to 2when = , i.e., when

the point is at the top of the wheel. And it is leastequal to 0when = 0
and = 2. These results agree with what we see when we look at the wheel
of a moving bicycle: a blur at the top and distinct spokes at the bottom.

Figure 8.10. A rolling wheel

The length of the cycloid as the parameter varies from 0 to some angle
is
Z Z Z /2
L(0, ) = 2 sin(t/2) dt = 4 sin(t/2) d(t/2) = 4 sin( ) d
t=0 t=0 =0
= 4 4 cos(/2), 0 2.

In particular, a full arch of the cycloid has length 8.

The cycloid has amazing properties. Upside down, it is the brachis-
tochrone, the curve of steepest descent, meaning that it is the curve between
two given points along which a bead slides (without friction) most quickly.
Upside down, it is also the tautochrone, meaning that a bead starting from
any point slides (without friction) to the bottom in the same amount of time.
392 8 Parametrized Curves

For another property of the cycloid, suppose that a weight swings from a string
4 units long suspended at the origin, between two upside-down cycloids. The
rightmore upside-down cycloid is

C() = ( sin , cos 1), 0 2.

Thus the weights position when it is swinging to the right is (for 0 )

C ()
() = C() + (4 L(0, ))
|C ()|
(1 cos , sin )
= ( sin , cos 1) + 4 cos(/2)
2 sin(/2)
= ( sin , cos 1) + 2 cot(/2)(1 cos , sin ).

But since 0 we may carry out the following calculation, in which all
quantities under square root signs are nonnegative and so is the evaluation of
the square root at the last step,
q
1
cos(/2) 2 (1 + cos )
cot(/2) = =q
sin(/2) 1
(1 cos )
2
s r
(1 + cos )2 (1 + cos )2
= =
(1 cos )(1 + cos ) 1 cos2
1 + cos
= .
sin
And so now
1 + cos
() = ( sin , cos 1) + 2 (1 cos , sin )
sin
= ( sin , cos 1) + 2(sin , 1 cos )
= ( + sin , 3 cos ).

Shift the weights position rightward by and upward by 2,

() + (, 2) = ( + + sin , 1 cos ), 0 .

On the other hand, the right half of the original upside-down cycloid is

C( + ) = ( + sin( + ), cos( + ) 1)
= ( + + sin , 1 cos ), 0 2.

These are identical: () + (, 2) = C( + ) for 0 . That is, the

weight swings along the trace of a cycloid congruent to the two others. Since
the the upside-down cycloid is the tautochrone, this idea was used by Huygens
 8.2 Parametrized Curves 393

to attempt to design pendulum-clocks that would work on ships despite their

complicated motion.
The area under one arch of the cycloid is the integral
Z 2
y(x) dx
x=0

where y(x) is the function that takes the x-coordinate of a point of the cycloid
and returns its y-coordinate. As the cycloid parameter varies from 0 to 2,
so does the x-coordinate of the cycloid-point,

x = x() = sin ,

and the parametrization of the cycloid tells us that even without knowing y(x),
we know that
y(x()) = 1 cos .
Thus the area under one arch of the cycloid is
Z 2 Z 2 Z 2
y(x) dx = y(x())x () d = (1 cos )2 d,
x=0 =0 =0

and routine calculation shows that the area is 3.

A parametrization for the conchoid of Nicomedes is

: (/2, /2) R2 , () = (b sec + d)(cos , sin )

where now the line L is {x = b}, rotating the conchoid a quarter turn clockwise
from before, and where the parameter is the usual angle from the polar
coordinate system. Any point (x, y) on the conchoid satisfies the equation

(x2 + y 2 )(x b)2 = d2 x2 .

A parametrization for the cissoid of Diocles is

 
2 2at2 2at3
: R R , (t) = , .
1 + t2 1 + t2

where the parameter t is tan with being the usual angle from the polar
coordinate system.

Exercises

8.2.1. (a) Let : I Rn be a regular curve that doesnt pass through the
origin, but has a point (t0 ) of nearest approach to the origin. Show that the
position vector (t0 ) and the velocity vector (t0 ) are orthogonal. (Hint: If
u, v : I Rn are differentiable then hu, vi = hu , vi + hu, v ithis follows
394 8 Parametrized Curves

quickly from the one-variable product rule.) Does the result agree with your
geometric intuition?
(b) Find a regular curve : I Rn that does not pass through the
origin and does not have a point of nearest approach to the origin. Does an
example exist with I compact?

8.2.2. Let be a regular parametrized curve with (t) = 0 for all t I.

What is the nature of ?

8.2.3. Let : I Rn be a parametrized curve and let v Rn be a fixed

vector. Assume that h (t), vi = 0 for all t I and that h(to ), vi = 0 for
some t0 I. Prove that h(t), vi = 0 for all t I. What is the geometric
idea?

8.2.4. (a) Verify the parametrization of the conchoid given in the section.
(b) Verify the relation (x2 + y 2 )(x b)2 = d2 x2 satisfied by points on the
conchoid.

8.2.5. (a) Verify the parametrization of the cissoid given in the section. Is this
parametrization regular? What happens to (t) and (t) as t ?
(b) Verify Newtons organic generation of the cissoid.

8.3 Parametrization by Arc Length

Recall that the trace of a curve is the set of points on the curve. Thinking
of a curve as time-dependent traversal makes it clear that different curves
may well have the same trace. That is, different curves can describe different
motions along the same path. For example, the curves

: [0, 2] R2 , (t) = (cos t, sin t)

2
: [0, 2] R , (t) = (cos 5t, sin 5t)
2
: [0, 2] R , (t) = (cos t, sin t)
2
: [0, log(2 + 1)] R , (t) = (cos(et 1), sin(et 1))

all have the unit circle as their trace, but their traversals of the circle are
different: traverses it once counterclockwise at unit speed, traverses it five
times counterclockwise at speed 5, traverses it once clockwise at unit speed,
and traverses it once counterclockwise at increasing speed.
Among the four traversals, and are somehow basically the same, mov-
ing from the same starting point to the same ending point in the same direc-
tion, never stopping or backing up. The similarity suggests that we should be
able to modify one into the other. On the other hand, and seem essentially
different from and from each another. The following definition describes the
idea of adjusting a curve without changing its traversal in any essential way.
 8.3 Parametrization by Arc Length 395

Definition 8.3.1 (Equivalence of Curves). Two curves : I Rn and

: I Rn are equivalent, written

if there exists a mapping : I I , smooth with smooth inverse, with > 0

on I, such that
= .

For example, consider the mapping

: [0, 2] [0, log(2 + 1)], (s) = log(s + 1).

This mapping is differentiable and so is its inverse,

1 : [0, log(2 + 1)] [0, 2], 1 (t) = et 1.

Also, (s) = 1/(s + 1) is positive for all s I. Again recalling the examples
and , the calculation

( )(s) = (cos(elog(s+1) 1), sin(elog(s+1) 1)) = (cos s, sin s) = (s)

shows that as expected.

A similar calculation with 1 shows that also . This symmetry is a
particular instance of a general rule, whose proof is an exercise in formalism.

Proposition 8.3.2 (Properties of Equivalence). Let , , and be

curves. Then
(1) .
(2) If then .
(3) If and then .
In words, the relation is reflexive, symmetric, and transitive.

Among a family of equivalent regular curves, one is canonical: the curve

that traverses at unit speed. Recall that the arc length of a curve from t
to t is Z t
L(t, t ) = | ( )| d.
=t

Definition 8.3.3 (Parametrization by Arc Length). The curve is

parametrized by arc length if for all points in s, s I with s < s ,

L(s, s ) = s s.

Equivalently, is parametrized by arc length if | (s)| = 1 for all s I.

396 8 Parametrized Curves

As in the definition just given, we adopt the convention that a curve

parametrized by arc length is by default denoted rather than , and its
parameter denoted s rather than t.
To justify the intuition that every regular curve is equivalent to some curve
parametrized by arc length, we need two familiar theorems. The first version
of the one-variable Fundamental Theorem of Integral Calculus says,
Let f : [a, b] R be continuous. Define a function
Z x
F : [a, b] R, F (x) = f.
a

Then F is differentiable on [a, b] and F = f .
And the Inverse Function Theorem for One Variable says,
Let f : I R have a continuous derivative on I with f (x) 6= 0
for all x I. Then the image of f is an interval I , and f has a
differentiable inverse g : I I. For each y I , the derivative
of the inverse at y is given by the formula g (y) = 1/f (x) where
x = g(y).
These theorems let us reparametrize any regular curve by arc length.

Proposition 8.3.4. Any regular curve is equivalent to a curve parametrized

by arc length.

Proof. Let : I Rn be regular. Thus we are tacitly assuming that

is smooth, so that in particular is continuous. Pick any parameter value
t0 I and let p0 = (t0 ) be the corresponding point on the curve. Define the
arc length function : I R by the formula
Z t
(t) = | ( )| d.
=t0

By the Fundamental Theorem, is differentiable and (t) = | (t)|. Thus

is continuous and never vanishes, so by the Inverse Function Theorem has
a differentiable inverse 1 : I I for some interval I . Define a new curve
: I Rn by = 1 . Thus and are equivalent, and the following
diagram commutes (meaning that either path around the triangle yields the
same result):
I



'
Rn .
7



I
 8.3 Parametrization by Arc Length 397

For all t I, letting s = (t), the chain rule gives an equality of vectors,
(t) = ( ) (t) = (s) (t) = (s) | (t)|,
and then taking absolute values gives an equality of scalars,
| (t)| = | (s)| | (t)|.
Since | (t)| > 0 for all t because is regular, it follows that
| (s)| = 1 for all s I .
Thus is parametrized by arc length.

So every regular curve is equivalent to a curve parametrized by arc length.
The next question about a regular curve is whether its equivalent curve that
is parametrized by arc length is unique. The answer is: Essentially yes. The
only choice is the starting point, determined by the choice of t0 in the proof.
Explicitly reparametrizing by arc length can be a nuisance since it requires
computing the inverse function 1 that we invoked in the abstract during
the course of reparametrizing. (This function can be doubly hard to write
down in elementary terms, because not only is it an inverse function but
furthermore it is the inverse function of a forward function defined as an
integral.) Since the theory guarantees that each regular curve is equivalent to
a curve parametrized by arc length, when we prove theorems in the sequel
we may assume that we are given such curves. But on the other hand, since
reparametrizing is nontrivial computationally, we want the formulas that we
will derive later in the chapter not to assume parametrization by arc length,
so that we can apply them to regular curves in general.

Exercises
8.3.1. Show that the equivalence on curves is reflexive, symmetric, and
transitive.
8.3.2. The parametrized curve
: [0, +) R2 , (t) = (aebt cos t, aebt sin t)
(where a > 0 and b < 0 are real constants) is called a logarithmic spiral.
(a) Show that as t +, (t) spirals in toward the origin.
(b) Show that as t +, L(0, t) remains bounded. Thus the spiral has
finite length.
8.3.3. Explicitly reparametrize each curve : I Rn with a curve :
I Rn parametrized by arc length.
(a) The ray : R>0 Rn given by (t) = t2 v where v is some fixed
nonzero vector.
(b) The circle : R R2 given by (t) = (cos et , sin et ).
(c) The helix : [0, 2] R3 given by (t) = (a cos t, a sin t, bt).
(d) The cycloid : [/2, 3/2] R2 given by (t) = (t sin t, 1 cos t).
398 8 Parametrized Curves

8.4 Plane Curves: Curvature

Let : I R2 be a plane curve parametrized by arc length s. We next spec-
ify a natural coordinate system at each point of . Its tangent vector T (s)
is
T = .
So to first order the curve is moving in the T -direction. Its normal vec-
tor N (s) is the 90-degree counterclockwise rotation of T (s). Thus the Frenet
frame {T, N } is a positive basis of R2 consisting of orthogonal unit vectors.
Before proceeding, we need to establish two handy little facts that hold in any
dimension n.
Lemma 8.4.1. (a) Let v : I Rn be a smooth mapping such that |v(t)| = c
(where c is constant) for all t. Then
hv, v i = 0.
(b) Let v, w : I Rn be smooth mappings such that hv(t), w(t)i = c (where
c is constant) for all t. Then
hw , vi = hv , wi.

Proof. (a) hv, v i = 21 hv, vi = 0. (b) hw , vi + hv , wi = hv, wi = 0.

We return to dimension n = 2. Part (a) of the lemma shows that the
derivative of the tangent vector is some scalar multiple of the normal vector,
T = N, = (s) R.
The scalar-valued function (s) is the curvature of . The curvature can be
positive, negative, or zero depending on whether to second order the curve is
bending counterclockwise toward N , clockwise away from N , or not at all.
In particular, if r is a positive real number then the curve
(s) = r(cos(s/r), sin(s/r)), sR
is a circle of radius r parametrized by arc length. The tangent vector T (s) =
(s) and its derivative T (s) = (s) are
1 1
T = ( sin(s/r), cos(s/r)), T =
( cos(s/r), sin(s/r)) = N,
r r
showing that the curvature of the circle is the reciprocal of its radius,
1
(s) =.
r
In general, the absolute curvature of a curve is the reciprocal of the radius of
the best-fitting circle to . We omit the proof of this.
Plausibly, if we are told only that is some curve parametrized by arc
length, that begins at some point p0 with initial tangent vector T0 , and that
has curvature function (s), then we can reproduce the curve . This is true
but beyond our scope to show here. Nonetheless it explains that:
 8.4 Plane Curves: Curvature 399

The combination of a set of initial conditions and the local information

of the curvature at each point of a curve is enough to recover the curve
itself, a global object.
Hence the local informationthe curvatureis of interest.
To see how the Frenet frame continually adjusts itself as is traversed, we
differentiate T and N . Since these are orthogonal unit vectors, their derivatives
resolve nicely into components via the inner product,

T = hT , T iT + hT , N iN
N = hN , T iT + hN , N iN.

The condition T = N shows that the top row inner products are hT , T i = 0
and hT , N i = . Since N is a unit vector hN , N i = 0 by part (a) of the lemma,
and since T and N are orthogonal hN , T i = hT , N i = by part (b). Thus
the Frenet equations for a curve parametrized by arc length are
    
T 0 T
= .
N 0 N

The geometric idea is that as we move along the curve at unit speed, the
Frenet frame continually adjusts itself so that its first vector is tangent to
the curve in the direction of motion and the second vector is ninety degrees
counterclockwise to the first. The curvature is the rate (positive, negative, or
zero) at which the first vector is bending toward the second, while the second
vector preserves the ninety-degree angle between them by bending away from
the first vector as much as the first vector is bending toward it.
Since = T and thus = T , the first and second derivatives of any
curve parametrized by arc length are expressed in terms of the Frenet frame,
    
10 T
= .
0 N

This matrix relation shows that the local canonical form of a such a curve is,
up to quadratic order,
1
(s0 + s) (s0 ) + s (s0 ) + s2 (s0 )
2

= (s0 ) + sT + s2 N.
2
That is, in (T, N )-coordinates the curve is locally (s, (/2)s2 ), a parabola at
the origin that opens upward or downward or not at all, depending on .
If we view the curve in local coordinates as we traverse its length at unit
speed, we see the parabola change its shape as varies, possibly narrowing
and widening, or opening to a horizontal line and then bending the other way.
400 8 Parametrized Curves

This periscope-view of , along with knowing (s) and (s) for one value s
in the parameter domain, determine entirely.
We want a curvature formula for any regular smooth plane curve, not
necessarily parametrized by arc length,

: I R2 .

To derive the formula, recall that the reparametrization of by arc length is

the curve characterized by a relation involving the arc length function of ,
Z t
= , where (t) = | | and so = | |.

By the Chain Rule, and then by the Product Rule and again the Chain Rule,

= ( ) ,
= ( ) + ( ) ( )2 .

These relations and the earlier expressions of and in terms of the Frenet
frame combine to give
        
0 0 10 T
= 2 = 2 .
0 N

Take determinants, recalling that = | |,

det( , ) = | |3 .

Thus the curvature is

det( , ) x y x y
= =
3/2 ( = (x, y) regular).
| |3 (x )2 + (y )2
In particular, if a curve is parametrized by arc length then its curvature in
coordinates is

= det( , ) = x y x y ( = (x, y) parametrized by arc length).

The fact that a plane curve lies on a circle if and only if its curvature
is constant cries out to be true. (If it isnt, then our definitions must be
misguided.) And it is easy to prove using global coordinates. However, we
prove it by working with the Frenet frame, in anticipation of the less obvious
result for space curves to follow in the next section.
Proposition 8.4.2. Let : I R2 be regular. Then

lies on a circle (s) is a nonzero constant for all s I.

When these conditions hold, || = 1/ where > 0 is the radius of the circle.
 8.4 Plane Curves: Curvature 401

Proof. We may assume that is parametrized by arc length.

( = ) We will zoom in on the global condition that lies on a circle,
differentiating repeatedly and using the Frenet frame as our coordinate sys-
tem. In the argument, and its derivatives depend on the parameter s, and
so does the curvature , but we omit s from the notation in order to keep the
presentation light. We are given that for some fixed point p R2 and some
fixed radius > 0,
| p| = .
And by the nature of the Frenet frame, p decomposes as

p = h p, T iT + h p, N iN. (8.1)

Since | p| is constant, Lemma 8.4.1(a) gives h p, i = 0, and now

Lemma 8.4.1(b) gives hp, i = h , i = 1. Since = T and = N ,
these calculations have shown that h p, T i = 0 and h p, N i = 1/ with
6= 0. Thus (8.1) is simply

p = (1/)N.

But since |p| = , it follows that 1/ = is constant, and so is constant,

as desired.
( = ) Assume that (s) is a nonzero constant. To show that p =
(1/)N , compute (using the Frenet equation N = T ) the derivative

( + (1/)N ) = T + (1/)(T ) = 0.

So +(1/)N is indeed some fixed vector p, and p = (1/)N as expected.

It follows that lies on the circle of radius 1/|| centered at p.

Since N = (1/) the previous proof has shown that the differential
equation
p = (1/)2
arises from uniform circular motion of radius 1/||.

Exercises

8.4.1. (a) Let a and b be positive. Find the curvature of the ellipse (t) =
(a cos(t), b sin(t)) for t R.
(b) Let a be positive and b be negative. Find the curvature of the loga-
rithmic spiral (t) = (aebt cos t, aebt sin t) for t 0.

8.4.2. Let : I R2 be parametrized by arc length. Fix any unit vector

v R2 , and define a function

: I R
402 8 Parametrized Curves

by the conditions

cos((s)) = hT (s), vi, sin((s)) = hN (s), vi.

Thus is the angle that the curve makes with the fixed direction v. Show
that = . Thus our notion of curvature does indeed measure the rate at
which is turning.

8.5 Space Curves: Curvature and Torsion

Now we discuss space curves similarly to the discussion of plane curves at the
end of the previous section. Let : I R3 be parametrized by arc length s.
Its tangent vector T (s) is
T = .
So to first order the curve is moving in the T -direction. Whenever T is
nonzero, the curves curvature (s) and normal vector N (s) are defined
by the conditions
T = N, > 0.
(Be aware that although the same equation T = N appeared in the context
of plane curves, something different is happening now. For plane curves, N
was defined as the 90-degree counterclockwise rotation of T , and the condi-
tion hT, T i = 1 forced T to be normal to T and hence some scalar multiple
of N . The scalar was then given the name , and could be positive, neg-
ative, or zero depending on whether to second order the curve was bending
toward N , away from N , or not at all. But now, for space curves, the con-
ditions T = N and > 0 define both N and , assuming that T 6= 0.
Again by Lemma 8.4.1(a) T is normal to T , and so N is normal to T , but
now it makes no sense to speak of N being counterclockwise to T , and now
is positive.) Assume that T is always nonzero. Then the curves binormal
vector is
B = T N.
Thus, the Frenet frame {T, N, B} is a positive basis of R3 consisting of or-
thogonal unit vectors.
We want to differentiate T , N , and B. The derivatives resolve into com-
ponents,

T = hT , T iT + hT , N iN + hT , BiB
N = hN , T iT + hN , N iN + hN , BiB
B = hB , T iT + hB , N iN + hB , BiB.

The definition
T = N
 8.5 Space Curves: Curvature and Torsion 403

shows that the top row inner products are

hT , T i = 0, hT , N i = , hT , Bi = 0.

And since N and B are unit vectors, the other two diagonal inner products
also vanish by Lemma 8.4.1(a),

hN , N i = hB , Bi = 0.

Lemma 8.4.1(b) shows that the first inner product of the second row is the
negative of the second inner product of the first row,

hN , T i = hT , N i = ,

and so only the third inner product of the second row is a new quantity,

N = T + B for the scalar function = hN , Bi.

The function is the torsion of . It can be positive, negative, or zero,

depending on whether to third order the curve is twisting out of the (T, N )-
plane toward B, away from B, or not at all.. Similarly, the first and second
inner products of the third row are the negatives of the third inner products
of the first and second rows,

hB , T i = hT , Bi = 0, hB , N i = hN , Bi = .

All of the derivatives computed so far gather into the Frenet equations,

T 0 0 T
N = 0 N .
B 0 0 B

The geometric idea is that as we move along the curve, the bending of the
first natural coordinate determines the second natural coordinate; the second
natural coordinate bends away from the first as much as the first is bending
toward it, in order to preserve the ninety-degree angle between them; the
remaining bending of the second coordinate is toward or away from the third
remaining orthogonal coordinate, which bends away from or toward from the
second coordinate at the same rate, in order to preserve the ninety-degree
angle between them.
The relations = T and = T = N and = (N ) = N + N ,
and the second Frenet equation N = T + B combine to show that

1 0 0 T
= 0 0 N .
2 B

This relation shows that the local canonical form of a such a curve is, up to
third order,
404 8 Parametrized Curves

1 1
(s0 + s) (s0 ) + s (s0 ) + s2 (s0 ) + s3 (s0 )
2 6
1 1
= (s0 ) + sT + s2 N + s3 (2 T + N + B)
 2  6  
2 2 3 3
= (s0 ) + s s3 T + s + s N+ s B.
6 2 6 6
In planar cross sections,
In the (T, N )-plane the curve is locally (s, (/2)s2 ), a parabola opening
upward at the origin (see figure 8.11, viewing the curve down the positive
B-axis).
In the (T, B)-plane the curve is locally (s, ( /6)s3 ), a cubic curve inflect-
ing at the origin, rising from left to right if > 0 and falling if < 0 (see
figure 8.12, viewing the figure up the negative N -axis).
In the (N, B)-plane the curve is locally ((/2)s2 , ( /6)s3 ), a curve in the
right half plane with a cusp at the origin (see figure 8.13, viewing the curve
down the positive T -axis).
The relation of the curve to all three local coordinate axes is shown in fig-
ure 8.14.

Figure 8.11. Space curve in local coordinates, from above

Let be a regular curve, not necessarily parametrized by arc length. We

want formulas for its curvature and torsion. Let be the reparametrization
of by arc length, so that = where is the arc length of . By the
chain rule,
= ( ) ,
= ( ) + ( )2 ,
= ( ) + 3( ) + ( )3 .
 8.5 Space Curves: Curvature and Torsion 405
B

Figure 8.12. Space curve in local coordinates, from the side

Figure 8.13. Space curve in local coordinates, from the front

These relations and the earlier expressions of and in terms of the Frenet
frame combine to give

0 0 0 0 1 0 0 T
= 2 0 = 2 0 0 0 N .
3 3 3 3 2 B
2
Thus = T (T + N ) = 3 B, and since = | | this gives the
curvature,
406 8 Parametrized Curves

Figure 8.14. Space curve in local coordinates

| |
= .
| |3
Similarly, det( , , ) = 6 2 , giving the torsion,

det( , , )
= .
| |2

As mentioned, the counterpart of Proposition 8.4.2 for space curves is

considerably less obvious. Local measurements answer the global question of
whether we are moving on a sphere.
Theorem 8.5.1. Let : I R3 be regular, with curvature and torsion func-
tions , never zero. Consider the reciprocal curvature and torsion functions,

r = 1/, t = 1/.

Assume also that never vanishes. Then

lies on a sphere r2 + (r t)2 is constant.

When these conditions hold, r2 + (r t)2 = 2 where > 0 is the radius of the
sphere.
Proof. We may assume that is parametrized by arc length.
( = ) As in the proof of Proposition 8.4.2, we zoom in on the global
condition that lies on a sphere, differentiating repeatedly and using the
 8.5 Space Curves: Curvature and Torsion 407

Frenet frame. We are given that for some fixed point p R3 and some fixed
radius > 0,
| p| = .
And by the nature of the Frenet frame, p decomposes as
p = h p, T iT + h p, N iN + h p, BiB. (8.2)
Since | p| is constant, Lemma 8.4.1(a) gives h p, i = 0; next,
Lemma 8.4.1(b) and the fact that is parametrized by arc length com-
bine to give h p, i = h , i = 1; and now Lemma 8.4.1(b) and
then Lemma 8.4.1(a) (again using the parametrization by arc length) give
h p, i = h , i = 0. Since = T and = N and =
2 T + N + B, the first two calculations have shown that h p, T i = 0
and h p, N i = 1/ with 6= 0, and the third one therefore has shown
that
0 = h p, 2 T + N + Bi = / + h p, Bi,
from which h p, Bi = /(2 ). Thus the description (8.2) of p in the
Frenet frame is
p = (1/)N + /(2 )B.
Because we have defined r = 1/, so that r = /2 , and t = 1/ ,
p = rN r tB.
And thus r2 + (r t)2 = 2 .
( = ) We expect that = p rN r tB. So let = + rN + r tB and
compute using the Frenet equations and various other results,
= T + r N + r(T + B) + (r t + r t )B r t N
= (1 r)T + (r r t )N + (r + r t + r t )B
r 
= + r t + r t B.
t
2 2
But r + (r t) is constant, so its derivative is zero,
r 
0 = 2rr + 2r t(r t + r t ) = 2r t + r t + r t .
t
Thus = 0 (here is where we use the hypothesis that never vanishes:
it prevents r from vanishing) and so indeed is some fixed vector p. Thus
= p rN + r tB as expected, and | p|2 is the constant r2 + (r t)2 .

Exercise
8.5.1. (a) Let a and b be positive. Compute the curvature and the torsion
of the helix (t) = (a cos t, a sin t, bt).
(b) How do and behave if a is held constant and b ?
(c) How do and behave if a is held constant and b 0?
(d) How do and behave if b is held constant and a ?
(e) How do and behave if b is held constant and a 0?
408 8 Parametrized Curves

8.6 General Frenet Frames and Curvatures

This section extends the Frenet frame to any number of dimensions. As with
plane curves and space curves, the basic idea is to take the derivatives of the
curve and straighten them out, giving rise to a coordinate system of orthogonal
unit vectors where each new direction takes one more derivativeand hence
one more degree of the curves behaviorinto account. The result is a local
coordinate system natural to the curve at each of its points.
Let n 2 be an integer, and let

: I Rn

be a regular curve. For any t I define the first Frenet vector of at t to

be (suppressing the t from the notation for brevity)

F1 = /| |.

Thus F1 is a unit vector pointing in the same direction as the tangent vector
of at t.
Assuming that F1 never vanishes and that n 3, next define the first
curvature 1 (t) of at t and the second Frenet vector F2 (t) of at t by the
conditions
F1 = 1 F2 , 1 > 0, |F2 | = 1.
Since |F1 | = 1 for all t, it follows from Lemma 8.4.1(a) that hF2 , F1 i = 0.
Because hF2 , F1 i = 0, Lemma 8.4.1(b) gives hF2 , F1 i = hF1 , F2 i = 1 .
Assuming that F2 + 1 F1 never vanishes and that n 4, define the second
curvature 2 (t) and the third Frenet vector F3 (t) by the conditions

F2 = 1 F1 + 2 F3 , 2 > 0, |F3 | = 1.

Then hF3 , F1 i = 0 because 1 F1 is the F1 -component of F2 . Again by

Lemma 8.4.1(a), hF2 , F2 i = 0, so that (since also hF1 , F2 i = 0) hF3 , F2 i = 0.
In general, suppose that 2 k n 2, and suppose that we have t-
dependent orthogonal unit Frenet vectors F1 , . . . , Fk and t-dependent posi-
tive curvature functions 1 , . . . , k1 such that (defining 0 F0 = 0 for con-
venience)
Fj = j1 Fj1 + j Fj+1 , j = 1, , k 1.
Since hFk , Fk i = 1, it follows by Lemma 8.4.1(a) that hFk , Fk i = 0. And since
hFk , Fj i = 0 for j = 1, , k 1, it follows by Lemma 8.4.1(b) that for such j,

hFk , Fj i = hFj , Fk i
= hj1 Fj1 + j Fj+1 , Fk i
(
0 if j = 1, , k 2,
=
k1 if j = k 1.
 8.6 General Frenet Frames and Curvatures 409

So, assuming that Fk 6= k1 Fk1 , define k and Fk+1 by the conditions

Fk = k1 Fk1 + k Fk+1 , k > 0, |Fk+1 | = 1.

Then the relation k Fk+1 = Fk + k1 Fk1 shows that hFk+1 , Fj i = 0 for j =

1, , k. Use this process, assuming the nonvanishing that is needed, until
defining n2 and Fn1 . Thus if n = 2 then the process consists only of
defining F1 , if n = 3 then the process also defines 1 and F2 , if n = 4 then
the process further defines 2 and F3 , and so on.
Finally, define the nth Frenet vector Fn as the unique unit vector orthog-
onal to F1 through Fn1 such that det(F1 , F2 , , Fn ) > 0, and then define
the (n 1)st curvature n1 by the condition

Fn1 = n2 Fn2 + n1 Fn .

The (n 1)st curvature need not be positive. By Lemma 8.4.1(b) yet again
we have Fn = n1 Fn1 , and so the Frenet equations are

F1 0 1 F1
F2 1 0 2 F2

F3 2 0 3 F3

.. .. .. .. ..
. = . . . . .

. .. .. .. .
.. . . . .
.
F n2 0 n1 Fn1
n1

Fn n1 0 Fn

The first n 1 Frenet vectors and the first n 2 curvatures can also be
obtained by applying the Gram-Schmidt process (see exercise 2.2.16) to the
vectors , , (n1) .
The Frenet vectors and the curvatures are independent of parametrization.
: Ie Rn be a second curve equivalent to . That is,
To see this, let

where : I Ie is smooth and has a smooth inverse, and > 0 on I. By

the Chain Rule,

(t) =
(t) (t) where t = (t).

(t) point in the same direction (because (t) > 0) and so

Thus (t) and
the corresponding first Frenet vectors are equal,

F1 (t) = Fe1 (t).

Since the curvatures and the rest of the Frenet vectors are described in terms
of derivatives of the first Frenet vector with respect to its variable, it follows
410 8 Parametrized Curves

that the Frenet vectors and the curvatures are independent of parametrization
as claimed,
Fei (t) = Fi (t) for i = 1, , n
and
i (t) = i (t)
for i = 1, , n 1.

Since the curvatures describe the curve in local terms, they should be
unaffected by passing the curve through a rigid motion. The remainder of this
section establishes this invariance property of curvature, partly because doing
so provides us an excuse to describe the rigid motions of Euclidean space.

Definition 8.6.1. The square matrix A Mn (R) is orthogonal if At A = I.

That is, A is orthogonal if A is invertible and its transpose is its inverse. The
set of n-by-n orthogonal matrices is denoted On (R).

It is straightforward to check (exercise 8.6.2) that

the identity matrix I is orthogonal,
if A and B are orthogonal then so is the product AB,
and if A is orthogonal then so is the inverse A1 .
These three facts, along with the fact that matrix multiplication is associative,
show that the orthogonal matrices form a group under matrix multiplication.
Some examples of orthogonal matrices are
   
1 0 cos sin
, for any R.
0 1 sin cos

Orthogonal matrices are characterized by the property that they preserve

inner products. That is, the following equivalence holds:

A On (R) hAx, Ayi = hx, yi for all x, y Rn

(Exercise 8.6.2(a)). Consequently, multiplying vectors by an orthogonal ma-

trix A preserves their lengths and the angles between them. Also, if A Mn (R)
is orthogonal then det A = 1 (Exercise 8.6.2(b)).
The orthogonal matrices of determinant 1 form the special orthogonal
group, denoted SOn (R). These matrices not only preserve length and angle,
but in addition they preserve orientation. Thus
(
hAx, Ayi = hx, yi, x, y Rn ,
A SOn (R)
det(Ax1 , , Axn ) = det(x1 , , xn ), x1 , , xn Rn .

It is straightforward to check that SOn (R) forms a subgroup of On (R).

Definition 8.6.2. A bijective mapping R : Rn Rn is called rigid if

hR(x) R(p), R(y) R(p)i = hx p, y pi for all p, x, y Rn .

 8.6 General Frenet Frames and Curvatures 411

That is, rigid maps preserve the geometry of vector differences. The next
proposition characterizes rigid mappings.

Proposition 8.6.3. The mapping R : Rn Rn is rigid if and only if R

takes the form R(x) = Ax + b with A On (R) and b Rn .

Proof. Verifying that any mapping R(x) = Ax + b where A On (R) and

b Rn is rigid is exercise 8.6.3.
Now let R : Rn Rn be rigid. Define a related mapping

S : Rn Rn , S(x) = R(x) R(0).

It suffices to show that S(x) = Ax for some A On (R). A small calculation

shows that S preserves inner products: for any x, y Rn ,

hS(x), S(y)i = hR(x) R(0), R(y) R(0)i = hx 0, y 0i = hx, yi.

Especially, if {e1 , . . . , en } is the standard basis of Rn then {S(e1 ), . . . , S(en )}

is again an orthonormal basis of Rn . Furthermore, hS(x), S(ei )i = hx, ei i for
any x Rn and for i = 1, . . . , n. That is,

S(x1 , . . . , xn ) = x1 S(e1 ) + + xn S(en ).

This shows that S(x) = Ax where A has columns S(e1 ), . . . , S(en ). Since
hS(ei ), S(ej )i = hei , ej i for i, j {1, . . . , n}, in fact A On (R) as desired.

Definition 8.6.4. A congruence is a rigid map R(x) = Ax + b where A is

special orthogonal.

With congruences understood, it is easy to show that they preserve cur-

vatures. Consider a regular curve

: I Rn ,

let R(x) = Ax + b be a congruence,and define a second curve

: I Rn ,
= R .

Then for any t I,

(t) = R ((t)) (t) = A (t).

Thus the first Frenet vectors of the two curves satisfy the relation

Fe1 = AF1 ,

and similarly for their derivatives,

1 Fe2 = Fe1 = (AF1 ) = AF1 = A1 F2 = 1 AF2 ,

412 8 Parametrized Curves

1 and 1 are positive and |Fe2 | = 1 = |F2 | = |AF2 |,

so that since

1 = 1 and Fe2 = AF2 .

Similarly,

i = i , i = 1, , n 1
and
Fei = AFi , i = 1, , n.
We need A to be special orthogonal rather than just orthogonal in order that
this argument apply to the last Frenet vector and the last curvature. If A is
orthogonal but not special orthogonal then Fen = AFn and en1 = n1 .

Exercises

8.6.1. Are the following matrices orthogonal?

  1 0 2  
cos sin 1 ab
, 0 5 0, .
sin cos 5 2 0 1 0d

(b) Confirm that the identity matrix I is orthogonal, that if A and B are
orthogonal then so is the product AB, and that if A is orthogonal then so is
its inverse A1 .

8.6.2. (a) Prove that any matrix A Mn (R) is orthogonal if and only if
hAx, Ayi = hx, yi for all x, y Rn . (The fact that hv, wi = v t w essentially
gives ( = ). For ( = ), show that At A has (i, j)th entry hAei , Aej i for
i, j = 1, , n, and recall that In is the matrix whose (i, j)th entry is ij .)
(b) Prove that any matrix A On (R) has determinant det A = 1.

8.6.3. Prove that any mapping R(x) = Ax + b where A On (R) and b Rn

is rigid.

8.7 Summary

(To be written.)
9
Integration of Differential Forms

The integration of differential forms over surfaces is characteristic of a fully

developed mathematical theory: it starts from carefully preconfigured defini-
tions and proceeds to one central theorem, whose proof is purely mechanical
because of how the definitions are rigged. Furthermore, much of the work is
algebraic, even though the theorem appears analytical. Since the motivation
for the definitions is not immediately obvious, the early stages of working
through such a body of material can feel unenlightening, but the payoff lies in
the lucidity of the later arguments and the power of the end result. The main
theorem here is often called Stokess Theorem, but in fact it is a generaliza-
tion not only of the classical Stokess Theorem (which is not due to Stokes; he
just liked to put it on his exams), but also of other nineteenth century results
called the Divergence Theorem (or Gauss Theorem) and Greens Theorem,
and even of the Fundamental Theorem of Integral Calculus. In fact, a better
name for the theorem to be presented here is the General FTIC.
The definitions of a surface and of the integral of a function over a surface
are given in section 9.1. Formulas for particular integrals called flow and flux
integrals are derived in section 9.2. The theory to follow is designed partly to
handle such integrals easily. The definitions of a differential form and of the
integral of a differential form over a surface are given in section 9.3, and the
definitions are illustrated by examples in sections 9.4 and 9.5. Sections 9.6
through 9.9 explain the algebraic rules of how to add differential forms and
multiply them by scalars, how to multiply differential forms, how to differen-
tiate them, and how to pass them through changes of variable. A Change of
Variable Theorem for differential forms follows automatically in section 9.10.
A construction of antiderivatives of forms is given in section 9.11. Returning to
surfaces, sections 9.12 and 9.13 define a special class of surfaces called cubes,
and a geometric boundary operator from cubes to cubes of lower dimension.
The General FTIC is proved in section 9.14. Section 9.15 sketches how it
leads to another proof of the classical Change of Variable Theorem. Finally,
section 9.16 explains how the classical vector integration theorems are special
414 9 Integration of Differential Forms

cases of the General FTIC, and section 9.17 takes a closer look at some of the
quantities that arise in this context.

9.1 Integration of Functions Over Surfaces

Having studied integration over solid regions in Rn , i.e., over subsets of Rn

with positive n-dimensional volume, we face the new problem of how to inte-
grate over surfaces of lower dimension in Rn . For example the circle in R2 is
one-dimensional, and the torus surface in R3 is two-dimensional. Each of these
sets has volume zero as a subset of its ambient space, in which it is curving
around. In general, whatever the yet-undefined notion of a k-dimensional sub-
set of Rn means, such objects will have volume zero when k < n, and so any
attempt to integrate over them in the sense of chapter 6 will give an integral
of zero and a dull state of affairs. Instead, the idea is to parametrize surfaces
in Rn and then define integration over a parametrized surface in terms of
integration over a non-curved parameter space.

Definition 9.1.1 (Parametrized Surface). Let A be an open subset of Rn .

A k-surface in A is a smooth mapping

: D A,

where D is a compact connected subset of Rk whose boundary has volume zero.

The set D is called the parameter domain of .

See figure 9.1. Here are some points to note about Definition 9.1.1:
Recall that a subset A of Rn is called open if its complement is closed.
The definitions in this chapter need the environment of an open subset
rather than all of Rn in order to allow for functions that are not defined
everywhere. For instance, the reciprocal modulus function

1/| | : Rn {0} R

is defined only on surfaces that avoid the origin. In most of the examples, A
will be all of Rn , but exercise 9.11.1 will touch on how the subject becomes
more nuanced when it is not.
Recall also that compact means closed and bounded. Connected means
that D consists of only one piece, as discussed informally in section 2.4.
And as discussed informally in section 6.5 and formally in section 6.8, the
boundary of a set consists of all points simultaneously near the set and
near its complementroughly speaking, its edge. Typically D will be some
region that is easy to integrate over, such as a box, whose compactness,
connectedness, and small boundary are self-evident.
 9.1 Integration of Functions Over Surfaces 415

The word smooth in the definition means that the mapping extends
to some open superset of D in Rk , on which it has continuous partial
derivatives of all orders. Each such partial derivative is therefore again
smooth. All mappings in this chapter are assumed to be smooth.
When we compute, coordinates in parameter space will usually be written
as (u1 , , uk ), and coordinates in Rn as (x1 , , xn ).
It may be disconcerting that a surface is by definition a mapping rather
than a set, but this is for good reason. Just as the integration of chapter 6
was facilitated by distinguishing between functions and their outputs, the
integration of this chapter is facilitated by viewing the surfaces over which
we integrate as mappings rather than their images.
A parametrized curve, as in Definition 8.2.1, is precisely a 1-surface.

z
v y

u

Figure 9.1. A surface

When k = 0, Definition 9.1.1 is a little tricky. By convention, R0 is the

set of all points with no coordinates, each of the no coordinates being a real
number. (Our definition of Rn at the beginning of chapter 2 danced around
this issue by requiring that n be positive.) There is exactly one such point,
the point (). That is, R0 consists of a single point, naturally called 0 even
though it is not (0). A 0-surface in Rn is thus a mapping

p : R0 Rn , p (0) = p,

where p is some point in Rn . In other words, p simply parametrizes the

point p. At the other dimensional extreme, if k = n then any compact con-
416 9 Integration of Differential Forms

nected subset D of Rn naturally defines a corresponding n-surface in Rn by

trivially parametrizing itself,

: D Rn , (u) = u for all u D.

Thus Definition 9.1.1 of a surface as a mapping is silly in the particular cases

of k = 0 and k = n, when it amounts to parametrizing points by using the
empty point as a parameter domain, or parametrizing solids by taking them
to be their own parameter domains and having the identity mapping map
them to themselves. But for intermediate values of k, i.e., 0 < k < n, we
are going to integrate over k-dimensional subsets of Rn by traversing them,
and parametrizing is the natural way to do so. Especially, a 1-surface is a
parametrized curve, and a 2-surface is a parametrized surface in the usual
sense of surface as in figure 9.1.
Let A be an open subset of Rn , let : D A be a k-surface in A, and
let f : A R be a smooth function. As mentioned above, if k < n then
the integral of f over (D) in the sense of chapter 6 is zero because (D)
is lower-dimensional than its ambient space Rn . However, the integral of f
over can be defined more insightfully.
For each point u of the parameter domain D, the n-by-k derivative matrix
(u) has as its columns vectors that are naturally viewed as tangent vectors
to at (u), the jth column being tangent to the curve in that arises from
motion in the jth direction of the parameter domain. In symbols, the matrix
is  
(u) = v1 vk nk ,
where each column vector vj is

Dj 1 (u)
..
vj = Dj (u) = . .
Dj n (u) n1

The parallelepiped spanned by these vectors (see figure 9.2) has a naturally
defined k-dimensional volume.
Definition 9.1.2 (Volume of a Parallelepiped). Let v1 , , vk be vectors
in Rn . Let V be the n-by-k matrix with these vectors as its columns. Then the
k-volume of the parallelepiped spanned by the {vj } is
q
volk (P(v1 , , vk )) = det(V T V ) . (9.1)

In coordinates, this formula is

q

volk (P(v1 , , vk )) = det [vi vj ]i,j=1, ,k , (9.2)

where vi vj is the inner product of vi and vj .

 9.1 Integration of Functions Over Surfaces 417

v
z

x
u

Figure 9.2. Tangent parallelepiped

The matrix V in this definition is n-by-k and its transpose V T is k-by-n, so

neither of them need be square. But the product V T V is square, k-by-k, and
this is the matrix whose determinant is being taken. Equation (9.2) follows
immediately from (9.1) because
T
v1 v1 v1 v1 vk
  .. = [v v ]
V T V = ... v1 vk = ... ..
. . i j i,j=1, ,k .
vkT vk v1 vk vk
For example, if k = 1 and : [a, b] Rn is a 1-surface (i.e., a curve)
in Rn , then its derivative matrix at a point u of [a, b] has one column,

1 (u)

(u) = ... .
n (u)
Consequently, formula (9.2) is
p
length( (u)) = (u) (u).

That is, Definition 9.1.2 for k = 1 specializes to the definition of | | as
from section 2.2. (Here and throughout the chapter, we drop the notational
convention that curves named are parametrized by arc length; thus no as-
sumption is present that | | = 1.) At the other extreme, if k = n then
formula (9.1) is
voln (P(v1 , , vn )) = | det(v1 , , vn )|.
That is, Definition 9.1.2 for k = n recovers the interpretation of | det | as
volume from section 3.8. When k = 2, formula (9.2) is
418 9 Integration of Differential Forms
p
area(P(v1 , v2 )) = |v1 |2 |v2 |2 (v1 v2 )2
p
= |v1 |2 |v2 |2 (1 cos2 12 )
= |v1 | |v2 | | sin 12 |,

giving the familiar formula for the area of a parallelogram. When k = 2 and
also n = 3, we can study the formula further by working in coordinates.
Consider two vectors u = (xu , yu , zu ) and v = (xv , yv , zv ). An elementary
calculation shows that the quantity under the square root in the previous
display works out to

|u|2 |v|2 (u v)2 = |u v|2 .

So when k = 2 and n = 3, Definition 9.1.2 subsumes the familiar formula

area(P(v1 , v2 )) = |v1 v2 |.

Here is an argument that (9.2) is the appropriate formula for the k-

dimensional volume of the parallelepiped spanned by the vectors v1 , , vk
in Rn . (The fact that the vectors are tangent vectors to a k-surface is irrele-
vant to this discussion.) Results from linear algebra guarantee that there exist
vectors vk+1 , , vn in Rn such that
each of vk+1 through vn is a unit vector orthogonal to all the other vj ,
det(v1 , , vn ) 0.
Recall the notation in Definition 9.1.2 that V is the n-by-k matrix with
columns v1 , . . . , vk . Augment V to an n-by-n matrix W by adding the re-
maining vj as columns too,
   
W = v1 vn = V vk+1 vn .

The scalar det(W ) is the n-dimensional volume of the parallelepiped spanned

by v1 , . . . , vn . But by the properties of vk+1 through vn , this scalar should
also be the k-dimensional volume of the the parallelepiped spanned by v1 , . . . ,
vk . That is, the natural definition is (using the second property of v1 , . . . , vn
for the second equality to follow)
p q
volk (P(v1 , , vk )) = det(W ) = (det W )2 = det(W T ) det(W )
q
= det(W T W ).

The first property of v1 , . . . , vn shows that

 
V T V 0k(nk)
WT W = ,
0(nk)k Ink

so that det(W T W ) = det(V T V ) and the natural definition becomes the de-
sired formula,
 9.1 Integration of Functions Over Surfaces 419
q
volk (P(v1 , , vk )) = det(V T V ).
The argument here generalizes the ideas used in section 3.10 to suggest a
formula for the area of a 2-dimensional parallelogram in R3 as a 3-by-3 deter-
minant. Thus the coordinate calculation sketched in the previous paragraph
to recover the relation between parallelogram area and cross product length
in R3 was unnecessary.
With k-dimensional volume in hand, we can naturally define the integral
of a function over a k-surface.
Definition 9.1.3 (Integral of a Function over a Surface). Let A be an
open subset of Rn . Let : D A be a k-surface in A. Let f : (D) R
be a function such that f is smooth. Then the integral of f over is
Z Z
f= (f ) volk (P(D1 , , Dk )).
D

In particular, the k-dimensional volume of is

Z Z
volk () = 1= volk (P(D1 , , Dk )).
D

By Definition 9.1.2 the k-volume factor in the surface integral is

q q
volk (P(D1 , , Dk )) = det( T ) = det([Di Dj ]i,j=1, ,k ) .

The idea of Definition 9.1.3 is that as a parameter u traverses the parameter

domain D, the composition f samples the function f over the surface,
while the k-volume factor makes the integral the limit of sums of many f -
weighted small tangent parallelepiped k-volumes over the surface rather than
the limit of sums of many (f )-weighted small box volumes over the pa-
rameter domain. (See figure 9.3.) The k-volume factor itself is not small, as
seen in figure 9.2, but it is the ratio of the small parallelepiped k-volume to
the small box volume shown in figure 9.3.
For example, let r be a positive real number and consider a 2-surface in R3 ,

: [0, 2] [0, ] R3 , (, ) = (r cos sin , r sin sin , r cos ).

This surface is the 2-sphere of radius r. Since the sphere is a surface of revo-
lution, its area is readily computed by methods from a first calculus course,
but we do so with the ideas of this section to demonstrate their use. The
derivative vectors are

r sin sin r cos cos
v1 = r cos sin , v2 = r sin cos ,
0 r sin

and so the integrand of the surface area integral is

420 9 Integration of Differential Forms

v z

u
x

Figure 9.3. Integrating over a surface

p q
|v1 |2 |v 2 |2 (v1 v2 )2 = r4 sin2 = r2 sin

(note that sin 0 since [0, ]). Therefore the area is

Z 2 Z
area() = r2 sin = 4r2 .
=0 =0

The fact that the sphere-area magnification factor r2 sin is the familiar vol-
ume magnification factor for spherical coordinates is clear geometrically: to
traverse the sphere, the spherical coordinates and vary while r stays con-
stant, and when r does vary it moves orthogonally to the sphere-surface so
that the incremental volume is the incremental surface-area times the incre-
mental radius-change. Indeed, the vectors v1 and v2 from a few displays back
are simply the second and third columns of the spherical change of variable
derivative matrix. The reader can enjoy checking that the first column of the
spherical change of variable derivative matrix is indeed a unit vector orthog-
onal to the second and third columns.
The integral in Definition 9.1.3 seems to depend on the surface as a
parametrization rather than merely as a set, but in fact the integral is un-
affected by reasonable changes of parametrization, because of the Change
of Variable Theorem. To see this, let A be an open subset of Rn , and let
: D A and : D e A be k-surfaces in A. Suppose that there exists
a smoothly invertible mapping T : D D e such that T = . In other
words, T is smooth, T is invertible, its inverse is also smooth, and the follow-
ing diagram commutes (meaning that either path around the triangle yields
the same result):
 9.1 Integration of Functions Over Surfaces 421

D



&
T 8A
qqqqq
qqq
 qqqq
Deq
When such a mapping T exists, is called a reparametrization of .
Let f : A R be any smooth function. Then the integral of f over the
reparametrization of is
Z q
(f ) det( T ).
e
D

By the Change of Variable Theorem, since D e = T (D), this integral is

Z q

(f T ) det ( T )T ( T ) | det(T )|.
D

p q
But | det(T )| = det(T )2 = det(T T ) det(T ), so this becomes
Z q

(f T ) det T T ( T )T ( T ) T ,
D

and by the general matrix rule B T AT AB = (AB)T AB, this is in turn

Z q

(f T ) det (( T )T )T ( T )T .
D

Finally, since T = , the Chain Rule shows that we have

Z q

(f ) det T ,
D

giving the integral of f over the original surface as desired.

Exercises

9.1.1. Consider two vectors u = (xu , yu , zu ) and v = (xv , yv , zv ). Calculate

that |u|2 |v|2 (u v)2 = |u v|2 .

9.1.2. Consider two vectors u = (xu , yu , zu ) and v = (xv , yv , zv ). Calculate

that the area of the parallelogram spanned by u and v is the square root
of the sum of the squares of the areas of the parallelograms shadows in the
(x, y)-plane, the (y, z)-plane, and the (z, x)-plane.
422 9 Integration of Differential Forms

9.1.3. Let f (x, y, z) = x2 + yz.

(a) Integrate f over the box B = [0, 1]3 .
(b) Integrate f over the parametrized curve

: [0, 2] R3 , (t) = (cos t, sin t, t).

(c) Integrate f over the parametrized surface

S : [0, 1]2 R3 , S(u, v) = (u + v, u v, v).

(d) Integrate f over the parametrized solid

V : [0, 1]3 R3 , V (u, v, w) = (u + v, v w, u + w).

9.1.4. Find the surface area of the upper half of the cone at fixed angle from
the z-axis, extended outward to radius a. That is, the surface is the image of
the spherical coordinate mapping with fixed at some value between 0 and
as varies from 0 to a and varies from 0 to 2.

9.1.5. (a) Let D Rk be a parameter domain, and let f : D R be a

smooth function. Recall from exercise 2.4.3 that the graph of f is a subset
of Rk+1 ,
G(f ) = {(u, f (u)) : u D}.
Note that f is a 1-surface in R, while the surface that captures the idea of the
graph of f as a k-surface in Rk+1 is not f itself but rather

: D Rk+1 , (u) = (u, f (u)).

Derive a formula for the k-dimensional volume of . In particular, show that

when k = 2 the formula is
Z p
area() = 1 + (D1 f )2 + (D2 f )2 .
D

(b) What is the area of the graph of the function f : D R (where D

is the unit disk in the plane) given by f (x, y) = x2 + y 2 ?

9.2 Flow and Flux Integrals

Let A be an open subset of Rn . A mapping F : A Rn is also called a
vector field on A. (The usage of field here is unrelated to the field axioms.) If
: I A is a curve in A and u is a point of I, then the flow of F along
at u is the scalar component of F ((u)) tangent to at (u). If : D A is
an (n1)-surface in A and u is a point of D, then the flux of F through at u
is the scalar component of F normal to at (u). Surface integrals involving
the flow or the flux of a vector field arise naturally. If F is viewed as a force
 9.2 Flow and Flux Integrals 423

field then its flow integrals, also called line integrals, measure the work of
moving along curves in A. If F is viewed as a velocity field describing the
motion of some fluid then its flux integrals measure the rate at which fluid
passes through permeable membranes in A. Each of the classical theorems
of vector integral calculus to be proved at the end of this chapter involves a
flow integral or a flux integral.
Flow and flux integrals have a more convenient form than the general
integral of a function over a surface, in that the k-volume factor from Defini-
tion 9.1.3 (an unpleasant square root) cancels, and what remains is naturally
expressed in terms of determinants of the derivatives of the component func-
tions of . These formulas rapidly become complicated, so the point of this
section is only to see what form they take.
Working first in two dimensions, consider a vector field,

F = (F1 , F2 ) : R2 R2 ,

and a curve,
= (1 , 2 ) : [a, b] R2 .
Assuming that the derivative is always nonzero but not assuming that
is parametrized by arc length, the unit tangent vector to at the point (u),
pointing in the direction of the traversal, is
(u)
Tb((u)) = .
| (u)|
Note that the denominator is the length factor in Definition 9.1.3. The parallel
component of F ((u)) along Tb((u)) has magnitude (F Tb)((u)). (See exer-
cise 2.2.15.) Therefore the net flow of F along in the direction of traversal
R
is F Tb. By Definition 9.1.3 this flow integral is
Z Z b Z b
(u)
F Tb = F ((u)) | (u)| = F ((u)) (u), (9.3)
u=a | (u)| u=a

and the length factor has canceled. In coordinates, the flow integral is
Z Z b


F Tb = (F1 )1 + (F2 )2 (u). (9.4)
u=a

On the other hand, for any vector (x, y) R2 , define (x, y) = (y, x).
(This seemingly ad hoc procedure of negating one of the vector entries and
then exchanging them will be revisited soon as a particular manifestation of
a general idea.) The unit normal vector to the curve at the point (u), at
angle /2 counterclockwise from Tb((u)), is

b ((u)) = (u) .
N
| (u)|
424 9 Integration of Differential Forms

Therefore the net flux of F through counterclockwise to the direction of

traversal is the flux integral
Z Z b
b
F N = F ((u)) (u) , (9.5)
u=a

or, in coordinates,
Z Z b

b=
F N (F2 )1 (F1 )2 (u). (9.6)
u=a

Next let n = 3 and modify the vector field F suitably to

F = (F1 , F2 , F3 ) : R3 R3 .

The intrinsic expression (9.3) for the flow integral of F along a curve remains
unchanged in R3 , making the 3-dimensional counterpart of (9.4) in coordinates
obvious,
Z Z b


F Tb = (F1 )1 + (F2 )2 + (F3 )3 (u).
u=a

As for the flux integral, consider a 2-surface in R3 ,

= (1 , 2 , 3 ) : D R3 .

Assuming that the two columns D1 and D2 of the derivative matrix are
always linearly independent, a unit normal to the surface at the point (u)
(where now u = (u1 , u2 )) is obtained from their cross product,

b ((u)) = D1 (u) D2 (u) .

N
|D1 (u) D2 (u)|
By property CP6 of the cross product, the denominator in this expression is
the area of the parallelogram spanned by D1 (u) and D2 (u), and this is
the area factor in Definition 9.1.3 of the surface integral. Therefore this factor
cancels in the flux integral of F through in the N b -direction,
Z Z
F Nb= F ((u)) (D1 (u) D2 (u)), (9.7)
uD

or, in coordinates,

Z Z (F1 )(D1 2 D2 3 D1 3 D2 2 )
b=
F N +(F2 )(D1 3 D2 1 D1 1 D2 3 ) (u). (9.8)
uD
+(F3 )(D1 1 D2 2 D1 2 D2 1 )

Whereas the 2-dimensional flow and flux integrands and the 3-dimensional
flow integrand involved derivatives j of the 1-surface , the integrand here
 9.2 Flow and Flux Integrals 425

contains the determinants of all 2-by-2 subblocks of the 3-by-2 derivative

matrix of the 2-surface ,

D1 1 D2 1
= D1 2 D2 2 .
D1 3 D2 3

The subdeterminants give a hint about the general picture. Nonetheless, (9.8)
is foreboding enough that we should pause and think before trying to compute
more formulas.
For general n, formula (9.3) for the flow integral of a vector field along a
curve generalizes transparently,
Z Z b Z b Xn



b
F T =
(F ) (u) = (Fi )i (u). (9.9)
u=a u=a i=1

But the generalization of formulas (9.5) through (9.8) to a formula for the flux
integral of a vector field in Rn through an (n 1)-surface is not so obvious.
Based on (9.7) the intrinsic formula should be
Z Z


F N b= (F ) (D1 Dn1 ) (u), (9.10)
uD

where the (n 1)-fold cross product on Rn is analogous to the 2-fold cross

product on R3 from section 3.10. That is, the cross product should be orthog-
onal to each of the multiplicand-vectors, its length should be their (n 1)-
dimensional volume, and when the multiplicands are linearly independent,
they should combine with their cross product to form a positive basis of Rn .
Such a cross product exists by methods virtually identical to section 3.10.
What is special to three dimensions is that the cross product is binary, i.e., it
is a twofold product. In coordinates, a mnemonic formula for the cross product
in R3 , viewing the vectors as columns, is

e1
v1 v2 = det v1 v2 e2 .
e3
This formula appeared in row form in section 3.10, and it makes the corre-
sponding formula for the cross product of n 1 vectors in Rn inevitable,

e1
.. .
v1 vn1 = det v v
1 n1. (9.11)
en

For example, in R2 a single vector v = (x, y) has a sort of cross product,

 
x e1
v = det = (y, x)
y e2
426 9 Integration of Differential Forms

This is the formula that appeared with no explanation as part of the flux
integral in R2 . That is, the generalization (9.10) of the 3-dimensional flux
integral to higher dimensions also subsumes the 2-dimensional case. Returning
to Rn , the cross product of the vectors D1 (u),. . . ,Dn1 (u) is

e1
.
(D1 Dn1 )(u) = det D (u) D 1 (u) .. .
n1
en

This determinant can be understood better by considering the data in the

matrix as rows. Recall that for i = 1, , n, the ith row of the n-by-(n 1)
derivative matrix is the derivative matrix of the ith component function
of ,  
i (u) = D1 i (u) Dn1 i (u) .
In terms of these component function derivatives, the general cross product is

1 (u) e1 e1 1 (u)
.. = (1)n1 det .. ..
(D1 Dn1 )(u) = det ... . . .
n (u) en en n (u)

2 (u) 1 (u) 1 (u)
3 (u) 3 (u) 2 (u)


4 (u) 4 (u)
= (1)n1 det e1 det e2 + det 4 (u) e3 +
.. .. ..
. . .
n (u) n (u) n (u)

1 (u)
..

n .
X i1 (u)
= (1)n1 (1)i1 det
(u) ei .
i=1 i+1
..
.
n (u)

Thus finally, the general flux integral in coordinates is

1
..
.
Z Z n
X i1

b = (1)n1
F N (1) (Fi ) det
i1
(u). (9.12)
uD i=1 i+1
.
..
n

The integrand here contains the determinants of all (n1)-by-(n1) subblocks

of the n-by-(n 1) derivative matrix of the (n 1)-surface . The best way
 9.2 Flow and Flux Integrals 427

to understand the notation of (9.12) is to derive (9.6) and (9.8) from it by

setting n = 2 and then n = 3.
We end the section by mentioning one more integral. Let k = 2 and let
n = 4, and consider a 2-surface in R4 ,

= (1 , 2 , 3 , 4 ) : D R4 .

Note that is a 4-by-2 matrix,

1 D1 1 D2 1
D2 2
= 2 D1 2 ,
3 = D1 3 D2 3
4 D1 4 D2 4

so that any two of its rows form a square matrix. Consider also any six smooth
functions
F1,2 , F1,3 , F1,4 , F2,3 , F2,4 , F3,4 : R4 R.
Then we can define an integral,
     
1 1
Z (F1,2 ) det + (F1,3 ) det + (F1,4 ) det 1
2 3 4
      (u).
2 2 3
uD
+(F2,3 ) det + (F2,4 ) det + (F3,4 ) det
3 4 4
(9.13)
Since the surface is not 1-dimensional, this is not a flow integral. And since
is not (n 1)-dimensional, it is not a flux integral either. Nonetheless, since
the integrand contains the determinants of all 2-by-2 subblocks of the 4-by-2
derivative matrix of the 2-surface , it is clearly cut from the same cloth as
the flow and flux integrands of this section. The ideas of this chapter will
encompass this integral and many others in the same vein.
As promised at the beginning of the section, the k-volume factor has can-
celed in flow and flux integrals, and the remaining integrand features determi-
nants of the derivatives of the component functions of the surface of integra-
tion. Rather than analyze such cluttered integrals, the method of this chapter
is to abstract their key properties into symbol-patterns, and then work with
the patterns algebraically instead. An analysis tracking all the details of the
original setup would be excruciating to follow, not to mention being unimag-
inable to recreate ourselves. Instead, we will work insightfully, economy of
ideas leading to ease of execution. Since the definitions to follow do indeed
distill the essence of vector integration, they will enable us to think fluently
about the phenomena that we encounter. This is real progress in methodol-
ogy, much less laborious than the classical approach. Indeed, having seen the
modern argument, it is unimaginable to want to recreate the older one.
428 9 Integration of Differential Forms

Exercises

9.2.1. Show that the n-dimensional cross product defined by a formula

in (9.11) satisfies the property

hv1 vn1 , wi = det(v1 , , vn1 , w) for all w Rn .

As in section 3.10, this property characterizes the cross product uniquely.

Are there significant differences between deriving the properties of the cross
product from its characterization (cf. Proposition 3.10.2) in n dimensions
rather than in 3?

9.2.2. Derive equations (9.6) and (9.8) from equation (9.12).

9.3 Differential Forms Syntactically and Operationally

We need objects to integrate over surfaces, objects whose integrals encompass

at least the general flow integral (9.9) and flux integral (9.12) of the previous
section. Let A be an open subset of Rn . The objects are called differential
forms of order k on A or simply k-forms on A. Thus a k-form is some
sort of mapping
: {k-surfaces in A} R.
R
Naturally the value () will be denoted . The definition of a k-form will
come in two parts. The first is syntactic: it doesnt say what a k-form is as
a function of k-surfaces, only what kind of name a k-form can have. This
definition requires some preliminary vocabulary: An ordered k-tuple from
{1, , n} is a vector

(i1 , , ik ) with each ij {1, , n}.

For example, the ordered 3-tuples from {1, 2} are

(1, 1, 1), (1, 1, 2), (1, 2, 1), (1, 2, 2), (2, 1, 1), (2, 1, 2), (2, 2, 1), (2, 2, 2).

A sum over the ordered k-tuples from {1, , n} means simply a sum of terms
with each term corresponding to a distinct k-tuple. Thus we may think of an
ordered k-tuple (i1 , , ik ) as a sort of multiple index or multiple subscript,
and for this reason we often will abbreviate it to I. These multiple subscripts
will figure prominently throughout the chapter, so you should get comfortable
with them. Exercise 9.3.1 provides some practice.

Definition 9.3.1 (Syntax of Differential Forms). Let A be an open subset

of Rn . A 0-form on A is a smooth function f : A R. For k 1, a k-form
on A is an element of the form
 9.3 Differential Forms Syntactically and Operationally 429
n
X
f(i1 , ,ik ) dxi1 dxik ,
i1 , ,ik =1

or X
fI dxI ,
I

where each I = (i1 , , ik ) is an ordered k-tuple from {1, , n} and each fI

is a smooth function fI : A R.
Make the convention that the empty set I = is the only ordered 0-tuple
from {1, , n}, and that the corresponding empty product dx is 1. Then
the definition of a k-form for k 1 in Definition 9.3.1 also makes sense for
k = 0, and it subsumes the special definition that was given for k = 0.
For example, a differential form for n = 3 and k = 1 is
ex+y+z dx + sin(yz) dy + x2 z dz,
and a differential form for n = 2 and k = 2 is
y dx dx + ex dx dy + y cos x dy dx,
with the missing dy dy term tacitly understood to have the zero function as
its coefficient-function f(2,2) (x, y), and hence to be zero itself. The expression
1
dx
x
is a 1-form on the open subset A = {x R : x 6= 0} of R, but it is not a
1-form on all of R. The hybrid expression
z dx dy + ex dz
is not a differential form because it mixes an order 2 term and an order 1
term.
Before completing the definition of differential form, we need one more
piece of terminology. If M is an n-by-k matrix and I = (i1 , , ik ) is an
ordered k-tuple from {1, , n}, then MI denotes the square k-by-k matrix
comprising the Ith rows of M . For example, if

12
M = 3 4 ,
56
and if I = (3, 1), then  
56
MI = .
12
The second part of the definition of a k-form explains how to integrate it over
a k-surface. In this definition, a differential form in the sense of Definition 9.3.1
is called a syntactic differential form.
430 9 Integration of Differential Forms

Definition 9.3.2 (Integration of Differential Forms). Let A be an open

subset of Rn . For k = 0, a syntactic 0-form = f on A gives rise to a
function of 0-surfaces in A, also called ,

: {0-surfaces in A} R,

defined by the rule that for any point p A,

(p ) = f (p).

That is, integrating over a one-point surface consists

P simply of evaluating f
at the point. For k 1, a syntactic k-form = I fI dxI on A gives rise to
a function of k-surfaces in A, also called ,

: {k-surfaces in A} R,

defined by the rule that for any k-surface : D A,

Z X
() = (fI ) det I . (9.14)
D I

For all k, the integral of over is defined to be (),

Z
= ().

Formula (9.14), defining (), is the key for everything to follow in this
chapter. It defines an integral over the image (D), which may have volume
zero in Rn , by pulling backthis term will later be defined preciselyto an
integral over the parameter domain D, which is a full-dimensional set in Rk
and hence has positive k-dimensional volume.
Under Definition 9.3.2, the integral of a differential form over a surface
depends on the surface as a mapping, i.e., as a parametrization. However, it
is a straightforward exercise to show that that the Multivariable Change of
Variable Theorem implies that the integral is unaffected by reasonable changes
of parametrization.
Returning to formula (9.14): despite looking like the flux integral (9.12),
it may initially be impenetrable to the reader who (like the author) does not
assimilate notation quickly. The next two sections will illustrate the formula
in specific instances, after which its general workings should be clear. Before
long, you will have an operational understanding of the definition.
Operational understanding should be complemented by structural under-
standing. The fact that the formal consequences of Definitions 9.3.1 and 9.3.2
subsume the main results of classical integral vector calculus still doesnt
explain these ad hoc definitions conceptually. For everything to play out so
nicely, the definitions must somehow be natural rather than merely clever,
and a structural sense of why they work so well might let us extend the ideas
 9.4 Examples: 1-forms 431

to other contexts rather than simply tracking them. Indeed, differential forms
fit into a mathematical structure called a cotangent bundle, with each differ-
ential form being a section of the bundle. The construction of the cotangent
bundle involves the dual space of the alternation of a tensor product, all of
these formidable-sounding technologies being utterly Platonic mathematical
objects. However, understanding this language requires an investment in ideas
and abstraction, and in the authors judgment the startup cost is much higher
without some experience first. Hence the focus of the chapter is purely op-
erational. Since formula (9.14) may be opaque to the reader for now, the
first order of business is to render it transparent by working easy concrete
examples.

Exercises

9.3.1. Write out all ordered k-tuples from {1, , n} in the cases n = 4, k = 1;
n = 3, k = 2. In general, how many ordered k-tuples I = (i1 , , ik ) from
{1, , n} are there? How many of these are increasing, meaning that i1 <
< ik ? Write out all increasing k-tuples from {1, 2, 3, 4} for k = 1, 2, 3, 4.
P
9.3.2. An expression = I fI dxI where the sum is over only increasing
k-tuples from {1, , n} is called a standard presentation of . Write out
explicitly what a standard presentation for a k-form on R4 looks like for
k = 0, 1, 2, 3, 4.

9.4 Examples: 1-forms

A k-form is a function of k-surfaces. That is, one can think of a k-form as

a set of instructions: given Ra k-surface , carries out some procedure on
to produce a real number, .
For example, let
= x dy and = y dz,
both 1-forms on R3 . A 1-surface in R3 is a curve,

= (1 , 2 , 3 ) : [a, b] R3 ,

with 3-by-1 derivative matrix

 
1

= 2 .
3

For any such curve, is the instructions, integrate 1 2 over the parameter
domain [a, b], and similarly instructs to integrate 2 3 . You should work
through applying formula (9.14) to and to see how it produces these di-
rections. Note that x and y are being treated as functions on R3 for example,
432 9 Integration of Differential Forms

x(a, b, c) = a for all (a, b, c),

so that x = 1 .
To see and work on a specific curve, consider the helix

H : [0, 2] R3 , H(t) = (a cos t, a sin t, bt).

Its derivative matrix is

a sin t
H (t) = a cos t for all t [0, 2].
b

Thus by (9.14),
Z Z 2 Z Z 2
= a cos t a cos t = a2 and = a sin t b = 0.
H t=0 H t=0

Looking at the projections of the helix in the

R (x, y)-plane
R and the (y, z)-plane
suggests that these are the right values for H x dy and H y dz if we interpret
the symbols x dy and y dz as in one-variable calculus. (See figure 9.4.)

z z

y y

x x

Figure 9.4. Integrating 1-forms over a helix

For another example, let

= dx,
a 1-form on R3 , and consider any curve

: [a, b] R3 , (t) = (1 (t), 2 (t), 3 (t)).

 9.4 Examples: 1-forms 433

Then Z Z Z
b b
= (1 ) 1 = 1 = 1 (b) 1 (a).
a a

A change of notation makes this example more telling. Rewrite the component
functions of the curve as x, y, and z,

: [a, b] R3 , (t) = (x(t), y(t), z(t)).

So now x is not a function on R3 as in the previous example, but a function

on [a, b]. The integral rewrites as
Z Z b
= x = x(b) x(a).
a

That is, the form dx does indeed measure change in x along curves. As a set
of instructions it simply says to evaluate the x-coordinate difference from the
initial point on the curve to the final point. Think of dx as a Pac-Man. You
give it a curve, it runs along the curve and gives you back your score: the net
change in x along the curve. Returning to the helix H, it is now clear with no
further work that
Z Z Z
dx = 0, dy = 0, dz = 2b.
H H H

It is a good practice with formula (9.14) to confirm these values.

To generalize the previous example, let A be an open subset of Rn , let
f : A R be any smooth function, and associate a 1-form to f ,

= D1 f dx1 + + Dn f dxn .

Then for any curve : [a, b] A,

Z Z b
= (D1 f )1 + + (Dn f )n
a
Z b
= (f ) by the chain rule in coordinates
a
b
= (f )a
= f ((b)) f ((a)).

That is, the form measures change in f along curves. Indeed, is classically
called the total differential of f . It is tempting to give the name df , i.e., to
define
df = D1 f dx1 + + Dn f dxn .
Soon we will do so as part of a more general definition.
434 9 Integration of Differential Forms

(Recall the chain rule: If A Rn is open, then for any smooth : [a, b]
A and f : A R,

(f ) (t) = f ((t)) (t)

1 (t)
 
= D1 f ((t)) Dn f ((t)) ...
n (t)
n
X
= Di f ((t))i (t)
i=1
" n
#
X
= (Di f )i (t),
i=1
Pn
so indeed (f ) = i=1 (Di f )i .)
Continuing to generalize, consider now a 1-form that does not necessarily
arise from differentiation,

= F1 dx1 + + Fn dxn .

For any curve : [a, b] Rn the integral of over is

Z Z b n
X

= (Fi )i (u),
u=a i=1

and this is the general flow integral (9.9) of the vector field (F1 , , Fn )
along . That is, the flow integrals from section 9.2 are precisely the integrals
of 1-forms.

Exercises
R
9.4.1. Let = x dy y dx, a 1-form on R2 . Evaluate
for the following
curves.
(a) : [1, 1] R2 , (t) = (t2 1, t3 t);
(b) : [0, 2] R2 , (t) = (t, t2 ).
R
9.4.2. Let = z dx + x2 dy + y dz, a 1-form on R3 . Evaluate
for the
following two curves.
(a) : [1, 1] R3 , (t) = (t, at2 , bt3 );
(b) : [0, 2] R3 , (t) = (a cos t, a sin t, bt).

9.4.3. (a) Let = f dy where f : R2 R depends only on y. That is,

f (x, y) = (y) for some : R R. Show that for any curve = (1 , 2 ) :
[a, b] R2 ,
Z Z 2 (b)
= .
2 (a)
 9.5 Examples: 2-forms on R3 435

(b) Let = Rf dx + g dy where f depends only on x and g depends only

on y. Show that = 0 whenever : [a, b] R2 is a closed curve, meaning
that (b) = (a).

9.5 Examples: 2-forms on R3

To get a more complete sense of what formula (9.14) is doing, we need to study
a case with k > 1, i.e., integration on surfaces of more than one dimension.
Fortunately, the case n = 3, k = 2 is rich enough in geometry to understand
in general how k-forms on n-space work.
Consider figure 9.5. The figure shows a 2-surface in R3 ,

= (1 , 2 , 3 ) : D R3 .

The parameter domain D has been partitioned into subrectangles, and the
image (D) has been divided up into subpatches by mapping the grid lines
in D over to it via . The subrectangle J of D maps to the subpatch B of (D),
which in turn has been projected down to its shadow B(1,2) in the (x, y)-
plane. The point (uJ , vJ ) resides in J, and its image under is (uJ , vJ ) =
(xB , yB , zB ).

B12 x

Figure 9.5. 2-surface in 3-space

436 9 Integration of Differential Forms

Note that B(1,2) = (1 , 2 )(J). Rewrite this as

B(1,2) = (1,2) (J).

That is, B(1,2) is the image of J under the (1, 2) component functions of . If
J is small then results on determinants give

area(B(1,2) ) | det (1,2) (uJ , vJ )| area(J).

Thus, the magnification factor between subrectangles of D and (x, y)-projected

subpatches of (D) is (up to sign) the factor det I from formula (9.14) for I =
(1, 2). The sign is somehow keeping track of the orientation of the projected
patch, which would be reversed under projection onto the (y, x)-plane. (See
figure 9.6.)

y
x

B21
x

B12 y

Figure 9.6. Projected patch and its reversal

Let = f dx dy, a 2-form on R3 , where f : R3 R is a smooth

function. By (9.14) and Riemann sum approximation,
Z Z
= (f ) det (1,2)
D
X
(f )(uJ , vJ ) det (1,2) (uJ , vJ )area(J)
J
X

f (xB , yB , zB ) area(B(1,2) ) .
B

This calculation gives a geometric

R interpretation of what it means to integrate
f dx dy over : to evaluate f dx dy, traverse the set (D) and measure
projected, oriented area in the (x, y)-plane, weighted by the density function f .
The interpretation is analogous for forms with dy dz, and so on.
For an illustrative example, consider the forms dxdy, dz dx, and dy dz
integrated over the arch surface
 9.5 Examples: 2-forms on R3 437

: [1, 1] [0, 1] R3 , (u, v) = (u, v, 1 u2 ).

(See figure 9.7.) The (x, y)-shadows

R of B1 , B2 have the same areas as J1 , J2
and positive orientation, so dx dy should be equal to area(D), i.e., 2.
(SeeR the left half of figure 9.8.) The (z, x)-shadows of B1 , B2 have area zero,
so dz dx should be an emphatic 0. (See the right half of figure 9.8.) The
(y,
R z)-shadows of B1 , B2 have the same area but opposite orientations, so

dy dz should be 0 by some cancellation on opposite sides of the (y, z)-
plane or equivalently, cancellation in the u-direction of the parameter domain.
(See figure 9.9.)

v

z
y
J1 J2
u
B2

Figure 9.7. An arch

Integrating with formula (9.14) confirms this intuition. Since

1 0
(u, v) = 0 1 ,
2u 0

we have
Z Z Z 1 Z 1  
10
dx dy = det (1,2) = det = 2,
D v=0 u=1 01

and similarly
Z Z Z 1 Z 1 
 Z Z
2u 0
dz dx = det (3,1) = det = 0 = 0,
D v=0 u=1 1 0 v u
438 9 Integration of Differential Forms

y x

B1 B2 B2

x z

B1

Figure 9.8. (x, y)-shadows and (z, x)-shadows

z z

B1 B2

y y

Figure 9.9. (y, z)-shadows

Z Z Z 1 Z 1   Z Z
0 1
dy dz = det (2,3) = det = 2u = 0.
D v=0 u=1 2u 0 v u

Note how the first integral reduces to integrating 1 over the parameter do-
main, the second integral vanishes because its integrand is zero, and the third
integral vanishes because of cancellation in the u-direction. All three of these
behaviors confirm our geometric insight into how forms should behave.
Since the differential form dx dy measures projected area in the (x, y)-
plane, the integral Z
z dx dy

should give the volume under the arch. And indeed formula (9.14) gives
Z Z
z dx dy = (1 u2 ) 1,
(u,v)D
 9.5 Examples: 2-forms on R3 439

which is the volume. Specifically, the integral is

Z Z 1 Z 1 1

z dx dy = (1 u2 ) = 1 (2 u3 /3 ) = 4/3.
v=0 u=1 1

Similarly, since dy dz measures oriented projected area in the (y, z)-plane,

integrating the differential form x dy dz should also give the volume under
the arch. Here the interesting feature is that for x > 0 the form will multiply
the positive distance from the (y, z)-plane to the arch by positive (y, z)-area,
while for x < 0 the form will multiply the negative distance from the plane to
the arch by negative (y, z)-area, again measuring a positive quantity. To see
explicitly that the integral is again 4/3, compute:
Z Z 1 Z 1 1

x dy dz = u 2u = 1 (2/3)u3  = 4/3.
v=0 u=1 1

With these examples, the meaning of a k-form = f dxI on n-space is

fairly clear:
Integrating over a surface : D Rn means traversing the set
(D) and measuring oriented, k-dimensional volume of the projection
of (D) into k-space RI , weighted by the density function f .
This interpretation explains the results from integrating various 1-forms over
the helix in the previous section. Those integrals deserve reviewing in light of
this interpretation.
As the last example of this section, consider a 2-form on R3 ,

= F1 dx2 dx3 + F2 dx3 dx1 + F3 dx1 dx2 .

For any 2-surface : D R3 the integral of over is

Z Z (F1 )(D1 2 D2 3 D1 3 D2 2 )

= +(F2 )(D1 3 D2 1 D1 1 D2 3 ) (u),
uD
+(F3 )(D1 1 D2 2 D1 2 D2 1 )

and this is the flux integral (9.8) of the vector field (F1 , F2 , F3 ) through .
A straightforward generalization of this example shows that the general in-
tegral of an (n 1)-form over an (n 1)-surface in Rn is the general flux
integral (9.12). That is, the flux integrals from section 9.2 are precisely the
integrals of (n 1)-forms.
Along with the last example of the previous section, this raises the ques-
tion: Why bother with k-forms for values of k other than 1 and n 1, and
maybe also 0 and n? The answer is that the amalgamation of k-forms for
all values of k has a coherent algebraic structure, making the whole easier to
study than its parts. The remainder of the chapter is largely an elaboration
of this point.
440 9 Integration of Differential Forms

After this discussion of the mechanics and meaning of integrating forms,

you should be ready to prove a result that has already been mentioned: inte-
gration of forms reduces to ordinary integration when k = n, and integration
of forms is unaffected by reasonable changes of parametrization. These points
are covered in the next set of exercises.

Exercises

9.5.1. Let a be a positive number. Consider a 2-surface in R3 ,

: [0, a] [0, ] R3 , (r, ) = (r cos , r sin , r2 ).

Sketch thisRsurface, noting that varies from 0 to , not from 0 to 2. Try to

determine dx dy by geometrical reasoning, and then check your answer
by using (9.14) to evaluate the integral. Do the same for dy dz and dz dx.
Do the same for z dx dy y dz dx.
R
9.5.2. Let = x dy dz + y dx dy, a 2-form on R3 . Evaluate when
is the 2-surface (a) : [0, 1] [0, 1] R3 , (u, v) = (u + v, u2 v 2 , uv); (b)
: [0, 2] [0, 1] R3 , (u, v) = (v cos u, v sin u, u).
9.5.3. Consider a 2-form on R4 ,

= F1,2 dx1 dx2 + F1,3 dx1 dx3 + F1,4 dx1 dx4

+ F2,3 dx2 dx3 + F2,4 dx2 dx4 + F3,4 dx3 dx4 .

Show that for any 2-surface : D R4 , the integral of over is given by

formula (9.13) from near the end of section 9.2.
9.5.4. This exercise proves that integration of k-forms on Rn reduces to stan-
dard integration when k = n
Let D Rn be compact and connected. Define the corresponding natural
parametrization, : D Rn , by (u1 , , un ) = (u1 , , un ). (This is
how to turn a set in Rn , where we can integrate functions, into the corre-
sponding surface, where we can integrate n-forms.) Let = f dx1 dxn ,
an n-form on Rn . Use (9.14) to show that
Z Z
= f.
D

Your solution should use the basic properties of but not the highly sub-
stantive Change of Variable
R Theorem. Note that in particular if f = 1, then
= dx1 dxn and = vol(D), explaining why in this case is called
the volume form.
Thus in Rn , we may from now on blur the distinction between integrating
the function f over a set and integrating the n-form = f dxI over a surface,
provided that I = (1, , n) (i.e., the dxi factors appear in canonical order),
and provided that the surface is parametrized trivially.
 9.5 Examples: 2-forms on R3 441

9.5.5. This exercise proves that because of the Change of Variable Theorem,
the integration of differential forms is invariant under orientation-preserving
reparametrizations of a surface.
Let A be an open subset of Rn . Let : D A and : D e A
be k-surfaces in A. Suppose that there exists a smoothly invertible mapping
T : D D e such that T = . In other words, T is smooth, T is invertible,
its inverse is also smooth, and the following diagram commutes:

D



&
T A
qqqq8
q
qqq
 qqqq
Deq

If det T > 0 on D then the surface is called an orientation-preserving

reparametrization of , while if det T < 0 on D then is an orientation-
reversing reparametrization of .
(a) Let be a reparametrization as just defined. Let S = T 1 : De D,
a smooth mapping. Starting from the relation (S T )(u) = id(u) for all u D
(where id is the identity mapping on D), differentiate, use the chain rule, and
take determinants to show that det T (u) 6= 0 for all u D.
(b) Assume now that the reparametrization is orientation-preserving.
For any n-by-k matrix M and any ordered k-tuple I from {1, , n}, recall
that MI denotes the k-by-k matrix comprising the Ith rows of M . If N is a
k-by-k matrix, prove the equality

(M N )I = MI N.

In words, this says that

the Ith rows of (M times N ) are (the Ith rows of M ) times N .

(Suggestion: Do it first for the case I = i, that is, I denotes a single row.)
(c) Use the chain rule and part (b) to show that for any I,

det I (u) = det I (T (u)) det T (u) for all u D.

(d) Let = f (x) dxI , a k-form on A. Show that

Z Z
= (f ) det I .
T (D)

Explain why the Change of Variable Theorem shows that

Z Z


= (f ) det I T det T .
D
442 9 Integration of Differential Forms

Explain why this shows that

Z Z
= .

What would the conclusion be for orientation-reversing ?

(e)
P Do the results from (d) remain valid if has the more general form
= I fI dxI ?

9.6 Algebra of Forms: Basic Properties

One advantage of forms over earlier set-ups of vector integral calculus is that
one can do much of the necessary work with them algebraically. That is,
crucial properties will follow from purely rule-driven symbolic manipulation
rather than geometric intuition or close analysis.
Let A be an open subset of Rn . Since k-forms on A are functions (functions
of k-surfaces), they come with an inherent notion of equality. The meaning of

1 = 2

is that 1 () = 2 () for all k-surfaces in A. In particular, the meaning of

= 0 is that () = 0 for all , where the first 0 is a form while the second
is a real number. Addition of k-forms is defined naturally,

(1 + 2 )() = 1 () + 2 () for all 1 , 2 , ,

where the first + lies between two forms, the second between two real
numbers. Similarly, the definition of scalar multiplication is

(c)() = c(()) for all c, , .

The addition of forms here is compatible with the twofold use of summation
in the definition of forms and how they integrate. Addition and scalar multi-
plication of forms inherit all the vector space properties from corresponding
properties of addition and multiplication in the real numbers, showing that
the set of all k-forms on A forms a vector space. Proving familiar looking facts
about addition and scalar multiplication of forms reduces quickly to citing the
analogous facts in R. For example, (1) = for any k-form (where the
second minus sign denotes additive inverse), because for any k-surface ,

( + (1))() = () + ((1))() = () + (1)(()) = 0,

the last equality holding since (1)x = x for all real numbers x.
Forms have other algebraic properties that are less familiar. For example,
on R2 , dy dx = dx dy. This rule follows from the skew symmetry of the
determinant: For any 2-surface : D R2 ,
 9.7 Algebra of Forms: Multiplication 443
Z Z
(dy dx)() = det (2,1) = det (1,2) = (dx dy)().
D D

More generally, given two k-tuples I and J from {1, , n}, dxJ = dxI if J
is obtained from I by an odd number of transpositions. Thus for example,

dz dy dx = dx dy dz

since (3, 2, 1) is obtained from (1, 2, 3) by swapping the first and third entries.
Showing this reduces again to the skew symmetry of the determinant. As a
special case, dxI = 0 whenever the k-tuple I has two matching entries. This
rule holds because exchanging those matching entries has no effect on I but
negates dxI , and so dxI = dxI , forcing dxI = 0. One can also verify directly
that dxI = 0 if I has matching entries by referring back to the fact that the
determinant of a matrix with matching rows vanishes.
Using these rules (dy dx = dx dy, dx dx = 0 and their generaliza-
tions), one quickly convinces oneself that every k-form can be written
X
= fI dxI (sum only over increasing I),
I

where a k-tuple I = (i1 , , ik ) is called increasing if i1 < < ik , as

mentioned in exercise 9.3.1. This is the standard presentation for mentioned
in Exercise 9.3.2. It is not hard to show that the standard presentation for
is unique. In particular, is identically zero as a function of surfaces if and
only if has standard presentation 0.
The next few sections will define certain operations on forms and develop
rules of algebra for manipulating the forms under these operations. Like other
rules of algebra, they will be unfamiliar at first and deserve to be scrutinized
critically, but eventually they should become second nature and you should
find yourself skipping steps fluently.

Exercise

9.6.1. Show that if is a k-form on Rn that satisfies = , then = 0.

9.7 Algebra of Forms: Multiplication

Given a k-tuple and an -tuple, both from {1, , n},

I = (i1 , , ik ) and J = (j1 , , j ),

define their concatenation (I, J), a (k+)-tuple from {1, , n}, in the obvious
way,
(I, J) = (i1 , , ik , j1 , , j ).
444 9 Integration of Differential Forms

Also, if f, g : A R are functions on an open subset of Rn then their

product f g is the function

f g : A R, (f g)(x) = f (x)g(x).
n
Definition
P 9.7.1 (Wedge
P Product). Let A be an open subset of R . If
= I fI dxI and = J gJ dxJ are respectively a k-form and an -form
on A, then their wedge product is a (k + )-form on A,
X
= fI gJ dx(I,J) .
I,J

That is, the wedge product is formed by following the usual distributive law
and wedge-concatenating the dx-terms.

For convenient notation, let k (A) denote the vector space of k-forms
on A. Thus the wedge product is a mapping,

: k (A) (A) k+ (A).

For example, a wedge product of a 1-form and a 2-form on R3 is

(f1 dx+f2 dy + f3 dz) (g1 dy dz + g2 dz dx + g3 dx dy)

= f1 g1 dx dy dz + f1 g2 dx dz dx + f1 g3 dx dx dy
+ f2 g1 dy dy dz + f2 g2 dy dz dx + f2 g3 dy dx dy
+ f3 g1 dz dy dz + f3 g2 dz dz dx + f3 g3 dz dx dy
= (f1 g1 + f2 g2 + f3 g3 ) dx dy dz.

This example shows that the wedge product automatically encodes the inner
product in R3 , and the idea generalizes easily to Rn . For another example, a
wedge product of two 1-forms on R3 is

(xu dx + yu dy + zu dz) (xv dx + yv dy + zv dz)

= (yu zv zu yv ) dy dz
+ (zu xv xu zv ) dz dx
+ (xu yv yu xv ) dx dy.

Comparing this to the formula for the cross product in section 3.10 shows
that the wedge product automatically encodes the cross product. Similarly, a
wedge product of two 1-forms on R2 is

(a dx + b dy) (c dx + d dy) = (ad bc) dx dy,

showing that the wedge product encodes the 2-by-2 determinant as well.
Lemma 9.9.2 to follow will show that it encodes the general n-by-n deter-
minant.
 9.7 Algebra of Forms: Multiplication 445

Naturally the wedge in Definition 9.7.1 is the same as the one in Defi-
nition 9.3.1. There is no conflict in now saying that the two wedges are the
same, since each wedge in the earlier definition sits between two 1-forms and
the definition attached no meaning to the wedge symbol. Definition 9.3.1
also juxtaposes functions (0-forms) and dxI terms (k-forms) without putting
a wedge between them, and it is still unclear what sort of multiplication
that juxtaposition connotes. In fact, it is also a wedge product, but when we
wedge-multiply a 0-form and a k-form we usually suppress the wedge. A basic
property of the wedge, its skew symmetry, will explain why in a moment.

Proposition 9.7.2 (Properties of the Wedge Product). Let A be an

open subset of Rn . The wedge product has the following properties.
(1) The wedge product distributes over form addition: for all k (A) and
1 , 2 (A),
(1 + 2 ) = 1 + 2 .
(2) The wedge product is associative: for all k (A), (A), and
m (A),
( ) = ( ).
(3) The wedge product is skew symmetric: for all k (A) and (A),

= (1)k .

The proof is an exercise. The unfamiliar (and hence interesting) property

is the third one. The essence of its proof is to show that for any k-tuple I and
any -tuple J,
dxJ dxI = (1)k dxI dxJ .
This formula follows from counting transpositions.
Note that the skew symmetry of the wedge product reduces to symmetry
(i.e., commutativity) when either of the forms being multiplied is a 0-form.
The symmetry is why one generally doesnt bother writing the wedge when a
0-form is involved. In fact, the wedge symbol is unnecessary in all cases, and
typically in multivariable calculus one sees, for example,

dx dy dz rather than dx dy dz.

Indeed, we could use mere juxtaposition to denote form-multiplication, but

because this new multiplication obeys unfamiliar rules we give it a new symbol
to remind us of its novel properties as we study it.
Also, the special case of multiplying a constant function c and a k-form
is consistent with scalar multiplication of c (viewed now as a real number)
and . Thus all of our notions of multiplication are compatible.
446 9 Integration of Differential Forms

Exercises

9.7.1. Find a wedge product of two differential forms that encodes the inner
product of R4 .

9.7.2. Find a wedge product of three differential forms that encodes the 3-by-3
determinant.

9.7.3. Prove the properties of the wedge product.

9.7.4. Prove that (1 + 2 ) = 1 + 2 for all 1 , 2 k (A) and

(A). (Use skew symmetry, distributivity, and skew symmetry again.)

9.8 Algebra of Forms: Differentiation

Definition 9.8.1 (Derivative of a Differential Form). Let A be an open
subset of Rn . For each integer k 0 define the derivative mapping,

d : k (A) k+1 (A),

by the rules
n
X
df = Di f dxi for a 0-form f ,
i=1
X X
d = dfI dxI for a k-form = fI dxI .
I I

For example, we saw in section 9.4 that for a function f , the 1-form

df = D1 f dx1 + + Dn f dxn

is the form that measures change in f along curves. To practice this new kind
of function-differentiation in a specific case, define the function

1 : R3 R

to be projection onto the first coordinate,

1 (x, y, z) = x for all (x, y, z) R3 .

Then by the definition of the derivative,

d1 = D1 1 dx + D2 1 dy + D3 1 dz = dx. (9.15)

This calculation is purely routine. In practice, however, one often blurs the
distinction between the name of a function and its output, for instance speak-
ing of the function x2 rather than the function f : R R where f (x) = x2
 9.8 Algebra of Forms: Differentiation 447

or the squaring function on R. Such loose nomenclature is usually harmless

enough and indeed downright essential in any explicit calculation where we
compute using a functions values. But if we blur the distinction here between
the function 1 and its output x then the calculation of d1 in (9.15) rewrites
as
dx = dx. (!)
This is not tautological: the two sides have different meanings. The left side
is the operator d acting on the projection function x, while the right side is a
single entity, the 1-form denoted dx. The equation is better written

d(x) = dx.

However it is written, this equality ensures that there is no possible conflict

between naming the differential operator d and using this same letter as part
of the definition of differential form.
Similarly, for a function of one variable f : R R, the definition of d
immediately says that
df
df = dx,
dx
where the single, indivisible symbol df /dx is the Leibniz notation for the
derivative of f . This relation, which is sometimes presented in first semester
calculus with nebulous meanings attached to df and dx, and which canNOT
be proved by cancellation, is now a relation between 1-forms that follows from
the definition of d. The moral is that the operator d has been so named to
make such vague, undefined formulas into definitions and theorems. For more
examples of differentiation, if

= x dy y dx

then according to Definition 9.8.1,

d = (D1 x dx + D2 x dy) dy (D1 y dx + D2 y dy) dx = 2 dx dy.

And if
= x dy dz + y dz dx + z dx dy
then
d = 3 dx dy dz.
The differentiation operator d commutes with sums and scalar multiples.
That is, if 1 , 2 are k-forms and c is a constant then

d(c1 + 2 ) = c d1 + d2 .

More interesting are the following two theorems about form differentiation.
448 9 Integration of Differential Forms

Theorem 9.8.2 (Product Rule for Differential Forms). Let A be an

open subset of Rn . Let and be respectively a k-form and an -form on A.
Then
d( ) = d + (1)k d.

Proof. Start with the case of 0-forms f and g. Then

n
X
d(f g) = Di (f g) dxi
i=1
Xn
= (Di f g + f Di g) dxi
i=1
n
! n
!
X X
= Di f dxi g+f Di g dxi
i=1 i=1
= df g + f dg.

Next consider a k-form and an -form with one term each, fI dxI and gJ dxJ .
Then

d(fI dxI gJ dxJ ) = d(fI gJ dxI dxJ ) by definition of multiplication

= d(fI gJ ) dxI dxJ by definition of d
= (dfI gJ + fI dgJ ) dxI dxJ by the result for 0-forms
= dfI (gJ dxI ) dxJ by distributivity
+ fI (dgJ dxI ) dxJ and associativity of
0k
= dfI (1) (dxI gJ ) dxJ
by skew symmetry
+ fI (1)1k (dxI dgJ ) dxJ
= d(fI dxI ) gJ dxJ by associativity and symmetry
k and definition of d.
+ (1) fI dxI d(gJ dxJ )
P P
Finally in the general
P case = I I and = J J , where each I = fI dxI
and each j = J gJ dxJ , quoting the one-term result at the third equality,
!
X X X
d( ) = d I J = d(I J )
I J I,J
X
k
= (dI J + (1) I dJ )
I,J
X X X X
=d I J + (1)k I d J
I J I J
= d + (1)k d.

 9.8 Algebra of Forms: Differentiation 449

Because the last step in this proof consisted only of pushing sums tediously
through the other operations, typically it will be omitted from now on, and
proofs will be carried out for the case of one-term forms.
Consider a function f (x, y) on R2 . Its derivative is
df = D1 f (x, y) dx + D2 f (x, y) dy,
and its second derivative is in turn
d2 f = d(df ) = d(D1 f (x, y) dx) + d(D2 f (x, y) dy)
= D11 f (x, y) dx dx + D12 f (x, y) dy dx
+ D21 f (x, y) dx dy + D22 f (x, y) dy dy.
The dx dx term and the dy dy term are both 0. And the other two terms
sum to 0 because the mixed partial derivatives D12 f (x, y) and D21 f (x, y) are
equal while dy dx and dx dy are opposite. Overall, then,
d2 f = 0.
This phenomenon of the second derivative vanishing is completely general.
Theorem 9.8.3 (Nilpotence of d). Let A be an open subset of Rn . Then
d2 = 0 for any form k (A), where d2 means d d. In other words,
d2 = 0.
Proof. For a 0-form f ,
n
X
df = Di f dxi ,
i=1
and so
n
X X
d2 f = d(df ) = d(Di f ) dxi = Dij f dxj dxi .
i=1 i,j

All terms with i = j cancel since dxi dxi = 0, and the rest of the terms cancel
pairwise since for i 6= j, Dji f = Dij f (equality of mixed partial derivatives)
and dxi dxj = dxj dxi (skew symmetry of the wedge product). Thus
d2 f = 0.
Also, for a k-form dxI with constant coefficient function 1,
d(dxI ) = d(1dxI ) = (d1) dxI = 0.
Next, for a one-term k-form = f dxI ,
d = df dxI
and so by the first two calculations,
d2 = d(df dxI ) = d2 f dxI + (1)1 df d(dxI ) = 0 + 0 = 0.
For a general k-form, pass sums and d2 s through each other.

450 9 Integration of Differential Forms

A form is called

exact if = d for some form

and
closed if d = 0.
Theorem 9.8.3 shows that:

Every exact form is closed.

The converse question, whether every closed form is exact, is more subtle. We
will discuss it in section 9.11.

Exercises

9.8.1. Let = f dx + g dy + h dz. Show that

d = (D2 h D3 g) dy dz + (D3 f D1 h) dz dx + (D1 g D2 f ) dx dy.

9.8.2. Let = f dy dz + g dz dx + h dx dy. Evaluate d.

9.8.3. Differential forms of orders 0, 1, 2, 3 on R3 are written

0 = ,
1 = f1 dx + f2 dy + f3 dz,
2 = g1 dy dz + g2 dz dx + g3 dx dy,
3 = h dx dy dz.

(a) For a 0-form , what are the coefficients fi of d in terms of ?

(b) For a 1-form 1 , what are the coefficients gi of d1 in terms of the
coefficients fi of 1 ?
(c) For a 2-form 2 , what is the coefficient h of d2 in terms of the coeffi-
cients gi of 2 ?

9.8.4. Classical vector analysis features the operator

= (D1 , D2 , D3 ),

where the Di are familiar partial derivative operators. Thus, for a function
: R3 R,
= (D1 , D2 , D3 ).
Similarly, for a mapping F = (f1 , f2 , f3 ) : R3 R3 , F is defined in the
symbolically appropriate way, and for a mapping G = (g1 , g2 , g3 ) : R3 R3 ,
so is h, Gi. Write down explicitly the vector-valued mapping F and the
function h, Gi for F and G as just described. The vector-valued mapping
is the gradient of from section 4.8,
 9.8 Algebra of Forms: Differentiation 451

grad = .

The vector-valued mapping F is the curl of F ,

curl F = F.

And the scalar-valued function h, Gi is the divergence of G,

div G = h, Gi.

9.8.5. Continuing with the notation of the previous two problems, introduce
correspondences between the classical scalar-vector environment and the en-
vironment of differential forms, as follows. Let
~ = (dx, dy, dz),
ds
~ = (dy dz, dz dx, dx dy),
dn
dV = dx dy dz.

Let id be the mapping that takes each function : R3 R to itself, but

with the output-copy of viewed as a 0-form. Let ds ~ be the mapping that
takes each vector-valued mapping F = (f1 , f2 , f3 ) to the 1-form
~ = f1 dx + f2 dy + f3 dz.
F ds
~ be the mapping that takes each vector-valued mapping G = (g1 , g2 , g3 )
Let dn
to the 2-form
~ = g1 dy dz + g2 dz dx + g3 dx dy.
G dn

And let dV be the mapping that takes each function h to the 3-form

h dV = h dx dy dz.

Combine the previous problems to verify that the following diagram com-
mutes, meaning that either path around each square yields the same result.
(Do each square separately, e.g., for the middle square start from an arbitrary
(f1 , f2 , f3 ) with no assumption that it is the gradient of some function .)

grad curl div

/ (f1 , f2 , f3 ) / (g1 , g2 , g3 ) /h

~
ds ~
dn
id dV
 
 f1 dx g1 dy dz 
d d d
/ +f2 dy / +g2 dz dx / h dx dy dz
+f3 dz +g3 dx dy

Thus the form-differentiation operator d, specialized to three dimensions, uni-

fies the classical gradient, divergence, and curl operators.
452 9 Integration of Differential Forms

9.8.6. Two of these operators are zero:

curl grad, div curl, div grad.

Explain, using the diagram from the preceding exercise and the nilpotence
of d. For a function : R3 R, write out the harmonic equation (or
Laplaces equation), which does not automatically hold for all but turns
out to be an interesting condition,

div(grad ) = 0.

9.9 Algebra of Forms: the Pullback

Recall the Change of Variable Theorem from chapter 6: Given a change of
variable mapping now called T (rather than as in chapter 6), and given a
function f on the range space of T , the appropriate function to integrate over
the domain is obtained by composing with T and multiplying by an absolute
determinant factor,
Z Z
f= (f T ) | det T |.
T (D) D

A generalization to forms of the notion of composing with T lets us similarly

transfer formsrather than functionsfrom the range space of a mapping T
to the domain. This generalization will naturally include a determinant factor
that is no longer encumbered by absolute value signs. The next section will
show that integration of differential forms is inherently invariant under change
of variable.
We start with some examples. The familiar polar coordinate mapping from
(r, )-space to (x, y)-space is
call
T (r, ) = (r cos , r sin ) = (x, y).

Using this formula, and thinking of T as mapping from (r, )-space forward to
(x, y)-space, any form on (x, y)-space can naturally be converted back into a
form on (r, )-space, simply by substituting r cos for x and r sin for y. If the
form on (x, y)-space is named then the form on (r, )-space is denoted T .
For example, the 2-form that gives area on (x, y)-space,

= dx dy,

has a naturally corresponding 2-form on (r, )-space,

T = d(r cos ) d(r sin ).

Working out the derivatives and then the wedge shows that
 9.9 Algebra of Forms: the Pullback 453

T = (cos dr r sin d) (sin dr + r cos d)

= r dr d.

Thus (now dropping the wedges from the notation), this process has converted
dx dy into r dr d as required by the Change of Variable Theorem.
For another example, continue to let T denote the polar coordinate map-
ping, and consider a 1-form on (x, y)-space (for (x, y) 6= (0, 0)),
x dy y dx
= .
x2 + y 2
The corresponding 1-form on (r, ) space (for r > 0) is
r cos d(r sin ) r sin d(r cos )
T = .
(r cos )2 + (r sin )2
Here the differentiations give

d(r sin ) = sin dr + r cos d, d(r cos ) = cos dr r sin d,

and so the form on (r, )-space is

r cos (sin dr + r cos d) r sin (cos dr r sin d)
T = = d.
r2
This result suggests that integrating over a curve in (x, y)-space will return
the change in angle along the curve. For example, integrating counterclock-
wise over the unit circle should return 2.
(Geometrically, let : I R2 {0} be a parametrized curve, let
p = (x, y) = (t) be a point on the curve, and view the unary cross product
(x, y) = (y, x) as a vector originating at p, pointing in the direction of in-
creasing polar angle . The tangent vector (t) = (x (t), y (t)) has component
length along the unary cross product vector as follows,
h(x , y ), (y, x)i xy yx
=p .
|(y, x)| x2 + y 2

(See figure 9.10.) To infinitesimalize this, multiply it by dt, and then, to make
the resulting form measure infinitesimal change in the polar angle along the
curve, we also need to divide by the distance from the origin to get altogether
(x dy y dx)/(x2 + y 2 ).)
For a third example, again start with the 1-form
x dy y dx
= ,
x2 + y 2
but this time consider a different change of variable mapping,
call
T (u, v) = (u2 v 2 , 2uv) = (x, y).
454 9 Integration of Differential Forms
y

(x , y )

(x, y)

(x, y)

Figure 9.10. Angular component of the tangent vector

The 1-form on (u, v)-space (for (u, v) 6= (0, 0)) corresponding to is now

(u2 v 2 ) d(2uv) 2uv d(u2 v 2 )

T = .
(u2 v 2 )2 + (2uv)2
The derivatives are

d(2uv) = 2(v du + u dv), d(u2 v 2 ) = 2(u du v dv),

and so
(u2 v 2 )(v du + u dv) 2uv(u du v dv)
T = 2
(u2 + v 2 )2
((u2 v 2 )v 2u2 v) du + ((u2 v 2 )u + 2uv 2 ) dv
=2
(u2 + v 2 )2
u dv v du
=2 .
u2 + v 2
Thus T is essentially the original form, except that it is doubled and now
it is a form on (u, v)-space. The result of the calculation stems from the fact
that T is the complex square mapping, which doubles angles. The original
form , which measures change of angle in (x, y)-space, has transformed back
to the form that measures twice the change of angle in (u, v)-space. Integrating
T along a curve in (u, v)-space that misses the origin returns twice the
change in angle along this curve, and this is the change in angle along the
image-curve T in (x, y)-space.
 9.9 Algebra of Forms: the Pullback 455

Given a mapping, the natural process of changing variables in a differen-

tial form on the range of the mapping to produce a differential form on the
domain of the mapping is called pulling the differential form back through the
mapping. The general definition is as follows.

Definition 9.9.1 (Pullback of a Differential Form). Let k be a nonneg-

ative integer. Let A be an open subset of Rn , and let B be an open subset
of Rm . Let
T = (T1 , , Tm ) : A B
be a smooth mapping. Then T gives rise to a pullback mapping of k-forms
in the other direction,
T : k (B) k (A).
Let the coordinates on Rn be (x1 , , xn ), and let the coordinates on Rm
be (y1 , , ym ). For each k-tuple I = (i1 , , ik ) from {1, , m}, let dTI
denote dTi1 dTik . Then the pullback of any k-form on B,
X
= fI dyI ,

is X
T = (fI T ) dTI .
I

Since each Tij is a function on A, each dTij is a 1-form on A and the

definition makes sense. As usual, when k = 0, the empty products dyI and dTI
are interpreted as 1, and the pullback is simply composition,

T f = f T.

As the examples before the definition have shown, computing pullbacks is easy
and purely mechanical: given a form in terms of ys and dys, its pullback
T comes from replacing each yi in by the expression Ti (x1 , , xn ) and
then working out the resulting ds and wedges.
The fact that pulling the form dx dy back through the polar coordinate
mapping produced the factor r from the Change of Variable Theorem is no
coincidence.

Lemma 9.9.2 (WedgeDeterminant Lemma). Define an n-form valued

function on n-tuples of n-vectors as follows. For any n vectors in Rn ,

a1 = (a11 , a12 , , a1n ),

a2 = (a21 , a22 , , a2n ),
..
.
an = (an1 , an2 , , ann ),

create the corresponding 1-forms,

456 9 Integration of Differential Forms

1 = a11 dx1 + a12 dx2 + + a1n dxn ,

2 = a21 dx1 + a22 dx2 + + a2n dxn ,
..
.
n = an1 dx1 + an2 dx2 + + ann dxn ,

and then define

(a1 , a2 , , an ) = 1 2 n .

Then
(a1 , a2 , , an ) = det(a1 , a2 , , an ) dx1 dxn .
That is, = det dx(1, ,n) .

We have already seen this result for n = 2 in section 9.7 and for n = 3 in
exercise 9.7.2.

Proof. The only increasing n-tuple from {1, , n} is (1, , n). As a product
of n 1-forms on Rn , (a1 , a2 , , an ) is an n-form on Rn , and therefore it is
a scalar-valued function (a1 , a2 , , an ) times dx(1, ,n) . The relation

(a1 , a2 , , an )dx(1, ,n) = 1 2 n

where i is the inner product ai (dx1 , , dxn ) for each i combines with
various properties of the wedge product to show that the following three con-
ditions hold:
The function is linear in each of its vector variables, e.g.,

2 , , an ) = (a1 , a2 , , an ) + (a1 , a
(a1 , a2 + a 2 , , an )

and
(a1 , ca2 , , an ) = c (a1 , a2 , , an ).
The function is skew symmetric, i.e., transposing two of its vector vari-
ables changes its sign.
The function is normalized, i.e., (e1 , e2 , , en ) = 1.
The determinant is the unique function satisfying these three conditions,
so = det.

Theorem 9.9.3 (PullbackDeterminant Theorem). Let A be an open

subset of Rn , and let B be an open subset of Rm . Let T : A B be a smooth
mapping. Let Rn have coordinates (x1 , , xn ), and let Rm have coordinates
(y1 , , ym ). Let I = (i1 , , in ) be an n-tuple from {1, , m}. Then

T dyI = det TI dx1 dxn .

 9.9 Algebra of Forms: the Pullback 457

Proof. By definition,

T dyI = dTI = dTi1 dTin

= (D1 Ti1 dx1 + + Dn Ti1 dxn )
(D1 Ti2 dx1 + + Dn Ti2 dxn )
..
.
(D1 Tin dx1 + + Dn Tin dxn ).

The right side is precisely (Ti1 , Ti2 , , Tin ), so the lemma completes the
proof.

In particular, when m = n and I = (1, , n), the theorem says that

T (dy1 dyn ) = det T dx1 dxn ,

confirming the polar coordinate example early in this section. Similarly, if T

is the spherical coordinate mapping,

T (, , ) = ( cos sin , sin sin , cos ),

then the theorem tells us that

T (dx dy dz) = 2 sin d d d.

You may want to verify this directly to get a better feel for the pullback
and the lemma. In general, the PullbackDeterminant Theorem can be a big
time-saver for computing pullbacks when the degree of the form equals the
dimension of the domain space. Instead of multiplying out lots of wedge prod-
ucts, simply compute the relevant subdeterminant of a derivative matrix.
What makes the integration of differential forms invariant under change of
variable is that the pullback operator commutes with everything else in sight.
Theorem 9.9.4 (Properties of the Pullback). Let A be an open subset
of Rn , and let B be an open subset of Rm . Let T = (T1 , , Tm ) : A B
be a smooth mapping. Then:
(1) For all 1 , 2 , k (B) and c R,

T (1 + 2 ) = T 1 + T 2 ,
T (c) = c T .

(2) For all k (B) and (B),

T ( ) = (T ) (T ).

(3) For all k (B),

T (d) = d(T ).
458 9 Integration of Differential Forms

That is, the pullback is linear, the pullback is multiplicative (meaning that
it preserves products), and the pullback of the derivative is the derivative of
the pullback. The results in the theorem can be expressed in commutative
diagrams, as in exercise 9.8.5. Part (2) says that the following diagram com-
mutes:
(T ,T )
k (B) (B) / k (A) (A)



T
k+ (B) / k+ (A),

and part (3) says that the following diagram commutes:

T
k (B) / k (A)

d d


T
k+1 (B) / k+1 (A).

All of this is especially gratifying because the pullback itself is entirely natural.
Furthermore, the proofs are straightforward: all we need to do is compute, ap-
ply definitions, and recognize definitions. The only obstacle is that the process
requires patience.

Proof. (1) Is immediate from the definition.

(2) For one-term forms f dyI and g dyJ ,

T (f dyI g dyJ ) = T (f g dy(I,J) ) by definition of multiplication

= (f g) T dT(I,J) by definition of the pullback
= f T dTI g T dTJ since (f g) T = (f T )(g T )
= T (f dyI ) T (g dyJ ) by definition of the pullback.

The result on multi-term forms follows from this and (1).

(3) For a 0-form f : Rm R, compute that
 9.9 Algebra of Forms: the Pullback 459
m
X

T (df ) = T ( Di f dyi ) applying the definition of d
i=1
m
X applying the definition
= (Di f T ) dTi
i=1
of the pullback
m
X n
X
= Di f T Dj Ti dxj applying the definition of d
i=1 j=1
n
"m #
X X
= (Di f T ) Dj Ti dxj interchanging the sums
j=1 i=1
Xn
= Dj (f T ) dxj recognizing the chain rule
j=1

= d(f T ) recognizing the definition of d

= d(T f ) recognizing the pullback.

For a one-term k-form f dyI , d(f dyI ) = df dyI , so by (2) and the result for
0-forms,

T (d(f dyI )) = T (df dyI ) applying the definition of d

= T df T dyI since pullback and wedge commute
= d(T f ) T dyI by the just-established result
= d(f T ) dTI by definition of the pullback, twice
= d(f T dTI ) recognizing the definition of d
= d(T (f dyI )) recognizing the pullback.

The multi-term result follows from this and (1).

The pullback also behaves naturally with respect to composition.

Theorem 9.9.5 (Contravariance of the Pullback). Let A be an open

subset of Rn , let B be an open subset of Rm , and let C be an open subset
of R . Let T : A B and S : B C be smooth mappings. Then for any
form k (C),
(S T ) = (T S ).

This peculiar-looking resultthat the pullback of a composition is the com-

position of the pullbacks, but in reverse orderis grammatically inevitable.
Again, a commutative diagram expresses the idea:

S T
k (C) / k (B) / k
4 (A).
(ST )
460 9 Integration of Differential Forms

Proof. For a 0-form f : C R, the result is simply the associativity of

composition,

(S T ) f = f (S T ) = (f S) T = T (S f ) = (T S )f.

Let (z1 , , z ) be coordinates on R . Any one-term 1-form dzq (where q is

an integer from {1, , }) can be viewed as d(zq ), with d the differentiation
operator and zq the q-th projection function. Thus

(S T ) dzq = d((S T ) zq ) since derivative commutes with pullback

= d((T S )zq ) from just above, since zq is a function
= d(T (S zq )) by definition of composition
= T (d(S zq )) since derivative commutes with pullback
= T (S dzq ) since derivative commutes with pullback
= (T S )dzq by definition of composition.

Since any k-form is a sum of wedge products of 0-forms and 1-forms, and since
the pullback passes through sums and products, the general case follows.
Recapitulating the section: To pull a differential form back though a map
is to change variables in the form naturally. Because the wedge product has
the determinant wired into it, so does the pullback. Because the pullback is
natural, it commutes with addition, scalar multiplication, wedge multiplica-
tion, and differentiation of forms, and it anticommutes with composition of
forms. That is, everything that we are doing is preserved under change of
variables.
The results of this section are the technical heart of the chapter. The
reader is encouraged to contrast their systematic algebraic proofs with the
tricky analytic estimates in the main proofs of chapter 6. The work of this
section will allow the pending proof of the General Fundamental Theorem of
Integral Calculus to be carried out by algebra, an improvement over hand-
waving geometry or tortuous analysis. The classical integration theorems of
the nineteenth century will follow without recourse to the classical procedure
of cutting a big curvy object into many pieces and then approximating each
small piece by a straight piece instead. The classical procedure is either im-
precise or Byzantine, but for those willing to think algebraically, the modern
procedure is accurate and clear.
We end this section by revisiting the third example from its beginning.
Recall that we considered the 1-form
x dy y dx
=
x2 + y 2
and the complex square mapping
call
T (u, v) = (u2 v 2 , 2uv) = (x, y),
 9.9 Algebra of Forms: the Pullback 461

and we computed that the pullback T was twice , but written in (u, v)-
coordinates. Now we obtain the same result more conceptually in light of the
results of this section. The idea is that since measures change in angle, which
doubles under the complex square mapping, the result will be obvious in polar
coordinates, and furthermore the pullback behaves so well under changes of
variable that the corresponding result for cartesian coordinates will follow
easily as well. Thus, consider the polar coordinate mapping
call
: R>0 R R2 \{(0, 0)}, (r, ) = (r cos , r sin ) = (u, v).

In polar coordinates the complex square mapping re-expresses itself as

call
S : R>0 R R>0 R, S(r, ) = (r2 , 2) = (r, ).

And the polar coordinate mapping also applies to the polar coordinates that
are output by the complex square mapping,

= ( r sin ) call
= (x, y).
: R>0 R R2 \{(0, 0)}, (
r, ) r cos ,

Thus we have a commutative diagram

/ R2 \{(0, 0)}
R>0 R

S T
 
/ R2 \{(0, 0)}.
R>0 R

In terms of differential forms and pullbacks we have the resulting diagram

1 (R>0 R) o 1 (R2 \{(0, 0)})
O O
S T


1 (R>0 R) o 1 (R2 \{(0, 0)}).

Now, to find T where = (x dy y dx)/(x2 + y 2 ), recall that pulls back

through the polar coordinate mapping to d, and recall that = 2. Thus we
have in the second diagram

d(2) o T
O O

d o

Since d(2) = 2 d, the sought-for pullback T must be the (u, v)-form that
pulls back through the polar coordinate mapping to 2 d. And so T should
be the double of , but with u and v in place of x and y,
462 9 Integration of Differential Forms

u dv v du
T = 2 .
u2 + v 2
This is the value of T that we computed mechanically at the beginning of
the section. Indeed, note that this second derivation of T makes no reference
whatsoever to the formula T (u, v) = (u2 v 2 , 2uv), only to the fact that in
polar coordinates the complex square mapping squares the radius and doubles
the angle.
Similarly, we can use these ideas to pull the area-form = dx dy back
through T . Indeed, dx dy pulls back through the polar coordinate mapping
which pulls back through S to r2 d(r2 ) d(2) = 4r3 dr d.
r d,
to r d
Thus we have a commutative diagram

4r3 drO d o T
O

r d o
r d

So T must pull back through the polar coordinate mapping to 4r3 dr d.

Since the area-form du dv pulls back to r dr d, the answer is the area
form du dv multiplied
by 4r2 in (u, v)-coordinates. That is, since r in (u, v)-
coordinates is u2 + v 2 ,

T = T (dx dy) = 4(u2 + v 2 ) du dv.

This formula for T can be verified directly by purely mechanical computa-

tion.

Exercises
call
9.9.1. Define S : R2 R2 by S(u, v) = (u + v, uv) = (x, y). Let =
x2 dy + y 2 dx and = xy dx, forms on (x, y)-space.
(a) Compute , S (u, v), and (use the PullbackDeterminant Theorem)

S ( ).
(b) Compute S , S , and S S . How do you check the last of
these? Which of the three commutative diagrams from the section is relevant
here?
(c) Compute d and S (d).
(d) Compute d(S ). How do you check this? Which commutative diagram
is relevant?
call
(e) Define T : R2 R2 by T (s, t) = (s t, set ) = (u, v). Compute
T (S ).
(f) What is the composite mapping S T ? Compute (S T ) . How do
you check this, and which commutative diagram is relevant?
 9.9 Algebra of Forms: the Pullback 463

9.9.2. Recall the two forms from the beginning (and the end) of the section,
x dy y dx
= , = dx dy.
x2 + y 2
Consider a mapping from the nonzero points of (u, v)-space to nonzero points
of (x, y)-space.
 
u v call
T (u, v) = , = (x, y).
u2 + v 2 u2 + v 2
As at the end of the section, in light of the fact that T is the complex reciprocal
mapping, determine what T and T must be. If you wish, confirm your
answers by computing them mechanically as at the beginning of the section.
9.9.3. Consider a differential form on the punctured (x, y)-plane,
x dx + y dy
= p .
x2 + y 2

(a) Pull back through the polar coordinate mapping from the end of the
section,
( = (
r, ) r cos , call
r sin ) = (x, y).
R
In light of the value of the pullback, what must be the integral where
is a parametrized curve in the punctured (x, y)-plane?
(b) In light of part (a), pull back through the complex square mapping
from the section,
call
T (u, v) = (u2 v 2 , 2uv) = (x, y),
by using diagrams rather than by relying heavily on computation. Check your
answer by computation if you wish.
(c) Similarly to part (a), pull back through the complex reciprocal map-
ping from the previous exercise,
 
u v call
T (u, v) = , = (x, y).
u2 + v 2 u2 + v 2
by using diagrams. Check your answer by computation if you wish.
(d) Let k be an integer. The relation x + iy = (u + iv)k determines (x, y)
as a function T (u, v). Pull the forms and from the previous exercise and
the form from this exercise back through T , with no reference to any ex-
plicit formula for T . The results should in particular reproduce your previous
answers for k = 2 and k = 1.
9.9.4. Let A = R3 {0}. Let r be a fixed positive real number. Consider a
2-surface in A,

: [0, 2] [0, ] A, (, ) = (r cos sin , r sin sin , r cos ).

464 9 Integration of Differential Forms

Consider also a 2-form on A,

= (x/r) dy dz (y/r) dz dx (z/r) dx dy.

Compute the derivative matrix (, ), and use the Pullback-Determinant

Theorem three times to compute the pullback . Compare your answer
to the integrand of the surface integral near the end of section 9.1 used to
compute the volume of the sphere of radius r. (It follows that is the area-
form for the particular surface in this exercise, but not that is a general
area-form for all surfaces.)

9.10 Change of Variable for Differential Forms

The definition of integration and the algebra of forms combine to make a

Change of Variable Theorem for differential forms a triviality. First, a theorem
of independent interest allows us to replace any integral of a differential form
over a parametrized surface with an integral over the trivial parametrization
of the surfaces parameter domain.

Theorem 9.10.1 (Pullback Theorem). Let A be an open subset of Rn .

Let be a k-form on A and let : D A be a k-surface in A. Define a
k-surface in Rk ,

D : D Rk , D (u) = u for all u D.

Then Z Z
= .
D

Proof. As usual, just do the case of a one-term form, = f dxI . Then

Z Z
f dxI = (f ) det I by definition, per (9.14)
D
Z
= (f ) det I du1 duk by Exercise 9.5.4
D
Z
= (f ) dxI by Theorem 9.9.3
D
Z
= (f dxI ) by definition of pullback.
D

The general Change of Variable Theorem for differential forms follows im-
mediately from the Pullback Theorem and the contravariance of the pullback.
 9.10 Change of Variable for Differential Forms 465

Theorem 9.10.2 (Change of Variable for Differential Forms). Let A

be an open subset of Rn , and let B be an open subset of Rm . Let T : A B
be a smooth mapping. For any k-surface in A, : D A, the composition
T : D B is thus a k-surface in B. Let be a k-form on B. Then
Z Z
= T .
T

Proof. Let D : D Rk be as above. Then

Z Z Z Z
= (T ) = (T ) = T .
T D D

The Pullback Theorem is essentially equivalent to the definition of inte-

gration once one has the PullbackDeterminant Theorem. Thus, a logically
equivalent route to ours through this material is to define integration of a k-
form in k-space as ordinary integration, and integration of a k-form in n-space
for k < n via the pullback. Doing so would have been a little tidier (there
would not be two notions of integration when k = n whose compatibility needs
to be verified), but the approach here has the advantage that one can start
integrating immediately before developing all the algebra.

Exercise
call
9.10.1. Let T : R2 R2 be given by T (x1 , x2 ) = (x21 x22 , 2x1 x2 ) = (y1 , y2 ).
Let be the curve : [0, 1] R2 given by (t) = (1, t) mapping the unit
interval into (x1 , x2 )-space, and let T be the corresponding curve mapping
into (y1 , y2 )-space. Let = y1 dy2 , a 1-form
R on (y1 , y2 )-space.
(a) Compute T , and then compute T by using formula (9.14).
(b) Compute RT , the pullback of by T .
(c) Compute T by using formula (9.14). What theorem says that the
answer here is the same as (a)?
(d) Let = dy1 dy2 , the area form on (y1 , y2 )-space. Compute T .
(e) A rectangle in the first quadrant of (x1 , x2 )-space,

R = {(x1 , x2 ) : a1 x1 b1 , a2 x2 b2 },

Rgets taken to some indeterminate patch B = T (R) by T . Find the area of B,

B
, using (d). (This exercise abuses notation slightly, identifying R with its
natural parametrization and B with the corresponding surface T R.)
(f) Why does this exercise require that R lie in the first quadrant? Can
the restriction be weakened?
466 9 Integration of Differential Forms

9.11 Closed Forms, Exact Forms, and Homotopy

Let be a differential form. Recall the terminology that

is exact if = d for some

and
is closed if d = 0.
The nilpotence of d (the rule d2 = 0 from Theorem 9.8.3) shows that any exact
form is closed. We now show that under certain conditions, the converse is
true as well, i.e., under certain conditions any closed differential form can be
antidifferentiated.
A homotopy of a set is a process of deforming the set to a single point,
the deformation taking place entirely within the original set. For example,
consider the open ball
A = {x Rn : |x| < 1}.
A mapping that shrinks the ball to its center as one unit of time elapses is

h : [0, 1] A A, h(t, x) = tx.

The idea geometrically is that at time t = 1, h is the identity mapping so

that h(1, A) = A, while at any intermediate time t (0, 1), h(t, A) = tA is
a scaled-down copy of the ball, and finally at time t = 0, h(0, A) = {0} and
the ball has shrunk to its center. (So here we have let time flow from t = 1
to t = 0 for convenience.)
However, the geometric story just told is slightly misleading. We could
replace the ball A in the previous example by all of Euclidean space Rn , and
the map
h : [0, 1] Rn Rn , h(t, x) = tx
would still contract Rn to {0} in the sense that each point x Rn is moved
by h to 0 as t varies from 1 to 0. However, at any intermediate time t (0, 1),
h(t, Rn ) = tRn = Rn is still all of Euclidean space. Although every point of Rn
is moved steadily by h to 0, h does not shrink the set Rn as a whole until
the very end of the process, when space collapses instantaneously to a point.
Any given point x of Rn is taken close to the origin once the time t is close
enough to 0, but the required smallness of t depends on x; for no positive t,
however close to 0, is all of Rn taken close to the origin simultaneously. The
relevant language here is that homotopy is a convergent process that need not
be uniformly convergent, analogously to how a continuous function need not
be uniformly continuous. The mental movie that we naturally have of a set
shrinking to a point depicts a uniformly convergent process, and so it doesnt
fully capture homotopy.
For another example, consider the annulus

A = {x R2 : 1 < |x| < 2}.

 9.11 Closed Forms, Exact Forms, and Homotopy 467

Plausibly there is no homotopy of the annulus, meaning that the annulus can
not be shrunk to a point by a continuous process that takes place entirely
within the annulus. But proving that there is no homotopy of the annulus is
not trivial. We will return to this point in exercise 9.11.1.
The formal definition of a homotopy is as follows.

Definition 9.11.1 (Homotopy, Contractible Set). Let A be an open sub-

set of Rn . Let be a positive number and let

B = (, 1 + ) A,

an open subset of Rn+1 . A homotopy of A is a a smooth mapping

h : B A

such that for some point p of A,

( )
h(0, x) = p
for all x A.
h(1, x) = x

An open subset A of Rn that has a homotopy is called contractible.

Again, the idea is that B is a sort of cylinder over A, and that at one end
of the cylinder the homotopy gives an undisturbed copy of A while by the
other end of the cylinder the homotopy has compressed A down to a point.
This section proves the following result.

Theorem 9.11.2 (Poincar e). Let A be a contractible subset of Rn , and

let k 1 be an integer. Then every closed k-form on A is exact.

To prepare for the proof of theorem, we consider a cylinder over A,

B = (, 1 + ) A,

but for now we make no reference to the pending homotopy that will have B
as its domain. Recall that the differentiation operator d increments the degree
of any differential form. Now, by contrast, we define a linear operator that
takes differential forms on B and returns differential forms of one degree lower
on A. Let the coordinates on B be (t, x) = (t, x1 , , xn ) with t viewed as
the zeroth coordinate.

Definition 9.11.3. For any positive integer k, define a linear mapping of

differential forms,

c : k (B) k1 (A), k = 1, 2, 3, ,

as follows: c acts on any one-term form that contains dt by integrating its

component function in the t-direction and suppressing its dt, and c annihilates
468 9 Integration of Differential Forms

differential forms that dont contain dt. That is, letting I denote (k 1)-tuples
and J denote k-tuples, all tuples being from {1, , n},
!
X X X Z 1 
c gI (t, x) dt dxI + gJ (t, x) dxJ = gI (t, x) dxI .
I J I t=0

With c in hand, we have two degree-preserving mappings from differen-

tial forms on B to differential forms on A, the compositions of c and the
differentiation operator d in either order,

cd, dc : k (B) k (A), k = 1, 2, 3, .

However, note that cd proceeds from k (B) to k (A) via k+1 (B), while dc
proceeds via k1 (A). To analyze the two compositions, compute first that
for a one-term differential form that contains dt,
n
!
X
(cd)(g(t, x) dt dxI ) = c Di g(t, x) dxi dt dxI
i=1
n
!
X
=c Di g(t, x) dt dx(i,I)
i=1
n Z
X 1 
= Di g(t, x) dx(i,I) ,
i=1 t=0

while, using the fact that xi -derivatives pass through t-integrals for the third
equality to follow,
Z 1  
(dc)(g(t, x) dt dxI ) = d g(t, x) dxI
t=0
n
X Z 1 
= Di g(t, x) dx(i,I)
i=1 t=0

Xn Z 1 
= Di g(t, x) dx(i,I) .
i=1 t=0

Thus cd + dc annihilates forms that contain dt. On the other hand, for a
one-term differential form without dt,

n
X
(cd)(g(t, x) dxJ ) = c D0 g(t, x) dt dxJ + Dj g(t, x) dx(j,J)
j=1
Z 1 
= D0 g(t, x) dxJ
t=0
= (g(1, x) g(0, x)) dxJ ,
 9.11 Closed Forms, Exact Forms, and Homotopy 469

while
(dc)(g(t, x) dxJ ) = d(0) = 0.
That is, cd + dc replaces each coefficient function g(t, x) in forms without dt
by g(1, x) g(0, x), a function of x only.
To notate the effect of cd+dc more tidily, define the two natural mappings
from A to the cross-sections of B where the pending homotopy of A will end
and where it will begin,
( )
0 (x) = (0, x)
0 , 1 : A B, .
1 (x) = (1, x)

Because 0 and 1 have ranges where t is constant, and because they dont
affect x, their pullbacks,

0 , 1 : k (B) k (A), k = 0, 1, 2, ,

act correspondingly by replacing t with a constant and dt with 0 while pre-

serving x and dx,

0 (g(t, x) dt dxI ) = 0, 1 (g(t, x) dt dxI ) = 0

and

0 (g(t, x) dxJ ) = g(0, x) dxJ , 1 (g(t, x) dxJ ) = g(1, x) dxJ .

It follows that our calculations rephrase as Poincar

es Identity,

(cd + dc) = (1 0 ), k (B), k = 1, 2, 3, .

With Poincares Identity established, we prove Poincares Theorem.

Proof (of Theorem 9.11.2). We have an open subset A of Rn , a point p of A,

a cylinder B = (, 1 + ) A for some positive number , and a homotopy

h : B A.

So also we have the corresponding pullback

h : k (A) k (B), k = 0, 1, 2, .

Let k 1, and consider a closed form k (A). Then h k (B) and

ch k1 (A). We show that ch is an antiderivative of by comput-
ing the quantity (cd + dc)h in two ways. First, because the pullback and
boundary operators commute and because d = 0,

(cd + dc)h = ch d + dch = d(ch ).

Second, by Poincares Identity and the contravariance of the pullback,

470 9 Integration of Differential Forms

(cd + dc)h = (1 0 )h = ((h 1 ) (h 0 ) ).

But (h 1 )(x) = h(1, x) = x and (h 0 )(x) = h(0, x) = p, i.e., h 1 is

the identity mapping and h 0 is a constant mapping, so that (h 1 ) has
no effect on while (h 0 ) annihilates the dxs of (which are present
since k 1), thus annihilating . In sum, the second computation gives . So
the computations combine to give

d(ch ) = .

That is, is exact as desired.

Note that the process of antidifferentiating by taking ch moves from A

up to the larger space B and then back down to A. In terms of algebra, the
process inserts ts into by pulling it back through the homotopy and then
strips them out in a different way by applying the c operator.
We end the section with an example. Consider any closed form on R2 ,

= f (x, y) dx + g(x, y) dy, D2 f = D1 g.

Pull back through the homotopy h(t, x, y) = (tx, ty) of R2 to get

h = f (tx, ty) d(tx) + g(tx, ty) d(ty)

= (xf (tx, ty) + yg(tx, ty)) dt + tf (tx, ty) dx + tg(tx, ty) dy.

Apply c to h in turn to get

Z 1



ch = xf (tx, ty) + yg(tx, ty) .
t=0

This function must have derivative . To verify that it does, compute that its
first partial derivative is
Z 1



D1 ch (x, y) = f (tx, ty) + xD1 (f (tx, ty)) + yD1 (g(tx, ty)) .
t=0

By the Chain Rule and then by the fact that D1 g = D2 f , the first partial
derivative is therefore
Z 1


D1 ch (x, y) = f (tx, ty) + xD1 f (tx, ty)t + yD1 g(tx, ty)t
t=0
Z 1 Z 1
= f (tx, ty) + t(xD1 f (tx, ty) + yD2 f (tx, ty)).
t=0 t=0
R1
The last integral takes the form t=0 u v where u(t) = t and v(t) = f (tx, ty).
And so finally, integrating by parts gives
 9.11 Closed Forms, Exact Forms, and Homotopy 471
Z 1 1 Z 1

D1 ch (x, y) = f (tx, ty) + tf (tx, ty) f (tx, ty)
t=0 t=0 t=0
= f (x, y).

Similarly D2 ch (x, y) = g(x, y), so that indeed

d(ch ) = f (x, y) dx + g(x, y) dy = .

Exercises

9.11.1. (a) Here is a special case of showing that a closed form is exact without
recourse to Poincares Theorem. A function f : R3 R is called homoge-
neous of degree k if

f (tx, ty, tz) = tk f (x, y, z) for all t R and (x, y, z) R3 .

Such a function must satisfy Eulers identity,

xD1 f + yD2 f + zD3 f = kf.

Suppose that = f1 dx + f2 dy + f3 dz is a closed 1-form whose coefficient

functions are all homogeneous of degree k where k 0. Show that = d
where
1
= (xf1 + yf2 + zf3 ).
k+1
(Suggestion: first check only the dx term of d, remembering that is closed.
The other two terms will work out similarly by symmetry.)
(b) Here is a closed form that is not exact. Let

x dy y dx
= ,
x2 + y 2

a 1-form on the punctured plane A = R2 {(0, 0)}. Show that is closed.

Compute that integrating around the counterclockwise unit circle,

: [0, 2] A, (t) = (cos t, sin t),

gives a nonzero answer. Explain why this shows that there is no 0-form (i.e.,
function) on the punctured plane such that = d.
(c) Use part (b) to show that there cannot exist a homotopy of the punc-
tured plane. How does this nonexistence relate to the example of the annulus
at the beginning of the section?

9.11.2. Let = f (x, y, z) dy dz + g(x, y, z) dz dx + h(x, y, z) dx dy be a

closed form on R3 . (Here h does not denote a homotopy.) Find an antideriva-
tive of .
472 9 Integration of Differential Forms

9.12 Cubes and Chains

Sections 9.7 through 9.9 introduced algebraic operators on differential forms:
the wedge product, the derivative, and the pullback. The next section will
introduce a geometric operator on surfaces. The first thing to do is specialize
the definition of a surface a bit. As usual, let [0, 1] denote the unit interval.
For k 0 the unit k-box is the cartesian product
[0, 1]k = [0, 1] [0, 1] = {(u1 , , uk ) : ui [0, 1] for i = 1, , k}.
As mentioned in section 9.3, when k = 0 this means the one-point set whose
point is ().
Definition 9.12.1 (Singular Cube, Standard Cube). Let A be an open
subset of Rn . A singular k-cube in A is a surface whose parameter domain
is the unit k-box,
: [0, 1]k A.
In particular, the standard k-cube is
k : [0, 1]k Rk , k (u) = u for all u [0, 1]k .
As with Definition 9.1.1 of a surface, now a cube is by definition a mapping,
and in particular a 0-cube is the parametrization of a point. In practice, we
often blur the distinction between a mapping and its image, and under this
blurring the word cube now encompasses noncubical objects such as a torus-
surface (which is a singular 2-cube in R3 ) or a solid sphere (a singular 3-cube
in R3 ). The next definition allows us to consider more than one cube at a
time. The purpose is to integrate over several cubes in succession, integrating
over each of them a prescribed number of times.
Definition 9.12.2 (Chain). Let A be an open subset of Rn . A k-chain in
A is a finite formal linear combination
X
C= s (s) ,
s

where each s is an integer and each (s) is a singular k-cube in A. (The

surface subscript is in parentheses only to distinguish it from a component
function subscript.)
For example, if , and are singular k-cubes in Rn then
2 3 + 23
is a k-chain in Rn . This k-chain is not the singular k-cube that maps points u
to 2(u) 3 (u) + 23 (u) in Rn . The term formal linear combination in the
definition means that we dont actually carry out any additions and scalings.
Rather, the coefficients s are to be interpreted as integration multiplicities.
A k-chain, like a k-form, is a set of instructions.
 9.12 Cubes and Chains 473

Definition 9.12.3 (Integral of a k-form over a k-chain in n-space). Let

A be an open subset of Rn . Let
X
C= s (s)
s

be a k-chain in A, and let be a k-form on A. Then the integral of over C

is Z X Z
= s ,
C s (s)

This definition can be written more suggestively as

Z X Z
P
= s .
s (s) s (s)

Although C is a formal linear combination, the operations on the right of the

equality are literal addition and multiplication in R. For example, let a and b
be points in Rn , and let a and b be the corresponding 0-cubes. Then for
any 0-form on Rn , = f : Rn R,
Z
= f (b) f (a).
b a

One can define predictable rules for addition and scalar multiplication (integer
scalars) of chains, all of which will pass through the integral sign tautologically.
Especially, the Change of Variable Theorem for differential forms extends from
integrals over surfaces to integrals over chains,
Z Z
= T .
T C C

We will quote this formula in the proof of the General FTIC.

Also, if C is a chain in A and T : A B is a mapping, then we can
naturally compose them to get a chain in B by passing sums and constant
multiples through T . That is,
X X
if C = s (s) then T C = s (T (s) ).
s s

Exercises

9.12.1. Let A be an open subset of Rn . Consider the inner product-like func-

tion (called a pairing)

h , i : {k-chains in A} {k-forms on A} R

defined by the rule

474 9 Integration of Differential Forms
Z
hC, i = for all suitable k-chains C and k-forms .
C

Show that this inner product is bilinear, meaning that for all suitable chains
C and Ci , all suitable forms and i , and all constants ci ,
X X
h ci Ci , i = ci hCi , i
i i

and X X
hC, c i i i = ci hC, i i.
i i
It makes no sense to speak of symmetry of this pairing since the arguments
may not be exchanged.
Do you think the pairing is nondegenerate, meaning that for any fixed
chain C, if hC, i = 0 for all forms then C must be 0, and for any fixed
form , if hC, i = 0 for all chains C then must be 0?
9.12.2. Let A be an open subset of Rn , let B be an open subset of Rm , and
let k 0. Any smooth mapping T : A B gives rise via composition to a
corresponding pushforward mapping from k-surfaces in A to k-surfaces in B,
T : {k-surfaces in A} {k-surfaces in B}, T = T .
In more detail, since a k-surface in A takes the form : D A where
D Rk is a parameter domain, the pushforward mapping is
T
( : D A) 7

(T : D B).
Using the pairing-notation of the previous exercise, a result from earlier in
this chapter renotates as
hT , i = h, T i for all suitable and .
Which result? Note that the renotation shows that the pushforward and pull-
back are like a pair of adjoint operators in the sense of linear algebra.

9.13 Geometry of Chains: the Boundary Operator

This section defines an operator that takes k-chains to (k 1)-chains. The
idea is to traverse the edge of each singular k-cube in the chain, with suitable
multiplicity and orientation. The following definition gives three rules that
say how to do so. The first rule reduces taking the boundary of a k-chain
to taking the boundary of its constituent singular k-cubes. The second rule
reduces taking the boundary of a singular k-cube to taking the boundary
of the standard k-cube. The third rule, giving the procedure for taking the
boundary of the standard k-cube, is the substance of the definition. It is best
understood by working through specific cases.
 9.13 Geometry of Chains: the Boundary Operator 475

Definition 9.13.1 (Boundary). Let A be an open subset of Rn . For each

k 1, define the boundary mapping

: {k-chains in A} {(k 1)-chains in A}

by the properties:
P
(1) For any k-chain s (s) ,
X  X
s (s) = s (s) .

(2) For any singular k-cube ,

= k .

(The composition here is of the sort defined at the end of the previous
section.)
(3) Define mappings from the standard (k1)-cube to the faces of the standard
k-cube as follows: for any i {1, , n} and {0, 1}, the mapping to
the face where the ith coordinate equals is

ki, : [0, 1]k1 [0, 1]k ,

given by

ki, (u1 , , uk1 ) = (u1 , , ui1 , , ui , , uk1 ).

Then
k X
X 1
k = (1)i+ ki, . (9.16)
i=1 =0

In property (2) the composition symbol Phas been generalized a little

from its ordinary usage. Since k is a chain s (s) , the composition
k
P
is defined as the corresponding chain s (s) . The compositions in
the sum make sense since by property (3), each (s) maps [0, 1]k1 into [0, 1]k .
To remember the definition of ki, in (9.16), read its name as:

Of k variables, set the ith to ,

or just set the ith variable to . The idea of formula (9.16) is that for each of
the directions in k-space (i = 1, , k), the standard k-cube has two faces with
normal vectors in the ith direction ( = 0, 1), and we should take these two
faces with opposite orientations in order to make both normal vectors point
outward. Unlike differentiation, which increments the degree of the form it
acts on, the boundary operator decrements chain dimension.
For example, the boundary of the standard 1-cube is given by (9.16)
476 9 Integration of Differential Forms

1 = 11,0 + 11,1 .

That is, the boundary is the right endpoint of [0, 1] with a plus and the left
endpoint with a minus. (See figure 9.11. The figures for this section show the
images of the various mappings involved, with symbols added as a reminder
that the images are being traversed by the mappings.) One consequence of
this is that the familiar formula from the one-variable Fundamental Theorem
of Integral Calculus, Z 1
f = f (1) f (0),
0
is now expressed suggestively in the notation of differential forms as
Z Z
df = f.
1 1

As for the boundary of a singular 1-cube : [0, 1] Rn (i.e., a curve in

space) with (0) = a and (1) = b, property (2) of the boundary definition
gives
= 1 = 11,0 + 11,1 .
Thus the boundary is the curves endpoint b with a plus and the start-point a
with a minus. The last example of section 9.4 now also takes on a more
suggestive expression, Z Z
df = f.

Figure 9.11. Standard 1-cube and its boundary

The boundary of the standard 2-cube is again given by (9.16)

2 = 21,0 + 21,1 + 22,0 22,1 .

This chain traverses the boundary square of [0, 1]2 once counterclockwise. (See
figure 9.12.) Next consider a singular 2-cube that parametrizes the unit disk,

: [0, 1]2 R2 , (r, ) = (r cos 2, r sin 2).

By property (2), = 2 . This chain traverses the boundary circle

once counterclockwise, two radial traversals cancel, and there is a degen-
erate mapping to the centerpoint. (See figure 9.13.) Changing to (r, ) =
(r cos 2, r sin 2) also parametrizes the unit disk, but now traverses
the boundary circle clockwise.
 9.13 Geometry of Chains: the Boundary Operator 477

Figure 9.12. Standard 2-cube and its boundary

Figure 9.13. Boundary of a singular 2-cube

The boundary of the standard 3-cube is, by (9.16),

3 = 31,0 + 31,1 + 32,0 32,1 33,0 + 33,1 .

This chain traverses the faces of [0, 1]3 , oriented positively if we look at them
from outside the solid cube. (See figure 9.14.)
The second boundary of the standard 2-cube works out by cancellation to

2 2 = 0.

(See the left side of figure 9.15.) And the second boundary of the standard
3-cube similarly is
2 3 = 0.
(See the right side of figure 9.15.) These two examples suggest that the nota-
tional counterpart to the nilpotence of d is also true,

2 = 0.

The nilpotence of is is indeed a theorem, and it is readily shown by a

double sum calculation in which terms cancel pairwise. But it will also follow
immediately from the main theorem of the chapter, the General FTIC, which
478 9 Integration of Differential Forms

Figure 9.14. Boundary of the standard 3-cube

states that in a precise sense the differentiation operator d and the boundary
operator are complementary. Their complementary nature is why they are
notated so similarly.

z
+
+
y

+
+

Figure 9.15. Second boundaries

 9.13 Geometry of Chains: the Boundary Operator 479

Since integration is invariant under reparametrization, you neednt be too

formal in computing boundaries once you understand how they work on stan-
dard cubes. The boundary of the unit square (the 2-cube), for example, is
adequately described as its edge traversed counterclockwise at unit speed,
and so the boundary of any singular 2-cube from the unit square into Rn is
simply the restriction of to the edge of the square with appropriate traver-
sal, or any orientation-preserving reparametrization thereof. In particular, any
rectangle in R2 can be obtained by scaling and translating the unit square in
an orientation-preserving fashion, so the boundary of such a rectangle is, as
one would hope, its edge, counterclockwise. More generally, a singular 2-cube
in R3 is a sort of parametrized membrane floating in space, and its boundary
is just its edge, traversed in the direction inherited from the parametriza-
tion, as we saw for the disk. Without the parametrization, neither direction
of traversing the membranes edge in Rn for any n > 2 is naturally preferable
to the other. Similarly in R3 , the boundary of the unit cube is its six faces,
oriented to look positive from outside the cube. In other words, an acceptable
coordinate system for a boundary face of the cube is two orthonormal vectors
whose cross product is an outward unit normal to the cube. The boundary of
any singular 3-cube : [0, 1]3 R3 is the restriction of to the boundary
faces of [0, 1]3 .
For example, consider the surface

: [0, a] [0, 2] [0, b] R3

given by the cyclindrical coordinate mapping,

(r, , z) = (r cos , r sin , z).

Although the parametrizing box is not literally [0, 1]3 , we grant ourselves
license to treat the upper limits of the parameters as 1 in determining the
signs in the formula

= (31,0 + 31,a + 32,0 32,2 33,0 + 31,b ).

Here we also grant ourselves license to use chain-addition inside the paren-
thesis rather than compose six times. The boundary components, unsigned,
are

( 31,0 )(, z) = (0, 0, z),

( 31,a )(, z) = (a cos , a sin , z),
( 32,0 )(r, z) = (r, 0, z),
( 32,2 )(r, z) = (r, 0, z),
( 33,0 )(r, ) = (r cos , r sin , 0),
( 31,b )(r, ) = (r cos , r sin , b).
480 9 Integration of Differential Forms

The first component maps to the z-axis from 0 to b, which is trivial as a 2-

surface in the sense that integrating any 2-form over it will give 0. The second
component maps to the vertical outside of cylinder (the label on the can), and
the positive sign that it carries connotes that the associated normal vector at
each point of the vertical outside of the cylinder point outwards. The third and
fourth components map to a solid a-by-b rectangle in the (x, z)-plane, inside
the solid cylinder; these are not trivial as 2-surfaces, but the components
carry opposite signs and so they cancel. The last two components map to the
bottom and the top of the cylinder, and the signs that they carry connote that
in each case the natural normal vector points outward. Thus the boundary of
the cylinder is as expected.
For another example, let B3 denote the solid unit ball in R3 . Let a, b, c
be positive numbers, and consider the surface that dilates the ball to the
associated solid ellipsoid

: B3 R3 , (x, y, z) = (ax, by, cz).

Since the parameter domain of is not a box, is not a singular 3-cube

even under the looser grammar that we have granted ourselves. Thus, to
compute the boundary of formally we should pre-parametrize B3 from a
box using the spherical coordinate system, : [0, 1] [0, 2] [0, ] B3 ,
and then understand that really means ( ) = 3 , where the
notation 3 is being stretched a little as in the previous example since the
parameter domain of isnt literally [0, 1]3 . Inevitably, the boundary of the
ball works out to be its spherical skin, although it is unfortunately oriented so
that the natural normal vector points inward in consequence of our spherical
coordinate system reversing orientation. (See exercise 9.13.3.) Consequently
the boundary of the ellipsoid is its skin as well.

Exercises

9.13.1. Define a singular k-cube called the simplex, : [0, 1]k Rk , by

k1
Y
(u1 , , uk ) = (u1 , (1 u1 )u2 , (1 u1 )(1 u2 )u3 , , (1 ui )uk ).
i=1
Pk Qk
(a) Show that if (x1 , , xk ) = (u1 , , uk ) then i=1 xi = 1 i=1 (1
ui ).
(b) Show that the image of lies in the set (also called the simplex)
k
X
S = {(x1 , , xk ) : x1 0, , xk 0, xi 1}.
i=1

(In fact, the image is all of the simplex, but showing this would take us too
far afield.)
 9.14 The General Fundamental Theorem of Integral Calculus 481

(c) For each of the values k = 1, 2, 3, do the following. Calculate (the

result is a (k 1)-chain). Graph by graphing each (k 1)-cube in the
chain and indicating its coefficient (+1 or 1) beneath the graph. Each graph
should show [0, 1]k1 and Rk .
3
p shell H : D R where
9.13.2. Describe the boundary of the hemispherical
2 2 2
D is the unit disk in R and H(x, y) = (x, y, 1 x y ). (You might
parametrize D from [0, 1]2 and then compute the boundary of the composition,
or you might simply push D from the section through H.)
9.13.3. Describe the boundary of the solid unit upper hemisphere

H = {(x, y, z) R3 : x2 + y 2 + z 2 1, z 0}.

(Since H is being described as a set, parametrize it.)

9.13.4. Describe the boundary of the paraboloid : D R3 where again
D is the unit disk in R2 and

(u, v) = (u, v, u2 + v 2 ).

9.13.5. Describe the boundary of : [0, 2] [0, ] R3 where

(, ) = (cos sin , sin sin , cos ).

(Before going straight to calculations, it will help to understand the geometry

of the problem, especially the interpretation of and in the image-space R3 .)
9.13.6. Describe the boundary of : [0, 1] [0, 2] [0, ] R3 where

(, , ) = ( cos sin , sin sin , cos ).

(Again, first make sure that you understand the geometry of the problem, es-
pecially the interpretation of the parametrizing variables in the image-space.)
How does this exercise combine with the result 2 = 0 to bear on exer-
cise 9.13.5?
9.13.7. Fix constants 0 < a < b. Describe the boundary of : [0, 2][0, 2]
[0, 1] R3 where (u, v, t) = (cos u(b + at cos v), sin u(b + at cos v), at sin v).
(First understand the geometry, especially the interpretation of u, v, and t in
the image-space.)

9.14 The General Fundamental Theorem of Integral

Calculus
As mentioned in the previous section, the algebraic encoding d of the deriva-
tive (an analytic operator) and the algebraic encoding of the boundary (a
geometric operator) are complementary with respect to integration:
482 9 Integration of Differential Forms

Theorem 9.14.1 (General FTIC). Let A be an open subset of Rn . Let C

be a k-chain in A, and let be a (k 1)-form on A. Then
Z Z
d = . (9.17)
C C

Before proving the theorem, we study two examples. First, suppose that
k = n = 1, and that the 1-chain C is a singular 1-cube : [0, 1] R taking
0 and 1 to some points a and b. Then the theorem says that for any suitable
smooth function f ,
Z b
f (x) dx = f (b) f (a).
a
This is the one-variable Fundamental Theorem of Integral Calculus. Thus,
whatever else we are doing, we are indeed generalizing it.
Second, to study a simple case involving more than one variable, suppose
that C = 2 (the standard 2-cube) and = f (x, y) dy for some smooth
function f : [0, 1]2 R. The derivative in the left side of (9.17) works out
to
d = D1 f (x, y) dx dy,
Exercise 9.5.4 says that we may drop the wedges from the integral of this
2-form over the full-dimensional surface 2 in 2-space to obtain a chapter 6
function-integral, and so the left side of (9.17) works out to
Z Z Z
d = D1 f (x, y) dx dy = D1 f.
2 2 [0,1]2

Meanwhile, on the right side of (9.17), the boundary 2 has four pieces,
but on the two horizontal pieces dy is zero since y is constant. Thus only the
integrals over the two vertical pieces contribute, giving
Z Z 1 Z 1 Z 1
= f (1, u) f (0, u) = f (1, u) f (0, u).
2 u=0 u=0 u=0

By the one-variable Fundamental Theorem, the integrand is

Z 1
f (1, u) f (0, u) = D1 f (t, u),
t=0

and so by Fubinis Theorem, the integral is

Z 1 Z 1 Z
D1 f (t, u) = D1 f.
u=0 t=0 [0,1]2
R
Thus both sides of (9.17) work out to [0,1]2 D1 f , making them equal as de-
sired, and the General FTIC holds in this case. The first step of its proof is
essentially the same process as in this example.
 9.14 The General Fundamental Theorem of Integral Calculus 483
R R
Proof. Recall that we want to establish formula (9.17), C d = C , where
C is a k-chain and is a (k 1)-form. Begin with the special case where C is
the standard k-cube,
C = k ,
c j dxk , where x = (x1 , , xk )
and takes the form = f (x) dx1 dx
and the c means to omit the term. Thus

= f (x) dxJ where J = (1, , b

j, , k).
R
To evaluate the left side C d of (9.17), we need to compute d. In this
special case

d = Dj f (x) dxj dxJ = (1)j1 Dj f dx(1, ,k) ,

and so by exercise 9.5.4, the left side reduces to the function-integral of the
jth partial derivative over the unit box,
Z Z Z
d = (1)j1 Dj f dx(1, ,k) = (1)j1 Dj f. (9.18)
k k [0,1]k
R
To evaluate the right side C
of (9.17), we need to examine the boundary

k X
X 1
k = (1)i+ ki, ,
i=1 =0

where ki, (u1 , , uk1 ) = (u1 , , ui1 , , ui , , uk1 ). Note that

1 0 0 0 0
0 1 0 0 0

.. .. . . .. .. ..
. . . . . .

0 0 1 0 0
(i, ) =
k
0 0
.
0 0 0
0 0 0 1 0

. . .. .. . . ..
.. .. . . . .
0 0 0 0 1

This derivative matrix is k-by-(k 1), consisting of the identity matrix except
that zeros have been inserted at the ith row, displacing everything from there
downwards. Meanwhile, recall that J = (1, , bj, , k), where the omitted
index j is fixed throughout this calculation. It follows that as the index i of
summation varies, the determinant of the Jth rows of the matrix is
(
k 1 if i = j,
det(i, )J =
0 if i 6= j.
484 9 Integration of Differential Forms

That is, the integral of = f (x) dxJ can be nonzero only for the two terms
in the boundary chain k with i = j, parametrizing the two boundary faces
whose normal vectors point in the direction missing from dxJ :
Z Z
f (x) dxJ = f (x) dxJ
k (1)j+1 (k k
j,1 j,0 )
Z
= (1)j+1 (f kj,1 ) 1 (f kj,0 ) 1.
[0,1]k1

Here the last equality follows from the definition of integration over chains
and the defining formula (9.14). For any point u = (u1 , , uk1 ) [0, 1]k1 ,
the integrand can be rewritten as an integral of the jth partial derivative by
the one-variable Fundamental Theorem of Integral Calculus,

(f kj,1 f kj,0 )(u)

= f (u1 , , uj1 , 1, uj , , uk1 ) f (u1 , , uj1 , 0, uj , , uk1 )
Z
= Dj f (u1 , , uj1 , t, uj , , uk1 ).
t[0,1]

Therefore, the right side of (9.17) is

Z Z Z
= (1)j+1 Dj f (u1 , , uj1 , t, uj , , uk1 ).
k u[0,1]k1 t[0,1]

By Fubinis Theorem this is equal to the right side of (9.18), and so the
General FTIC is proved in the special case.
The rest of the proof is handled effortlessly by the machinery of forms and
chains. A general (k 1)-form on [0, 1]k is
k
X
= j , c j dxk .
each j = fj (x) dx1 dx
j=1
P
Each j is a form of the type covered by the special case, and d = j dj .
So, continuing to integrate over the standard k-cube, and citing the special
case just shown for the crucial equality in the middle,
Z Z X XZ
d = dj = dj
k k j j k
XZ Z X Z
= j = j = .
j k k j k

Thus the theorem holds for a general form when C = k .

For a singular k-cube in A and for any (k 1)-form on A, we now
have
 9.14 The General Fundamental Theorem of Integral Calculus 485
Z Z
d = (d) by the Pullback Theorem
k
Z
= d( ) since derivative commutes with pullback
k
Z
= since the result holds on k
k
Z
by the Change of Variable Theorem for
=
k differential forms, extended to chains
Z
= by definition of boundary.

So the result holds for singular

P cubes.
Finally, for a k-chain C = s s (s) in A and for any (k 1)-form on A,
Z Z X Z X Z
d = P d = s d = s ,
C s s (s) s (s) s (s)

with the third equality due to the result for singular cubes, and the calculation
continues
X Z Z Z Z
s = P = P
= .
s (s) s s (s) ( s s (s) ) C

This completes the proof.

The beauty of this argument is that the only analytic results that it uses
are the one-variable FTIC and Fubinis Theorem, and the only geometry that
it uses is the definition of the boundary of a standard k-cube. All the twisting
and turning of k-surfaces in n-space is filtered out automatically by the algebra
of differential forms.
Computationally, the General FTIC will sometimes give you a choice be-
tween evaluating two integrals, one of which may be easier to work. Note that
the integral of lower dimension may not be the preferable one, however; for
example, integrating over a solid 3-cube may be quicker than integrating over
the six faces of its boundary.
Conceptually the General FTIC is exciting because it allows the possi-
bility of evaluating an integral over a region by antidifferentiating and then
integrating only over the boundary of the region instead.

Exercises

9.14.1. Similarly to the second example before the proof of the General FTIC,
show that the theorem holds when C = 3 and = f (x, y, z) dz dx.
2
9.14.2.
R Prove as a corollary to the General FTIC that = 0, in the sense
that 2 C = 0 for all forms .
486 9 Integration of Differential Forms

9.14.3. Let C be a k-chain in Rn , f : Rn R a function, and a (k 1)-

form on Rn . Use the General FTIC to prove a generalization of the formula
for integration by parts,
Z Z Z
f d = f df .
C C C

4
9.14.4. Let be a 4-chain in R with boundary . Suitably specialize the
General FTIC to prove the identity
Z
f1 dy dz dw + f2 dz dw dx + f3 dw dx dy + f4 dx dy dz

Z
= (D1 f1 D2 f2 + D3 f3 D4 f4 ) dx dy dz dw.

Here the order of the variables is (x, y, z, w).

9.15 Classical Change of Variable Revisited

The most technical argument in these notes is the proof of the classical Change
of Variable Theorem (Theorem 6.7.1) in sections 6.8 and 6.9. The analytic re-
sults contributing to the proof were the one-variable Change of Variable The-
orem and Fubinis Theorem, and of these, the one-variable Change of Variable
Theorem is a consequence of the one-variable FTIC. Meanwhile, the analytic
results contributing to the proof of the General FTIC were the one-variable
FTIC and Fubinis Theorem. Thus the proofs of the multivariable classical
Change of Variable Theorem and of the General FTIC rely on the same anal-
ysis. However, the proof of the General FTIC was easy. Now, with the General
FTIC in hand, we revisit the classical Change of Variable Theorem, sketching
the light, graceful proof that it deserves in turn.
The first issue to address is that the classical Change of Variable Theorem
has been quoted in this chapter, and so if we now propose to revisit its proof
then we must take care not to argue in a circle. In fact, our only uses of the
classical Change of Variable Theorem in this chapter were to prove that inte-
grals of functions over surfaces are independent of reparametrization (the end
of section 9.1) and that integrals of differential forms over surfaces are indepen-
dent of orientation-preserving reparametrization (exercise 9.5.5). The proof of
the General FTIC requires neither the classical Change of Variable Theorem
nor independence of parametrization. Thus this chapter could have proceeded
without the classical Change of Variable theorem, but then requiring us to
remember that all of its results were provisionally parametrization-dependent.
A schematic layout of the ideas is shown in figure 9.16.
Nonetheless, even the parametrization-dependent General FTIC, which we
may grant ourselves without the classical Change of Variable Theorem, is a
powerful result, and in particular it leads to the conceptually different proof
 9.15 Classical Change of Variable Revisited 487

FTIC CoV
+3 Fubini
(n = 1) (n = 1)




s{
CoV
(n > 1)


Z
f ind. of param.

Z
ind. of orient.-pres. param.

'/ FTIC ow
(general)

Figure 9.16. Layout of the main results as established so far

of the classical Change of Variable Theorem. Once the theorem is proved, we

may conclude that the chapters results are independent of parametrization
after all. Being patient gives us the same results more easily. The provisional
new layout of the ideas is shown in figure 9.17. The improved organization is
clear.
Let J be a box in Rn , and consider a smooth change of variable mapping

: J Rn .

(See figure 9.18.) Assume that

det > 0 everywhere on J.

To prove the classical Change of Variable Theorem, we need to show that the
following formula holds for any smooth function f : (J) R:
Z Z
f = (f ) det .
(J) J

View the mapping as a singular n-cube in Rn . (Since J need not be the unit
box, the definition of a singular n-cube is being extended here slightly to allow
any box as the domain. The boundary operator extends correspondingly, as
488 9 Integration of Differential Forms

FTIC CoV
3+ Fubini
(n = 1) (n = 1)





+#
s{
FTIC
(general)


CoV
(n > 1)


Z
f ind. of param.

Z
ind. of orient.-pres. param.

Figure 9.17. Provisional layout of the main results after this section

discussed at the end of section 9.13.) Consider the trivial parametrization of

the image of the cube,

(J) : (J) Rn , (J) (x) = x for all x (J).

Let J be the trivial parametrization of J. In the language of differential

forms, the formula that we need to show is
Z Z
= where = f (x) dx. (9.19)
(J) J

Here x = (x1 , , xn ) and dx = dx1 dxn , and the pullback in the right
side of the equality is = (f )(x) det (x) dx. (Note that applying the
Pullback Theorem (Theorem 9.10.1) reduces the desired formula to
Z Z
= ,
(J)

i.e., to independence of parametrization, the one result in this chapter that

relied on the classical Change of Variable Theorem.) The starting idea of this
section is to try to derive (9.19) from the General FTIC.
 9.15 Classical Change of Variable Revisited 489

Figure 9.18. The singular cube

To see how this might be done, begin by reviewing the derivation of the
one-variable Change of Variable Theorem from the one-variable FTIC, dis-
playing the calculation in two parts,
Z (b) Z (b)
f= F = F ((b)) F ((a)) (9.20)
(a) (a)

and Z Z
b b

(f ) = (F ) = (F )(b) (F )(a). (9.21)
a a
Since the right sides are equal, so are the left sides, giving the theorem. Here
the first version Rof the one-variable FTIC (Theorem 6.4.1) provides the an-
x
tiderivative F = (a) f of f .
Now, starting from the integral on the left side of the desired equal-
ity (9.19), attempt to pattern-match the calculation (9.20) without yet wor-
rying about whether the steps are justified or even meaningful,
Z Z Z
= d = . (9.22)
(J) (J) (J)

Similarly, the integral on the right side of (9.19) looks like the integral at the
beginning of the calculation (9.21), so pattern-match again,
Z Z Z

= d( ) = . (9.23)
J J J

Thus it suffices to show that the right sides are equal,

Z Z
= .
(J) J

This formula looks like the desired (9.19) but with (n1)-dimensional integrals
of (n 1)-forms. Perhaps we are discovering a proof of the multivariable
490 9 Integration of Differential Forms

Change of Variable Theorem by induction on the number of variables. But we

need to check whether the calculation is sensible.
Just as the one-variable calculation rewrote f as F , the putative multi-
variable calculation has rewritten as d, but this needs justification. Recall
that = f (x) dx. Although (J) is not a box, an application of Theorem 6.4.1
to the first variable shows that in the small, f takes the form
f (x1 , x2 , , xn ) = D1 F (x1 , x2 , , xn ).
Consequently the in our calculation can be taken as
= F (x) dx2 dxn ,
provided that whatever we are doing is on a small enough scale. So now we
assume that the box J is small enough that the argument of this paragraph
applies at each point of the nonbox (J). We can do this by partitioning the
original box J finely enough into subboxes J and then carrying out the proof
for each subbox. (Alternatively, by Proposition 6.9.3 we may assume that f
is identically 1 and then take F (x) = x1 . Or, to avoid any specific calculation
we may assume that the box J is small enough that (J) is contractible,
and then has an antiderivative by Poincares Theorem, Theorem 9.11.2.)
Once we have , the objects in (9.23) are noncontroversial and the steps are
clear, except perhaps the tacit exchange of the derivative and the pullback
in the first step of pattern-matching.
R R The remaining issue is what to make of
the symbol-pattern (J) d = (J) in (9.22). Recall that (J) is the
trivial parametrization of (J). However, in dimension n > 1, (J) is not a
box, so (J) is not a cube, and so (J) has no meaning. Even if we know
the topological boundary of (J) (the points arbitrarily close to (J) and
to its complement), the topological boundary inherits no canonical traversal
from the trivial parametrization. The calculation is not sensible.
A 1999 article by Peter Lax in the American Mathematical Monthly shows
how to solve this problem. Recall that we we are working with a mapping
: J Rn .
The box J is compact, and hence so is its continuous image (J). Therefore
some large box B contains them both. If J is small enough then because
det > 0 on J, it follows from some analysis that extends to a mapping
: B Rn
such that
is the original on J,
takes the complement of J in B to the complement of (B) in B,
is the identity mapping on the boundary of B.
(See figure 9.19.) Furthermore, the n-form on the original (J) can be
modified into a form on the larger set B such that
 9.15 Classical Change of Variable Revisited 491

is the original on (J),

= 0 essentially everywhere off the original (J).
And now that the nonbox (J) has been replaced by the box B, the calculation
of the antiderivative form such that = d works in the large.

Figure 9.19. The extension of to B

Let B denote the trivial parametrization of B. Then the properties of

and show that the desired equality (9.19) has become
Z Z
= ,
B B

the integrals on both sides now being taken over the same box B. Again
pattern-matching the one-variable proof shows that the integral on the left
side is Z Z Z
= d =
B B B
and the integral on the right side is
Z Z Z

= d( ) = ,
B B B

where everything here makes sense. Thus the problem is reduced to proving
that Z Z
= ,
B B
And now the desired equality is immediate: since is the identity mapping on
the boundary of B, the pullback in the right-side integral of the previous
display does nothing, and the two integrals are equal. (See exercise 9.15.1 for a
slight variant of this argument.) The multivariable argument has ended exactly
492 9 Integration of Differential Forms

as the one-variable argument did. We did not need to argue by induction after
all.
In sum, the General FTIC lets us side-step the traditional proof of the
classical Change of Variable Theorem, by expanding the environment of the
problem to a larger box and then reducing the scope of the question to the
larger boxs boundary. On the boundary there is no longer any difference
between the two quantities that we want to be equal, and so we are done.
The reader may well object that the argument here is only heuristic, and
that there is no reason to believe that its missing technical details will be any
less onerous than those of the usual proof the classical Change of Variable
Theorem. The difficulty of the usual proof is that it involves nonboxes, while
the analytic details of how this argument proceeds from the nonbox (J) to
a box B were not given. Along with the extensions of and to B being
invoked, the partitioning of J into small enough subboxes was handwaved.
Furthermore, the change of variable mapping is assumed here to be smooth,
whereas in Theorem 6.7.1 it need only be C 1 . But none of these matters is
serious. A second article by Lax, written in response to such objections, shows
how to take care of them. Although some analysis is admittedly being elided
here, the new argument nonetheless feels more graceful to the author of these
notes than the older one.

Exercise

9.15.1. Show that in the argument at the end of the section, we could instead
reason about the integral on the right side that
Z Z Z
= d = .
B
R R
Thus the problem is reduced to proving that B
=
. Explain why
the desired equality is immediate.

9.16 The Classical Theorems

The classical integration theorems of vector calculus arise from specializing n
and k in the General FTIC. As already noted, the values n = k = 1 give the
one-variable FTIC,
Z b
df
dx = f (b) f (a).
a dx
If k = 1 but n is left arbitrary then the result is familiar from section 9.4. For
any curve : [0, 1] Rn , let a = (0) and b = (1). Then
Z
f f
dx1 + + dxn = f (b) f (a).
x 1 x n
 9.16 The Classical Theorems 493

Setting n = 2, k = 2 gives Greens Theorem: Let A be an open subset

of R2 . For any singular 2-cube in A and functions f, g : A R,
ZZ   Z
g f
dx dy = f dx + g dy.
x y

The double integral sign is used on the left side of Greens Theorem to em-
phasize that the integral is two-dimensional. Naturally the classical statement
doesnt refer to a singular cube or include a wedge. Instead, the idea classi-
cally is to view as a set in the plane and require a traversal of (also
viewed as a set) such that is always to the left as one moves along .
Other than this, the boundary integral is independent of how the boundary is
traversed because the whole theory is invariant under orientation-preserving
reparametrization. (See figure 9.20.)

Figure 9.20. Traversing the boundary in Greens Theorem

Greens Theorem has two geometric interpretations. To understand them,

first let A R2 be open and think of a vector-valued mapping F~ : A R2
as defining a fluid flow in A. Define two related scalar-valued functions on A,

curl F~ = D1 F2 D2 F1 and div F~ = D1 F1 + D2 F2 .

These are two-dimensional versions of the quantities from exercises 9.8.4

and 9.8.5. Now consider a point p in A. Note that curl F~ (p) and div F~ (p)
depend only on the derivatives of F~ at p, not on F~ (p) itself. So replacing F~
by F~ F~ (p), we may assume that F~ (p) = 0, i.e., the fluid flow is stationary
at p. Recall that D1 F2 is the rate of change of the vertical component of F
with respect to change in the horizontal component of its input, and D2 F1 is
the rate of change of the horizontal component of F with respect to change
494 9 Integration of Differential Forms

in the vertical component of its input. The left side of figure 9.21 shows a
scenario where the two terms D1 F2 and D2 F1 of (curl F~ )(p) are positive.
The figure illustrates why curl F~ is interpreted as measuring the vorticity of F~
at p, its tendency to rotate a paddle-wheel at p counterclockwise. Similarly,
D1 F1 is the rate of change of the horizontal component of F with respect
to change in the horizontal component of its input, and D2 F2 is the rate of
change of the vertical component of F with respect to change in the vertical
component of its input. The right side of figure 9.21 shows a scenario where
the terms of (div F~ )(p) are positive. The figure illustrates why div F~ is viewed
as measuring the extent that fluid is spreading out from p, i.e., how much fluid
is being pumped into or drained out of the system at the point. Specifically,
the left side of the figure shows the vector field

F~ (x, y) = (y, x)

whose curl and divergence at the origin are

(curl F~ )(0) = 2, (div F~ )(0) = 0,

and the right side shows (with some artistic license taken to make the figure
legible rather than accurate) the vector field

F~ (x, y) = (x, y)

whose curl and divergence at the origin are

(curl F~ )(0) = 0, (div F~ )(0) = 2.

Figure 9.21. Positive curl and positive divergence

For the two geometric interpretations of Greens Theorem, introduce the

notation
 9.16 The Classical Theorems 495

dA = dx dy, ~ = (dx, dy),

ds ~ = (dy, dx).
dn

The form-vectors ds ~ and dn ~ on are viewed respectively as differential

increment around the boundary and differential outward normal (see ex-
ercise 9.16.1), while dA is differential area. Then setting F~ = (f, g) and
F~ = (g, f ) respectively shows that Greens Theorem says
ZZ Z ZZ Z
curl F~ dA = ~
F~ ds and div F~ dA = ~
F~ dn.

The resulting two interpretations are

the net counterclockwise vorticity of F~ throughout
equals the net flow of F~ counterclockwise around
and
the net positive rate of creation of fluid by F~ throughout
equals the net flux of F~ outward through .
These interpretations appeal strongly to physical intuition.
We can also bring dimensional analysis to bear on the the integrals in
Greens Theorem. Again view the vector field F~ as a velocity field describing
a fluid flow. Thus each component function of F~ carries units of length over
time (for instance, m/s). The partial derivatives that make up curl F~ and div F~
are derivatives with respect to space-variables, so the curl and the divergence
carry units of reciprocal time (1/s). The units of the area-integral on the left
side of Greens theorem are thus area over time (1/s m2 = m2 /s), as are the
units of the path-integral on the right side (m/s m = m2 /s as well). Thus
both integrals measure area per unit of time. If the fluid is incompressible then
area of fluid is proportional to mass of fluid, and so both integrals essentially
measure fluid per unit of time: the amount of fluid being created throughout
the region per unit of time, and the amount of fluid passing through the
boundary per unit of time; or the amount of fluid circulating throughout the
region per unit of time, and the amount of fluid flowing along the boundary
per unit of time.
The physical interpretations of divergence and curl will be discussed more
carefully in the next section.
Setting n = 3, k = 2 gives Stokess Theorem: Let A be an open subset
of R3 . For a singular 2-cube in A and functions f, g, h : A R,
ZZ      
h g f h g f
dy dz + dz dx + dx dy
y z z x x y
Z
= f dx + g dy + h dz.

Introduce the notation

496 9 Integration of Differential Forms

~ = (dx, dy, dz)

ds and ~ = (dy dz, dz dx, dx dy),
dn

and for a vector-valued mapping F~ : R3 R3 define

curl F~ = (D2 F3 D3 F2 , D3 F1 D1 F3 , D1 F2 D2 F1 ).

Then setting F~ = (f, g, h) shows that Stokess Theorem is

ZZ Z
~ ~
curl F dn = ~
F~ ds

As with Greens Theorem, the classical statement doesnt refer to a singular

cube or include a wedge. Instead, is an orientable two-dimensional set in
space, and its boundary is traversed counterclockwise about its normal
vectors. The integrals in the previous display are both independent of how
and are parametrized, provided that the geometry is just described.
To interpret Stokess Theorem, think of a mapping F~ : R3 R3 as de-
scribing a fluid flow in space. The mapping curl F~ is interpreted as measuring
the local vorticity of F~ around each positive coordinate direction. The form-
~ on is viewed as differential outward normal, while ds
vector dn ~ on is
viewed as differential increment around the boundary. Thus the interpretation
of Stokess Theorem is a 3-dimensional version of the first interpretation of
Greens Theorem,
the net tangent vorticity of F~ throughout
equals the net flow of F~ around .

Setting n = 3, k = 3 gives the Divergence Theorem (or Gausss The-

orem): Let A be an open subset of R3 . For a singular 3-cube in A and
functions f, g, h : A R,
ZZZ   ZZ
f g h
+ + dx dy dz = f dy dz + g dz dx + h dx dy.
x y z

Introduce the notation

dV = dx dy dz,
and for a vector-valued mapping F~ : R3 R3 define

div F~ = D1 F1 + D2 F2 + D3 F3 .

Then setting F~ = (f, g, h) shows that the Divergence Theorem is

ZZZ ZZ
~
div F dV = ~
F~ dn

Thus the interpretation of the Divergence Theorem is a 3-dimensional version

of the second interpretation of Greens Theorem,
 9.16 The Classical Theorems 497

the net positive creation of fluid by F~ throughout

equals the net flux of F~ outward through .
Again, the classical theorem views and as sets, so long as whatever
parametrization of is used to compute the right-side integral has the same
orientation as the boundary of the parametrization of used to compute the
left-side integral.

Exercises
9.16.1. (a) Let : [0, 1] R2 , t 7 (t) be a curve, and recall the form-
vectors on R2 ds~ = (dx, dy), dn~ = (dy, dx). Compute the pullbacks (ds)~
~ and explain why these are interpreted as differential tangent and
and (dn)
normal vectors to .
(b) Let : [0, 1] R3 , t 7 (t) be a curve and : [0, 1]2 R3 ,
~ = (dx, dy, dz),
(u, v) 7 (u, v) a surface, and recall the form-vectors on R3 ds
~ ~
dn = (dydz, dzdx, dxdy). Compute the pullbacks (ds) and (dn) ~ and
explain why these are interpreted respectively as differential tangent vector
to and differential normal vector to .
9.16.2. Use Greens Theorem to show that for a planar region ,
Z Z
area() = x dy = y dx.

Thus one can measure the area of a planar set by traversing its bound-
ary. (This principle was used to construct ingenious area-measuring machines
called planimeters before Greens Theorem was ever written down.)
9.16.3. Let H be the upper unit hemispherical shell,
H = {(x, y, z) R3 : x2 + y 2 + z 2 = 1, z 0}.
Define a vector-valued function on R3 ,
F (x, y, z) = (x + y + z, xy + yz + zx, xyz).
RR
Use Stokess Theorem to calculate H curl F dn.~

9.16.4. Use the Divergence Theorem to evaluate

Z
x2 dy dz + y 2 dz dx + z 2 dx dy,
H

where H is the boundary of the solid unit hemisphere

H = {(x, y, z) R3 : x2 + y 2 + z 2 1, z 0}.
(Thus H is the union of the unit disk in the (x, y)-plane and the unit upper
hemispherical shell.) Feel free to cancel terms by citing symmetry if youre
confident of what youre doing.
498 9 Integration of Differential Forms

9.16.5. Let g and h be functions on R3 . Recall the operator = (D1 , D2 , D3 ),

which takes scalar-valued functions to vector-valued functions. As usual, define
the Laplacian operator to be = D11 + D22 + D33 . From an earlier exercise,
= div grad.
(a) Prove that div (g h) = g h + g h.
(b) If D is a closed compact subset of R3 with positively oriented bound-
ary D, prove that
ZZZ ZZ
(g h + g h) dV = ~
g h dn.
D D

(Here n is the unit outward normal to D and hn is the directional derivative

of h in the direction of n.) Interchange g and h and subtract the resulting
formula from the first one to get
ZZZ ZZ
(g h h g) dV = ~
(g h h g) dn.
D D

These two formulas are Greens identities.

(c) Assume that h is harmonic, meaning that it satisfies the harmonic
equation h = 0.
Take g = h and use Greens first identity to conclude that if h = 0 on the
boundary D then h = 0 on all of D.
Take g = 1 and use Greens second identity to show that
ZZ
~ = 0.
h dn
D

What does this say about harmonic functions and flux?

9.17 Divergence and Curl in Polar Coordinates

The picture-explanations given in the previous section to interpret the diver-

gence and the curl are not entirely satisfying. Working with the polar coor-
dinate system further quantifies the ideas and makes them more coherent by
applying to both operators in the same way.
Rather than study the divergence and the curl of a vector field Fe at a
general point p, we may study the divergence and the curl of the modified
vector field
F (x) = Fe (x + p) Fe (p)
at the convenient particular point 0, at which the value of F is 0 as well. That
is, we may normalize the point p to be 0 by prepending a translation of the
domain, and we also may normalize F (0) to 0 by postpending a translation
of the range. With this in mind, let A R2 be an open set that contains the
 9.17 Divergence and Curl in Polar Coordinates 499

origin, and let F be a continuous vector field on A that is stationary at the

origin,
F = (f1 , f2 ) : A R2 , F (0) = 0.
At any point other than the origin, F resolves into a radial component and
an angular component. Specifically,

F = Fr + F ,

where

Fr = fr r, fr = F r, r = (cos , sin ) = (x, y)/|(x, y)|,

F = f ,
f = F , = r = ( sin , cos ) = (y, x)/|(x, y)|.

(Recall that the unary cross product (x, y) = (y, x) in R2 rotates vectors
90 degrees counterclockwise.) Here fr is positive if Fr points outward and
negative if Fr points inward, and f is positive if F points counterclockwise
and negative if F points clockwise. Since F (0) = 0, the resolution of F into
radial and angular components extends continuously to the origin, fr (0) =
f (0) = 0, so that Fr (0) = F (0) = 0 even though r and are undefined at
the origin.
The goal of this section is to express the divergence and the curl of F
at the origin in terms of the polar coordinate system derivatives that seem
naturally suited to describe them, the radial derivative of the scalar radial
component of F ,
fr (r cos , r sin )
Dr fr (0) = lim+ ,
r0 r
and the radial derivative of the scalar angular component of F ,
f (r cos , r sin )
Dr f (0) = lim+ .
r0 r
However, matters arent as simple here as one might hope. For one thing,
the limits are stringent in the sense that they must always exist and take
the same values regardless of how behaves as r 0+ . Also, although F
is differentiable at the origin if its vector radial and angular components Fr
and F are differentiable at the origin, the converse is not true. So first we
need sufficient conditions for the converse, i.e., sufficient conditions for the
components to be differentiable at the origin. Necessary conditions are always
easier to find, so Proposition 9.17.1 will do so, and then Proposition 9.17.2 will
show that the necessary conditions are sufficient. The conditions in question
are the CauchyRiemann equations,

D1 f1 (0) = D2 f2 (0),
D1 f2 (0) = D2 f1 (0).
500 9 Integration of Differential Forms

When the CauchyRiemann equations hold, we can describe the divergence

and the curl of F at the origin in polar terms, as desired. This will be the
content of Theorem 9.17.3.
Before we proceed to the details, a brief geometric discussion of the
CauchyRiemann equations may be helpful. The equation D1 f1 = D2 f2 de-
scribes the left side of figure 9.22, in which the radial component of F on
the horizontal axis is growing at the same rate as the radial component on
the vertical axis. Similarly, the equation D2 f1 = D1 f2 describes the right
side of the figure, in which the angular component on the vertical axis is
growing at the same rate as the angular component on the horizontal axis.
Combined with differentiability at the origin, these two conditions will imply
that moving outward in any direction, the radial component of F is growing
at the same rate as it is on the axes, and similarly for the angular component.
Thus the two limits that define the radial derivatives of the radial and angular
components of F at 0 (these were displayed in the previous paragraph) are
indeed independent of . An example of this situation, with radial and angular
components both present, is shown in figure 9.23.

Figure 9.22. Geometry of the CauchyRiemann equations individually

As mentioned, the necessity of the CauchyRiemann equations is the nat-

ural starting point.

Proposition 9.17.1 (Polar Differentiability Implies Differentiability

and the CauchyRiemann Equations). Let A R2 be an open set that
contains the origin, and let F be a continuous vector field on A that is sta-
tionary at the origin,

F = (f1 , f2 ) : A R2 , F (0) = 0.
 9.17 Divergence and Curl in Polar Coordinates 501

Figure 9.23. Geometry of the CauchyRiemann equations together

Assume that the vector radial and angular components Fr and F of F are
differentiable at the origin. Then F is differentiable at the origin, and the
CauchyRiemann equations hold at the origin.

For example, the vector field F (x, y) = (x, 0) is differentiable at the ori-
gin, but since D1 f1 (0) = 1 and D2 f2 (0) = 0, it does not satisfy the Cauchy
Riemann equations, and so the derivatives of the radial and angular compo-
nents of F at the origin do not exist.

Proof. As already noted, the differentiability of F at the origin is immediate

since F = Fr + F and the sum of differentiable mappings is again differen-
tiable. We need to establish the CauchyRiemann equations.
The radial component Fr is stationary at the origin, and we are given
that it is differentiable at the origin. By the componentwise nature of differ-
entiability, the first component Fr,1 of Fr is differentiable at the origin, and
so necessarily both partial derivatives of Fr,1 exist at 0. Since Fr,1 vanishes
on the y-axis, the second partial derivative is 0. Thus the differentiability
criterion for the first component of Fr is

Fr,1 (h, k) hD1 Fr,1 (0) = o(h, k).

To further study the condition in the previous display, use the formula
( f (x,y)
r
(x, y) if (x, y) 6= 0,
Fr (x, y) = |(x,y)|
0 if (x, y) = 0

to substitute h fr (h, k)/|(h, k)| for Fr,1 (h, k). Also, because F is angular, F,1
vanishes on the x-axis, and so D1 F,1 (0) = 0; thus, since f1 = Fr,1 + F,1 ,
we may substitute D1 f1 (0) for D1 Fr,1 (0) as well. Altogether the condition
becomes
502 9 Integration of Differential Forms

h(fr (h, k)/|(x, y)| D1 f1 (0)) = o(h, k).

A similar argument using the second component Fr,2 of Fr shows that

k(fr (h, k)/|(x, y)| D2 f2 (0)) = o(h, k).

And so we have shown the first CauchyRiemann equation and a little more,
fr (h, k)
lim = D1 f1 (0) = D2 f2 (0).
(h,k)0 |(h, k)|

For the second CauchyRiemann equation we could essentially repeat the

argument just given, but a quicker way is to consider the radial component

of the vector field F = f r fr ,
(
f (x,y)
(x, y) if (x, y) 6= 0,
(F )r (x, y) = |(x,y)|
0 if (x, y) = 0.

This radial component is differentiable at the origin since it is a rotation of the

angular component of the original F , which we are given to be differentiable at
the origin. And F = (f2 , f1 ) in Cartesian coordinates, so as just argued,

f (h, k)
lim = D1 f2 (0) = D2 f1 (0).
(h,k)0 |(h, k)|

This last display encompasses the second CauchyRiemann equation at the

origin.
Note that the argument has used the full strength of the hypotheses, i.e.,
it has used the differentiability at the origin of each component function of Fr
and each component function of F .

As mentioned, the converse to Proposition 9.17.1 holds too.

Proposition 9.17.2 (Differentiability and the CauchyRiemann

Equations Imply Polar Differentiability). Let A R2 be an open set
that contains the origin, and let F be a continuous vector field on A that is
stationary at the origin,

F = (f1 , f2 ) : A R2 , F (0) = 0.

Assume that F is differentiable at the origin, and assume that the Cauchy
Riemann equations hold at the origin. Then the vector radial and angular
components Fr and F are differentiable at the origin.

Proof. Let a = D1 f1 (0) and let b = D1 f2 (0). By the CauchyRiemann equa-

tions, also a = D2 f2 (0) and b = D2 f1 (0), so that the Jacobian matrix of F
at 0 is  
a b
F (0) = .
b a
 9.17 Divergence and Curl in Polar Coordinates 503

The condition that F is differentiable at 0 is

F (h, k) (ah bk, bh + ak) = o(h, k).

Decompose the quantity in the previous display into radial and angular com-
ponents,



F (h, k) (ah bk, bh + ak) = Fr (h, k) a(h, k) + F (h, k) b(k, h) .

Since the components are at most as long as the vector,

Fr (h, k) a(h, k) = o(h, k) and F (h, k) b(k, h) = o(h, k).

That is, Fr and F are differentiable at the origin with respective Jacobian
matrices    
a0 0 b
Fr (0) = and F (0) = .
0a b 0
This completes the proof.

Now we can return to the divergence and the curl.

Theorem 9.17.3 (Divergence and Curl in Polar Coordinates). Let

A R2 be a region of R2 containing the origin, and let F be a continuous
vector field on A that is stationary at the origin,

F = (f1 , f2 ) : A R2 , F (0) = 0.

Assume that F is differentiable at the origin and that the CauchyRiemann

equations hold at the origin. Then the radial derivatives of the scalar radial
and angular components of F at the origin,

fr (r cos , r sin )
Dr fr (0) = lim+
r0 r
and
f (r cos , r sin )
Dr f (0) = lim ,
r0+ r
both exist independently of how behaves as r shrinks to 0. Furthermore,
the divergence of F at the origin is twice the radial derivative of the radial
component,
(div F )(0) = 2Dr fr (0),
and the curl of F at the origin is twice the radial derivative of the angular
component,
(curl F )(0) = 2Dr f (0).
504 9 Integration of Differential Forms

Proof. By Proposition 9.17.2, the angular and radial components of F are

differentiable at the origin, so that the hypotheses of Proposition 9.17.1 are
met. The first limit in the statement of the theorem was calculated in the
proof of Proposition 9.17.1,
fr (h, k)
Dr fr (0) = lim = D1 f1 (0) = D2 f2 (0).
(h,k)0 |(h, k)|
This makes the formula for the divergence immediate,

(div F )(0) = D1 f1 (0) + D2 f2 (0) = 2Dr fr (0).

Similarly, again recalling the the proof of Proposition 9.17.1,

f (h, k)
Dr f (0) = lim = D1 f2 (0) = D2 f1 (0),
(h,k)0 |(h, k)|

so that
(curl F )(0) = D1 f2 (0) D2 f1 (0) = 2Dr f (0).


If F is a velocity field then the limit in the formula
f (r cos , r sin )
(curl F )(0) = 2 lim+
r0 r
has the interpretation of the angular velocity of F at the origin. That is:
When the CauchyRiemann equations hold, the curl is twice the an-
gular velocity.
Indeed, the angular velocity away from the origin is by definition the rate
of increase of the polar angle with the motion of F . This is not the counter-
clockwise component f , but rather = f /r, i.e., is the function called g
in the proof of Proposition 9.17.1. To understand this, think of a uniformly
spinning disk such as a record on a turntable. At each point except the center,
the angular velocity is the same. But the speed of motion is not constant over
the disk, it is the angular velocity times the distance from the center. That is,
the angular velocity is the speed divided by the radius, as claimed. In these
terms, the proof showed that the angular velocity extends continuously to 0,
and that (curl F )(0) is twice the extended value (0).
Also, if F is a velocity field then the right side of the formula
fr (r cos , r sin )
(div F )(0) = 2 lim+
r0 r
has the interpretation of the flux density of F at the origin. That is:
When the CauchyRiemann equations hold, the divergence is the flux
density.
 9.17 Divergence and Curl in Polar Coordinates 505

To understand this, think of a planar region of incompressible fluid about the

origin, and let r be a positive number small enough that the fluid fills the area
inside the circle of radius r. Suppose that new fluid being added throughout
the interior of the circle, at rate c per unit of area. Thus fluid is being added
to the area inside the circle at total rate r2 c. Here c is called the flux density
over the circle and it is is measured in reciprocal time units, while the units
of r2 c are area over time. Since the fluid is incompressible, r2 c is also the
rate at which fluid is passing normally outward through the circle. And since
the circle has circumference 2r, fluid is therefore passing normally outward
through each point of the circle with radial velocity

r2 c rc
fr (r cos , r sin ) = = .
2r 2
Consequently,
fr (r cos , r sin )
2 = c.
r
Now let r shrink to 0. The left side of the display goes to the divergence of F
at 0, and the right side becomes the continuous extension to radius 0 of the
flux density over the circle of radius r. That is, the divergence is the flux
density when fluid is being added at a single point.

Exercises

9.17.1. Put R2 into correspondence with the complex number field C as fol-
lows:  
x
x + i y.
y
Show that the correspondence extends to
  
a b x
(a + i b)(x + i y).
b a y

Show also that the correspondence preserves absolute value, i.e.,

 
x
| | = |x + i y|,
y

where the first absolute value is on R2 and the second one on C.

9.17.2. Let A R2 be an open set that contains the origin, and let F :
A R2 be a vector field on A that is stationary at the origin. Define a
complex-valued function of a complex variable corresponding to F ,

f (x + iy) = f1 (x, y) + if2 (x, y), (x, y) A.

Then f is called complex-differentiable at 0 if the following limit exists:

506 9 Integration of Differential Forms

f (z + z) f (z)
lim .
z0 z
The limit is denoted f (z).
(a) Suppose that f is complex-differentiable at 0. Compute f (z) first by
letting z go to 0 along the x-axis, and again by letting z go to 0 along
the y-axis. Explain how your calculation shows that the CauchyRiemann
equations hold at 0.
(b) Show also that if f is complex differentiable at 0 then F is vector
differentiable at 0, meaning differentiable in the usual sense. Suppose that f
is complex-differentiable at 0, and that f (0) = rei . Show that

(div F )(0) = 2r cos , (curl F )(0) = 2r sin .

(c) Suppose that F is vector-differentiable at 0 and that the Cauchy

Riemann equations hold at 0. Show that f is complex-differentiable at 0.

9.18 Summary

The bulk of the ideas introduced in this chapter are algebraic. Even so, the
General FTIC subsumes the three classical integration theorems of vector
calculus, and it eases the proof of the classical Change of Variable Theorem.
Index

absolute value of a vector, 32 CauchyRiemann equations, 499

absolute value on Cck (Rn ), 356 CauchySchwarz Inequality, 33
properties of, 357 centroid, 306
addition of matrices, 76 chain, 472
affine mapping, 75 Chain Rule, 151
annulus, 313 in coordinates, 164
ArithmeticGeometric Mean Inequality, change of scale principle, 323
242 change of variable mapping
for cylindrical coordinates, 317
ball, 51 for polar coordinates, 312
n-dimensional of radius r, 323 for spherical coordinates, 318
closed, 206 Change of Variable Theorem
basis of Rn , 28
for differential forms, 465
negative, 124
for multiple integrals, 315
positive, 124
for one-variable integrals, 285
standard, 27
characteristic function of a set, 289
Bernoullis Inequality, 4
characteristic polynomial, 187
BolzanoWeierstrass Property
circle in a field, 382
in R, 54
cissoid of Diocles, 385
in Rn , 54
boundary closed set, 52
of a closed ball, 206 closure of a set, 334
boundary mapping, 475 common refinement of two partitions,
boundary point of a set, 335 265
bounded compact set, 55
function, 258, 262 complement of a set, 205
mapping, 262 completeness of the real number system,
set, 53 2
box as a binary search criterion, 2
compact, 262 as a monotonic sequence criterion, 3
face of, 115 as a set-bound criterion, 2
unit, 114 component of a vector
normal, 40
cartesian product, 261 parallel, 40
508 Index

componentwise nature of linearity, 68 parameter interval, 388

conchoid of Nicomedes, 383 parameter value, 388
congruence, 411 parametrization by arc length, 395
connected set, 57 point, 388
continuity rectifiable, 390
at a point, 44 regular, 389
componentwise nature of, 46 speed, 389
- at a point, 273 tangent vector, 389
- on a set, 273 torsion, 403
of the partial differentiation mapping, trace, 388
358 velocity vector, 389
on a set, 44 cycloid, 390
pointwise, 277 cylindrical coordinates, 316
preservation under composition, 45
preservation under uniform limit, 375 derivative
sequential at a point, 273 linearity of the, 150
sequential on a set, 273 of a convolution, 364
uniform on a set, 277 of a differential form, 446
vector space properties of, 45 of a mapping, 146
contractible space, 467 uniqueness of, 148
convergence determinant, 92, 105
componentwise nature of, 44 of a basis of Rn , 124
linearity of, 44 properties of, 106
convolution, 363 difference of sets, 362
approximation of Cc0 (Rn )-functions differentiability
by, 372 componentwise nature of, 149
approximation of Cck (Rn )-functions differential form, 428
by, 373 closed, 450
approximation of Ic (Rn )-functions derivative of, 446
by, 377 exact, 450
Critical Point Theorem, 179 pullback of, 455
cross product standard presentation of, 431
definition of, 128 differential forms
in Rn , 426 integration of, 430
properties of, 128 meaning of integration of, 439
cube Product Rule for, 448
singular, 472 syntax of, 428
standard, 472 wedge product of, 444
curl of a vector-valued mapping, 451 differentiation under the integral sign,
in polar coordinates, 503 308
curve, parametrized, 388 Dirac delta function, 354, 369
arc length, 389 directional derivative, 189
binormal vector, 402 discrete set, 58
curvature, 398, 402 distance in Rn , 37
equivalence, 395 properties of, 37
Frenet equations, 399, 403 divergence of a vector-valued mapping,
Frenet frame, 398 451
Frenet vector, 408 in polar coordinates, 503
normal vector, 398, 402 Divergence Theorem, 496
 Index 509

geometric interpretation of, 496 helix, 390

H
olders Inequality, 244
equivalence relation, 395 homogeneous function, 168, 471
Euclidean number, 380 homotopy, 467
Euclidean space, 23
Euclids Least Area Problem, 237 identity matrix, 78
Eulers identity for homogeneous Implicit Function Theorem, 227
functions, 471 induction theorem, 3
Extreme Value Theorem, 4, 56 inner product
of two vectors, 31
field axioms, 1 properties of, 31
flow integral Integrability Criterion, 269
2-dimensional integral
coordinate, 423 iterated, 298
intrinsic, 423 lower, of a function over a box, 267
3-dimensional of a Cc0 (Rn )-function, 363
coordinate, 424 of a differential form over a chain, 473
general, 425 of a differential form over a surface,
flux integral 430
2-dimensional of a function over a box, 267
coordinate, 424 of a function over a surface, 419
intrinsic, 424 upper, of a function over a box, 267
3-dimensional integral curve, 195
coordinate, 424 interior point of a set, 145
intrinsic, 424 Intermediate Value Theorem, 4
general interval
coordinate, 426 compact, 257
intrinsic, 425 Inverse Function Theorem, 210
Fubinis Theorem, 299 inverse image characterization of
Fundamental Theorem of Integral continuity, 206
Calculus inverse image of a mapping, 206
first version, 5 isolated point of a set, 52
General, 482
one-variable Jacobian matrix, 146
first version, 283
second version, 284 Known-Integrable Function, 295
second version, 5 known-integrable function, 374

gamma function, 325 Lagrange Multiplier Condition, 244

Gausss Theorem, 496 Laplaces equation, 172, 452
gradient, 189 Laplacian operator, 498
gradient of a scalar-valued function, 450 length
GramSchmidt process, 41 of an interval, 258
graph of a function, 291 level set
Greens identities, 498 of a function, 191
Greens Theorem, 493 of a mapping, 217
geometric interpretations of, 495 limit of a function, 59
limit point of a set, 52
harmonic equation, 452 line in R3 , 131
510 Index

line in a field, 382 matrix-by-vector, 69

linear combination, 28 scalar-by-matrix, 76
linear equation
homogeneous, 91 nilpotence of d, 449
inhomogeneous, 91 norm of a linear mapping, 75
Linear Invertibility Theorem, 99 open set, 205
linear mapping, 64 order axioms, 2
transpose of, 74 ordered k-tuple from {1, , n}, 428
lower sum, 259, 263 increasing, 431
orthogonal vectors, 38
mapping
affine, 75 Pappuss Theorem, 323
component functions of, 46 parallelepiped, 116
from Rn to Rm , 41 face of, 117
from A to Rm , 41 parameter domain of a surface, 414
linear, 64 partial derivative, 157
null, 60 mixed, 169
rigid, 410 second order, 169
matrix, 69 partition
echelon, 88 of a box, 262
elementary, 83 of an interval, 258
indefinite, 183 path-connected set, 59
lower triangular, 90 permutation, 103
negative definite, 183 inversion in, 103
new column of, 88 sign of, 103
old column of, 88 Persistence of Inequality, 49
orthogonal, 100, 410 plane in R3 , 131
positive definite, 183 point in a field, 382
recombine, 83 polar coordinates, 311
scale, 84 Product Rule
second derivative, 179 for differential forms, 448
skew-symmetric, 100 product rule
transpose of, 80 multivariable, 153
transposition, 84 pullback
triangular, 97 contravariance of, 459
matrix inversion algorithm, 89 properties of, 457
matrix multiplication pullback of a differential form, 455
properties of, 79 Pullback Theorem, 464
Max/min Test PullbackDeterminant Theorem, 456
general, 187 pulse function, 361
two-variable, 184 pushforward of a surface, 474
Mean Value Theorem, 5 quotient rule
modulus multivariable, 153
of a vector, 32
properties of, 33 real number system, 1
mollifying kernel, 363 refinement of a partition, 264
multilinear function, 156 reparametrization of a surface, 421
multiplication orientation-preserving, 441
matrix-by-matrix, 77 orientation-reversing, 441
 Index 511

saddle point, 181 Taylor remainder, 8

sawtooth function, 370 integral form, 20
scalar multiplication, 25 Lagrange form, 12
sequence, 42 Taylors Theorem, 11
convergence of, 43 test approximate identity, 368
limit of, 43 test function, 355
null, 42 The Continuous Image of a Compact
sequential characterization Set is Compact, 56
of bounded sets, 54 topological property, 57
of closed sets, 52 transpose
of compact sets, 55 of a linear mapping, 74
of limit points, 52 of a matrix, 80
simplex, 480 Triangle Inequality, 35
n-dimensional of side r, 310 full, 36
unit in R3 , 306
Size Bounds for a vector, 37 upper sum, 259, 263
smooth function, 353
smooth mapping, 415 Vandermonde matrix, 111
Snells Law, 238 vector addition, 24
special orthogonal group, 410 vector space axioms, 26
spherical coordinates, 318 vector space over a field, 27
Stokess Theorem, 495
volume
geometric interpretation of, 496
k-dimensional in Rn
subbox of a partition, 262
of a parallelepiped, 416
subinterval of a partition, 258
of a surface, 419
subsequence, 54
of a box, 262
substitution
of a set, 289
forward, in one-variable integrals, 286
volume form, 440
inverse, in one-variable integrals, 286
Volume Zero Criterion, 291
sum of sets, 362
support of a function, 355
surface, 414 wedge product, 444
properties of, 445
Taylor polynomial, 7 skew symmetry of, 445