Академический Документы
Профессиональный Документы
Культура Документы
Editors
F. W. Gehring
P. R. Halmos
Advisory Board
C. DePrima
I. Herstein
J. Kiefer
Wendell Fleming
Functions of
Several Variables
2nd Edition
Springer-Verlag
New York Heidelberg Berlin
Wende II Fleming
Brown University
Department of Mathematics
Providence, Rhode Island 02912
Editorial Board
F. W. Gehring P. R. Halmas
University of Michigan University of California
Department of Mathematics Department of Mathematics
Ann Arbor, Michigan 48104 Santa Barbara, California 93106
Vll
Preface
Wendell H. Fleming
viii
Contents
Chapter 1
Euclidean spaces I
1.1 The real number system 2
1.2 Euclidean P 5
1.3 Elementary geometry of En 10
1.4 Basic topological notions in En 14
*1.5 Convex sets 19
Chapter 2
Elementary topology of En 28
2.1 Functions 28
2.2 Limits and continuity of transformations 31
2.3 Sequences in En 37
2.4 Bolzano-Weierstrass theorem 43
2.5 Relative neighborhoods, continuous transformations 47
2.6 Topological spaces 50
2.7 Connectedness 56
2.8 Compactness 60
2.9 Metric spaces 62
2.10 Spaces of continuous functions 67
*2.11 Noneuclidean norms on En 70
Chapter 3
Differentiation of real-valued functions 76
3.1 Directional and partial derivatives 76
3.2 Linear functions 79
ix
Contents
Chapter 4
Vector-valued functions of several variables 119
4.1 Linear transformations 119
4.2 Affine transformations 125
4.3 Differentiable transformations 128
4.4 Composition 134
4.5 The inverse function theorem 140
4.6 The implicit function theorem 147
4.7 Manifolds 153
4.8 The multiplier rule 161
Chapter 5
Integration 167
5.1 Intervals 168
5.2 Measure 170
5.3 Integrals over En 181
5.4 Integrals over bounded sets 186
5.5 Iterated integrals 190
5.6 Integrals of continuous functions 200
5.7 Change of measure under affine transformations 206
5.8 Transformation of integrals 209
5.9 Coordinate systems in En 216
5.10 Measurable sets and functions; further properties 222
5.11 Integrals: general definition, convergence theorems 227
5.12 Differentiation under the integral sign 237
5.13 LP-spaces 240
Chapter 6
Curves and line integrals 245
6.1 Derivatives 245
6.2 Curves in En 247
6.3 Differential I-forms 253
6.4 Line integrals 258
*6.5 Gradient method 265
*6.6 Integrating factors; thermal systems 268
Chapter 7
Exterior algebra and differential calculus 275
7.1 Covectors and differential forms of degree 2 276
7.2 Alternating multilinear functions 283
x
Contents
Chapter 8
Integration on manifolds 321
8.1 Regular transformations 322
8.2 Coordinate systems on manifolds 329
8.3 Measure and integration on manifolds 334
8.4 The divergence theorem 340
*8.5 Fluid flow 350
8.6 Orientations 353
8.7 Integrals of r-forms 356
8.8 Stokes's formula 362
8.9 Regular transformations on submanifolds 367
8.10 Closed and exact differential forms 369
8.11 Motion of a particle 375
8.12 Motion of several particles 380
Appendix I
Axioms for a vector space 383
Appendix 2
Mean value theorem; Taylor's theorem 385
Appendix 3
Review of Riemann integration 386
Appendix 4
Monotone functions 388
References 389
Answers to problems 391
Index 405
Xl
1
Euclidean spaces
This book is concerned with the differential and integral calculus of functions
of several variables. For this purpose one needs first to know some basic
properties of euclidean space of arbitrary finite dimension n. We begin
this chapter with a brief review of the real numbers and the elements of
vector algebra and geometry of such spaces. Later in the chapter the concepts
of neighborhood, open set, and closed set are introduced. These constitute
the basis for studying what are called topological properties of n-dimen-
sional space.
Format
The word "Theorem" has been reserved for what the author considers
the most important results. Results of lesser depth or interest are labeled
"Proposition." The symbol 0 indicates the end of the proof of a theorem or
proposition. Occasionally part of a proof is left to the reader as a homework
exercise. The sections marked with an asterisk (*) may be omitted without
disrupting the organization. References are given at the end of the book.
We presume that the reader is acquainted with the most elementary aspects
of set theory. The symbols
E,~, U, (\, - , c
stand, respectively, for is an element of, is not an element of, union, intersection,
difference, and inclusion. Sets ordinarily are denoted by capital italicized
letters. A set is described either by listing its elements or by some property
characterizing them. Thus {2, 5, 7} is the set whose elements are the three
numbers 2, 5, and 7. If S is a set and n a property pertaining to elements of S,
then {p E S: n} denotes the set of all PES with property n. For example,
if Z = {l, 2, ... } is the set of natural numbers, then S = {x E Z: x = 2y - 1
1 Euclidean spaces
Axiom I
(a) Any two real numbers have a sum x + y and a product xy, which are
also real numbers. Moreover,
Commutative law x +y= y + x, xy = yx,
Associative law x + (y + z) = (x + y) + z, x(yz) = (xy)z,
Distributive law x(y + z) = xy + xz
for every x, y, and z.
(b) There are two (distinct) real numbers 0 and 1, which are identity
elements under addition and multiplication, respectively:
x + = ° x, xl = x
for every x.
(c) Every real number x has an inverse - x with respect to addition, and
if x =1= 0, an inverse x - 1 with respect to multiplication:
x + (-x) = 0,
Axiom II. There is a relation < between real numbers such that:
(a) For every pair of numbers x and y, exactly one of the following
alternatives holds: x < y, x = y, y < x.
(b) w < x and x < y imply w < y (transitive law).
(c) x < y implies x + z < y + z for every z.
(d) x < y implies xz < yz whenever < z. °
From Axioms I and II follow all of the ordinary laws of arithmetic. In
algebra any set with two operations (usually called addition and multiplica-
tion) having the properties listed in Axiom I is called afield. A field is called
ordered if there is in it a relation < satisfying Axiom II.
2
1.1 The real number system
The real numbers form an ordered field. However, this is by no means the
only ordered field. For example, the rational numbers also form an ordered
field. We recall that x rational means that x = p/q, where p and q are integers
and q "# O. Yet another axiom is needed to characterize the real number
system. This axiom can be introduced in several ways. Perhaps the simplest
of these is Axiom III, stated below in terms of least upper bounds.
In Section 2.3 we state other axioms that turn out to be equivalent to
Axiom III. We should warn the reader that Axiom III is more subtle than
Axioms I and II, and that one becomes aware only gradually of its impli-
cations. However, this axiom is the foundation stone for some of the most
important theorems in calculus.
Let S be a nonempty set of real numbers. If there is a number c such that
x s c for every XES, then c is called an upper bound for S. If c is an upper
bound for Sand b ~ c, then b is also an upper bound for S.
Axiom III. Any set S of real numbers that has an upper bound has a least
upper bound.
The least upper bound for S is denoted by sup S. If S has no upper bound,
then we set sup S = + 00.
A number d is a lower bound for S if d s x for every XES. If S has a
lower bound, then (Problem 2) S has a greatest lower bound. It is denoted
by inf S. If S has no lower bound, then we set inf S = - 00.
EXAMPLE I. Let S = {I, 2, 3, ... }, the set of positive integers. Then sup S =
+ 00 and inf S = 1.
are called closed and open, respectively, and have a as greatest lower bound.
The corresponding intervals ( - 00, b], (- 00, b) have b as least upper bound.
Let S be a set that has an upper bound. Example 2 shows that the number
sup S need not belong to S. If sup S does happen to be an element of S, then
it is the largest element of S and we write "max S" instead of "sup S."
Similarly, if S is bounded below and inf S is an element of S, then we write for
it "min S."
3
I Euclidean spaces
The real number system also satisfies the archimedean property. This means
that for every e > 0, x > 0 there exists a positive integer m such that x < me.
To prove it, suppose to the contrary that for some pair e, x of positive
numbers, me ~ x for every m = 1, 2, .... Then x is an upper bound for the
set S = {e, 2e, 3e, .. .}. Let c = sup S. Then (m + l)e ~ c and therefore
me ~ c - e, for each m = 1, 2, .... Hence c - e is an upper bound for S
smaller than sup S, a contradiction. This proves the archimedean property.
We shall not prove that there actually is a system satisfying Axioms I,
II, and III. There are two well-known methods of constructing the real
number system, starting from the rational numbers. One of them is the
method of Dedekind cuts and the other is Cantor's method of Cauchy
sequences.
Axioms I, II, and III characterize the real numbers; in other words,
any two systems satisfying these three axioms are essentially the same. To
put this more precisely in algebraic language, any two ordered fields satisfying
Axiom III are isomorphic.
For proofs of these facts, refer to the book by Birkhoff and McLane
[2, Chapter III].
PROBLEMS
l. Find the least upper bound sup S and greatest lower bound inf S of each of the
following sets:
(a) {x: x 2 - 3x + 2 < O}.
[Note: In this book, "Show that ... " and "Prove that ... " both mean "give
a valid mathematical proof."J
4
1.2 Euclidean En
1.2 Euclidean En
In this book we denote the real number system by E I. Let us now define the
space E", whose elements are n-tuples of real numbers. The elements of En
will be called vectors.
5
I Euclidean spaces
i= 1
The vector space En with this inner product is called euclidean n-space. Other
inner products in En are considered in Section 2.11.
The euclidean norm (or length) of a vector x is
Ixl = (x' X)1/2.
It is positive, except when x = 0, and satisfies the following two important
inequalities. Fo~ every x, y E En,
(1.1 ) Ix' yl ~ Ixllyl (Cauchy's inequality),
(1.2) Ix + yl ~ Ixl + Iyl (triangle inequality).
6
1.2 Euclidean E"
(x + y) . (x + y) = x.x + 2x . Y + Y . y.
From Cauchy's inequality,
Ix + yl2 ~ Ixl2 + 21xllyl + lyl2,
or
(1.3)
for every choice of scalars e I, ... , em and of vectors Xl' ... , x"'. We recall
(Appendix A.1) that L j Cjx j is called a linear combination of Xl' ... , x m •
The euclidean distance between x and y is Ix - y I. If x, y, and z are vectors,
then
x - z = (x - y) + (y - z).
(1.4) Ix - zl ~ Ix - yl + Iy - zl,
which justifies the name" triangle inequality" (see Figure 1.1).
x ______\
y
o
Figure 1.1
7
1 Euclidean spaces
.....
y
.... ........ Ix - yl
........
.... ..... x
o
Figure 1.2
Orthonormal bases
En is an n-dimensional vector space, and any linearly independent set of
vectors {VI"'" vn} with n elements is a basis for it.
A basis {VI"'" vn } for En is called orthonormal if Vi . Vj = bij' where
8
1.2 Euclidean En
Taking the inner product of each side with Vi and using the formula Vi • Vj =
bij' we obtain
The coefficients c i (1.6) are just the components of x with respect to the
orthonormal basis vectors.
PROBLEMS
4. Show that 2:7= I IXil ~ J~ lxi, for any x = (Xl, ...• xn). [Hint: First suppose that
Xi;:: O. Use Equation (1.1) with yi = I.J
5. Show that 21xl2 + 21yl2 = Ix + yl2 + Ix - Y12. What does this say about
parallelograms (see Figure 1.3)?
x +y
o
Figure 1.3
9
1 Euclidean spaces
11. Let l ' be a vector subspace of En, of dimension k; and consider its orthogonal
complement
10
1.3 Elementary geometry of En
Figure 1.4
Hyperplanes, half-spaces
Given two points XI' X 2 E En, consider the set
c = z· e 1 = Zl, c = z· (e 1 + 2e 2 ) = Zl + 2Z2,
C = Z • (e 2 + 3e 3 ) = Z2 + 3z 3, C = Z • (e 3 + 4e 4 ) = z3 + 4Z4.
11
1 Euclidean spaces
From these equations, the components Zl, ... , Z4 of the vector z satisfy
~
:::::-::::: H = {x: z· x ~ c}
P={x:z·x=c}
Figure 1.5
Convex sets
This concept is defined as follows.
Definition. Let KeEn. Then K is a convex set if the line segment joining
any two points of K is contained in K (Figure 1.6).
12
1.3 Elementary geometry of E"
8 ......
XI
I
X
Convex
X
2
~2 XI X
Not convex
Figure 1.6
E" itself is a convex set. The empty set and sets with just one point trivially
satisfy the definition; hence they are convex. The reader should be able to
think of several kinds of geometric objects such as lines, planes, spherical
balls, regular solids, and so on, which appear to be convex sets. However,
geometric intuition is not always a reliable guide, especially in four or more
dimensions. In any case, intuition is no substitute for a proof that the set in
question is actually convex. To show that a set K is convex directly from the
definition, we must verify that for every Xl' X2 E K and t E [0,1], the point
X = tX l + (l - t)x 2 also belongs to K. In the definition, we assumed that
Xl i= x 2 . But if XI = x 2 , it is trivial that X E K. since X = XI = x 2 .
EXAMPLE 3. Let V be an open spherical n-ball, namely, V = {x: Ix - xol < «:5},
for some Xo and «:5 > O. To show that V is a convex set, we proceed as in
Example 2. Let XI' x 2 E V and X = tX I + (1 - t)x 2 , where t E [0,1]. Then
Hence X E U.
13
I Euclidean spaces
PROBLEMS
1. Let n = 3. Find the plane that contains the three points e l , e2, and e 3 - 3e l .
Sketch its intersection with the first octant in E3.
2. (a) Find the hyperplane in E4 containing the four points 0, e l + e 2, e l - e 2 + 2e 3 ,
3e4 - e 2 ·
(b) Find the value of t for which t(e l - e 2) + (l - t)e4 is in this hyperplane.
3. Let i denote the line in E4 through e l - e 3 and -e l + e 2 + 2e4 . Find the hyper-
plane P through e l - e 3 to which i is perpendicular.
4. Let "Y = {(x, y, z): 2x + 3y - z = a}. Show that "Y is a 2-dimensional vector
subspace of E 3 , and find a basis for 'Or. ("Y is a vector subspace of En if x, y E "Y imply
x + Y E "Yand ex E "Y for any scalar c.)
5. Let"Y = {x:z'x = O}, where z i= Oisgiven. Show that "Yisan(n - I)-dimensional
vector subspace of En, and find a basis for "Y.
6. Show that {x: Ix - xII = Ix - x 2 1}, where Xl and X2 are given points in En, is a
hyperplane.
7. Show that {x: Ix - xII = e Ix - x 21}, where Xl and X2 are given points in En and
o < e < I, is an (n - I)-sphere.
8. Let x o , Xl' ... , Xn- I be such that XI - Xo, ... , Xn- I - Xo are linearly independent.
Prove that there is exactly one hyperplane containing Xo , XI, ... , Xn - I.
9. A set P = {x: Zi' X = ei for i = I, ... , n - k}, where ZI, ... ' Zn-k are linearly
independent vectors, is called a k-piane in En. Let Xo , Xl> ... , Xk be such that
XI - Xo , ... , Xk - Xo are linearly independent. Prove that there is exactly one
k-plane containing x o , Xl' .. ·' X k·
10. Prove that any line in En is a convex set.
11. Show that K is a convex set by directly applying the definition. Sketch K in the
cases n = 1,2,3.
(a) K = {X:IXII + ... + Ix nl:5: I}.
(b) K = {x = eiv i + ... + env n, 0:5: ei :5: I for i = I, ... , n}, where {VI' ... , vn}
is a basis for En. This is the n-paralleiepiped spanned by VI' ... ' Vn with 0 as a
vertex.
12. Let P be a hyperplane. Prove that the line through any two points of P is contained
in P. Why does this imply that P is a convex set?
14
1.4 Basic topological notions in En
ofx contains at least one point of A and at least one point of A then x is a
C,
frontier point of A.
An interior point of A necessarily is a point of A, and an exterior point
must be a point of A However, a frontier point may belong either to
C•
A or to AC.
Figure 1.7
15
1 Euclidean spaces
It is always true that int A c A. If these two sets are the same, then A is
called an open set.
Proposition 1.1. If A and B are open sets, then A u B and A n B are open.
PROOF. To show that A u B is open, we show that every point of A u B is
interior to A u B. Let x E A u B. We must find a neighborhood U of x
such that U c A u B. By the definition of the union of two sets, either
x E A or x E B.lfx E A, then there is a neighborhood U ofx such that U c A.
Since A c A u B, U c A u B. Similarly, if x E B there is a neighborhood of
x contained in A u B. This proves that A u B is open.
If x E A n B, then x has neighborhoods U 1, U 2 such that U 1 C A and
U 2 c B. Let U 3 = U I n U 2 . Then U 3 is a neighborhood of x and U 3 c
A n B. Therefore A n B is open. D
16
1.4 Basic topological notions in E"
n
I'E1
These sets are, respectively, the union and intersection of the indexed collec-
tion. Iff is a finite set, then the indexed collection is called finite. For instance,
if j = {I, 2, ... , m}, we write the collection as AI' A 2, ... , Am. If j =
{I, 2, ... }, then the indexed collection is an infinite sequence of sets and is
written AI' A 2, ... , or [Am], m = 1,2, ... In that case the union is written
A I U A2 U ... or U;;;~ I Am' with similar notations for the intersection.
Proposition 1.2. The union of any indexed collection of open sets is open.
The intersection of any finite indexed collection of open sets is open.
EXAMPLE 7. Let A = {x: a :S x < b}. The points a and b are frontier points
of A, with a E A, b ¢ A. The set A is neither open nor closed. In the notation
of Section 1.1, A = [a, b). Its interior and closure are respectively the open
interval (a, b) and the closed interval [a, b].
We recall that
(A n BY = A e U Be, (A u BY = A e n Be.
17
1 Euclidean spaces
If each set of the collection is indexed by itself (taking.f = 21, AA = A), then
this definition of union and intersection agrees with the one for indexed collec-
tions. Propositions 1.2 and 1.3 remain valid for unindexed collections.
PROBLEMS
18
1.5 Convex sets
6. Show that:
(a) fr A = fr(A C). (b) cI A = cI(cI A).
(c) fr A = cI A n cI(A C ). (d) int A = (cI(A C ))'.
7. Show by giving examples that the following are in general false:
(a) int(cI A) = int A. (b) fr(fr A) = fr A.
8. Let A be open and B closed. Show that A - B is open, and that B - A is closed.
9. Show that:
(a) int(A n B) = (int A) n (int B).
(b) cI(A u B) = (cI A) u (cI B). [Hint: Part (a).]
10. Show that:
(a) int(A u B) =:> (int A) u (int B).
{b) cI(A n B) c (cI A) n (cI B).
Give examples in which = does not hold.
Proposition 1.4. If K b ... ,Km are convex sets, then their intersection
Kl n··· n Km is convex.
PROOF. Let Xb x 2 be any two points of Kl n .. , n K m, Xl i= x 2 . Let I
denote the line segment joining Xl and X 2 . For eachj = 1, ... , m, Xl. X2 E K j'
Since K j is convex I c K j for each j = 1, ...• m. Thus I c KIn··· n Km.
D
A convex polytope is the set of all points X that satisfy a given finite system
of linear inequalities of the form zj· X 2': cj , j = 1, ... , m. The theory of
linear programming is concerned with the problem of maximizing or
minimizing a linear function subject to such a system of linear inequalities.
It has various interesting economic and engineering applications [to, 13].
In Section 3.6 it is shown that the maximum and minimum values of a linear
function must occur at "extreme points" of K, at least if K is compact.
19
1 Euclidean spaces
In the proof of Proposition 1.4 we did not really use the fact that the num-
ber of sets K j is finite. Therefore we have:
/
/
/ / Supporting line
Figure 1.8
20
1.5 Convex sets
Given y E fr B, let
Hy = {x:y.x~ 1},
Py = {x: y. x = I},
so that P y is the hyperplane bounding H y' By Cauchy's inequality and the
fact that Iyl = I,
y·x ~ Iyllxl = Ixl.
Equality holds if and only if x is a positive scalar multiple of y. Hence B c Hy
and B n P y consists only of y. The supporting hyperplane to B at y is P y
(Figure 1.9).
Figure 1.9
Again let K be any nonempty, closed convex set that is a proper subset
of En(K #- En). Let Jf K denote the collection of all closed half-spaces H such
that K cHand the hyperplane P bounding H is supporting for K. For
instance, the collection Jf B in the above example consists of the various half-
spaces Hy for all possible choices of y E fr B.
The notation
Theorem 1.1. K = n
HE.KK
H.
21
1 Euclidean spaces
Figure 1.10
22
1.5 Convex sets
Figure 1.11
23
1 Euclidean spaces
Figure 1.12
X = L tjx j
j=O
of the vertices X o , XI' ... ,xr (problem 4). The numbers to, tl, ... ,t' are cll-lled
the barycentric coordinates of x. The (r - 1)-dimensional face opposite the
vertex Xi is the set of points of the r-simplex with t i = 0.
For example, the vertices Xo , XI' Xi of a triangle have barycentric coordi-
nates (1,0,0), (0, 1,0), (0,0, 1), respectively. The midpoint of the face dpposite
X o has barycentric coordinates (0, 1, 1). The interior points of the triangle
have barycentric coordinates (to, tl, t 2 ), all of which are strictly positive. In
each case to + t l + t 2 = 1.
The simplex with vertices 0, e I, ... , en is called the standard n-simplex.
It will be denoted by L, and is of use in Section 5.7 in connection with integra-
tion. The barycentric coordinates (to, tl, ... , t n) of a point X E L are given
by t i = Xi for i = 1, ... , n, to = 1 - (x I + ... + xn).
24
1.5 Convex sets
Let
sj = tj - r:x.c j for j = 1, ... , m,
where r:x. is a positive number chosen so that sj 2: 0 for each j = 1, ... , m
and Sk = 0 for some k. Explicitly,
Then
25
1 Euclidean spaces
Figure 1.13
PROBLEMS
1. Show that each of the following subsets of £2 is closed and convex by writing it as
the intersection of closed half-planes:
(a) The regular hexagon with center (0, 0) and e 1 as one vertex.
(b) {(x,y):y ~ lxi, -I s x s I}.
(c) {(x, y): y s log x, x > O}.
(d) {(x, y): 0 s y s sin x, 0 s x S rr}.
2. Write the standard n-simplex as the intersection of n +1 closed half-spaces.
Illustrate for n = 2 and n = 3.
26
1.5 Convex sets
5. Prove that a supporting hyperplane for a closed convex set K can contain no
interior point of K.
6. Let K be any convex set. Prove that its interior and its closure are also convex sets.
7. The barycenter of an r-simplex is the point at which the barycentric coordinates
are equal, to = t l = ... = t'.
(a) Show that the barycenter of a triangle is at the intersection of the medians.
(b) State and prove a corresponding result for r 2 3.
8. Let X be a convex combination of XI' ... , Xm and let Xj be a convex combination
of Yjl,"" Yjmj,j = 1, ... , m. Show that X is a convex combination of ZI,"" zP'
which are the distinct elements of the set {Yik: k = I, ... , mj,j = I, ... , m}.
9. Let S be any subset of En. The set S of all convex combinations of points of S is the
convex hull of S.
(a) Using Problem 8, show that S is convex.
(b) Using Proposition 1.6, show that if K is convex and S c K, then S c K. Thus
the convex hull is the smallest convex set containing S.
10. Given Xo and b > 0, let C = {x: Ix i - x~1 ::::; 0, i = I, ... , n}, an n-cube with
center Xo and side length 20. The vertices of C are those X with IXi - Xo I = 0 for
i = I, ... , n. Show that C is the convex hull of its set of vertices. [Hint: Use induction
on n.]
11. Let K be a closed subset of En such that both K and its complement En - K are
nonempty convex sets. Prove that K is a half-space.
12. Let A and B be convex subsets of En. The join of A and B is the set of all X such that
X lies on a line segment with one endpoint in A and the other in B. Show that the
join of A and B is a convex set.
27
2
Elementary topology of En
2.1 Functions
One should think of a function f as assigning to each element p of some set
San elementf(p) of another set T. The elementf(p) is called the value off
at p. However, this is not a satisfactory definition of" function " because ofthe
ambiguity of the word "assigns." We give a more careful definition, in terms of
cartesian product sets.
28
2.1 Functions
29
2 Elementary topology of En
for every pES. If <P is a constant function, <p(p) = c for every pES, then we
write cf instead of <pf.
Restriction of a function
Often one is interested only in the values of a function f for elements of
some subset A of its domain. The restriction off to A is the function with
domain A and the same values asfthere. It is denoted by f I A. Thus
flA = {(p,f(p)):PEA}.
For instance, if a real-valued function f is integrated over an interval IcE 1,
then it is only f II which is important. The values off outside I do not affect
the integral.
Images, inverse images
Let fbe a function from a set S into a set T. The image under f of a set A c S
is the setf(A) = {f(p): pEA}. It is a subset of T, and in fact the restriction
f IA is a function from A onto f(A). The inverse image of a set BeT is the
setf -1(B) = {p: f(p) E B}. It is a subset of S.
EXAMPLE 1. Let f(x) = Xl. Then f([ - 2,2]) = [0,4], f(El) = [0,00),
f - 1([ 1, 3]) = [ -)3, - 1] u [1, ~/3]. The function f is not univalent since
f( -x) = f(x).
for any A c S. Consider any pEA. Then f(P) E f(A) by definition of f(A).
Take B = f(A) in the definition of inverse image set above. Then
p E f -1(f(A)). Since this is true for each pEA, we get (*).
PROBLEMS
1. (a) Let f(x) = cos x. Find fW ),f([ -71:/4,71:/2]), r 1([0, I]).
(b) Let g = f I[0,71:]. Find gIrO, 71:]), g-I([O, I]). Is g univalent?
2. The equations s = (Xl + y2)1/2, t = x - y define a transformation f from E2 into E2,
such that fIx, y) = (s, t). Let A = {(x, y): x 2 + l :s; a 2 }, where a > 0 is given.
(a) Find f(A).
(b) Find f- I(A).
30
2.2 Limits and continuity of transformations
4. Let f be a function from S into T. Show that, for any subsets A and B of S:
(a) f(A v B) = f(A) v f(B).
(b) f(A n B) c f(A) n f(B).
(c) Iff is univalent, then f(A n B) = f(A) n f(B).
5. Let f be a function from S into T. Show that, for any subsets C and D of T:
(a) r l(C v D) = r l(C) v r l(D).
(b) r l(C n D) = r l(C) n r l(D).
(c) r l(DC) = [f-1(D)J.
6. Let Sand T be sets, and let n(p, q) = p for all pES, q E T. The function n projects
S x Tonto S. Let ReS x T be a relation. Show that R is a function if and only if
n IR is univalent and onto S.
7. Let 1 :::;; s :::;; n - 1. Let us regard En as the cartesian product ES x En- s, and write
x = (x', XU), where x' = (Xl, ... , XS ), XU = (X S + 1, ... , xn). Let n(x) = x' be the
projection of En onto P. Show that n(A) is an open subset of ES if A is an open subset
of En.
8. Let AcEs, B c En- s, and regard the cartesian product A x B as a subset of En,
as in Problem 7.
(a) Show that A x B is open if both A and B are open.
(b) Show that A x B is closed if both A and B are closed.
31
2 Elementary topology of En
V ~f(U)
D
....-
f
~
Figure 2.1
If we let 8 and c5 denote the radii of V and U respectively, then the defi-
nition may be rephrased as follows: f(x) ...... Yo as x ...... Xo if for every 8 > 0
there exists c5 > 0 such that I f(x) - Yo I < 8 whenever 0 < Ix - Xo I < c5.
The number c5 depends, of course, on 8 and may also depend on Xo. Given 8
and x o , there is a largest possible c5. However, there is ordinarily no reason
to try to calculate it.
Let us first show that limits behave properly with respect to sums and
products. Let f and g have the same domain and values in the same euclidean
Em.
32
2.2 Limits and continuity of transformations
Limits of components
Since f has values in Em,
f(x) = (f I(X), ... , fm(x))
where.t is a real valued function called the ith component of f(with respect
to the standard basis of Em). For instance, in Problem 2, Section 2.1,f I(X, y) =
(x 2 + y2)1/2 and f2(X, y) = X - y. The next proposition shows that con-
vergence of f(x) to Yo is equivalent to convergence of each component
fi(X) to the corresponding component of yo.
Proposition 2.2. Yo = Iimx~xo f(x) if and only if y~ = limx~xo fi(X) for each
i = 1, ... ,m.
33
2 Elementary topology of E"
°
PROOF. Let V be any neighborhood of Yo. There exists (j >
f(x) E V whenever < Ix - xol < (j. IfO < It I < (jjIvl, then
° such that
EXAMPLE I. Let j(x, y) = X2/(X 2 + y2), (x, y) oF (0,0), and let Xo = (0,0).
° ° °
Taking v = e 1 = (1, O),j(r, 0) = 1 for every t oF 0. Hencej(t, 0) -+ 1 as t -+ 0.
Similarly, taking v = e 2 = (0, 1),J(0, t) = for every t oF andj(O, t) -+ as
t -+ 0. Since these limits are different,fhas no limit at (0,0).
As we show below, if Xo = (x o, Yo) oF (0,0), then j(x, y) -+ j(xo, Yo) as
(x, y) -+ (xo, Yo). Thus (0, 0) is the only point where there is no limit.
EXAMPLE 2. Let j(x, y) = (y2 - X)2/(y4 + x 2), (x, y) oF (0,0), and again let
Xo = (0,0). Consider any v = (h, k) oF (0,0). Then
(tk 2 - hf
f(th, tk) = t2k4 + h2
Continuity
Let us now suppose that Xo is an interior point of the domain D of the trans-
formation f.
EXAMPLE 3. Let f(x, y) = (xy)I/3. Show directly from the definition that j is
° °
continuous at (0, 0). Denote the radii of the neighborhoods U and V by (j and
e, respectively. We must show that given e > there exists (j > such that
Ij(x, y) - j(O, 0) I < e whenever x 2 + y2 < (j2. Note that j(O, 0) = 0.
34
2.2 Limits and continuity of transformations
Proposition 2.4. Let f and g be continuous at Xo. Then f + g and fg are con-
tinuous at Xo. If g(xo) '# 0, then fg - 1 is continuous at Xo'
°
EXAMPLE 7. A rational function f(x) = P(x)/Q(x), where P and Q are poly-
nomials, is continuous at each point where Q(x) '# by Proposition 2.4. For
instance, in Examples 1 and 2,fis continuous at each (x, y) '# (0,0).
35
2 Elementary topology of En
Limits at 00
Let us call a set of the form {x : Ix I > b} a punctured neighborhood of 00. The
definition of" limit at 00 "then reads: Yo = limlxl ~ x f(x) iffor every neighbor-
hood V of Yo there exists a punctured neighborhood U of 00 such that
f(U) c V.
Whenfis real valued we say that limx~xo f(x) = + 00 if for every C > 0
there is a punctured neighborhood U of Xo such that f(x) > C whenever
x E U. The definition of "limx~xo f(x) = - 00" is similar.
PROBLEMS
(a) lim
X4 + y4
-2--2' (b) lim
xi
-2--4'
(x.y)-(o.O) X +y (x.y)-(o.O) X +y
. (x'x t )(x'x 2 ) .
(c) hm , where XI and X2 are given vectors not O.
X'X
7. (a) Letg(x) = If(x)la where a > O. Suppose that f is continuous at Xo and f(xo) = O.
Show that g is continuous at Xo.
(b) Use (a) and Problem 3 to give another proof that the function in Example 3 is
continuous at (0, 0).
36
2.3 Sequences in En
2.3 Sequences in En
An infinite sequence is a function whose domain is the set of positive integers.
For brevity, we use the term "sequence" to mean infinite sequence. In this
section let us consider sequences with values in En. It is customary to denote
by Xm the value of the function at the integer m = 1, 2, ... , and to call Xm the
mth term of the sequence. The sequence itself is denoted by x I' x 2 , ... , or for
brevity by [xml It must not be confused with the set {XI' X2,"'} whose ele-
ments are the terms of the sequence. This set may be finite or infinite. For
t
instance if Xm = (- 1 then the sequence is - 1, 1, - 1, ... , and the set
{XI' X2""} has only two elements -1 and 1.
Definition. Suppose that for every E > 0 there exists a positive integer N
such that IXm - Xo I < E for every m ;?: N. Then Xo is the limit of the
sequence [xm].
37
2 Elementary topology of E"
We leave the proof of Propositions 2.6 and 2.7 to the reader (Problem 6).
The proofs of the corresponding Propositions 2.1 and 2.2 in Section 2.2 can be
adapted for this purpose.
In this section we prove three important results (Theorems 2.1-2.3)
that depend on the least upper bound property of the real numbers
(Axiom III, Section 1.1).
The first of these results is about convergence of monotone sequences of
real numbers.
A sequence [xm] of real numbers is called monotone if either x I :$; X2 :$;
X3 :$; ... or XI 2': X2 2': X3 2': .... In the first instance the sequence is non-
decreasing, in the second non increasing. A sequence [xm] is bounded if there is
a number C such that IXm I :$; C for every m = 1,2, ....
EXAMPLE I. Let 0 < a < I. The sequence a, a 2 , a 3 , ..• of its powers is de-
creasing, and is bounded below by O. The limit of the sequence is b =
inf{a,a 2 ,a 3 , ••• }. Since b:$; a m+ l , a-1b:$; am for m = 1,2, .... Therefore
a-1b is a lower bound for {a, a 2 , a 3 , .• • }. Since b is the greatest lower bound,
a-1b'S:; b. However, a-1b 2': b since 0 < a < I; and hence a-1b = b. This
implies that b = O.
1 1 1
Ym :$; 1 + 1 + - + - + ... + ~- < 3.
2 4 2m - I
38
2.3 Sequences in En
Xm = 1+ 1+ ~2! (1 - 2.)
m
+~
3!
(1 - 2.)(1
m
- ~)
m
°
Definition. If for every G > there exists a positive integer N such that
Ix, - xml < G for every I, m ~ N, then [x",] is a Cauchy sequence.
EXAMPLE 4. Let Xm = L~= I(COS k)/k 2 . Let us show that [xm] is a Cauchy
sequence. If 1 ~ m, then
I cos k
X l - Xm = L -k 2
k=m+1
IXI - xml S fm
1dt
2"
t
= m -1 - 1- I < m -1 .
Fortunately, it turns out that the concepts of Cauchy sequence and con-
vergent sequence are equivalent, for sequences in En.
39
2 Elementary topology of En
e e
IXm - Yol ~ IXm - xNI + IXN - Yol < 2" + 2: = e.
40
2.3 Sequences in En
For the next theorem we need the concept of diameter of a set. A nonempty
set A is called bounded if there exists C > 0 such that Ix I ~ C for all x E A.
The diameter of a bounded set A is
diamA = sup{lx - YI:x,YEA}.
Theorem 2.3 (Cantor). Let [AmJ be a sequence of closed sets such that
Al ::::> A2 ::::> ••• and 0 = Iim m_ oc diam Am· Then Al n A2 n··· contains
a single point.
PROOF. For each m = 1,2, ... let Xm be a point of Am. Let us show that
[xmJ is a Cauchy sequence. Given e > 0, there exists N such that diam AN < e.
If I, m 2:: N, then x,, Xm E AN since A, c AN' Am C AN' Therefore
Ix, - xml ~ diam AN < e.
By Theorem 2.2 the sequence [xmJ has a limit Xo. For each I = 1,2, ... ,
xm E A, for every m 2:: I smce Am C A" Since A, is closed, Xo E A, (Problem 5).
Since this is true for each I, Xo E Al n A2 n··· .
IfXEAI n A2 n ... , then xEAm and
o ~ Ix - xol ~ diamA m, m = 1,2, ...
Since diam Am -+ 0 as m -+ 00, Ix - xol = 0 and x = xo. D
Axiom 1111
(a) EI has the archimedean property.
(b) Every Cauchy sequence in EI has a limit.
Then Axioms III and III I are equivalent in the presence of Axioms I and II,
Section 1.1. This means that if either III or III I is taken as an axiom, then the
other can be proved as a theorem. We have shown that III implies 111 1 , To
show that III I implies III, we first note that if III I is taken as an axiom, then
Theorem 2.3 remains valid (with the same proof). It then suffices to deduce
the least upper bound property from the archimedean property and Theorem
2.3. To prove this, let S be any subset of EI that is bounded above, and
let c be some upper bound for S. Let us define a sequence of closed intervals
41
2 Elementary topology of En
Figure 2.2
11,1 2 "" as follows: Let a be some point of S, and II = [a, c]. Divide
II at the midpoint (a + c)/2 into two congruent closed intervals. If (a + c)/2
is an upper bound for S, let 12 be the left-hand interval, otherwise let 12
be the right-hand interval. In general, suppose m ~ 1 and 1m has been
defined. If the midpoint of 1m is an upper bound for S, let Im+ 1 be the left half
of 1m , otherwise the right half. The archimedean property implies that for
any x ~ 0, the sequence [m- 1x] tends to 0 as m -+ 00. Since 0 ~ 2- m ~ m- 1,
the sequence [2 - mx] also tends to O. Let x = 2(c - a). Now I I :::l 12 :::l .•.
and the length of 1m is 2 - mx. By Theorem 2.3, lin 12 n ... contains a single
point Xo. By the construction, Xo = sup S (see Figure 2.2).
Infinite series
Formally, an infinite series is an expression written Ik=
I X k or XI + X 2 + ....
To be more precise, with any sequence [Xk] is associated another sequence
[sm], where Sm = XI + ... + Xm is called the mth partial sum. This pair of
sequences defines an infinite series. If the sequence of partial sums has a
limit s, then the series is convergent and s is its sum. This is denoted by s =
Xl + X2 + .... If the sequence of partial sums has no limit, then the series
is divergent.
If s = XI + X 2 + "', t = YI + Y2 + "', then s + t = (XI + yd +
(x 2 + Y2) + ... and cs = (ex l ) + (cx 2 ) + ... for any scalar c. This follows
from the definition and Proposition 2.6. Some further elementary properties
are given in Problems 7(c) and 8.
PROBLEMS
In Problems 1 and 2 you may use the results of Problems 9 and 10.
t. Find the limit if it exists.
(a) Xm = (2m - 2- m)!(3m + rm).
(b) Xm = sin(mn/2).
(e) Xm = sin mn.
(d) Xm = ((m + l)/(m - l)t. [Hint: (1 + l/m)m -> e as m -> w.]
(e) Xm = ((m 2 + l)/(m 2 - l))m.
42
2.4 Bolzano- Weierstrass theorem
3. Show that a sequence [xm] has at most one limit Xo' [Hint: If Yo were another
limit, let e = IXo - Yo 1/2.]
4. (a) Let A be a closed set. Show that ifx m E A for m = I, 2•... and Xo = Iimm_~ Xm,
then Xo E A.
(b) If A is not closed, show that there exists a sequence [xm]. with Xm E A for
m = 1,2, ... , converging to a limit Xo ¢= A.
5. Show that if Xm E A for every m ;:::: I and Xo = Iim m_ ~ Xm' then Xo E cI A.
6. (a) Prove Proposition 2.6.
(b) Prove Proposition 2.7.
7. (Comparison tests.) Show that:
(a) If 0 ~ Xm ~ Ym for every m ;:::: I and Ym -> 0 as m -> C1J, then Xm -> 0 as m -> 'CIJ.
(b) If [xm], [ym] are nondecreasing sequences such that Xm ~ Ym for each m = 1,2, ...
and Ym -> Y as m -> C1J, then [xm] has a limit x ~ y.
(c) If 0 ~ Xm ~ Ym for every m = 1,2, ... and t = YI + Y2 + ... , then the series
XI + Xl + ... converges with sum s ~ t.
8. An infinite series XI + Xl + ... converges absolutely if the series of nonnegative
numbers IxII + Ixli + ... converges. Prove that any absolutely convergent infinite
series is convergent. [Hint: Show that the sequence [sm] of partial sums is Cauchy.]
9. Show that if a > 0, then
(a) Iim m_ oc a 1 / m = 1. (b) lim m_ x amlm! = O.
(c) Iim m_ oc (x m)l/m = 1 provided Iim m_ x Xm = a.
[Hints: For part (a) reduce to the case 0 < a < 1. By Example 1, if b < 1 then
a ~ bmfor only finitely many m. For part (b), compare with the sequence [elm] for
suitable e and suitable I in Problem 7(a).]
10. Let Xo = Iim m_ x X m, Yo = limm_~ Ym, and assume that Ym of. 0 for m = 0, 1,2, ....
Show that xolYo = Iim m_ r xmlYm. [Hint: By (c) of Proposition 2.6 it suffices to
show that yO' I = Iim m_ ~ y;;; I.]
11. Show that Xo = Yo in Example 3.
43
2 Elementary topology of En
I = { x: IXl. - x~ I
. a
s "2' i =
}
1, ... , n •
where
X = (Xl, ... ,Xn ),
is called an n-cube with center Xo and side length a. If A is any bounded set,
then A c I for some n-cube I.
44
2.4 Bolzano-Weierstrass theorem
f-
rr 112
1 13 114
Figure 2.3
Corollary l. Let See be closed and bounded. Then every infinite set A c S
has at least one accumulation point Xo E S.
PROOF. Since S is a bounded set and A c S, A is a bounded set. By the
Bolzano-Weierstrass theorem, A has an accumulation point Xo. By Proposi-
tion 2.8, Xo E cl A. Since S is closed, cl A c S. Thus Xo E S. 0
45
2 Elementary topology of En
Let us consider the case when A is an infinite set. Then A c A I since each
Am c A l' Since A 1 is bounded, A is bounded. Let Xo be an accumulation
point of A. As in the proof of Corollary 1, we have Xo E A 1 since A I is a closed
set and A c A I' For each m = 1,2, ... , Xo is also an accumulation point of
the set {xm' Xm + I, ... }. Since this set is contained in Am and Am is closed, we
get in the same way Xo E Am. Since this is true for each m, Xo E n~= I Am. D
PROBLEMS
46
2.5 Relative neighborhoods, continuous transformations
A,
0
,..., A2
i --,
AI
I I n I
A3 A3 A3 A3
Figure 2.4
6. (Subsequences.) Let [xmJ be a sequence, and Yl = x"" for / = 1,2, ... , where
ml < mz < .... Then [ytJ is called a subsequence of [xm].
(a) Show that any bounded sequence in En has a convergent subsequence.
(b) A set S is called sequentially compact if: any bounded sequence [xmJ, with Xm E S
for m = 1,2, ... ,has a subsequence [ytJ such that Yl --+ Yo as /--+ 00, Yo E S. Show
that a nonempty set SeEn is sequentially compact if and only if S is closed and
bounded.
7. Let y be any frontier point of a closed convex set K. Show that K has a supporting
hyperplane P that contains y. [Hint: Let {Ym} be a sequence of points exterior to K
such that Ym tends to Y as m --+ 00. Let xm be a point of K nearest to Ym and
Then IUmI = 1 and Xm tends to y as m --+ ex;. By the proof of Theorem 1.1 there is a
supporting hyperplane of the form {x: Urn' (x - xm) = O}. Let U be an accumulation
point of the bounded set {Ub u z , ... } and P = {x: U' (x - y) = O}.J
47
2 Elementary topology of En
U = Sn W S
Figure 2.5
Theorem 2.4. Let f be continuous on a closed, bounded set S. Then f(S) is also a
closed, bounded set.
48
2.5 Relative neighborhoods, continuous transformations
Next suppose that f(S) is not closed. Then there exists Yo E cl[f(S)] - f(S).
For m = 1,2, ... there exists Ym E f(S) such that Ym -+ Yo as m -+ 00. Choose
zmES with Ym = f(zm). As before, the set B = {ZI,Z2''''} has an accumu-
lation point Zo E S. Since Yo ¢ f(S), Yo =1= f(zo). Let V be a neighborhood of
f(zo) of radius tl f(zo) - Yo I. Since f is continuous at Zo, there exists a relative
neighborhood U of Zo with f(U) c V. In particular, Ym = f(zm) is in V for
infinitely many m for which Zm E U. This contradicts the fact that Ym -+ Yo as
m ~ 00. Thus f (S) is closed. D
PROBLEMS
3. Let S = [aJbl
(a) Show that the relative neighborhoods of a are the half open intervals [a, c) with
a < c < b, and S.
(b) Leta < Xo < ~a + b). Show that the relative neighborhoodsofx o areas follows:
(xo - 15, Xo + b) if 0 < 15 < Xo - a; [a, Xo + b) if Xo - a :-:; c5 < b - Xo; S.
(c) Describe the relative neighborhoods in the remaining cases ~a + b) :-:; Xo :-:; b.
4. Define the projection 11: from En onto P as in Problem 7, Section 2.1.
(a) Show that 11:(A) is closed and bounded if A is closed and bounded.
(b) Give an example ofa closed set A such that n(A) is not closed.
49
2 Elementary topology of En
5. (a) Let f be continuous on S, and let SIC S. Show that the restriction f lSI is
continuous on S l'
(b) Let f(x) = I-x if x~O and f(x)=O if x <0. Let SI = [0,00), S= El.
Show that f lSI is continuous, but f is not continuous at each point of S l'
6. Let f be a transformation with domain S. Show that f is continuous at Xo if and only
if f(x o) = limm~oc f(x m) for every sequence [xm] such that Xm E S for m = 1,2, ...
and Xm ..... Xo as m ..... 00.
7. Let f be continuous on En. Suppose, moreover, that f(x) > 0 for all x # 0, and that
f(cx) = cf(x) for any x and c > O. Show that there exist a > 0 and b > 0 such that
alxl ::; f(x) ::; blxl. [Hint: First consider {x: Ixl = I}.]
8. (Uniform continuity.) A transformation f is uniformly continuous on SeEn if given
[; > 0 thert! exists b > 0 (depending only on [;) such that If(x) - f(y) I < [; for every
x, YES with Ix - y I < b. Show that if S is closed and bounded then every f con-
tinuous on S is uniformly continuous on S. [Hint: If not, then there exists [; > 0
and for m = 1,2, ... , x m, Ym E S such that If(xm) - f(Ym) I ~ [; and IXm - Yml ::; 11m.
Let Xo be an accumulation point of {Xlo Xl," .}. Show that the continuity of fat
Xo is contradicted.]
Definition. Let S be a nonempty set. For every pES let OIJ p be a collection of
subsets of S called neighborhoods of p such that:
(1) Every point p has at least one neighborhood.
(2) Every neighborhood of p contains p.
(3) If Uland U 2 are neighborhoods of p, then there is a neighborhood
U 3 of p such that U 3 CUI n U 2 .
(4) If U is a neighborhood of p and q E U, then there is a neighborhood
V of q such that V c U.
50
2.6 Topological spaces
EXAMPLE I. Let S = En, and as in Section 1.4, let JIf x be the collection of all
open spherical n-balls with center x. Clearly, Axioms (1) and (2) of the
definition above are satisfied, and in (3) we may take U 3 = U I n U 2.
Axiom (4) is verified in Example 1, Section 1.4. Thus En is a topological space.
Axiom (4) (in the preceding definition) guarantees that any neighborhood
is an open set. Propositions 1.1-1.3 in Section 1.4 about unions and inter-
sections of open sets (or closed sets) remain true in any topological space.
The proofs are almost the same as before.
It often happens that two different collections of neighborhoods JIf P' JIf~
lead to the same collection of open subsets of S. In En we need not have started
with spherical neighborhoods. For instance, the neighborhoods obtained
from any noneuclidean norm on En lead to the same open sets as in Section
1.4. (cf. Section 2.11).
The open sets, and not the particular kinds of neighborhoods from which
they were obtained, determine all of the topological properties of S. Thus we
say that the collections 0If p' JIf~ define the same topology on S if they lead to the
same collection of open sets.
51
2 Elementary topology of En
Continuous functions
The concept of continuous function from one topological space into another
is a natural extension of that already considered for transformations between
euclidean spaces.
If SeEn, T = Em. and S is given the relative topology, then this definition
of continuity agrees with that in Section 2.5.
Proposition 2.4'. Let f and 9 be real valued functions, continuous at Po. Then
f + 9 and fg are continuous at Po· If g(po) =t- 0, then fg - 1 is continuous at Po .
We leave the proof of Propositions 2.4' and 2.5' to the reader (Problems 11
and 12).
Let us next specialize to real valued functions f and special choices for B.
52
2.6 Topological spaces
By applying Corollary 2 to -fand -c, the set {p :f{p) < c} is open and
{p :f{p) s c} is closed. Corollary 2 often provides a convenient way to show
that a given set is open or closed.
Composites
Let f be a function from S into T, and g from R into S. The composite fo g
is defined by
(fa g)(r) = f[g{r)] for every r E R.
Figure 2.6
53
2 Elementary topology of En
Subspaces
Let So be a topological space and S a nonempty subset of So. By disregarding
the complement So - S, S becomes a topological space in the following
way. If PES, then a neighborhood of p relative to S is a set U = S n W, where
W is a neighborhood in So of p. Axioms (I) through (4) (in the first definition
in this section) are satisfied. For instance, to prove (3) let U 10 U 2 be relative
neighborhoods of p. Then U I = S n WI' U 2 = S n W2 , where WI and W2
are neighborhoods in So of p. Since So is a topological space, there is a neigh-
borhood W3 in So of p with W3 c WI n W2 • Then U 3 = S n W3 is a relative
neighborhood of p and U 3 CUI n U 2. This topology is called the relative
topology induced on S by the topology of So, and S is a topological subspace
of So.
In particular, if So = En, this is the relative topology on S mentioned in
Section 2.5 and Example 2 (above).
The following proposition gives a convenient way to describe the sets that
are open relative to S, or closed relative to S.
54
2.6 Topological spaces
Homeomorphisms
Let f be a univalent function from a topological space S onto a topological
space T. Then f has an inverse f - 1, whose domain is T. For each q E Tits
value f - l(q) is the unique pES such that f(p) = q. If both f and its inverse
f -1 are continuous functions, thenfis a homeomorphism. If there is a homeo-
morphism from S onto T, then Sand T are homeomorphic topological spaces.
In topology, homeomorphic spaces Sand T may be regarded as indis-
tinguishable. Every topological property enjoyed by S is also enjoyed by T.
For a deeper introduction to the literature on the general theory of
topological spaces, we refer to the work by Kelley [14].
PROBLEMS
1. Consider En with o/i x the collection of neighborhoods of x (open spherical n-balls
with center x), and o/i~ the collection of" neighborhoods" in Example 3.
(a) Verify that En with the collections o/i~ of "neighborhoods" satisfies Axioms (I)
through (4) for a topological space.
(b) Show that each U' E o/i~ contains some U EJU x; and each U E JU x contains
some U' E 'ft~.
(c) Using (b), show that the collections Yli x , J7t~ lead to the same collection of open
sets, and hence define the same topology on En.
2. Use Corollary 2 to show that each of the following sets is closed:
(a) {x: -2 ~ x ~ 2, x 3 - X ;::: O}.
(b) {(x,y):x 4 + y4 = I}.
(c) {x: Yo' x ~ Ixl}, Yo a given vector.
°
4. Let So = E2. Find whether A is open relative to S, closed relative to S, or neither.
(a) S = {(x, y): < x 2 + i
< I}, A = {(x, y) E S: x 2 ~ y2}.
(b) S = {(x,y):x 2 + y2 = 4},A = {(X,Y)ES:X 2 < i}.
5. Let S be an open subset of So. Show that the relatively open sets are just those open
subsets of So contained in S.
55
2 Elementary topology of En
9. Consider the following nonstandard topology on the plane E2: In this topology
the" 6-neighborhood" of (xo, Yo) is the set {(x, y) : Xo - 6 < x < Xo + 6, y = Yo}.
(a) Verify Axioms (1) through (4).
(b) Show that if D c E2 is open in the usual sense, then D is open in this topology,
but not conversely.
(c) Let fix, y) = g(x)h(y), where g and h have domain El and g is continuous in
the usual topology of £1. Show that f is continuous in this topology.
10. Let S = {(x,y):x 2 + y2 = I}, T = {(x,y):lxl + Iyl = I}, both with the relative
topology. Show that Sand T are homeomorphic topological spaces, by finding a
homeomorphism f from S onto T.
11. Prove Proposition 2.4'.
12. Prove Proposition 2.5'.
2.7 Connectedness
From the intuitive point of view, a set should be regarded as connected if it
consists of one piece. Tluus an interval on the real line E1 is connected, while
the set [0, 1] u [2, 3] is disconnected. For more complicated sets, intuition
is not a reliable guide.
56
2.7 Connectedness
EXAMPLE 2. Let S = {(x, y): i :-: :; Xl, X #- O}, So = £1. To show that S is
disconnected, let A = S (\ H + , B = S (\ H _ , where H + , H _ are the open
half-planes H + = {(x, y): x > O}, H _ = {(x, y): x < O}.
Let us next characterize completely the connected subsets of £1.
J
~~
{Z ~'
v
[x, y]
Figure 2.7
In Example 2, Section 1.1, four types of finite interval and four types of
semi-infinite interval were considered. Each of these is an interval according
to the definition just given. Moreover, £1 is an interval, and any set {x} with
a single point is an interval. These 10 types of intervals are the only possibilities
(Problem 5).
57
2 Elementary topology of En
58
2.7 Connectedness
PROBLEMS
1. Show from the definition that the following are disconnected subsets of the plane e:
(a) The hyperbola x 2 - y2 = 1.
(b) Any finite subset of E2 with at least two elements.
(c) {(x,y):x 2 + y2 S 4,y # x + I}.
2. Show that a topological space S is disconnected if and only if there exists a con-
tinuous function f on S, with f(p) = 0 or f(p) = I for all PES, but f(p) not constant
on S.
3. Let SI and S2 be connected subsets of En, such that Sin S2 is not empty. Show that
SI u S2 is connected.
5. Let J be an interval. Show that either J = E I , J has a single point, or J is of one of the
eight types listed in Example 2, Section 1.1.
6. Instead of Axiom III. (least upper bound property) about the real numbers, take
as an axiom the property that EI is connected. Prove Axiom III as a theorem.
[Hint: Let A = {all upper bounds of S} and B = A'. Show that B is open; and if S
has no least upper bound, then A is open.]
8. Let gl be a path from PI to P2 and gz a path from pz to P3' Let h(t) = gl(2t) if
Os t s 1. and h(t) = gz(2t - I) if! s t s 1. Show that h is a path from PI to P3'
9. Let D c en be open. By polygonal path in D from x to y let us mean a path g in D
from x to y with the following property: There exist to, t I, . . . , t m such that 0 =
to < t I < ... < tm- I < tm = I and g(t) = g(td + (t - tdvk if tk s t S tk+ I' where
Vk = (tu I - tk)-I[g(tu d - g(tk)J, k = 0, I .... , In - 1.
Let g be a polygonal path in D from x to y, and U be any convex set such that
y E U and U c D. Using Problem 8, find a polygonal path in D from x to any point
ZE U.
10. Let D c En be open. Given xED, let A = {y: there is a polygonal path in D from
x to y} and let B = D - A. Using Problem 9, show that A and B are open sets and
A is not empty. [Hint: Any neighborhood is convex.] If B is not empty, then D is
disconnected.
11. A set S is totally disconnected if no connected subset of S contains more than one
point. Show that the following are totally disconnected subsets of EI :
(a) S = {rational numbers}.
(b) S = C, the Cantor set (Section 2.4, Problem 5).
(c) Show that EI would be totally disconnected if we gave EI the nonstandard
topology in Problem 8, Section 2.6.
59
2 Elementary topology of En
2.8 Compactness
Let us next introduce another property, called compactness, which a subset
of a topological space may possess. For a set S c P, compactness of S
turns out to be equivalent to S being closed and bounded. Therefore, Theorem
2.4 can be rederived in a more elegant way.
Compactness is defined in terms of open coverings. Let S be a subset of a
topological space So. A collection mof su bsets of So is a copering of S if
eVery point of S belongs to some set A Em, that is, if S c UAE~A.If21' c m
and m' is also a covering of S, then m' is a subcovering. If every A Emis an
open subset of So, then mis an open copering of S.
We are interested in whether every open covering m
of S has a finite
subcovering; i.e. whether finitely many sets A" ... , Am E mexist such that
SeA, u··· u Am.
60
2.8 Compactness
The converse to the Heine- Borel theorem is true. To prove it, suppose first
that S is unbounded. Let Am be the neighborhood of 0 of radius m = 1,2, ....
Then {A I, A 2, ... } is an open covering of S with no finite subcovering. If S
is not closed, let Xo E fr S - S and Am = {x: Ix - xol > 11m}. Then
{A I, A z , . .. } is an open covering of S with no finite subcovering. This proves
the converse.
We summarize these results as follows.
Theorem 2.9. A set SeEn is compact if and only ifS is closed and bounded.
The reader is cautioned that Theorem 2.9 depends heavily on the finite
dimensionality of En. The concepts in this theorem all make sense in any
normed vector space (Section 2.9). However, closed, bounded subsets of an
infinite dimensional normed vector space need not be compact. For an
example, see Problem 4, Section 2.9.
Let us next show that the continuous image of a compact set S is compact.
Consider a function ffrom So into some topological space T, and S c So.
61
2 Elementary topology of En
PROOF. Consider first the case S = So. Let flJ be any open covering of I(S).
Since 1 is continuous,j - '(B) is open for every B E flJ. The collection of sets
1 - l(B) is an open covering of S. Since S is compact, a finite subcollection
{f - l(B 1), ... , 1 - l(Bm}} covers S. Then {B 1, ... , Bn.} is a finite subcollection
of flJ which covers I(S). Hence I(S) is compact. The case S c So reduces to
this one by considering S as a topological space with the relative topology
and by replacing 1 with the restriction 1 IS. 0
PROBLEMS
62
2.9 Metric spaces
Although the metrics d and d are different, they will give rise to the same
topology on the sphere (the relative topology).
63
2 Elementary topology of En
EXAMPLE 1 (continued). Theorem 2.2 states that En, with the euclidean
distance is a complete metric space. If SeEn with the euclidean distance, then
S is a complete metric space if and only if S is a closed subset of P. To see this,
suppose that S is closed. Let [xmJ be any Cauchy sequence, with Xm E S for
m = 1, 2, .... According to Theorem 2.2, the sequence [xmJ has a limit Xo.
Since S is closed, Xo E S (Problem 4a, Section 2.3). Thus S is a complete
metric space. On the other hand, if S is not closed, then there exists a sequence
[xmJ, with Xm E S for m = 1,2, ... ,converging to a limit Xo ¢ S (Problem 4b,
Section 2.3). The sequence [xmJ is Cauchy, but has no limit in S. Thus S is
not a complete metric space.
In Sections 2.10 and 5.13 we see that certain metric spaces whose elements
are functions are complete.
64
2.9 Metric spaces
Definition. A normed vector space is a vector space "fI together with a real
valued function I II with domain "fI, such that:
(1)
(2)
°
Ilull > for all u E "fI, u -=1= O.
I cu I = Ie III u I for every real c and u E "fl.
(3) Ilu + vii s Ilull + Ilvll for every u, v E "fl.
From Property (2) with c = 0, we have 11011 = 0, where 0 is the zero element
of "fl. From (2), (3), and induction on m, one can show that
IIJ/jUjll s JllcjlllU)1
for every choice of real numbers e\ ... , em and u I, ... , Urn E "fl.
We define the distance between u, v E "fI by
d(u, v) = Ilu - vii.
Then "fI is a metric space. Property (3) implies the triangle inequality (iii).
EXAMPLE 4. Let "fI = En, with the norm of a vector x = (Xl, ... ,xn) equal to
I xI = max { Ix I I, ... , Ixn I}.
Neighborhoods ofx in the metric d(x, y) = Ilx - yll are n-cubes (see Example
3, Section 2.6). This is just one of many nonstandard norms which can be
placed on En. It is seen in Section 2.11 that all norms on En lead to the usual
topology of En.
If Y = (i, ... , ym, ... ) is another such sequence, then the sum is defined by
x + y = (Xl + i, ... ,xm + ym, ... ).
Similarly, cx = (exl, ... , cx m , •.• ). Let "fI be the space of such sequences
for which the sum of squares of the components xm is finite. As norm we take
The reader should check that "fI is a normed vector space. Just as for En, a
set A c "fI is called bounded if there is a positive number C such that
Ilxll s C for all x EA. Unlike En, there exist closed bounded subsets of "fI
which are not compact (Problem 4).
A normed vector space "fI which is complete (in the metric d(u, v) =
Ilu - viI) is called a Banach space. The space En, with any norm, is a finite
dimensional Banach space. Examples of infinite dimensional Banach spaces
65
2 Elementary topology of En
are the sequence space in Example 5, the space <f6'(S) in Section 2.10, and the
U spaces in Section 5.13. We refer to the work by Taylor [22J for an intro-
duction to the theory of Banach spaces and their key role in current mathe-
matical analysis.
PROBLEMS
d = d(p, q)
(p, q) 1 + d(p, q)'
Show that d satisfies properties (i), (ii), and (iii) for a metric space.
4. Let ,,//" be as in Example 5.
(a) Verify Properties (1) through (3) for the norm I I in this example.
(b) Show that ,,/," is a complete metric space.
(c) For I = 1,2, ... let e, = (0, ... ,0, 1,0, ...) where I - 1 zeros precede the 1.
Let A = {e b e 2 , ••. }. Show that A is bounded and closed.
(d) Show that the set A in part (c) is not compact. [Hint: Let U, = {x: Ilx - e,ll < I}.
The collection {U I, U 2, ... } covers A.]
5. For any compact subsets A, B of En, let d(A, B) be the smallest number a with the
following property: for every x E B there exists YEA such that Ix - y I :::; a, and for
every YEA there exists x E B such that Ix - YI :::; a. Show d is a metric on the space
whose elements are all compact subsets of En.
°
6. Let S be a compact metric space, with metric d.
(a) Show that for any b > there is a finite set So c S such that any pES is distant
less than b from some q E So [i.e., d(p, q) < b].
(b) A set A c S is countable if either A is a finite set or A = {PI' pz, ... } fOl: some
infinite sequence [Pm] in S, Pm oF P, for I oF m. Show that there is a countable
set A with cl A = S.
7. A topological space So is called a Hausdorffspace if So has the property that for every
p, q E So (p oF q) there exist a neighborhood U of p and a neighborhood V of q such
that U ( l V is empty.
(a) Show that any metric space is a Hausdorff space.
(b) Show that any compact set S c So is closed, if So is a Hausdorff space.
(c) Let f be continuous and univalent from a compact space S onto a Hausdorff
space T. Show thatf- I is continuous from Tonto S. [Hint: Show that U- I )-I(B)
is closed if B is closed.]
66
2.10 Spaces of continuous functions
---r-----------------L----x
Figure 2.8
67
2 Elementary topology of En
Definition. Let fl, f2, ... be a sequence of functions, such that fm E £J6{S)
for m = 1,2, .... Then fm tends to f uniformly on S as m -+ 00 if
Ilfm - fll -+ 0 as m -+ 00.
EXAMPLE I. Let S = [0, 1], and let fm{x) = xm (mth power of x). Then
fm(x) -+ 0 as m -+ 00 for 0 ~ x < 1. Since fm(1) = 1, fm(l) -+ 1 as m -+ 00.
The sequence [fm] converges pointwise to f, where f(x) = 0 for 0 ~ x < 1,
f(1) = 1. However, the convergence is not uniform. This follows from
Theorem 2.11, since f is not a continuous function. A direct proof can also
be given (Problem 1).
Proposition 2.13. The space £J6{S), with the metric d(f, g) = Ilf - gil, is a
complete metric space.
PROOF. Let fl,f2, ... be any Cauchy sequence in £J6{S). Given I> > 0 there
exists N such that II}; - fmll < I> for every I, m ~ N. Since Ih(p) - fm{P) I ~
II}; - fmll for each pES,
Ih(p) - fm(P) I < I> for every I, m ~ N.
This shows that, for fixed p, the sequence of real numbers fl(P),f2(P), ...
is Cauchy. By Theorem 2.2, this sequence has a limit, which we denote by
f(P). This defines the function f. Let us show that Ilfm - fll -+ 0 as m -+ 00.
Given I> > 0 there exists N 1 such that I fm - };II < 1>/2 for every I, m ~ N l'
Since fm{P) -+ f(p) as m -+ 00, there exists N 2 > N 1 such that Ih(p) - f{p) I
< 1>/2 for every I ~ N 2' Note that N 2 may depend on both P and 1>, but N 1
depends only on 1>. Then
fm{P) - f{p) = [fm{P) - h(p)] + [};{p) - f{p)],
Ifm{P) - f{p) I ~ Ifm{P) - h(p)1 + Ih(p) - f(p)l·
68
2.10 Spaces of continuous functions
Theorem 2.12. The space rc(S), with the metric d(f, g) = Ilf - gil, is a
complete metric space.
PROOF. Let Um], m = 1,2, ... , be any Cauchy sequence in rc(S). Since
rc(S) c &4(S), the sequence is Cauchy in &4(S). By Proposition 2.13 &4(S) is
complete. Thus there exists a bounded function f such that I fm - f I -> 0
69
2 Elementary topology of en
PROBLEMS
l. Show directly from the definition of uniform convergence that the sequence [fm] in
Example 1 does not converge uniformly.
2. Let fm(x) = m2 x if 0 S x S m- I , j~(x) = m(2 - mx) if m- I S x S 2m- I , and
fm(x) = 0 for all other x.
(a) Find II fm II.
(b) Show that fm tends to 0 pointwise for each x, but that the convergence is not
uniform.
3. Let S be a finite set with n elements. Explain how '??(S) can be identified with the space
in Example 4, Section 2.9.
4. Letf(x) = D")= I (sin kx)/k 2 . Use Theorem 2.11 to show that f is continuous on EI.
5. Let SeEn be compact, and let Y be a compact subset of '??(S). Show that:
(a) There exists a number C such that II f II s C for all fEY.
(b) Given e > 0 there exists b > 0 depending on e (but not on f), such that
If(x) - f(y)1 < e for all fE.'!' and all X,YES satisfying Ix - yl < b. [Hint:
Problem 8, Section 2.5 and Problem 6a, Section 2.9.]
70
2.11 Noneuclidean norms on En
EXAMPLE I. Let
i= 1
The n-balls with respect to this norm are convex polytopes. For example, if
n = 2, the closed unit 2-ball {x : I xI ~ I} is the square with vertices e1, e2 ,
-e 1, -e 2 . Compare with Problem lla, Section 1.3.
The function B is called bilinear. Let us assume that B(x, x) > 0 for every
x :f. O. This means that the matrix (Ci) is positive definite. With B we associate
a quadratic norm, given by
71
2 Elementary topology of En
where ret, ... , en} is the standard basis for En. Then
and therefore
Ilxll ~ Mlxl·
From this Ilx - yll ~ Mix - yl for every x and y, which implies that
I I is a continuous function. Therefore it has a minimum value m on the
compact set {x : Ix I = I},
m = min{llxll : Ixl = I}.
By Axiom (1) for a norm m > O. If x = 0, all terms in (2.3) are O. Given
any x =1= 0, let e = Ixl-t. Then lexl = elxl = 1, and hence llexll ~ m. By
Axiom (2) lIexll = Jclllxll, from which
Ilxll ~ mlxl. o
From Proposition 2.14,
mix - yl ~ IIx - yll ~ Mix - yl
for every x and y. This says that the ratio of the I II-distance to the euclidean
distance is bounded between m and M. Let us call a set D II II-open if every
Xo ED has some neighborhood with respect to this norm which is contained
in D. If Ix - xol < (jjM, then IIx - xoll < (j. Therefore the (j-neighborhood
of Xo with respect to the norm II I contains the ordinary euclidean ((jjM)-
neighborhood of Xo. If D is I II-open, then every Xo E D has a euclidean
neighborhood contained in D. Hence D is open in the ordinary sense.
Similarly, the euclidean (j-neighborhood of Xo contains the (m(j)-neighbor-
hood of Xo with respect to the norm I II. It follows that every set which is
open in the ordinary sense is also I II-open. Thus a set is I II-open if and only
if it is open in the usual sense. Since the open subsets of a topological space
determine the topology (Section 2.6), all norms on En lead to the same top-
ology of En.
72
2.11 Noneuclidean norms on En
Theorem 2.13. Let K be any set with Properties (i) through (iv). Let 11011 = 0,
andJor every x "# 0, let
1
(2.5) IIxll = - - - - - - - , -
max{t: tx E K}
Then II II is a norm, and (2.4) holds.
PROOF. By (i) and (iv) there exist rl >
Iyl
and r2 > °
such that y E K if
srI and y ¢ K if Iyl > r2 • Hence given x "# 0, tx E K if It I s rl/lxl,
°
and tx ¢ Kif It I > r2/1xl (see Figure 2.9). Let
Sx = {t: tx E K}.
°
Then Sx contains the (r II Ix I)-neighborhood of and is bounded above by
r2/1 x I. Since K is a closed set, Sx is also closed. Hence Sx has a largest element
max Sx, which is positive. This shows that IIxll = Ilmax Sx is well-defined
and is positive. Moreover, by (ii) and the fact that 0 E K, the line segment
between 0 and any point of K is contained in K. Therefore max Sx 2: 1 if and
only if x E K, which says that (2.4) holds.
It remains to verify Axioms (2) and (3) for a norm. By Property (iii),
11- x II = II x II. It is left to the reader to check that if c > 0, then
I
max Sex = - max Sx,
c
73
2 Elementary topology of En
Figure 2.9
and consequently Ilexll = ellxll. Then Axiom (2) holds. For (3) we may
assume that x "# 0 and y "# O. Let
t
t = max Sy, U=--.
s +t
Observe that 0 < u < 1 and
1 1 1
- = - + - = Ilxll + Ilyll·
su s t
A little manipulation shows that su = (1 - u)t. Consequently,
su(x + y) = u(sx) + (1- u)ty.
Since sx, ty E K and K is a convex set by Property (ii), su(x + y) E K. There-
fore su S max Sx+y, and
1
Ilxll + Ilyll = -
su
z Ilx + YII·
This verifies Axiom (3). o
PROBLEMS
2. The ellipse K = {(x, y): x 2 + xy + 4y2 ~ I} has Properties (i) through (iv) in
Theorem 2.13. .
(a) For what (quadratic) norm is it the closed unit 2-ball?
(b) Find Ile l - e2 11.
74
2.11 Noneuclidean norms on En
Ii
4. Let p ~ 1, and K = {x: = i IXi Ip ~ I}.
(a) Show that K is convex. [Hint: Problem 3.]
(b) Show that K satisfies Properties (i) through (iv) in Theorem 2.13.
(c) Show that the norm associated with K in Theorem 2.13 is Ilxll = [Ii= i IxiIP]i/ p•
Note. For this norm the inequality Ilx + yll ~ Ilxll + Ilyll is called
Minkowski's inequality. There is a related inequality for integrals, which we
prove in Section 5.13.
5. A seminorm on En is a real valued function f satisfying: f(x) ~ 0 for every x;
f(cx) = Ie If(x) for every c and x; and f(x + y) ~ f(x) + f(y) for every x and y.
(a) Let f be a seminorm and K = {x: f(x) ~ I}. Show that K is closed and satisfies
Properties (ii) through (iv). Show that K is compact if and only if f is a norm.
[Hint: First prove that f is continuous.]
(b) Conversely, let K be any closed set satisfying Properties (ii) through (iv). Let
f(x) = 0 if x = 0 or if the line through 0 and x is contained in K. Otherwise, let
1
f(x) = - - , - - - - - - - c -
max{t: tx E K}
as in (2.5). Show that f is a seminorm.
(c) Let n = 3 and f(x, y, z) = Ixl + 21YI. Sketch K and show that f is a semi norm.
75
3
Differentiation of
real valued functions
We now begin the differential calculus for real valued functions of several
variables. The first step is to define the notions of directional derivative
and partial derivative. Then the concept of differentiable function is intro-
duced, by linear approximation to the increments of a function. Taylor's
formula with remainder is obtained for functions of class e(q); such functions
have continuous partial derivatives of orders 1,2, ... , q. It is then applied
to problems of relative extrema and to the characterization of convex
functions of class e(2).
The chain rule for partial derivatives is postponed to Chapter 4, since it
is a natural corollary of the composite function theorem for vector-valued
functions to be proved there.
76
3.1 Directional and partial derivatives
(3.2)
Figure 3.1
77
3 Differentiation of real valued functions
Thus
provided the limit exists. Stated in less precise terms, /;(x) is the derivative
taken with respect to the ith variable while holding all other variables fixed.
The symbol}; denotes the real valued function whose value at X is };(x).
Its domain is the set of points where f has an ith partial derivative.
For purposes of brevity, we occasionally abuse the notation by writing
}; for the value /;(x) at some particular x. In each such instance this abuse is
indicated either explicitly or by the context.
EXAMPLE 4. Let f(x) = ljJ[g(x)] for every XED. Suppose that the ith partial
derivative of g at Xo and the derivative of IjJ at g(xo) exist. By the composite
function theorem for functions of one variable
(3.4)
78
3.2 Linear functions
PROBLEMS
Unless otherwise stated, the domain D off is en for the particular n indicated
in the problem.
1. In each case find the partial derivatives of f.
(a) f(x, y) = x log(xy), D = {(x, y): xy > O}.
(b) f(x, y, z) = (x 2 + 2y2 + Z)3.
(c) f(x) = x • x.
2. Let f(x, y) = (x - 1)2 - i. Find the derivative of fat e 2 in any direction v, using
the definition of directional derivative.
3. Let f(x, y) = 2xy(x 2 + y2)-1/2, if (x, y) =I (0,0), and frO, 0) = O. Find the derivative
of fat (0, 0) in any direction v.
4. Let f(x, y) = (xy)I/3. (a) Using the definition of directional derivative, show that
fl(O,O) = f2(0, 0) = 0, and that ±e l , ±e 2 are the only directions in which the
derivative at (0, 0) exists. (b) Show that f is continuous at (0, 0).
5. Let f(x, y, z) = 1x + y + z I. If Xo = (xo, Yo, zo) is such that Xo + Yo + Zo = 0, find
those directions v in which the derivative at Xo exists.
6. Show that the derivative off at Xo in the direction - v is the negative of the derivative
at Xo in the direction v.
These two conditions are equivalent to the single condition L(cx + dy) =
cL(x) + dL(y) for every x, y E En and scalars c, d. By induction, if L is linear,
for every m, Xl> ... , xm E En, and scalars cl, ... , cm. In words, this states
that" L of a linear combination of Xl> •.• , Xm is the corresponding linear
combination of L(x l ), ... , L(xm)."
If ai' ... , an are real numbers, then the function L defined by
(3.6)
for every X E En, is linear. Conversely, if L is a linear function, let
ai = L(e i ), i = 1, ... , n.
For each x we have
x = xle l + ... + xnen.
Applying (3.5) with cj = x j , Xj = e j , and m = n, we get
L(x) = Xl L(e l ) + ... + xnL(en).
This has the form (3.6).
Let us denote the right-hand side of (3.6) by a' x. We have proved the
following.
Proposition 3.1. A real valued Junction L is linear if and only if there exist
real numbers a I, ... , an such that L(x) = a . x Jor every x E P. 0
The object a is called a covector and al' ... ' an are its components.
Covectors are not elements of En, but belong to the n-dimensional vector
space (En)* dual to En. This is explained in more detail later in the present
section.
Note that we have written ai with subscripts, while the components Xi of a
vector x are written with superscripts. We call the n-tuple a = (aI' ... , an)
a co vector, and ai' ... , an are its components. The number a' x is called
the scalar product of the co vector a and vector x. If a and bare covectors,
then the sum a + b and product of a by a scalar c are defined by
80
3.2 Linear functions
Xi(e) = 3~,
where 3~ = 3ij is Kronecker's delta. Let us denote by ei the covector cor-
responding to the linear function Xi. The ith component of ei is 1, and the
other components are O. We call {et, ... , en} the standard basis for (En)*.
It is dual to the standard basis {e l , ... , en} for En, in the sense explained in
Appendix A.l.
The notation is chosen so that for every formula about vectors there is
a corresponding formula about co vectors obtained by interchanging sub-
scripts and superscripts. For instance, the components of a vector x satisfy
Xi = Xi(X) = ei ° x. The corresponding formula for the components of a
covector a is ai = a ° ei . In (En)* a euclidean inner product and norm are
defined in the same way as in En. These facts are summarized in the table
that follows.
Vectors Covectors
i= 1
a°b= I
i= 1
aibi
n
Scalar product a ° x = Iaixi
i= 1
a ° ei = ai
For some purposes, one can avoid the use of covectors by the simple
device of raising indices.lfa = (al,"" an) is a covector, consider the vector
Y = (i, ... , yn) such that i = ai for i = 1, ... , n. Then a ° x = yo x for any
vector x, where ° on the left side means the scalar product and on the right
side, the euclidean inner product. However, this device is often unsatisfactory.
The difference in behavior of vectors and covectors under linear transfor-
mations was already mentioned. Moreover, it is sometimes convenient to
81
3 Differentiation of real valued functions
give En a noneuclidean inner product. This affects the formula for changing
covectors into vectors. (See the distinction between the differential and
gradient of a function, Section 3.3.)
PROBLEMS
4. Let {XI' ... , x"} be a basis for E". Define L by the formula
L(clxl + ... + cox") = c"
for every c I, ... , cO. Show that:
(a) L is a linear function.
(b) The set P = {x: L(x) = O} is the (n - I)-dimensional vector subspace spanned
by {XI' ... , x"_ d.
5. Let II II be any norm on E" (Section 2.11). Define on (E")* the dual norm as follows.
For every covector a,
Iiall = max{a' x: Ilxll = I}.
(a) Verify that the dual norm satisfies Properties (1), (2), and (3) (Section 2.9).
(b) Show that IIxll = max{a' x: lIall = I}. [Hint: Problem 7, Section 2.4.]
82
3.3 Differentiable functions
(3.7) I·
h~
f(xo + h) - f(xo) - L(h) -
Ihl - .
°
Let us show that iff is differentiable at x o , then f has a derivative at Xo
in every direction v. Taking h = tv, (3.7) implies that
n
(3.9) df(x o) . h = L .fi(xO)h i•
i=1
83
3 Differentiation of real valued functions
=
(x. ;')
C~L.;,~t1r- (x, z)
I
I
I
I
I I 2
~'----: - :----- x
,I ,
I '
:
X,
~
I XO =0 = !( x o). ; = ! (x )
;' '" =0 + d/ (x o) ' (x - xol
Figure 3.2
except at (0, 0). Moreover'!l andf2 are continuous functions except at (0, 0).
By Theorem 3.2 in Section 3.4,! is differentiable at any (xo, Yo) # (0, 0). The
components of df(xo, Yo) are fl (xo, Yo) and f2(XO, Yo). The equation for the
tangent plane at (xo, Yo, f(xo, Yo)) is
84
3.3 Differentiable functions
for every h such that 0 < Ihl < b. This shows that
lim f(xo + h) = f(x o),
h-+O
EXAMPLE 2. Let
2xy2
f(x, y) =
x
2+
y
4' if (x, y) =1= (0, 0), and f(O, 0) = o.
There are two cases to consider. If cos e =1= 0, then the derivative at (0, 0) in
the direction (cos e, sin e) is
lim f(t cos e, t sin e) - f(O,O) = lim 2 cos esin 2 e = 2 sin 2 e
1-+0 t 1-+0 cos 2 e+ t 2 sin 4 e cos (J •
85
3 Differentiation of real valued functions
If cos 8 = 0, then f(t cos 8, t sin 8) = 0 for every t and the directional
derivative at (0,0) is O. However,J(y2, y) = 1 for every y # O. Sincef(O, 0) = 0,
f is not continuous at (0, 0). By Theorem 3.1,f is not differentiable at (0, 0).
Let us next state a proposition which, although of no interest in itself, is
useful later.
Proposition 3.2. Let 4>(t) = f(x o + th). Then for elWY t such that f is differ-
entiable at Xo + th,
4>'(t) = df(x o + th) . h.
PRCX)F. If h = 0 the result is trivial. If h # 0, then
o= lim f(x o + th + tt) - f(xo + th) - df(xo + th) . tt .
q-O Ittl
In particular, let tt = rho Then
o = lim 4>(t + r) - 4>(t) - r df(x o + th) . h
r '
Mean value theorem. Let f be continuous at every point x of the line segment I
joining Xo and Xo + h, with f differentiable at each point of I except perhaps
the endpoints. Then there exists a number s E (0, 1) such that
f(x o + h) - f(x o) = df(xo + sh) . h.
PROOF. Let 4>(t) = f(x o + th), as in Proposition 3.2. Then 4> is continuous
on [0, 1], and
4>(1) = f(x o + h),
4>'(t) = df(x o + th) . h, if 0 < t < 1.
By the mean value theorem for functions of one variable (Section A.2) there
exists s E (0, 1) such that 4>(1) - 4>(0) = 4>'(s). D
86
3.3 Differentiable functions
df(x) = L/;(x)e i
i= 1
87
3 Differentiation of real valued functions
Note. This definition of the gradient vector is correct only if we use the
euclidean inner product in En. Suppose that En is given some other inner
product B(x, y) = r::,j= 1 cijxii, with (Ci) a symmetric, positive definite
matrix (see Section 2.11). Let us denote the gradient vector with respect to
this inner product by grad B f(x). We require that grad B f(x) satisfy
B(gradBf(x), h) = df(x)' h, for all h E En.
To find the components of the vector z = gradBf(x), let h = e i in (3.9). Then
2>
n
j= 1
ijz i = !;(x), i = 1, ... , n.
This is a system of linear equations for the components z1, ... ,zn of
grad B f(x). Let (c ij ) denote the inverse of the matrix (cij)' Then
n
Zi = I ciiJ';{x), i = 1, ... , n,
j= 1
n
(3.12) gradBf(x) = I cijjj(x)e i ·
i.j= 1
PROBLEMS
2. Using the formula df(x o)' v for directional derivative, find the derivative of fat
Xo in the direction v.
(a) f(x, y) = xy, Xo = (1,3), v = (2/j-S, -1/jS).
(b) f(x, y) = x exp(xy), Xo = e l - e z , v = (l/j2)(e l + ez).
(c) f(x, y, z) = ax z + bi + cz z, Xo = e l , v = e 3 .
3. Letf(x, y) = log(x z + 2y + I) + J~ cos(tZ)dt, y > -!-
(a) Find df(x, y).
(b) Find approximately f(O.03, 0.03).
5. In Problem 3, Section 3.1, show that f is continuous at (0, 0), but not differentiable
at (0, 0),
6. Let f(x, y) = 2xyZ /(x z + y4), if (x, y) # (0,0), and f(O, 0) = 0, as in Example 2.
(a) Show that - I :$ f(x, y) :$ I for every (x, y).
(b) Find {(x, y): f(x, y) = l} and {(x, y): f(x, y) = -I}.
(c) Find {(x, y): grad f(x, y) = (0, On.
(d) Find {(x, y): f(x, y) = c} for any c, and illustrate with a sketch.
88
3.4 Functions of class C(q)
7. Let f and 9 be differentiable at Xo. (a) Prove that the sum f + 9 is differentiable
at xo, and d(f + g)(x o) = df(x o) + dg(xo). (b) Prove that the product fg is dif-
ferentiable at x o , and d(fg)(xo) = f(xojdg(xo) + g(xojdf(x o). [Hint: Recall the
proof for n = 1.J
8. (Euler's formula.) Let p be a real number. A functionf is called homogeneous of degree
p if f(tx) = tPf(x) for every x "" 0 and t > O. Let f be differentiable for all x "" O.
Show that if f is homogeneous of degree p, then
df(x) . x = pf(x)
for every x "" 0, and conversely. [Hint: Let ¢(t) = f(tx) and use Proposition 3.2
with Xo = O. For the converse, show that for fixed x, ¢(t)t- P is a constant.J
9. Let Q(x) = I7.
j = I CijXixj, where Cij = C ji and Q(x) > 0 for every x "" O. Let
f(x) = [Q(X)JP/2. Calculate df(x) and verify Euler's formula for this function.
10. Let f be continuous at Xo and 9 differentiable at Xo with g(x o) = O. Show that the
product fg is differentiable at Xo.
If the partial derivativesfl(x), ... , fn(x) exist for every XED andfl, ... ,fn
are continuous functions on D, thenfis afunction of class C(i).
i = 1, ... , n - 1.
89
3 Differentiation of real valued functions
x"
--+---------------x
Figure 3.3
Using the above inequalities, the triangle inequality, and the fact that
Ih" I ~ Ihi, 161
~ Ihi, we get
EXAMPLE I. IfJand 9 are functions of class C(1) with the same domain D, then
J + 9 is of class C(1). Using the product rule from elementary calculus, the
EXAMPLE 2. The composite of two functions of class C(1) is also of class C(1).
Suppose that J = I/J 0 g, where I/J and 9 are of class C(1). By Formula (3.4),
}; = (I/J' g)gi' Since I/J' and 9 are continuous, their composite I/J' 9 is
0 0
Ii/x) or
a2J
ax j axi (x).
For example, if J(x, y) = x2 l then J,(x, y) = 2xl, Jll(X, y) = 2y 3,
Jdx, y) = 6xl·
If all of the partial derivatives Ii/x), i, j = 1, ... , n, exist at every xED
and each};j is a continuous function, thenJis called aJunction oj class C(2).
By the corollary to Theorem 3.2, ifJis of class C(2) then J" ... ,fn are continuous.
Hence any function of class C(2) is also of class C(').
The partial derivatives of J of order q = 3,4, ... are defined similarly,
wherever they exist. The notation for partial derivative at x, first in the
direction ei " second in the direction ej2' and so on, is
1 S i, S n, 1= 1, ... , q.
Definition. If all of the qth order derivatives of J exist at every XED and
each };" ... , iq is a continuous function on D, then J is a Junction oj class C<q),
91
3 Differentiation of real valued functions
Theorem 3.3. Iff is of class e(1) and both Jij and jji are continuous, i =1= j, then
Jij = hi.
PROOF. Suppose first that n = 2. We need to show thatf12 = f2l. Let (xo, Yo)
be any point of D, and 15 0 > 0 such that the bo-neighborhood of (xo, Yo)
is contained in D. For 0 < u <bo/fi let
1
A(u) = 2 [f(xo
u
+ u, Yo + u) - f(xo, Yo + u) - f(xo + u, Yo) + f(xo, Yo)]·
A(u) is sometimes called the second difference quotient. Let
g(x) = f(x, Yo + u) - f(x, Yo)
for every x such that (x, Yo + u), (x, Yo) E D. The domain of g is an open
subset of £1 which contains the closed interval [xo, Xo + u]. Moreover,
g'(x) = fl(X, Yo + u) - fl(X, Yo). Thusg is of class e(1) sincefl is continuous,
and
1
A(u) = 2 [g(xo + u) - g(xo)]·
u
Applying the mean value theorem to g, there exists ~ E (xo, Xo + u) such that
A(u) = ~u g'@ = ~u [fl(~' Yo + u) - fl(~' Yo)]
(see Figure 3.4). Of course the number ~ depends on u. Let
h(y) = fl(~' y)
for every y such that (~, y) E D. The domain of h is open and contains
[Yo, Yo + u]. Moreover, h'(y) = fde, y), h is of class e(1) since f12 is con-
tinuous, and
1
A(u) = - [h(yo + u) - h(yo)].
u
92
3.4 Functions of class C(q)
----...
y
(Xo,Yo+U) (xo+u,Yo+u)
1
I
(~, 1'/)1
1'/ ---- ---~
I
i (Xo + u, Yo)
I
I
----~----------~-----x
Figure 3.4
93
3 Differentiation of real valued functions
n
r/>(q)(t) =. ~ h, ..... iq(X O + thW', ... , hiq .
11, •.•• Iq= 1
The sum in the last formula has nq terms. It is taken over all integers ii'
i2, ... , iq, such that 1 ::; i 1 , • •. , iq ::; n.
By Taylor's formula for functions of one variable (Section A.2) there
exists s E (0, 1) such that
1 1
r/>(1) = r/>(O) + r/>'(O) + - r/>"(O) + ... + r/>(q-l)(O) + - r/>(q)(s).
2! (q - 1)! q!
But r/>(l) = f(x), r/>(O) = f(x o), and we have, by substitution, Taylor's
formula with remainder:
1 ~ ..
+ ... + ( -1)!. ~ _ h, ..... iq~,(xo)h'I .. ·h'q~1 + Ri x ),
q It •.•• , l q _ l - 1
Notice that the first terms on the right-hand side are just the first degree
approximationf(x o) + df(x o)' h tof(x) considered in Section 3.3.
If we ignore the remainder Rix), the right-hand side is a polynomial in
hi, ... , hn of degree q - 1. If the remainder is small this polynomial furnishes
an approximation to f(x). One can give an explicit estimate for the error, in
terms of bounds for the qth-order partial derivatives off Suppose that K is
sh E K for
satisfy
°: ;
a convex subset of D, such that Xo E K and x = Xo + hE K. Then Xo +
s ::; 1. Moreover, suppose that all qth-order partial derivatives
94
3.4 Functions of class CCq)
fl = 2xy, -
f 2- X2, fll = 2y,
f12 = f21 = 2x, fl1 2 = fl 2 1 = f2 11 = 2,
95
3 Differentiation of real valued functions
and all other partial derivatives are 0. R 4 (x, y) = 0, and Taylor's formula
becomes
f(x, y) = f(l, -1) + J~(x - 1) + fiY + 1)
1 2
+ 21 [fl1(x - 1) + 2fdx - 1)(y + 1)]
1 2
+ 31 [3fI12(X - 1) (y + 1)],
where the partial derivatives on the right-hand side are evaluated at (1, -1).
Thus
x 2 y = -1 - 2(x - 1) + (y + 1) - (x - 1)2 + 2(x - l)(y + 1)
+ (x - 1)2(y + 1).
°
EXAMPLE 7. In Example 3, let p > be an integer. Let A = [0, ex)), DI = EI,
F(x) = x P for all x. Then f is of class C(q) on A for any q, and F is an extension
of class c<q) of f IA.
96
3.4 Functions of class C(q)
class e(q) on En such that F(x) = f(x) for every x E A. Hence f is of class cq)
on A.
Actually, what one needs to assume about A to apply Whitney's theorem
is the following: Every Xo E A has a neighborhood U such that any pair of
points x, Y E U n B can be joined in B by a polygon of length no more than
c Ix - YI, where c ~ 1 depends only on U. If A is convex, then the line segment
joining x and y lies in B, and one may take c = 1.
For other extension theorems of Whitney, see Trans. Amer. Math. Soc.
36 (1934) and Bull. Amer. Math. Soc. 50 (1944).
we may put the corresponding infinite series. This infinite series is called the
Taylor series for f(x) at Xo.
If K is a convex subset of D and Xo E K, then the following is a sufficient
condition that f(x) be the sum of its Taylor series for every x E K. Suppose
that there is a positive number M whose qth power bounds every qth-order
partial derivative of f, namely,
(3.16)
for every x E K, q = 1,2, ... , and 1 ::; iI' ... , iq ::; n. Then (3.15) holds,
with e = Mq:
q.1 '
where B = Mnl/2Ihl. Since Bqjq! -+ °as q -+ 00,
lim Rix) = °
for every x E K, provided inequalities (3.16) hold.
A function f is called analytic if every Xo E D has a neighborhood U Xo
such that the Taylor series at Xo converges to f(x) for every x E U xo . We
have just proved the following:
° if x ::; 0.
97
3 Differentiation of real valued functions
Let us show that this function is of class C(oo) and that f(q)(O) = 0 for every
q = 1,2, .... For x i= 0 the derivatives f(q)(x) can be computed by elemen-
tary calculus, and each f(q) is continuous on El - {O}. It is at the point 0
where f must be examined. Now
f'(x) = x 3 ~ exp(-~)
x~
= 2x· ~ exp(- ~).
X4 x 2
PROBLEMS
1. Expand J(x, y, z) = xyz by Taylor's formula about Xo = (1, -1,0), with q = 4.
2. Let J(x, y) = X-I cos y, x > 0, and Xo = (1,0).
(a) Expand J(x, y) by Taylor's formula about xo , with q = 2, and find an estimate for
IR 2 (x, Y)I.
(b) Show that R/'(. y) --+ 0 as q --+ 00 for (x. y) in some open set containing Xo.
98
3.5 Relative extrema
3. Let f(x, y) = !/t(ax + by), where a and b are scalars and !/t is of class C(q) in some
open set containing 0. Show that Taylor's formula about (0, 0) becomes
where (j) is the binomial coefficient (which equals the number of j-element subsets
of a set with m elements).
4. Let f be continuous on an open set D and of class C(l) on D - {xo}. Suppose, more-
over, that Ii = lim x _ xo };(x) exists for each i = 1, ... , n. Prove that Ii = };(xo), and
consequently that f is of class C(l) on D. State and prove a corresponding result in
case q > 1. [Hint: Apply the mean value theorem to f(x o + te;) - f(x o).
5. Prove the statement made in Example 3. [Hint: Problem 4 with Xo = O.J
6. Let f(x) = Xk sin( Ijx) if x i= 0, and frO) = 0. Show that:
(a) If k = 0, then f is discontinuous at 0.
(b) If k = 1, then f is of class C(O) but not differentiable at O.
(c) If k = 2, then f is differentiable but not of class C(l).
(d) What can you say for k ;:::: 3?
7. Let f(x, y) = xy(x 2 - y2)j(X 2 + i), if (x, y) i= (0,0), and frO, 0) = 0.
(a) If (x, y) i= (0, 0), find fdx, y) and f21 (x, y) by elementary calculus, and verify
that they are equal.
(b) Using Problem 4 show that f1(0, 0) = f2(0, 0) = 0 and f is of class C(l).
(c) Using the definition of partial derivative, show that fI2(0, 0) and f21 (0,0) exist
but are not equal. Why does this not contradict Theorem 3.3?
8. Given nand q, how many solutions of the equation i 1 + ... + in = q are there with
ii' ... , in nonnegative integers? With i 1 , ... , in positive integers? What does this
say about the number of different qth-order partial derivatives of a function of class
Clq)?
!(x o) = min{f(x): x E A}
is the minimum value of f on A. (Of course, there need not be any such
point xo. However, if A is a compact set, then by Theorem 2.5, any
continuous function has an absolute minimum at some point of A.)
If f(x o) < f(x) for every x E A except x o , then f has a strict absolute
minimum at Xo.
99
3 Differentiation of real valued functions
100
3.5 Relative extrema
derivative is
1
grad f(x) • v(x) = 1grad f(x) 1 grad f(x) • grad f(x) = 1 grad f(x) I·
(a) (b)
Figure 3.5
101
3 Differentiation of real valued functions
102
3.5 Relative extrema
°
for some ho E En. Then there exists a neighborhood V I of x o , V I C V,
such that Q(y, h o) < for all y E V I. Take x = Xo + ch o where Ic I is small
enough that x E V I, and y = Xo + sh. Then h = ch o: and since Q(x o + sh, )
is quadratic,
Q(xo + sh, h) = c2Q(X O + sh, h o) < O.
But Q(x o + sh, h) z 0, a contradiction. This proves (a).
To prove (a'), suppose that Q(x o , ) > O. Using Problem 8 and the fact
that the functions Iij are continuous at X o , there exists a neighborhood V of
Xo such that V c A and Q(y, ) > 0 for every y E U. Taking y = Xo + sh,
we find from (3.19) that I(x) > I(x o ) for every x E V, x #- Xo. This proves
that I has a strict relative minimum at Xo. Statements (b), (b') follow respec-
tively from (a), (a') by considering -I. D
Criterion I
In this case, with n = 2,
Q(x, y, h, k) = I11h2 + 2I12hk + I22k2
where we have written (h, k) for (h I, h 2) and Iij for Iij(x, y).
Since (x, y) is a nondegenerate critical point, d 2(x, y) = II d22 - I~2 #- 0.
Consider first the case when 111122 - Ii2 > 0. Then the quadratic equation
°
Q(x, y, h, k) = has no roots (h, k) except (0,0). Therefore, Q(x, y, h, k) has
the same sign for all (h, k) #- (0,0). We have Q(x, y, h, 0) = Illh 2, Q(x, y, 0, k)
= I22k2. The sign of 111 and 122 determines whether Q(x, y, , ) > 0 or
Q(x, y, , ) < O.
By Theorem 3.4 (a'), (b') we find that a critical point (x, y) is a point of
relative minimum if III > 0,122 > 0, II d22 - Ii2 > 0, relative maximum
if 111 < 0,f22 < 0,f11122 - Ii2 > 0.
If II d22 - Ii2 < 0, then {(h, k): Q(x, y, h, k) = O} consists of two lines
intersecting at (0,0). They divide the (h, k)-plane into four parts, on two of
which Q(x, y, h, k) > 0 and on the other two of which Q(x, y, h, k) < O. In
this case, Q(x, y, , ) is indefinite. A critical point where 111122 - Ii2 < 0 is
called a saddle point. The function illustrated by Figure 3.5(b) has one point
of absolute maximum, one of relative maximum, and one saddle point.
103
3 Differentiation of real valued functions
EXAMPLE. Let f(x, y) = 2y2 - x(x - 1)2 for every (x, y) E £2 and A = £2.
This function has two critical points, (1, 0) and (1,0). We find that
fll = 4 - 6x, f22 = 4, fllf22 - fI2 = 16 - 24x.
The point H, 0) gives a relative minimum and (1, 0) is a saddle point. In this
example it is instructive to find the level sets {(x, y): f(x, y) = c}. They are
indicated in Figure 3.6 for the critical values - 247 = fH, 0),0 = f(1, 0) and
for nearby values of c.
c= -r.,
----Hr-H~-----r.**-+H_-------------x
Figure 3.6
The point (t, 0) of relative minimum is an isolated point of the level set
containing it. For - n < c < 0 the level set has two parts. The one that
n.
encloses (t, 0) resembles a small ellipse if c is near - This can be attributed
to the fact that near (t, 0), f(x, y) is approximated by the first two nonzero
terms in its Taylor expansion about (t, 0), namely,
f(t, 0) + tQH, 0, x -1, y) = -n + (x -1)2 + 2y2.
The level sets of this quadratic function are ellipses with center H, 0)- if
c > - 247. Similarly, f(x, y) is approximated by -(x - 1)2 + 2y2 near the
saddle point (1,0). The level sets -(x - 1)2 + 2y2 = c are hyperbolas if
c =F O. Near (1,0) the level sets of f resemble these hyperbolas. For c = 0
we get the lines fly = ±(x - 1) tangent to the level set offat (1, 0).
Let us now state two criteria for Q(x, ) > 0 (or Q(x, ) < 0) that hold
for any dimension n. By Theorem 3.4 these imply that a critical point x
gives a relative minimum (or relative maximum).
Criterion II
For any n let
104
3.5 Relative extrema
These are called the principal minor determinants of the matrix (fiix)). The
mth principal minor dm(x) is the determinant of the matrix obtained by
deleting the last n-m rows and columns. The determinant dix) is the Hessian
of fat x.
Let us state without proof the following criterion:
For a proof of the first of these two statements, see the work by Bocher [3J,
especially pp. 140 and 147. The second follows from the first by considering
- Q. Here iff is an abbreviation for" if and only if."
Criterion II is fairly convenient for small values of n, but becomes unwieldy
for larger ones. This is because of the very large number of operations required
to calculate the determinant of an m x m matrix even for moderately small m.
Criterion III
In linear algebra it is shown that any quadratic form can be written as a
linear combination of squares by suitably choosing a new orthonormal
basis for P. This fact is also proved in Section 4.8 below. Therefore
n
(3.20) Q(x, h) = L Ai(X) [1Ji(x)] 2,
i= I
where for each hE En, 1JI(X), ... , 1Jn(x) are the components of h with respect
to some orthonormal basis {vI(x), ... , vn(x)} for En,
n
h= L 1J i(X)Vi(X).
i=1
The numbers AI (x), ... , An(X) are just the characteristic values of the matrix
(fi}{x)).
If Ai(X) > 0 for each i = 1, ... , n, then from (3.20) Q(x, h) > 0 unless
1Ji(X) = 0 for each i (that is, unless h = 0). In this case Q(x, ) is positive
definite. Conversely, ifh = Vi(X), then Q(x, h) = Ai(X). Therefore, ifQ(x, ) > 0,
then in particular Q(x, vi(x)) > 0, and Ai(X) > 0. This proves the first of the
following statements:
The second is proved in the same way. Replacing on both sides" >0" by
"~O" we get a criterion for nonnegative semidefiniteness, and replacing
"<0" by" :::;;0," one for nonpositive semidefiniteness.
If n is fairly large it is better, instead of Criterion II, to try some numerical
method for putting Q(x, ) in the form (3.20).
105
3 Differentiation of real valued functions
Boundary extrema
If f is continuous on a compact set A, then f has absolute extrema on A.
They may occur either at interior or at boundary points of A. If an absolute
maximum occurs at an interior point Xo of A, then Xo is among the relative
maxima in int A. We can try to find it by Theorem 3.4. However, Theorem 3.4
does not apply at boundary points of A.
If Xo E fr A and Xo gives an absolute maximum, then f(x) ~ f(xo) for
every x E A, and in particular for every x E fr A. Therefore Xo also gives an
absolute maximum among points of fr A. If fr A is sufficiently smooth, the
Lagrange multiplier rule (Section 4.8) can be applied.
In Section 3.6 we discuss extrema of linear functions, for which calculus
is of no use.
PROBLEMS
3. Let f(x, y, z) = x 2 + y2 - Z2. Show that f has one critical point, which does not
give a relative extremum. Describe the level sets.
4. Let f(x, y, z) = x2 + 3y2 + 2Z2 - 2xy + 2xz. Show that 0 is the minimum value
off
5. Given XI"", Xm , find the point X where Ij~ I Ix - xjl2 has 'an absolute minimum,
and find the minimum value.
6. (a) In Problem I(a) find the (absolute) maximum and minimum values of f on
the circular disk x 2 + l ::s; 1.
(b) Do the same for I(c).
7. (a) Let f(x) = I/I(a . x), where 1/1 is of class e(2) and a =I- O. Find all critical points,
and show that every critical point is degenerate.
(b) Illustrate this result in case f(x, y) = (x _ y)2.
106
3.6 Convex and concave functions
8. Let g(h) = D.j= I cijhih j . Assume that g > 0, that is, that g(h) > 0 for every h =1= O.
(a) Show that there exists a number m > 0 such that g(h) ~ m Ih 12 for every h.
[Hint: The polynomial g is continuous and has a positive minimum value m
on the unit (n - I)-sphere {h: Ihl = 1}.]
(b) SupposethatlCij - cijl < w- 2 foreachi,j = 1, ... ,n.LetG(h) = L7.
j =1 Cijhih •
j
where each Yi E U. Show that if U is small enough, detC/:iy;)) =1= 0 and consequently
the system of equations (*) has only the solution x - Xo = 0, a contradiction.]
Definition. The function f is convex on K if, for every x I, X z E K and t E [0, 1],
(3.2la) f(tx I + (1 - t)xz) ~ t{(xd + (1 - t)f(xz).
The assumption that K is a convex set is needed to ensure that the point
tX I + (1 - t)x z belongs to the domain of f. In order to see the geometric
meaning of convexity, let us denote points of En + I by (Xl, ... , x n , z) or, for
short, by (x, z). Let
K+ = {(x, z): x E K, z ?: f(x)}.
If Xl = x z , then (3.2la) holds trivially. Therefore, suppose that Xl =F Xz .
Let I denote the line segment in En + l joining (XI,f(X I )) and (xz,f(x z)).
Points of I are of the form
(tXI + (1 - t)xz, tf(xd + (l - t)f(x z )),
where t E [0,1]. Inequality (3.2la) says that such points belong to K+.
Therefore, the definition says geometrically that the line segment I is con-
tained in K+ for every pair of points Xl' Xz E K.
107
3 Differentiation of real valued functions
PROOF. Let f be convex on K. Let (Xl> ZI)' (X2' Z2) E K+, (XI' ZI) -# (X2' Z2),
and l' be the line segment joining them. Let (x, z) be any point of 1'. Then
X = tX I + (1 - t)X2'
Z = tZ I + (l - t)Z2,
I
(x, j(x))
I
---r--~~~--~~--~--x
XI X X2
Y
K
Figure 3.7
108
3.6 Convex and concave functions
The same proof shows that {X: f(x) < c} is also convex. The converse to
Proposition 3.5 is false; for example, let f be any increasing function with
domain EI. Then Kc is either all of EI, a semiinfinite interval, or the empty
set. In each case Kc is convex. However, f need not be a convex function; for
instance, if f(x) = x 3 , then f is not convex on EI.
EXAMPLE I. Let A be any nonempty closed subset of En. For every x, let
f(x) be the distance from X to A, namely,
f(x) = min{lx - yl: YEA}.
Let us show that the function f so defined is convex on En if and only if A is a
convex seLIf f is convex on En, then Ko = A and A is convex by Proposition
3.5 with c = O. Conversely, assume that A is convex. Given XI' Xl E En and
t E [0, 1], let YI be a point of A nearest x I, Y1 a point of A nearest Xl' and
X = tX I + (1 - t)x l , y = ty I + (l - t)y 1 .
(Actually, one can say "the nearest point" rather than" a nearest point"
since the set A is convex. This fact is not needed here.) Since A is convex,
YEA. By definition of f, f(x) ~ IX - YI. Then
f(x) ~ It(x i - YI) + (l - t)(x l - Yl)1,
f(x) ~ tix i - YII + (1 - t)lx z - Yll = tf(x l ) + (1 - t)f(xz)·
Hence f is a convex function on En.
In particular, let A consist of a single point Xo. Then f(x) = Ix - Xo I,
and this function is convex on En
Concave functions
The definition of concave function is obtained by reversing the inequality
sign in (3.21a): f is concave on K if, for every xl> Xz E K and t E [0, 1],
(3.21b)
If strict inequality holds whenever XI i= Xz and 0 < t < 1, then f is strictly
concave on K.
There are propositions about concave functions corresponding to Propo-
sitions 3.4 and 3.5 for convex functions. In them K + must be replaced by
K- = {(x, z): x E K, z ~ f(x)},
and Kc by
K C = {xEK:f(x);:::: c}.
A function f is concave on K if and only if - f is convex on K. By using this
fact, or by repeating the proofs, it is easy to prove these propositions about
concave functions.
Many useful inequalities can be obtained from (3.21a) or (3.21 b) by judi-
ciously choosing the function f and the number t.
109
3 Differentiation of real valued functions
PROOF. Let Xo be any point of K, and d the distance from Xo to the boundary
of K (d = + 00 if K = En). Let C be an n-cube with center Xo and side length
2c5, where n'/2c5 < d. Let V denote the set of vertices of C (see Problem to,
Section 1.5). V is a finite set. Let
Xo +u
Xo - u
---2«5 ,I
Figure 3.8
Let x be any point such that 0 < Ix - Xo I < c5, and define Xo + u, Xo - u
on the line through Xo and x as in Figure 3.8. Let us write x as a convex com-
bination of Xo + u and xo, and Xo as a convex combination of x and Xo - u.
If t = c5 - 1 Ix - Xo I, then
x = t(x o + u) + (1 - t)xo,
1 t
Xo = -1- x + -1- (xo - u).
+t +t
Since J is convex,
J(x) ~ tf(xo + u) + (1 - t)J(x o) ~ tM + (1 - t)J(x o),
1 t J(x) + tM
J(xo) ~ -l-J(x) + -l-J(x o - u) ~ 1 .
+t +t +t
The inequalities give
110
3.6 Convex and concave functions
or
(3.23)
The estimate (3.23) shows that 1 is continuous at Xo. [Note: This proof was
suggested by F. 1. Almgren.] 0
EXAMPLE °
2. Let K = [0, 1] and I(x) = x if < x ::; 1,/(0) = 1. Then 1 is
convex on K but is discontinuous at the left endpoint 0.
PROOF. Let 1 be convex on K, and let xo, x be any two points of K. Let
h = x- Xo and t E (0, 1). By definition of convex function,
III
3 Differentiation of real valued functions
°
But this is just the inequality (3.21a) in the definition of convex function. We
assumed that t E (0, 1), but if t = or 1, (3.2la) trivially holds. Therefore
f is convex on K. 0
Therefore
tdf(x o) • h < t[f(x o + h) - f(x o)].
112
3.6 Convex and concave functions
Let us next prove a theorem that provides a convenient test for concavity
or convexity of a function of class e(2). We recall the function Q defined by
(3.18).
By Proposition 3.6a,fis convex on K. On the other hand, ifit is not true that
Q(x, ) ~ 0 for every x E K, then Q(x o, ho) < 0 for some Xo E K and ho -=I- O.
Since f is of class e(2l, Q( ,ho) is continuous on K. Hence there exists b > 0
such that Q(y, ho) < 0 for every y in the b-neighborhood of xo. Let h = cho,
where c > 0 is small enough that Ih I < b, and let x = Xo + h. Since
Q(x o + sh, ) is quadratic,
114
3.6 Convex and concave functions
EXAMPLE 5. Let f(x) = ¢[g(x)J, where: (a) 9 is convex and of class e(2)
on K; (b) ¢ is of class e(2) on an interval I such that g(K) c I, with
¢'(u) :2: 0, ¢"(u) :2: 0 for all u E I. By the formula in one-variable calculus for
derivatives of composites,
};(x) = ¢'[g(X)Jgi(X).
By using this theorem again and the product rule,
};ix) = ¢'[g(x)Jgiix) + ¢"[g(x)Jgi(x)gix).
Writing for short gi for gi(X), and so on,
n n
Q(x, h) = ¢'[g(x)J L 9;jh ih j + ¢"[g(x)J L gigjhih j.
i,j=1 i,j=1
The first term on the right-hand side is nonnegative since ¢'[g(x)J :2: 0
and 9 is convex on K. The second term is also nonnegative since ¢"[g(x)J :2: 0
and
EXAMPLE 6. Let f(x, y) = i(x 3 + y3) + xy. Then fll = X, f22 = y, fll f22
- fi2 = xy - 1. Hence f is strictly convex on the part of the first quadrant
above the hyperbola xy = 1, and strictly concave on the part of the third
quadrant below this hyperbola.
115
3 Differentiation of real valued functions
Theorem 3.7. Let f be differentiable and convex on an open convex set K and
Xo E K a critical point. Then f has an absolute minimum at xo.
PROOF. Since df(x o) = 0, f(x) 2': f(x o) for every x E K by Proposition 3.6a.
o
Similarly, any differentiable concave function has an absolute maximum at
any critical point.
°
of f. Since f(x) = a' X = alx l + ... + anx n, the partial derivatives are
/;(x) = ai. Iff has a critical point, then a = 0 and f(x) = for every x.
Let us assume that a =f. 0 and consider the problem of extremum on a
convex polytope K (Section 1.5). If K is contained in {x: Xi 2': 0, i = 1, ... , n},
this is a problem in linear programming and has various interesting applica-
tions [9, 13].
For simplicity let us assume that K is compact. The extrema of f must
occur on the boundary fr K. Let us show that they can be found by consider-
ing only certain points of fr K, called extreme points.
Proposition 3.8. Let K be compact and convex. Then every point of K is a con-
vex combination of extreme points of K.
PROOF. Let us proceed by induction on the dimension n. If n = 1, then
K is an interval or a single point. Suppose that the proposition is true in
dimension n - 1. Let Xo E K. If Xo is a boundary point, then by Problem 7,
Section 2.4, K has a supporting hyperplane P containing xo. By an isometry
of En (see Section 4.2), we may arrange that Xo = 0 and the equation of P is
116
3.6 Convex and concave functions
But f(x) = C, and since each ti > 0 we must have f(x) = C for j = 1, ... , m.
Thus Xl' •.• , Xm E K I . Conversely, if f(x j ) = C for eachj and X is a convex
combination of Xl, ... , X m, then f(x) = C. 0
PROBLEMS
117
3 Differentiation of real valued functions
4. Using Problem 3, find the largest a such that f is convex on x 2 + y2 :::; a2.
(a) f(x, y) = 10g(1 + x 2 + i). (b) f(x, y) = sin(x 2 + y2).
5. Using Example 5, show that each of the following functions is convex on En:
(a) f(x) = IxI P, P ~ 1. (b) f(x) = (1 + X' x)x·x.
(c) f(x) = (1 + IxI2)P/2, p ~ 1. [Hint: First consider p = 1.]
6. Let f(x) = xtjJ(x) and g(x) = tjJ(1/x), where tjJ has a second derivative tjJ"(x) for
every x > 0. Show that f is convex on (0, (0) if and only if g is convex on (0, (0).
7. Let x > 0, y > 0, °: :;
t :::; 1. Show that tx
increasing, concave function.]
+ (1 - t)y ~ x'yl-'. [Hint: Log is an
f(J/Xi) ~ J/f(X)
for every XI"'" Xm E K and scalars t 1, ... ,tm such that each ti ~ and
t 1 + ... + t m = 1. [Note: For convex functions the sense of the inequality is
°
reversed.]
9. (a) Generalizing Problem 7, show that if Xl>"" Xm are positive numbers,
for j = 1, ... , m, and t l + ... + tm. = 1, then
°: :; ti
(b) Prove that the geometric mean is no more than the arithmetic mean, namely.
Xl + ... + Xm ( )l/m
~ X\"'X m •
m
10. Show that if f and g are convex on K, then f + g is convex on K.
11. (a) Let f and g be convex on K, and let h(x) = max[f(x), g(x)] for every X E K.
Show that h is convex on K. [Hint: Use Proposition 3.4.]
(b) Illustrate for the case f(x) = Ix-II, g(x) = x/2.
12. Let f be both convex and concave on En. Show that there exist a and b such that
f(x) = a . x + b for every x E En.
13. Let f be continuous on K, and assume that f(!<x I + x 2)) :::; !f(XI) + !f(x 2) for
every Xl' x 2 E K. Show that f is convex on K. [Hint: First show (3.21a) when
t = j/2k where j = 0, 1, ... , 2k and k is a positive integer.]
14. Let K be closed and convex, and P a supporting hyperplane for K. Show that any
extreme point of K ('\ P is an extreme point of K.
15. Let K be a closed convex polytope (not necessarily compact) and f be a linear
function such that f(x) is bounded above on K. Show that f has an absolute
maximum on K.
118
4
Vector-valued functions
of several variables
119
4 Vector-valued functions of several variables
We also recall from Section A.1 the concept of vector subspace. To verify that
a set PeEn is a vector subspace of En, one must show that x, YEP implies
x + YEP and that x E P implies cx E P for any real c. A vector subspace P
has a dimension p, and 0 ~ p ~ n. If p = 0 then P = {O}. If p = 1, then P is
a line containing 0; if p = 2, P is a plane containing 0, and so on.
If L is a linear transformation, then the set L(E') is a vector subspace of
En. To prove this, let x, YE L(E'). Then x = L(s), Y = L(t) for some s, tEE'.
But x + Y = L(s) + L(t) = L(s + t). Hence x + YE L(E'). Similarly, if
x E L(E') then cx E L(E') for every scalar c. The dimension p of the vector
space L(E') is called the rank of L.
The kernel of Lis {t: L(t) = O}. It is a vector subspace of E' (Problem 2).
The dimension v of the kernel is called the nullity ofL. The rank and nullity
are related by [12, p. 66J
(4.1) p+v=r.
The matrix of L
Let us denote the standard basis vectors for E' by E1, ... , E, and those for
En by et> ... ,en • Thus vectors t = (t 1, ... ,t') and x = (x1, ... ,xn) can be
written, respectively, as
, n
t = L
j= 1
tjE j , X = L xiei ·
i= 1
With these bases is associated a matrix (C}) for L with n rows and r columns,
in the following way. Let Vj = L(E j ), j = 1, ... , r. The vectors VI"'" V, are
the columns of the matrix. Let C} be the element in the ith row of column j.
Let us show that, if x = L(t), then the components satisfy
,
(4.2) x -
i _ 'L.
" Cjt
i j,I
' -- 1 , ... , n.
j= 1
,
(4.4) L(t) = L tjv j'
j=l
120
4.1 Linear transformations
i = 1, ... , n.
By (4.2), the components of Wi are the entries C i1 , •.• , c~ of the ith row of the
matrix (c~). For that reason wI, ... , wn are called the row covectors of the
matrix (c~).
v1 V2 V3 V,
w1 c: d d c,1
w2 ci d c,2
w3 d c,3
wn c~ ... . .. . .. ,
en
By (4.4) the column vectors v1, ... , V, span L(E'). The rank p equals the
largest number of linearly independent column vectors of the matrix. Since
row rank equals column rank [12, p. 105J, P is also the largest number of
linearly independent row covectors of the matrix.
Composition
Let L be linear from e into En, and M linear from En into £P. The composite
MoL is linear. Its matrix is the product of the matrices of M and L
(Problem 4).
The case r = n
Let 1 denote the identity linear transformation, I(t) = t for every tEe. Its
matrix is (b~), which has 1 for each element of the principal diagonal and 0
elsewhere. L is nonsingular if it has rank p = n, and singular if p < n. A non-
singular linear transformation L has an inverse L - 1, which is also a linear
transformation with
L-1oL=LoL- 1 =1.
121
4 Vector-valued functions of several variables
det L = det(c)).
Among the properties of determinants we recall:
By (4.1), L is singular if and only if v > O. But v > 0 means that L(t) = 0 for
some t =1= O. Therefore, from (4.2) the system of homogeneous linear equations
n
(4.5) 0= L C)ti, i = 1, .. . ,n
i= 1
The linear transformation L takes the standard basis vectors £1' £2 into
V1 = 2e 1 + 3e 2 , V2 = e 1 - e 2 • Since det L = -5 =1= 0, Lis nonsingular (see
Figure 4.1).
'<
Figure 4.1
122
4.1 Linear transformations
( 2 -5)
-1 O·
Since det M = - 5 #- 0, M is also nonsingular. The composite is found by
(M 0 L)(s, t) = M(2s + t,3s - t) = [2(2s + t) - 5(3s - t)]E I - (2s + t)E2'
(M 0 L)(s, t) = (-lIs + 7t)E I - (2s + t)E 2 •
The matrix of MoL is
(-11 7).
-2 -1
As expected, det MoL = 25 = det M det L.
EXAMPLE 2. Let n = r = 3, and let L be the linear transformation that takes
the standard basis vectors t l , £2, £3 respectively into
VI = e l + 2e 2 - e3 ,
The matrix is
-1
1 -1)
4 ,
o -1
which has VI' v2, V3 as column vectors. The determinant is 0, and thereforeL is
singular. In fact, V3is a linear combination ofv I and V2, namely, v3 = VI + 2v 2.
Since VI and v2 are linearly independent, the rank ofL is 2. L(E 3 ) is the plane
containing 0, vI, V2' By (4.1) the kernel has dimension 1. It is found by solving
the system (4.5) of homogeneous linear equations. One solution is tl =
£1 + 2£2 - £3' The kernel consists of all scalar multiples of t l .
123
4 Vector-valued functions of several variables
as follows. We have
r
Wi = L C~&j, i = 1, ... , n,
j= I
where {& I, ... , &r} is the standard basis for (E')*. If we write b = L *(a), then
the components of b satisfy
n
(4.7) hj = LaiC), j = 1, ... , r.
i= I
Ifone regards the co vector a as a row vector, then (4.7) states that bis obtained
by multiplying a on the right by the matrix (c~).
Note that Formulas (4.6) and (4.7) are statements about covectors corre-
sponding to Formulas (4.4) and (4.2) about vectors. We say that (4.6) is dual
to (4.4), and (4.7) dual to (4.2). Further aspects of the duality between Land
L* are mentioned in Problem 5.
PROBLEMS
124
4.2 Affine transformations
Isometries of En
Let g be a transformation from en into en. If g preserves the distance between
each pair of points, then g is called an isometry.
which says that g preserves the euclidean inner product. Let Vj = g(E),
j = 1, ... , n. Then IVjl = 1 and from (4.9)
for each i,j = 1, ... , n where, as in Section 1.2, bij is Kronecker's delta.
Hence VI"'" Vn form an orthonormal basis for P. Let us show that g is a
linear transformation. For each s, t, we have from (4.9)
125
4 Vector-valued functions of several variables
and hence
[g(s + t) - g(s) - g(t)] . Vj = 0,
for eachj = 1, ... n. The vector g(s + t) - g(s) - g(t) has component 0 with
respect to each basis vector Vj' Hence
126
4.2 Affine transformations
IfL is linear from E" into E", let V denote the linear transformation whose
matrix (c{) is obtained by exchanging rows and columns of the matrix (c~) of
L. We call V the adjoint of L. It is characterized by the formula
(4.10) Y. L(t) = V(y) . t
for every y, tEE" (see Problem 5).
Let us apply (4.10) with y = L(s). We get
L(s) . L(t) = (V 0 L)(s) . t.
From this equation, L is orthogonal if and only if s· t = (V L)(s)' t for 0
(. J
(tl, ... ,t") into S(t) = (tl, ... ,t"-I, -t"). S is a reflection of E" about the
hyperplane t" = O. Its matrix is
Two such reflections take each t into itself; that is, So S = I. Hence S =
S-I = st. If M is any orthogonal transformation with det M = -1, then
L = S 0 M is a rotation of E" about 0 and
M=S-loL=SoL.
Thus any orthogonal transformation is either a rotation or the composite of
S and a rotation.
EXAMPLE 3. Let n = 2, and L be a rotation of the plane E2 about (0, 0). Since
I vII = 1, VI = (cos 8)e 1 + (sin 8)e 2 for some 8 E [0, 2n). Since L is a rotation,
V 2 = (- sin 8)e 1 + (cos 8)e 2. The matrix is
(
COS 8
sin 8
-sin
cos 8
8).
The angle of rotation is 8.
127
4 Vector-valued functions of several variables
PROBLEMS
1. Let n = r = 2.
(a) Describe geometrically the linear transformation L with matrix G~).
(b) Find SoL and LoS, where S is the same as in Example 2. Show that both are
rotations of E2 about (0, 0).
3. Let Land M be rotations of En about O. Show that L - I and MoL are also rotations.
4. (a) Show that the composite of two affine transformations is also affine.
(b) Which affine transformations are univalent?
5. Prove (4.10).
For formulas for the differentials of products, see Problem 12, Section 4.4.
In order to find the matrix associated with the differential Dg(to) of a
linear transformation g, let us first prove the following proposition. For
each t E~,
g(t) = gl(t)e l + .. , + g"(t)e".
The real valued functions gl, ... , g" are the components of g. Their partial
derivatives are denoted by g~ or ogijotj.
The partial derivatives of g are defined just as for real valued functions:
() l' g(to
gJ,t o = 1m
+ Sl:) - g(to)
, j = 1, ... , r.
8-0 S
129
4 Vector-valued functions of several variables
We see later that the Jacobian often plays the same role in the calculus
offunctions of several variables as the derivative does in the case r = n = 1.
In particular, this is so in the theorems about inverses (Section 4.5) and trans-
forming mUltiple integrals (Section 5.8). If the Jacobian is 0 at a point to,
then Dg(to) is singular. This suggests some kind of irregularity in the behavior
of g near to. In order to exclude such irregularities we repeatedly make the
assumption in later sections that the Jacobian is not O.
n
Transformation g = (g 1, ••• ,gn) = L giei
i= 1
Differential at t Dg(t)
Matrix of Dg(t) (g;{t)), i,j = 1, ... , n
Column vectors git), j = 1, ... , r
Row covectors dgi(t), i = 1, ... , n
Jacobian (r = n) Jg(t) = det(g~{t))
. ogi n. . r..
g~{t) = oti' gJ{t) = i~/~{t)e;, dg'(t) = j~l g~{t)f/
EXAMPLE I. If n = 1, then Dg(t) has the single row co vector dg(t). We may
identify Dg(t) with dg(t). If dg(t) =F 0, the rank of Dg(t) is 1; otherwise it is O.
x = S2 + t2
Y = 2st
Figure 4.2
This set is the vertical line segment shown in Figure 4.2. By letting a take
all possible nonnegative values, one gets a collection of line segments cover-
ing Q. If a = 0, the line segment degenerates to the point (0,0). This shows
that Q = g(d).
In Example 3, gl(S, t) = S2 + t 2, g2(S, t) = 2st. The partial derivatives of
gl appear in the first row of the matrix of Dg(s, t), and those of g2 in the
second row. This matrix is therefore
( 2S 2t).
2t 2s
Jg(s, t) = det ( 2S
2t
2t)
2s = 4(S2 - t 2 ).
°
Note that Jg(s, t) =F if S2 =F t 2 • In Section 4.5 we show that g has a local
inverse near any such point (s, t). The points (s, is) where S2 = t 2 map under
g onto the boundary half-lines y = ix, x ~ 0, on the right-hand side of
Figure 4.2. The transformation g has no local inverse near (s, is).
Let us now suppose that the domain d of g is an open set.
Definition. If the components gl, ... ,gn are of class c<q), q 2 0, then g is a
transformation of class c<q). Similarly, if gl, ... ,gn are of class c<q) on
Bed, then g is of class C(q) on B.
131
4 Vector-valued functions of several variables
Proposition 4.4. Let g be differentiable at to. Then given e > 0 there exists a
neighborhood no of to sueh that no c ~ and
(4.15) Ig(t) - g(t o) I ~ (II Dg(to)ll + e)lt - tol
for every t E no.
PROOF. LetL = Dg(to)andsetg(t) = g(t) - L(t). SinceDL(to) = L,Dg(to) = 0
(the zero linear transformation). By (4.11), in which g is replaced by g, there
is a neighborhood no of to such that
(*)
for every t E no. But
get) - g(to) = [L(t) - L(to)] + [get) - g(to)]·
From (*), (4.14), and the triangle inequality we get (4.15). o
If g is of class C(1), there is a stronger version of Proposition 4.4.
132
4.3 Differentiable transformations
Proposition 4.5. Let g be of class C(1) and to E~. Then given t: > 0 there
exists a neighborhood Q of to such that Q c ~ and
(4.16) Ig(s) - g(t) I ~ (II Dg(to)ll + t:)ls - tl
for every s, t E Q.
PROOF. Let g be as before. The row covectors d?/(t o) are all O. Since the partial
derivatives of g are continuous, given E > 0 there is a neighborhood Q
of to such that Idgi(u) I < t:/n for every u E Q and i = 1, ... , n. By Corollary 1,
Section 3.3, for every s, t E Q,
. . t:
Ig'(s) - g'(t) I ~ -Is - tl,
n
n
(**) Ig(s) - g(t) I ~ L Igi(s) - gi(t) I ~ t:ls - tl.
i= 1
From (**) and (4.14) we obtain (4.16) in the same way as before. 0
PROBLEMS
1. In Example 3, find:
(a) The image of any vertical line s = c.
(b) The inverse image of any line y = mx through the origin.
(c) The image of the circular disk bounded by C in Figure 4.2.
2. Letg(s, t) = Is - tie] + Is + tle 2,L\ = E2. Find g(E2) and answer questions (a), (b),
and (c) in Problem 1 for this transformation.
6. (a) Let g be affine, g(t) = L(t) + Xo for every t E g. Show that Dg(t) = L for every
tEE'.
(b) Let g be a differentiable transformation such that Dg is a constant function
and L\ is a connected open set. Show that g is the restriction to A of an affine
transformation.
133
4 Vector-valued functions of several variables
7. Show that:
(a) IILII > 0 unless L has rank O.
(b) IlcLl! = 1c111LII·
(c) IILI + L211 ~ IILIII + IIL211.
(d) 11M LII ~ IIMIIIILII.
0
(e) The vector space ;t'(E', E") of all linear L from E' into E" has dimension tlr.
8. Another norm for linear transformations, which we denote by III III, is defined as
follows:
(1)
and
(2)
II. Let g be of class C(I). The maximum rank possible for Dg(t) is min(r, n), Show that
{t: rank Dg(t) = miner, n)} is open.
4.4 Composition
We now derive a rule for differentiating the composite of two differentiable
transformations. As corollaries of the basic formula (4.17) we then obtain a
formula for lacobians and the chain rule for partial derivatives.
Let g be a transformation from an open set Ll c Er into an open set
DeE", and let f be a transformation from D into £P.
134
4.4 Composition
PROOF. Let L = Dg(t o), M = Df(xo). Let us first prove the theorem for two
special cases.
Case J. M = O. Let us show that
whenever Ix - xol < 1]. Let 3 = min{C- I I],3 0 }. Then we get by taking
x = g(to + k) and using
f[g(to + k)] = F(to + k), f[g(to)] = F(t o),
Since F(t) = M[g(t)] in the special case when f is linear, (**) implies in this
case that F is differentiable at to, with DF(t o) = MoL.
In the general case, let f = f - M. Then Df(x o) = 0, and Case 1 applies
to the transformation F = fog. Moreover, Case 2 applies to G = M g. 0
135
4 Vector-valued functions of several variables
This is called the chain rule for partial derivatives. We rewrite it in more
compact notation as:
Note that F 1, ... , F r are the components of the covector dF(t o), and
f1' ... ,f" the components of the covector df(x o). According to Formula
(4.7), the chain rule (4.18) states that dF(t o) = L*[df(x o)] when L = Dg(to).
Thus, the chain rule can be regarded as a particular case of the formula for
changing the components of a covector under the dual L * of the linear
transformation L.
Another suggestive form for the important formula (4.18) is obtained by
writing it with the other notation for partial derivatives:
of of ogl of og"
-j= -1- + ... + - -j j = 1, ... , r.
ot ox ot
j ox" ot '
Let us next consider the case when F and g are vector-valued functions
of one real variable (r = 1). Their derivatives are denoted by F'(to) and g'(to).
If we identify the p x 1 matrix of DF(t o) with F'(to), regarded as a column
vector, and the n x 1 matrix of Dg(t o) with g'(to), then we get from (4.17):
for every t E L\ and j = 1, ... , r. Since all of the functions h, g, g) are con-
tinuous, each partial derivative F j is continuous. Hence F is of class C(l).
If q ~ 2, then repeated application of the chain rule shows that F is of class
c<q) and gives formulas for calculating its partial derivatives of orders 1,
2, ... ,q.
In case p > 1 and f, g are of class C(q), the preceding discussion shows
that the components F1, ... ,FP of F are of class C(q), since FI = flog for
each I = 1, ... , p. Therefore F is of class C(q). We have proved:
136
4.4 Composition
EXAMPLE 2. Let F(x) = f[x, g(x)], wherefand 9 are of class e(2). In this case
gl(X) = X, g2(X) = g(x), and the formula in Example 1 becomes
The expression on the left-hand side is called the Laplacian off The partial
differential equationfll + f22 = 0 is called Laplace's equation. Its solutions
are called harmonic functions. The formula above expresses the Laplacian in
polar coordinates.
In this example
gl(r,O) = r cos e, g2(r, e) = r sin e.
Using the chain rule, we get
F II = cos e(fll cos e + fl2 sin e) + sin e(f21 cos e + f22 sin e),
F 22 = - r sin e[fll ( - r sin 0) + f12(r cos e)] + r cos e[f21 ( - r sin e)
+ f2z(r cos e)] - fIr cos 0 - f2r sin e.
137
4 Vector-valued functions of several variables
PROBLEMS
Assume that all functions which occur in these problems are of class e(2).
1. Let F(x, y) = f(x, xy). Find the mixed partial derivative F 12'
2. Let F(x, y) = f[x, y, g(x, y)]. Express the partial derivatives of F of orders 1 and 2
in terms of those of f and g.
(b) Use it to find (oF/os)(so, to) and (oF/ot)(so, to) in Problem 3(a).
138
4.4 Composition
5. Let f(x, y) = 4>(x - cy) + ljJ(x + cy), where c is a scalar. Show that f22 = c2fll'
6. Let
n=4 and
Let
Show that
[Note: The partial differential equation fnn = C2(f1l + ... + fn-I,n- tl is called
the wave equation in n variables. Problem 5 gives D'Alembert's solution for n = 2.
Solutions of the type in Problem 6 are called spherical waves.]
7. Suppose that f satisfies the partial differential equation f2 = fll + bf, where b
is a scalar. Let F(x, y) = exp(-by)f(x, y). Show that F2 = F II .
8. Let F = f L, where L is a linear transformation with matrix (ej). Show that the
0
j, I = I, ... , r.
(a) Show that if M and L are Lorentz, then MoL and L - I are also Lorentz.
[Hint: S2 = I.]
(b) Show that L is Lorentz if and only if S = L' SoL. 0
i= 1 i= 1
11. Show that if L is Lorentz, then F I I + ... + Fn-I,n-I - Fnn = fll + ... +
fn-I,n-I - fnn· In other words, the wave operator is invariant under Lorentz
transformations (c = 1).
12. Use the composite function theorem to establish the following rule for differentials
of inner products. Let q, and \jI be differentiable at to, both with values in En. Show
that q, . \jI is differentiable at to and that d(q, . \jI)(t o) = q,(to)' D\jI(t o) + \jI(to)' Dq,(t o).
[Note: Ifv E En and L is a linear transformation from E' into En, then v . L denotes
the real valued linear function such that (v . L)(k) = v . L(k) for all k E E'.] [Hint:
Take g = (q" \jI) with values in E2n and f(x, y) = x . y.]
139
4 Vector-valued functions of several variables
(4.20) 1
g - 1'(x ) = g'(t)' 'f x = g(t ).
1
In two or more dimensions the Jacobian Jg(t) takes the place of the deriva-
tive g'(t). However, the situation is by no means as simple as before. First of
differentiability. Second, and more important, the fact that Jg(t) ¥- does
not imply that g has an inverse. (see Example 2 below). However, the non-
°
all, we have to assume that g is at least of class C(1), a stronger condition than
140
4.5 The inverse function theorem
--
g
Figure 4.3
Corollary. Let g be of class C(1), and suppose that Jg(t) i= 0 for all tEA.
Then the image g(B) of any open set B c A is an open set.
(*) s=
Jx+Y+2 Jx"=Y '
141
4 Vector-valued functions of several variables
The restriction glL\o has all of the properties (1)-(4) of the inverse function
theorem. Its inverse is f = (fl,j2) wherefl(x,y) andfZ(x,y) are on the
right side of equations (*). The open set g(L\o) is the interior of the quadrant
Q in Figure 4.2.
- g
• (s, t - 2n)
I
Figure 4.4
142
4.5 The inverse function theorem
Regular transformations
By adding to the hypotheses of the corollary above the assumption that g
is univalent we obtain a property that we call regularity.
143
4 Vector-valued functions of several variables
L cj)[ml(t)
00
Lemma 2. Let g be of class C(l),from ~ into En, with Jg(t 1 ) =I- 0, tl E~. Then
there exist a neighborhood n oft 1 with n c ~,a neighborhood U afx 1 =
get!), and afunction F defined on U, such that:
(i) The restriction gin is univalent.
(ii) U c g(n) and F(U) c n.
(iii) g[F(x)J = xfor all x E U.
(iv) F is differentiable at Xl' and
DF(x 1) = [Dg(t 1)J-l.
144
4.5 The inverse function theorem
145
4 Vector-valued functions of several variables
where ij = %(11 L -111)-1. Given any e > 0 choose c such that 0 < c .:::; t
and
where t = f(x). Since each g~ is continuous and fis continuous, the composite
g~0 f is continuous. Let (y~) be a nonsingular matrix. Cramer's rule [12J
146
4.6 The implicit function theorem
C<l).Ifg is of class e(2), then each g~ is of class e(l). By Corollary 3, Section 4.4,
g~ f is of class e(1). Hence each f~ is of class e(1l, and f of class e(2). By
0
PROBLEMS
1. Determine whether Jg(t) # 0 for all t E~. Find g(~). If g is univalent, find g-I
explicitly.
(a) g(t) = t + Xo (a translation), ~ = En.
(b) g(s, t) = (s + 2t)el + (s - t)e2, ~ = E2.
(c) g(s, t) = (S2 - S - 2)e , + 3te 2, ~ = E2.
(d) g(s, t) = (S2 - t 2)el + ste2, ~ = E2 - {(O,O)}.
(e) g(s, t) = (log St)el + 1/(s2 + t 2)e2' ~ = {(s, t): 0 < t < s}.
2. Let g(t) = t 4 + 2t 2, ~ = (0, 00). Find g- I.
3. Letr = n = 3,andg(s, t, u) = (u cos st)el + (u sin st)e2 + (s + u)r J . Theng(1:1 + 1: 3)
= e, + 2e3' Let f be a local inverse of g such that f(e , + 2e 3) = 1:, + 1:3' Find
Df(e, + 2e 3) using (4) of the inverse function theorem.
4. In Example 2, what are the images of horizontal straight lines? Show that g is a
conformal transformation (Problem 9, Section 4.3), and hence that the images of
vertical and horizontal straight lines intersect at right angles. Illustrate with a sketch.
5. Let g(s, t) = (exp s cos t)e, + (exp s sin t)e 2 and ~ = E2.
(a) Show that Jg(s, t) # 0 for all (s, t), but g is not univalent.
(b) Let ,1 = {(s, t): 0 < t < 2n}. Show that the restriction of g to ,1 is univalent,
and find its inverse.
(c) Find g(E2).
(d) Show that g is conformal.
6. Let ~ be an open convex set and g a differentiable transformation such that
I? j=' g~{t)hihj > 0 for
every t E ~ and h # O. Show that g is univalent. [Hint:
Suppose that g(td = g(t 2). Let h = t2 - t" f(t) = [g(t) - g(t!)] . h, and apply the
mean value theorem to f.] This result is due to H. Nikaid6.
7. Suppose that g is of class C(1), with Jg(t) # 0, for all t E ~. Given x I/o g(~), let l{I(t) =
Ix - g(tW. Show that dl{l(t) # 0 for all t E ~.
8. Let g be of class cn. Suppose that there is a number c > 0 such that Ig(s) - g(t) I :2:
cis - tl for all s, tEEn. Show that:
(a) g is univalent.
(b) Jg(t) # 0 for all tEEn.
(c) g(En) = En. [Hint: Use Problem 7, and show that l{I(t) has a minimum.]
147
4 Vector-valued functions of several variables
derivatives <1>1 (xo, Yo), <l>2(XO, Yo) is not O. This is a special case ofthe implicit
function theorem below.
More generally, suppose that 1 :::; m < n. Consider a transformation
(4.24)
from an open set D c En into Em. The implicit function theorem ensures that
the equation eIl(x) = 0 determines m components of x as functions of the
remaining n - m components, in a neighborhood of any xo such that
eIl(xo) = 0 and D<I>(x o) has maximum rank m.
The matrix of DeIl(xo) is m x n, with elements the partial derivatives
<I>~{xo), i = 1, ... , m,j = 1, ... , n.
If DeIl(xo) has maximum rank m, then some set of m columns of its matrix
is linearly independent. For the present, let us assume that the last m columns
are linearly independent. Let r = n - m, and let
(4.25)
denote the determinant of the sub matrix with these columns. We also let
denote the vectors obtained by taking only the first r components of x and
xo. We seek to write the remaining components as x,+ I = ¢1(X) for I =
x
1, ... , m, and for in some open set R containing xo.
When this is possible
we write for brevity q, = (¢l, ... , ¢m) and x = (x, q,(x)).
Implicit function theorem. Let ell be of class C(q) from an open set D c En
into Em, where q ~ 1 and 1 :::; m < n. Let Xo E D be such that eIl(x o) = 0 and
jell(x o) #- O. Then there exist a neighborhood U ofx o , an open set R c E'
containing x o , and q, = (¢ 1, ... , ¢m) of class c(q) on R such that:
PROOF. Since ell is at least of class C(1), the Jacobian jell is a continuous
function. By assumption it is not zero at X o , and therefore is not zero for x
in some neighborhood U 0 of Xo'
Let us consider the transformation f, with domain U 0 and values in En,
148
4.6 The implicit function theorem
1 o 0; 0 o
o I
I
I
I
o I 0 o
----------1----------
I
I
I
I
, I
--t---'-~"-4-J'--
xo't' x', ... , x'
'--y-----'
R
X·+ 1 = cj/(t)
Figure 4.5
Writing (x, 0) for (x 1, ... , x', 0, ... , 0), let (see Figure 4.5)
R = {x: (x, 0) E f(Un.
Since f(U) is an open set, R is open. For every X E R, let
¢'(X) = g'+ '(X, 0), 1= 1, ... , m.
Then x E U and ~~x) = 0 if and only if x E Rand f(x) = (x, 0). Since fl U and
g are inverses, f(x} = (x, 0) if and only if x = g(x, 0). 0
149
4 Vector-valued functions of several variables
°
solving the equation <I>(x, y) = for y. We get ¢(x) = ±(x 2 - 1)1/2, the
°
sign depending on whether Yo > or Yo < 0. The interval R is the projection
°°
of H n V onto the x-axis.
If Yo = 0, Xo = ± 1, then <l>2(XO, Yo) = and the reasoning above fails.
However, at these points, <l>l(XO, Yo) # and (xo, Yo) has a neighborhood
V such that H n V = {(ljJ(y), y): y E R}, where ljJ(y) = ±(l + 1)1/2 and R
the projection of H n V onto the y axis.
(**)
In the formulas (**) the partial derivatives of <I> are evaluated at (x, y, ¢(x, y».
To calculate the second-order partial derivatives ¢11' ¢12, ¢22' the chain
rule is applied again. For instance, taking the partial derivative with respect
to the second variable in the first equations (**), we get
<1>12 + <1>13 ¢2 + [<1>32 + <1>33 ¢2]¢1 + <1>3 ¢12 = 0.
Substituting the expressions for CPt> ¢2 obtained above and solving for ¢12,
we get
A.. = _ (<1>3)2<1>12 - <1>2<1>3<1>13 - <1>1<1>3<1>32 + <1>1<1>2<1>33
'1-'12 (<1>3)3
Let m = 2, n = 3, r = n - m = 1. Writing <I> = (<I>, '1') rather than
(<1>1, <1>2), and cp, ljJ rather than cpl, ¢2, we have
<I>[x, cp(x), I/I(x)] = 0,
'I'[x, cp(x), I/I{x)] = 0,
°
and <1>2'1'3 - <1>3'1'2 # for every x E R. The partial derivatives in question
are evaluated at (x, cp(x), I/I(x». By the chain rule
<I> 1 + <1>2 cp' + <1>31/1' = 0,
'1'1 + 'l'2¢' + '1'31/1' = 0,
150
4.6 The implicit function theorem
cp' = <1>3 'P 1 - <1>1 'P 3, 1/1' = <1>1 'P 2 - <1>2 'P 1.
<1>2 'P 3 - <1>3 'P 2 <1>2 'P 3 - <1>3 'P 2
The second derivatives cp", 1/1" can be found by another application of the
chain rule.
For convenience we assumed in (4.25) that the last m columns of the
m x n matrix of partial derivatives (<I>l(x o)) were linearly independent. More
generally, one need merely suppose that some set of m columns is linearly
independent, in other words, that the linear transformation D«J)(x o) has
maximum rank m. Let us suppose that columnsjl,j2' .,. ,jm form a linearly
independent set, where we may suppose that j 1 < h < ... < jm. Let
i b ... , ir be those integers between 1 and n not included among j 1, ... ,jm'
with i 1 < ... < i r . For brevity let us write A for the r-tuple of integers
(il"'" ir) and Xl for the r-tuple (XiI, ... , xir).
The implicit function theorem now states, roughly speaking, that locally
the equation «J)(x) = 0 determines XiI, ... , xim as functions of Xl. More pre-
cisely, U, R, and CPl, ... , cpm exist as before such that
and
{x E U: «J)(X) = O} = {X: Xl E R, xi! = CPl(X l ) for I = 1, ... , m}.
0(<1>1, <1>2)
0(x\x 4 ) #-0 atx o .
EXAMPLE 3. Let
<I>(x, y, Z) = X2 + i + Z2 - 2xz - 4.
Then
<l>1(X, y, Z) = 2x - 2z,
<l>2(X, y, Z) = 2y,
<l>3(X, y, Z) = 2z - 2x.
151
4 Vector-valued functions of several variables
PROBLEMS
152
4.7 Manifolds
4.7 Manifolds
The word manifold is used in mathematics to describe a topological space that
locally is "like" euclidean E', for some r called the dimension of the manifold.
For instance, a circle is locally like EI. Such geometric figures in E3 as ellip-
soids, cylinders, and tori are locally like E2. A cone is not locally like E2 near
its vertex.
We approach the idea of manifold from a rather concrete viewpoint.
For us, a manifold M is a subset of some euclidean En that can locally be
described by an equation <I>(x) = 0, where D<I>(x) must have maximum rank.
Another definition of manifold can be given abstractly in terms of coordinate
systems. It has the advantage that one need not presuppose that M is a subset
of some euclidean space. This will be discussed in Chapter 8.
Mn U = {x E U: <I>(x) = O}.
°
Let us indicate how an r-manifold M is locally like E'. First assume that
jq,(x o) #- with j<l>(x) as in (4.25), and define f as in the proof of the implicit
function theorem. The neighborhood of Xo chosen in Section 4.6 need not
coincide with the neighborhood U in the definition of manifold. In this section
let us denote the former neighborhood by U I rather than U, and let us denote
the set R in Section 4.6 by R I . We may suppose that U leU. Now flU I is a
regular transformation (Section 4.5), and
153
4 Vector-valued functions of several variables
M
f(V I )
- f
j(MnVtl
Figure 4.6
EXAMPLE 3. Let M = {(x, y): x 2 = y2}. This set consists of the two lines
y = ±x, and is not a I-manifold. Roughly speaking, M is not like EI near
the crossing point (0,0). More precisely, if M were a I-manifold, then by the
implicit function theorem the following would be true: Each (xo, Yo) E M has
a neighborhood U 1 such that either M nUl = {(x, 4>(x)): x E Rd or
M nUl = {(t/t(y), y): y E R 2 }, where R I , R2 are open. In the present example,
(0,0) has no such neighborhood.
154
4.7 Manifolds
every Xo E M. By Problem 11, Section 4.3, {x: D<I>(x) has rank m} is open.
Hence any Xo E M has a neighborhood U such that D<I>(x) has rank m for
every x E U, and M n U = {x E U: <I>(x) = O}.
Definition. When (4.26) holds, we say that M is the r-manifold determined by <1>.
°
metrically. For brevity let us set x, = Xo + tho Then h is a tangent vector if,
for some b > 0, there exists y, E M whenever < It I < b, such that
X, = Xo + Ih
Figure 4.7
If we set ",(t) = y, for 0 < It I < <5, and ",(0) = x o, then (*) states that ""(0) = h
(see Figure 4.7).
Let T(xo) denote the set of all tangent vectors at Xo. It is called the tangent
space to M at Xo. If r is the dimension of M, then it is plausible that the
tangent space is a vector space of dimension r. Let us show that this is true.
Let V and «I> be the same as in the definition of manifold.
Theorem 4.3. The tangent space T(xo) is the kernel of the linear transformation
D<Jl(xo)·
Since D«I>(x o) has rank m, the kernel T(xo) is a vector subspace of P with
dimension r = n - m.
PROOF OF THEOREM 4.3. Let L = D«I>(x o). We must show that h is a tangent
vector if and only if L(h) = O.
Let hE T(xo). Let '" be as in the definition of tangent vector. Then
<Jl["'(t)J = 0 for every t E ( - 15, <5). Calculating the derivative of <Jl by 0 '"
156
4.7 Manifolds
Figure 4.8
By definition of f,
N(h) = hi, i = 1•... , r,
N+r(h) = d<l>I(XO)' h, 1= 1, ...• m.
Since h is in the kernel of D«D(x o), N+r(h) = O. Thus A(h) = (ft, 0), and
h = A- 1(h, 0) = l\I'(O). D
Corollary 1. The gradients grad <l>l(X O)' ... , grad <l>m(xo) form a basis for
the space of normal vectors to Mat xo.
Tangent r-planes
The r-plane tangent to M at Xo is
{x: x = Xo + h, h E T(x o)}
(see Figure 4.9). The terms tangent line, tangent plane, and tangent hyperplane
are used when r = 1,2, and n - 1, respectively. The tangent r-plane is
{x: grad <l>1(XO)' (x - x o) = 0 for 1= 1, ... , m}.
157
4 Vector-valued functions of several variables
Xo +h
Figure 4.9
EXAMPLE 8. Let
M = {(x,y,z):x 2 + y2 + Z2 - 2xz - 4 = OJ.
According to Example 3, Section 4.6, M is the 2-manifold determined by the
function ct> of that example. Then
grad Cl>(x, y, z) = 2(x - Z)el + 2ye2 + 2(z - x)e3'
Let us find the spaces of tangent and normal vectors to M at (2, )3, 1).
gradCl>(2,)3, 1) = 2e 1 + 2,,/3e 2 - 2e 3.
This is a normal vector, and any other is a scalar multiple of it. The tangent
vectors h satisfy
the partial derivatives being evaluated at (xo, Yo, zo). The tangent space
T(xo, Yo, zo) consists of all scalar multiples of d.
158
4.7 Manifolds
(x, y, z) near (xo, Yo, zo), the tangent planes cannot coincide at points on the
intersection M.
This idea can be extended to the intersection of manifolds of other
dimensions.
*Intersections of manifolds
Let M be an r-manifold and N an s-manifold, with M n N nonempty. Let us
assume that r + s > n. Let TI(x o) denote the tangent space to M at a point
Xo E M n N, and T2 (x O) the tangent space to N at xo. There exist a neighbor-
hood U I of Xo and «I> = (<1>1, ... ,<I>n-r) such that D«I>(x) has rank n - r for
every x E U I and
M nUl = {x E U I: «I>(x) = O}.
In the same way, there exists a neighborhood U 2 ofx o and 'I' = ('PI, ... , 'P n - s)
PROBLEMS
In Problems 1-6 and 11, one can show that the set in question is a manifold
by verifying (4.26) for suitable «1>.
1. Let F(x, y) = exp(x 2 + 2y2 + 2). Find the level sets and determine which are
I-manifolds.
2. (a) Show that if c =I- 0, then the hyperboloid x 2 + l - 4Z2 = c is a 2-manifold.
Is the cone x 2 + l = 4Z2 a 2-manifold?
(b) Find the tangent plane at 2e l - e 2 + e 3 to the hyperboloid x 2 + l - 4Z2 = 1.
3. Let M = {(x,y,z):xy = 0,x 2 + l + Z2 = I, Z =I- ±I}. Show that M is a 1-
manifold. Sketch M.
4. Let f be of class C(1) on an open set A c E2. Let M = {(x, y,f(x, y)): (x, y) E A}.
(a) Show that M is a 2-manifold.
(b) Show that (Ux, y), f2(X, y), -I) is a normal vector to M at (x, y, f(x, y)).
(c) Show that the equation for a tangent plane agrees with the one in Section 3.3.
(d) State the corresponding results when f is a function of n variables.
159
4 Vector-valued functions of several variables
L cijxixb = 1.
i,j= 1
L c)e i + U -l)n'
i,j= 1
where (c~) is the matrix ofL. Let O(n) be the set of orthogonal transformations of En.
(a) Show that O(n) is a manifold of dimension !n(n - 1) and class Cleo). [Hints: It
suffices to show that I has a neighborhood U with the properties required in
the definition of manifold. Use Proposition 4.1. Here I is the identity trans-
formation.
(b) Show that L is a tangent vector to O(n) at I if and only ifL' = - L. (Such a linear
transformation is called skew symmetric.)
160
4.8 The multiplier rule
(c) Let SO(n) be the set of an rotations of En about O. Show that SO(n) is a relatively
open subset of O(n), and hence SO(n) is also a manifold of dimension in(n - 1).
Show that SO(n) is the largest connected subset of O(n) which contains I. [Hint:
Using induction on n, show that any L E SO(n) can be joined with I be a path
in SO(n).]
Xo = 20" Yo = - ;:;,-,
.;.0'
Zo = 20" a
±j2.
Therefore Xo = ±(j212)(e1 - e 2 + e 3), depending on which of the two
possible values for a is used. Since f is continuous and M is a compact set,f
has a maximum and a minimum value on M. One of the two critical points
obtained by the multiplier rule must give the maximum and the other the
minimum. Since
and
162
4.8 The multiplier rule
there is a multiplier 0' such that v1 is a critical point of F. Then grad f(x) =
2L(x) since the matrix (c}) is symmetric. Hence
0= grad F(v 1 ) = 2L(v 1) - 20'V1'
Hence L(v 1) = (TV 1, which shows that 0' is a characteristic value and v1 a
characteristic vector. Since f(x) = L(x)' x,
A1 = !(v 1) = L(v 1)' V1 = O'V 1 . v1,
and since V1 . v1 = 1,0' = A1'
We next let
M2 = {x:lxl = l,x'v 1 = A},
A2 = max{f(x): x E M 2}'
and V2 E M2 be a point such that f(v 2 ) = ).2' We have added another con-
straint. Hence M2 c M 1 and A2 :$ A1' Obviouslyv 2 . V1 = O. For k = 3, ... ,n,
let
Mk = {x:lxl = l,x'vi = 0 fori = 1, ... ,k -I},
(4.28) Ak = max{f(x): x E M k },
163
4 Vector-valued functions of several variables
Since Vi • Vj = 0 for i #- j, we get upon taking the inner product with Vj'
o= 2L(v k)· Vj + O"j+ 1, ifj < k,
0"1 = L(v k )· V k = f(v k ),
Theorem 4.4. The characteristic values of the symmetric matrix (c~) are
Ab ... , An· For each i = 1, ... , n, Vi is a characteristic vector corresponding
to Ai. If ~ 1, ... , ~n denote the components of x with respect to the ortho-
normal basis {VI' ... ' Vn}, then
(4.29) f(x) = Al(~1)2 + ... + An(~n)2
for every x E En.
PROOF. The first two statements have already been proved. To prove the
third, we have
n
X = L ~iV;,
i= 1
n n
°
PROOF. If each Ai is positive, then by (4.29), f(x) > whenever x
Conversely, iff(x) > for every x =1= 0, then Ai = f(v;) > for each i.
°° =1= o.
D
I
I
...... /
----+-------~/r.,~,------~---x
" /
B / ......
/
/ ,,
/
/
/
/
Figure 4.10
PROBLEMS
165
4 Vector-valued functions of several variables
7. Letf(x) = xIX Z ... xn and M = {x: Xl + ... + xn = 1, Xi> 0 for i = 1, ... , n}.
(a) Show that f(x) ::; n-n for every x E M, with equality if Xl = ... = xn = n- I.
[Hint: First show that f has an absolute maximum on M. Apply the multiplier
rule to log f, which has a maximum at the same point where f has one.]
(b) Using (a), prove that the geometric mean of n positive numbers is no more
than their arithmetic mean. See Problem 9(b), Section 3.6.
8. Let p > 1 and p' be the number such that p - I + (p') - I = 1. Let I x II be as in
Problem 4, Section 2.11, and for each co vector a let II a II = max {a . x : II x II = I}.
Show that
n )IIP'
Iiall = ( IlaY
[Note: For these norms the inequality la . x I ::; Ilallllxll is called Holder's inequality,
A related inequality for integrals is given in Section 5.13].
to. Let IILII be defined as in Section 4.3. Show that IILll z is the largest characteristic
value ofL' 0 L, [Hint: Use (4,10) with y = L(t)].
11. (Second derivative test for constrained relative maxima.) Let f and <I> be of class e(2),
and let Q(x, h) = D,
j= I Fiix)hihj where F is as in the multiplier rule. Show that:
(a) If f IM has a relative maximum at x o , then Q(x o , h) ;::: 0 for every hE T(xo).
(b) If Q(x o, h) > 0 for every hE T(x o), h ¥- 0, then f IM has a strict relative
maximum at xo' [Hints: See the proof of Theorem 3,4, Set hI = t-I(y, - x o),
and show that Iim,~o Q(x o, hI) = Q(xo, h),]
166
5
Integration
167
5 Integration
or
If A = En, we write simply Sf d"~. The symbols dv" and dv,,(x) are used
after an integral sign merely for con venience and for traditional reasons. They
will have no significance by themselves.
Ifn = 1, then we write, as is customary, SA f(x)dx instead of SA f(x)dVI(x),
and if A = [a, b], we write S~f(x)dx..
5.1 Intervals
What is the n-dimensional measure of a subset A of En? To start with,
let us consider the simplest possible case-where A is an n-dimensional
interval.
Xo;
I
I
I
_ _~_ _ _ _~________- L_ _ _ _ Xl
Figure 5.1
168
5.1 Intervals
The n-dimensional measure V(I) of I is the product of the lengths of the inter-
vals J l' ... , J n •
We next define the measure of a set that is a finite union of n-dimensional
intervals. For this purpose the idea of grid of hyperplanes is introduced.
Grids
For each i = 1, ... , n let us take a finite set of real numbers; let the elements
of these sets be denoted by x~, where
and mj + 1 is the number of elements of the ith set. Let P~ be the hyperplane
with equation Xi = x~, and let 0 be the union of all these hyperplanes.
Such a set 0 is called a grid of hyperplanes. A grid divides En into a finite
number of n-dimensional intervals, called intervals of 0, and a finite number
of unbounded sets. The latter could be called semiinfinite intervals of 0, but
we have no occasion to do so. The intervals of 0 have the form I =
J 1 x···xJ n where Jj=[x~i,x~i+l]and the integersjl, ... ,jn may be
chosen arbitrarily subject to 1 -5, jj -5, mj. There are mi'" mn intervals
of the grid, and for convenience we have taken them to be closed (see Figure
5.2).
X2
a
2
x h+l
X71
xiJ + 1 Xl
x},
I
Figure 5.2
Let us call a set Y a figure if Y is the union of certain intervals II' ... , I p
of some grid O. The measure of Y is
(5.1 ) V( Y) = V(I d + ... + V(I p).
There are many possible choices for O. Consequently, we must show that
V(Y) depends only on Yand not on the particular grid chosen. Let us call
0' a refinement of 0 if 0 cO'. It is easy to show that V(Y) is unchanged
if 0' is obtained by adding one hyperplane to 0, and hence by induction if
o is replaced by any refinement of it. Now let 0 and 0' be any two grids such
that Y is the union of intervals of 0 and also the union of intervals 0'. Then
o u 0' is a refinement of both. Consequently, V(Y) is the same whether
o or 0' is used.
169
5 Integration
This same reasoning shows that if Yand Z are figures, then Yand Z can
be written as unions of intervals of the same grid n. Therefore Y u Z is a
figure. Moreover,
(5.2) V(Yu Z) ~ V(Y) + V(Z).
PROBLEMS
5.2 Measure
We now define measure in the Lebesgue sense for a large class of subsets of
En. A set A in this class is called measurable, and the measure of A is denoted
by V(A). Some main properties of measurable sets and of measure are
stated in Theorems 5.1 through 5.3. In studying the present section, the
reader should first concentrate on understanding the definitions and the
statements of these main theorems. A careful study of the various lemmas
and propositions in the section might be postponed.
We begin by defining the measure of a bounded set A. This is done in two
stages. First, the measure of an open set G is defined by approximating G
from within by figures, and that of a compact set K by approximating K from
without by figures. In the second stage, A is approximated from within by
compact sets and from without by open sets. This two-stage approximation
process is an important feature of the Lebesgue theory of measure.
There is an older theory of measure due to Jordan. In this theory A is
approximated simultaneously from within and without by figures. The
Jordan theory is unsatisfactory for several reasons. Among them is the fact
170
5.2 Measure
I I
~
~
\ G
-
y -
V
V -
--
~ /
"- ./
I
Figure 5.3
that the class of sets to which it applies is too small. For instance, there are
compact sets to which the Jordan theory does not assign any measure.
Let G be an open set. If Y is a figure contained in G, then the measure
of G must be more than V( Y). It is defined to be the least upper bound of the
set of all such numbers V(Y) (see Figure 5.3).
Now let K be a compact set. If Z is any figure whose interior int Z contains
K, then V(Z) must exceed the measure of K.
Definition. The measure of a compact set K is V(K) = inf{ V(Z) : K c int Z}.
171
5 Integration
Lemma 1. Let K be a compact set and G an open set such that KeG. Then
there is a figure Y such that K c int Yand Y c G.
We are interested in those bounded sets A whose outer and inner measures
are equal.
Definition. A bounded set A is called measurable if its outer and inner measures
are equal. If A is measurable, then the number
V(A) = I:::(A) = V(A)
Let us show that bounded open sets and compact sets are measurable
and that the new definition of their measures agrees with the previous one.
Let H be a bounded open set. If G is open and H c G, then any figure Y
contained in H is also contained in G. By the definition of measure for open
sets, V(H) ~ V(G). When G = H, equality holds. From the definition of
outer measure, V(H) = V(H). Given t; > 0, there is a figure Y c H with
V(H) - t; < V(Y). Since Y is a compact set, V(Y) ~ I:::(H). Since this is true
for every t; > 0, V(H) ~ I:::(H). But !:::(H) ~ V(H), and therefore
!:::(H) = V(H) = V(H).
172
5.2 Measure
Let us next show that the union, intersection, and difference of two
bounded measurable sets is measurable (Proposition 5.2). For this a series of
preparatory lemmas is needed.
173
5 Integration
This shows that V(K) + V(L) is a lower bound for {V(W) : K u LeW},
and hence is no more than the greatest lower bound V(K u L). D
174
5.2 Measure
and the two sets on the right-hand side are measurable and do not intersect.
By Proposition 5.1a, A u B is measurable. 0
Lemma 3b. Let G I , G z , ... be open sets each of which has finite measure.
IfG = G I u Gz u ···,then
00
V(G):::; L V(Gd·
k=1
175
5 Integration
PROOF. Let Y c G. The figure Y is a compact set and G I' G2' ... form an
open covering of Y. Hence Y c G I U ... u Gm for some m. By Lemma 3a
and induction on m,
m
L V(A k) ~ !:::(A).
k=1
Since this is true for each m,
OC!
L V(Ad ~ [(A).
°
k= I
On the other hand, given £ > let Gk be an open set such that Ak c Gk and
V(G k ) < V(A k ) + £r\ k = 1,2, ... , and let G = G I U G2 U .... Then
A c G and therefore V(A) ~ V(G). By Lemma 3b,
ro 00
and L 2- k = 1. Since this is true for any s > 0, A is measurable and (5.9)
holds. D
If in Proposition 5.1 b the sets A I, A 2, ... are not disjoint, then the equality
(5.9) becomes an inequality. This is part of Theorem 5.2 below.
Unbounded sets
For a possibly unbounded set A the concepts of A being measurable and
measure of A are defined as follows. Let Ur = {x: Ix I < r}. A set A is called
measurable if A n Uris measurable for every r > 0. The measure of A is
V(A) = lim V(A n Ur ).
r-+ + oc'
176
5.2 Measure
Theorem 5.1
(a) Every open set G is measurable.
(b) If A is any measurable set, then its complement A C is measurable.
(c) Let AI' A z , ... be measurable sets. Then the union Al U A z U ...
and the intersection A I n A z n ... are measurable sets.
Theorem 5.2
(a) IfYis afigure, then its measure V(Y) is given by the elementary formula
(5.1).
(b) If G is open, then V(G) = sup{V(y): Y c G}, where Y denotes a
figure.
(c) If K is compact, then V(K) = inf{ V(Z): K c int Z}, where int Z
denotes the interior of a figure Z.
(d) If A = A I U A z U ... , where Ak is measurable for each k = 1,2, ... ,
then
x
(5.10) V(A) ~ L V(Ad·
k=1
PROOF OF THEOREM 5.2. Part (a) of Theorem 5.2 was verified earlier (see
Example 1), and parts (b), (c) are just the definitions. Let us prove (d) when
A is bounded and postpone the proof for unbounded A to Section 5.10.
177
5 Integration
Define the disjoint sets B I' B l ' . .. as in the proof of Theorem 5.1. By
Proposition 5.1 b,
rJJ
V(A) = L V(Bd·
k= I
V(G) = L V(Ik)'
k=1
The third possibility is that G contains some half-line. In that case V(G) =
+ 00.
°
The sets of measure playa special role. They turn out to be negligible
in integration theory, and for that reason we shall call them null sets.
Corollary 2. If AI' A l , ... are null sets, then AI U Az U ... is a null set.
If B c A and A is a null set, then B is a null set.
PROOF. By (5.10),
° °
rJJ
~ V(A I U Al U ... ) ~ L = 0,
k= I
which proves the first assertion. If A is a bounded null set and B c A, then
EXAMPLE 4. Let A be the set of rational numbers in the interval (0, 1). Then
A is countable. For instance, one can write
178
5.2 Measure
Theorem 5.3
(a) Let A l , A 2 , ••• be a nondecreasing sequence of measurable sets. Then
(5.11 )
(5.12)
PROOF. Let us prove the theorem under the assumption that there is a
spherical ball U such that A, c U for each v = 1,2, .... This restriction
will be removed in Section 5.10. To prove (a), define B10 B 2 , • " as in the
proofs of Theorems 5.1 and 5.2. Then
We get (5.11) by taking the limit as v ~ 00. To get (5.12), we apply (5.11) to
the nondecreasing sequence of sets C,. = U - A,., and note that V(C) =
V(U) - V(A.) since A. c U. D
EXAMPLE 6. To see the need for the assumption V(A tl < 00 in Theorem
5.3(b), let n = 1 and A. = [v, 00). Then Al :::J A2 :::J ••• and V(A,.) = + 00
for each v = 1,2, .... However, Al n A2 n ... is empty, and hence
VeAl n A2 n···) = O.
179
5 Integration
PROBLEMS
4. Let A = Al U A2 U ... , where Ak = {(x, y): x = 11k, 0 :5: Y :5: I} for k = 1,2, ...
Show that V2 (A) = o.
5. Let Ao be the circular disk with center (0, 0) and radius 1. For k = 1,2, ... , let Ak
be the circular disk with center (1 - 4- k)e1 and radius 4- k- l • Let A =Ao-
(AI u A2 U ... ). Find V2(A).
6. Prove Lemma 1. [Hint: Consider the collection of all intervals I such that leG.
The interiors int I of these intervals form an open covering of K.]
9. Let A = {XI' x2 , ••. } be a countable subset of (0, 1). Given 0 < Il < 1, let Ilk =
1l2- k-Ik = (Xk - Ilk' Xk + Ilk), and G = II
l, U 12 U ....
(a) Show that VI(G) :5: Il.
(b) In particular, let A be the set of rational numbers in (0,1). Let K = [0,1] - G.
Then K is a compact subset of the irrational numbers. Show that VI(K) ~ 1 - Il.
(c) Show that K = fr K.
180
5.3 Integrals over En
(If V(A) = + 00, this means that for every C > 0 there is a compact set K c A
with V(K) ?: c.)
(5.13) f¢ dV = J1 Ck V(Ak)'
EXAMPLE 1. Let ¢(x) = kim for x E [(k - 1)lm, kim), k = 1,2, ... , m, and
¢(x) = 0 for x ¢ [0, 1). Then Ak = [(k - 1)lm, kim), Ck = kim, and
_1 +
m2 m2 f + ... + ~ = m + 1.
¢ dx =
m2 2m
2..
We write dx instead of dV1 in case n = 1. Note that for large m the integral
is approximately!, which is the area of the triangle bounded by the lines
y = x, y = 0, and x = 1 in E2.
f ¢ dV = f
k=lm
~ V(Ad = n I ~ ~(~)2 - (~)2J
k=lm~m m
f ¢ dV = n3
m
[2 f
k= 1
k2 - I kJ = -;m [2m(m + 1)(2m
k= 1 6
+ 1) - m(m + I)J
2
f A.
'I'
d
V
= n(m + 1)(4m - 1)
6m 2 •
For large m the integral is approximately 2n13. which is the volume of the
solid in £3 bounded above by the cone z = (x 2 + l)1/2, below by the plane
z = 0, and with lateral boundary the cylinder x 2 + y2 = 1.
181
5 Integration
Figure 5.4
Proposition 5.3. Let ¢ and t/I be step functions, and c any real number. Then
¢ + t/I and c¢ are step functions. Moreover,
f t/I dV =
ppm
I~O dl V(BI) = I~O k~O dl V(Ak n Bl )·
°
Definition. Let fbe real valued with domain P. The support off is the smallest
closed set K such that f(x) = for every x rf: K.
We are now ready to define upper and lower integrals for a bounded
function f that has compact support. The upper integral of f is denoted
182
5.3 Integrals over E"
1
by f dV. If <P is any step function such that <P ~ f, then J<P dV is an upper
estimate for it. We take the greatest lower bound of the set of all such numbers
J<P dV.
Definition. The upper integral over En of a bounded function f with compact
support is
JfdV~ Jf dV.
(5.16)
(5.17c) f f dV = f g dV if f ~ g.
183
5 Integration
PROOF. Given e > 0 there exists a step function l/Jl ~ f such that
fl/J2 dV < J
9 dV + ~.
Then l/Jl + l/J2 is a step function and l/Jl + l/J2 ~ f + g. Consequently,
J(J+g)dV< Jf dV + Jgdv+e.
(S.18a)
Similarly,
*Riemann integral
If in (S.1Sa) only elementary step functions l/J are allowed, then the upper
Riemann integral off is obtained. The lower Riemann integral off is obtained
by allowing only elementary step functions", in (S.1Sb). Let us denote upper
184
5.3 Integrals over En
(5.19) ~(f) ~ J ~J
f dV f dV ~ S(f).
If §.(f) = S(f), then f is called Riemann integrable. Their common value S(f)
is the Riemann integral of f. From (5.19), if f is Riemann integrable, then f is
integrable [in the sense of (5.16)] and
PROBLEMS
f ¢ dV2 =
rs(r +s-
2
1)
.
3. Let a unit square be divided into a square of side (4m + 1)-1 in the center and 2m
annular figures of equal width (4m + 1)-1 surrounding it, as shown in Figure 5.5.
e2r-----------------~
I
I
I
I
----~----
2m
Figure 5.5
185
5 Integration
Let cp(x, y) = 0 for (x, y) in the small square or outside the large square. Let cp(x, y) =
(-I)kk in the kth annular figure, k = I, ... , 2m. Show that
f cp dV2
= 8m(2m
(4m +
+ 1)
1)2 .
What is this approximately when m is large?
4. (a) Show that if f is integrable, then S (ef)dV = e Sf dV. [Hint: Show that this is
true if e ~ 0, and that - S g dV = 1(-g)dV for every g. If e < 0, set g = ef and
g = -efJ
1
(b) Show that f dV ::;; 1g dV if f ::;; g.
The function fA is bounded and has compact support. The values off outside
A should contribute nothing to the integral of f over A.
(5.21)
186
5.4 Integrals over bounded sets
Therefore (1) and (2) follow from Proposition 5.4. Let 1A denote the function
with the value 1 on A and otherwise O. It is a step function, called the
characteristic function of A, and by (5.13) with m = 1, C1 = 1, S 1A dV = V(A).
This establishes (3). For (4) we have fA S gA and apply Proposition 5.4.
To prove (5) we have
f(x) s I f(x) I s C for every x E A.
Hence from (4)
f A
f dV = r
JintA
fdV= I
clA
fdV.
Let us next show that under quite mild assumptions about f, the integral
exists. It is for this purpose that the idea of measurable function is introduced.
Let f have domain En.
Definition. If {x: f(x) > c} is a measurable set for every scalar c, then f is a
measurable function.
187
5 Integration
kim
(k - J)/m
A,
Figure 5.6
188
5.4 Integrals over bounded sets
The construction of the step functions 4>, ljJ in the proof just completed
is a key to understanding the power of the Lebesgue theory of integrals. In
Figure 5.6 we first subdivided the vertical axis, and then obtained the sets
Ak on which 4> and ljJ are defined. From their definition, 4>(x) - ljJ(x) is
uniformly small for large m. In fact, 4>(x) - ljJ(x) :::;; 11m for all x E En. This
construction is quite different from that used in defining the Riemann integral.
As mentioned at the end of Section 5.3, the latter begins with subdividing En
into intervals, on each of which elementary step functions approximating I
from above and below must be constant. Thus, the Lebesgue theory sub-
divides the I-dimensional range of I, while the Riemann theory subdivides
the n-dimensional domain of f.
To apply this key lemma, we let A be a bounded measurable set, and I
be a function whose domain contains A. If {x E A : I(x) > c} is measurable
for every scalar c, then we call I measurable on A.
189
5 Integration
PROBLEMS
1. Let f(x) = 2x - x 2 if 0 ::; x ::; 2, f(x) = 0 otherwise. Using the notation in the
proof of Lemma 1, describe the sets AI, ... , Am. Sketch the step functions 1> and '"
in the case m = 4.
2. In each case show that f is integrable over A.
(a) f(x~ = x 2 exp x, A = [0, a].
(b) f(x) = sin(l/x) if x =I 0, frO) = 5, A = [ -1, 1].
(c) f(x, y) = (x 4 - y2)/(X 2 - y), A = {(x, y): Ixl ~ 1, Iyl ~ 1, x 2 =I y}.
(d) f(x) = 0 if x is irrational, fix) = l/q if x = p/q where p and q are integers with
no common factor, A = (0, 1].
(e) f(x) = 1 if x is irrational, fix) = 0 if x is rational, A = [a, b].
3. For each part of Problem 2 describe the sets {x E A: f(x) > c}.
4. Show that if f, g, and h are integrable over A and I fIx) - g(x) I ::; h(x) for every
x E A, then ISA f dV - SA g dVI ::; SA h dV.
5. Let f be of class C(2) on [0, a] and b = max {f"(x) I : 0 :$; x ::; a}. Let g(x) =
frO) + !,(O)x. Using Problem 4, show that IS~ f dx - afrO) - a 2f'(0)/21 ::; a 3 b/6.
Use this result to estimate S6/2 exp( -x 2 /2)dx.
6. (Mean value theorem for integrals.) Let A be compact and connected. Let f be
continuous on A and g be integrable over A with g(x) ;::: 0 for every x E A. Prove
that there exists x* E A such that
fA
fg dV = f(x*) f A
g dV.
[Hint: Let C and c ·be the maximum and minimum values of f on A. Then
cg ::; fg ::; Cg. Use (2) and (4) of Theorem 5.4 and the intermediate value theorem.]
190
5.5 Iterated integrals
x"
A(x') { -----QI•
I
:
,
A ,
:
I
----1--,
I I
I
I
I I I
, I I
I I I
I I I
,I ,I
--+---~~----L------X'
~
R
Figure 5.7
(5.23) f A
f(x)dv,,(x) = r {f
JR A(x')
f(x', xlI)dv,,-s(XII)}dV.(X')'
g(x') = f A(x')
f(x', x")dv,,-ix").
The outer integral on the right-hand side of (5.23) is the integral of 9 over R.
The set R is compact, since R is the projection of the compact set A on ES •
It is easily shown that A(x') is compact for each x', and that f(x', ) is con-
tinuous on A(x'). However, we do not need to use these facts in the proof of
Theorem 5.6. The function 9 need not be continuous. However, it is shown
in the course of proving Theorem 5.6 that 9 is integrable over R. In the proof
of Theorem 5.6, the following fact about monotone sequences of functions is
used. For each v = 1,2, ... let Fv be a bounded measurable function with
191
5 Integration
I F dv" = v-oo
lim IFv dv".
192
5.5 Iterated integrals
Figure 5.8
If X rt A, then since A is closed there exists v(x) such that d. < dist(x, A)
for every v ~ v(x). In this case 0 = F(x) = F .(x) when v ~ v(x). Thus (**)
holds for every x rt A. Suppose that x E A. Since f is continuous, given
e > 0 there exists ~(x) > 0 such that I f(y) - f(x) I < e for every YEA
such that Iy - x I < ~(x). Choose v(x) such that d. < ~(x) for every v ~ v(x).
Then F(x) ~ F .(x) < F(x) + e for every v ~ v(x). Therefore (**) also
holds when x E A.
By the monotone sequences theorem
f F dv" = lim
\'-+ 00
fF" dv".
For each x', the functions F,,(x', ) form a monotone sequence tending to
F(x', ). Let
G.(x') = f
F.(x', )dv,,-., G(x') = fF(x', )dv,,-s'
lim
'-00
fG. d~ = fG d~.
Since F. is an elementary step function, we have by (*) with <I> = F.,
Therefore
(5.23a)
Since F = fA' S F dv" = SA f dv". For x' E R, the inner integral on the right-
hand side of (5.23) is G(x'); and G(x') = 0 for x rt R. Thus, S G d~ equals the
193
5 Integration
outer integral on the right-hand side of (5.23). From (5.23a), Theorem 5.6 is
proved in case f ~ o.
To remove the assumption f~ 0, write f = (f + C) - C, where C is
the maximum value of If I on A. Then f is the difference of nonnegative con-
tinuous functions for each of which (5.23) holds. By subtraction, (5.23)
holds for f. D
It is not essential in Theorem 5.6 that x' = (Xl, ... , XS). One can equally
well take integers il < ... < is and
wherejl < ... < jn-s are those integers between 1 and n not included among
iI, ... , is·
In particular, let n = 2. Then s = 1, and we can take either x' = x or x' = y.
To avoid writing parentheses we write Sdx Sf dy instead of S{j f dy}dx.
Then
where the iterated integrals are taken over the appropriate subsets of El.
Many authors write the iterated integral as IS f dy dx, but this notation
would lead to confusion when we come to the exterior differential calculus
in later chapters.
The iterated integral is usually easier to evaluate when taken in one of the
two possible orders than in the other order.
f L:1 dx (x 2 + y)dy.
Then f(x, y) = x 2 + y and A is as shown (Figure 5.9). Evaluating the inner
integral first, we get
il
o dy _)X2 + y)dx =
f,;y i1(X33-
0 + xy )/,;y
_,;ydY ="38 i1
0 y3j2 dy 16
= 15·
194
5.5 Iterated integrals
--~----~~~~----+--x
-\
Figure 5.9
f
A
f dV3 = f
JR
{f A(x.y)
f dZ}dVi X , y),
J
and A(x, y) is the interval [0, 4 - x 2 - y] (see Figure 5.1 0.) Writing the
integral over R as an iterated integral, we get
x2
{ f dV3 = L2 dx L 4- dy LJ4-XLYfdZ.
There are five other possible ways of writing SA f dV3 as an iterated triple
integral. For instance,
x
Figure 5.10
195
5 Integration
In the same way, for any n the integral can be written in n! possible ways
as an n-fold iterated integral.
As a special case of Theorem 5.6, one can take f(x) = 1 for all x E A. One
gets the following formula for the measure of A.
Corollary
x"
A(u) f ------
__l~----~L---~----_XI
ue, ee,
Figure 5.11
EXAMPLE 3. Let A be the n-simplex with vertices 0, eel' ... , ee n , where e > 0.
For e = 1 this is the standard n-simplex (Section 1.5). Let us show by induction
on n that v,,(A) = en/n!. If n = 1, then A = [0, e] and Vl(A) = e. Assuming
the result in dimension n - 1, we apply (5.25) (see Figure 5.11). Now
°
A -
I e(e - u)n-l
v,,(A) = o(_ n1)'
(e - u)n Ie en
. du = - n(n _ 1)'. 0 = "n.
196
5.5 Iterated integrals
Then
A(x') = [h(x'), H(x')]
and
i = 1, ... , n.
Let m = (ml, ... , mn). The centroid of A is the point x such that m = v,,(A)x,
provided v,,(A) > O. The components of x satisfy
. mi
X' = v,,(A)' i = 1, ... , n.
i = 1, ... , n.
If p is constant, then x* = x.
197
5 Integration
Ix"
------+4----+--+---x'
e,
H(u)
Figure 5.12
ml = r
JH
Xl dY,.(x) = II
0
U du r
JH(U)
1 dy"-l(X")
= IXn - l II
o
u(I - U 2 )(n- 1)/2 du = l.
IXn -l
n+
Thus
198
5.5 Iterated integrals
PROBLEMS
1. Find the area and also the centroid of:
(a) {(X,Y):X 2 ::;; y::;; x + 2}. (b) {(x,y): Iyl- I ::;; x::;; jt=7}.
2. Express the iterated integral
as an integral over a set A c E2, and then as an iterated integral in the opposite
order. Evaluate it.
3. Express as an iterated triple integral:
Show that IX4 = n 2/2. (In Section 5.9 we give a general formula for IX,.)
8. Write points of E'+ 1 as (x, z), where x = (xl, ... , x'). Let
A = {(x, z):O::;; z::;; 1 -lxI2}.
199
5 Integration
If the set of numbers on the right-hand side has an upper bound, then f is
integrable over A. Otherwise, the integral diverges to + 00. The definition has
two immediate consequences, which we state in the following form.
L (g + h)dV = Lg dV +L h dV ~ Lg dV + L h dV.
tjt(r) = JA(r)
f dV.
r fdV~ J fdV~ I.
J
K A(r)
200
5.6 Integrals of continuous functions
fa
oo f dx = lim
r-+ + 00
fr f dx.
a
f oo
1
x-Pdx =--
P- 1
1
Figure 5.13
°
tinuous on A, f ~ 0, but f may be unbounded on any neighborhood of xo.
For each J > let A'(J) = {x E K : I x - Xo I ~ J} (Figure 5.l3). Each of the
sets A'(J) is compact, and
(5.28) fA
f dV = lim
b~O +
f A'(b)
f dV.
The proof of (5.28) is similar to that for (5.27). The same formula holds if f
is continuous on A = K - L, where L is any closed set. In this case A'(J)
is the set of points of K distant at least J from L.
EXAMPLE 3 (continued). Let A = (0, (0) and f(x) = x - p. Taking A I = [1, (0),
A2 = (0, 1], the integral over Al exists only if p > 1, and over A2 only if
p < 1. Hence SO' x- P dx diverges to + 00 for every p.
(5.30)
If both f and A are bounded, the new definition of integral agrees with
that in Section 5.4 (Problem 11).
Figure 5.14
202
5.6 Integrals of continuous functions
° °
and by (b) of Lemma 1 so is their sum If I = f+ + f-. Conversely, let If I
be integrable over A. Since ~ f+ ~ If I, ~ f- ~ I f I, by (a) of Lemma 1
f+ and f- are integrable over A. Hence f is integrable over A. D
f.En-K
Ix - Xo 1- P d V
EXAMPLE 5. Let
The function r is called the gamma function. Let f(x) = xu - 1 exp( - x). Since
°
f(x) ~ x u - 1 and p = 1 - u < 1, f is integrable over (0, 1] by comparison
with x- p. For any b, x b exp( - x) --+ as x --+ + 00 and therefore is bounded
on [1, 00). Letting b = u + 1 we see that there is a number C such that
f(x) ~ C/x 2 for every x E [1, 00). Thus f is integrable over [1, 00) and over
(0, 1], therefore over (0, 00).
The gamma function generalizes the factorial. First of all,
Integrating by parts,
which gives
(533) f(u + 1) = uf(u).
The integration by parts over (0, (0) is justified by taking it first over intervals
[<5, IJ and [1, r J and letting <5 --+ 0+, r --+ + 00. In particular, if m is an integer,
then f(m + 1) = mf(m). Since r(l) = 1,
(5.34) f(m + 1) = m!.
Iterated integrals
Theorem 5.6 regarding iterated integrals has the following extension (we
defer the proof to the end of Section 5.11). Let J be continuous on A. If J is
integrable over A, then Formula (5.23) holds. In case J ~ O. then a stronger
statement can be made. If either side of (5.23) diverges to + 00, then so does
the other. Therefore, if J ~ 0, one way to show that J is integrable over A is
to show that the iterated integral exists.
fJ o
OC!
dV2 = 4 f dy
fOC!
0
(1
x dx
2
+x +y
2)2 = 2
fOC! dy
0
-1- - 2 =
+y
7r.
The reader is cautioned that ifJhas both positive and negative values on A,
then one cannot conclude that J is integrable from the fact that the iterated
integral exists.
204
5.6 Integrals of continuous functions
The iterated integral in the opposite order does not exist, hence f is not
integrable over A.
PROBLEMS
(a) f p-=-t'
2
1
dx
x2 - 1
[Hint: Let cf>(x) = 1/Jx+l as in Example 4.]
(b) f"12
-"12
Isinxl-Pdx. [Hint:x/sinx--+ 1 as x--+ 0.]
(d) foo
-00
31:
~x(lxl
dx
+ 1)
.
(f) fAX
~ dV2(x, y), A =
-2--2
+y
E2 - {(O,O)}
(g) f
E'
exp(-Ix - yl)
1 + Ix _ yl2 dV2(x,y).
2. Show that {(x, y, z): 0:5: z :5: (x 2 + y2)jxy, 0 < x :5: 1, 0 < y :5: I} has infinite
volume.
3. Show that the volume of {(x, y, z): 0:5: z :5: Ixylexp( _x 2 - y2)} is 1.
4. Show that Jj(x, y)dV2(x, y) exists if f is bounded, continuous, and I f(x, y) I :5:
C(1 + x 2 + y2)- p12, P > 2.
5. If lim,~oo S. . . ,J(x)dx exists but f is not integrable on EI, then this limit is called
the Cauchy principal value. Find the Cauchy principal value:
(a) f(x) = x/(1 + x 2 ). (c) Any odd function f, f( - x) = - f(x)
(b) f(x) = x + 1/(1 + x 2 ). for every x.
6. Let A be open, and f a function continuous and integrable on A. For r > 0, let
A(r) = {x E A : Ixl :::; r, dist(x, A C) ~ r- 1 }. Show that:
(a) A(r) is compact.
(b) SA f dV = lim,~oo SAl') f dV.
7. Show that each of the following integrals over En converges.
(a) S(lx12 + 1)- PI2 dV(x),p > n.
(b) S Ixl-1x1dV(x).
205
5 Integration
r
= [1,00), but
lim
\'-+00
f
Kv
f(x)dx f:- lim
r-+co 1
f(x)dx.
10. (a) Assume that for every real number c, both of the sets {x E A : f(x) < c} and
{x E A : f(x) > c} are open relative to A. Show that f is continuous on A.
[Hint: Show that {x E A: c < f(x) < d} is open relative to A.J
(b) Using (a) and Corollary 2, Section 2.6, prove: if f is continuous on A, then f+
is continuous on A.
11. Let A be a bounded, measurable set, and let f be bounded and continuous on A.
Show that SA f dV, as defined by (5.30), is the same as in Section 5.4. [Hint: Suppose
first that f :2: O. For any compact set K c A, we have, taking the integral over A
in the previous sense, 0 ~ SA f dV - SK f dV ~ C[V(A - K)J provided f(x) ~ C
for every x E A. But V(A - K) = V(A) - V(K) is arbitrarily small. Apply this to
f +, f -, and subtract.J
12. (Difficult.) Let f be continuous on an open set D. Assume that the integrals of
f+ and f- over D both diverge to + 00. Show that given any number I there is a
sequence of compact sets KI c K z c ... such that D = KI U K z U ... and
lim v _ ro JKJ dV = t.
PROOF. First of all, Lemma 1 is valid for the following particular kinds of
linear transformations. Points in the domain of L are denoted by t.
(1) For some k and I, the transformation L merely interchanges the com-
ponents t k and t ' of t. Then V[L(l)J = V(l) and det L = -1.
206
5.7 Change of measure under affine transformations
(2) The matrix of L is diagonal. If x = L(t), then Xi = elti, where c~, ... , c~
are the diagonal elements. We have I = 1 I X ... x l n , where e~ch li
is a I-dimensional interval. Then L(I) = 1'1 X ... x 1~, where 1; is a
I-dimensional interval of length lell times the length of l i . Since
det L = c~ ... c~, we have V[L(I)] = Idet LI V(I).
(3) There exist k and I, k ¥- I, such that Xk = t k + d, Xi = t i for i ¥- k.
For notational simplicity we take k = 1, I = 2, and regard En as E2 X En-2.
Then 1=1' x 1", where l' = 11 X 1 2 , 1" = 13 X ... x In. Let L' be the
transformation from E2 into E2, such that
L'(t l , t 2) = (t l + ct 2, t 2).
Then L'(I') = P', where P' is a parallelogram of the same area as 1', and
L(I) = P' x 1". By (5.24),
v,,[L(I)] = V2(P') v,,- 2(1") = V2(1') v,,- 2(1") = v,,(I).
Since det L = 1, Lemma 1 is valid for L of type (3).
Next, we observe that if M and N are any two linear transformations for
which the theorem holds, then
V[(M 0 N)(K)J = Idet M I V[N(K)]
= Idet M" det N IV(K),
and (det M)(det N) = det MoN. Hence the theorem also holds for their
composite MoN.
If N has row covectors Wi, ... , wn and M is of Type (3), then the kth
row covector of M 0 N is wk + cw l and the others are unchanged. The kth
column vector of NoM is Vk + CV 1 and the others are unchanged, where
Vb"', Vn are the column vectors of N. A transformation M of type (1) when
applied on the left interchanges the kth and Ith row covectors ofN, and when
applied on the right interchanges the kth and Ith column vectors. Moreover,
the inverse of a transformation of Type (1) or (3) is of the same type.
Now let L be any linear transformation. Then
where Mb ... , Mp are of Types (1), (2), and (3). Since Lemma 1 holds for
each M i , it is valid for L. D
207
5 Integration
Measure of an n-simplex
Let S be an n-simplex with vertices X O, Xb •.. , Xn (see Section 1.5). Let
Vi = Xi - X O ' The vectors Vb"" Vn are linearly independent. Let L be the
linear transformation with Vb"" Vn as column vectors, and let g(t) =
L(t) + x o ' As before, let k be the standard n-simplex. If t = (t I, . . . , tn) E k,
let to = 1 - (t 1 + ... + tn) and let x = g(t). Then
n n
X = Xo + L ti(X i - x o) = L tiXi'
i= 1 i=O
Measure of an n-parallelepiped
Given Xo and linearly independent vectors Xl' ... , X n , let
P = {X: X = Xo + t
1= 1
tiv i, 0 :s; t i :s; 1, i = 1, ... , n}.
208
5.8 Transformation of integrals
-----
1:2 t--------,
Xo + V2
Figure 5.15
PROBLEMS
1
X6. .xl ... X~)
.
V(S) =- (
det·· ..
n! Xo x~ x~
1 1 1
[Hint: Subtract the first column from each of the other columns. This does not
change the determinant.]
4. Let (C}) be a positive definite symmetric n x n matrix, and let )'10 ... ' An be its
characteristic values. Show that
where IXn is the measure of the unit n-ball. [Hint: See Theorem 4.4.J
5. Show that if Iv;I ~ C for each i = 1, ... , nand P is a parallelepiped spanned by
VI, ••. , Vn , then V(P) ~ c. [Hint: Use induction, a suitable rotation of En, and the
method of slices.]
209
5 Integration
( \
I
----
B g
/ to
m
--rI
~
"-
Figure 5.16
f.
g(1)
f dV ~ f[g(t o)] IJg(to) I V(I).
Let
4>(t) = f[g(t)] IJg(t) I.
Since f is continuous and g is of class C<l), the composite f o g is continuous
and IJg I is continuous. Hence 4> is continuous. The integral of 4> over I is
approximately 4>(t o)V(I), which is the right-hand side of the above expression.
The figure Z is the union of small intervals II, ... , I p. Let tk E I k • We
should have approximately
f.
g(Z)
f dV ~ t 4>(t )v(h) ~ f 4> dV.
k=l
k
Z
The following theorem states that the formula is correct, and not merely for
compact sets.
210
5.8 Transformation of integrals
(5.38) If(X)dV(X)
A
= f g-l(A)
f[g(t)] I Jg(t)IdV(t),
To prove Theorem 5.8 let us first prove three lemmas. In these lemmas I
denotes an n-dimensional cube. It is convenient to take I half-open to the
right. Then I = cl I - N, where N is a certain compact set composed of
(n - I)-dimensional faces of I. Hence g(I) = g(cl /) - g(N) is the difference
of compact sets, and so it is measurable. Let / be the side length of I, d its
diameter, and I' the concentric closed n-cube of side length (1 + 1:)/. Let G, L,
Xo be as defined at the beginning of the section.
s.
_f--- ['
--
t.
• to
Figure 5.17
211
5 Integration
f 11(1)
f dV < r¢ dV +
JI
B V(I)
Then a 2 °
since
(a - ~)(b - ~) - (a
°
f 2 0, and b > since g is regular. Let rjJ(~, r) =
°
+ ~)b(1 + rt. Then rjJ(O, 0) = and rjJ is continuous.
°
Hence IrjJ(~, r) I < B for every (~, r) in some neighborhood V of (0,0). Choose
some ~ > 0, r > small enough that (~, r) E Vand ~ < b. Then
(*) (a + ~)(1·+ rtb < (a - ~)(b - ~) + B.
f 11(1)
f dV s (a + ~)V[g(I)],
and by Lemma 1, V[g(I)] s V[G(I')]. By Theorem 5.7,
V[G(I')] = b V(I') = b(1 + rtV(I).
Therefore
(**) fg(l)
f dV s (a + ~)b(1 + rtV(I).
On the other hand,
(***) i ¢ dV 2 (a - Wb - ~)V(I).
212
5.8 Transformation of integrals
f 11(1)
f dV 5, rcP dV.
J[
PROOF. Suppose this is false for some n-cube ]0. Then
c= f g(10)
f dV -
Jr cP dV
[0
Choose some suchj and let]1 = ]j' By dividing]1 into 2n congruent n-cubes
and repeating the argument, we obtain ]2. Continuing, we obtain a sequence
of n-cubes]1 :::> ]2 :::> ••• such that
f.
g(1')
fdV- r cPdV~c2-ln,
J['
I = 1,2, ... ,
and the diameter of]1 tends to 0 as 1-+ 00. By Theorem 2.3, (cl]l) n (cl ]2) n·· .
contains just one point to, Let e < c/V(I°). Then
eV(I') = er'nv(I°) < c2- ln .
213
5 Integration
of those meeting B, and so on. Then Z 1 :::J Z 2 :::J . •• and their intersection
is B. There exists Vo such that Z. c Ll for all v ~ Vo.
Applying Lemma 3 to each of these congruent disjoint n-cubes comprising
Z., and adding, we get
ff
A
dV s, fg(Zv)
f dV s, r cf> dV
Jz v
{f dV s, Lcf> dV.
But g-l is also a regular transformation. Interchanging the roles of A
and B,
rf dV = f
JK g-'(K)
cf> dV s, rcf> dV.
JB
By (5.26),
(#)
214
5.8 Transformation of integrals
If either f or cjJ is integrable, then so is the other, and (5.38) holds. This
completes the proof of Theorem 5.8. 0
Corollary 1
EXAMPLE 1. Let A = {(x, y): x > 0, y > 0,0 < xy < 3, x < y < 2x},f(x, y)
= y2, g(s, t) = Pte 1 + foe 2 for s > 0, t > O. We show that g is univalent
by solving the equations
x- ~
- '-it' y= V r;t
~1
explicitly, and find that s = xy, t = y/x. Since Jg(s, t) = 1/2t #- 0, g is regular.
Moreover, the part in A of the hyperbola xy = c,O < c < 3, corresponds to
the segment s = c, 1 < t < 2 in B. Hence B is as shown in Figure 5.18, and
f.
P
f(x)dV(x) = Idet L I f.
10
f[g(t)]dV(t).
215
5 Integration
2~-----, g
B
--~--------------x
3
Figure 5.18
PROBLEMS
3. Find the second moment about (0, 0) of the parallelogram with vertices (0,0),
e l + e z , - 2e l + 3e z , - e l + 4e z . [Hint: Let g be the linear transformation L with
column vectors e l + ez , -2e l + 3e z .J
4. Let B be a compact set, x its centroid, and g an affine transformation. Show that
g(x) is the centroid of g(B).
5. Let A be symmetric about 0, that is, x E A implies - x E A. Let f be integrable over
A and f( - x) = - f(x) for every x E A. Show that SA f dV = 0.
6. Let g be of class e(l) on d and K be a compact subset of d. Show that there is a
number e such that Ig(s) - g(t) I ~ Cis - tl for every s, t E K. [Hint: Proposition
4.5, Section 4.3.J
7. Let g be regular and K, e as in Problem 6. Show that if Be int K, then V[g(B)J :-:;
eV(B). [Hint: For t E B the partial derivatives ofg satisfy Ig;(t) I :-:; c. Use Problem 5,
Section 5.7, to see that IJg(t) I :-:; e.J
8. Let g be a regular transformation from an open set d onto D. Let B = g-I(A), where
A is a measurable subset of D. Show that B is measurable. [Hints: First consider the
case when A c LcD, where L is compact. Also, g-l is regular. See Problem 7.
Find compact sets LI c L z c··· with union D.J
216
5.9 Coordinate systems in En
Lf(X, y)dV2 (x, y) = Saf[r cos 8, r sin e]r dV2 (r, e),
where g(B) = A - N. Since N is a null set, the integral over A is the same
as the integral over A - N.
5. (Spherical coordinates in E"). For n = 2, this is the polar coordinate system.
e
Proceeding inductively, let r = lxi, 1 be the angle from the positive x 1 _
axis to x; more precisely,
e1 = cos- 1(x 1jr), 0 < 8 1 < n,
and
217
5 Integration
where 13n is a number not depending on C/>, r1' or rz· To find 13n, set c/> == 1.
Then
where an is the measure of the unit n-ball. Hence 13n = nan, which turns out
to be the (n - 1)-dimensional measure of the unit (n - 1)-sphere.
6. (Cylindrical coordinates in E3.) Let (R, e) be the polar coordinate system.
Then (R, e, Z) is a coordinate system for D = E3 - {(x, 0, z): x ~ O}.
The equations x = r cos (), y = r sin (), z = z relate the cylindrical and
standard cartesian coordinates of a point XED. In a similar way cylindrical
coordinates can be introduced in En.
7. The idea of barycentric coordinates (Section 1.5) does not agree precisely
with the definition in this section. However, let to, t 1 , ••• , t n be the bary-
centric coordinates of x, with respect to the vertices xo ' Xl> •..• Xn of
an n-simplex S. Let g be the affine transformation defined at the end of
Section 5.7, and let f = g-l. Then t 1 , ••• , t n are the coordinates ofx in the
affine coordinate system f, and to = 1 - (t 1 + ... + tn).
218
5.9 Coordinate systems in En
r(u)r(v) = 2 2- u- v {
OO
S 2U-1 exp( - ~)dS {OOt 2V - 1 exp( - ~)dt.
Writing the iterated integral as an integral over the first quadrant Q and
introducing polar coordinates,
= 22- u- v f
0
OO r2(u+v)-1 exp ( - 2
r2)dr f"/2
0 (cos O)2U-1(sin O)2v-1 dO.
The function B is called the beta function, and we have just shown that
(5.42) B( ) = r(u)r(v)
u, v r(u + vr
(5.43)
°
integer m = 0, -1, -2, ... , one can use (5.44) to define r(u). For instance,
if -1 < u < 0, then < u + 1 < 1 and by definition r(u) = r(u + l)ju.
Next r(u) is defined for - 2 < u < -1, and so on. The gamma function can
also be defined for complex values of u [23, p. 148].
219
5 Integration
To obtain another expression for B(u, v), set cos 2 e = g(e) = z and apply
the transformation formula. Then Ig'(e)1 = 2 cos sin and e e,
(n/2
B(u, v) = Jo (cos 2 8)U-l(sin2 8)V-12 cos esin ede,
A variety of integrals can be reduced to either (S.41) or (S.4S) and hence can
be evaluated in terms of the gamma function (see Problem 11).
PROBLEMS
220
5.9 Coordinate systems in E"
Figure 5.19
12. Let Q(x) = Ii.i= I C}Xixi > 0 for every x '" 0, where c~ = c{ for i,j = 1, ... , n.
f
Show that
exp[ - Q(x)/2]dv,,(x) = (2n)n/2[det(c~n- 1/2.
[Hint: Make a suitable orthogonal transformation.]
221
5 Integration
PROOF. Let
v--+ 00
These limits exist since ami ~ aml ~ ... and alv ~ al v ~ .. , . Moreover,
b1 ~ bl ~ ... and Cl ~ C2 ~ ... , Let
b = lim bm , c = lim cv.
For every m and v, amv ~ c. Hence bm ~ c for every m, and it follows that
b ~ c. Similarly, c ~ b. 0
222
5.10 Measurable sets and functions; further properties
By Lemma 1,
V(A) = lim V(A.) .
..... 00
223
5 Integration
V(Dv) S L V(Ak)'
k=1
By (5.11),
00
PROOF
{x:g(x) > e} = o
l'= 1
Then Z1 ~ Z2 ~ .... The limit of the monotone sequence Z1, Z2, ... is called
the lower limit of the sequence y\, Y2,"" and is denoted by lim inf\'~Ci y,,,
Similarly, if
Wl' = sup{yl" Yl'+ 1, . . . }, v = 1,2, ... ,
then WI ~ W2 ~ .... The limit of the monotone sequence WI' W2' .,. is the
upper limit lim supl'~OO Yl" Since Zl' ~ Wl' for each v = 1,2, ... , we must have
Equality holds if and only if the sequence Yl' Y2' ... has a limit.
then f is measurable.
PROOF. For each v = 1,2, ... let hix) = inf{fix),/l'+ I(X), ... }. By (3b)
each hl' is measurable; hi ~ h2 ~ ... , and by the definition of "lim inf,"
f(x) = sup{h 1(x), h 2(x), ... }. Therefore f is measurable. D
225
5 Integration
j - 1 j
if - - < f(x) ~-,
v v
where j is an integer, and 1 - v2 ~ j ~ v2 . This construction is like that in
Section 5.4, Lemma t. Then
f(x) = lim fv(x)
v-+ 00
for every x. Defining gv in the same way,
g(x) = lim gv(x).
For each v the functions fv and gv are measurable and take a finite number of
values. It is easy to show that their sum fv + gv is measurable. Since for
every x
f(x) + g(x) = lim [J.(x) + gv(x)],
v-+ 00
f + g is measurable by (5). o
PROBLEMS
1. Let amv = m(m + v)-t. Show that (5.47) does not hold. Which assumption in Lemma
1 fails to be satisfied?
2. Let 1 and g be measurable real-valued functions. Show that their product is measur-
able. [Hint: Show that the square of a measurable function is measurable. Then
21g = (f + gf - 12 - g2.]
3. Let A be a measurable set with V(A) < 00. Show that there exists a measurable set
B c A with V(B) = tv(A).
4. Show that:
(a) If A is measurable, N is a null set, and A - NcB c A u N, then B is measurable
and V(A) = V(B). [Hint: First consider the case of bounded sets.]
(b) If1 is measurable and f(x) = g(x) except for x in a null set N, then g is measurable.
226
5.11 Integrals: General definition, convergence theorems
(5.49) II dV = lim
r--+ + 00
r I dV.
JV r
If the limit is finite, then I is integrable (over En). Otherwise the left-
hand side of (5.49) equals + 00. If 0 ~ I ~ g, then JI dV ~ Jg dV.
(b) Let I be measurable and I ~ O. For any t > 0 consider the function
Jsuch that
J(x) = min {f(x), t} for every x.
(5.51)
lim f<Pv dV = O.
v--+ 00
PROOF. There exists e such that <Pv(x) ~ e for every x and v = 1, hence also
for v = 2,3, .... Given c > 0 let c = c/2V(K). Let
Av = {x: <Pv(x) > c}.
By hypothesis K ::J Al ::J A2 ::J ... and Al II A2 II ... is the empty set.
Since V(A I ) is finite, we may apply Formula (5.12), obtaining
lim V(Av) = V(AI II A2 II···) = O.
v--+ 00
Let Vo be such that V(Av) < el2e for every v ~ vo. Then
f <pvdv~eV(Av)<~,
Av
ifv~vo,
while for every v,
r
JK - Av
<Pv dV ~ cV(K - Av) ~ cV(K) = -2c .
Hence for every v ~ Vo ,
o
228
5.11 Integrals: General definition, convergence theorems
PROOF. We have shown in Section 5.10 that the limit f is measurable. First
let us assume that f is bounded and has compact support. Let 4>v = f - fv.
Since 0 ~ fv ~ f, each function fv is also bounded and has compact support.
Moreover,
amv = fUrn
fv d V, m, v = 1, 2, ...
and apply Lemma 1 in Section 5.10. Finally, iffis unbounded observe that
for each m = 1,2, ... ,
(5.53) Iu + g)dV = If dV + I g dV
229
5 Integration
The sequencesfl' f2"'" gl' g2' ... , fl + gl' f2 + g2,'" are nondecreasing
and tend respectively to f, g, f + g. By the monotone sequences theorem, the
corollary is true when f ~ 0, g :?: O.
In the general case,
Os,(f+gt s,f+ +g+.
Since f+ and g+ are integrable, so is (f + g)+. Similarly (f + g)- is inte-
grable. Then
f+ + g+ = (f + g)+ + cp,
f- + g- = (f + g)- + cP,
where cp :?: O. Since the corollary is true for nonnegative functions,
f(Cf)dV = c f f dV.
Corollary 3. Let fl, f2, ... be measurable, fl be integrable, fl :?: f2 :?: ... :?: 0,
and let f(x) = limv-+oo fv(x) for every x. Then (S.S2) holds.
PROOF. Let gv = fl - fv and apply the theorem to the nondecreasing
sequence gl, g2 ... which has fl - f as limit. D
For sequences that are not necessarily monotone, the result is as follows.
Fatou's lemma. Let /.. be measurable and /.. ~ 0 for each v = 1, 2, .... Then
230
5.11 Integrals: General definition, convergence theorems
m = v, v + 1, ... ,
J f dV = v-+
lim Jh v dV.
00
Since (**) holds for each v, the limit is no more than the right-hand side of
(**). 0
V(Am) ~~ Jf dV.
231
5 Integration
we have
lim inf Igv dV = lim inf Ifv dV - II/> dV.
v-+ 00 v-+ 00
Applying Fatou's lemma to the sequence gl, g2,'·' and adding JI/> dV to
. each side, we again get (S.S4a).
Similarly, if for each v, fv :::; I/> where I/> is integrable, then
I
f dV = lim Jfv dV.
v .... 00
PROOF. Since - 9 :::; fv :::; 9 and both - 9 and 9 are integrable, we can apply
(S.54a) and (S.S4b). But
lim inf fv(x) = lim sup fv(x) = f(x)
v.... 00 v.... 00
But lim inf :::; lim sup, and hence JJ. dV tends to Jf dV. D
Note: In (b) it suffices that Ifv(x) I :::; g(x) almost everywhere. One can
redefine J.(x) to be 0 on the null set where this inequality does not hold
without changing Jfv dV.
232
5.11 Integrals: General definition, convergence theorems
{ f dV = ffAdV.
f Fv dV = {fv dV,
If, for every x E A, 0 s fl(X) S f2(X) S '" and fv(x) tends to f(x) as v ~ 00,
then 0:$ F 1(x) :$ F 2(X):$ .,. and F,,(x) tends to F(x) for every x E P.
Applying the monotone sequence theorem to the sequence F 1- F 2 _ . . . _ we get
J
A
f dV = lim
v-oo
J
A
fv d V.
The proofs of Fatou's lemma and the dominated convergence theorem for
integrals over A are similar.
Corollary 4. Let A be a measurable set of finite measure and f1' f2' ... measur-
able on A. Assume that:
(a) limv-+oo f.(x) = f(x) almost everywhere in A.
(b) There is a number C such that IUx)1 :$ C for every x E A and v =
1,2, .... Then
f A
f dV = lim
v-c:c
f
A
fv dV.
233
5 Integration
PROOF. Let g(x) = C for every x. Since V(A) is finite, g is integrable over A.
The corollary is then a special case of the dominated convergence theorem
for integrals over A. 0
(5.55) f A
f dV = lim
v~oo
f
Av
f dV.
PROOF. To prove (a), letfv = fAy, v = 1,2, .... Then limv-+oo fv(x) = f(x) for
every x E A and Ifv(x) 1 :::;; If(x)l. The conclusion follows from the dominated
convergence theorem, with g = 1f I.
The proof of (b) is similar. 0
f
A
f dV = f
A-N
f dV = lim
v--+oo
r f dV.
JKy
Thus fA f dV :::;; s. Hence fA f dV = s. o
The right-hand side of (5.56) was taken in Section 5.6 as the definition.
Corollary 6 shows that the definition there agrees with the one in the present
section in case f ~ O. Since the procedure for defining the integral when f also
has negative values was the same in both sections [see (5.30) and (5.51)],
the two definitions agree in general.
234
5.11 Integrals: General definition, convergence theorems
Iterated integrals
Formula (5.23) expresses fA f dV as an iterated integral. In Theorem 5.6
we showed that (5.23) is correct if A is compact and f is continuous on A.
Actually, (5.23) is correct if A is any measurable set and f integrable over A.
This is called Fubini's theorem. Let us show that Fubini's theorem is correct
if A is measurable and f is continuous on A. For the general case, when f is
integrable but not necessarily continuous on A, we refer to the work by
McShane and Botts [18, p. 143].
Consider first f continuous on A and f : : : O. Let K 1, K 2, ... and N be as
in the proof of Corollary 6. Let A = K 1 U K 2 U· .. ; then A = Au N. Let
gv(x') = r
JKv(X')
f(x', x")dv,,-.(x"), g(x') = f-
A(x')
f(x', x")dv,,_ix")
where the sets Kv(x'), A(x') are defined by (5.22) with A replaced by K., A.
By Theorem 5.6,
iKv
f(x)dv,,(x) = i
R
gv(x')dY.(x').
We leave it to the reader to show that v,,-s[N(x')] = 0, for almost all x' E R
(Problem 11). Since A(x') = A(x') u N(x'), we have for almost all x'
g(x') = fA(x')
f(x', x")dv,,_ix").
Then (*) becomes the desired formula (5.23). The proof shows that if either
side of (*) diverges to + 00, then so does the other side, provided f : : : O.
If f is continuous and integrable on A, with both positive and negative
values, we write f = f + - f - with f + , f - as above. Since both f + and f-
satisfy (5.23), f does also.
Note. If A is a-compact (Problem 5, Section 5.10), then we can take
A = A and avoid mentioning the sets N(x') in the proof.
PROBLEMS
1. Let f be integrable, and f(x) = g(x) almost everywhere. Show that 9 is integrable,
and that Sf dV = J9 dV. [Hint: This is known from Section 5.4 if f and 9 are
bounded and have compact supports.]
235
5: Integration
2. Let f(x, y) = (x 2 + y2)- p/2 if 0 < x 2 + l < 1; and f(x, y) = 0 otherwise (p > 0).
Find If and ft! dV2 • What happens as t -+ oo?
5. Let fv(x) = v if x E (0, V-I) and fv(x) = 0 otherwise, v = 1,2, ... Show that
lim v.... oo fv(x) = 0 for every x, but f fv dx = 1. Why does this not contradict
Lebesgue's dominated convergence theorem?
6. Let fv(x) = v/(x 2 + v2), v = 1,2, ... Show that 0 ~ fv(x) ~ 1, limv.... oo !v(x) = 0
f
for every x, and fv dx = 1t. Why does this not contradict the dominated con-
vergence theorem?
7. Let fl,f2' ... be functions on EI, which satisfy the hypotheses of Lebesgue's
dominated convergence theorem. Let Fv(x) = fo
fv(t)dt, F(x) = fo
f(t)dt. Show that
Fv tends to F uniformly on EI as v -+ 00. [Note: As is customary notation in
elementary calculus, SO = fro.
x] if x :2: 0, and SO = - f~ if x < 0.]
8. (a) Let fl> f2' .. , be integrable over a measurable set A, and assume that
Lk()= I fA 1h 1dV is finite. Let G(x) =
I~ I I h(x) I· Since the terms of this series
are nonnegative, it either converges or diverges to + 00 for every x E A. Show
that G is integrable over A and hence G(x) is finite for almost every x E A.
[Hint: Apply the monotone sequences theorem to the sequence G I , G2, ... ,
where Gv = l(fl)AI + ... + 1(fv)A I.]
(b) By (a) the series I~ 1 h(x) converges absolutely for almost every x'= A. Let
F(x) be the sum of the series. Show that fA F dV = Ik'=
I fA h dV. [Hint:
Apply the dominated convergence theorem to the sequence Fl. F2, ... , where
F, = (fl)A + '" + (fv)A']
9. Let fl' f2,'" be integrable over A, where A has finite measure. Assume that
1h(x) 1 ~ Ck for every x E A and k = 1, 2, ... ,and that the series ~a;.1 Ck converges.
f
10. Suppose that fv :2: 0 for v = 1,2, ... and that fv dV -+ 0 as v -+ 00. Show that
lim infv.... oo !v(x) = 0 almost everywhere. [Hint: Problem 3, Fatou's lemma.]
11. Write x = (x', x") with x' E ES , x" E en- s. Suppose that v,,(N) = O. Let N(x') =
{x": (x', x") EN}. Show that v,,-s[N(x')] = 0 except for x' E N' where V.(N') = O.
[Hint: First take N bounded. Let G I ::::> G2 ::::> •• , be open sets such that N c Gv
for v = 1,2, ... and v,,(G v) -+ 0 as v -+ 00. Let 9v(x') = v,,-s[Gv(x')]. Then v,,(G v) =
f
9v dV.. Apply the monotone sequences theorem to 91' 92, ....
236
5.12: Differentiation under the integral sign
(5.57) o
at j
ff
Af(X, t)dv,,(x) ofj (x, t)dV(x),
= A at j = 1, ... ,1.
Thus cf>(t m ) --+ cf>(t o) as m --+ 00. Since this is true for every sequence in B
tending to to, cf> is continuous at to. 0
PROOF. Let to E B, and let c5 > 0 be such that B contains the interval
(to - c5, to + (5). If 0 < u < c5, then of/at is integrable over A x [to, to + u].
This set is measurable, and by the iterated integral theorem,
(*) f'O
10
+
dt f A of
U
at (x, t)dv,,(x) = f A {f
10
'o u
+ of (x, t)dt }dv,,(x).
at
By the fundamental theorem of calculus the inner integral on the right-hand
side is f(x, to + u) - f(x, to). If we let t{t(t) denote the right-hand side of
237
5 Integration
(**) f to + u
to
"'(t)dt = (Wo + u) - (Wo)·
Similarly (**) is true when -~ < u < O. Lemma 1 implies that", is con-
tinuous; hence by the fundamental theorem of calculus, "'(to) = </J'(t o). 0
If, instead of an open set, B is a closed interval [a, b], then the proof of
Lemma 2 shows that (5.59) is still true provided </J'(t) means the one-sided
derivative at the endpoints.
Theorem 5.9. Let A be measurable and B open. Let f and of/oti, j = 1, ... , I,
be continuous on A x B and satisfy
If(x, t)1 ::;; g(x), I:~ (x, t)1 ::;; hJ{x) for every x E A, t E B,
where g and hI' ... , hi are integrable over A. Then the function </J in (5.58)
is of class CU) on B and its partial derivatives are given by (5.57).
PROOF. Applying Lemma 2 with t 1, ••• , ti-I, ti - 1 , ••• , t' fixed, we get
o</J
oti (t) = f A
of (x, t)dv,,(x), ]. = 1, ... , l.
oti
EXAMPLE 1. Let </J(t) = If X-I exp( -xt)dx, t > O. Find </J'(t). Using (5.59),
</J'(t) = - f <XJ
1
exp(-xt)dx = -t- 1 exp(-t).
If B is an interval (a, (0), a > 0, the hypotheses of Lemma 2 are satisfied with
g(x) = h(x) = exp( -ax). The formula for </J'(t) is correct for all t in any such
interval, and hence for every t > O.
238
5.12 Differentiation under the integral sign
PROBLEMS
provided f and of/at are continuous on [ao, bo] x B where B is open, that
ao ~ a(t) ~ bo and ao ~ b(t) ~ bo for every t E B, and that the functions a, bare
of class C(1). [Hint: Let G(x, t) = go
f(s, t)ds, so that oG/ox = f Calculate the
derivative of G[b(t), t] - G[a(t), t].]
5. Let c/J(t) = Slog t X-I exp(x 2t 2)dx, t > 1. Find c/J'(t).
6. Let c/J(x, t) = S~~~: f(s)ds, where c > 0 and f is of class C(1) on El. Show that
Wc/J/ot 2) = c2Wc/J/OX 2).
7. Let c/J(x, y) = H; dt S~ f(s, t)ds, x> 0, y > O. Show that if f is continuous on
{(x, y): x 2 0, Y 2 O}, then (02cpjOX oy) = f
8. Let
(a) Show that c/J(x) = i-Jn exp( -2Ixl) for all x. [Hint: For x> 0, show that
l{>'(x) = -2l{>(x) using the substitution s = l/t.J
(b) Note that application of (5.59) gives a false result at x = O. Why is this not
surprising?
9. Let c/J(t) = SO' f(x, t)dx, for t > 0, where f(x, t) = exp( - xt)x - 1 sin x. Show that:
(a) c/J'(t) = -1/(1 + t 2).
(b) c/J(t) -+ 0 as t -+ + 00. [Hint: Apply the dominated convergence theorem with
fv(x) = f(x, tv) where 1 ~ t 1 ~ t2 ~ ... and tv -+ + 00 as v -+ 00.]
(c) lim
,~oo
f' 0
sin x dx
x
=~.
2
(d) foo Isin x IdX = +
0 x
00.
[Hints: From (a) and (b), c/J(t) = n/2 - tan -1 t.] Integrate by parts to show that
11 00
f(x, t)dx I ~ ~ for suitable c.
By the dominated convergence theorem,
lim
t-+O +
f'0
f(x, t)dx = f'
0
-
x
x dx.
sin-
239
5 Integration
5.13 LP-Spaces
The concept of normed vector space is defined in Section 2.9. The spaces
~(S) of bounded, continuous functions in Section 2.10 furnish one interesting
class of infinite dimensional normed vector spaces. We now consider another
class, the spaces U(A), which are defined as follows.
240
5.13: U-Spaces
s- 1
(**) ~ [F(X)]l/p L 2- k+r < 21-k[F(x)]l/p.
r=O
Since the right-hand side approaches 0 as k --+ 00, the sequence of real
numbers gl(X), g2(X)1 ... is Cauchy. Let
f(x) = lim gk(X). if x E A - B,
k-oo
241
5 Integration
If v;;::: Nk> then Ilgk - fvllp < 2- k - klp (recall the definition of gk)' Using
Minkowski's inequality
Ilfv - flip ~ I~~ - gkllp + Ilgk - flip.
Since the right-hand side tends to ° as k ~ 00, this shows that
t llp ~ 1 + ~p (t - 1) = ~ + ~.
p p
Setting t = uPv- P', where u ;;::: 0, v > 0, we find since 1 - p' = - p'lp that
uP vP'
(5.62) uv ~ -p + "p
°
Obviously, this inequality also holds when v = 0.
If I f I p = 0, then f(x) = almost everywhere in A and both sides of
Holder's inequality are 0. Similarly both sides are if Ilgllp' = 0. Suppose °
that Ilfllp > 0, Ilgll p' > 0, and let
1 = (1/llfll p )f,
Then
II liP dV = IlgIP' dV = 1,
242
5.13 LP-Spaces
By Holder's inequality,
I· g = {Ig dv".
243
5 Integration
PROBLEMS
1. Let f(x) = Ix 1-'. Show that if A is the unit n-ball, then f E U(A) for p < n/a. but not
for p ~ n/a.. [Hint: See Section 5.6, Example 4.]
2. Letf(x) = x- I (logX)-2,A = (O,t).ShowthatfEU(A)forp = l,but not forp > 1.
3. Let 1 ::; q < p. Using Holder's inequality show that if A has finite measure, then
f E U(A) implies that f E U(A). Give an example to show that it is necessary to
assume A has finite measure. [Hint: Apply HOlder's inequality to the functions Ifl q
and 1 with p replaced by p/q.]
4. Prove properties (a) and (b) for the p-norm.
5. Let A = [0, 1].
°
(a) Describe a sequence I I, 12 , ••• of closed intervals such that V1(I v) --+ as v --+ 00,
°
but every x E [0,1] belongs to infinitely many Iv'
°
(b) Let fv(x) = 1 for x E Iv> fv(x) = for x ¢: Iv' Show that I fvll p --+ as v --+ 00, for
1 ::; p < 00, but fv(x) tends to a limit as v --+ 00 for no x E [0, 1].
244
6
Curves and line integrals
6.1 Derivatives
Let g be a function from a set J c El into En. Let t be an interior point of J.
Then the derivative of g at t is the vector
. 1
(6.1) g'(t) = hm - [g(t + u) - g(t)],
U-'O u
245
6 Curves and line integrals
Here f . g is the real-valued function whose value at each t is the inner product
f(t)· g(t).
The derivative has a geometric interpretation as a tangent vector. Let
us suppose that 1 is an interval. As t traverses 1 from left to right, the point
g(t) traverses some curve in En. A precise definition of the term "curve" is
given in Section 6.2.
Let us assume that to is a point of 1 at which g'(to) # O. Let
Xo = g(to), Vo = g'(to),
x=g(t), y=xo+uv o ,
where t = to + u and Iu I is small enough that tEl. The ratio of the distances
I x - y I and I x - Xo I may be written, upon multiplying numerator and de-
nominator by III u I, in the form
Ix-yl
Ix - xol
11
= ~ [g(t)
'I 1
- g(to)] - g(to) IOlu)[g(t) - g(t o)] I'
Hence
. Ix - yl 1
hm---- = 0 - , - = o.
u ... olx-xol Ig(to)1
This justifies calling Vo a tangent vector at Xo, and the line through Xo and
Xo + Vo a tangent line at Xo (Figure 6.1). Note that we have used the assump-
tion that g'(t 0) # O.
Figure 6.1
The number t is often called a parameter. It need not have any geometric
or physical significance. However, if n = 3 and t happens to denote time in
a physical problem, then g'(t) is the velocity vector.
A vector-valued function g has components gl, ... , gn, which are the
real valued functions such that
n
g(t) = L g'(t)ei
i= 1
for every tEl. Ifg'(t) exists, then the ith component U-1[gi(t + u) - gi(t)]
of the expression on the right side of (6.1) tends to gi'{t) as u -+ 0, by Propo-
sition 2.2, and
n
(6.3) g'{t) = L gi'(t)e i ·
i= 1
246
6.2 Curves in En
Conversely, if gi'(t) exists for each i = 1, ... , n, then g'(t) exists and is given
by (6.3).
PROBLEMS
1. Find the tangent line at 2- I / Ze l - 2 1/ Ze 1 to the ellipse represented by g(t) =
(cos t)e l + (2 sin t)e z , J = [0, 2n]. Illustrate with a sketch.
6.2 Curves in En
Let g be a function from an interval 1 c E1 into En. Then g(t) traverses a
curve in En as the" parameter" t traverses 1. It is better not to call g itself
a curve. Instead one should regard any vector-valued function f obtained from
g by a suitable change of parameter as representing the same curve as g. We
define a curve as an equivalence class of equivalent parametric represen-
tations. To simplify matters we first consider only curves with continuously
changing tangents.
Let us now be more precise. Let us for simplicity assume that 1 = [a, b],
a closed bounded interval, and that the components gt, ... , gn are of class
e(1) on [a, b]. By gi'(a) and gi'(b) we mean, respectively, right-hand and left-
hand derivatives. They are equal to the derivatives at a and b of any class e(1)
extension of gi to an open set containing [a, b].
247
6 Curves and line integrals
EXAMPLE I. Let g(t) = tel + tZe z , 1 ::; t ::; 2. Then gl(t) = t, gZ(t) = t Z,
g'(t) = e l + 2te z #- O. Hence g is a parametric representation of class C(1) on
the interval [1, 2]. In fact, it represents the arc of the parabola y = X Z between
(1, 1) and (2,4), traversed from left to right (Figure 6.2). If we let f(r) = (exp r)e l
+ (exp 2r)e z ' 0'::; r ::; log 2, then f also represents this same parabolic arc.
In effect, fis obtained from g by the parameter change t = exp r. It is reason-
able to regard f and g as equivalent, and we do so.
y
I
I
(2,4)
\
,, /
/
I
---~""-+-","------- x
x = t, Y = t2
Figure 6.2
Now let g be any parametric representation of class C(l) on [a, b]. Let ¢
be any real-valued function of class C(1) on some closed interval [ex,f3] such
that
(6.4) ¢'(r) > 0 for every r E [ex, 13], ¢(ex) = a, ¢(f3) = b.
Let f be the composite of g and ¢, denoted by f = g ¢. Then
0
248
6.2 Curves in En
Definition. We say that f is equivalent to g if there exists <p satisfying the above
conditions such that f = g <p. 0
By requiring that the components gl, ... ,gn be of class C(q), q ;;::: 2, and
allowing only parameter changes <p of class C(q), the notion of curve of class
C(q) can be defined in the same way. To study curvature of curves one needs
to assume class C(2) at least [25]. However, for present purposes we need only
class C(l). From now on we say "curve" instead of "curve of class C(l)," and
"representation" instead of "parametric representation of class C(1)."
Each curve has an infinite number of representations. If g is one such
representation, then each parameter change <p leads to another. It is often
highly advantageous to make a judicious choice of parameter. In a physical
problem, time (measured according to some preassigned scale) may be the
preferred parameter. For certain curves one of the components Xl, ... , Xn
can be taken as the parameter. For example, if the first component gl'(t) of
the tangent vector g'(t) is everywhere positive, then gl has an inverse (Section
A.4). Let us take for <p the inverse of gl. Formally this amounts simply to
solving the equation Xl = gl(t) for t, obtaining t = <p(x l ). Set r = Xl. Then
Xl is the new parameter andfl(x l ) = Xl. Figure 6.3 illustrates this situation
for n = 2.
A curve y is to be regarded as the path traversed by a moving point, and
we have not excluded the possibility that y passes through the same point
several times. The multiplicity of a point x is the number of points t E [a, b]
such that g(t) = x. The multiplicity does not depend on the particular
representation g chosen for y, since any <p satisfying (6.4) is a univalent
XJ
I (x,f(x)) I
I I
--~--~aL-------7b----x
Figure 6.3
249
6 Curves and line integrals
g(b)
Figure 6.4
function, namely, </J( r I) of- </J( r 2) ifr I of- r 2' The trace of y is the set of points
of positive multiplicity, that is, the set of points through which y passes at
least once. If x has multiplicity 1, then x is called a simple point. If every
point of the trace is simple, then y is called a simple arc.
The point g(a) is called the initial endpoint of y and g(b) the final endpoint.
If g(a) = g(b), then y is called a closed curve. A closed curve is called simple
if every point of the trace is simple except g(a), which has multiplicity 2
(Figure 6.4).
(6.6) 1= flg'(t)ldt.
If f is equivalent to g, then
by (6.5) and the theorem about change of variables in integrals (Section A.3).
Thus I does not depend on the particular representation chosen for y.
250
6.2 Curves in E"
°
the velocity vector, that is, the speed of motion.
Since S'(t) > the equation s = S(t) can be solved for t. More precisely,
the function S has an inverse tjJ of class C(l) on [0,1]. Let G = go tjJ. Then G
is the standard representation of y. From (6.5)
Ig'[tjJ(s)] I'
251
6 Curves and line integrals
we find that
(6.7a) IG'(s) I = 1
for every s E [0, I]. Hence G'(s) is a unit tangent vector at the point G(s). If we
write dxilds for Gi'(s), then (6.7a) can be rewritten
(6.7b)
EXAMPLE 3 (continued). Let m > O. Then I g'(t) I = m, S(t) = mt. Solving the
equation s = S(t) for t, we obtain the standard representation
G(s) = (cos s)e l + (sin s)e 2 , o ::; s ::; 2mrr.
Piecewise smooth curves
It is not difficult to adapt the preceding discussion to curves of class C1)
except for a finite number of corners and cusps. By a parametric representation
of a piecewise smooth curve is meant a continuous function g on an interval
[a, b] with the following property: There exist to, t l , ... , tp with
PROBLEMS
1. Which of the following represent simple arcs? Simple closed curves? Illustrate with
a sketch.
(a) g(t) = (a cos t)e l + (b sin t)e2, a> 0, b > 0, J = [0, 2n].
(b) Same as (a) except J = [-n, 2n].
(c) g(t) = (-cosh t)e l + (sinh t)e 2 , J = [ -I, 1] (see Section 3.5 for the definition
of cosh and sinh).
252
6.3 Differential I-forms
2. (a) Let I' be represented by f(x) = xe l + f(x)e z , a ::; x ::; b, where f is of class e(1)
r
on [a, b]. Show that
1= Jl + [f'(x)r dx.
a
253
6 Curves and line integrals
Thus
(6.9)
For short we usually say" I-form" instead of" differential form of degree
1." In Chapter 7 differential forms of any degree r = 0, 1,2, ... ,n are defined.
The value of co at x is denoted by co(x). It is the covector
(6.10)
where, as in Section 3.2, e l , ... ,en are the standard basis covectors.
A I-form co is a constant form if there is a covector a such that c.o(x) = a
for every XED. In particular, for each i = 1, ... , n let us consider the con-
stant I-form with value e i . This I-form is denoted by dXi. Since (En)* is a
vector space, the sum co + ~ oftwo functions co and' with the same domain D
and values in (En)* is defined (Section 2.1). Similarly, the productfco is defined
iffis a real-valued function and co is a I-form, with the same domain D. In
particular, widx i is the I-form whose value at each x is wi(x)e i. From (6.10),
WI dx l + ... + Wn dxn is the I-form whose value at each x is co(x). Therefore
Formula (6.8) is correct.
254
6.3 Differential I-forms
Some authors define df as the real valued function whose domain is the
cartesian product D x En and whose value at each pair (x, h) is the number
df(x) . h. Knowing df(x), one can find df(x)' h for every hE En, and vice
versa. Hence this definition is equivalent to the one we have given.
Iffand g are differentiable functions with the same domain D, then
d(f + g) = df + dg, d(fg) = f dg + g df
These formulas follow from Problem 7, Section 3.3. Similarly, writing e for
the constant function with value e,
de = 0
where 0 denotes the "zero form" whose value is 0 everywhere. If D is con-
nected, then, conversely, df = 0 implies that f is a constant function. This is
just a restatement of Corollary 2, Section 3.3.
If L is a linear function, then dL is a constant I-form. For let L(x) = a' x,
as in Proposition 3.1. Then the ith partial derivative of Lis ai and dL(x) = a
for every x.
In particular, the standard cartesian coordinate functions Xl, ... , xn
(Section 3.2) are linear. In fact Xi(X) = e i • x = Xi, and dXi(x) = e i . Hence
dX i is just the constant I-form which we have denoted by dXi. The common
practice of writing dx i instead of dX i arises from the habit of confusing nota-
tionally a function with its value at some particular point x, in this case of
confusing Xi with Xi = Xi(X). Nevertheless, following custom, we adhere to
the notation dXi.
255
6 Curves and line integrals
In (6.11) we may as well suppose that i < j. Thus the definition says in
effect that ro is a closed differential form if its components W 1 , . .. ,Wn satisfy
these n(n - 1)/2 conditions. For instance, if n = 2 let us write dx and dy
instead of dx 1 and dx 2, and
M(x, y) = w 1(x, y), N(x, y) = w 2(x, y).
The expression for a I-form is then
ro = M dx + N dy.
The condition that ro be closed is that the components M and N satisfy
oM oN
oy ox'
We have shown that every exact I-form is closed. The converse is false,
as Example 2 (below) shows. It is comparatively easy to check whether con-
ditions (6.11) are satisfied. Therefore, it is very desirable to find some addi-
tional condition that will guarantee that the converse holds. Such a condition
is that the domain D be simply connected, a term to be defined in Section 8.10.
We prove in Chapter 8 that if D is simply connected, then every closed I-form
with domain D is exact. It turns out that any convex set, and in particular P,
is simply connected. For n = 2, an open, connected set D is simply connected
if and only if, roughly speaking, D has no holes.
We have been careful to distinguish notationally between functions and
their values. One can scarcely attain a sound knowledge of calculus until this
distinction is recognized. Nevertheless, in examples we sometimes abuse the
notation for brevity. For instance, d(x 2y) = 2xydx + x 2 dy is short for the
statement "df = f1 dx + f2 dy, wheref(x, y) = X2y,j1(X, y) = 2xy,jix, y) =
x 2 for every (x, y) E £2."
for every (x, y). Hence ro is a closed I-form. Since E2 is connected and simply
connected, ro = df where f is determined up to the addition of a constant
function. The functionfcan be found by partial integration with respect to the
first variable, as follows:
of
ox (x, y) = M(x, y) = 2xy, f(x, y) = x 2y + 4>(y),
256
6.3 Differential I-forms
Of course these equations hold for every (x, y) E E2. Then ¢'(y) = 2y, and
¢(y) = y2 + c for every y, where the "constant of integration" c is a number
that may be chosen arbitrarily. Hence for every (x, y) E E2,
!(x,y) = x 2 y + y2 + c.
y
M(x,y) = - 2 2'
X +y
x
N(x,y) = 2 2'
X +y
A computation shows that oM/oy = oN/ox in D, hence co is closed. Let us
show that co is not exact. Let Dl be the open subset of D obtained by deleting
°
the positive x-axis. For every (x, y) E Dl let 0(x, y) be the angle from the
positive x-axis to (x, y), < 0(x, y) < 2n (Figure 6.5). Using elementary
calculus, we find that in D 1 , d0 = co [Problem 6(a)]. If there were a function
! of class e(2) on D such that co = df, then upon restricting! to D 1 we would
have d(f - 0) = O. Since Dl is connected,f - 0 would be constant on D 1 •
This would imply that 0 can be continuously extended across the positive
x-axis, which is false. Hence co is not exact.
(0,0)
Figure 6.5
257
6 Curves and line integrals
PROBLEMS
1. Let n = 1. Give a precise interpretation of the formula df/dx = f' from elementary
calculus. [Hint: The quotient of two real valued functions is defined wherever the
denominator does not have the value 0.]
i= 1
is an exact I-form. [Hint: Let h(u) = ug(u) for every u EEl. The function h has an
antiderivative.]
By reasoning such as that indicated in Section 6.2, the sum (*) tends to the
integral (6.12a) as Il -+ O. This integral is called a line integral.
258
6.4 Line integrals
The line integral exists, since if g is piecewise of class C(1) the integrand
in (6.12a) is bounded and has a finite number of discontinuities. Iff is equiva-
f:
lent to g, then using (6.5)
Hence the line integral does not depend on the particular representation
g chosen for y.
Line integrals have an important role in many parts of mathematical
analysis, such as in the theory of complex analytic functions. Several funda-
mental physical concepts are also expressed in terms of line integrals.
The concept of work is mentioned at the end of the present section. In Section
6.6 we discuss thermal systems, and in Section 8.5, circulation in a flowing
fluid.
L
The notation for line integral is roo Writing out the scalar product
in (6.12a),
(6.12b)
EXAMPLE I. Let y be the semicircle with center (0,0) and endpoints ± ae2'
directed from - ae 2 to ae 2 . Let us evaluate fl' x dy - y dx. Points (x, y) of the
semicircle satisfy the equations x = a cos 0, y = a sin 0, where - n/2 ~
o ~ n/2. The most convenient representation for y is on [ - n/2, n/2] with
g(O) = (a cos O)e l + (a sin O)e2' Then
f
y
(x dy - y dx) = [ 12
-1</2
(dY dX)
x dO - Y dO dO =
[12
-1<12
a 2 dO = a 2 n.
259
6 Curves and line integrals
for any pair of I-forms ro, ~, and scalar c. Let g represent Y on [a, b], and let
</> be of class C(1) on [0:,13] with
(6.14) f - Y
ro - -fro
l'
Let Yi"'" Yp be piecewise smooth curves such that the final endpoint
of Yj is the initial endpoint of Yj+ 1 for j = I, ... , p - 1. Let Y be obtained
by "joining together" the curves Yi"'" Yp' More precisely, let us divide
[0,1] into p subintervals [U - l)/p,j/p] of the same length. Each curve Yj
has a representation on an interval [aj' bj]' By a linear change of parameter
we may assume that aj = (j - 1)/p, bj = j/p. Let gj be such a represen-
tation of Yj for each j = 1, ... , p. Then glJ/p) = gj+iU!P). Let g be the
function such that g(t) = git) for t E [(j - l)/p, j/p]' Then g is a parametric
representation that is piecewise of class C(1), and Y is the curve that g repre-
sents. Let us call Y the sum of these curves and write Y = Yi + ... + Yp'
Since an ordinary integral over [a, b] is the sum of the integrals over the sub-
intervals [(j - l)/p, j/p] , we have
(6.\5) f Y'+"'+Yp
co = f y,
co + ... + f ro.
Yp
f Y
ro = I 4
j=i
f
Yj
ro.
y
Y3
d •
Y4 f . fY2
C
y,
x
a b
Figure 6.6
260
6.4 Line integrals
Let us now consider the case when ro is an exact differential form, ro = df,
where J is of class C(l) on an open set D containing the trace of y. By the
definition (6.12a),
{ dJ = f dJ[g(t)J . g'(t)dt.
fu
of calculus,
Xo
Figure 6.7
261
6 Curves and line integrals
PROOF. We have seen that (1) implies (2) in Theorem 6.1. If Yl and Y2 have the
same endpoints, then Yl - Y2 is closed. If (2) holds, then
o= f
VI-)'2
ro = f f
VI
ro -
)'2
roo
f(x) = fro,
)'
where Y is any curve lying in D with initial endpoint Xo and final endpoint x.
Since we are assuming (3) in Theorem 6.1, it does not matter which curve with
these properties is chosen. Let us show that df = roo
xo
)' I
x x + ue l
Figure 6.8
Let gj(t) = x + tej, !/1m = wlx + teJ Then gj represents Yj on [0, u], and
f
hence
i = 1, ... , n.
Therefore df(x) = ro(x). Since this is true for every xED, ro = df D
262
6.4 Line integrals
Work
Let D be an open connected subset of E3. In mechanics the idea offorce field
is considered. A force field assigns at each XED a linear function, which we
call the force covector acting at x and denote by ro(x). If h is a "small dis-
placement" from x, then the work done moving a particle along the line
segment from x to x + h is approximately ro (x) . h. The force field is the
differential form ro of degree I whose value at each XED is the force co vector
ro(x).
For present purposes it is more natural to regard force as a co vector rather
than a vector. However, one can also consider the force vector F(x) =
wl(x)e l + w2(x)e 2 + w3(x)e3 with the same components as ro(x). This
simple device for changing covectors into vectors is justified since we use the
standard euclidean inner product. If En is given another inner product
B(x, y), then the components of ro(x) and F(x) would be related by
Wi = I cijFj,
j= 1
where (Ci) is the matrix of Band (c ij) = (ciT I. Compare with (2.2) and
(3.12).
Let y b" a piecewise smooth curve lying in D. Using the notation in
Section6.2,withforsakeofsimplicitysj = tj_l,thevectorhj = g(t) - g(tj-d
is a displacement from g(t j _ I ), which is small if f.1 is small. The work done
going along y from g(t j _ l ) to g(t j } should be approximately ro[g(t j _ d] . hj'
This suggests the following.
w = fro.
)'
w = {ro[G(S)]' G'(s)ds.
The expression ro[G(s)] . G'(s) is called the component of the field at G(s) in
the direction of the unit tangent vector G'(s) to y.
A force field ro of class e(l) is called conservative if ro is closed. In Section
8.10 we define the concept of simply connected open set. It is shown there
that any closed I-form ro is exact if D is simply connected. If ro is exact and
ro = dj, thenfis a potential of the field roo If Dis connected,fis determined up
to the addition of a constant function.
263
6 Curves and line integrals
PROBLEMS
1. Evaluate -!- Lx
dy - y dx in case:
(a) y bounds the triangle shown in Figure 6.9.
(b) y is represented by g(t) = (a cos t)el + (b sin t)e2, 0 ::; t ::; 2n, where a, b > O.
Your answer should be the area of the set enclosed by y. This is a very special case
of Green's theorem (Section 8.7).
(0, O)-----~(a, 0)
Figure 6.9
2. Let gl(t) = tel + (2t - l)e 2, I::; t ::; 2, and g2(t) = (t + I)e l + (t 2 + t + 1)e2,
0::; t ::; 1. Evaluate L. (0 and L, (0 for each of the I-forms (a), (c), and (d) in Problem
3, Section 6.3, where YI and Y2 are the curves represented by gl and g2, respectively.
r
3. Let n = I, (0 = f dx, and Y be the interval [a, b] directed from a to b. Show that
fy
(0 =
a
f(x)dx.
4. Let D c £2 be open and simply connected, and u, v functions of class C( I) that satisfy
au av av au
ax ay' ax - ay"
[Note: These two first-order partial differential equations are called the Cauchy-
Riemann equations. They are fundamental to the theory of complex analytic
functions.]
Show that for any closed curve y lying in D,
f y
u dx - v dy = Q, f y
v dx + u dy = O.
5. Show that the force field (0 = l/I(p2)(X dx + y dy + z dz) is conservative, where 1/1 is
any function of class C(I) on £1 and p2 = x 2 + y2 + Z2. Find its potential f, if
f(O) = o.
(Other line integrals.) The following problems deal not with integrals of
1-forms, but with some other types of line integrals that often occur.
264
6.5 Gradient method
I f ds = ff[g(t)]lg'(t)ldt.
y •
Show that this integral does not depend on the particular representation g chosen
for y. In particular, if arc length is the parameter, then
I ).
f ds = II 0
f[G(s)]ds.
L.
7. The moment of inertia of a curve y about a point Xo is Ix - Xo 12 ds. Find the moment
of inertia about 0 of the line segment in E3 joining e l and e 2 + 2e3'
8. The centroid x of a curve y is the point such that
_. Jy Xi ds
x'=-- i = 1, ... ,n,
I '
f f y
W =
•
W[g(t), g'(t)]dt.
Show that:
(a) This integral does not depend on the particular representation g of y.
L.
(b) If W(x, h) = ro(x)' h, then L W = ro: and if W(x, h) = f(x)lhl, W = L f ds. L.
(c) Let W(x, h) = Ilhll, where II II is any norm on En (Section 2.11). Then W is L.
called the length of y with respect to this norm. Show that if y is the line segment
joining Xl and x2 , then the length is Ilxl - x211.
i = 1, ... , n,
265
6 Curves and line integrals
Figure 6.10
Suppose that grad f(x o) # O. Since grad f is continuous there exists a neigh-
borhood U of Xo and m > 0 such that 1grad f(x) 1 ~ m for every X E U. There
exists t 1 such that g(t) E U for every t ~ t 1. By the fundamental theorem of
266
6.5 Gradient method
4>(t2) = 4>(tl) + f
II
'24>,(t)dt ~ 4>Ud + m2 (t2 - ttl·
The right-hand side tends to +x as t2 ---+ + x, but 4>(t2) ::os: f(x o). This is a
contradiction. Hence grad f(x o) = O. 0
PROOF. By Problem 8, Section 3.5, there exists In > such that Q(x o , h) ::os:
~ m 1h 12 for every h. Since f is of class C(2) there is a neighborhood U of Xo
°
Xo + h EO U and s EO (O, 1). Since f;(x o) =
Si EO (0, 1) such that
1
°
such that fJxo + sh) - h/xo) ::os: m/2n2 for i, j = 1, ... ,11, whenever
1
If X EO U and h = X - x o , then
m
L
n . .
gradf(x)' (x - x o) = f;)xo + Sih)hlhJ ::os: - -lhI 2 .
i.j= I 2
Now let l/1(t) = 1 g(t) - Xo 12 , the square of the distance from x o , Taking
x = g(t) and using (6.17), we have provided g(t) EO U
1/1'(t) = 2[g(t) - xo]' g'(t)::os: -mlg(t) - x 0 1
2,
267
6 Curves and line integrals
Figure 6.11
Figure 6.10 indicates the behavior of the level sets and gradient trajectories
near a nondegenerate maximum. The situation is similar near a nondegenerate
minimum. The gradient trajectories in that case are followed as t --+ - 00.
Near a saddle point (n = 2) the behavior of the trajectories is indicated
in Figure 6.11.
whose solutions are gl(t) = Xl exp(2t), g2(t) = Yl exp( -2t). The trajectories
lie on the hyperbolas xy = k orthogonal to the level sets x 2 - i = c, and on
the coordinate axes. Only trajectories starting from points (0, Yl) lead to the
saddle point (0,0).
PROBLEMS
268
6.6 Integrating factors; thermal systems
Definition. A function 4> is an integrating factor if 4>(x) #- 0 for all XED and
4>ro = dffor some functionf
EXAMPLE I. Let D = {(x, y): x > O}, ro = 2y dx + x dy. Then ro is not exact,
but
xro = 2xydx + x 2 dy = d(x 2 y).
If 4>ro = df and 4>(x) #- 0, then the curves on which ro = 0 are just those
on whichf(x) is constant. To see this, note thatf[g(t)] is constant if and only
if df = 0 on y. But df = 0 on y if and only ro = 0 on y.
Note. In the terminology of differential equations, a curve y on which
ro = 0 is called a solution curve for the differential equation ro = O. Such a
differential equation is called of Pfaffian type. Since ro(x) #- 0, we have also
df(x) #- 0 if there is an integrating factor 4>. Thus the level sets Be =
{x ED: f(x) = c} are (n - I)-manifolds (see Section 4.7). Every solution
curve lies in some level set Be.
Let us now assume that w 1(x) = 1, i.e., that ro has the special type
n
(6.19) ~ = L Wi dxi .
i= 2
269
6 Curves and line integrals
(6.20) xl - x6 = - f~·
1"
(a) Ifro has an integrating factor in QI' then ro satisfies property (P) in QI'
(b) If ro satisfies property (P) in QI, then there exists an open set ~ containing
x* such that ~ c Q I and ro has an integrating factor in~.
For proof of Theorem 6.2(b), to simplify the notation let us suppose that
x* = O. Suppose that ro has property (P) in QI' Let us define F(u, z) for (u, z)
in a suitable neighborhood Q 2 of 0 as follows (Q 2 c Qd. Let y be any curve
lying in QI, such that ro = 0 on y and y has endpoints (u, 0), (Xl, z) (see Figure
6.12). We set F(u, z) = Xl. Let us show that F(u, z) does not depend on the
particular choice of y. Suppose that y is another such curve, with endpoints
(u, 0) and (Xl, z). The curve YI obtained by traversing y in the opposite sense
followed by y is of type .r. By Property (P), xI = XI.
270
6.6 Integrating factors; thermal systems
__________~~~-----------Xl
(U,O)
Figure 6.12
In order to show that F(u, z) is of class e(l), we make a particular choice for
'Y and appeal to a theorem about differential equations. This curve 'Y is
represented on [0,1] by g(t) = (gl(t), tz), where gl is the solution of the dif-
ferential equation
dg 1
L Wj[g o ::;; t
n 1 .
- = - (t), tz]x', ::;; 1,
dt j=2
with gl(O) = U, Z = (x 2 , ••• , xn). Then (0 = 0 on 'Y and 'Y has endpoints
(u,O), (gl(l), z). Thus F(u, z) = gl(l). By [11, Section 1.3], F is of class e(1)
on a suitable neighborhood il2 of O.
Let us next show that
(6.21 )
of
ox j (u, z) = -
I
Wj(x , z), i = 2, ... ,n,
where Xl = F(u, z). For this purpose, given v > 0 sufficiently small, consider
the curve represented on [0, 1] by G(t) = (GI(t), z + tvej), where
dG I
dt = -Wj[G1(t), Z + tve;]v,
with GI(O) = Xl. Then (0 = 0 on this curve; its endpoints are (Xl, z), (G I(1),
z + ve;). Thus F(u, z + vej) = GI(l). By the mean value theorem,
F(u, z + ve;) - F(u, z) GI(l) - GI(O)
v v
= -Wj[GI(s),z + sve;]
for some s E (0, 1). When v --. 0 we get (6.21).
Let (0# be the I-form obtained by substituting, in (6.19), Xl = F(u, z) and
replacing dx 1 by dF. Then by (6.21),
#
= dF + L.. wjdx = -duo
~ j of
(0
j=2 au
Since F(u, 0) = u, we have
(6.22)
271
6 Curves and line integrals
Consider the transformation F(u, z) = (F(u, z), z) from O 2 into En. By (6.22),
JF(u, 0) = 1. The inverse function theorem implies that 0 has a neighborhood
o c O2 such that FlO has an inverse of class C(1), defined on an open set
~ = F(O) containing F(O) = O. The inverse has the form (FIO)-I(X) =
(f(x), z), where f(x) is that u for which F(u, z) = Xl. We set
of 1
r(x) = ou (f(x), z), 4>(x) = r(x)·
Thermal systems
The subject of thermodynamics is concerned with relationships between heat
and other forms of energy such as mechanical, chemical, electrical, or
magnetic. It deals with gross effects such as temperature and pressure,
without reference to the underlying atomic structure of the matter being
studied.
The following mathematical model is used to describe various thermal
systems [16]. A possible state of the system is a vector x = (Xl, ... ,xn). The
number Xl represents the total energy of the system, and x 2 , • •• ,xn other
quantities such as volume and mass. Let D denote the set of possible states.
It is assumed that D is open, convex, and that xED implies ex E D for any
c > 0. A I-form ro = dx l + ~oftype(6.l9)isgiven. The coefficientsw2' ... ,Wn
are functions on D homogeneous of degree 0, namely, Wi(CX) = wi(x) for
C > 0. [Note: In thermodynamics literature the components Xi are called
extensive variables and the functions Wi' intensive variables.] The physical
significance of ~ is as follows. If a thermal system in a state x is displaced to a
new state x + h, then for Ih I small, ~(x)· h represents approximately the
work done (cf. Section 6.4). Let us now suppose that the states of a thermal
system change with time t, and let x = g(t) be the state at time t E [a, b].
L
Let y be the curve represented by g. Then ~ is the work done in moving the
thermal system along y. On the other hand,
L dx l = gl(b) - gl(a)
L
is the change in total energy. The sum ro is interpreted as the heat energy
transferred to the system from outside it.
The curve y is called adiabatic if ro = 0 on y. In other words, no heat is
transferred as the thermal system states move along an adiabatic curve y.
As above, let Z = (x 2 , ••• , xn) denote the vector of system states, omitting
energy Xl, and denote by Xo = (x6, Zo), Xl = (xL Zl), the initial, final states
as the system moves along y. It is assumed that:
272
6.6 Integrating factors; thermal systems
(6.23) f f
y
ro = k
y
dS = k[S(x,) - S(xo)]·
We have
aT aIJ
ax' = r'[IJ(x)] ax' > O.
where P is pressure and J,l is called the chemical potential. If ro = T dS, then
¢= liT is an integrating factor and
as as P as J,l
au T' av T' aN T
Since S is homogeneous of degree 1, Euler's formula (Problem 8, Section 3.3)
gives
Therefore
(6.24)
S = U + PV - J,lN
T .
For an ideal gas, it is required that
U U
(6.25) P = a V' T=h-
N'
for suitable constants a, h.
Formula (6.18) gives three equations that ¢ = liT must satisfy. One of
these holds by (6.25). The other two equations determine J,l, and then S,
from (6.24):
a V 1 U
(6.26) S = -Nlog- + -Nlog- + CN
b N b N
where C is any constant (see Problem 4).
PROBLEMS
[In other words, for fixed u, F(u, y) satisfies the differential equation dx/dy =
- N(x, y).] Following the last part of the proof ofTheorem 6.2 (b), show that co = r df
in some open set ~ containing (x*, y*).
3. Let n = 2 and co = dU + P dV = T dS.
(a) Show that S = T-1(U + PV).
(b) Find Tif P = aUV- 1 •
4. In Example 2, derive (6.26).
274
Exterior algebra and
differential calculus
7
275
7 Exterior algebra and differential calculus
276
7.1 Covectors and differential forms of degree 2
(h, k). For fixed h, let B(h, ) denote the function whose value at k is B(h, k).
Similarly, for fixed k, B( , k) is the function whose value at h is B(h, k). The
function B is bilinear if B(h, ) and B( , k) are linear functions for every h, k.
We have
n n
h= L hiej, k = L kie i .
i= 1 j= 1
If B is bilinear, then
n n
B(h, k) = L hiB(ej, k) = L hikjB(ej, eJ
i= 1 i.j= 1
Let
(7.1) i,j=l, ... ,n;
then for every (h, k)
n
(7.2) B(h, k) = L Wijhik j.
i, j= 1
or
(7.3) B(h, k) = L wiihikj - hjk i).
i<j
Conversely, given n(n - 1)/2 numbers Wij, i < j, Formula (7.3) defines an
alternating bilinear function.
It is convenient to introduce a different name and notation for alternating
bilinear functions.
277
7 Exterior algebra and differential calculus
The function eij is alternating and bilinear, for any i,j = 1, ... , n. Thus eij
is a 2-covector. We also have
(7.4)
Since (7.3) holds for all (h, k), we have upon writing ro instead of B to denote a
2-covector,
(7.5) ro = L.
" wijeij.
j<j
Conversely, given n(n - 1)/2 numbers Wjj' i < j, Formula (7.5) defines an
alternating bilinear function. The 2-covectors eij, i < j, are linearly indepen-
dent (Problem 2). These n(n - 1)/2 2-covectors form a basis for (E'2)*, called
the standard basis. In particular, (E 2)* is of dimension n(n - 1)/2.
a = L aje j, b = L.
" b·e
} '
j
i= I j= I
we then set
n
The values of ro are 2-covectors. The same Greek letters ro, ~, 1], ... used
above to denote 2-covectors are now used to denote differential forms.
278
7.1 Co vectors and differential forms of degree 2
One way to get a 2-form is by multiplying two I-forms, using the product
A. If ~ and 1) are I-forms, then ~ A 1) is the 2-form obtained by multiplying
their values at each x:
(~ A 1))(X) = ~(x) A 1)(X).
In particular, dXi denotes the I-form, whose (constant) value at any x is the
I-covector ei . Then dx i A dx j has the constant value ei A ej = eij. By (7.5),
ro(x) = L wij(x)eij
i<j
By (7.4), we have
(7.7)
We have
279
7 Exterior algebra and differential calculus
In Section 6.3 we defined the concepts of closed and exact differential forms
of degree 1. By (6.11) and (7.8), a I-form co of class e( l) is closed if and only if
dco = O. Iffis of class e(2), then dfis exact (hence closed) and of class e(l).
Therefore d(df) = O.
dco = dM /\ dx + dN /\ dy = ( oN
ox -
OM)dx
a:y /\ dy.
The definition of dco given above may at first seem rather arbitrary. We
show at the end of Section 7.4 that the exterior differential is uniquely speci-
fied by a few reasonable properties that may be required of it. However, let
us now show how to arrive at dco in another way, starting from a special case
of Stokes's formula. We require that Stokes's formula hold for any parallel-
ogram (in Section 8.8 it is proved under much more general circumstances).
The following discussion is intended purely as motivation, and can be
omitted if desired. Consider a parallelogram K with vertices x o , Xo + h,
Xo + k, Xo + h + k, and let
g(s, t) = Xo + sh + tk.
Then K = g(I), where I = [O,IJ x [0, IJ is a square in the (s, t) plane. The
pair of vectors (h, k), written in the given order, assigns an orientation to K
(we postpone a formal definition of the concept of orientation until Section
7.5). The integral of a 2-form 'lover an oriented parallelogram K is denoted
by SKo'l. It is defined as follows. Let 'I(x) be the value of 'I at x; 'I(x) is a 2-
covector, that is, an alternating bilinear function, which assigns to the pair
(h, k) the real number 'I(x)(h, k). Let
280
7.1 Co vectors and differential forms of degree 2
Figure 7.1
The first integral on the right side is the contribution from the second and
fourth segments of y, starting from xo, and the second integral the contri-
bution from the first and third segments. Stokes's formula asserts that
(7.9) f)'
ro = (0'1,
JK
for a suitable 2-form 'I. Let us show that 'I = dro.
Now
n
ro(x). k = L wi(x)k i
i= 1
where W 1, •.• , Wn are the components of the I-form roo By the fundamental
theorem of calculus
to
Wi[g(1, t)J - Wi[g(O, t)J = Jo os Wi[g(S, t)]ds
t og (1
= Jo dWi [g(S, t)] • os ds = Jo dWi[g(S, t)] . h ds.
Similarly,
281
7 Exterior algebra and differential calculus
By comparing these formulas with (*) and (**) we see that, if(7.9) holds for any
parallelogram K, we must require
n
T)(x)(h, k) = L {(dWi(X)' h)k i - (dWi(X)' k)hi}.
i= 1
But then
T)(x)(h, k) =
i= 1
f {± ow;
ox
(hik i -
i= 1
kihi)}.
which is just the formula preceding (7.8) for dID. Thus T) = dID.
PROBLEMS
1. Following Example 1, write out for n = 2 and n = 4 the standard basis for (E~)* and
express any 2-covector as a linear combination of the basis covectors e ii.
2. Show that the 2-covectors eii, 1 :0; i < j :0; n, are linearly independent.
3. Let n = 3. Simplify:
(a) (e l + el ) /\ (e! - 3e 3 ).
(b) (dy - x dz) /\ (xy dx + 3 dy + z dz).
4. Let n = 2. Find the exterior differential of:
(a) xly dy - xyl dx.
(b) x dy + y dx.
(c) f(x)dx + g(y)dy.
(d) fIx, y)dy.
282
7.2 Alternating multilinear functions
where
(7.11)
Interchanges
Let S be some set. For our purposes we shall take either S = En or S =
{l, 2, ... ,n}.If(PI"" ,Pr) and (P'l"" ,p~) are r-tuples of elements of S, let us
say that the second r-tuple is obtained from the first by interchanging Ps
and PI if p~ = PIO P; = Pso and P; = PI for I # s, t.
EXAMPLE 1. The triple of vectors (h 3 , h2' hI) is obtained from (hI' h2' h3) by
interchanging h3 and hI' The 4-tuple of integers (1,5, 3, 7) is obtained from
the 4-tuple (1, 7, 3, 5) by interchanging 5 and 7.
We recall that an r-tuple (hI, ... ,hr) is called linearly dependent if there
exist scalars c l , ... ,cr, not all 0, such that clhl + ... + crhr = O.
283
7 Exterior algebra and differential calculus
PROOF. First of all, the conclusion is true if some vector in the r-tuple is
repeated. For instance, suppose that hi = h2. Since M is alternating,
M(hb h2,h3,···,hr) = -M(h 2,h l ,h 3,···,hr)·
Then M(hb hi, h3' ... , hr) is its own negative, and must be O.
Suppose for instance that hr is a linear combination of the vectors pre-
ceding it,
hr = clhl + ... + cr-Ih r _ l ·
Since M(h l , ... , hr- I, ) is alinearfunction,
r- I
In the lth term on the right-hand side, the vector hi is repeated, and hence
each term is O. Thus M(h l , ... , hr) = O. 0
284
7.2 Alternating multilinear functions
285
7 Exterior algebra and differential calculus
where the numbers W A are defined by (7.11) and the alternating r-linear
functions e A by (7.14). The notation [A.] indicates a sum over all increasing
r-tuples A. = (i l' ... , ir)'
PROOF. Let Nt equal the right-hand side ~of (7.16). For each J.1 = (jl' ... ,jr),
-
M(ej" ... , ejJ "A
WAe (e
= L. iI ,···, ejJ
[A]
But M and Nt are r-Iinear and have the same value w il at (eiI , ... ,ejJ for
each J.1. By (7.10) M = M. 0
Let us return to the special case r = n mentioned earlier. The only increas-
ing n-tuple of integers between 1 and n is (1,2, ... ,n). By Proposition 7.2,
any alternating, n-linear function has the form M = ce l ''' n, where c = W l "'n
is a scalar. In the notation of (7.12), e l .. · n = D.
PROBLEMS
1. Let n = 5. Find
286
7.3 Multicovectors
3. Show that:
(a) b~ = b~.
7.3 M ulticovectors
Let us now introduce a different name and a different notation for alter-
nating, multilinear functions.
287
7 Exterior algebra and differential calculus
Products
In (E~)* we define the euclidean inner product
ro'~=LWl(l
Il]
and set Irol 2 = ro' roo The standard basis elements are orthonormal with
respect to this inner product.
Another important product is the exterior product, denoted by the symbol
1\. The exterior product of an r-covector and an s-covector is an (r + s)-
covector, defined as follows: If
A = (i" ... , ir ), v = (jl"" ,js),
let us write A, v for the (r + s)-tuple
ro 1\ ~= L wl("e l 1\ e.
[lll,,]
288
7.3 Multicovectors
PROOF. The proof of (1) and (2) is almost immediate from the definition and
is left to the reader (Problem 5). To prove (3),
(7.19) (e A /\ e) /\ eP = e A /\ (e V /\ ep).
289
7 Exterior algebra and differential calculus
Since both sides of (7.18) change sign under interchanges in .Ie or v, formula
(7.18) also holds for non increasing r-tuples. Thus (7.20) is valid regardless
of whether .Ie is increasing.
Let us next give two useful formulas, which hold if ro is the exterior
product of l-covectors.
Note that the matrices whose determinants appear on the right side of
(7.21) and (7.22) are r x r. The second of them is obtained by deleting all
columns except jl, ... ,jr in the r x n matrix whose row co vectors are
a l , ... , a r .
PROOF. By (7.14) and (7.20), Formula (7.21) is valid if at, ... , a' are standard
basis co vectors, a k = e ik • Formula (7.21) then holds in general, since both
sides of(7.21) are multilinear in a I, ... , a r . Ifh, = ej" then wI' = ro(eh , ... , ej )
and a k • h, = a',.
Thus (7.22) follows from (7.21). D
*Remarks. The exterior product has been defined in terms of the standard
bases. It is not clear that it is "coordinate free," in other words, that the same
exterior product would be obtained starting from different bases. However,
let us add Formula (7.21) to the list [(1)-(4)] of properties in Proposition 7.3.
Then 1\ is the only product with these five properties. To see this, we first
observe that Formula (7.20) is a special case of (7.21), upon taking a k = eik
and using (7.14). Moreover, (7.18) is a consequence of(7.20) and the associative
law (4). Once the product is known for basis elements, Properties (1) and (2)
determine it in general. Thus 1\ is the only product with Properties (1)
through (4) and (7.21). In fact, (3) can be omitted from the list since it follows
from the other four. Since none of these five properties refers 10 bases, the
exterior product is coordinate free.
*Note about terminology. A multilinear function M of degree r and domain
En X ••. X En is often called a covariant tensor ofrank r. An r-covector is then
called an alternating covariant tensor of rank r.
290
7.4 Differential forms
The sum of an r-covector and an s-covector has been defined only when
r = s. However, one may form the direct sum
(An)* = (E~)* EB (E1)* EB ... EB (E~)* EB ... ,
where we agree that (E~)* is the scalar field. The exterior product induces a
product in (An)*, which is then an algebra over the real numbers. This algebra
is called the exterior algebra of (En)* [4]. (A n)* is sometimes called the
covariant Grassmann algebra or the covariant alternating tensor algebra
of En.
PROBLEMS
1. Write down the standard basis for (E~)* for each r = 1,2,3,4. Find all products
e A /\ e" where A = (i) and v = (j, k, I) IS an increasing triple.
2. Let n = 3. Simplify:
(a) (2e l - e 2 ) /\ (3e 2 + e 3 ). (b) e 21 e 23 .
/\
(c) (e l - e 2 + 3e 3 ) /\ e 21 . (d) (e 23 + e 31 ) /\ (5e l - e 2 ).
3. Let n = 5. Simplify:
(a) e 253 /\ (e 14 + e42 ).
4. Show that ifLp] cAe A = 0, then c, = 0 for every increasing I.. [Hint: See (7.15).J
5. Prove (1) and (2) of Proposition 7.3.
6. Prove the associative law (4) of Proposition 7.3, using (I), (2), and (7.19).
7. Show that 0>"/\ ~ /\ '1 = -'1 1\ ~ /\ 0> if 0> has degree r. '1 has degree t. and both r, t
are odd.
<0 = I Wi dxi ,
i= 1
The values of <0 are r-covectors. The Greek letters <0 and ~ used in Section
7.3 to denote r-covectors are now used to denote differential forms. The
context indicates clearly which is intended.
291
7 Exterior algebra and differential calculus
for every XED. Similarly, iffis a real valued function, thenfco is the r-form
such that
(fco)(x) = f(x)ro(x)
for every XED.
Theorem 7.1. Let co, ~, and l) be differential forms, and f a real valued function,
all with the same domain D. Then
(1) (co + ~) /\ l) = co /\ l) + ~ /\ l).
(2) (fco) /\ ~ = f(co /\ ~).
(3) ~ /\ co = ( - l)"co /\ ~, if co has degree r and ~ has degree s.
(4) (~ /\ co) /\ l) = ~ /\ (co /\ l)).
PROOF. This is immediate from the definitions above, and Proposition 7.3. 0
We recall that dx i is the I-form with constant value ei . Similarly, for any
r-tuple A. = (iI' ... ,ir ) the r-form dXil /\ '" /\ dxir has constant value eA.
Hence if co is an r-form, then
(7.23) co = I W A dXi! /\ '" /\ dx ir
[AJ
where the value of W A at x is wix). Using Problem 4, Section 7.2, one can
also write
1 ~. .
co = , L. WA dX" /\ ... /\ dx'r.
r. ;.
We say that an r-form co is of class C(q) if the functions W A in (7.23) are
of class C(q).
We recall that iffis a O-form of class C(1), then dfis the I-form
df= fl dx 1 + ... + fndxn,
where flo ... ,fn are the partial derivatives. In particular, if co is an r-form of
class C(1), then W A is a O-form of class C1) and dw;. is defined.
Definition. Let co be an r-form of class C(1). The exterior differential dco is the
(r + I)-form defined by the formula
(7.24) dco = L,dw A /\ dXil /\ ... /\ dxir.
[AJ
292
7.4 Differential forms
dro = df 1\ dx 1\ dy = ( of
ox dx
of
+ oy dy + oz dz
of)
1\ dx 1\ dy,
of of
dro = OZ dz 1\ dx 1\ dy = OZ dx 1\ dy 1\ dz.
EXAMPLE 2. If r = n, then
ro = f dx 1 1\ ... 1\ dxn
where f = ro 1 "'n' Since dro is an (n + 1)-form, dro = O.
Theorem 7.2. The exterior differential has the following properties:
(1) d(ro + ~) = dro + d~, if ro and ~ are rlorms of class e(l).
(2) d(ro 1\ ~) = dro 1\ ~ + ( -1 Yro 1\ d~, if ro is an rlorm and ~ is an slorm,
both of class e(1).
(3) d(dro) = 0 ifro is an rlorm of class e(2).
f 1\ ~ = f~. Similarly, if s
[If r = 0, we agree that = 0 then ro 1\ f = fro.
The theorem remains true if r = 0 or s = 0.]
293
7 Exterior algebra and differential calculus
[A)[V)
while
L (dw" 1\ E") 1\ (C,EV) = [L dW A 1\ E"] 1\ [L ("Ev] = dro 1\ ~.
WM W M
Similarly,
(- I)' L (w"E") 1\ (d(,. 1\ EV) = (-I)'ro 1\ d~,
[A)[v)
a2f a2f } . .
d(df) = L ( a i a j - a ja i Xl 1\ dx J = O.
i<j X X X X
The form EA has constant coefficients and hence dE" = O. Using the product
rule (2), d(df 1\ E") = O. Takingf = .w Aand using (I),
= Ldw A 1\ EA.
[A)
294
7.5 MuItivectors
Thus d'ro = dro for every ro of class e(1). In particular, this proves that the
exterior differential d is "coordinate free."
PROBLEMS
d(P dy /\ dz + Q dz /\ dx + R dx /\ dy) =
JP + -JQ + -JR) dx /\ dy /\ dz.
( -::;-
ex Jy Jz
4. (a) Show that if co and ~ are closed differential forms, then co /\ ~ is closed.
(b) Show that if co is closed and ~ is exact, then co /\ ~ is exact.
6. A function f is an integrating factor for a I-form co if f(x) #- 0 for every XED and fco
is exact. Show that if co has an integrating factor then co /\ dco = O.
8. Let COl, ... , coP be I-forms such that co i = If= I fJ dg< i = I, ... , p. Assume that
the functions f~ are of class e(l), the gi are of class e(2), and that the I-covectors
col(x), ... , coP(x) are linearly independent for every XED. Find I-forms 9~ such that
dco i = If=19~ /\ co i . [Hint: The p x p matrix (f~(x)) must be nonsingular.]
Note. Conversely if dco i is a linear combination of COl, ... ,coP with coefficients
I-forms 9~, then locally functions f~, gi as above can be found. This result is called the
Frobenius integration theorem and has important applications in geometry and dif-
ferential equations [8, p. 97]. We give a proof for p = I in Section 7.10.
7.5 Multivectors
Let us now define objects called multivectors of degree r, or, for brevity,
r-vectors. They have a role dual to the r-covectors. Formally, one passes from
multicovectors to multicovectors by merely exchanging subscripts and super-
scripts. Multivectors are usually denoted by ~,p, .... We are interested
295
7 Exterior algebra and differential calculus
The case r =2
By I-vector we mean a vector hE En. To motivate the general treatment of
r-vectors, let us begin with a discussion of 2-vectors. They are defined as
alternating bilinear functions on (En)* x (En)*. This is a special case of the
definition of r-vector given below. Let E'2 denote the space of 2-vectors. In
exactly the same way as in Section 7.1, one introduces 2-vectors eij such that
eji = - eij' e jj = 0, and the n(n - 1)/2 2-vectors eij for i < j form a basis for
E'2. This is called the standard basis for E'2; we see below that the basis for
(E'2)* of 2-covectors eij is dual to it.
Let us denote the component of a 2-vector 12 with respect to eij by (Xij.
Thus
(7.25)
This formula is dual to (7.5); note that subscripts and superscripts have been
exchanged.
The exterior product h 1\ k of vectors h, k is defined, as in Section 7.1,
by agreeing that e j /\ ej = e jj , and that 1\ distributes with addition and scalar
multiplication. We are particularly interested in decomposable 2-vectors,
namely, those that can be written as 12 = h 1\ k for suitable vectors h, k.
The reason is that several interesting geometric quantities are associated
with a decomposable 2-vector.
Let hand k be linearly independent vectors, and P the 2-dimensional
vector subspace of En that they span:
P = {x = sh + tk, s, t real}.
We call the pair (h, k), in the given order, aframe for P. Suppose that (hi, k')
is another frame for P. Then, for suitable real numbers s, t, u, v
296
7.5 Multivectors
(h', k') are frames for the same 2-dimensional subspace P. Thus, although
P has many frames (h, k), the associated 2-vectors h 1\ k are determined by P
up to a scalar multiple.
The concept of orientation is essential in defining integrals of differential
forms in Section 8.7. As a preliminary step, we define orientation for vector
subspaces of En in the present section. Any frame (h, k) determines an orienta-
tion for the 2-dimensional subspace P spanned by (h, k). Let us agree that
another frame (h', k') determines the same orientation for P as (h, k) if c > 0
in (7.26), and the opposite orientation if c < O. This is a somewhat clumsy
way to define orientation. We give in Formula (7.29) below an equivalent
definition in terms of 2-vectors, which is the one used in this book.
ex' P = L
i<j
rt. ij{3ii,
297
7 Exterior algebra and differential calculus
Let e). be the r-vector defined by the following formula dual to (7.14):
(7.31)
for every r-tuple (a 1 , .•• , ar ) of covectors.
Let E~ denote the set of all r-vectors. Then E~ satisfies the axioms for a
vector space. It consists of 0 only if r > n. For 1 :s; r :s; n the r-vectors
298
7.5 MuItivectors
Table 7.1
r-vectors r-covectors
Elements of E~ (E~)*
e). with A increasing form the standard basis for E~. The number IX). is the
component of (l with respect to e)..
The inner product (l • Pof two r-vectors, and the exterior product (l /\ P
of an r-vector (l and an s-vector Pare defined by the formulas dual to those
in Section 7.3. In each instance subscripts are replaced by superscripts and
vice versa. The exterior product of multi vectors has the same properties
listed in Proposition 7.3. The scalar product ro . (l of an r-covector ro and an
r-vector (l is defined in the third from last line of the Table 7.1. The last two
lines of the table are particular cases of the formula for ro . (l.
The formulas in the second line and in the last two lines of Table 7.l are
valid regardless of whether Aand J1 are increasing, since they are known to be
valid for increasing r-tuples, and both sides of each formula change sign
under interchanges.
The reader should compare this table with the corresponding table for
r = 1 (Section 3.2). According to the definition (Section A.l), the dual space
of E~ consists of all real valued linear functions F with domain E~. The dual
space may be identified with (E~)* in the following way. Given an r-covector
ro, let F((l) = ro . (l for every (l E E~. This establishes an isomorphism between
(E~)* and the dual space of E~. The next to last line of the table implies that
the standard bases for E~ and (E~)* are dual.
299
7 Exterior algebra and differential calculus
Proposition 7.5. IfCl = hi /\ ... /\ h" then for every r-covector ro,
ro . Cl = ro(hb ... , hr ).
Proposition 7.6. If ro· Cl = ro· Pfor every decomposable r-covector ro, then
Cl = p.
PROOF. The standard basis r-covectors e A are decomposable. Hence for
every increasing A.,
D
300
7.5 Multivectors
The only difference between the notions of basis and frame is that the
latter takes into account the order in which the basis vectors hI> ... , hr are
written.
Theorem 7.3
(a) An r-tuple (hi" .. , hr) is linearly dependent if and only ifh l /\ ... /\ hr = O.
(b) Let PeEn be an r-dimensional vector subspace and (hi' ... , hr), (h 'l , ... , h~)
be any twoframesfor P. Then there is a scalar c such that
(7.34) h'l /\ ... /\ h~ = chi /\ ... /\ hr·
(c) Conversely, if (hi, ... , hr) and (h'l' ... , h~) are frames which satisfy (7.34)
for some scalar c, then they are frames for the same vector subspace P.
(see Figure 7.2).
/"\
/' \
// \
h~ / \
\
h2 -\ ' \ - - - - -7
\ /
/
/
/
/
o
Figure 7.2
301
7 Exterior algebra and differential calculus
let a I, ... ,ar be covectors such that ak • hi = <5~, and let 0) = aI /\ ... /\ ar •
By (7.32),
0). IX = det(<5n = 1.
Hence IX #- O.
h; = L cihm, 1 = 1, ... , r.
m=1
where IX' = h'l /\ ... /\ h~. The matrix on the right is the product of the
matrices (a k • hm) and (ci). Hence if c = det(ci),
0) • IX' = c det(a k • hm ) = co)· IX = 0) • (CIX).
This is true for every decomposable 0); hence IX' = CIX by Proposition 7.6.
To prove (c), again let IX = hi /\ ... /\ h" IX' = h'l /\ ... /\ h~. By hypoth-
esis, {hi' ... , hr } and {h'I' ... , h~} are linearly independent sets. It suffices to
show that each h; is a linear combination of hi' ... , hr. Suppose that this is
false for some I, say for 1 = 1. Then {hi' ... , h" h'd is a linearly independent
set. Let a l , ... , a r + 1 be such that ak • hm = <5~ for k, m = 1, ... , r + 1, where
we have set h'l = hr+ I' Let 0) = a l /\ . . . /\ a r . Then 0). a = 1, but 0) • IX' = 0
since the elements a k • h'l = ()~+ 1 of the first column of the r x r matrix
(a k • h'l) are O. This contradicts the assumption that IX' = CIX, C #- O. D
302
7.5 MuItivectors
K = {x: x = Xo + ±
k;1
tkhk' 0 ~ tk ~ 1, k = 1, ... , r}
is the r-parallelepiped spanned by hI> ... , hr with Xo as vertex.
By part (a) of Theorem 7.3, v,.(K) = 0 if and only if (hI> ... ,hr) is linearly
dependent.
To show that the definition (7.35) of r-measure for parallelepipeds is
reasonable, let us show that v,.(K) is the product of the lengths of the vectors
hl>"" hr in case these vectors are mutually orthogonal. If VI"", Vr are
vectors, let a k be the covector with the same components as the vector Vk •
Then a I !\ ... !\ a r is the r-covector with the same components as the r-
vector VI !\ ... !\ Vr . By (7.32),
(v I !\ ... !\ vr)· (hi !\ ... !\ hr) = det(v k • hi),
where the· now denotes inner product. In particular, let Vk = hk. Then
(7.36)
Taking square roots, we get a formula for v,.(K). If hi' ... , hr are mutually
orthogonal, then hk ' hi = 0 for k "# I and det(h k ' hi) = Ih I 12 •• ·lh,.1 2 . In
this case v,.(K) = Ihll·· 'Ihrl as required.
S = {x:x = Xo + i
k;1
tkhk,tk 2': Ofork = 1, ... ,r, i
k;1
tk ~ I}.
303
7 Exterior algebra and differential calculus
Orientations
Let P be an r-dimensional vector subspace of E", and (hi"'" h,) a frame for
P. The r-vector
hi /\ ... /\ h,
(10 = ----=----'---
Ihi/\ ... /\ h, I
has P as its r-space, and I(10 I = 1. We call (10 an orientation for P:
Definition. A decomposable r-vector (10 is an orientation for P if I(10 I = 1 and
P is the r-space of (10 •
There are two orientations for P. If (10 is one of them, then - (10 is the other.
Two frames (hb ... ,h,), (h/b ... ,h~) for P determine the same orientation for
P if and only if c > 0 in (7.34). Note that c is the determinant of the r x r
matrix (ci) in the proof of Theorem 7.3(b).
If two vectors in the frame (hi' ... ,h,) are interchanged, then the exterior
product hi /\ ... /\ h, changes sign. Thus the orientation of a frame changes
under interchanges.
We now have a criterion that shows when two frames lead to the same
r-vector.
Corollary. Let (h b ... , h,), (h~, ... , h~) be frames. Then hi /\ ... /\ h, =
h/l /\ ... /\ h~ if and only if these frames span the same r-space P, have
the same orientation, and their parallelepipeds with 0 as vertex have the same
r-measure.
The case r = n
In this case P = E" and ±e l ... " are the two orientations. Let us call e l ... "
the standard, or positive, orientation of E" and - e l ... " the negative orientation
of E". When r < n, we do not attempt to call one orientation of P positive
and the other negative. If (hi' ... ,h") is a frame for E", then by (7.33)
(7.38)
The frame has positive orientation if det(h7) > 0 and negative orientation
if det(M) < O. For small values of n, (7.38) sometimes furnishes a convenient
way to find the determinant of the matrix (hf), since the exterior product on
the left side of (7.38) is readily computed.
304
7.5 Multivectors
PROBLEMS
1. Simplify (n = 6):
(a) e 3 1\ e s 1\ e 24 .
(b) e 2 1\ e 3 1\ e 62 .
(c) e l 1\ (e 14 + e64 ).
(d) (e l + 3e4 - e6) 1\ (2e23 + e 36 ) 1\ e4S '
(e) (ell + e 13 ) 1\ (e 34 + e 2S ) 1\ (eS6 + e46 ).
2. Evaluate the indicated scalar products (n = 4).
(a) (e l + e 2) . (e l + e 2).
(b) e l2 . e 34 .
(c) eI34'(e431 + 3e I24 ).
(d) (e l - e4) 1\ (e 2 + e4). (e l + 2e 4) 1\ (e 2 - 2e 4).
3. Using Theorem 7.3 show that (2el + e 3, e 2 + e4, e l + e4, e 3 + e4) is a frame for
E4. What is its orientation? Find the determinant of the matrix that has these vectors
as columns, using (7.38).
4. Show that (e l + e4 , e 2 + e s , e 3 + e 6, e l + e s , e 2 + e6 , e 3 - e4) is a frame for E6.
What is its orientation?
5. Show that (e l - e 2, e 2 - e3) and (3e l - e 2 - 2e 3, 2e l - e 2 - e 3) are frames for
the same vector subspace of E3. Do their orientations agree?
6. Find the area of the triangle with vertices 2e 3, e l - e 2 + 2e 3, e l + 3e 3.
7. Find the volume of the 3-simplex in E4 with vertices 0, e l - e 3, e 2, e 3 + 2e 4.
8. Let K be an r-parallelepiped spanned by hi' ... , hr with 0 as vertex. For each
increasing A. let Kl = Xl(K) where Xl is the projection onto the r-space of el'
(Xl leaves the components xit, ... , xir of any x unchanged and replaces each of the
other components by 0.) Let rl = hi 1\ •.• 1\ hr. Show that locll = V.(K l ) and
hence [v.(K)]2 = Ill] [V.(Kl)Y Illustrate for n = 3 and r = 1,2.
9. Show that:
(a) If rl is decomposable, then rl 1\ rl = O.
(b) If rl and P are decomposable r-vectors, then (rl + P) 1\ (rl + P) = 2rl 1\ P if r
is even and is 0 if r is odd.
(c) The 2-vector el2 + e 34 is not decomposable. [Hint: Use (a).]
10. Let rl and P be decomposable nonzero 2-vectors, and P, Q be their respective
2-spaces. Show that if P 11 Q = {O}, then rl + P is not decomposable; and if P 11 Q
is a line through 0, then rl + P is decomposable and rl #- cpo [Hints: In the first
instance rl = h 1\ k, P = h' 1\ k', where {h, k, h', k'} is a linearly independent set.
In the second rl = h 1\ k, P = h 1\ k', where h E P 11 Q.]
11. Let rl = h 1\ k, rl #- O. Show that the matrix (oc ii ) has rank 2. [Hint: Show that each
column vector of the matrix is a linear combination of hand k.]
12. Let (x o, XI' ... , x r) be an (r + I)-tuple such that the vectors XI - x o, ... , xr - Xo
are linearly independent. Such an (r + I)-tuple defines an oriented r-simplex.
Its r-vector is I/r !(XI - xo) 1\ •.• 1\ (x r - xo). Let Pi be the (r - I)-vector of the
ith oriented face (x o, XI' ... , Xi-I, Xi + I"'" xr). Show that
r
I(-I)iPi = O.
i=O
305
7 Exterior algebra and differential calculus
hI = L(k I )
L2 (k l 1\ k 2 ) = hI 1\ h2
o o
Figure 7.3
-
At the end of the section we show that Lr is well defined by Formula
(7.39). The linear transformations L, are said to be induced by L.
There is a geometric way to think about L,. Let Q be the vector subspace
of Em spanned by {kb' .. ,k,}, and let hi = L(k i), P = L(Q). Then P is the
vector subspace of En spanned by {hl' ... ,h,}. Let us suppose that Q and P
are both of dimension r. [Note: This holds in particular if r = m, m ::; n,
and L has maximum possible rank r. This case is encountered repeatedly in
Chapter 8.J When Q and P have dimension r, (k l , ... , k,) is a frame for Q
and (h l , ... , h,) a frame for P. Moreover, Q is the r-space of the r-vector
p = kl 1\ ... 1\ k,; and P is the r-space of the r-vector (X = hl 1\ ... 1\ h, =
L,(P) according to (7.39).
Let I be the r-parallelepiped spanned by (k t, ... , k,) with 0 as vertex,
and let K = L(l). Then K is the r-parallelepiped spanned by (h io ... , h,),
with 0 as vertex (see Figme 7.3). Consider the ratio of r-measures:
By (7.35), " = I(X II P 1- 1. Let us show that" depends only on the r-space Q
and on L. Let (k'l,"" k~) be another frame for Q, and h; = L(kJ Then
(h'l" .. , h~) is a frame for P. Let P' = k'l 1\ ... 1\ k~, (x' = h'l 1\ ... 1\ h~ =
L,(P'). By Theorem 7.3, P' = cp for some real c =f. O. Since L, is linear,
(X' = L,(cP) = cL,(P) = c(x.
306
7.6 Induced linear transformations
Therefore
This defines the induced linear transformation Lr for 1 < r < m. We set
LI = L. If r > m, then E~ has the single element 0 and Lr(O) = O.
Let us show that for every PE E~, Y E E:,
(7.44)
Let /1 = Ub"" jr) and v = (k b .. ·, ks) be increasing. Then
L r (E JJ. ) 1\ L s (E v) = v·) 1 1\'" 1\ v·Jr 1\ Vk 1 1\'" 1\ vk
5 •
If any integer is repeated in the (r + s)-tuple (/1, v), then this is O. Otherwise,
the right-hand side is (-1)PL r +.{E t ) where r is the increasing (r + s)-tupJe
obtained from (/1, v) by p interchanges. Since Ell. \' = (-l)PE, and Lr+s is
307
7 Exterior algebra and differential calculus
By (7.33), c; is the Ath component ofvit 1\ .,. 1\ Vjr' By (7.42) and (7.43)
(7.45)
308
7.7 Transformation law for differential forms
PROBLEMS
if ro E (E~)*, ~ E (E~)*.
309
7 Exterior algebra and differential calculus
00'
Figure 7.4
for every t E 11, which proves (2). By the chain rule (Section 4.3), d(f 0 g)(t) =
Lj[df(x)]. By (7.49) the right-hand side is (df)~(t). Thus (3) holds. Recall
that dx i stands for dXi, where Xi(X) = Xi for each x. Then gi = Xi 0 g and
from (3) withf = Xi,
(7.50) (dx i)# = di.
Then (4) follows from this and (2). To prove (5) we have, from (1) through (4);
oo~ = L OJ;. 0 g dg i1 1\ .,. 1\ dg ir ,
[A]
The matrix of Dg(t) is (g~{t», and the row covectors are dg 1 (t), .. . , dg"(t).
By (7.22), the }lth component of dg i1 1\ ... 1\ dg ir is g~. Therefore
(7.51) (dXil 1\ '" 1\ dxir)~ = L9~dtil 1\ ... 1\ dtir.
[1']
310
7.8 The adjoint and codifferential
PROBLEMS
1. Let m = 2, n = 3, and g(s, t) = (st)e\ + (s cos t)e 2 + (exp t)e 3 . Find 0>' if:
(a) 0) = x dy. (b) 0) = y dz /\ dx.
(c) 0) = dx /\ dy /\ dz. (d) 0) = f (a O-form).
(7.52) C = I(WAog)g~.
IAJ
(b) Show that 0)" = 0, if Dg(t) has rank < r for each t E!l and 0) is of degree r.
[Note: In tensor language a differential form of degree r, being an r-covector-
valued function, is called a covariant alternating tensor field of rank r. Formula
(7.52) is the transformation law for the components of such a tensor field.]
3. Let n = m = 2, r = 1,0) = M dx + N dy. Find explicitly dO) and 0)" and verify that
(dO))" = dO)".
4. Let n = m = 3, g(s, t, tI) = (s cos t)e l + (s sin t)e 2 + tle 3 . Find:
(a) (f dx /\ dy /\ dz)". (b) (x dy /\ dz)".
311
7 Exterior algebra and differential calculus
n=3
r=2
01 = hI 1\ h2 o
Figure 7.5
Let tX = hi /\ ... /\ hn _ I' If tX = 0, then we set *tX = O. If tX =1= 0, then *tX will
turn out to be the vector h with the following three properties: (I) h is a vector
normal to the (n - I)-space P spanned by hi' ... ,hn - I ; (2) (h, hi' ... ,hn - I )
is a positively oriented frame for En; (3) Ih I = ItX I. Condition (3) says that the
length ofh equals v,,- I (K), where K is the (n - I )-parallelepiped with vertex 0
spanned by hi, ... ,hn - I (see Figure 7.5).
With this in mind, let us define * first for the standard basis (n - 1)-
vectors. Let i' = (1,2, ... , i-I, i + I, ... , n). Since i-I interchanges will
change the n-tuple (i, i') into the increasing n-tuple (1, ... ,n),
(-I)i-Ie i /\ ei' = e l .. · n ·
Therefore we set
(7.53) *e i· = (_l)i-I ei .
We want the operation * to behave linearly.
For any (n - I)-vector ex = LI= 1 r/e j ., let *ex be the vector h =
II=I ai'(*e i·). Its components are
(7.54) i = I, ... , n.
Proposition 7.7. The vector h = *tX defined by (7.54) has properties (I), (2), (3)
above.
PROOF. Given a frame (hi"'" hn - I ), these three properties determine a
vector, which we denote temporarily by 11. Let (h'b ... , h~_ I) be an orthogonal
frame for P, hI. . hi = 0 if k =1= I. Then ex is a scalar multiple b ofh'J /\ ... /\ h~-l'
and replacing h'l by bh'l we may assume that tX = h'l /\ ... /\ h~ - I' By (l),
11· hI, = 0 for each k = I, ... , n - 1. Therefore
111/\ tXl = Illllh'II'''lh~_11 = IlllltXl.
312
7.8 The adjoint and codifferential
1:
n /\ a -_(~7_i)
i~lnei /\ (~j')
j~ll1. ej' =
[~-ii'
i~lhll. (-1) i-lJ e1··· n ,
313
7 Exterior algebra and differential calculus
The remainder of this section is not used in Chapter 8. Let us define *ex
for any r-vector ex when 0 ~ r ~ n. If r = 0 or n, we set
*c = ce l ... n,
If 0 < r < n, let A = (i1>' .. , i,) be any increasing r-tuple, and let A.' =
U1>' .. , jn-,) be the increasing (n - r)-tuple whose entries are those integers
jl between 1 and n which do not appear in A. Let
;: 1 ... n
BA = UA',A .
It is ± 1, depending on whether an odd or an even number of interchanges
puts A.', A in increasing ordeLIf ex is any r-vector, then its adjoint is the (n - r)-
vector *ex whose components satisfy
(7.60) (*ex( = exABA.
If r = n - 1 and A = ii, then A.' = (i), BA = (_1)i- 1, and (7.60) agrees with
(7.54). From the definition (7.60),
*(ex + P) = *ex + *P, *(cex) = c*ex.
Moreover, *ex = 0 if and only if ex = O. Thus the operation * gives an iso-
morphism between E~ and E~_,. This isomorphism preserves inner products.
In fact, if ex and pare r-vectors, then
*ex *P = L (*ex)"'(*P( = L (BA)2ex AfJA.
0
[i·1 [AI
and
(7.61)
314
7.8 The adjoint and codifTerentia]
315
7 Exterior algebra and differential calculus
PROBLEMS
ro = L" w·o d
I Xi d
/\ ... /\ X i-I /\ dX i+ I /\ ..• /\ dX "
,
;=1
4. Let L be a linear transformation from E" into E" such that v,,-I(K) = v,,_I[L(K)]
for every (11 - I)-parallelepiped K. Show that L is an isometry.
7. Show that the components of dro satisfy (dro)1. = Li!: (-I)j-I OWl.jJx ij , where Aj
is the r-tuple (i[o ... , ij-I, ij + [0 ••• , ir+ I)'
8. If ~ is a I-form and ro is an r-form let ~. ro be the (r - I)-form such that *(~ . ro) =
(_l)"-I~ /\ (*ro). Show that:
(a) d(fro) = fdro + df ·ro. (b) (~'ro)v = D=I CW,.. i·
316
7.9 Special results for n = 3
(see Figure 7.5). The cross product distributes with vector addition and scalar
multiplication, and h2 x hi = - hi X h 2 . However, it is not associative.
The triple scalar product of three vectors is denoted by [hi' h2' h 3]. It is
given by
(7.66)
Its absolute value equals Ihi /\ h2 /\ h31, which is the volume of the paral-
lelepiped spanned by hi' h2' h3 with vertex O. The sign of the triple scalar
product is positive if (hI> h2' h3) is a positively oriented frame for E3 and
negative if this frame is negatively oriented.
Whenn = 3,r(n - r)isalwaysevenand(-l)r(n-r) = l.Then*(h l x h 2) =
hi /\ h 2. Using Problem 5(a), Section 7.8,
hi /\ h2 /\ h3 = (hi x h 2)· h 3e 123 ,
which gives another formula for the triple scalar product:
[hi' h2' h 3] = (hi x h2)· h 3.
The cross product of two co vectors, or of two I-forms, is given by
(7.67) co x ~ = *(co /\ ~).
EXAMPLE I. Show that div(curl co) = 0 for every I-form co of class C 2 ). Using
the fact that (1* = *d (Formula 7.64),
div(curl co) = d(*dco) = *d(dco) = *0 = o.
PROBLEMS
317
7 Exterior algebra and differential calculus
(e) curl(curl ro) = d(div ro) - Lapl ro, where Lapl(M dx + N dy + 0 dz) =
(Lapl M)dx + (Lap I Njdy + (Lapl O)dz and Lapl f is the Laplacian of the
function f.
(f) ~. curl ro - ro' curl ~ = div(ro x ~). [Hint: By the dual to (7.61), ~. *dro =
*(~ /\ dro).]
5. Show that
(a) hi x (h z x h3) = (hi' h3)hz - (hi' hz)h 3. [Hint: Since both sides are trilinear
in (hi, hz, h3)' it suffices to prove this when hi' hz, h3 are standard basis vectors.
Use Problem 3.]
(b) The cross product is not associative.
(c) (hi x h z ) X (h3 x h4 ) = [hi' h1' h4 ]h 3 - [hi' hz, h3]h4 .
6. Let ro = (E I dx l + E1 dx 1 + E3 dx 3 ) /\ dx 4 + BI dx z /\ dx 3 + B z dx 3 /\ dx l +
B3 dx l /\ dx 2 , where the functions Bi, Ei are of class C(1) on an open subset of E4.
Show that dro = 0 if and only if curl E + JB/Jx 4 = 0, div B = O. Here curl and
div are taken in the variables (Xl, xZ, x 3). {Note: The equation dro = 0 represents
one-half of Maxwell's equations for an electromagnetic field in free space. The
functions E I , E z , E3 represent the electrical components of the field and B I , B z , B3
the components of a magnetic induction vector. There is a similar equation repre-
senting the other half of Maxwell's equations [S, p. 45], T. Frankel, 1974, Amer. Math.
Monthly, 81, 343-34S]}.
Theorem 7.5
(a) If co has an integrating factor in D, then co /\ dco = O.
(b) Suppose that co /\ dco = O. Then, for any x* ED, there exists an open
set L1 such that x* E L1 and co has an integrating factor in L1.
The proof of (a) is an easy exercise, which we leave to the reader (see
Problem 6, Section 7.4). As already mentioned (Section 6.6), in proving (b) it
suffices to suppose that co has the special form
n
(7.69) co = dx 1 + L wi dxi .
i=2
318
7.10 Integrating factors (continued)
PROOF. For sufficiently small Q there exist I-forms 11 1, ... , 11" continuous
on Q such that {l1 l (x), ... , l1"(x)} is an orthonormal basis for (E")* and
111 = lrol-lro(see Problem I). Then
~(x) = L (iix)l1i(X) /\ lli(X),
i<i
where (iix) is the component of ((x) with respect to the (i,j)th element of the
corresponding orthonormal basis {l1 i(X) /\ lli(X), i < j} for (E~)*. Since
ro /\ ~ = 0,
o = 111(X) /\ ~(x) = L (iiX)l1l(X) /\ ll i(x) /\ ll i(x).
2Si<i
with gl(O) = u, for 0 :::; t :::; 1, (u, z) E Q2, where Q 2 is some neighborhood of O.
By definition, gl(l) = F(u, z). Moreover, gl(t) = F(u, tz) (see Problem 2).
Hence
dg l
(7.71) -d
t
= L Fi(u, tz)x'
i=2
" .
where Fi = OF/OXi. Let H(t, u, z) = (F(u, tz), tz). By (7.70) and (7.71) we have
" "
ro' = F I du +L F;(x i dt + t dx i ) + L Wi H(x i dt + t dxiJ, 0
i=2 i=2
319
7 Exterior algebra and differential calculus
Ol/li = '/"'/'.
at '1''1','
1. Let ro be a continuous I-form with ro(O) "# O. Show that there exist a neighborhood n
of 0 and I-forms 1) I, ... , 1)n continuous on n such that 1)i(X) . 1)i(x) = c5ij, i,j = I, ... , n
and 1)1(X) = Iro(x)I-lro(X), for all x E n. Here c5ij = I if i = j, c5ij = 0 if i "# j. [Hint:
Apply the Gram-Schmidt process to {ro l , ... , ron} where rol = ro and ro 2 , ••• , ron
are chosen so that the covectors rol(O), ... , ron(O) are linearly independent.]
2. Let gl(t) be as in (7.70), and g'(l) = F(u, z). Show that g'(k) = F(u, kz) for 0::5: k ::5: 1.
[Hint: Consider the change of variable , = kt.]
320
8
Integration on manifolds
321
8 Integration on manifolds
Since r S n, Condition (3) states that the differential Dg(t) has maximum
possible rank. We see below that (3) can be replaced by an equivalent con-
dition (3'). When r = 1, Condition (3) states that g'(t) i= O. This is the same
condition imposed in Section 6.2 in defining parametric representations of
curves. The concepts of Section 6.2, such as equivalence, could be generalized
to r > 1. However, this is not done in this book.
PROOF. Let {tt. ... , t r} be the standard basis for Er. Since ~ is open, it
contains the <5-neighborhood of to for some <5 > O. For lsi < <5, let "'(s) =
g(to + std. Then", is of class C(l), and "'(s) E M.(One may think of", as
representing a curve on M through x o, obtained by fixing the components
t j = tb of t for j = 2, ... , r.) By definition of tangent vector (Section 4.7),
""(0) is a tangent vector to M at xo. But ""(0) = gl (to) by definition of partial
derivative. Thus gl (to) E TM(XO)' Similarly, the partial derivatives g2(t O), ... ,
322
8.1 Regular transformations
gr(t O) are tangent vectors to M at Xo. Since Dg(to) has rank r, the vectors
gl (to), ... , gr(t O) are linearly independent. Hence they form a basis for
TM(x O)' which is an r-dimensional vector subspace of En. 0
The vectors og/os, og/ot are linearly independent. By Proposition 8.1 they
form a basis for the tangent space at x = g(s, t). For instance, at the point
g(2, n13) = e j + J3 e z + 2e 3 , this basis is
By the same method of proof as for Proposition 8.1 one can show that
h = Dg(toHk) is a tangent vector to M at Xo, for any k E E' (see Problem 6).
The special choices k = tj, j = 1, ... , r, gave the basis in Proposition 8.1
for the space TM(x O) of tangent vectors at Xo. Thus we have:
Corollary. TM(xo) is the image of E' under the linear transformation Dg(to).
(8.1)
By Theorem 7.3, j"g(t) > 0 if and only if the vectors gj(t), ... , gr(t) are
linearly independent. Since these vectors are the columns of the matrix of
Dg(t), their linear independence is equivalent to the fact that Dg(t) has
maximum rank r. Thus condition (3) in the definition of regular transforma-
tion is equivalent to
323
8 Integration on manifolds
L V(Kd = L j'g(tk)~(Id,
k=l k=l
g -
Xo = g(t o)
h, = ag,(to)
Figure 8.1
324
8.1 Regular transformations
figures, this rough argument should make plausible the definition (8.2) of
~(A).
EXAMPLE I (continued). Let us find j'g(s, t). From the rules of exterior
multiplication
og og . .
os /\ at = [(cos t)e l + (sm t)e 2 + e 3J /\ [( -s sm t)e l + (s cos t)e 2J
og og .
os /\ at = se l2 - (s sm t)e31 - (s cos t)e23 ,
og og ocp ocp
ox /\ oy = e l2 - oy e 31 - ox e 23 ,
325
8 Integration on manifolds
=
g(.x, y)
~r------t-----y
x
(Jh)
R
Figure 8.2
Since the last line is positive, /g(x, y) > 0, Thus g is a regular transformation
from R onto S, The area of a set A c S lying above a measurable set B c R is
Let S = g(R). Then S is the r-manifold determined by «I> = (<1>1, ' .. , <l>n-r),
where <l>1(X) = Xl - cjJl(XA) for XED with D = {x: XA E R}. Then g is of class
°
C(1) and univalent. The explicit formula for /g(XA) is complicated. However,
8g 8g
-8 .- /\ '"
Xl!
/\ -8 . = e,.\
Xlr
+ other terms.
This r-vector is not 0 since its Ath component (the coefficient of e,.\) is 1. Hence
/g(x,.\) > 0, which shows that g is regular from R onto S. If we set XA(X) = x\
then X"\ projects En onto E' and g = (X"\ IS)-l.
326
8.1 Regular transformations
The case r = n
Let q, be a transformation of class C(l) from ~ c E' into E'. In this chapter we
call q, a fiat transformation. Condition (3) in the definition of regular trans-
formation above is equivalent to Jq,(t) i= 0 for every t E ~, where Jq,(t) is
the Jacobian. For flat transformations, this definition of regularity agrees
with that in Section 4.5. Regular flat transformations will be an essential
tool in relating overlapping coordinate systems on a manifold M. As a
first step in this direction, we show the following.
PROOF. Since g and q, are both C(l) and univalent, their composite g is also
of class Cl) and univalent. By the composite function theorem (Section 4.3),
Dg(t) = Dg[q,(t)] 0 Dq,(t). Since Dq,(t) is a nonsingular linear transformation
from E' onto E', and Dg[q,(t)] is a linear transformation of rank r, Dg(t)
also has rank r. Thus g is regular. It remains to prove (8.3). Given t, let
L = Dg[q,(t)]. If kb"" kr are linearly independent vectors in E' and
hj = L(k),j = 1, ... , r, then
By (7.38),
q,l(t) 1\ ... 1\ q,r(t) = det ¢~(t)el"'r
= Jq,(t)e 1 ··· r·
327
8 Integration on manifolds
~s
X'JlG
Figure 8.3
WR
PROOF. Consider any to E ~, and let Xo = g(to). It suffices to show that a
neighborhood Q of to exists such that g(Q) is a relatively open subset of M
containing Xo. If A = (iI' ... ,ir ) is an increasing r-tuple, then let Xi. =
(XiI, ... , xir), and XA(X) = XA as in Example 3 above.
By the implicit function theorem, there exist an increasing r-tuple A and a
neighborhood U of Xo with the following properties. Let S = M n U and
R = XA(S). Then R is an open subset of E' containing x o , XA IS is univalent,
and G = (XAIS)-l is a regular transformation from R onto S. Note that G
is the kind of regular transformation denoted by g in Example 3.
Let gA = X A g. Note that gA is the flat transformation obtained by
0
must also have rank r for any t E ~. Thus gA IQ is a regular flat transformation
from Q into R. By the corollary to the inverse function theorem (Section 4.5),
gA(Q) is an open subset of R. By the way G is constructed, G(R 1 ) is open
relative to M, for any open Rl c R. But g(Q) = G(R 1) when Rl = gA(Q).
Thus g(Q) is a relatively open subset of M containing Xo. 0
328
8.2 Coordinate systems on manifolds
PROBLEMS
1. For each of the following transformations from L1 c £2 into £3, find /g(s, t) and
g(L1). Show that g is univalent.
(a) g(s, t) = (s + t)e l + (s - 3t)e 2 + (-2s + 2t + 2)e 3 , L1 = {(s, t):O < 5 + t < I,
5>0,t>0}.
°
(b) g(p, e) = (p cos 1X)e1 + (p sin IX cos e)e2 + (p sin IX sin e)e3' < IX < rr/2 (IX fixed),
L1 = (0, if)) x (0, 2rr).
(c) g(s, t) = ste l + 5e 2 + te 3 , L1 = E2.
2. In each case find the tangent space Ts(xol to S = g(L1) at the indicated point Xo.
(a) In Problem I(a), take Xo = jel - je2 + 2e 3 ·
(b) In Problem I(b), take Xo = (cos lX)e l + )2/2 sin lX(e 2 + e3 ).
(c) In Problem I(c), take Xo = e l + e2 + e3 .
3. Let g(t) = (cos t)e l + (sin 2t)e 2 , L1 = (0, 3rr/2). Sketch g(L1). Show that g is univalent,
that /g(t) > 0, but that g(L1) is not a I-manifold.
V2 (A) = f {[- -
B
0(g2, g3)J2
0(5, t)
[0(g3, gl)J2
+ --
0(5, t)
[o(gl, g2)J2}1/2
+ --
c'l(s, t)
dV2 (s, t).
7. Let n = 4, r = 2, A = (1,2).
(a) Show that in Example 3,
/g(XI, x 2 ) = [1 + Igrad ¢lll + Igrad ¢212 + (¢l¢i - ¢i¢1)2]1/ 2 .
(b) Find V2 (A) if
A = {xEE4 :x 3 = (X I )2 - (X 2)2,X4 = 2XIX2, (X I )2 + (X 2)2:::; I).
329
8 Integration on manifolds
One might prefer a criterion in terms of F rather than its inverse, for F
to be a coordinate system. Such a criterion is given in Problem 5. On the
other hand, the inverse g rather than F is ordinarily used for calculations in
the sections to follow.
When r = n, we called in Section 5.9 any regular flat transformation
F = (Fl, ... , Fn) a coordinate system. The inverse g = F- 1 of a regular
flat transformation is regular. Hence the present definition is consistent with
that in Section 5.9, when r = n.
330
8.2 Coordinate systems on manifolds
Coordinate changes
We recall from Proposition 8.2 that the composite g = go q, of a regular
transformation g and a regular flat transformation q, is regular. If .:i and fl
are the respective domains and if .:i = q,(fl), then g(~) = g(fl). If we denote
this set by S, then S is a coordinate patch on M. Moreover, F = g- 1 and F
= g- 1 are both coordinate systems for S. In this way, any regular flat trans-
formation leads to a change of coordinates.
EXAMPLE I (continued). In this example, g(r, e) = g(r cos e, r sin e). Let us
replace fl in part (a) by fll = E2 - {(x, 0): x > O}. Such a change makes no
difference in computing integrals over a part of fl, since the half-line that has
been removed has 2-dimensional measure O. For (x, Y) E fll' let R(x, y) =
(x 2 + y2)1/2 and 8(x, y) the angle from the positive x-axis to (x, y) with
o < 8(x, y) < 2n. Then q, = (R, 0) and g = go q,.
Let us next consider overlapping coordinate patches Sand S on M,
with respective coordinate systems F and F. Let us show that on S n S
these coordinate systems are related by such a regular flat transformation
q,. Let
flo = F(S n S), ~o = F(S n S),
and g = F- 1 , g = F- 1 • Figure 8.4 shows the appropriate choice for q,.
fiat transformation from flo onto ~ 0, and g(t) = g[q,(t)] for all t E flo·
PROOF. The final assertion holds by construction. It remains to show that
q, is regular. Since q, is univalent by construction, it suffices to show that any
331
8 Integration on manifolds
Figure 8.4
332
8.2 Coordinate systems on manifolds
PROBLEMS
1. Let M = {(x,y,z):x 2 + 2y + Z2 = 3,z > O,y > Ixl}. Let F = (XIM, YIM) and
F = (X 1M, Z IM). Describe ~,,i, g, g, and q, (see Figure 8.4).
2. Let M = {(y2 + Z2, y, z): y > O} and let F(x, y, z) = (y + z, exp z) for (x, y, z) EM.
Show that F is a coordinate system for M and find F(M). [Hint: First take y and z
as coordinates on M and then find a suitable coordinate change q, giving the system
F.J
3. In Example 1, take (p,8) as coordinates on S, where p = (x 2 + l + Z2)1I2. Find
a regular flat transformation q, taking (p, 8) into (r, 8). Show that t = (g q,r I is
0
DF(x) ITM(x) is a linear transformation of rank r for each XES. Show that ~ = F(S)
is open and g = (FISr I is regular. [Hint: Consider G as in the proof of Proposition
8.3. Show that FoG is a regular flat transformation if R is a sufficiently small open
set containing x~.J
6. (Stereographic projection.) Let M be the sphere x 2 + l + (z - 1)2 = I. For each
X = (x, y, z) EM except the "north pole" 2e 3 , let (s, t, 0) be the point where the line
through 2e 3 and x meets the plane z = O. Let F(x) = (s, t).
(a) Show that F is a coordinate system for M - {2e3}'
(b) Let hlo h2 be tangent vectors to M at x, and let k/ = DF(x)(h/), I = 1,2. Show
that the angle between kl and k2 equals the angle between hi and h2.
7. Let fl, ... , 1', <1>1, ... , <1>n-r be functions of class C(l) on an open set D. Suppose
that F = (f I IS, ... , 1'1 S) is a coordinate system for S, that S = {x ED: cJ)(x) = O},
and that DcJ)(x) has rank n - r for every XED. Show that each Xo E S has a neighbor-
hood U such that (fl, .. . ,1', <1>1, ... , <1>n-r) restricted to U is a coordinate system
for U.
8. Let I :c:; r < n, and let jt(n, r) = {cx E E~: Icxl = I, cx is decomposable}. Identify
E~ with Em, and show that A(n, r) is a manifold of dimension r(n - r). [Hint:
Given CX o E A(n, r), let (VI"'" vn) be an orthonormal frame for En such that CXo =
v I 1\ ... 1\ Yr' Show that if CX is in a small enough neighborhood of cx o, then cx can
be uniquely written in the form
333
8 Integration on manifolds
The integral over B is taken in the sense of Section 5.6. By (8.1), j'g(t) =
Ig,(t) 1\ ... 1\ gr(t)l. Since gis ofcIass C<'), j'g is continuous. Hence(jo g)j'g
is continuous. If f 2: 0, then the integral over B either exists or diverges to
+ 00. When the latter occurs we agree that the integral of f over A also
diverges to + 00.
We must show that the integral does not depend on the particular choice
of coordinate system. Let S be another coordinate patch such that A c S,
and let F be a coordinate system for S. Let us adopt the notation of Figure 8.4.
Then g(t) = g[q,(t)] for all t E do and A = g(B) = gal), where B c do and
334
8.3 Measure and integration on manifolds
= Lf[g(t)]/g[«j)(t)]dV,(t),
as required.
EXAMPLE 2.Let M be a I-manifold and B a closed interval [a, b]. Using the
terminology of Section (6.2), A is the trace of the simple arc y represented on
[a, bJ by g. Formula (8.4) becomes
f f(x)dV (x) = f.
2 f(x)dV2 (x) = J
1C/2
d4>
J21C
f[g(4), 8)Jsin 4> d8.
H g(B) 0 0
335
8 Integration on manifolds
(8.5) f f dYr = f= f
A k 1 AnSk
f dYr.
Partition of unity
Let us recall from Section 5.3 that the support of a function t/I is the smallest
closed set outside of which t/I(x) = O. Let us first find for every Xo and l5 > 0
a function t/I of class C(OO) on En such that t/I(x) > 0 on the l5-neighborhood of
Xo and the support of t/I is the closure of that neighborhood. In fact, let
h(x) = 0, Ixl;:::: 1.
From the example at the end of Section 3.4 and the composite function
theorem, h is of class C( (0) on E 1 . Let
336
8.3 Measure and integration on manifolds
f I dv,. f
A
=
AnK
I dv,..
In particular, if {7r 1, ...• 7r m} is a partition of unity. then for any I the support
of I7rk is compact and lies in some coordinate patch. Hence the integral of
I7rk is defined.
(8.6)
fA
I dv,. = f f I7rk dv,..
k= 1 A
337
8 Integration on manifolds
Since the support of fXI is contained in some coordinate patch, its integral
over A can be written according to (8.4) as an integral over a set BeE'.
By Theorem 5.4, the integral over B of a finite sum is the sum of the integrals.
Hence
1= 1, ... , p,
Since the right -hand sides are equal, the integral off over A does not depend
on which partition of unity is chosen.
If f(x) = 1 for every x EM, then the integral gives the r-dimensional
measure
~(A)= f Jnkd~.
k= 1 A
When A is a subset of some coordinate patch, this agrees with the previous
definition. If ~(A) = 0, then A is called an r-null set. The integral has the
same elementary properties listed in Theorem 5.4 for r = n (Problem 11).
Moreover, ~ is countably additive (Problem 10).
(4) If K is any compact subset of M, then the support of 1tk meets K for
only finitely many k.
The sum in (3) is now an infinite series. However, on any compact set only
finitely many terms are different from 0. Every manifold has a partition
of unity. This can be proved by an elaboration of the proof of Proposition 8.6,
which we do not give.
Let f be continuous on A. Then f is called integrable over A if
Lk"= 1 SA I fink d~ is finite. If f is integrable over A, then its integral is
338
8.3 Measure and integration on manifolds
L
*Note. The line integral f ds was defined in Section 6.4 without requiring
that y be simple. If y is not simple, then its trace A = g([a, b]) need not be
contained in a I-manifold. There is a more general notion of r-dimensional
measure and integral for sets that are not necessarily subsets of an r-manifold.
The general formula, which becomes for simple arcs the one in Example 2, is
where N(x) is the multiplicity of the point x. For any r ;::: I there is a similar
formula
PROBLEMS
x'. = ~(~)
1 f x'dv,.(x),
. i = 1, ... , n.
v,.
A A
339
8 Integration on manifolds
I
the center of mass are
3. Show that 1e3 is the centroid of the hemisphere H in Example 3. Use spherical
coordinates on H.
4. Find the second moment about the z-axis of:
(a) The sphere x 2 + y2 + Z2 = 1.
(b) The triangle with vertices e l , e 2 , e 3 .
5. Let xo, XI' ... , x, be the vertices of an r-simplex ~. Show that the centroid of
~ is (r + 1)-I(XO + XI + ... + x,).
°
6. (Surfaces of revolution.) Let y be a simple arc (or simple closed curve) lying in the
half y > of the (x, y) plane. From Section 6.2, y has a standard representation G
on [0, I], where I is the length and IG'(s)1 = 1 for 0:::; s :::; I. Let g(s, t) = GI(s)e l +
G 2 (s) [(cos t)e 2 + (sin t)e 3 ], and let M = g((O, I) x [0,2n]).
(a) Prove Pappus's theorem: V2(M) = 2ny/, where (x, y) is the centroid of y.
(b) Find the area of a torus (doughnut) of major radius rl and minor radius r2'
7. Let S = {x: Ixi = I} be the unit (n - I)-sphere in En. Show that the (n - 1)-
measure of the "zone" {x E S: a < xn < b} depends only on the difference b - a
when n = 3, but this is false when II i= 3. Assume that - I :::; a < b :::; 1.
8. Let v(r) = IXnrn be the n-dimensional measure of a spherical n-ball of radius r. Show
that v'(r) is the (n - I)-measure of its boundary. [Hint: Spherical coordinates
(Section 5.9).] [Note: IXn was calculated in (5.46). If fJn = v,,- I [unit (n - I)-sphere],
then Pn = IllXn ·]
9. Let M be a r-manifold. Show that:
(a) Any set A open relative to M is r-measurable.
(b) If A is an r-measurable subset of M, then M - A is r-measurable.
(c) If A I, A 2, ... are r-measurable subsets of M, then A I U A2 U ... is r-measurable.
10. Let M be a compact r-manifold. Let A = Al U A2 U···, where Aio A 2, ... are
disjoint r-measurable subsetsofM. Show that v,.(A) = v,.(A I ) + v,.(A 2 ) + ···[Hint:
Use a partition of unity.]
11. Prove that the statements (1 )-(7) obtained by replacing n by r everywhere in Theorem
5.4 are true for integrals over r-measurable subsets of a compact r-manifold.
340
8.4 The divergence theorem
~
· r~ = L...
d IV ac..
i= 1 ax'
The divergence theorem equates the integral of div ~ over a set D with an
integral over the boundary fr D, provided D belongs to a suitably restricted
class of sets called regular domains.
EXAMPLE I. Let D = {x: 1 x 1 < 1 or 1 < 1 x 1 < 2}. Then fr D is the union of
two concentric (n - I)-spheres of radii 1 and 2. However, D is on both sides
of the inner (n - 1)-sphere.
Actually, this example is rather artificial. If D is the interior of its closure,
D is bounded, and fr D is an (n - 1)-manifold, then using the implicit function
theorem it can be shown that D is a regular domain.
From the definition, all exterior normals at x are positive scalar multiples
of any particular one. We shall be principally concerned with the unit exterior
normal, which will be denoted by v(x)( 1 v(x) 1 = 1).
Let us show that the exterior unit normal v is a continuous function on
fr D, if D is a regular domain. Let Xo be any point of fr D, and U, <I> as in the
definition above. By Corollary 1, Section 4.7, the vector n(x) = grad <I>(x) is
normal to fr D at x E (fr D) n U.
Let t/J(t) = <I>(x + tn(x)). Then t/J(O) = 0 and
There exists b > 0 such that t/J(t) < 0 for -b < t < 0 and t/J(t) > 0 for
o < t < b. Therefore grad <I>(x) is an exterior normal at x. The vector
v(x) = igrad<l>(x)I- 1 grad <I>(x)
341
8 Integration on manifolds
fr D
Figure 8.5
(8.7) r
lrrD
~(x)· v(x)dv,,_ 1(x) = r div ~(x)dv,,(x)
JD
The number ~(x) . v(x) is called the (exterior) normal component of the
co vector ~(x). Let us defer the proof until later in the present section. The
somewhat restrictive assumption that fr D is a C(1) manifold is made to
simplify the proof. The theorem is still true if fr D is not a manifold but
instead consists of a finite number of pieces of class C(1) intersecting in sets of
dimension n - 2. For example, if D is the interior of an n-cube then the
pieces are the faces, which are cubes of dimension n - 1 and intersect in
(n - 2)-dimensional cubes. This more general form ofthe divergence theorem
will be precisely stated at the end of the section. For certain special kinds of
sets D there is an easy proof of the theorem (Problem 3).
Note. In applying (8.7) we sometimes ignore the distinction between
vectors and covectors. If F is vector-valued with the same components as
~ (pi = Ci, i = 1, ... , n). then we set div F = div ~. On the left side of (8.7),
~ . v becomes F . v, the euclidean inner product of F and v.
For n = 3, the divergence theorem is often called Gauss's or Ostrogradsky's
theorem. It has various interesting physical interpretations. Let ~ be a force
field acting in some open set Do C E3. For each x E Do, ~(x) is the force
acting at x (Section 6.4). For notational simplicity, let us set M = fr D through-
out the discussion to follow. The number fM ~(x) . v(x)dV2(x) is called the
outward flux across the boundary M. The divergence theorem expresses
the outward flux as a volume integral over D. If D has small diameter and
contains x o , then the outward flux is approximately V3(D)div ~(xo). To
make this statement more precise let us prove the following.
342
8.4 The divergence theorem
f(x o) = .
hm
diamD-+O
1
-(-)
v" D
iD
f(x)dv,,(x).
In words, this formula says that given c > there exists D >
if D is any open set of diameter less than Dwith Xo ED, then
° ° such that
Corollary. Let ~ be of class C(1) on an open set Do. Then ~ is divergence free
if and only if
(*) L~(x)· v(X)dv,,_l(X) = 0, M = fr D,
343
8 Integration on manifolds
*Flows in En
°
Let t denote time, and imagine that points of E" move in such a way that a
point at x when t = moves to another position y at time t. For instance,
the movement might be due to the flow of a fluid in E3 (Section 8.5). If we
set y = Tt(x), then T t is a flat transformation from some portion of En into
E" for each t. Let us require that Tt+s = Tr Ts. Thus a point initially at x
0
Figure 8.6
{Tr } is a flow.
If D is an open subset of Do, then the flow takes points initially in D into
points of Dr = Tr(D) at time t. Let v(t) = v,,(Dr)' The derivative v'(t) is the rate
of change of the n-dimensional measure of Dr. Let Wr = (aTrlat). Then
provided cl D c Do. We leave the proof to the reader (Problem 4). In par-
ticular, (8.9) implies that the n-dimensional measure vet) is constant if
°
div Wo(x) = for all XED.
Green's formulas
Formulas (8.10) and (8.11) are consequences of the divergence theorem and
are useful in many applications. Letfbe of class e(2) on cl D, and let M =
fr D as above. Let!v(x) denote the derivative offin the direction of the exterior
normal at x E M, namely,
fv(x) = df(x) • vex).
344
8.4 The divergence theorem
Let cfJ be another function of class e(2) on cl D, and let ~(x) = cfJ(x}df(x}.
Then
~(x) . v(x} = cfJ(x}f.(x),
div ~ =.2:
n a ( af
-ai cfJ -ai ) = dcfJ . df + cfJ Lapl f
,= I X X
When n = 3 the right-hand side often has (except for a suitable multiplicative
constant) the physical interpretation of energy.
Ufis harmonic andf(x} = 0 for every x E M, then from (8.12) the integral
of the nonnegative continuous function Idf 12 is O. Hence df(x} = 0 for every
x E cl D. Given Xo E D let Xl be a point of M nearest Xo. The line joining Xo
and Xl lies in cl D, and from the mean value theoremfis constant on it. Since
f(x l } = 0, we must havef(x o} = O. Thusf(x) = 0 for every x E cl D.
Suppose that f and g are both of class e(2) on cl D and harmonic, and that
f(x} = g(x) for every X E M. Then cfJ(x} = f(x) - g(x) = 0 for X E M and cfJ is
harmonic. Hence cfJ(x} = 0, and f(x) = g(x), for every X E cl D. This shows
that there is at most one harmonic function of class e(2) on cl D with given
values on the boundary M. It is more difficult to show that there is in fact a
harmonic function f with given boundary values. This is called Dirichlet's
problem. Uthe boundary dataf IM are merely continuous, thenfis continuous
on cl D and of class C(2) and harmonic on D. [15, Chapter XI]. If the boundary
data are smooth enough, then f is of class e(2) and harmonic on cl D. For
instance this is true if M is of class e(3) andfis of class e(3) on M.
Let us now turn to the proof of the divergence theorem. The proof pro-
ceeds by first considering two particular cases (Lemmas 2 and 3). The general
result is reduced to these via a partition of unity. As in Chapter 5, .f f dv"
denotes the integral off over all of En.
345
8 Integration on manifolds
Lemma 2. Let ~ be a 1{orm of class e(t) on En such that ~ has compact support.
J
Then div ~ dv" = O.
PROOF. Let 1 sis n. By the iterated integrals theorem
where Xi' = (Xl, ... ,Xi-I, Xi+ I, ... ,xn). Since (i has compact support,
the inner integral is 0 by the fundamental theorem of calculus. Therefore
J O(i/OXi dv" = O. Summing from 1 to n we get the lemma. D
In Lemma 3 we write (as in Section 5.5) x' = (Xl, ... ,xn- I ) instead of xn'.
We return to the situation depicted in Figure 8.2. Let R c En - I be an open
set, and ¢ of class e(t) on R with ¢(x) > O. Let S = {(x', ¢(x')): x' E R}, and
v(x) = (vl(x), ... , vn(x)) the normal vector to S at x with [v(x)[ = 1 and
vn(x) > O. Let e be the cylindrical region bounded above by S and below by
the hyperplane xn = 0, namely,
e= {(x', xn): 0 < xn < ¢(x'), x' E R}.
L~::dv" = i(nVndv,,-I'
f ox: dv"
c
o(
=
{{
JR Jo ox: dxn
(",(X') iJ( }
dv,,_I(x').
for any x = (x', ¢(x') E S. To see this, let <I>(x) = xn - ¢(x') as in Example 2,
Section 8.1. Then
grad <I>(x) = en - grad ¢(x'),
[grad <I>(x) [ = (1 + [grad ¢(X'W)I/2.
Since v(x) = [grad <l>(x) [- I grad <I>(x), we get (**). In the same way as for
Example 1, Section 8.3, iffis continuous on S, then
346
8.4 The divergence theorem
Let us next show that the divergence theorem is correct for 1-forms with
sufficiently small support.
provided ~ has support contained in Vi' One then gets (8.7) by summing
from i = 1 to n, and taking ~ with support contained in V = V In··· n Vn.
For notational simplicity, consider i = n. The same proof applies when i < n.
By the implicit function theorem, there is a relatively open set S c fr D
with Xo E S, such that
S = {(x', 4>(x')): x' E R},
with R c en-I an open set and 4> of class e(l). Since xZ > and xZ = 4>(x o),
°
we may choose S small enough that 4>(x') > for all x E R. Let II> have the
°
properties in the definition of regular domain. Since grad lI>(x o) is a positive
scalar multiple of v(x o) and vn(xo) > 0, the partial derivative II>n(XO) is
°
positive. Let us choose a neighborhood Vof Xo sufficiently small that x' E R,
xn > 0, and II>n{x) > for any x = (x', xn) E V. Then lI>(x', .) is an increasing
function on the interval Ix' = {x n : (x', xn) E V} and <l>(x', 4>(x')) =0. Let
Vn c V be a neighborhood of Xo small enough that (x', 4>(x')) E V for any
x E Vn. For xED n Vn' we have <l>(x) < 0, <l>(x', 4>(x')) = 0, and <l>(x', )
increasing on I",. Thus, xn < 4>(x'). Therefore, D n VII c e with e as in
Lemma 3 (see Figure 8.7).
If ( has support contained in V n' then
i (nVndv,,_l = S(nvndv,,-I,
I i
frD S
i D
a(n
-a
X
n dv"
a(n
ox
=
Dn Un
-;;-n dv" =
C
a'n
-a
x
n dv".
Figure 8.7
By Lemma 4,
JfrD
r (1l:k~). V dv,,-I = r div(1l:k~)dv", k =
JD
1, ... , m.
r
JrD
~.vdv,,_1 = f Jr
k= I frD
(1l:k~)·vdv,,-I'
348
8.4 The divergence theorem
Figure 8.8
±f ~'Vdv,.-l
k= 1 Ak
= r div~dv,.
JD
provided ~ is of class C1) on cI D (see Figure 8.8). Let us say that such a set D
has a boundary which is piecewise of class C(l).
PROBLEMS
3. Prove the divergence theorem directly from the fundamental theorem of calculus
when Dis:
(a) The unit n-cube {x: 0 < Xi < I, i = 1, ... , n}.
(b) The standard n-simplex.
4. Consider a flow {T,}, and define W" v(t) = v,,(D,) as in the discussion preceding (8.9).
Show that:
(a) JT,(x) > 0 for all x E Do and t ~ o.
(b) (OjiJt)JT, = div W 0 when t = O.
(c) (iJ/iJt)JT, = JT, div Wo.
(d) Formula (8.9) holds.
5. Let f(x, y) = t log(x 2 + y2), (x, Y) #- (0,0). Let D = {(x, y): 0 < x 2 + y2 < a2}, and
let ~ = df Show that the left side of (8.7) is 271:. but the right side is o. Why does not
this contradict the divergence theorem?
349
8 Integration on manifolds
[Hint: Apply the second Green's formula with D and f as above and let a -+ O+.J
f3n is as in Problem 8, Section 8.3.
9. Let g(t)= Xo + L(t), where L is a rotation of En. Show that both sides of (8.7) equal
the corresponding integrals in which ~ is replaced by ~" = Ii= 1 ((i g)dgi and D 0
by fj = g-I(D).
- dm
dt
= r (pv). v d V
JrrD 2•
- ~7 = L div(pv)dV3'
On the other hand, by differentiating under the integral sign (Section 5.12),
dm
dt =
f. Jiop
D dV3 '
350
8.5 Fluid flow
For each to the functions -div(pv) and op/ot have the same integral over
every regular D with cl D c Do. By Lemma I, Section 8.4, for every Xo E Do
these functions have the same value at (xo, to). In other words,
~= -div(pv).
If the density p is constant, then the fluid is incompressible. Thus for incom-
pressible fluids div v = 0 at every time t.
In the above discussion the only physical principle used is conservation
of mass. To proceed further, some additional description of the physical
properties of the flow is needed. Let us consider the following simple model
of fluid flow. It is assumed that: (1) the fluid is incompressible, (2) the velocity
v(x) at a point x is not time dependent, and (3) there exists a function J of
class e(2) on Do such that v = grad!
Flows that satisfy (2) are called steady. The function J in (3) is called a
velocity potential. To better understand Condition (3), let us introduce the
concept of circulation along a curve y lying in the region Do of flow. Let us
choose arc length s as parameter on y, and let G be the corresponding
standard representation of yon [0, I], where I is the length of y (Section 6.2).
Then G'(s) is a unit tangent vector at the point x = G(s) on y. The tangential
component of the flow is v[G(s)] . G'(s), and the circulation along y is
c = f~V[G(S)]' G'(s)ds.
The circulation can be rewritten as follows. Consider the I-form
( = Vi dx + v2 dy + v3 dz, where vl(x), v2 (x), v3 (x) are the components of the
velocity v(x). Then
c = L~.
Condition (3) states that ~ = dJ is an exact differential form. By Theorem
6.1, Condition (3) is equivalent to the statement that the circulation along
any closed curve y lying in D is O.
A weaker assumption than (3) is that: (3') ~ is a closed differential form
(d~ = 0). The flow is called irrotational if (3') holds, for reasons indicated in
Section 8.8. Theorem 8.4 (Section 8.10), implies that (3) and (3') are equivalent
if Do is simply connected.
351
8 Integration on manifolds
352
8.6 Orientations
;:
~n Figure 8.9
:
~
flow. In particular, the solution does not account for eddies downstream from
the obstacle, in which the velocity vector differs greatly in direction from the
general direction of flow.
PROBLEMS
8.6 Orientations
Let M be an r-manifold. For each x E M, the tangent space TM(x) is an r-
dimensional vector subspace of En. According to Section 7.5 TM(x) has two
possible orientations, each of which is an r-vector of norm 1. If one of these
orientations is denoted by o(x), then the other is - o(x). We would like to
choose the orientation for TM(x) consistently on M: in other words, we want
the function 0 whose value at x is o(x) to be continuous on M.
353
8 Integration on manifolds
Figure 8.10
The case r = n
Here the n-manifold M is an open subset of En. The possible values for
o(x) are ±e l ... n • If M is connected, then o(x) must be constant. If o(x) =
e 1 " ' n for every x E M, then M is positively oriented; and if o(x) = -e 1 ... "
for every x E M, then M is negatively oriented.
The case r = n - 1
If M is an (n - I)-manifold in E", then the adjoint n(x) = *o(x) is a unit normal
vector to Mat x. The condition that M be orientable is that a unit normal can
be chosen continuously on M. If D is an open set that is a regular domain
(Section 8.4), then the exterior unit normal orients fr D. We call this the
positive orientation on fr D. If M is not the boundary of an open set, then M
may not be orientable. This is shown by the following famous surface.
Figure 8.11
354
8.6 Orientations
Figure 8.12
where OA(X) is the component of o(x) with respect to the standard basis
r-vector e Aand A. = (i1" .. ' ir) denotes an increasing r-tuple as in Section 7.5.
As in Section 8.1, let gA = (git, ... ,gir) be the flat transformation obtained by
considering only those components of g.
Proposition 8.8. The induced orientation agrees with o(x), at x = get), if OA(X)
and the Jacobian JgA(t) have the same sign. If OA(X) and JgA(t) have opposite
signs, then the two orientations are opposite.
PROOF. By (7.33) with hj = git), the Ath component of gl(t) A •.• A g.(t) is
JgA(t). Therefore oc~(x) = [,Ig(t)r IJgA(t). Since llo(X) = ±o(x), llo(X) = o(x)
if and only if ocMx) = OA(X). D
355
8 Integration on manifolds
PROBLEMS
3. Let M = {(x, y, z): l = x 2 + Z2 + I, y < O}, and 0 the orientation for M such that
031(X, y, z) < O. Let (r,O) be polar coordinates in the (x, z) plane, and g(r, 0) =
(r cos O)el - (r2 + 1)1/2e 2 + (r sin O)e3' Use Proposition 8.8 to show that the induced
orientation agrees with o.
4. Suppose that M is the r-manifold determined by <1>, in the sense that M satisfies
(4.26), Section 4.7. Show that Mis orientable.
(8.14) f f AO
ro =
A
ro(x) • o(x)dv,.(x),
356
8.7 Integrals of r-forms
(4) If Iro(x) I ::; C for every x E A, then ISA o rol ::; CY,.(A).
(5) SAO ro = SAr ro + SA~ ro if A = Al U A2 and Al n A2 is empty.
These follow at once from corresponding elementary properties of the
right-hand sid\! of (8.14) (see Problem 11, Section 8.3).
For instance, in (3),
The case r = n
Let A + denote A with the pOSItIve orientation el "' n of En. Let ro =
f dx l /\ ... /\ dxn be an n-form. Then
ro(x) • e l ... n = WI ... n(x) = f(x),
and (8.14) becomes
(8.15) f A+
fdxl /\ ... /\ dxn = fA
fdv".
The left-hand side of(8.15) changes sign if either the orientation of A is reversed
or two differentials dXi and dx i are interchanged. For instance, if n = 2 then
fA+
f dx /\ dy = f
A
f dV2 ,
A- f
fdx /\ dy =
A+
fdy /\ dx = f A
fdV2 • -f
Let us next show how to rewrite JAo ro if r < n and A is an r-measurable
subset of some coordinate patch S. Let g be regular from an open set ~ c E'
onto S. We recall from Section 7.7 that ro~ is the corresponding r-form on ~,
obtained formally by substituting everywhere g(t) for x and dg i for dXi.
(8.16) fAO
ro = r ro~
JB + '
A = g(B),
357
8 Integration on manifolds
PROOF. Given t, let x = g(t) and L = Dg(t). Recall that git) = L(E), where
{E 1, ... , Er} denotes the standard basis for Er. By (7.39),
gl(t) /\ ... /\ gr(t) = L r(E l /\ ... /\ Er) = L r(E 1 ... J
By (7.49), ro#(t) = L:Cro(x)J. Since E1"' r is the positive orientation for E',
J 23( 0) 8(g2,g3)
g r, = 8(r, 0) = r.
Compare with Example 1, Section 7.7. Since Jg B and 0 23 are both positive,
Proposition 8.8 implies that 0 agrees with the orientation induced by g.
Ifro = Pdx /\ dy + Qdz /\ dx + Rdy /\ dz, then
ro# = po g(dx /\ dy)# + Q a g(dz /\ dx)# + R a g(dy /\ dzt
358
8.7 Integrals of r-forms
We may use Proposition 8.9 here. Note that g(B) differs from A by a portion
of the curve on M corresponding to = 0 or 2n. The curve contributes e
nothing to integrals over A. For instance, let ol = xz dx /\ dy. Then ol' =
- 1'3 sin 2 e dr /\ de. By (8.16),
f AO
ro = ( (- 1'3 sin 2 e)dr /\ de = - ( r3 sin 2 e dVir, e)
JB + JB
= - (2"sin2 e de (\3 dr = - ~.
Jo Jo 4
Example 2 suggests how Proposition 8.8 and 8.9 can be generally applied.
By Formula (7.23) it suffices to consider ol = f dX i, /\ ... /\ dx ir , since any
r-form is the sum of terms of this type. Then
(dX i, /\ ... /\ dx ir )' = dg i' /\ ... /\ dg ir = Jgi'(t)dt 1 /\ •.• /\ dtr.
If OA(X) and JgA(t) have the same sign, then
(8.17) fAO
fdxi, /\ ... /\ dx ir = ( fog JgAdv,..
JB
(8.18) ( ol = ( dOl.
JaD + J D+
( ol = ( Ol-Odv,,-l = ( ~-Vdv,,-b
JaD + 1rrD JrrD
( dOl = ( div ~ dx 1 dxn = ( div ~ dv".
JD JD JD
/\ .. , /\
+ +
359
8 Integration on manifolds
The case n = 2
Suppose that fr D = C 1 u··· U Cm' where each Ck is the trace of a simple
closed curve 'l'k> and C 1, ••. , Cm are disjoint. The orientation is chosen by
selecting the unit tangent vector v(x, y} so that (v(x, y), v(x, y» is a positively
oriented orthonormal frame for E2. Intuitively speaking, this means that as
the boundary is traversed, D is always on the left. Then
i f oD+
0> =
YI
0> + ... + fYm
0>.
Figure 8.13
V2 (D} = 1
-2 icD+
xdy - ydx.
PROBLEMS
1. Let II c E" have the positive orientation and let g be a regular flat transformation
from II into E".
(a) Show that g induces the positive orientation on g(ll) if and only if Jg(t) > 0
for every tEll.
(b) Show that (8.16) becomes
f A+
f dx l /\ ••. /\ dx" = f
B+
fog Jg dt l /\ ••. /\ dt",
360
8.7 Integrals of r-forms
2. Let A = {(x, y, z) : y = x 2 + Z2, Y :::; 4}, oriented so that 031(X) > O. Evaluate:
(a) SAO z dx /\ dy. (b) SAO exp y dz /\ dx.
[Hint: Use polar coordinates in the (x, z)-plane.]
f AO
Z2 dy /\ dz.
5. Let n = 4 and M = {x: (X I)2 + (X 2)2 = 1, (X 3)2 + (X4)2 = I}. Let gis, t) = (cos S)el +
(sin S)e2 + (cos t)e3 + (sin t)e 4, 0 :::; s, t :::; 2n.
(a) Find the orientation 0 induced by g from the positive orientation of £2.
(b) Evaluate
fMO
dX3 /\ dx 4 + Xl x 3 dx l /\ dx4.
6. Let A = g(B), where g is as in Problem 2, Section 8.6, and 0 the orientation induced
by g. Using (8.16), show that
fAO
Pix, y, z)dx /\ dy = f B
Pix, y, <p(x, y))dVl(x, y).
11. Suppose that D = {(x, y): fix) < y < g(x), a < x < b} = {(x, y): <p(y) < x < ",(y),
e < y < d}. Show directly from the fundamental theorem of calculus and properties
of line integrals that
f cD+
N dy = f· aN dVl ,
D ax
fiJD+
M dx = - f ayaM dV
D
1·
Adding, we get the Green's theorem for regular domains of this special type (see
Figure 8.14).
361
8 Integration on manifolds
'----0
c ----,-
I I
I I
I :
~~--~a----------~b---x
Figure 8.14
12. The winding number of a closed curve y in E2 about a point (xo, Yo) not in the trace
of y is
_ 1
w(xo, Yo) - -
f
(x - xo)dy - (y - yo)dx
2 2 •
2n y (x - xo) + (y - Yo)
Let y be the positively oriented boundary of a regular domain D.
(a) Show that w(xo, Yo) = 1 if(xo, Yo) E D. [Hint: Apply Green's theorem to
DE = {(x, y) ED: (x - x o)2 + (y - Yo)2 2 e},
where e < dist[(x o, Yo), fr D]. Note that m = 2 in Formula (8.19).]
(b) Show that w(xo, Yo) = 0 if (xo, Yo) rt cI D.
362
8.8 Stokes's formula
Stokes's formula:
(8.20)
The case r = 2, n = 3
Here co = P dx + Q dy + R dz is a I-form and dco is a 2-form. In this case,
(8.20) equates the integral of dco over an oriented surface A in £3 with the
integral over the boundary of A oriented consistently with the orientation on
A. Figure 8.15 illustrates a case when 8Ao is a single simple closed curve y.
The set A is the image under g of B, and the orientation 0 on A is that induced
by g from the positive orientation on B. The curve y bounding B is taken
onto y by g, with corresponding orientations.
--+------------------s --~-------------------y
Figure 8.15
e e.
At the last step we used cos 2 = 1 - sin 2 In this example, one also has
L co = SilB. COl, whereoB+ is the perimeterofthe rectangle B = (0,1) x (0,2n:)
363
8 Integration on manifolds
(8.21) f f
cAo
ro =
A
curl ro(x) . n(x)dV2 (x).
The name Stokes's formula was traditionally applied to (8.21) and not its
generalization (8.20).
This is proved using a lemma similar to that for the proof of the corresponding
formula (8.8) for the divergence. Note that curl ro = 0 if and only if ro is
closed (dro = 0).
*Note 011 fluid flow. In Section 8.5 we defined the circulation c of a fluid
along a curve y lying in the region Do of flow. We saw that c = L ~, where ~ is
a I-form with the same components as the velocity vector. By Stokes's
formula, c = fAo d~ if Y = aAo and A c Do. We call the flow irrotational if
d~ = o. Thus irrotational flows are those for which the circulation is 0 along
any simple closed curve y that bounds such a surface A.
Some generalizations. Let M be an orientable manifold of class C(2).
We stated Stokes's formula above in case cl A is contained in some co-
ordinate patch. By using partitions of unity, this restriction can be removed.
364
8.8 Stokes's formula
f dco = 0
JMo
for every (r - 1){orm co of class C( 1) on M.
PROOF. Let {1t 1 , ••• , 1tm} be a partition of unity for M. Let g(k) be a regular
transformation of class e(2) from an open ·set.:1k c E' onto a coordinate patch
Sk containing the support of 1tk. Then
f d(1tkco) = ±
JMo
f
4t;
[d(1t kco)r = ± f
4t;
d(1t kCO)1 = 0,
by the divergence theorem and the fact that (1tk CO)1 = 0 near fr.:1 k since
1tk(X)= 0 outside a compact subset of Sk. Since Lk 1tk = 1 and Lk d1t k =
d(Lk 1tk) = 0,
Since M has empty boundary relative to itself, one would expect to obtain
o on the left-hand side of (8.20) when A = M. Proposition 8.10 states that
this is correct.
Now let M be any orient able r-manifold of class e(2). Let us call a relatively
open set A c M a regular domain on M if:
(1) cl A is a compact subset of M;
(2) the boundary K of A relative to M is an (r - 1)-manifold of class C(2);
(3) A is the interior, relative to M, of cl A.
Condition (3) rules out oddities of the sort in Example 1, Section 8.4, in
which A is on both sides of K. Problem 8 gives a condition equivalent to (3),
stated in terms of coordinate systems on M.
Let 0 be an orientation for M. If S is a coordinate patch that intersects K,
then K n S can be oriented consistently with 0, as explained earlier in the
present section. It can be shown that these orientations agree in overlapping
coordinate patches. This defines an orientation pon K. Let K with orientation
p be denoted by vAG.
Theorem 8.1. Let A be a regular domain on M, and let co be an (r - 1){orm
of class C(1) on cl A. Then
(8.20) f ilAO
co = f
AO
dco.
This theorem can be proved using the divergence theorem and a partition
of unity in much the same way as for Proposition 8.10. We do not give the
details.
365
8 Integration on manifolds
We have assumed that M is of class e(2), but Theorem 8.1 is still valid for
manifolds of class e(l). Moreover, the relative boundary K may be piecewise
of class Cll in the sense explained at the end of Section 8.4. For instance, if
M is an r-plane and A an r-simplex contained in M, then the boundary of A
relative to M is piecewise of class e(l).
PROBLEMS
1. Let ro = yz dx + x dy + dz. Let}' be the unit circle in the xy-plane, oriented in the
L
counterclockwise direction. Calculate ro and SAO dro and verify that they are equal,
where the orientation 0 is chosen so that vAo = }' and:
(a) A is the disk x 2 + y2 < 1 in the xy-plane.
(b) A = {(x, y, 1 - x 2 - y2): x 2 + i
< I}.
2. Let ro = z exp( - y)dx + z dy + y dz. Evaluate SAO dro when A is:
(a) The ellipsoid x 2/a 2 + y2/b 2 + Z2/C2 = 1 oriented by the exterior normal.
(b) The square with vertices 0, e l + e2, J"2e
3 , e l + e2 + fie 3, oriented so that
023(X) > O.
(c) The paraboloid y = x 2 + Z2 oriented so that 031(X) > O.
6. Let (I be the r-vector of an r-simplex So and Po, PI' ... , Pr the (r - I)-vectors of its
oriented faces (Problem 12, Section 7.5). Show that
366
8.9 Regular transformations of submanifolds
Proposition 8.11. Let g be of class c(1) from an open set /1 c g into En. If
Dg(to) has rank r, then there exists a neighborhood Q of to such that g(Q)
is an r-manifold.
PROOF. Since Dg(to) has rank r, there exists A = (i1>'''' ir) such that
Jg).(t o) "# 0, where g). = (git, ... ,gir). By the inverse function theorem, to
has a neighborhood Q such that g).IQ is a regular flat transformation and
Ro = g)'(Q) is open. Then glQ = Go (g).IQ), with G as in Figure 8.3
(Section 8.1), and g(Q) = G(Ro) is an r-manifold. 0
Theorem 8.2. Let g satisfy conditions (1), (2), and (3) in the definition ofregular
transformation (Section 8.1). Moreover, let g be open. Then g(/1) is an
r-manifold.
Corollary. Ifg satisfies (1), (2), and (3) and g-' is continuous, then g(/1) is an
r-manifold.
PROOF. Let /1, c /1 be open. Then g(/1 1) = (g-')-'(/1,) is open relative to
g(/1) by Theorem 2.6. Therefore, g is an open transformation. 0
367
8 Integration on manifolds
Theorem 8.3. Let g be regular from an open set Ll c E' into an r-manifold M.
Let Nell and P = g(N). Let 1 :s:: s < r. Then
(a) P is an s-manifold if and only if N is an s-manifold.
(b) If N is an s-manifold and x = g(t), tEN, the tangent space Tp(x) is the
image under Dg(t) of TN(t).
(8.23) Y = g(Z),
PROBLEMS
1. Let X!. = X!. 0 1.. as in Part (iii) of the proof of Theorem 8.3. Show that DX!.(t) has
rank s for any t E B. Hint. [DG(x!.)r 1 = DX!.(xll T,,(xl.
2. Let Q and M be r-manifolds, with Q cErn, M c En. Let g be a transformation of
class c( 1) on an open set containing Q, with g( Q) eM. Call g IQ regular if g IQ is
univalent and Dg(t) ITQ(t) has rank r for each t E Q. Let N c Q and P = g(N). Prove
that statements (a) and (b) in Theorem 8.3 remain true provided glQ is regular.
369
8 Integration on manifolds
applies when r = 1. The second is that D be star-shaped and applies for any
degree r.
The results in this section do not depend on the rest of Chapter 8, except
for the use of Green's theorem in the easy special case of a rectangle.
Homotopies
Let f and g be transformations of class e(2) from a set B c Em into a set
A c Eft. We are interested in whether it is possible to smoothly interpolate
in A between f and g. If this is possible then f and g are called homotopic in A.
To state this more precisely, let us consider the subset [0, 1] x B of Em + 1.
If f(t) = Xo for every t E [a, b], then one should think intuitively that
Ys shrinks to the point Xo as s --+ 0+. When A is an open subset of E2 this is
possible roughly speaking provided Ys does not loop around any holes which
may be present in A. In Figure 8.16, A has two holes and the curves Ys in the
figure are not null homotopic in A.
370
8.10 Closed and exact differential forms
Figure 8.16
In case g'(t) =F 0, (g, co) is just another notation for the line integral of co along
the curve represented by g.
Proposition 8.13. Let co be closed. Iff and g are strictly homotopic in D, then
(f, co) = (g, co).
PROOF. Let co' be the I-form on the rectangle R = [0,1] x [a, b] induced
by the transformation H. Since dco = 0, dco' = (dco)' = 0. By Green's
theorem,
The integral over oR + is the sum of the integrals over the four segments
AI' . .. , A4 indicated in Figure 8.17. Now
(OHi OHi)
co' = 2:1
n
i=
COi 0 H(dx i)' = 2:1
n
i=
Wi 0 H --;1 ds
uS
+ --;1
ut
dt ,
Hand OHi/ot being evaluated at (1, t). Since H(I, t) = g(t), the right-hand
side is just (g,co). Similarly, since H(O, t) = f(t),
f co' =
)..
-(f,co).
371
8 Integration on manifolds
)'3
b
A4 R ;'2
a
).,
Figure 8.17
EXAMPLE 2. Let n = 2 and let D be the plane with (0,0) removed. Let 0> =
°
(x dy - Y dx)/(x 2 + y2). Formally,o> = dE>, where E>(x, y) is the angle from
the positive x-axis to (x, y), < E>(x, y) < 2n. However, E> is defined only in
the plane with a slit removed even though 0> is defined and of class Coo) in D.
° °
For each integer m -=F let gm(t) = (cos mt)e\ + (sin mt)e 2, ~ t ~ 2n. Then
<gm,O» = 2mn, which shows that gm and g, are not strictly homotopic in D
when m -=F l. The transformation gm represents the unit circle traversed Im I
°
times, counterclockwise if m > and clockwise if m < 0.
Theorem 8.4. If D is a simply connected open subset of En, then every closed
110rm with domain D is exact.
PROOF. By Theorem 6.1, it suffices to show that L 0> = ° for every piecewise
smooth closed curve y lying in D. Let g be a representation of such a curve y
on [0, 1], such that g is piecewise of class e( 1). There is a seq uence gl' g2 , ... of
transformations of class e(OO) on [0,1] such that: (1) gm(O) = gm(l) for m =
1,2, ... ; (2) gm(t) -+ g(t) for every t E [0,1], and g~(t) -+ g'(t) except at the
(finitely many) points of discontinuity of g', as m -+ 00; (3) Igm(t) I and Ig~(t) I
372
8.10 Closed and exact differential forms
(8.24)
(*)
II o dl) = -
II 0 ds /\ d'9 +
II asal)I
0 ds /\
o t t of
ot j Jo f ds = Jo ot j ds, j = 1, ... ,m,
d II
o
f ds = .fI(:t oII f dS)dt j = II0 ds /\ d'f
)=
j
373
8 Integration on manifolds
(8.26)
With this formula we can readily deduce a result about closed forms,
called Poincare's lemma.
Figure 8.18
374
8.11 Motion of a particle
closed, then eo + ~ is closed and ceo is closed for any scalar c. Thus ;zr(D) is
a vector space over E1. Similarly, let rffr(D) denote the vector space consisting
of all exact r-forms of the type eo = d~ where ~ is of class C(oo) on D. Then
rffr(D) c ;zr(D). According to DeRham's theorem, the quotient vector space
yt'r(D) = ;zr(D)jrffr(D) is isomorphic to the r-dimensional cohomology group
of D with real coefficients. (The homology and cohomology groups of a space
are defined in algebraic topology. They contain a great deal of topological
information about the space.) In particular, every closed r-form is exact if
and only if yt'r(D) = O.
PROBLEMS
1. Let D be the solid torus obtained by rotating the circular disk (y - a)2 + Z2 < b2,
o < b < a, about the z-axis. Let y be the circular path traversed by the center of the
disk. Show that Jy (x dy - Y dx)/(x 2 + i) i= O. Hence by Corollary 1, y is not null
homotopic in D.
2. Let S be the sphere x 2 + y2 + Z2 = a2 , oriented by the unit exterior normal. Let
ro = p-3(X dy /\ dz + y dz /\ dx + z dx /\ dy),
p2 = x 2 + i + Z2,
[Note: In the proof of Theorem 8.4, let IjJ be a periodic extension of gi', and let
g~(t) = l(O) + J~ IjJm(x)dx, i = 1, ... , n.]
375
8 Integration on manifolds
(see Problem 1). The product x 1\ F(x) is called the torque. When (8.30) holds,
u'(r») u'(r)
x 1\ F(x) = X 1\ ( - -r- x = - -r- x 1\ x= o.
Thus, angular momentum is constant during motion of the particle. This result
has the following two consequences.
376
8.11 Motion of a particle
o Xi = g(t;). i = 1. 2
Figure 8.19
°
Consider two times t I, t 2, with < t 2 - t 1 and 8(t 2) - 8(t d :::;; 2n. Let
y denote the curve along which the particle moves for t 1 :::;; t :::;; t 2, and B the
region shown in Figure 8.19. Then using Green's theorem
377
8 Integration on manifolds
cp(r)
--r----L--------~-------r
Figure 8.20
This is a first-order differential equation for r(t). Let us suppose that cf>(r) =
A2(2mr2)-1 + u(r) has the shape indicated in Figure 8.20.
Now (mj2)r2 = E - cf>(r), and hence cf>(r) s E. Consider the case E < Eo,
and r 1, r2 the two solutions of cf>(r) = E (Figure 8.20). Then r 1 S r(t) S r2.
We seek those solutions for which r(t) = 0 only for isolated values of t. By
separation of variables, one can express t = t(r) as a function of r on each
time interval where r(t) is monotone. In fact, on such an interval
Moreover, for any ro in this interval, r(t) is a periodic function with period
Tj2, where
378
8.11 Motion of a particle
If we take
_ ( 2EA2)1/2
e - 1 + mk 2 '
This is the equation of a conic section (Kepler's first law). For E < 0, e < 1.
In this case the curve is an ellipse. For E = 0 and E > 0, the curve is a para-
bola and hyperbola, respectively. When E < 0, the minor and major semi-
axes of the ellipse are
c A
rl = (1 _ e2)1/2 = (2mIEI)I/2
c k
rz = 1 _ e2 = 21 E I'
By Equation (8.32), with B the interior of the ellipse and t 2 - t I = T,
AT A
2m = TCrlrZ = TCr2 (2mIEI)I/2'
Therefore
PROBLEMS
1. (a) Let f and g be functions of class C(1) on [a, b]. Show that
d df dg
- (f 1\ g) =- 1\ g +f 1\ - .
dt dt dt
(b) Derive (8.31).
2. Suppose that (8.28) and (8.30) hold and that angular momentum is O.
(a) Show that the particle moves on a straight line through O.
(b) Let u(r) = r + ,-1. Show that the motion is periodic for any energy level E > 2.
379
8 Integration on manifolds
Then
N
(8.37) T[q(t), q(t)] = j~' ~j Iiy
380
8.12 Motion of several particles
is the kinetic energy at time t. The matrix (ak/(q)) is symmetric and positive
definite for each q. For k = 1, ... , n,
(838) oT ~ '1 ~ • oG
. ~ = L, aklq = L, mj x j '""""3lkj
uq 1=1 j=1 uq
(8.39) I ~
-
oapl 'p'l
L, - q q
2 p,l= 1 oqk
By a short calculation (Problem 3), the equations of motion (8.36) imply
(in the coordinates q)
These are called Lagrange's equations for the motion of the particles. The
total energy E = T + V is constant during the motion (Problem 2).
Lagrange's equations furnish a starting point for further development
of theory. Hamilton's equations furnish another useful formulation (Problem
5) [10; 17, Chapter 13].
PROBLEMS
2T =
naT It·]
I -;;;
aq
k=l
381
8 Integration on manifolds
oL k -I
(8.44) oil =
- -7tk' - , ••• , n•
(8.45)
aH
-a = q ,
"k
k = 1, ... , n.
7tk
[Note: 'I' is called the canonical transformation in mechanics, and (8.46) are the
equations of motion in Hamiltonian form.]
382
Appendix
383
Appendix
A basis for "Y is a linearly independent set that spans "Y. If "Y is finite
dimensional, then every basis B has the same number n of elements [12, p. 43].
The number n is the dimension of 'Or. If n = 0, then "f' has the single element
e. If n > 0 and B = {u 1 , ••• , un} is a basis for "Y, then every UE "Y can be
uniquely written as a linear combination
U = C1U1 + ... + cnu n.
A non empty set [ljJ c "Y is called a vector subspace of "Y if: u, v E [ljJ
implies U + v E [ljJ and U E [ljJ implies cu E [ljJ for any real c. In other words,
this says that [ljJ when provided with the addition and scalar multiplication in
"Y is a vector space.
Let "Y and "/fI be vector spaces. Let L be a function with domain "Y and
values· in "/fl. Then L is linear if
(a) L(u + v) = L(u) + L(v) for every u, v E "Y.
(b) L(cu) = cL(u) for every u E "Y and real c.
Let Land M be linear. The sum L +M is given by
(L + M)(u) = L(u) + M(u)
for every u E "Y. The function L + M has Properties (a) and (b), and thus
is linear. If c is a real number, then cL is the linear function given by (cL)(u) =
cL(u) for every u E "Y.
Let !£l("Y, "/fI) denote the set of all linear functions with domain "Y and
values in "/fI, together with these operations of sum of functions and multi-
plication offunctions by scalars. Then !£l("Y, "/fI) satisfies Axioms (1) through
(5) for a vector space. The zero element of !£l("Y, "/fI) is the function whose
value at every u E "Y is the zero element of "/fl.
384
A.2 Mean value theorem: Taylor's theorem
L(u) = I ai ci = L aiLi(u).
i= 1 i= 1
Mean value theorem. Let f be real valued and continuous on a closed interval
[a, b], and let the derivative f'(x) exist for every x E (a, b). Then there
exists c E (a, b) such that
f(b) - f(a) = f'(c)(b - a).
PROOF. Let m = [f(b) - f(a)J/(b - a) and let F(x) = f(b) - f(x) - m(b - x).
Then F is continuous on [a, b] and F'(x) = - f'(x) + m for x E (a, b). Since
[a, b] is compact, F has a maximum and a minimum value on [a, b]. If the
maximum value is positive, then since F(a) = F(b) = 0, the maximum must
occur at some Xl E (a, b). By elementary calculus F'(Xl) = 0 and we may
take c = Xl' Similarly, if the minimum value is negative we may take c = X2'
where F(X2) is the minimum value. If neither of these possibilities occurs, then
F(x) = 0 on [a, b] and c is arbitrary. 0
385
Appendix
Taylor's theorem with remainder. Let f, together with its deril'atil'es /"
/", ... ,f(q- I), be continuous on a closed interval [a, b] and letthe qth-order
derivative f(q)(x) exist for every x E (a, b). Then there exists c E (a, b) such
that
where
R = pq)(c) (b _ )q
q , a .
q.
PROOF. Let
f"(x)
G(x) = f(b) - f(x) - f'(x)(b - x) - 2! (b - xf
f(q-l)(X) K
- ... - (b - x)q- 1 - (b - x)q
-
(q-1)! q!'
where the number K is so chosen that G(a) = O. Then G(b) = 0 and, using
the product rule,
G'(x) = (b - X)q-l [ -f(q)(x) + K].
(q - 1)!
Repeating the reasoning in the proof of the mean value theorem, there exists
CE(a, b) such that G'(c) = O. Then pq)(c) = K. 0
where
a = to < t 1 < ... < t m- 1 < tm = b,
and
Jl = max{t 1 - to,t2 - t 1 ,···,tm - tm-d·
More generally, t~e Riemann integral exists for any bounded function
with a finite number of discontinuities. It agrees with the integral in Lebesgue's
sense, which is defined in Chapter 5 for a much wider class of functions.
386
AJ Review of Riemann integration
F(t) = ff(s)ds, as t S b.
Since f is continuous, given e > 0 there exists b > 0 such that f(t) - e <
f(s) < f(t) + e whenever Is - t I < b. Then if h < b,
)
. f( t-e<
F(t + h)h - F(t)
<
f( )
t+e.
Hence
In the theorem, F'(a) means the right-hand derivative and F'(b) means
the left-hand derivative.
The fundamental theorem says that F is an antiderivative of f If G is
any antiderivative of f, then G/(t) - F'(t) = 0 for every t E [a, b], and by
the mean value theorem G(t) - F(t) is constant on [a, b]. Thus G(t) - F(t) =
G(a) - F(a) = G(a), and upon setting t = b we obtain
387
Appendix
ff(t)dt = f f[<J>(r)]<J>'(r)dr.
This is the formula for change of variables in integrals. If <J>'(r) :::; 0 for every
r E [~, f3], then <J>(~) ~ <J>(f3). The same formula holds if we agree that
f = -i~
A.4 Monotone functions
Letfbe real valued with domain S eEl.
388
References
389
References
20. Nehari, Z. 1968. Introduction to Complex Analysis. Boston, Mass.: Allyn and Bacon.
21. De Rham, G. 1955. Varieties Differentiables, Actualites Scientifiques et Industrielles,
No. 1222, Paris: Hermann.
22. Taylor, A. 1958. Introduction to Functional Analysis. New York: Wiley.
23. Titchmarsh, E. C. 1939. The Theory of Functions. London: Oxford University Press.
24. Whitney, H. 1957. Geometric Integration Theory. Princeton: Princeton University
Press.
25. Willmore, T. 1. 1959. An Introduction to Differential Geometry. London: Oxford
University Press.
390
Answers to problems
SECTION 1.1
1 (a) 2, 1.
(b) fi, no lower bound.
(c) fi, -1.
(d) 0, _e- 1 .
(e) 1, t.
sup S E S for (b), (c), (e); inf S E S for (c), (d).
SECTION 1.2
1 4e 1 -+ e 3 + 3e4 , -2e 1 - e 3 + e4 , j30, j6, j6 j12,
2e l 6.
8 V4 = ±(j2jlO)(4e 1 + 3e l - 3e 3 + 4e4 ).
9 Jm, m = 1, 2, ... , n.
SECTION 1.3
1 {x:x 1 + Xl + 4x 3 = I}.
2 (a) {x:3x 1 - 3Xl - 3x 3 - X4 = O}.
(b) t = ~.
SECTION 1.4
2 (a) {x:O<lx-xol<b}, {x:lxl=Oorb}, {x:lxl=:;b}.
(b) Empty set, A, A.
391
Answers to problems
(c) A, the union of the half-lines y = 0, y = x + 1, x ~ -1, {(x, y): 0 :5: Y :5: x + 1,
x ~ -I}.
(d) A, the union of the circle x 2 + y2 = 1 and the line segment joining (0, 0) and
(1,0), the closed circular disk x 2 + y2 :5: 1.
(e) Empty set, E2, E2.
(f) Empty set, A, A.
(g) Empty set, A u {O}, A u {O}.
3 Open in (c), (d); closed in (b), (f).
SECTION 1.5
7 (b) The barycenter is at the intersection of the line segments that join the vertices
with the barycenters of the opposite (r - I)-dimensional faces.
SECTION 2.1
1 (a) [-1, 1], [0, 1], {x: x = 2mn: + y, y E [ - ~,~J m any integer}.
SECTION 2.2
(a) t.
(b) No limit.
(c) t.
(d) (1,5) = e l + 5e 2 •
(e) No limit.
Continuity: In (a), (b), (e) except at(O, 0). In (c) except at O. If we setj(O) = 1. then jis
also continuous at O. In (d) at every point of EI.
5 Continuous at any (x, y) with y of. 0, also at (0,0).
6 (a) O. (c) No limit.
(b) No limit. (d) 1.
SECTION 2.3
1 (a) O.
(b) No limit.
(c) o.
2 (a) (-t, 0).
(b) No limit.
(c) (1, 0).
392
Answers to problems
SECTION 2.4
1 (a) ± 1.
(b) No accumulation points.
SECTION 2.5
1 (a) S is not closed.
(b) S is not bounded.
SECTION 2.6
3 (a) Both open and closed relative to S.
(b) Closed relative to S.
(c) Neither.
4 (a) Closed relative to S.
(b) Open relative to S.
SECTION 2.11
(a) The square {(x, y): Ixl :-:; 1, Iyl :-:; I}.
2 (a) II(x, y)11 = [x 2 + xy + 41]1/2.
(b) 2.
SECTION 3.1
(a) fl(X, y) = 1 + log(xy), f2(X, y) = x/y
(b) fl(x, y, z) = 6x(x 2 + 2yZ + z)Z, fz(x, y, z) = 12y(xz + 21 + z)Z,
f3(X, y, z) = 3(x Z + 2yZ + z)Z
(c) j;(x) = 2Xi
SECTION 3.2
1 (a) a = e l + e Z + 2e 3.
(b) The plane x + y + 2z = c in E3. The line through
c- 2 c
el + ez + - - e 3 and - e3.
2 2
393
Answers to problems
SECTION 3.3
1 It has the equation 4x + 5y + z + 4 = o.
2 (a) )5.
(b) I/J2e.
(c) o.
3 (a) [2x(x Z + 2y + 1)-1 + cos(xZ)]e l + 2(x z + 2y + 1)- l e 2.
(b) 0.09.
4 (a) xo.
(b) lxi-Ix.
(c) 2(xo· x)xo.
6 (b) {(x, y): x = yZ, x#- O}, {(x, y): x = - yZ, x#- O}.
(c) The union of the sets in 6(b) and the x-axis, with (0,0) excluded since f is not
differentiable there.
(d) {(x, y): x = yZ(c- 1 ± )c- 2 - I), x#- O} if lei s 1, c #- 0; the x- and y-axes if
c = O.
SECTION 3.4
1 xyz = -z + (y + l)z - (x - l)z + (x - l)(y + 1)z
2 (a) f(x, y) = 1 - (x - 1) + Rz(x, y);
IRz(x,y)1 s m- 3 [(x - 1)z + lJ where m = min {x, I}.
7 (c) fdO,O) = -1, f21(O,O) = 1.
SECTION 3.5
1 (a) Maximum at -}el.
(b) Saddle point at -e l + 2e z .
(c) Maximum at each point where xy = nl2 + 2mn; minimum at each point where
xy = -n12 + 2mn, m any integer. Saddle point at 0
(d) Saddle points at 0 and at e l .
2 (a) Maximum at -e l + e z , saddle point at 1< -el + e z).
(b) Saddle points at -e z and at e l + e 2 .
(c) Saddle points at mne l , m any integer.
5 x = (1Im)(x I + ... + xm). The minimum value is
6 (a) i, -2.
(b) sin 1, -sin 1.
394
Answers to problems
7 (a) Let e be any number such that IjI'(e) = O. Then an points x on the hyperplane
{x: a· x = e} are critical.
(b) All points on the line y = x are critical.
SECTION 3.6
1 If f(x) = aox4 + a 1x 3 + a2x2 + a3x + a4, then ao > 0, 8aoa2 ~ 3ai.
2 (a) Neither.
(b) Concave, not strictly.
(c) Convex if p :-;; 0 or p ~ 1. Concave if 0 :-;; p :-;; 1. Not strictly.
(d) Strictly convex.
(e) Neither. However, f is strictly convex on each half of {(x, y): 2xy < -I}.
4 (a) a = 1
(b) a 2 satisfies the equation cot a2 = 2a 2, a2 < n12.
SECTION 4.1
1 ( 1 1
-2 0 1
5). The rank is 2. The kernel consists of an scalar multiples of
tl - I1t2 + 2t3'
3 (a) The diagonal elements are e l , ... , en; all other elements are O.
(b) (L -I )i(X) = (ei)-I Xi provided e i #- 0 for every i.
SECTION 4.2
1 (a) Reflection in the line s = t.
(b) They are rotations through angle 3n12, nl2 respectively.
2 L is not a rotation.
4 (b) If g(t) = L(t) + x o , then L must be nonsingular.
SECTION 4.3
1 (a) The parabola x = c 2 + y2/4e 2 if c #- 0; the positive x-axis if e = O.
(b) The lines t = s(k ± Jk2=l) if k = 11m, m2 < 1; the line t = s if m = 1; the
line t = - s if m = - 1; {(O, O)} if m 2 > 1.
(c) {(x, y) E Q: x :-;; a2 }.
SECTION 4.4
F 12 = xfl2 + Xyf22 + f2' the partial derivatives of f being evaluated at (x, xy).
2 FI = fl + f3gl, F2 = f2 + f3g2'
Fll = fll + 2f13g1 + f33(gl)2 + f3gll'
F12 = f12 + f1392 + f32g1 + f33glg2 + f3912,
F 22 = f22 + 2f23g2 + fdg2)2 + f3g22'
3 (a) 162
(b) -1>I@1>'H).
SECTION 4.5
5 (b) (gl.1)-I(X, y) = log R(x, Y)EI + 8(x, y)E2' where R(x, y) = (x 2 + yl)I/2, 8(x, y)
is the angle from the positive x-axis to (x, y).
(c) g(E2) = E2 - {(O,O)}.
396
Answers to problems
SECTION 4.6
1 (j/ = -<I>d<l>2' 4Y" = -[(<1>2)2<1>11 - 2<1>1<1>2<1>12 + (<I>d 2<1>22]/(<I>2)3.
2 4YII = -[(<1>2)2<1>11 - 2<1>1<1>2<1>12 + (<I>d 2<1>22]/(<I>d 3.
3 4YI = -1, 1/11 = 0, 4Y2 = 0, 1/12 = 1 at (-1,1).
SECTION 4.7
1 {(x, y): F(x, y) = c} is an ellipse if log c > 2, is the one point set {(O, O)} if log c = 2,
and is empty iflog c < 2. Any ellipse is a I-manifold.
2 (a) The cone is not a 2-manifold.
(b) 2(x - 2) - (y + 1) + 4(z - I) = o.
7 No.
SECTION 4.8
1 l
3 JI4/3.
9 (a) Al = A2 = I, A3 = - 1.
SECTION 5.1
2 V(Y) = 4, ViZ) = 12, V(Yu Z) = 13, V(Y n Z) = 3.
3 V(IIU[2)=¥, V(Iln[2)=!.
4 (e - l)eXP(~) I/m .
m exp(1/m) - 1
SECTION 5.2
5 239n/240.
SECTION 5.3
1 (a) Unbounded; (- 00, OJ.
(b) Bounded; £2.
(c) Bounded; £2.
(d) Bounded; {(x,y):y2:2: x 2 , Ixl + Iyl:::; I}.
3 1.
397
Answers to problems
SECTION 5.5
1 (a) t, (t, I)·
(b) 1 + n/2, (x, 0) where x = 2/(6 + 3n).
2 (1 - 3e- 2 )/4.
6 (b) (e - If.
(e) 2 - 2n.
SECTION 5.6
1 (a) Exists.
(b) Exists if p < 1; divergent if p ~ 1.
(e) Exists.
(d) Exists.
(e) Exists if p < 1, p + q > 1; divergent otherwise.
(f) Diverges.
(g) Diverges.
5 (a) O.
(b) n.
(e) O.
SECTION 5.7
1
6'
2 7.
SECTION 5.8
2 2 log 2.
3 1p-.
SECTION 5.9
1 2a s/15.
2 i.
3 t.
398
Answers to problems
4 I X/2
0 0 0
de
I[COS 6+ sin 6) - 1
r dr
I9(r cos 6, r sin 6)
f[r cos e, r sin e, z Jdz.
4n
5 - (a 2 _ b2)3/2.
3
7 n/2.
10 n 2 /4.
11 (a) trH)r(t)freh
(b) r(iW(1)fr(i).
(c) ~ r(a : I}
1
(d) ( - -
)c+' r(c + I).
d+1
(e) r(k + l)fr(n + k + 1).
SECTION 5.11
2 If
t
dV2 = ~
2-p
+ nt'-2IP(~)
2-p
if p of- 2;
If
t
dV2 = n(l + log t) if p = 2.
SECTION 5.12
1 (a) 2 tan -'(lft).
1 + exp(nt)
(b) t2 + 1
SECTION 6.1
1 Y = 2(x - fi)·
2 Y - 1 = 1{x - I), z - 1 = 1<x - 1).
3 fi( -e, + e2 )·
399
Answers to problems
SECTION 6.2
1 (a) Simple closed curve.
(b) Neither.
(c) Simple arc.
2 (b) -1-;[(4 + 9b)3!2 - 8].
SECTION 6.3
2 M2 = NI, N3 = O2 ,0 1 = M 3.
3 (a) f(x, y) = tx 2y + c.
(b) Not exact.
(c) Not exact.
(d) f(x, y) = x/y - y/x + cjJ(x, y), where cjJ is constant on each of the four quadrants
into which the coordinate axes divide E2.
SECTION 6.4
1 (a) taco
(b) nab.
2 (a) Ii.
(e) 2,¥.
(d) -i.
5 f(x, y, z) = tcjJ(p2), where cjJ(u) = J~ I/J(v)dv.
7 2J6.
8 (a) ne 3 •
(b) tJ2n 3 + 2J2n.
SECTION 6.6
3 (b) T = Cpa!(1 +a).
SECTION 7.1
3 (a) _e 12 - 3e 13 - 3e23
(b) -xy dx 1\ dy + x 2y dx 1\ dz + (3x + z)dy 1\ dz.
4 (a) 4xy dx 1\ dy.
(b) O.
(c) O.
(d) (of/ox)dx 1\ dy.
400
Answers to problems
SECTION 7.2
1 -1, 0, 1, 0, O.
2 -1.
SECTION 7.3
2 (a) 6e l2 + 2e l3 _ e 23 .
(b) O.
(c) _3e I23 .
(d) 4e 123.
3 (a) _eI2345.
(b) e l235 _ e l234 _ e 1345 .
SECTION 7.4
1 (a) 2xy sin(xi)dx /\ dy /\ dz.
(b) 3 dx /\ dy /\ dz.
3 (a) ~ i:
n i= I
(_I)i+ IXi dx l /\ ... /\ dX i - 1 /\ dx i + I /\ ... /\ dx".
SECTION 7.5
(a) -e 2345 .
(b) O.
(c) -e I46 •
(d) 2e 6 , + e 2, - 2e l " where i' = (I, ... , i - I , i + 1, ... ,6).
(e) 2e l ... 6 •
2 (a) 2.
(b) O.
(c) -1.
(d) - 3.
3 Negative; - 1
4 Negative.
5 No.
6 -fi12.
401
Answers to problems
SECTION 7.6
1 (a) Ifb = L*(a), then b l = al - a2 + 2a 3, b2 = -2al + 3a3'
(b) If ~ = CE12, then L2(~) = CV l A V2 = c(2e 12 + 7e13 - 3e23 ), where VI and v2
are the column vectors.
(c) If (0 = W12e12 + w13 e 13 + W23e23, then L!((O) = (-2W12 + 7W13 - 3W23)E12.
(d) O.
(e) j6i.
SECTION 7.7
1 (a) st(cos t ds - s sin t dt).
(b) -st cos t exp t ds A dt.
(c) 0
(d) f(st, s cos t, exp t).
3 (01 = M g dg l
0 + No g dg 2 • (d(O)' = (-aN - -OM) 0
O(gl, g2)
g -~- - ds A dt.
iJy iJx o(s, t)
4 (a) fog s ds A dt A duo
(b) s cos t sin t ds A du + S2 cos 2 t dt A duo
SECTION 7.9
SECTION 8.1
1 (a) 4.)3; the triangle with vertices 2e 3, e l + e 2, e l - 3e2 + 4e3'
(b) p sin IX; a cone from which points (x, y, 0), y ;::: 0, are deleted.
(c) [1 + S2 + t 2]1/2; the hyperboloid x = yz.
2 A basis for T..(x o) is:
(a) {e l + e 2 - 2e3, e l - 3e2 + 2e3}.
(c) {e l + e 2, e l + e 3}.
SECTION 8.2
1 8 = {(x, y): Ixl < 1, Ixl < y < (3 - x 2 )/2},
~ = {(x,z):lxl < 1,z > O,Z2 < 3 - x 2 - 2Ixl}.
g(x, y) = xe l + ye2 + (3 - x 2 - 2y)1/2e 3,
g(x, z) = xel + 1(3 - x 2 - Z2)e2 + ze 3,
4>(x, y) = (x, (3 - x 2 - 2y)1/2).
2 F(M) = {(s, t): t < exp s}.
402
Answers to problems
SECTION 8.3
2
4 (a) 8n/3.
(b) -!3/6.
6 (b) 4n 2 r l r z .
SECTION 8.4
1 The integrals equal O.
2 (a) -n/4. (b) n/4.
SECTION 8.5
2 v(x,y) = C l r- 2 (xe l + ye z ).
SECTION 8.6
1 (c) Four possible orientations; on each half of the hyperbola M use either orientation
v or orientation -v, with v as in part (a).
SECTION 8.7
2 (a) - 8n. (b) n(e 4 - 1).
3 t.
4 t.
5 (a) o(x) = [( - sin S)el + (cos s)ez] /\ [( - sin t)e3 + (cos t)ell
(b) n 2 •
8 112 ,
SECTION 8.8
1 n.
2 (a) O.
(b) J2(1 - e- I ).
(c) O.
403
Index
A c
Absolute curvature 253 Canonical transformation 382
Accumulation point 43 Cantor's
Adiabatic curve 272 function 180
Adjoint set 47
of a linear transformation 127 theorem 41
of a multicovector 314 Cartesian product 28
of a multivector 311 Cauchy
Affine transformation 125 convergence criterion 40
Almost everywhere 231 inequality 6
Alternating multilinear function 283 principal value 205
Analytic function (real) 97 sequence 37, 64
Angle between two vectors 7 Cauchy- Riemann equations 134
Archimedean property 4 Center of mass 197
Centroids 197, 265, 339
Chain rule 136
Change of variables in Riemann
B integrals 387
Characteristic
Ball, n-dimensional 12 function 187
unit 198, 220, 340 values and vectors 163
Banach space 65 Closed
Barycenter 27 differential form 265, 294
Barycentric coordinates 24 set 17, 51
Beta function 219 Closure of a set 15,51
Bilinear function 71 Codifferential 315
alternating 277 Compact
Bolzano-Weierstrass theorem 45 set 60
Boundary (or frontier) of a set 15,51 topological space 60
Bounded set 41 Comparison tests 43, 203
transformation (or function) 33 Complete metric space 64
405
Index
406
Index
407
Index
408
Index
409
Index
410
Index
T v
Tangent Vector 5
plane to a manifold 84, 157 Vector space 383
space 156, 323 bases 384
vector ISS, 246, 252 dimension 384
Taylor's subspaces 120, 384
formula 94
series 97
theorem 386 w
Tensor 290, 299 Wave equation 139
algebra 291 0' Alembert's solution, spherical
field 311 waves 139
Thermal system 272 Whitney's extension theorems 96
extensive, intensive variables 272 Work 263
Thermodynamics, first and second
laws 273
Topological space, axioms for 50
Topology of E"
basic notions 14
relative topology 51
Totally disconnected set 59
Transformation 29
affine 125
bounded 33
composition 134
conformal 134
continuous 34, 48
differentiable 128, 131
flat (regular) 327
Jacobian of a 129
linear 120
of class C<q) 131
of integrals 209
open 367
orthogonal 126
partial derivative of 129
regular 143, 322
Translation 125
Triangle inequality 6, 63
Triple scalar product 317
u
Uniform
continuity 50
convergence of sequences 68
distance, norm 67
Upper bound 3
411