Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Advanced Calculus

Download as pdf or txt
Download as pdf or txt
You are on page 1of 210

University of Chicago A Textbook for Advanced Calculus

John Boller and Paul J. Sally, Jr.

Chapter 0

Number Systems and Cardinality


Dans la pr esente Note, on va essayer de pr eciser une terminologie propre ` a l etude des ensembles abstraits. Cette etude a pour but principal d etendre les propri et es des ensembles lin eaires ` a des ensembles de plus en plus g en eraux, et par voie de cons equence, de diss equer ces propri et es et den rechercher pour chacune la v eritable origine. Il en r esulte que le fond des d emonstrations est pour ainsi dire donn e davance et que la dicult e consiste en grande partie a pr eciser dans quel cadre elles viendront se placer. Adopter une terminologie, cest donc tracer davance toute la th eorie. Au fur et ` a mesure du d eveloppement de cette th eorie, la terminolgie a vari e et variera encore. Mais il nest peut- etre pas inutile den proposer une, adapt ee ` a la situation pr esente. M. M. Fr echet, Extrait des Comptes rendus du Congr` es des Soci et es savantes en 1924.

0.1

The Integers

The set of natural numbers is the familiar collection N = {1, 2, 3, . . . , n, . . . }. It would be possible to rigorously develop the properties of the natural numbers deriving from the Peano postulates. We choose not to do this here, but we refer the interested reader to [La]. We do wish to take a more formal approach towards another familiar set of numbers, namely the integers. The integers form the collection {0, 1, 1, 2, 2, . . . }, which we study in elementary arithmetic. We denote the integers by the symbol Z (from the German word Zahlen ). The operations in the integers are addition (+) and multiplication (), and here are the rules. We expect that the reader is well versed in the arithmetic of the integers, but we are stating these properties explicitly for two reasons. First, these properties are used in arithmetic from the earliest grades, but are seldom justied. Second, these properties will be used to describe other algebraic structures that we will meet later. Rules of Arithmetic in Z 0.1.1 (A1) (M1) (A2) (M2) (A3) (M3) If If If If If If a, b Z, then a + b Z. a, b Z, then a b Z. a, b, c Z, then a + (b + c) = (a + b) + c. a, b, c Z, then a (b c) = (a b) c. a, b Z, then a + b = b + a. a, b Z, then a b = b a. 3 Closure Associativity Commutativity

(A4) (M4) (A5)

0 Z a Z, a + 0 = 0 + a = a. 1 Z a Z, a 1 = 1 a = a. a Z, a Z a + (a) = (a) + a = 0.

Identities Additive inverses

In general, elements in Z do not have multiplicative inverses in Z. That is, given an element a Z, we cannot necessarily nd another element b Z such that ab = 1. However, some integers do have multiplicative inverses, namely 1 and 1. The operations of addition and multiplication are tied together by the distributive law. (D) If a, b, c Z, then a (b + c) = (a b) + (a c). Without the distributive law, there would be no connection between addition and multiplication. The richness of the structure is embodied in the interaction between the two operations. Lets stop and investigate some of the implications of these 10 axioms. Facts 0.1.2 1. Additive identities are unique. Proof. Suppose that 0 and 0 are additive identities. Then 0 = 0 + 0 = 0 . 2. Multiplicative identities are unique. Proof. Exercise. (Hint: Use the same technique as above.) 3. Additive inverses are unique. Proof. Suppose that a Z and a + a = 0. Then a + (a + a ) = a + 0 = a. On the other hand, by associativity ((A2)), we have a + (a + a ) = ((a) + a) + a = 0 + a = a . Thus, a = a. 4. (Cancellation for addition) If a, b, c Z and a + b = a + c, then b = c. Proof. If a + b = a + c, then a + (a + b) = a + (a + c). By associativity ((A2)), ((a) + a) + b = ((a) + a) + c, and hence 0 + b = 0 + c, from which we conclude that b = c. Proof. We can write

5. If a Z, then a 0 = 0. a0 = a (0 + 0)

(a 0) + 0

= a0+a0

This is really quite something, and it emphasizes the role of the distributive law. What we have here is multiplication by the additive identity reproducing the additive identity. We have more interaction between multiplication and addition in the following statements. 6. If a Z, then (1) a = a.

by properties of the additive identity and the distributive law. Now cancel to get a 0 = 0.

Notice that this really says something. That is, the left-hand expression, (1) a, represents the additive inverse of the multiplicative identity multiplied by a. The right-hand side, a, on the other hand, represents the additive inverse of a.

Proof. We can write a + (1) a = 1 a + (1) a = (1 + (1)) a = 0 a = 0. But additive inverses are unique, so a = (1) a.

Notice that, when convenient, we drop the dot which signies multiplication. 4

Exercise 0.1.3 If a, b Z, then (a)b = a(b) = (ab). Exercise 0.1.4 If a, b Z, then (a)(b) = ab. Now, what other properties do the integers have? In the integers, cancellation for multiplication doesnt follow from the rst 10 axioms. Cancellation for multiplication should be familiar; many texts introduce it as an additional axiom for the integers in the following form. (C) If a, b, c Z with a = 0 and ab = ac, then b = c. Exercise 0.1.5 Why is a = 0 excluded? However, we will see shortly that because the integers are also ordered, cancellation in the integers is a consequence of the order properties. Exercise 0.1.6 Cancellation can be phrased in another way. Show that the statement if a, b Z and ab = 0, then either a = 0 or b = 0 is equivalent to cancellation. What else do we have for the integers? We have inequalities. The < sign should be familiar to you. It is subject to the following rules of order. (O1) If a, b Z, then one and only one of the following holds: a < b, a = b, or b < a. (Trichotomy) (O2) If a, b, c Z with a < b and b < c, then a < c. (Transitivity) (O3) If a, b, c Z and a < b, then a + c < b + c. (Addition) (O4) If a, b, c Z, a < b, and 0 < c, then ac < bc. (Multiplication by positive elements) We adopt the usual notation and terminology. That is, if a < b, we say that a is less than b. If a < b or a = b, we say that a is less than or equal to b and write a b. If a < b we may also write b > a and say that b is greater than a. The statement b a is now self-explanatory. Here are some examples of recreational exercises and facts which go with the order axioms. For these statements and the following exercises, let a, b, c Z. Facts 0.1.7 1. a > 0 i a < 0.

Proof. Suppose a > 0. Add a to both sides. Proof. Suppose a > 0. Then, since b > 0, ab > 0 b = 0.

2. If a > 0 and b > 0, then ab > 0.

3. If a > 0 and b < 0, then ab < 0. Proof. Suppose a > 0 and b < 0. Then b > 0 and a(b) = (ab) > 0. So ab < 0. 4. If a < 0 and b < 0, then ab > 0. Proof. If a < 0 and b < 0, then a > 0 and b > 0. Hence (a)(b) = ab > 0. 5. If a = 0, then a2 > 0. Proof. If a is greater then 0, use Fact 2. If a is less then 0, use Fact 4. 6. 1 > 0. Proof. 1 = 12 . 5

7. If a > b and c < 0, then ac < bc. Proof. If a > b, then a b > 0. Since c > 0, (c)(a b) = ac + bc > 0. Hence, bc > ac. 8. If a > b, then a < b. Proof. Let c = 1. Are you having fun yet? Good, try these exercises. Exercise 0.1.8 Suppose that 0 < a and 0 < b. Show that a < b i a2 < b2 . Exercise 0.1.9 Suppose that a < 0 and b < 0. Show that a < b i b2 < a2 . Exercise 0.1.10 Show that 2ab a2 + b2 . The set N of natural numbers is the set of positive elements in Z, that is, the set of elements which are greater than 0. It is clear that N is closed under addition and multiplication. If we add trichotomy, these properties lead to an alternate characterization of order. Exercise 0.1.11 Suppose now that we have only the rst 10 axioms for Z as well as the cancellation property (C). Let P be a set of integers with the following properties. 1. If a Z, then one and only one of the following holds: a P , a = 0, or a P . 2. If a, b P , then a + b P and ab P . For a, b Z, dene a < b if b a P . Show that this relation satises (O1)(O4). Moreover, if we have a relation that satises (O1)(O4), and we dene P = {a Z | a > 0}, then show that P satises properties 1 and 2 above. Exercise 0.1.12 Show that the cancellation property (C) can be proved using the axioms for addition and multiplication and the order axioms. So far, the integers have ve axioms for addition, four for multiplication, one for the distributive law, and four for order. There is one more axiom which plays a crucial role. It is called the Well-Ordering Principle. This Principle assures us that 1 is the smallest positive integer. This should not come as a surprise but we do need something to conrm this. In the rational numbers, which we construct in the next section, the rst fourteen axioms are satised, but there is actually no smallest positive element. Thus, we need to introduce the Well-Ordering Principle as an axiom for Z. ( 0.1.13 WO) Well-Ordering Principle for Z If A is a nonempty subset of the positive integers, then A has a least element. That is, there exists an element a0 A, such that for all a A, a0 a. That does it! We now have the 15 properties, and they completely characterize the integers. (For a proof of this, see Project 2 in this chapter.) Most of the work with the Well-Ordering Principle will be done later. However, here are a couple of facts which follow immediately from the Well-Ordering Principle. Facts 0.1.14 1. There are no integers between 0 and 1. Proof. Let A = {a Z | 0 < a < 1}. If A = , then it has a least element a0 which is in A. So, 2 0 < a0 < 1, and, by property (O4), 0 < a2 0 < a0 . But then a0 A and a0 is not the least element. 6

2. (Mathematical Induction) Let A be a set of positive integers such that 1 A, and if k A, then k + 1 A. Then A is the set of all positive integers.

c 1 < c, it follows that c 1 A. And, so, (c 1) + 1 = c is also in A, which is a contradiction. Exercise 0.1.15 If n and k are non-negative integers with n k , we dene the binomial coecient n k = n! k !(n k )!

Proof. Suppose there exists a positive integer which is not in A, and let A be the set of all such positive integers. Then A is a nonempty subset of the positive integers, and hence has a least element c. Now c > 1 since 1 A, and there is no integer between 0 and 1. So c 1 is an integer greater than 0. Since

n k

by

where n! = n(n 1) 2 1, and we set 0! = 1 (this will be explained later in the book when we discuss the Gamma function). Prove the Binomial Theorem : If a, b Z and n is a positive integer, then
n

(a + b)n =
k=0

n k nk a b . k

(Use Mathematical Induction.) Remark 0.1.16 Observe that the binomial coecient n k represents the number of ways of choosing k objects from n objects where order does not matter. The binomial coecient n k is the number of subsets n n of k elements in a set with n elements. Of course the binomial theorem implies that n k=0 k = 2 , the total number of subsets of a set with n elements. Exercise 0.1.17 i. Prove by induction that if A and B are nite sets, A with n elements and B with m elements, then A B has nm elements. ii. Prove by induction the corresponding result for a collection of k nite sets, where k > 2.

0.2

Equivalence Relations and the Construction of Q

Next we turn to the idea of a relation on a set. Here is the formal denition of a relation. Denition 0.2.1 A relation on a set X is a subset R of X X . For example, we can dene a relation on Z by setting R equal to {(a, b)|a, b Z and a < b}. Equivalence relations are a special type of relation, which we dene below. They appear everywhere in mathematics, and we really mean that. What an equivalence relation does is take a set and partition it into subsets. Some equivalence relations appear to be very natural, some appear to be supernatural, and others appear to make no sense at all. Denition 0.2.2 Let X be a set. An equivalence relation on X is a relation R on X such that

(ER1) For all a X , (a, a) R. (Reexive) (ER2) For a, b X , if (a, b) R, then (b, a) R. (Symmetric) (ER3) For a, b, c X , if (a, b), (b, c) R, then (a, c) R. (Transitive) The twiddle notation () is often used in mathematics. Here we use it as follows: if (a, b) R, we write a b. Then the denition of equivalence relation becomes 7

(ER1) For all a X , a a. (Reexive) (ER2) For a, b X , if a b then b a. (Symmetric) (ER3) For a, b, c X , if a b and b c, then a c. (Transitive) Again, speaking loosely, we can refer to as an equivalence relation on X . Exercise 0.2.3 Let R be a relation on X that satises the following two conditions. a. For all a X , (a, a) R. b. For a, b, c X if (a, b), (b, c) R, then (c, a) R. Show that R is an equivalence relation. Example 0.2.4 The most basic example of an equivalence relation is equality. That is, a b i a = b. Prove this, but please dont write anything. Example 0.2.5 If A and B are triangles in the plane, write A B if and only if A is similar to B . Example 0.2.6 Let n be an integer greater than or equal to 2. If a, b Z, we say that a b i a b is a multiple of n, that is, n divides a b. This last example requires a little more elucidation. So, we present a brief discussion about divisibility in Z. Denition 0.2.7 Suppose that a and b are integers. We say that a divides b, written a|b, if there is an element c Z such that b = ac. When a divides b, the number a is called a divisor of b. We need the following facts about divisibility. Facts 0.2.8 1. If a Z, then a|a. 2. If a|b then a| b. 3. If a|b and b|c, then a|c. These facts are easy to prove. For example, if a|b and b|c, there are integers h and k such that b = ha and c = kb. But then c = (hk )a, and since hk is an integer by axiom (M1), that does it. Exercise 0.2.9 Show that, if a Z, then a|0. Exercise 0.2.10 Show that, if a and b are integers such that a|b and b|a, then a = b. Exercise 0.2.11 Show that, if c|a and c|b, and s, t Z, then c|(sa + tb). There is one other type of integer that should be familiar to the reader. Denition 0.2.12 Let p be a positive integer greater than or equal to 2. We say that p is prime if the only positive divisors of p are 1 and p. If n is a positive integer greater than 2 which is not prime, then n is called composite . So, if n is composite there exist integers a and b both greater than or equal to 2, such that n = ab. Exercise 0.2.13 Let n be a positive integer greater than or equal to 2. Then there exists a prime p such that p divides n. 8

With this discussion of divisibility under our belt, we dene the notion of congruence in the integers. Let n be an integer greater than or equal to 2. For a, b Z, we say that a is congruent to b modulo n, a b (mod n), provided that n | a b. Exercise 0.2.14 For a xed integer n 2, we dene a b if and only if a b (mod n). Show that this is an equivalence relation on Z. Now we return to equivalence relations in general. The partitioning into subsets relative to an equivalence relation comes about as follows. If a X , we write C (a) = {b X | b a}. C (a) is called the class of a or the equivalence class containing a. Here are the properties of equivalence classes.

Theorem 0.2.15 (Properties of equivalence classes) 1. a C (a).

Proof. Reexivity.

2. If a b, then C (a) = C (b). Proof. Transitivity. 3. If a is not equivalent b (a b), then C (a) C (b) = . 4. C (a) = X.

Proof. If c C (a) C (b), then c a and c b, so a b. So C (a) C (b) = i C (a) = C (b).


a X

Proof. Use (1) above.

This all means that an equivalence relation on a set X partitions X into a collection of pairwise disjoint subsets. Although this looks quite special, its really not that impressive. For example, take a set X and break it up into pairwise disjoint nonempty subsets whose union is all of X . Then, for a, b X , dene a b if a and b are in the same subset. Exercise 0.2.16 Prove that this is an equivalence relation on X . One more example of an equivalence relation will prove useful for future developments. This is a method for constructing the rational numbers Q from the integers Z using the properties discussed in the last section. We consider the set F = {(a, b) | a, b Z and b = 0}. We are thinking (for example) of the pair (2, 3) as the fraction 2 3 . For (a, b), (c, d) F , we dene (a, b) (c, d) if ad = bc. Thus, for instance, (2, 3) (8, 12) (6, 9). Exercise 0.2.17 Show that is an equivalence relation on F . The set of equivalence classes determined by this equivalence relation is called the rational numbers and is denoted by Q. You should be extremely happy about this since it explains all that business about equivalent fractions that you encountered in elementary school. What a relief! We have several things to do with this example. First, we have to add and multiply rational numbers, that is, add and multiply equivalence classes. The fundamental principle to be established here is that, when we add or multiply equivalence classes, we do it by selecting an element from each equivalence class and adding or multiplying these. We must be certain that the result is independent of the representatives that we choose in the equivalence classes. For simplicity, we denote the class of (a, b) by {(a, b)} rather than C ((a, b)). 9

F or {(a, b)}, {(c, d)} Q, we dene addition and multiplication as follows. {(a, b)} + {(c, d)} = {(ad + bc, bd)}, and {(a, b)} {(c, d)} = {(ac, bd)}.

What we must establish is the fact that if (a, b) (a , b ) and (c, d) (c , d ), then (ad + bc, bd) (a d + b c , b d ) and (ac, bd) (a c , b d ). All this requires is a little elementary algebra, but, for your sake, well actually do one and you can do the other. Of course, we do the easier of the two and leave the more complicated one for you. So, here goes: (a, b) (a , b ) means that ab = a b, and (c, d) (c , d ) means that cd = c d. Multiplying the rst equality by cd , and then substituting cd = c d on the right hand side of the resulting equation, we get the desired equality acb d = a c bd.

Exercise 0.2.18 You do addition. Its messy. When we are dening some operation which combines equivalence classes, we often do this by choosing representatives from each class and then showing that it doesnt make any dierence which representatives are chosen. We have a formal name for this. We say that the operation under consideration is well-dened if the result is independent of the representatives chosen in the equivalence classes. Throughout this book, we will encounter equivalence relations on a regular basis. You will be fortunate enough to have the opportunity to prove that these are actually equivalence relations. What properties are satised by addition and multiplication as dened above? For example, what about the associativity of addition? We must prove that ({(a, b)}+{(c, d)})+{(e, f )} = {(a, b)}+({(c, d)}+{(e, f )}). Well, ({(a, b)} + {(c, d)}) + {(e, f )} = = {(ad + bc, bd)} + {(e, f )} {((ad + bc)f + (bd)e, (bd)f )}.

Now we use associativity and distributivity in Z to rearrange things in an appropriate fashion. This gives {(((ad)f +(bc)f )+(bd)e, (bd)f )}, and using the acrobatics of parentheses, we get {(a(df )+b(cf +de), b(df ))} = {(a, b)} + ({(c, d)} + {(e, f )}). This is all rather simple, that is, to prove various properties of addition and multiplication in Q, we reduce them to known properties from Z. Exercise 0.2.19 i. Prove the associative law for multiplication in Q. ii. Prove the commutative laws for addition and multiplication in Q. iii. Show that {(0, 1)} is an additive identity in Q. iv. Show that {(1, 1)} is a multiplicative identity in Q. v. Show that {(a, b)} is an additive inverse for {(a, b)}. vi. Prove the distributive law for Q. Notice here that if {(a, b)} = {(0, 1)}, that is, a = 0, then {(a, b)} {(b, a)} = {(1, 1)}. Thus, in Q, we have multiplicative inverses for nonzero elements. Lets tidy this up a bit. First of all, we have no intention of going around writing rational numbers as equivalence classes of ordered pairs of integers. So lets decide once and for all to write the rational number {(a, b)} as a b . Most of the time this fraction will be reduced to lowest terms, but, if it is not reduced to lowest terms, it will certainly be in the same equivalence class as a fraction which is reduced to lowest terms. Note that it is always possible to choose a fraction from the equivalence class which is in lowest terms because of the well-ordering principle in the integers, applied to the numerators. With this, addition and multiplication of rational numbers have their usual denition: 10

a + b a b

c d c d

= =

ad + bc , bd ac . bd

Now consider the axioms for the integers (A1)(A5), (M1)(M4), and (D). All of these hold for the rational numbers, and there is an additional multiplicative property, multiplicative inverses. (M5) If a = 0, then there is an element a1 such that aa1 = a1 a = 1. Remark 0.2.20 Note that the natural numbers N may be regarded as a subset of Z, and in turn the integers Z may be regarded as a subset of Q by identifying the integer n with the equivalence class n 1. The operations of addition and multiplication are sometimes called binary operations or internal laws of composition Denition 0.2.21 Let R be a non-empty set. An internal law of composition (ILC) on R is a map : R R R. If a, b R then we usually write ((a, b)) = a b Of course the more properties that are satised by internal laws of composition, the better life gets. Denition 0.2.22 A set with two internal laws of composition, + and , that satisfy (A1)(A5), (M1) (M4), and (D) is called a commutative ring with 1. If, in addition, cancellation (C) holds for multiplication, the commutative ring with 1 is called an integral domain. If (M5) also holds, the structure is called a eld . Note that the word commutative in this denition refers not to the commutativity of addition but to the commutativity of multiplication. Thus, in our latest terminology, Z is a integral domain and Q is a eld. What about cancellation for multiplication? This followed from order in Z, but for Q (or any eld for that matter) cancellation for multiplication holds automatically. Exercise 0.2.23 Prove that the cancellation law (C) holds in any eld. Exercise 0.2.24 Let X be a nonempty set and R = (X ), the power set of X . Show that R with symmetric dierence as addition and intersection as multiplication is a commutative ring with 1. When is R a eld? (See Appendix A for the relevant set-theoretic denitions.) There is another denition which will prove useful in our discussions about these various algebraic structures. Denition 0.2.25 Suppose that R is a commutative ring with 1. A subset R0 of R is a subring if R0 is a ring itself with the same operations of addition and multiplication as in R. We dont necessarily require that R0 have a multiplicative identity and in this case we call R0 simply a commutative ring. The same idea can be used to dene subintegral domain . Finally, if F is a eld and F0 is a subset of F , we say that F0 is a subeld if it is a eld with the same operations of addition and multiplication as in F . Exercise 0.2.26 i. Let R be a ring and R0 a non empty subset of R, show that R0 is a subring if and only if for any a, b R0 we have a b and ab in R0 . ii. If F is a eld and F0 is non-empty subset of F , are the properties in (i) enough to ensure that F0 is a subeld? 11

Just as we consider two sets to be the same when there exists a bijection between them, there is a notion of sameness for other mathematical structures dened by the existence of certain maps, called isomorphisms. Denition 0.2.27 Let R and R be two commuative rings with 1. We say that R and R are isomorphic if there exists a bijection : R R such that a. (x + y ) = (x) + (y ) for all x, y R; b. (xy ) = (x)(y ) for all x, y R. We call the map an isomorphism. What about order in Q? It is simple to extend the order from Z to Q. We do this using the notion of a set of positive elements. We say that a b Q is positive if ab > 0 in Z. Exercise 0.2.28 Show that the above notion of positivity in Q satises the properties in Exercise 0.1.11, or equivalently, the properties of order given in (O1)(O4). Denition 0.2.29 An integral domain or eld in which there is an order relation satisfying (O1)(O4) is called an ordered integral domain or ordered eld, respectively. See Project ?? for more about this. Denition 0.2.30 Suppose that R and R are ordered integral domains. We say that R and R are order isomorphic if there exists an isomorphism : R R such that if x, y R and x < y then (x) < (y ) in R . We call the map an order isomorphism. So what is this all about? We have rules for the integers, and the same rules, along with (M5), are satised by the rational numbers. Actually, there are lots of structures other than the integers and the rational numbers which have operations of addition, multiplication, and, sometimes, an order relation. For example, the real numbers R, the complex numbers C, the algebraic numbers A, the collection of n n matrices Mn (R), all satisfy some or all of these properties. We want to give two more examples before we leave this section. First, let n be a positive integer greater than or equal to 2 and consider the equivalence relation given in Example 0.2.6. What are the equivalence classes? For example, take n = 5. Then we have 5 classes. They are C (0) = 0 C (1) = 1 C (2) = 2 C (3) = 3 C (4) = 4 = {0, 5, 5, 10, 10, . . . } = {1, 6, 4, 11, 9, . . . } = {2, 7, 3, 12, 8, . . . }

= {3, 8, 2, 13, 7, . . . } = {4, 9, 1, 14, 6, . . . }.

Note that, in this example, we have simplied the notation of equivalence class by writing the equivalence class C (a) by a. Observe that 5 = 0, 6 = 1, etc. In general, for an arbitrary n, we will have n classes 0, 1, . . . , n 1. These are called the equivalence classes modulo n, or, for short, mod n. Moreover, for any integer a, we denote the equivalence class in which a lies by a. Of course, it is always true that a is equal to one of the classes 0, 1, . . . , n 1. Lets dene addition and multiplication mod n. Denition 0.2.31 Denote the set of equivalence classes 0, 1, . . . , n 1 by Zn . For a, b Zn , dene a + b = a + b and ab = ab. Exercise 0.2.32 i. Show that addition and multiplication in Zn are well-dened. ii. Show that, with these operations, Zn is a commutative ring with 1. 12

iii. Show that Zn cannot satisfy the order axioms no matter how > is dened. iv. Show that Z2 is a eld but Z4 is not. v. For p prime show that Zp is a eld. The second example is the real numbers denoted by R. A construction and complete discussion of the real numbers is given in the next chapter. We will see that the real numbers are an ordered eld which contains Q and has one additional property called the least upper bound property.

0.3

Countability

Our discussion of number systems would not be complete without mentioning innite sets. Indeed, most of the sets we deal with in analysis are innite. Moreover, any discussion of continuity and change must involve innite sets. Thus we are motivated to begin a formal discussion of what it means for a set to be innite. (See Appendix A for a consideration of the more elementary notions of set theory.) Denition 0.3.1 A set A is nite if A is empty or there exists n N such that there is a bijection f : A {1, 2, . . . , n}, where {1, 2, . . . , n} is the set of all natural numbers less than or equal to n. In this case, we say A has n elements. Exercise 0.3.2 If A is a nite set and B is a subset of A, show that B is a nite set. In addition show that if B is a proper subset then the number of elements in B is less then the number of elements in A. There is a natural and useful property of nite sets, which in fact will turn out to be a characterization of them: Theorem 0.3.3 If A is a nite set and B is a proper subset of A, then there is no bijection between B and A. Proof. Suppose A has n elements and B has m elements with m < n. Then the Pigeonhole Principle (see Appendix A.3) tells us that, for any function from A to B , there is an element of B which is the image of two dierent elements of A. Exercise 0.3.4 Show that the following are nite sets: i. The English alphabet. ii. The set of all possible twelve letter words made up of letters from the English alphabet. iii. The set of all subsets of a given nite set. This approach to things makes the denition of innite sets quite simple: Denition 0.3.5 An innite set is a set that is not nite. One of the most important characteristics of a set is its cardinality, which formalizes the notion of the size of the set. A thorough treatment of cardinality would take us too far aeld, but we can say what it means for two sets to have the same cardinality. Denition 0.3.6 The cardinal number of a nite set A is the number of elements in A, that is, the cardinal number of A is the natural number n if there is a bijection between A and {k N | 1 k n}. Denition 0.3.7 A set A has cardinality 0 (pronounced aleph null or aleph naught) if it can be put in one-to-one correspondence with N, that is, there is a bijection between the set and N. 13

In general, two sets have the same cardinality if they can be put in one-to-one correspondence with each other. Example 0.3.8 The set N has cardinality 0 (this should not come as a surprise). We will see later in this section that there are innite sets with cardinality other than 0 . Example 0.3.9 The set N {0} has cardinality 0 because the function f : N {0} N given by f (n) = n + 1 is a bijection. Example 0.3.10 The set Z has cardinality 0 because the function f : Z N given by f (z ) = is a bijection. There is a very useful theorem which asserts the existence of a one-to-one correspondence between two sets. This relieves us of the burden of constructing a bijection between two sets to show that they have the same cardinality. Theorem 0.3.11 (Schr oder-Bernstein) If A and B are sets, and there exist injections f : A B and g : B A, then there exists a bijection between A and B . Proof. First, we divide A into three disjoint subsets. For each x A, consider the list of elements Sx = {x, g 1 (x), f 1 g 1 (x), g 1 f 1 g 1 (x), . . . }. The elements of this sequence are called predecessors of x. Notice that in Sx , we start with x A. Then g 1 (x) B if g 1 (x) exists (x may not be in the image of g ). For each x A, exactly one of the three following possibilities occurs. 1. The list Sx is innite. 2. The last term in the list is an element of A. That is, the last term is of the form y = f 1 g 1 g 1 (x), and g 1 (y ) does not exist (i.e. y is not in the image of g ). In this case, we say that Sx stops in A. 3. The last term in the list is an element of B . That is, the last term is of the form z = g 1 f 1 g 1 (x) and f 1 (z ) does not exist (i.e. z is not in the image of f ). In this case, we say that Sx stops in B . Let the corresponding subsets of A be denoted by A1 , A2 , A3 . Similarly, dene the corresponding subsets of B . That is B1 B2 B3 = = = {y B | y has innitely many predecessors }, 2z + 2 2z 1 if z 0 if z < 0

{y B | the predecessors of y stop in A}, and {y B | the predecessors of y stop in B }.

bijections.

Now observe that f : A1 B1 , g : B1 A1 are both bijections. Also, g : B2 A2 and f : A3 B3 are

Exercise 0.3.12 Suppose A, B , and C are subsets of a set X such that A B C . Show that if A and C have the same cardinality, then A and B have the same cardinality. 14

Example 0.3.13 Q+ has cardinality 0 (recall that Q+ denotes the positive rational numbers). Here are three proofs: 1. This is a very common and very sloppy proof. However the underlying idea will stand us in good stead.
1 1 2 1

1 2

e 

3 2

G e

3 1

e e

4 1

5 2

7 2

G e e

1 3

1 4

e 

2 3

4 3

5 3

4 4 e 4 1 2 3 4 5 5 f 5 f 5  . . . . . . . . . . . .

6 G d 1  11 9 2 2 c 8 7 3 d 3  9 11 4 4 c 6 7 5 5 e . . . . . .

5 1

To nd a bijection between N and Q+ , we write all the positive fractions in a grid, with all fractions with denominator 1 in the rst row, all fractions with denominator 2 in the second row, all fractions with denominator 3 in the third row, etc. Now go through row by row and throw out all the fractions that arent written in lowest terms. Then, starting at the upper left hand corner, trace a path through all the remaining numbers as above.
1 We can count along the path we drew, assigning a natural number to each fraction. So 1 1 1, 2 2, 2 3 3 3, 4, 5, etc. This is a bijection. Therefore, Q is countable. Although this is a very + 1 1 2 common proof, the bijection is not at all obvious. It is very dicult to see, for example, which rational number corresponds to 1, 000, 000.

2. In this proof, well make use of the Schr oder-Bernstein Theorem. It is easy to inject N into Q+ ; simply send n to n. For the injection from Q+ to N, we consider Q+ to be the set of positive rational numbers written as fractions in base 10, and we let N be the natural numbers but written in base 11 with the numerals being 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, d. So, for example, the integer which is written 21 in base 10 is 1d in base 11, and 1222(10) = d11(11) . (Incidentally, we are writing the bases in base 10.) Now dene a an ...a2 a1 th function f : Q+ N by writing a digits of the numerator and b = bm ...b2 b1 , where ai and bi are the i a denominator (and, of course, integers between 0 and 9). Then, set f ( b ) = an . . . a2 a1 dbm . . . b2 b1 . The fraction a b will always be written in lowest terms. For instance, if we take the fraction 2/3 in base 10, then f (2/3) = 2d3 in base 11 which is the same as the integer 355 written in base 10. Each number which is the image of a fraction has one and only one d in it, so it is easy to see which fraction is represented by a given integer. According to Schr oder-Bernstein, two injections make a bijection, so Q+ is countable. 3. Write each positive fraction in lowest terms and factor the numerator and denominator into primes, n p1 1 p2 2 p 1 n so that p m , with pi = qj . If by chance p or q is 1, and cant be factored, write it as 1 . q = 1 2
q1 q2 qm

15

Then let f : Q+ N be dened by f


n 1 2 p 1 p2 pn 1 2 m q1 q2 qm 1 22 2m 1 n 21 1 22 1 = p2 q2 qm . p2 n q1 1 p2

In particular, note that if a Q+ is an integer, then f (a) = a2 .

Exercise 0.3.14 For this exercise, consider the function f in the third proof above. 1. Verify that f is a bijection. 2. Suppose that N = 10k for some integer k . Find
p q

Q such that f

p q

= N.

Exercise 0.3.15 Use any one of the above three proofs to show that Q has cardinality 0 . Exercise 0.3.16 Show that the natural numbers are an innite set. Exercise 0.3.17 Show that any set that has the same cardinal number as N is an innite set. Note: A set is called countable, or sometimes denumerable , if it has cardinality 0 (that is, if it is in one-to-one correspondence with the natural numbers). The term countable is used in several ways. Many people use it to refer to innite sets which are in one-to-one correspondence with N, while others include nite sets when they say countable. This is not something to get disturbed about. Usually, when we refer to a countable set, we mean countably innite (cardinality 0 ) . When we refer to a nite set, we will generally say A is a nite set. Exercise 0.3.18 Show that a subset of a countable set is countable or nite. Exercise 0.3.19 Show that the set of all polynomial functions with integer coecients is a countable set. Theorem 0.3.20 If A is an innite set, then A has a countable subset. Proof. Take any innite set A and choose an element a1 in A. Let A1 = A \ {a1 }. By the denition of innite set, A1 is innite. So we choose a2 in A1 and dene A2 = A \ {a1 , a2 }. Since A is not nite, we can continue to choose elements. Thus, if we have chosen a1 , . . . , an , we consider An = A \ {a1 , . . . , an }. Since A is innite, we can choose an element an+1 in An . Continuing inductively, we obtain our desired countable subset. Note that this countable set may be all of A.

Remark 0.3.21 There is some discussion among mathematicians as to whether the preceding proof involves the Axiom of Choice. The Axiom of Choice in its fullest form will be discussed below. However, one can make the argument that it requires some sort of choice mechanism to pick an element from a non-empty set. The technique that we use in the proof of Theorem 0.3.20 is sometimes referred to as the countable Axiom of Choice. We could pursue an alternate denition of an innite set. In fact, we could dene innite sets rst and then say that a nite set is a set that is not innite. We use Theorem 0.3.20 as motivation for the following. Redenition 0.3.22 A set is innite if there is a bijection between the set and one of its proper subsets. Redenition 0.3.23 A nite set is a set that is not innite. 16

To show the equivalence of the two denitions, recall that in Theorem 0.3.3 we showed there is no bijection between a nite set and any of its proper subsets. This means that if a set is innite by our new denition, it is not nite (hence, innite) by the old denition too. Next, lets show that any set that is innite by the old denition is bijective with one of its proper subsets. Proof. Say A is an innite set and B A is countable. Then we can write B = {b1 , b2 , . . . , bn , . . .}. Now dene f : A A \ {b1 } as follows: for a A \ B , f (a) = a, and for bi B , f (bi ) = bi+1 . Thus f is a bijection between A and A \ {b1 }. Therefore, our denitions are equivalent. We now turn to operations involving innite sets. Facts 0.3.24 1. If A1 and A2 are countable sets, then A1 A2 is a countable set. 2. If A1 , A2 , . . ., An are countable sets, then n j =1 Aj is a countable set. 3. Let {Aj }j N be a countable collection of countable sets. Then j N Aj is a countable set. Proof. We prove 3 only. You can prove the other two (or deduce them from 3). Write Aj = {aj,1 , aj,2 , . . . , aj,n , . . .}. We use the diagonal process, as in Example 0.3.13. Simply write A1 : a1,1 , a1,2 , . . . , a1,n , . . . A2 : a2,1 , a2,2 , . . . , a2,n , . . . . . . Am : am,1 , am,2 , . . . , am,n , . . . . . . Now count diagonally, ignoring repetitions. Now lets take a look at Cartesian products. It is clear from the ideas presented above that if A1 and A2 are countable, then A1 A2 is countable. Exercise 0.3.25 i. Show that if A1 , A2 , . . ., An are countable, then A1 A2 An is countable. ii. What can you say about the countable Cartesian product of countable sets? Next we look at the power set (A) for any set A. Theorem 0.3.26 If A is any set (including the empty set), there is no bijection between A and (A). Proof. This is clear if A is the empty set. Suppose that there is a bijection between A and (A). If a A, let Pa be the subset of A associated with it. Now consider the set B = {a|a Pa }. The set B must be associated to some element of A, which we creatively call b, so that B = Pb . Is b in B ? For b to be in B , we must have that b Pb . But B = Pb , so therefore b is not in B . But then b Pb , which means that b is in B . This is a contradiction. Therefore, there is no bijection between A and (A).

Denition 0.3.27 If A is a countable set, then the cardinality of (A) is denoted by c. 17

Exercise 0.3.28 Show that the denition of the cardinal number c does not depend on the choice of the countable set A. That is if A and B are countable sets then there is a bijection between (A) and (B ). Remark 0.3.29 At this point, we observe that if A is a countable set, A = {a1 , a2 , . . . , an , . . .}, then (A) is in one-to-one correspondence with the set of all functions from A to the set {0, 1}. This correspondence is dened as follows: If B is a subset of A, then we dene the map fB : A {0, 1} by fB (aj ) = 1 if aj is in B , 0 if aj is not in B . Observe that fB can be viewed as a binary expansion of a real number between 0 and 1. Exercise 0.3.30 Suppose that A is a nonempty set. Show that (A) is in one-to-one correspondence with the set of all functions from A to {0, 1}. Remark 0.3.31 Based on the reasoning in the previous exercise, if A is a nite set with n elements, the cardinality of (A) is 2n . We extend this reasoning to countable sets to write c = 20 . One of the most important sets of numbers that we deal with in this book is the collection of real numbers R. In the next chapter, we will go through the formal construction of the real numbers from the rational numbers. For the present discussion, we can just consider the set of real numbers to be the set of all innite decimals with the convention that no decimal expansion can end in repeating 9s. There are two things to show about the reals. The rst is the proof due to Cantor that the reals are uncountable, and the second is that the cardinality of the real numbers is in fact c. Theorem 0.3.32 The set of all real numbers between 0 and 1 is not countable. Proof. We rst note that the decimal expansion is unique with the exception of those that end in all nines. In this case, we always round up the digit which occurs before the sequence of nines. To prove that this set is not countable, we assume that it is, and list the real numbers between 0 and 1 vertically. a1 = 0.a1,1 a1,2 . . . a1,n . . . a2 = 0.a2,1 a2,2 . . . a2,n . . . . . . am = 0.am,1 am,2 . . . am,n . . . . . . We now proceed using a process similar to the one used in the proof of Theorem 0.3.26 to produce a real number between 0 and 1 which is not on our list. We construct a number b = 0.b1 b2 . . . bn . . . by proceeding diagonally down the list as follows: if a1,1 = 1, take b1 = 2. If a1,1 = 1, take b1 = 1. Next, if a2,2 = 1, take b2 = 2. If a2,2 = 1, take b2 = 1. Continuing this process, we see that the decimal b = 0.b1 b2 . . . bn . . . cannot be on our list, since it diers from each number we list in at least one digit. Consequently, the real numbers between 0 and 1 are not countable. Theorem 0.3.33 The cardinality of the real numbers between 0 and 1 is c = 20 . Proof. To write down an exact bijection between (N) and the real numbers between 0 and 1 requires some care. The standard way to do this is to write all real numbers between 0 and 1 in their binary expansion in such a way that no expansion terminates in all ones. In considering the corresponding subsets of N, we rst remove two specic subsets of (N). We remove the two collections Af = {C (N) | C is nite} and Acf = {D (N) | c D is nite}. The collection (N) \ (Af Acf ) is in one-to-one correspondence with all binary expansions which have an innite number of ones but do not terminate in all ones. We get the required bijection by Remark 0.3.29. We can place Af into one-to-one correspondence with the set of all nite binary expansions with 0 in the rst place, and Acf can be put into one-to-one correspondence with the set of all nite binary expansions with 1 in the rst place.

18

Exercise 0.3.34 Write down these last two bijections explicitly. Exercise 0.3.35 i. Prove that the countable union of sets of cardinality c again has cardinality c. ii. Prove that the set of all real numbers has cardinality c. iii. Prove that the set of irrational numbers in R has cardinality c. How big do cardinal numbers get? For instance, the power set of R is bigger than c. In fact, the power set of R can be identied with the set of all maps from R into {0, 1} just as we did above for the power set of N. Thus, the cardinality of (R) is 2c . The following theorem is interesting, and we include it because it illustrates the kinds of non-intuitive results that one encounters when dealing with innite sets. Theorem 0.3.36 There is a bijection between the unit interval and the unit square. Proof. Let and I = [0, 1][0, 1]. This seems like a great time to use Schr oder-Bernstein. The function f : I I 2 dened 2 by f (x) = (x, 0) is an injection. Dene the function g : I I by the rule g ((a0 .a1 a2 . . . an . . . , b0 .b1 b2 . . . bn . . .)) = (0.a0 b0 a1 b1 a2 b2 . . . an bn . . .), where a0 a1 a2 . . . an . . . and b0 .b1 b2 . . . bn . . . are decimal expansions of the coordinates of any point in I 2 (of course, the decimal expansion is prohibited from ending in all 9s). The function g : I 2 I is an injection. Therefore, there is a bijection between I and I 2 .
2

I = [0, 1] = {x R | 0 x 1}

0.4

Axiom of Choice

Denition 0.4.1 A partially ordered set is a set X with a relation which is reexive, transitive, and anti-symmetric (that means that if a b and b a, then a = b). A totally ordered set is a partially ordered set with the additional property that, for any two elements a, b X , either a b or b a. A well-ordered set is a totally ordered set in which any non-empty subset has a least element. Example 0.4.2 1. (N, ) is a totally ordered set, as are (Z, ), (Q, ) and (R, ). 2. Let X be a set, and let (X ) be the collection of all subsets of X . Then ((X ), ) is a partially ordered set. Denition 0.4.3 Let Y be a subset of a partially ordered set X . An upper bound for Y is an element a X such that y a for all y Y . A least upper bound for Y is an element b X such that b is an upper bound for Y and if a is an upper bound for Y , then b a. The least upper bound is sometimes abbreviated lub, and is also denoted as sup (supremum). You can gure out what a lower bound and greatest lower bound (glb) are. The greatest lower bound is also denoted by inf (inmum). Observe that a subset of a partially ordered set may not have an upper bound or a lower bound. Exercise 0.4.4 If a subset Y of a partially ordered set X has an upper bound, determine whether or not Y must have a least upper bound. If Y has a least upper bound, determine whether or not this least upper bound is unique. 19

Denition 0.4.5 In a partially ordered set, an element b is maximal if a b implies a = b . We turn now to one of the major topics of this chapter, the axiom of choice, and various logically equivalent statements. For many years, there has been considerable discussion among mathematicians about the use of the axiom of choice and the seemingly contradictory results that come along with it. We nd it indispensable in obtaining a number of results in mathematics. The Axiom of Choice 0.4.6 Given a collection C of sets which does not include the empty set, there exists a function : C C C C with the property that A C , (A) A. Another way of looking at this is as follows. Suppose {Ai }iI is a collection of non-empty sets indexed by an index set I . A choice function is then dened as a map : I iI Ai such that (i) Ai . The axiom of choice can then be rephrased. The Axiom of Choice 0.4.7 For every collection of nonempty sets there exists a choice function. The axiom of choice is equivalent to a number of other very useful statements which are not at all obvious. Here they are, in no particular order. Let X be a partially ordered set. The collection (X ) can be partially ordered by inclusion, see 0.4.2. This partial ordering on (X ) is used in some of the statements below. Hausdor Maximality Principle 0.4.8 Every partially ordered set X contains a totally ordered subset that is maximal with respect to the ordering on (X ). Zorns Lemma 0.4.9 If a non-empty partially ordered set has the property that every non-empty totally ordered subset has an upper bound, then the partially ordered set has a maximal element. Well-Ordering Principle 0.4.10 Every set can be well-ordered. The following lemma is slightly complicated, but it will allow us to prove the equivalence of the above statements with little trouble. Lemma 0.4.11 Suppose that (X, ) is a non-empty partially ordered set such that every non-empty totally ordered subset has a least upper bound. If f : X X is such that f (x) x for all x X , then there is some w X such that f (w) = w. Proof. First we reduce to the case when X contains a least element, call it b. In fact, if X is nonempty choose any b X and replace X by X = {x X | x b}. It is clear that X is stable under f (that is f (X ) X ) and has the same properties as X . We call a subset Y of X admissible if 1. b Y 2. f (Y ) Y 3. Every lub of a totally ordered subset of Y belongs to Y . X is certainly admissible, and the intersection of any family of admissible sets is admissible. Let W be the intersection of all admissible sets. The set {x|b x} is admissible, so if y W , then b y . We will now construct a totally ordered subset of W with the property that its least upper bound is a xed point of f . Consider the set P = {x W | if y W and y < x then f (y ) x}. Note that P is non-empty since b P . First we show that any element of P can be compared to any element of W and hence P is totally ordered. Now x an x P and dene Ax = {z W |z x or z f (x)}. We would like to show that Ax is admissible. 20

1. Obviously, b Ax since b x. 2. Suppose z Ax . There are three possibilities. If z < x, f (z ) x by the conditions of P , so f (z ) Ax . If z = x, f (z ) = f (x) f (x) so f (z ) Ax . If z f (x), then f (z ) z f (x) so f (z ) Ax . 3. Finally, let Y be a totally ordered non-empty subset of Ax , and let y0 be the lub of Y in X . Then y0 W , since W is admissible. If z x for all z Y then y0 x and hence y0 Ax . Otherwise z f (x) for some z Y , which implies y0 f (x), so y0 Ax . Thus, Ax is admissible. Since Ax is an admissible subset of W , Ax = W . Put another way, if x P and z W , then either z x or z f (x) x, and thus P is totally ordered. Therefore P has a least upper bound, call it x0 . Again x0 W and f (x0 ) W because W is admissible. We will now show f (x0 ) = x0 . First we claim x0 P . Indeed, if y W and y < x0 , then there exists x P with y < x x0 , whence f (y ) x x0 . Let y W and suppose y < f (x0 ). As we saw above Ax0 = W , so we have y x0 . If y = x0 , then f (y ) = f (x0 ) f (x0 ). If y < x0 , then f (y ) x0 f (x0 ). In either case, we nd f (x0 ) P. Hence f (x0 ) x0 f (x0 ). Whew! Theorem 0.4.12 (1) The Axiom of Choice, (2) Hausdor Maximality Principle, (3) Zorns Lemma, and (4) Well-Ordering Principle are all equivalent. Proof. We will show that (1) implies (2), which implies (3), which implies (4), which implies (1), and then we will be done. (1) (2) Take a non-empty partially ordered set (E, ). Make E , the family of totally ordered subsets of E , into a partially ordered set under inclusion. We wish to show that E has a maximal element (i.e., an element which is not smaller than any other element). So we will assume the opposite and reach a contradiction by applying Lemma 0.4.11. We must rst check to see if the lemma is applicable: Suppose F is a totally ordered subset of E . Then it has a least upper bound, namely F F F . Now, for a given e E , let Se = {x E|e x, e = x}. Then Se can never be the empty set, because that would mean that e is maximal. So we apply the axiom of choice by dening a function f : {Se |e E} E with the property that f (Se ) Se . Now dene g : E E by g (e) = f (Se ). This gives us that e g (e) for all e E , contradicting the lemma. (2) (3) Again, consider a partially ordered set (E, ). Now let x be the upper bound for E0 , a maximal totally ordered subset of E . Suppose that there is some y E such that y > x. Then E0 {y } is a totally ordered set containing E0 , contradicting our assumption of maximality. Exercise 0.4.13 Now you nish the proof. Show that Zorns Lemma implies the Well Ordering Principle, and that the Well Ordering Principle implies the Axiom of Choice.

21

0.5

Independent Projects

0.5.1 Basic Number TheoryThe following statements present a number of facts about elementary number theory. Your goal in this project is to prove them. 1. The division algorithm: if a, b Z and b = 0, then there is a unique pair q, r Z with a = qb + r and 0 r < |b |. 2. If M is a subset of Z which is closed under subtraction and contains a nonzero element, then M = {np | n Z}, where p is the least positive element of M . Denition 0.5.1 Let a, b Z. The greatest common divisor of a and b is the largest positive integer d such that d | a and d | b. We often denote the greatest common divisor of a and b by (a, b). We say that a and b are relatively prime if (a, b) = 1. 3. If a, b Z and d = (a, b), then there exist s, t Z such that d = sa + tb. 4. Euclids lemma: If p is prime and p|ab, then p|a or p|b. 5. If (a, c) = 1, and c|ab, then c|b. 6. If (a, c) = 1, a|m and c|m, then ac|m. 7. If a > 0 then (ab, ac) = a(b, c). 8. The integers Z have unique factorization, that is, if n is an integer greater than or equal to 2, then there exist unique distinct primes p1 , p2 , . . . , pk , with p1 < p2 < < pk , and positive integer exponents k 1 2 1 , 2 , . . . , k such that n = p 1 p2 pk .
k 1 2 9. If n is a positive integer greater than or equal to 2 with unique factorization n = p 1 p2 pk , then the number of positive divisors of n is

(1 + 1)(2 + 1) (k + 1). Denition 0.5.2 Let a, b, c Z, with c > 1. We say that a is congruent to b modulo c if c | (a b), and we denote ethos statement by a b (mod c). 10. If a b (mod m), then a b (mod m), a + x b + x (mod m), and ax bx (mod m) for every x Z. 11. If (c, m) = 1 and ca cb (mod m), then a b (mod m). 12. If (c, m) = 1, then cx b (mod m) has a unique solution x modulo m. In other words, the congruence is satised by some x Z, and if it is also satised by some other x Z, then x x (mod m). 13. If p is prime and c 0 (mod p), then cx b (mod p) has a unique solution x modulo p. 14. If a b (mod m) and c d (mod m), then a + c b + d (mod m) and ac bd (mod m). 15. If a, b, c Z and d = (a, b), then ax + by = c has integer solutions x and y if and only if d|c. Denition 0.5.3 LCM Let a, b Z. The least common multiple of a and b is the smallest positive integer d such that a | d and b | d. We often denote the least common multiple of a and b by [a, b]. 16. If a, b Z, then m[a, b] = [ma, mb] when m > 0. 17. If a and b are positive integers, then (a, b)[a, b] = ab. 22

18. If ca cb (mod m) and d = (c, m), then a b (mod

m d ).

19. If m, a, b Z, m > 1, the congruence ax b (mod m) is solvable if and only if (a, m)|b. There are exactly (a, m) solutions distinct modulo m. 20. If a, b, s, t Z are such that sa + tb = 1, then (a, b) = 1. Denition 0.5.4 Let m Z be greater than 1, and let P = {i Z | 1 i m 1, and (i, m) = 1}. A reduced residue system modulo m is a set Q of integers such that each of the integers in P is congruent modulo m to exactly one of the elements in Q. 21. The number of elements in a reduced residue system modulo m is independent of the representatives chosen. 22. If p is a prime and denotes Eulers function (where (a) is the number of integers between 0 and 1 ). a, inclusive, that are relatively prime to a), then (pn ) = pn pn1 = pn (1 p 23. The number of elements in a reduced residue system modulo m is (m). 24. If a1 , . . . , a(m) is a reduced residue system modulo m and (, m) = 1, then a1 , . . . , a(m) is a reduced residue system modulo m. 25. If m is a positive integer and (, m) = 1, then (m) 1 (mod m). 26. If d1 , . . . , dk are the positive divisors of n, then
k i=1

(di ) = n.

0.5.2 Ordered Integral DomainsThis project is designed to show that any ordered integral domain contains a copy of the integers. Thus, in particular, any ordered eld such as the rationals or real numbers contains a copy of the integers. Let R be an ordered integral domain. Denition 0.5.5 An inductive set in R is a subset S of R such that a. 1 S , and b. if x S , then x + 1 S . Example 0.5.6 i. R is an inductive subset of R. ii. S = {x R | x 1} is an inductive subset of R. Now dene N to be the intersection of all the inductive subsets of R. It is clear that N is an inductive subset of R. Of course, N is supposed to be the natural numbers. Since of all the axioms for a commutative ring with 1, as well as the order axioms, hold in R, we can use them freely in N . The following facts are easy to prove, so prove them. Facts 0.5.7 1. Suppose that S is a non-empty subset of N such that 1 S and if x S then x + 1 S , show that S = N . 2. Show that N is closed under addition. 3. Show that N is closed under multiplication. Hint: x x N and look at the set Mx = {y N | xy N }. Show that Mx is an inductive subset. 4. Show that the well ordering principle holds in N . 23

5. Show that all elements of N are positive. This is all ne, but where do we get the integers? Well, of course, we just tack on 0 and the negative natural numbers. Before nodding your head and shouting Hooray!, you must show that this new set Z = N {0} {n R | n N } is closed under multiplication and addition. 6. Show that if m, n N then m n Z . In particular if m N then m 1 Z . 7. Show that Z is closed under addition. 8. Show that Z is closed under multiplication. So we have that Z is an ordered integral domain in which the positive elements are well ordered. 9. Show that Z and the integers, Z, are order isomorphic. That is, there exists a bijection : Z Z such that (a) (x + y ) = (x) + (y ) for all x, y Z , (c) if x < y in Z , then (x) < (y ) in Z.

(b) (xy ) = (x)(y ) for all x, y Z , and

24

Chapter 1

The Real and Complex Numbers


Thus the System of Real Numbersthe denition of irrationals and the extension of the four species to the new numbersis established. The method has the advantage of simplicity in detail. It is well for the student, after a rst study of the method of Dedekind, to work it through in detail. He will then return to the former method with increased power and greater zest. The method of regular sequences is a middle-of-the-road method. It is an easy way to reach the mountain top. The traveller buys his ticket and takes the funicular. Many people prefer this mode of travel. But some like a sti climb over rocks and across streams, and such an ascent has its advantages if the heart is good and the muscles are strong. William Fogg Osgood Functions of Real Variables

In Chapter ??, we dened the integers and discussed their properties in some detail. We then constructed the rational numbers from the integers and observed that the rational numbers form an ordered eld. It follows from Project 0.2 that any eld which contains the integers must also contain the rationals as a subeld. Exercise 1.0.1 Prove that any eld that contains the integers contains the rationals as a subeld. In this chapter, we do several things. First, we introduce the real numbers by adding the Least Upper Bound Property to the axioms for an ordered eld. Second, despite Osgood, we construct the real numbers from the rational numbers by the method of Cauchy sequences. Third, we construct the complex numbers from the real numbers and prove a few useful theorems about complex numbers. Intermingled in all of this is a discussion of the elds of algebraic numbers and real algebraic numbers. As a project at the end of the chapter, we lead the reader through the construction of the real numbers via Dedekind cuts. In a second project, we study decimal expansions of real numbers.

1.1

The Least Upper Bound Property and the Real Numbers

Denition 1.1.1 Let F be an ordered eld. Let A be a nonempty subset of F . We say that A is bounded above if there is an element M F with the property that if x A, then x M . We call M an upper bound for A. Similarly, we say that A is bounded below if there is an element m F such that if x A, then m x. We call m a lower bound for A. We say that A is bounded if A is bounded above and A is bounded below.

25

Examples 1.1.2

i. Consider the subset A of Q: A= 1+ (1)n n nN .

Then A is bounded above by

3 2

and bounded below by 0.

ii. Let A = {x Q | 0 < x3 < 27}. Then A is bounded below by 0 and bounded above by 3. Exercise 1.1.3 Let a be a positive rational number. Let A = {x Q | x2 < a}. Show that A is bounded in Q. Denition 1.1.4 Let F be an ordered eld, and let A be a nonempty subset of F which is bounded above. We say that L F is a least upper bound for A if the following two conditions hold: a. L is an upper bound for A; b. if M is any upper bound for A, then L M . The denition of a greatest lower bound is similar with all of the inequalities reversed. Exercise 1.1.5 Show the least upper bound of a set is unique if it exists. Previously, we have discussed the real numbers in an informal way as a collection of decimal expansions, with the property that no expansion ends in all nines. Of course, this is not a formal denition of the real numbers, but it is common practice to work with the real numbers with this particular representation. We now give a formal denition of the real numbers which provides a working basis for proving theorems. Later in this chapter, starting with the rational numbers as an ordered eld we will give a precise construction of the real numbers as an ordered eld in which the least upper bound property holds. Denition 1.1.6 An ordered eld F has the least upper bound property if every nonempty subset A of F that is bounded above has a least upper bound. Exercise 1.1.7 Show that any two ordered elds with the least upper bound property are order isomorphic. We will see in Section ?? that an ordered eld with the least upper bound property exists. Thus, it makes sense to make the following denition. Denition 1.1.8 The real numbers are the (unique, up to order isomorphism) ordered eld that satises the least upper bound property. We will denote this eld by R. We say that the real numbers are an ordered eld with the least upper bound property. In many texts, the real numbers are dened as a complete ordered eld. This is actually a misuse of the word complete, which is dened in terms of the convergence of Cauchy sequences. This will be discussed later in this chapter. Exercise 1.1.9 Find the least upper bound in R of the sets in Exercise 1.1.3 and Example 1.1.2. Denition 1.1.10 An ordered eld F has the greatest lower bound property if every nonempty subset A of F that is bounded below has a greatest lower bound. That is, there exists an element of F such that: a. is a lower bound for A; b. if m is any lower bound for A, then m . Exercise 1.1.11 Prove that an ordered eld has the least upper bound property i it has the greatest lower bound property. 26

If L is the least upper bound of a set A, we write L = lub A or L = sup A (sup stands for supremum ). If is the greatest lower bound of a set A, we write = glb A or = inf A (inf stands for inmum ). Exercise 1.1.12 Let n be a positive integer that is not a perfect square. Let A = {x Q | x2 < n}. Show that a least upper bound in Q. Conclude that A is bounded in Q but has neither a greatest lower bound nor n exists in R , that is, there exists a real number a such that a2 = n. We have observed that the rational numbers are contained in R. A real number is irrational if it is not in Q. Fact 1.1.13 We can conclude from Exercise 1.1.12 that if n is a positive integer that is not a perfect square, then n exists in R and is irrational. Exercise 1.1.14 Suppose that A and B are bounded sets in R. Prove or disprove the following: i. lub(A B ) = max{lub(A), lub(B )}. ii. If A + B = {a + b | a A, b B }, then lub(A + B ) = lub(A) + lub(B ). iii. If the elements of A and B are positive and A B = {ab | a A, b B }, then lub(A B ) = lub(A)lub(B ). iv. Formulate the analogous problems for the greatest lower bound.

1.2

Consequences of the Least Upper Bound Property

We now present some facts that follow from the least upper bound property and the properties of the integers. The rst is the Archimedean property of the real numbers. Theorem 1.2.1 (Archimedean Property of R) If a and b are positive real numbers, then there exists a natural number n such that na > b. Proof. If a > b, take n = 1. If a = b, take n = 2. If a < b, consider the set S = {na | n N}. The set S = since a S . Suppose S is bounded above by b. Let L = lub S . Then, since a > 0, there exists an

element n0 a S such that L a < n0 a. But then L < (n0 + 1)a, which is a contradiction.

Corollary 1.2.2 If is a positive real number, there exists a natural number n such that

1 n

< .

Denition 1.2.3 Let F be an ordered eld. From Chapter 0, we know that Z F and by Exercise 1.0.1 we know Q F . We say that F is an Archimedean ordered eld if for every x F there exists N Z such that x < N . The elds Q and R are Archimedean ordered elds. Exercise 1.2.4 Let F be an Archimedean ordered eld. Show that F is order isomorphic to a subeld of R. Next, we show that every real number lies between two successive integers. Theorem 1.2.5 If a is a real number, then there exists an integer N such that N 1 a < N . so by the well ordering principle S has a least element N . Then N 1 / S , so N 1 a < N . We now show that there is a rational number between any two real numbers. 27 Proof. Let S = {n Z | n > a}. Then by the Archimedean property, S = . S is bounded below by a,

Theorem 1.2.6 If a and b are real numbers with a < b, there exists a rational number r = a < r < b.

p q

such that

1 < b a. Proof. From the Archimedean property of R (Corollary 1.2.2) there exists q N such that q Now consider the real number qa. By Theorem 1.2.5, there exists an integer p such that p 1 qa < p. It

follows that

p1 q

a< p q . This implies that

p q

1 q

a, or a <

p q

a+

1 q

< b.

Denition 1.2.7 A subset A of R is said to be dense in R if for any pair of real numbers a and b with a < b, there is an r A such that a < r < b. Corollary 1.2.8 The rational numbers are dense in the real numbers. How do the irrational numbers behave? Exercise 1.2.9 i. Show that any irrational number multiplied by any nonzero rational number is irrational. ii. Show that the product of two irrational numbers may be rational or irrational. Next we show that there is an irrational number between any two real numbers. Corollary 1.2.10 The irrational numbers are dense in R. Proof. Take a, b R such that a < b. We know that 2 is irrational and greater than 0. But then b a b a < . By Corollary 1.2.8, there exists a rational number p/q , with p = 0 such that <p . Thus q < 2 2 2 2 a < 2p/q < b, and 2p/q is irrational. The real numbers are the union of two disjoint sets, the rational numbers and the irrational numbers, and each of these sets is dense in R. Note that density implies nothing about cardinality since the rationals are countable and the irrationals are not, as shown in Section 0.3.

1.3

Rational Approximation

We have just shown that both the rational numbers and the irrational numbers are dense in the real numbers. But, really, how dense are they? It is reasonable to think that proximity for rational numbers can be measured in terms of the size of the denominator. To illustrate this, we ask the question, How close do two rational numbers have to be in order to be the same rational number? This is not a trick question it is designed to illustrate the principle mentioned above. Thus, if a/b, c/d Q and |a/b c/d| < 1/bd, then a/b = c/d. This idea can be encapsulated in the following theorem. Throughout this section, we shall assume that the denominator of a rational number is a positive integer and that the numerator and denominator are relatively prime. Theorem 1.3.1 If a/b is a xed rational number and p/q is a rational number such that 0 < |p/q a/b| < 1/mb for some positive integer m, then q > m. Proof. Simplify the subtraction of fractions and multiply both sides by bq . We now present several facts on rational approximation. For the rest of this section, we assume that the reader is familiar with the results contained in Project 0.5.1 at the end of Chapter ??. Exercise 1.3.2 Let a and b be relatively prime integers. Show that the equation ax + by = 1 has innitely many solutions (q, p) with q and p relatively prime integers. 28

Theorem 1.3.3 Let = a/b with a and b relatively prime and b = 1. Then there exist innitely many p/q Q such that |a/b p/q | < 1/q . q > 0. We then have |a/b p/q | = 1/bq < 1/q . Proof. Let (q, p) be a solution to the equation ax + by = 1. Then q = 0 since b = 1. We may assume

Remark 1.3.4 If b = 1 then the same result holds with < replaced by . The next theorem characterizes rational numbers in terms of rational approximation. We rst need the following exercise. Exercise 1.3.5 Let be a real number, and let and t be positive real numbers. Show that there exists only a nite number of rational numbers p/q with q < that satisfy | p/q | < 1/q t . Theorem 1.3.6 Let = a/b Q. Then there are only nitely many p/q so that | p/q | 1/q 2 . Proof. Suppose there are innitely many p/q satisfying the inequality. Then by the exercise above, q gets arbitrarily large. Thus there exists a p/q with q > b such that |a/b p/q | < 1/q 2 . This implies that

|aq bp| < b/q < 1, which is a contradiction.

We next consider rational approximation of irrational numbers. The question is: if is irrational, are there any rational numbers p/q satisfying the inequality | p/q | < 1/q 2 ? The armative answer follows from a theorem of Dirichlet on rational approximation of any real number. Theorem 1.3.7 (Dirichlet) Let be a real number and n a positive integer. Then there is a rational number p/q with 0 < q n satisfying the inequality 1 p . q (n + 1)q

Proof. If n = 1, then p/q = [] or p/q = [ + 1] satises | p/q | 1/2. (Recall that [] is the greatest integer less than or equal to ; for further details, see Appendix A.) Suppose that n 2. Consider the n + 2 numbers 0, [], 2 [2], . . . , n [n], 1 in the interval [0, 1]. Assume that the numbers in our list are distinct, which is the case if is irrational. By the pigeonhole principle, two of the numbers dier in absolute value by at most 1/(n + 1). If one of the numbers is 0 and the other is i [i], then i n, |i [i]| 1/(n + 1), and [i] 1 . i (n + 1)i

After [i]/i is reduced to lowest terms p/q , the rational number p/q satises the required inequality. Similarly, if the two numbers are j [j] and 1, then j n and reducing ([j] + 1)/j to lowest terms p/q , we have p/q satises the required inequality. Finally, if the two numbers are i [i] and j [j], where i < j , then 1 |j [j] (i [i])| = |(j i) + ([j] [i])| . n+1 Then 1 [j] [i] . ji (n + 1)(j i) Thus, after ([j] [i])/(j i) is reduced to lowest terms p/q , the rational number p/q satises the inequality because j i < n. 29

In the event that the n + 2 numbers are not distinct, then itself is a rational number with denominator at most n. For this case, either there exists 1 i n so that = or there exist 1 i < j n so that = [i] i

. Thus, if the numbers are not distinct the required inequality is trivially satised by itself.

[j] [i] ji

Corollary 1.3.8 Given any real number there is a rational number p/q such that | p/q | < 1/q 2 . Proof. This follows immediately from the theorem. Now comes the good news (or bad news depending on how you look at it). Theorem 1.3.9 If is irrational, then there are innitely many rational numbers p/q such that | p/q | < 1/q 2 . Proof. Suppose there are only a nite number of rational numbers p1 /q1 , p2 /q2 , . . . , pk /qk satisfying the inequality. Then, there is a positive integer n such that | pi /qi | > 1/(n + 1)qi for i = 1, 2, . . . , k . This contradicts Theorem 1.3.7 which asserts the existence of a rational number p/q satisfying q n and | p/q | < 1/(n + 1)q < 1/q 2 . So, there you have it, a real number is rational if and only if there exists only a nite number of rational numbers p/q such that | p/q | 1/q 2 . And a real number is irrational if and only if there exists an innite number of rational numbers p/q such that | p/q | 1/q 2 .

1.4

Intervals

At this stage we single out certain subsets of R which are called intervals. Denition 1.4.1 A subset of R is an interval if it falls into one of the following categories. a. For a, b R with a < b, the open interval (a, b) is dened by (a, b) = {x R | a < x < b}. b. For a, b R with a b, the closed interval [a, b], is dened by [a, b] = {x R | a x b}. c. For a, b R with a < b, the half open interval [a, b), is dened by [a, b) = {x R | a x < b}. d. For a, b R with a < b, the half open interval (a, b], is dened by (a, b] = {x R | a < x b}. e. For a R, the innite open interval (a, ), is dened by (a, ) = {x R | a < x}. f. For b R, the innite open interval (, b), is dened by (, b) = {x R | x < b}. g. For a R, the innite closed interval [a, ), is dened by [a, ) = {x R | a x}. h. For b R, the innite closed interval (, b], is dened by (, b] = {x R | x b}. i. R = (, ). 30

Denition 1.4.2 If x R a neighborhood of x is an open interval containing x. For many instances, it is useful to use symmetric neighborhoods. That is, if x R a symmetric neighborhood of x is an interval of the form (x , x + ), where > 0. These intervals, and their counterparts in other spaces, are used extensively throughout analysis. Exercise 1.4.3 Suppose that I is a subset of R, show that I is an interval if and only if for all a, b I , with a b, the closed interval [a, b] is contained in I . The notion of interval is valid in any ordered eld, and we will occasionally nd this useful. We end with this section with a theorem about intervals in R, which is called the Nested Intervals Theorem. Theorem 1.4.4 (Nested Intervals Theorem) Let ([an , bn ])nN be a nested sequence of closed bounded intervals in R. That is, for any n we have [an+1 , bn+1 ] [an , bn ], or equivalently, an an+1 bn+1 bn for all n. Then nN [an , bn ] = . Proof. Let A = {an | n N}. Then A is bounded above by b1 . If a = lubA, then a
nN [an , bn ].

The nested intervals property is actually not exclusive to the real numbers. In fact, it is really a theorem about a sequence of nested compact sets in a metric space. This result will be proved in the Chapter ??. There is often some confusion about the relationship between the nested interval theorem in R and the least upper bound property. Although our proof in R involves the least upper bound property, it can be done in alternate ways. Exercise 1.4.5 i. Give an example of a nested sequence of closed (but not necessarily bounded) intervals such that the intersection is empty. ii. Give an example of a nested sequence of bounded (but not necessarily closed) intervals such that the intersection is empty.

1.5

The Construction of the Real Numbers

We are now ready to proceed with the construction of the real numbers from the rational numbers using the fact that the rational numbers are the ordered eld constructed from Z in Chapter ??. We have already dened R as an ordered eld in which the least upper bound property holds. We now proceed to build such a eld starting from Q. Recall that the absolute value on Q is dened as follows: |a | = a a if a 0 if a < 0.

Also recall that the absolute value on Q satises the following three properties. 1. For any a Q, |a| 0, and |a| = 0 if and only if a = 0. 2. For any a, b Q, |ab| = |a||b|. 3. For any a, b Q, |a + b| |a| + |b| (triangle inequality). Exercise 1.5.1 Show that, for any a, b Q, we have ||a| |b|| |a b|. Denition 1.5.2 A sequence of rational numbers is a function f : N Q. We often denote such a sequence by (f (1), f (2), . . . , f (k ), . . .), or (a1 , a2 , . . . , ak , . . .), where ak = f (k ) for each k N. Most frequently, we will use (ak )kN for this same sequence. 31

Denition 1.5.3 A sequence (ak )kN of rational numbers is a Cauchy sequence in Q if, given any rational number r > 0, there exists an integer N such that if n, m N , then |an am | < r. Denition 1.5.4 A sequence (ak )kN converges in Q to a Q if, given any rational number r > 0, there exists an integer N such that, if n N , then |an a| < r. The rational number a is calld the limit of the sequence (ak )kN . Sometimes, we just say that the sequence (ak )kN converges in Q without mentioning the limit a. Exercise 1.5.5 If a sequence (ak )kN converges in Q, show that (ak )kN is a Cauchy sequence in Q. Exercise 1.5.6 Show that the limit a of a convergent sequence is unique. Denition 1.5.7 Let (ak )kN be a sequence of rational numbers. We say that (ak )kN is a bounded sequence if the set {ak | k N} is a bounded set in Q. Lemma 1.5.8 Let (ak )kN be a Cauchy sequence of rational numbers. Then (ak )kN is a bounded sequence. Proof. Let (ak )kN be a Cauchy sequence of rational numbers. Pick N N such that |an am | < 1 for n, m N . Then |an aN | < 1 for all n N , so that |an | < 1 + |aN | for all n N . Let M be the max of

|a1 |, |a2 |, . . . , |aN 1 |, 1 + |aN |. Then {|ak | | k N} is bounded above by M .

Let C denote the set of all Cauchy sequences of rational numbers. We dene addition and multiplication of Cauchy sequences term-wise, that is, (an )nN + (bn )nN = (an + bn )nN and (an )nN (bn )nN = (an bn )nN . Exercise 1.5.9 Show that the sum of two Cauchy sequences in Q is a Cauchy sequence in Q. Theorem 1.5.10 The product of two Cauchy sequences in Q is a Cauchy sequence in Q. Proof. Let (ak )kN and (bk )kN be Cauchy sequences in Q. By the lemma above, these sequences must be bounded. Let A and B be upper bounds for the sequences (|ak |)kN and (|bk |)kN , respectively. Let r > 0. Since (ak )kN is a Cauchy sequence, we can choose N1 such that if n, m > N1 , then |an am | < 2r B . Since (bk )kN is a Cauchy sequence, we can choose N2 such that if n, m > N2 , then |bn bm | < 2r . A Let N = max{N1 , N2 }. If n, m > N , then |a n b n a m b m | = |a n b n a n b m + a n b m a m b m | |an ||bn bm | + |bm ||an am | A|bn bm | + B |an am | < r.

Exercise 1.5.11 Show that, with addition and multiplication dened as above, C is a commutative ring with 1. Let I be the set of sequences (ak )kN in C with the property that, given any rational r > 0, there exists an integer N such that if n N , then |an | < r. The set I consists of Cauchy sequences that converge to 0. Lemma 1.5.12 Suppose (ak )kN C \ I , then there exists a positive rational number r and an integer N such that |an | r for all n N . 32

Proof. Suppose (ak )kN / I . Then there exists a rational number r > 0 such that |ak | 2r innitely often. Pick N N such that |an am | < r for n, m N . This implies that |an | > |am | r for n, m N . Fix an m N for which |am | 2r. Then for all n N , we have |an | > r. We can rephrase the statement of this lemma. We say that a property of rational numbers holds eventually for the terms of a sequence (ak )kN if there exists some N N such that the property holds for an whenever n N . So the lemma above says that for a Cauchy sequence (ak )kN that is not in I , there exists a positive rational number r such that |ak | is eventually greater than or equal to r. Exercise 1.5.13 Show that if a Cauchy sequence (ak )kN does not converge to 0, all the terms of the sequence eventually have the same sign. Denition 1.5.14 Let (ak )kN , and (bk )kN be Cauchy sequences in Q. We say that (ak )kN is equivalent to (bk )kN , denoted by (ak )kN (bk )kN , if (ck )kN = (ak bk )kN is in I . Exercise 1.5.15 Show that denes an equivalence relation on C . Denote by R the set of equivalence classes in C . We claim that, with appropriate denitions of addition and multiplication (already indicated above) and order (to be dened below), R is an ordered eld satisfying the least upper bound property. If (ak )kN is a Cauchy sequence, denote its equivalence class by [ak ]. As one might expect, the sum and product of equivalence classes are dened as follows: [ak ] + [bk ] = [ak + bk ] and [ak ][bk ] = [ak bk ]. Exercise 1.5.16 Show that addition and multiplication are well dened on R. Exercise 1.5.17 Show that R is a commutative ring with 1, with I as the additive identity and [ak ] such that ak = 1 for all k as the multiplicative identity. This follows easily from Exercise 1.5.11. Theorem 1.5.18 R is a eld. Proof. We need only show that multiplicative inverses exist for non-zero elements. So assume that [ak ] = I . Then, as we saw in Lemma 1.5.12, ak is eventually bounded below in absolute value. Hence, we can pick M N and c > 0 such that |ak | > c for all k M . Dene a sequence (bk )kN as follows: bk = 1 for k M , and bk = 1/ak for k > M . Observe that for n, m large enough, 1 |a n a m | 1 1 = 2 |a n a m |. an am |a n a m | c Hence, (bk )kN is a Cauchy sequence. It is clear by construction that [bk ] is the multiplicative inverse of [ak ].

The next step is to dene order on R. Let [ak ] and [bk ] represent distinct elements of R. Then [ck ] = [ak bk ] is not equal to I . Hence there exists N N such that all the terms of ck have the same sign for k N . Thus, either ak < bk for all k N or bk < ak for k N . We use this fact to dene an order on R. Denition 1.5.19 Let a = [ak ], b = [bk ] be distinct elements of R. We dene a < b if ak < bk eventually and b < a if bk < ak eventually. Exercise 1.5.20 Show that the order relation on R dened above is well-dened and makes R an ordered eld. 33

To nish this o, we must show that R is an Archimedean ordered eld that satises the least upper bound property. We will have then reached the Osgoods mountain top so we can dismount the funicular and ski happily down the slope. Dene a map i : Q R by sending r Q to the equivalence class of the constant sequence (r, r, . . . ). It is evident that this map is injective and order-preserving, so we may consider Q R as ordered elds. Theorem 1.5.21 The eld R is an Archimedean ordered eld. Proof. Suppose a R and a > 0. Let (ak )kN represents a. As noted above, the Cauchy sequence (ak )kN is bounded above by some integer N , that is, ak < N for all suciently large k . It follows that a is less than the integer (N, N, . . . ) in R (under the inclusion Q R). Theorem 1.5.22 The least upper bound property holds in R. Proof. Let A be a nonempty subset of R that is bounded above by, say, m. Then, by the Archimedean property, we can nd M Z with m M . Let a be in A and let n be an integer with n < a. For p N, set Sp = {k 2p | k Z and n k 2p M } {m}. Note that Sp = and is nite. Now let ap = min{x | x Sp and x is an upper bound for A}. Note that if p < q , then ap 2 p < aq ap , |ap aq | 2p for all p < q from which it follows that (ak )kN is a Cauchy sequence. Let L = [ak ]. We claim that L is a least upper bound for A. Suppose x A and x > L. Choose p such that 2p < (x L) (using the Archimedean property). Since ap 2p < aq for p < q and (ap ) is a decreasing Cauchy sequence, it follows that ap 2p L ap . In particular if we add 2p < x L and ap 2p L we obtain ap < x which is a contradiction. Therefore L is an upper bound for A. Suppose that H is an upper bound for A and H < L. Choose p such that 2p < L H . Take x A such that ap 2p < x. Then ap 2p < H . Adding, we get ap < L. But, as noted above, L ap for all

since, for example, ap 2p is not an upper bound for A, while aq is an upper bound. But this implies that

p N, so this is a contradiction. In this section, we have constructed an ordered eld R in which the least upper bound property holds. Following our discussion in Section ??, this means that R must in fact, up to order isomorphism, be the real numbers R. From now on, we will refer to the real numbers exclusively as R.

1.6

Convergence in R

We dene the absolute value on R in exactly the same manner as on Q. Denition 1.6.1 Suppose x R. The absolute value of x is dened by |x| = x if x 0, x if x < 0.

The following are the essential properties of the absolute value. Theorem 1.6.2 Properties of absolute value on R 1. For any x R, |x| 0, and |x| = 0 i x = 0. 2. For any x, y R, |xy | = |x||y |. 34

3. For any x, y R, |x + y | |x| + |y | (triangle inequality). Exercise 1.6.3 Prove the properties of the absolute value. With absolute value dened, we can talk about Cauchy and convergent sequences in R. Of course, we should rst dene what a sequence in R is. Denition 1.6.4 A sequence of real numbers is a function f : N R. As with rational numbers, we will usually denote a sequence of real numbers by (ak )kN , where ak = f (k ). Denition 1.6.5 A sequence (ak )kN of real numbers is convergent if there exists an element a R such that given any > 0, there exists N N such that k N implies that |ak a| < . We say that (ak )kN converges to a, and a is called the limit of the sequence (ak )kN . Symbolically, we write
k

lim ak = a.

We will often say that a sequence of real numbers is convergent without specic reference to the limit a. Note that N depends on the choice of . Exercise 1.6.6 Show that the limit a of a convergent sequence is unique. Denition 1.6.7 Given a sequence (ak )kN a subsequence is a sequence of the form (akj )j N , where (kj )j N is a strictly monotonic increasing sequence of natural numbers. In other words, a subsequence of a sequence contains some, but not necessarily all, terms of the original sequence in their original order. Exercise 1.6.8 Let (ak )kN be a convergent sequence. Show that any subsequence converges to the same limit. Denition 1.6.9 A sequence (ak )kN of real numbers is bounded if the set {ak | k N} is bounded. Exercise 1.6.10 Show that any convergent sequence of real numbers is bounded. Exercise 1.6.11 Find a bounded sequence of real numbers that is not convergent. Denition 1.6.12 A sequence (ak )kN of real numbers is monotonic increasing if ak ak+1 for all k N. A sequence (ak )kN of real numbers is strictly monotonic increasing if ak < ak+1 for all k N. Similar denitions hold for monotonic decreasing and strictly monotonic decreasing sequences with the inequalities reversed. A sequence is called monotonic if it is monotonic increasing or monotonic decreasing. The following lemmas fundamental in discussing convergence in R. Lemma 1.6.13 Let (ak )kN be a sequence in R. Then (ak )kN has a monotonic subsequence. Proof. Suppose (ak )kN does not have a monotonic increasing subsequence. Then, there exists n1 N such that an1 > ak for all k > n1 . Again, since (ak )k>n1 does not have a monotonic increasing subsequence, there exists n2 > n1 such that an2 > ak for all k > n2 . Moreover an1 > an2 . Continuing in this way, we obtain a strictly monotonic decreasing subsequence (an1 , an2 , . . .).

Lemma 1.6.14 Every bounded monotonic sequence in R converges to an element in R. 35

k > N , we have a ak aN > a . So limk ak = a.

Proof. Suppose (ak )kN is monotonic increasing and bounded. Let a be the least upper bound of the set {a1 , a2 , . . .}. For all > 0, there exists an N such that a < aN a. Since (ak )kN is increasing, if

Lemma 1.6.15 Every bounded sequence in R has a convergent subsequence. Exercise 1.6.16 Prove Lemma 1.6.15. Even if a bounded sequence of real numbers is not itself convergent, we can still say something about its long-term behavior. Denition 1.6.17 Let (ak )kN be a bounded sequence of real numbers. For each n N, dene bn = sup{ak | k n}, and cn = inf {ak | k n}. We dene the limit supremum of the sequence (ak )kN to be limn bn , and we denote this by lim supk ak . We dene the limit inmum of (ak )kN similarly: lim inf k ak = limn cn . Exercise 1.6.18 Show that if (ak )kN is a bounded sequence, then lim supk ak and lim inf k ak exist. Exercise 1.6.19 Show that if (ak )kN is convergent, then lim supk ak = lim inf k ak = limk ak . Exercise 1.6.20 Show that if lim supk ak = lim inf k , then (ak )kN is convergent, and limk ak = lim supk ak = lim inf k ak . Exercise 1.6.21 For the following sequences, compute lim sup and lim inf. i. ak = 1 + (1)k .
1 . ii. bk = (1)k + k

iii. ck = cos k . As we did for sequences of rational numbers, we can dene Cauchy sequences of real numbers. However, unlike the situation in the rational numbers, we will see that every Cauchy sequence of real numbers converges in R. Denition 1.6.22 (See Denition 1.5.3) A sequence (ak )kN in R is a Cauchy sequence if, given any > 0, there exists N N such that n, m N implies |am an | < . Exercise 1.6.23 i. Prove that every Cauchy sequence in R is bounded. ii. If (ak )kN is a Cauchy sequence in R, show that for any > 0 there exists a subsequence (akj )j N such that |akj akj+1 | < /2j +1 for j N. Theorem 1.6.24 (Cauchy Criterion) A sequence (ak )kN of real numbers is convergent if and only if it is a Cauchy sequence. Proof. We already did half of this in Q, (see Exercise ??) but we will do it again. First, we prove that if (ak )kN is convergent, then it is Cauchy. Suppose limk ak = a. Then, since the sequence converges, given > 0, there exists N N such that |an a| < 2 for all n N . Thus, if n, m Ne psilon, we have |a n a m | | a n a | + |a m a | < + = , 2 2

and so (ak )kN is a Cauchy sequence. Suppose now that (ak )kN is a Cauchy sequence in R. Then, by Exercise 1.6.23, (ak )kN is a bounded sequence, and hence by Lemma 1.6.15 has a convergent subsequence. Call the limit of this subsequence a. Then, since (ak )kN is Cauchy, it is clear that limk ak = a. 36

Exercise 1.6.25 Show that if (an )nN and (bn )nN are Cauchy sequences in R, then (an + bn )nN and (an bn )nN are Cauchy sequences in R. Denition 1.6.26 Let S be a subset of R. Then x R is an accumulation point of S if, for all > 0, we have ((x , x + ) \ {x}) S = . Remark 1.6.27 Thus, x is an accumulation point of S if every interval around x contains points of S other than x. Of course, x does not have to be an element of S in order to be an accumulation point of S . Exercise 1.6.28 Find the accumulation points of the following sets in R. i. S = (0, 1); ii. S = {(1)n + iii. S = Q; iv. S = Z; v. S is the set of rational numbers whose denominators are prime. Lemma 1.6.29 Let S be a subset of R. Then every neighborhood of an accumulation point of S contains innitely many points of S . Proof. Let x be an accumulation point of S . Given > 0, there is a point x1 (x , x + ) S such that x1 = x. Let 1 = |x x1 |. Then, there is a point x2 (x 1 , x + 1 ) S such that x2 = x. Iterating
1 n

| n N} ;

this procedure, we get an innite set of elements in S that is contained in (x , x + ). Exercise 1.6.30 You prove the converse. Now here is a Big Time Theorem.

Theorem 1.6.31 (BolzanoWeierstrass) Let S be a bounded, innite subset of R. Then S has an accumulation point in R. Proof. Pick an innite sequence (ak )kN of distinct elements of S . Then, by Lemma 1.6.15, (ak )kN has a convergent subsequence,(bkj )j N . If limj akj = b , then b is an accumulation point of S . Exercise 1.6.32 i. Find an innite subset of R which does not have an accumulation point. ii. Find a bounded subset of R which does not have an accumulation point. Denition 1.6.33 Let S be a subset of R. We say that S is an open set in R if, for each point x S , there is an > 0 (depending on x) such that (x , x + ) S . Denition 1.6.34 Let S R. We say S is a closed set in R if the complement of S is an open set in R. Note that the empty set and R are both open and closed subsets of R. Exercise 1.6.35 i. Show that and R are the only subsets of R that are both open and closed in R. 37

ii. Show that every nonempty open set in R can be written as a countable union of pairwise disjoint open intervals. iii. Show that an arbitrary union of open sets in R is open in R. iv. Show that a nite intersection of open sets in R is open in R. v. Show, by example, that an innite intersection of open sets is not necessarily open. vi. Show that an arbitrary intersection of closed sets in R is a closed set in R. vii. Show that a nite union of closed sets in R is a closed set in R. viii. Show, by example, that an innite union of closed sets in R is not necessarily a closed set in R. Exercise 1.6.36 Show that a subset of R is closed i it contains all its accumulation points. Exercise 1.6.37 We dene the Cantor set to be a subset of the closed interval [0, 1]. First, remove the open interval (1/3, 2/3) from [0, 1]. Next, remove the open intervals (1/9, 2/9) and (7/9, 8/9). At each step, remove middle third of the remaining closed intervals. Repeating this process a countable number of times, we are left with a subset of the closed interval [0, 1] called the Cantor set. Show that: i. the Cantor set is closed; ii. the Cantor set consist of all numbers in the closed [0, 1] whose ternary expansion consists of only 0s and 2s and may end in innitely many 2s; iii. the Cantor set is uncountable; iv. every point of the Cantor set is an accumulation point of the Cantor set; v. the complement of the Cantor set in [0, 1] is a dense subset of [0, 1]. The next theorem, the Heine-Borel theorem for R, is the second of the two basic topological theorems for the real numbers, the rst of which is the Bolzano-Weierstrass theorem. We shall see versions of these theorems again in Chapter ??. Theorem 1.6.38 (Heine-Borel) Let S be a closed and bounded subset of R. Given a collection {Ui }iI of open sets such that S iI Ui , there exists a nite subcollection U1 , . . . , Un of {Ui }iI such that S U1 . . . Un . Proof. Suppose that S is a nonempty, closed, bounded subset of R. If a = glb(S ) and b = lub(S ), then, since S is closed, a and b are in S , and S [a, b]. Let {Ui }iI be a collection of open sets such that S iI Ui . By adjoining the complement of S (if necessary), we obtain a collection U of open sets whose union contains [a, b]. Now let B = {x [a, b] | [a, x] is covered by a nite number of open sets in U }. Then B is nonempty since a B , and B is bounded above by b. Let c = lub(B ). Exercise 1.6.39 Prove that c B . (Hint: Prove that B is closed.) Suppose c < b. Let U be a nite subcollection of U that covers [a, c]. Because any element of U that contains c is open, there exists y such that c < y < b and [c, y ] is in the same open set that contains c. Thus [a, y ] is covered by U . This is a contradiction, and hence b must equal c. Thus [a, b] is covered by a nite number of open sets from U , and by throwing away the complement of S (if necessary), S is covered by a nite number of open sets from the original collection. The ideas contained in the Heine-Borel theorem are important enough to deserve their own denitions. We introduce them here in the context of R, but we will see them again in Chapter ?? when we talk about metric spaces. 38

Denition 1.6.40 Let A be a subset of R. An open covering of A is a collection of open sets {Ui }iI such that A iI Ui . Denition 1.6.41 Let A be a subset of R. We say that A is a compact set if every open covering of A has a nite subcovering. That is, if {Ui }iI is an open covering of A, there is a nite subcollection U1 , U2 , . . . , Un of the collection {Ui }iI such that A U1 U2 Un This notion of compactness is more subtle than it appears. It does not say that in order for a set to be compact, it must have a nite open covering. Note that this is true of any subset of R, since we can take the single set U1 = R to be a covering. The stress in the denition is that every open covering must have a nite subcovering. This is a big idea in analysis. It allows us to reduce certain arguments about innite sets to arguments about nite sets. We can rephrase the Heine-Borel theorem to say that closed and bounded subsets of R are compact. In fact, the converse of this statement is also true, namely, that all compact subsets of R are closed and bounded. Exercise 1.6.42 Show that a compact subset of R is both closed and bounded. Exercise 1.6.43 Let S = (a, b). Give an example of an open covering of S that does not have a nite subcovering. Exercise 1.6.44 Let S = Z. Give an example of an open covering of S that does not have a nite subcovering. There is another closely related notion regarding subsets of R. Denition 1.6.45 A subset A of R is sequentially compact if every innite sequence in A has a subsequence that converges to an element of A. The following theorem can be proved easily using the Bolzano-Weierstrass and Heine-Borel theorems in R. Exercise 1.6.46 A subset of R is compact if and only if it is sequentially compact. We will see in Chapter ?? that the same theorem is true in metric spaces. In Section ??, we give an indication of how this works in C.

1.7

The complex numbers C

To start this section, we give a somewhat inexact denition of complex numbers. This notion is often used as the denition of the complex numbers, but it does contain some ambiguity which we will rectify immediately. Denition 1.7.1 (Rural Denition) The set of complex numbers, C, is the collection of expressions of the form z = a + bi where a, b R, and i is a symbol which satises i2 = 1. If z = a + bi and w = c + di are in C, then we dene z + w = (a + c) + (b + d)i, and zw = (ac bd) + (bc + ad)i. Actually, one can go a long way with this denition if the symbol i with the property that i2 = 1 doesnt cause insomnia. In fact, though, once you assert that i2 = 1, you must accept the fact that (i)2 = 1, and hence there is some ambiguity in the choice of which of i and i is the square root of 1. This diculty is avoided in the following construction. We consider the Cartesian product R R with addition and multiplication dened by (a, b) + (c, d) = (a + c, b + d), (a, b)(c, d) = (ac bd, bc + ad). 39

Exercise 1.7.2 Show that R R with addition and multiplication as dened above is a eld, with (0, 0) as the additive identity, (1, 0) as the multiplicative identity, (a, b) = (a, b), and (a, b)1 = (a/(a2 + b2 ), b/(a2 + b2 )) if (a, b) = (0, 0). Denition 1.7.3 The eld of complex numbers is the set C = R R with the operations of addition and multiplication dened above. Note that R is isomorphic to the subeld of C given by {(a, 0) | a R}. If we set i = (0, 1), then i2 = (1, 0). Finally, to x things up real nice, we write (a, b) = (a, 0) + (b, 0)(0, 1), or, returning to our original rural denition, (a, b) = a + bi. The rst observation to make is that C cannot be made into an ordered eld. That is, it cannot satisfy the order axioms given in Section 0.1. This is immediate because in any ordered eld, if a = 0 then a2 > 0. This would imply that i2 = 1 > 0, but 12 = 1 > 0 and this is a contradiction. Exercise 1.7.4 Show that the eld of complex numbers is not isomorphic to the eld of real numbers. Denition 1.7.5 If z = a + bi with a, b R, we call a the real part of z and b the imaginary part of z . We write a = Re z and b = Im z . The complex number z is called pure imaginary if a = Re z = 0. Denition 1.7.6 If z = a + bi with a, b R, the complex conjugate of z , denoted z , is the complex number z = a bi. Exercise 1.7.7 Prove the following statements about complex conjugates: i. z R i z = z ; ii. z + w = z + w ; iii. zw = zw ; iv. zz R. Denition 1.7.8 If z = a + bi with a, b R, the absolute value of z is |z | = (zz ) 2 = (a2 + b2 ) 2 , where, of course, we mean the nonnegative square root in R. If z and w are complex numbers, then |z | and |w| are real numbers, and hence it makes sense to say that |z | < |w|. However, it makes no sense to say that z < w since C is not an ordered eld. Exercise 1.7.9 Show that if we identify z = a + bi with the point (a, b) R2 , then the absolute value of z is equal to the distance to the point (a, b) from (0, 0).
Im(z ) z =(Re(z), Im(z))
1 1

|z | Re(z )

40

Exercise 1.7.10 Show that the absolute value on C satises all the properties of the absolute value on R. 1. For any z C, we have |z | 0, and |z | = 0 i z = 0. 2. For any z, w C, we have |zw| = |z ||w|. 3. For any z, w C, we have |z + w| |z | + |w| (triangle inequality). Exercise 1.7.11 Why is the triangle inequality so named? Denition 1.7.12 ?? If z = x + iy C, z = 0, and r = |z |, then the polar form of z is z = r(cos + i sin ) where is the unique solution to the equations x = r cos , y = r sin , in the interval [0, 2 ). The angle is called the principal branch of the argument of z and is denoted Arg(z ). For z as above, we often write z = rei , where ei is dened to be cos + i sin . (In fact, cos + i sin is n the value of the complex exponential function f (z ) = ez , dened by the power series ez = n=0 z n! , when z = i. See Section ??.) Exercise 1.7.13 Show that there is a unique with 0 < 2 simultaneously satisfying cos = sin =
b a2 +b2
1 2

a a2 +b2

1 2

for a pair of real numbers a and b not both zero.

Exercise 1.7.14 Prove by induction that (ei )n = ein for n N. Exercise 1.7.15 Suppose that n N. Prove that, if z = e n , for k Z and 0 k n 1, then z n = 1. Such a z is called an n-th root of unity. Note that these n roots of unity are all distinct. Remark 1.7.16 The n-th roots of unity form a cyclic group of order n under multiplication. An n-th root of unity is primitive if it is a generator of this group. In fact, the primitive n-th roots of unity are those of the form e2ik/n where k and n are relatively prime. (See Project ??.) Proposition 1.7.17 If n > 1, the sum of the n distinct n-th roots of unity is 0. Proof. For any z C, Now let z = e
2i n 2ki

(1 z n ) = (1 z )(1 + z + z 2 + + z n1 ).

Exercise 1.7.18 Suppose z is a nonzero complex number, and write z = rei . Show that z has exactly n distinct complex n-th roots given by r1/n ei(2k+)/n for 0 k n 1.

1.8

Convergence in C

Now that we have an absolute value on C, we can dene the notion of Cauchy sequence and convergent sequence in C. Denition 1.8.1 A sequence (zk )kN of complex numbers is convergent if there exists an element z C such that the sequence satises the following property: given any > 0, there exists N N such that k N implies that |zk z | < . We say that (zk )kN converges to z , and z is called the limit of the sequence (zk )kN . Symbolically, we write lim zk = z.
k

41

We will often say that a sequence of complex numbers is convergent without specic reference to the limit z . Note that in the denition, is a positive real number (this is already implied by the use of the inequality symbol, but we repeat it here for emphasis). Note too that N depends on . As usual, the limit of a convergent sequence is unique. Denition 1.8.2 Let r be a positive real number, and let z0 C. The open ball of radius r with center at z0 is the set Br (z0 ) = {z C | |z z0 | < r}. (1.1) The closed ball of radius r with center z0 is the set B r (z0 ) = {z C | |z z0 | r}. (1.2)

The open balls and closed balls in C are the analogs of open and closed intervals in R. We can dene open and closed sets in C in a fashion similar to the denitions in R. Denition 1.8.3 Let S be a subset of C. We say that S is an open set in C if, for each point z S , there is an > 0 (depending on z ) such that B (z ) S . Denition 1.8.4 Let S C. We say that S is a closed set in C if the complement of S is an open set in C. Exercise 1.8.5 i. Show that the empty set and C are both open and closed subsets of C. ii. Show that no other subsets of C besides and C are both open and closed in C. iii. Show that an arbitrary union of open sets in C is an open set in C. iv. Show that a nite intersection of open sets in C is an open set in C. v. Show, by example, that an innite intersection of open sets in C need not be an open set in C. vi. Show that an arbitrary intersection of closed sets in C is a closed set in C. vii. Show that a nite union of closed sets in C is a closed set in C. viii. Show, by example, that an innite union of closed sets in C is not necessarily a closed set in C. Exercise 1.8.6 i. Let (a, b) and (c, d) be open intervals in R. Show that the open rectangle (a, b) (c, d) = {x + iy C | x (a, b), y (c, d)} is an open set in C. ii. Let [a, b] and [c, d] be closed intervals in R. Show that the closed rectangle [a, b] [c, d] = {x + iy C | x [a, b], y [c, d]} is a closed set in C. iii. Let z C, and let S be an open set containing z . Show that there exists an open rectangle R = (a, b) (c, d) such that z R and R S . Exercise 1.8.7 Consider the collection of open balls {Br (z )} in C where r Q and Re(z ), Im(z ) Q. Show that any open set in C can be written as a nite or countable union from this collection of sets. Exercise 1.8.8 Show, by example, that there are open sets in C that cannot be written as the countable union of pairwise disjoint open balls. Denition 1.8.9 Let A C. The set A is bounded if there exists r > 0 such that A Br (0). 42

Exercise 1.8.10 Dene the notion of a bounded sequence in C. Denition 1.8.11 (See Denition 1.6.22) A sequence (zk )kN in C is a Cauchy sequence if, given any > 0, there exists N N such that n, m N implies |zm zn | < . Exercise 1.8.12 Prove that every Cauchy sequence in C is bounded. Theorem 1.8.13 (Cauchy Criterion) A sequence (zk )kN of complex numbers is convergent if and only if it is a Cauchy sequence. Proof. The rst half of the proof is identical to the proof of Theorem 1.6.24. Suppose now that (zk )kN is a Cauchy sequence in C. Let zk = ak + bk i, where ak , bk R. Then |zm zn |2 = (am an )2 + (bm bn )2 . Hence, |am an | |zm zn |, and since (zk )kN is a Cauchy sequence, it follows that (ak )kN is a Cauchy sequence in R. Similarly, (bk )kN is a Cauchy sequence in R. If limk ak = a and limk bk = b, then limk zk = z where z = a + bi. Exercise 1.8.14 Show that every bounded sequence in C has a convergent subsequence. Denition 1.8.15 Let S be a subset of C. Then z is an accumulation point of S if, for all > 0 we have (B (z ) \ {z }) S = . Remark 1.8.16 Thus, z is an accumulation point of S if every open ball around z contains points of S other than z . Of course, z does not have to be an element of S in order to be an accumulation point of S . Exercise 1.8.17 Find the accumulation points of the following sets: i. S = {z C | |z | = 1} (this is the unit circle in C); ii. S = {z C | Re z > Im z }; iii. S = {a + bi | a, b Q}; iv. S = {a + bi | a, b Z};
1 v. S = { n + 1 mi

| n, m N}.

Exercise 1.8.18 i. Let S be a subset of C. Show that every open set containing an accumulation of S contains innitely many points of S . ii. (BolzanoWeierstrass Theorem for C) Prove that any bounded innite set in C has an accumulation point in C. Denition 1.8.19 Let S be a subset of C. An open covering of S is a collection of open sets {Ui }iI such that S iI Ui . Denition 1.8.20 Let S be a subset of C. We say that S is compact if every open covering of S has a nite subcovering. That is, if {Ui }iI is an open covering of S , there is a nite subcollection U1 , U2 , . . . , Un of the collection {Ui }iI such that S U1 U2 Un . Theorem 1.8.21 (Heine-Borel) If S is a closed and bounded subset of C, then S is compact. 43

Proof. For the purposes of this proof, we recall that C = {(x, y ) | x, y R} as in Denition ??. We prove it for S = [a, b] [c, d], where a, b, c, d R and a < b and c < d, and leave the general case as an exercise. Take a point x0 [a, b] and consider the set {x0 } [c, d]. We take an open set N C containing {x0 } [c, d]. We claim that there exists an open interval I containing x0 such that I [c, d] N . We see this as follows. For each point (x0 , y ) {x0 } [c, d], choose ry > 0 such that the open square (x0 ry , x0 + ry ) (y ry , y + ry ) N . (See exercise ??.??.) By intersecting these squares with {x0 } R, we get a collection of open intervals of the form (y ry , y + ry ) that cover [c, d]. By the Heine-Borel theorem in R, there exists a nite subcollection of these open intervals that covers the interval [c, d]. Hence the corresponding collection of open squares also covers {x0 } [c, d]. Let r be the minimum of the ry from this nite collection. Then I = (x0 r, x0 + r) is our desired interval. Now let {Uj }j J be an open covering of S . For each x [a, b], the collection {Uj }j J covers {x} [c, d]. As we did above, we choose a nite subcollection U1 , . . . , Un that covers {x} [c, d]. The open set Nx = U1 Un contains a set of the form Ix [c, d] by the preceding discussion, where Ix is an open interval containing x. The collection {Ix }x[a,b] covers [a, b], and hence by the Heine-Borel theorem for R, there exists a nite subcollection Ix1 , . . . , Ixm that covers [a, b]. We take our nite subcollection of the original open cover {Uj }j J to be {U | for some xi , the set U is one of the elements in the union that denes Nxi }.

Exercise 1.8.22 Prove the general case of the Heine-Borel theorem in C. (Hint: Take a closed bounded set in C and put it inside the product of two closed bounded intervals. Then use the result from the proof above.) Exercise 1.8.23 Show that a subset of C is closed i it contains all its accumulation points. Exercise 1.8.24 Dene the notion of sequentially compact for a subset of C, and show that a subset of C is sequentially compact if and only if it is closed and bounded. (See Denition 1.6.45.)

1.9

Innite Series

We assume that the reader has had at least an elementary introduction to innite series and their convergence properties. In fact, the theory of innite series actually reduces to the convergence of sequences, which we have covered thoroughly in this chapter. An innite series is expressed as a sum of an innite number of elements from some place where addition makes sense. These elements could be numbers, functions or what have you, so we begin with one-sided series of numbers. We take an innite series to be an expression of the form n=1 an , where the elements an come from a number system in which addition makes sense. So that we dont wander around aimlessly, lets x our number system to be the complex numbers, that is an C, with the possibility of restricting to the real numbers or even the rational numbers. In the denition, we have chosen to use the natural numbers as the index set, but in considering innite series we could start the summation with any integer n0 and write n=n0 an . Later, we will also consider two-sided series where the index set is the entire set of integers and we write an . If these expressions are going to have any meaning at all, we must look at the partial sums. Denition 1.9.1 If N SN = n=1 an . Examples 1.9.2 i. Let an = 1 for all n. Then SN = N . ii. Let an = 1/n. Then SN = 1 + 1/2 + + 1/N . 44
n=1

an is an innite series of complex numbers, the N -th partial sum of the series is

iii. Let an = 1/2n . Then SN = 1 1/2N . iv. Let an = (1)n+1 . In this case, SN = 1 if N is odd and 0 if N is even. v. Fix , with 0 < < 2 , and let an = ein /n. Then SN = without more information about . vi. Let an = sin n/n2 . In this case, SN =
N n=1 N in /n, n=1 e

which is the best we can do

sin(n )/n2 .
N

Denition 1.9.3 Let n=1 an be an innite series of complex numbers. If N N, we let SN = n=1 an . The sequence (SN )N N is called the sequence of partial sums. We say that the series n=1 an converges if the sequence of partial sums (SN )N N converges, and we call the sum of a convergent series n=1 an the number S to which the sequence (SN )N N converges. If the sequence (SN )N N does not converge we say that n=1 an diverges. Of course, since we are working in C, the series converges if and only if the sequence (SN )N N is a Cauchy sequence. That is, given > 0, there is a N N such that for n, m > N (assuming n > m), then n | k=m+1 an | < . Exercise 1.9.4 Determine which of the series in Example 1.9.2 converge. We are faced with two problems. The rst is, How do we tell if a series converges? The second is, If a series does converge, how do we nd the explicit sum? There is extensive literature about these two questions, but the fact is that the second question presents many more diculties than the rst. In Chapter ??, the theory of Fourier Series will provide some assistance for some important special cases. The most helpful series in all of this discussion is a geometric series. Denition 1.9.5 Let z be a complex number. The geometric series dened by z is Exercise 1.9.6 i. If N N and z = 1, show that SN = ii. If |z | < 1, show that limn z n = 0. iii. If |z | > 1, show that limn z n does not exist. Theorem 1.9.7 Consider the geometric series dened by a complex number z . If |z | < 1, then the series converges. If |z | > 1, then the series diverges. Exercise 1.9.8 i. Prove the theorem using the exercise above. ii. What can you say about a geometric series for which |z | = 1? Theorem 1.9.9 Suppose that a series Exercise 1.9.10 Prove this. The more useful phrasing of this theorem is often the contrapositive; namely, if the terms of a series do not go to zero, then the series must diverge. Note, however, that the property that limn an = 0 does not ensure that the series n=1 an converges. The most useful example is given above where an = 1/n. In this case, S1 = 1, S4 > 2, it is easy to check that S2n > n for n N, and hence the series n=1 1/n diverges. 45
n=1 N n=0 n=0

z n.

zn =

1z N +1 1z .

an converges. Show that limn an = 0.

Exercise 1.9.11 The series S = n=1 1/n is often called the harmonic series. We have just proved that this series diverges. Show that, by suitably eliminating an innite number of terms, the remaining sub-series can be made to converge to any positive real number. Denition 1.9.12 A series verges. Proposition 1.9.13 If
n=1 n=1

an of complex numbers converges absolutely if the series


n=1

n=1

|an | con-

an converges absolutely, then


n k=m+1

an converges. |ak |, by the Triangle Inequality.

Proof. This follows from the fact that |

ak |

n k=m+1

The converse to Proposition 1.9.13 is false, and is shown by the example n=1 (1)n+1 /n. This series n converges since | k=m+1 (1)k+1 /k | < 1/m. However as we have seen above the series does not converge absolutely. There are various tests to determine if a series converges. These include the comparison test, the ratio test, and the root test. The comparison test is often very useful, but its use depends on knowing ahead of time a series which converges. Theorem 1.9.14 (Comparison Test) Suppose that an > 0 for every n N and that n=1 an converges. If bn C satises |bn | an for all n, then the series n=1 bn converges absolutely and hence converges. Proof. For each N N, let SN = n=1 an , and let S = limN SN . Let TN = n=1 |bn |. Then for every N N, TN SN , and hence TN S . Thus, TN is a monotonic bounded sequence of real numbers, which must converge.
N N

Exercise 1.9.15 i. If the series to cs.


n=1

an converges to s and c is any constant, show that the series

n=1

can converges

ii. Suppose that n=1 an and n=1 bn are innite series. Suppose that an > 0 and bn > 0 for all n N and that limn an /bn = c > 0. Show that n=1 an converges if and only if n=1 bn converges. Exercise 1.9.16 i. If p R and p < 1, show that ii. If p R and p > 1, show that
N n=1 n=1

1/np diverges.

1/np converges. (Hint: Use the fact from elementary calculus that
N 1

1 np n=2 )

1 1 dx = p x p1

1 N p1

The most useful series for comparison is the geometric series dened by a real number r, with 0 < r < 1. Theorem 1.9.17 (Ratio Test) Suppose that n=1 an is a series of non-zero complex numbers. If r = limn |an+1 /an | exists, then the series converges absolutely if r < 1, and the series diverges if r > 1. Proof. Suppose limn |an+1 /an | = r < 1. If satises r < < 1, then there exists N N such that |an+1 |/|an | < for all n N . Consequently, |an | |aN |nN for all n N . The result follows from the Comparison Test. The second half of the theorem follows from Theorem ??.

46

Exercise 1.9.18 Give examples to show that if r = 1 in the statement of the Ratio Test, anything may happen. Our nal test for convergence is called the root test. This can be quite eective when the comparison test and ratio test fail. Theorem 1.9.19 (Root Test) Suppose that n=1 an is a series of complex numbers. Let r = lim supn |an |1/n . If r < 1, then the series converges absolutely. If r > 1, then the series diverges. Proof. Suppose that lim supn |an |1/n = r < 1. Pick such that that < r < 1. Then, there exists N N such that |an | n for all n N . The convergence of the series now follows from the comparison test. The second half of the theorem is left as an exercise.

Exercise 1.9.20 Give examples to show that if r = 1 in the statement of the Root Test, anything may happen. Exercise 1.9.21 Suppose that the ratio test applies to a series. That is, limn |an+1 |/|an | = r. Show that lim supn |an |1/n = r. Denition 1.9.22 Let z0 be a xed complex number. A complex power series around z0 is a series of the form n=0 an (z z0 )n , where the coecients an are in C for all n N. When this series converges, it converges to a function of the complex variable z . Exercise 1.9.23 Show that if the series converges absolutely for a complex number z then it also converges for a any complex number w such that |w z0 | |z z0 |, that is the series converges on the disk {w C | |w z0 | |z z0 |}. From this exercise, it follows that a complex power series around z0 that converges absolutely at any point other then z0 will have a disk of convergence of the form {z C | |z z0 | < r}. The supremum of all such r is called the radius of convergence of the power series. To determine the radius of convergence for a complex power series we use the convergence tests developed above, in particular the root test. Theorem 1.9.24 Suppose that lim supn |an |1/n = r. If r > 0, then the power series n=0 an (z z0 )n has a radius of convergence 1/r. If the number r = 0, the we say that the radius of convergence is innity, 1 and if the lim sup does not exist because |an | n is unbounded (r = ), we say that the radius of convergence is 0. Examples 1.9.25
n 1/n i. Consider the series = 1, and the power series converges absolutely n=0 n(z z0 ) . Then limn n for |z z0 | < 1, that is, the radius of convergence is 1.

ii. Consider the series n=1 nn (z z0 )n . Then limn (nn )1/n = , so the radius of convergence is 0 and the series converges only for z = z0 . Exercise 1.9.26 Determine the radius of convergence of the following power series: i. zn ; n! n=1 zn ; ln(n) n=2 nn n z . n! n=1 47

ii.

iii.

1.10

Algebraic notions in R and C

In this chapter, we have constructed the real numbers and showed that they are the unique ordered eld, up to order isomorphism, that satises the least upper bound property. We have also constructed the complex numbers, a eld containing a subeld that is isomorphic to the real numbers. Moreover, we have seen that the complex numbers cannot be ordered. While the question of when two elds are isomorphic is fundamental, we now pursue the more rened question of studying the ways in which a eld is isomorphic to itself. Let us start simply. Consider the eld of rational numbers and a function f : Q Q such that f (x + y ) = f (x) + f (y ). What can we say about f ? We have f (x + 0) = f (x) + f (0) = f (x), so f (0) = 0. Next we have, for n N, f (n) = f (1 + 1 + + 1) = f (1) + f (1) + + f (1) = nf (1).
n times n times

A similar argument shows that f (m/n) = (m/n)f (1) for any positive rational number m/n. Also, for any positive rational number r, we have 0 = f (0) = f (r + (r)) = f (r) + f (r), so that f (r) = f (r). Thus, f (r) = rf (1) for all r Q. Now suppose that we also want f (xy ) = f (x)f (y ) for x, y Q, that is, we want f to be an isomorphism of elds. Then, f (1) = f (1 1) = f (1)f (1), and if f is to be injective, we must have f (1) = 1. Thus, the function f can be none other than the identity function on Q (see Denition ??). What we have just done in the preceding paragraph is to show that the only eld isomorphism from Q to itself is the identity. In general, an isomorphism from a eld to itself is called an automorphism, a notion which we make precise with the following denition. Denition 1.10.1 Let F be a eld. An automorphism of F is a bijection, : F F , such that a. (x + y ) = (x) + (y ) for all x, y F , b. (xy ) = (x)(y ) for all x, y F . We denote the set of automorphisms of a eld F by Aut(F ). Exercise 1.10.2 Show that the set Aut(F ) of automorphisms of a eld F has the structure of a group under function composition (see Project 2.1). Lets see what happens for R. Let f : R R be an automorphism. By the exact same reasoning as for Q, we see that f (1) = 1, and in fact, f (r) = r for all r Q. The next thing to note here is that if a R and a = 0, then a2 > 0 and f (a2 ) = (f (a))2 , so f (a2 ) > 0. Since all positive real numbers have unique positive square roots, we can conclude that if c > 0, then f (c) > 0. Thus, if a < b, then f (a) < f (b) since b a > 0. Now take any real number c. If c Q, then f (c) = c. If c Q and f (c) = c, then there are two possibilities. If c < f (c), choose a rational number r such that c < r < f (c). Then f (c) < f (r) = r, which is a contradiction. If f (c) < c, we run into the same problem. So we conclude that f (c) = c for all c R. Theorem 1.10.3 The groups Aut(Q) and Aut(R) consist only of the identity. Exercise 1.10.4 Find a eld F such that Aut(F ) = {1}. Exercise 1.10.5 Find nontrivial elements of Aut(C). Exercise 1.10.6 i. Let F be a eld and let be an element of Aut(F ). Dene H = {x F | (x) = x}. Show that H is a subeld of F . ii. Suppose that F is a eld and that Q is a subeld of F . If Aut(F ), show that Q is a subeld of H . Exercise 1.10.7 48

i. Find Aut(Zp ) where p is a prime and Zp is the nite eld with p elements. ii. Let F = {a + b 2 | a, b Q}. Show that F is a eld and nd Aut(F ). This is the beginning of the subject called Galois theory, in which one of the goals is to determine Aut(F ) when F is a so called algebraic extension of Q. More generally, an algebraic extension of Q is a subeld of C all of whose elements are roots of polynomials with coecients in Q (see Denition ??). If R is a commutative ring with 1, we write R[x] for the collection of polynomials in the variable x with coecients in R. We can add and multiply polynomials in the usual manner, and this makes R[x] into a commutative ring with 1. Exercise 1.10.8 Show that Z[x], Q[x], R[x], and C[z ] are integral domains. Determine the elements in each of these domains which have multiplicative inverses. Denition 1.10.9 Let F be a eld. We say that F is algebraically closed if every nonconstant polynomial in F [x] has a root in F . That is, F is algebraically closed if, for every nonconstant p(x) F [x], there is an element r F such that p(r) = 0. The most important example of an algebraically closed eld is supplied by the Fundamental Theorem of Algebra, which states that the eld of complex numbers is algebraically closed. There is a semi-innite number of proofs of this theorem. We will present one of these in Project 3.10.4 using the properties of continuous functions developed in Chapter 3. Exercise 1.10.10 Let F be a eld and suppose that p(x) F [x]. Show that r is a root of p(x) if and only if (x r) is a factor of p(x). That is, we can write p(x) = (x r)q (x) for some q (x) F [x]. (Hint: Consider the division algorithm for polynomials.) Denition 1.10.11 Let A be the collection of all roots of polynomials in Z[x]. A is called the set of algebraic numbers in C. The set AR = A R is called the set of real algebraic numbers. A real number which is not a real algebraic number is called transcendental. Example 1.10.12 Among the more famous algebraic numbers are i and i. For real algebraic numbers, the most famous one is probably 2. The most famous transcendental numbers are and e. Exercise 1.10.13 Show that A and AR are elds. Exercise 1.10.14 Show that the eld A of algebraic numbers is countable. Remark 1.10.15 It follows from the exercise above that the eld AR of real algebraic numbers is countable and hence the set of transcendental numbers is uncountable. Exercise 1.10.16 Find nontrivial elements of Aut(AR ). Exercise 1.10.17 Find nontrivial elements of Aut(A) that are not on your list from the previous problem.

1.11
1.11.1

Independent Projects
Another construction of R

Denition 1.11.1 A subset of Q is said to be a cut (or a Dedekind cut ) if it satises the following: a. the set = and = Q; b. if r and s Q satises s < r, then s ; 49

c. if r , then there exists s Q with s > r and s . Let R denote the collection of all cuts. Denition 1.11.2 For , R, we dene + = {r + s | r and s }. Let 0 = {r Q | r < 0}. Exercise 1.11.3 If and are cuts, show that + is a cut, and also show that 0 is a cut. Exercise 1.11.4 Show that, with this addition, (R,+) is an abelian group with 0 as the identity element. We now dene an order on R. Denition 1.11.5 If , R, we say that < if is a proper subset of . Exercise 1.11.6 Show that the relation < satises the following properties: 1. if , R, then one and only one of the following holds: < , = , or < (Trichotomy); 2. if , , R with < and < , then < (Transitivity); 3. if , , R with < , then + < + (Additivity). It is now possible to dene the notions of bounded above, bounded below, bounded, upper bound, least upper bound, lower bound, and greatest lower bound, in Rm just as we did earlier in this Chapter. Exercise 1.11.7 Show that the least upper bound property holds in R, that is, if A is a nonempty subset of R which is bounded above, then A has a least upper bound in R. Next, we must dene multiplication in R. Denition 1.11.8 If , R with , > 0, then = {p Q | there are positive elements r and s such that p rs}. The next step is multiplication by 0, which is exactly as it should be, namely, for any R, we dene 0 = 0. If < 0 or < 0, or both, replace any negative element by its additive inverse and use the multiplication of positive elements to dene multiplication accordingly. For example, if < 0 and > 0, = [()( )]. Exercise 1.11.9 Show that R with addition, multiplication, and order as dened above is an ordered eld. (Hint: think carefully about how to dene the multiplicative inverse of a nonzero cut.) Exercise 1.11.10 Put it all together and show that R is an Archimedean ordered eld in which the least upper bound property holds.

50

1.11.2 Decimal Expansions of Real numbers In Chapter 0, we used a decimal representation of the real numbers to show that the real numbers between 0 and 1 form an uncountable set. In this project, we actually prove that every real number between 0 and 1 has a unique decimal expansion that does not terminate in all 9s. In addition, we discuss the fact that rational numbers have decimal expansions of three dierent types. The rst is those rational numbers whose denominators are are divisors of a power of 10, the second is those whose denominators are relatively prime to 10, and the third is an intermediate case. Since we know every real number lies between two consecutive integers (see Theorem 1.2.1), we start with a real number x such that 0 < x < 1. Let D = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9}. Assume rst that x is irrational. The construction proceeds as follows. Let a1 be the largest element of D which is less then 10x. Then 0 < x a1 /10 < 1/10. Let a2 be the largest integer in D less then 100x 10a1 . Proceeding as before we get 0 < x a1 /10 a2 /102 < 1/102. Continuing this process, we obtain a monotonic increasing sequence Sn = a1 /10 + a2 /102 + + an /10n , where aj D and 0 < x Sn < 1/10n. So we conclude that Sn an converges to x, and we get x = a1 /10 + a2 /102 + + an /10n + . . . = n=1 10 n . We call 0.a1 a2 . . . the decimal expansion of x. Exercise 1.11.11 Let x be a irrational number between 0 and 1. Show that the decimal expansion x is unique. We now turn to rational numbers between 0 and 1. We can apply the above procedure to rational numbers but with the possibility of equality in any of the inequalities above. Suppose that x has a terminating decimal expansion that is, there exists N so that an = 0 for all n > N and aN = 0. Then we can write x = a1 /10 + a2 /102 + + aN /10N . Exercise 1.11.12 i. Show that if r is a rational number in (0, 1), then the decimal expansion of r terminates if and only if the denominator of r has the form 2a 5b . where a and b are non-negative integers. ii. With r as above show that the last non-zero digit of r is in the m-th place where m = max{a, b}. Note that rational numbers with terminating decimal expansions are the only real numbers between 0 and 1 for which equality can occur in initial procedure. Next, consider a rational number r = p/q in (0, 1) for which q is relatively prime to 10. From Exercise ??.25, q divides 10(q) 1. Let n be the smallest natural number such that q divides 10n 1. Then (p/q )(10n 1) is an integer, which we denote by m. That is, m= p m p (10n 1) or = n . q q 10 1

Using results about geometric series from Section 1.9, we can now write m m m p = n = n (1 10n )1 = n (1 + 10n + 102n + . . .) = m/10n + m/102n + . . . . q 10 1 10 10 As 0 < p/q < 1 we have m < 10n . Thus the right hand side of the equation above gives us a periodic decimal expansion of p/q whose period has length at most n. Exercise 1.11.13 Prove that the period is exactly n. Exercise 1.11.14 Let p/q be a rational number between 0 and 1. If q and 10 are relatively prime, show that p/q has a unique periodic decimal expansion with the length of the period equal to the order of 10 mod q , that is, the smallest power of 10 that is congruent to 1 mod q . We now present the remaining case as an exercise. 51

Exercise 1.11.15 Let p/q be a rational number in (0, 1) with q = 2a 5b r, where r is relatively prime to 10. Let k = max{a, b}, and let n be the smallest positive integer such that r divides 10n 1. Show that after k digits, the decimal expansion of p/q is periodic of length n. We ask nally whether decimal expansions are unique. The answer is contained in the following exercise. Exercise 1.11.16 i. Consider the decimal 0.9999 . . . = converges to 1. That is, 0.9999 . . . = 1.
9 n=1 10n .

Show that this geometric series

ii. Show that every number that has a decimal expansion that ends in repeating nines can be written as a terminating decimal. iii. Show that the exceptional situation in part ?? is the only non-uniqueness that can occur.

52

Chapter 2

Linear Algebra
The presentation of the material will perhaps incur the reproach of being too dogmatic in its methods. To this possible objection, we would like to make two answers. Firstly, that what the student may learn here is not designed to help him with problems he has already met but with those he will have to cope with in the future; it is therefore impossible to motivate the denitions and theorems by applications of which the reader does not know the existence as yet. Secondly, that one of the important pedagogical problems which a teacher of beginners in mathematics has to solve is to impart to his students the technique of rigorous mathematical reasoning; this is an exercise in rectitude of thought, of which it would be futile to disguise the austerity. Claude Chevalley, Fundamental Concepts of Algebra

This is a book on analysis. However, almost all of the structures that we deal with in analysis have an underlying algebraic component, and an understanding of this algebraic component makes it a lot easier to discuss the analysis. Our approach is more user friendly than Chevalleys. We will nd that analysis is mostly about inequalities, while algebra is mostly about equalities. The fundamental algebraic ideas we discuss in this chapter concern vector spaces and linear algebra. Other algebraic structures that play a role in analysis include groups, rings, and elds. Some ideas about these were discussed in Chapters 0 and ??. More can be found in the Projects at the end of this chapter.

2.1

Fundamentals of Linear Algebra

The algebraic structures that are really fundamental for analysis are vector spaces (sometimes called linear spaces). Recall that a eld has been dened in Denition 0.2.22. Denition 2.1.1 Let F be a eld. A vector space over F is a triple (V, +, ) where (V, +) satises the axioms (A1)-(A5) of Chapter 0 and is a map from F V to V satisfying the following properties: a. if F and v V , then v V ; b. if F and v1 , v2 V , then (v1 + v2 ) = ( v1 ) + ( v2 ); c. if , F and v V , then ( + ) v = ( v) + ( v); d. if , F and v V , then ( ) v = ( v); 53

e. if 1 is the multiplicative identity in F and v V , then 1 v = v. The function (, v) v is called scalar multiplication and the elements of F are called scalars . We frequently suppress the dot (). For completeness, we restate axioms (A1)-(A5). (A1) If v1 , v2 V , then v1 + v2 V . (A2) If v1 , v2 , v3 V , then v1 + (v2 + v3 ) = (v1 + v2 ) + v3 . (A3) If v1 , v2 V , then v1 + v2 = v2 + v1 . (A4) There exists 0 V , such that for all v V , v + 0 = 0 + v = v. (A5) For every v V , there exists v V such that v + (v) = v + v = 0. Any structure satisfying (A1) (A5) above is called an abelian group (See Project 2.1). Hence, a vector space is an abelian group with scalar multiplication. Exercise 2.1.2 i. If 0 is the additive identity in F and 0 is the additive identity in V , show that 0 v = 0 for any v V . Note that this statement actually says something. It says that the additive identity in F has a property for scalar multiplication similar to multiplication by 0 in commutative rings. However, in this case, multiplication by the 0 in F gives 0 in V . ii. Show that condition e. does not follow from the other axioms. Examples 2.1.3 1. If F is a eld, V = {0} is a vector space over F . 2. If F is a eld, then F is a vector space over F with scalar multiplication being ordinary multiplication in F . 3. Let F be eld and let F n = {(x1 , x2 , . . . , xn ) | xj F, j = 1, 2, . . . , n}. Addition in F n is dened coordinatewise, that is, if x = (x1 , x2 , . . . , xn ) and y = (y1 , y2 , . . . , yn ), then x + y = (x1 + y1 , x2 + y2 , . . . , xn + yn ). If F and x = (x1 , x2 , . . . , xn ) F n , we set x = (x1 , x2 , . . . , xn ). Then F n is a vector space over F . 4. The ring of polynomial functions in one variable, F [x], forms a vector space over F . 5. Let X be a nonempty set, let F be a eld, and let V = F (X, F ) be the set of all functions from X to F . For f, g V , dene (f + g )(x) = f (x) + g (x), and for F dene (f )(x) = f (x). Then V is a vector space over F . 6. The real numbers R form a vector space over Q. Exercise 2.1.4 Check that the above examples satisfy the axioms for a vector space. Remark 2.1.5 What is a vector? Thats easy to answer. A vector is an element of a vector space. Lots of people describe a vector as a quantity having magnitude and direction. This is not particularly useful in most contexts (see, for example, 5 above) . However, in this chapter, when we do the geometric interpretation of vectors in n-dimensional Euclidean space, it will be helpful to think of vectors this way. Exercise 2.1.6 Consider the collection R[x] of all polynomial functions with coecients in R. Show that R[x] is a vector space over R and also over Q. 54

There are three fundamental notions which need to be discussed right at the beginning of our treatment of vector spaces. These are linear combinations, linear independence, and linear dependence. These notions are the heart and soul of elementary vector space theory. Denition 2.1.7 Let V be a vector space over a eld F , and let v1 , v2 , . . . , vm V . Then a vector of the form 1 v1 + 2 v2 + + m vm where 1 , 2 , . . . , m F is called a linear combination of v1 , v1 , . . . , vm . Remark 2.1.8 Note that the denition of linear combination refers to a sum of a nite number of vectors from the vector space V . Denition 2.1.9 Let V be a vector space over a eld F and v1 , v2 , . . . , vm be nonzero vectors in V . We say that the set {v1 , v2 , . . . , vm } is a linearly independent set if, for any scalars 1 , . . . , m F such that 1 v1 + 2 v2 + + m vm = 0, we have 1 = 2 = = m = 0. Example 2.1.10 A set containing a single nonzero vector forms a linearly independent set. Exercise 2.1.11 In the vector space F n over a eld F , set e1 = (1, 0, . . . , 0), e2 = (0, 1, 0, . . . , 0), and generally, ej = (0, 0, . . . , 1, . . . , 0), where the 1 is in the j th coordinate and 0 is in the other coordinates. Show that, for any k , 1 k n, the set {e1 , e2 , . . . , ek } is a linearly independent set. Exercise 2.1.12 Let {v1 , . . . , vm } be a linearly independent set in a vector space V . Show that any nonempty subset of this set is linearly independent. Remark 2.1.13 An innite set of vectors is said to be linearly independent if each nite subset is linearly independent. Exercise 2.1.14 Show that the set {1, x, x2 , . . . , xn , . . . } is a linearly independent set in Q[x]. Denition 2.1.15 Let V be a vector space over a eld F , and let v1 , v2 , . . . , vm be vectors in V . The set {v1 , v2 , . . . , vm } is a linearly dependent set if there exist scalars 1 , 2 , . . . , m F , not all zero, such that 1 v1 + 2 v2 + + m vm = 0, that is, a set is linearly dependent if it is not linearly independent. Exercise 2.1.16 i. Let V = F n and let v be any vector in V . Show that the set {e1 , e2 , . . . , en , v} is a linearly dependent set in V . ii. Show that, if {v1 , . . . , vm } is a set of vectors in a vector space V and one of these vectors is the zero vector, then the set is a linearly dependent set. iii. Let v1 and v2 be vectors in a vector space V . Show that the set {v1 , v2 } is a linearly dependent set i one of these vectors is a scalar multiple of the other. Lemma 2.1.17 Suppose that {v1 , . . . , vm } is a linearly dependent set of nonzero vectors in a vector space V over a eld F . Then there exists k , 1 < k m, such that vk is a linear combination of v1 , . . . , vk1 . 55

Proof. Note that {v1 } is a linearly independent set because v1 is a nonzero vector. Let k be the largest integer such that {v1 , . . . , vk1 } is a linearly independent set. Observe that k m because {v1 , . . . , vm } is a linearly dependent set. Then {v1 , . . . , vk } is a linearly dependent set, so there exist scalars 1 , . . . , k such that 1 v1 + + k vk = 0, and k = 0. Then vk = so the integer k gives the desired result. There are two additional ideas that are closely related. One is that of a spanning set, and the second is that of a basis. Denition 2.1.18 Let V be a vector space over a eld F , and let S be a set of vectors in V . The set S is a spanning set for V if every element of V can be written as a (nite) linear combination of the vectors in S . We say that the set S spans V . Example 2.1.19 Let F be a eld and let V = F n . The set S = {e1 , e2 , . . . , en } as dened in Exercise 2.1.11 spans V . In particular, if x = (x1 , x2 , . . . , xn ), then x = x1 e1 + x2 e2 + + xn en . Exercise 2.1.20 Let V = R3 over the eld F = R. Show that the set S = {v1 = (1, 2, 3), v2 = (4, 5, 6), v3 = (7, 8, 9)} does not span V by nding a vector v V that cannot be written as a linear combination of v1 , v2 , and v3 . Example 2.1.21 Let V = Q[x] be the vector space of polynomials dened over the eld F = Q. The collection of monomoials S = {1, x, x2 , . . . , xn , . . .} is a spanning set for V because every element in Q[x] can be written as a nite linear combination of these monomials. Note that no proper subset of S has this property. In fact, no nite set of vectors spans V . Exercise 2.1.22 Show that the vector space Q[x] is a countable set. Exercise 2.1.23 Let V be a vector space with a spanning set S . Suppose T is a set of vectors containing S . Show that T also spans V . The second of these big ideas is that of a basis, which combines the aspects of linear independent sets of vectors and spanning sets of vectors. Denition 2.1.24 A spanning set S is a basis for V if S is linearly independent. Example 2.1.25 Let F be a eld and let V = F n . The spanning set S = {e1 , e2 , . . . , en } is a basis for V . This is called the standard basis, or the canonical basis for V . Example 2.1.26 Let V = Q[x] over the eld F = Q. The spanning set S = {1, x, x2 , . . . , xn , . . .} is a basis for V . Exercise 2.1.27 Let V be a vector space over the eld F , and let S be a basis of V . Show that every vector v V can be written uniquely as a linear combination of vectors in S . Exercise 2.1.28 Show that no proper subset of a linearly independent set can be a spanning set. What we now prove is that if a vector space has a basis with a nite number of elements, then all bases have the same number of elements. This will allow us to dene the dimension of a vector space as the cardinality of a basis. The following lemma plays a crucial role. 56 1 2 k1 v1 vk1 , k v2 k k

Lemma 2.1.29 (Exchange Lemma) Suppose that V is a vector space over a eld F and that S = {u1 , . . . , um } is a spanning set for V . If the set T = {v1 , . . . , vn } is a linearly independent set in V , then n m. Proof. The idea of the proof is to replace the elements of S one-by-one with elements of T in such a way that at each stage, we still have a spanning set for V . If n > m then we will have a proper subset of T that spans V , which is impossible by Exercise 2.1.28. First, consider the collection S1 = {v1 } {u1 , . . . , um }. Since S spans V , the set S1 also spans V by Exercise 2.1.23. Either v1 = uj1 for some j1 , or S1 contains S as a proper subset, and hence is linearly dependent by Exercise 2.1.28. In the latter case, according to Lemma 2.1.17 there exists an element uj1 S1 which is a linear combination of v1 , u1 , . . . , uj1 1 . In either case, we dene the set S1 = {v1 , u1 , u2 , . . . , uj1 1 , uj1 +1 , . . . , um }. This set still spans V because S1 did, and uj1 was itself a linear combination of the elements of S1 . Let us iterate this procedure. Dene S2 = {v2 } {v1 , u1 , . . . , uj1 1 , uj1 +1 , . . . , um }. By the same reasoning, S2 is a spanning set. So, proceeding as above, we can nd j2 such that the set
S2 = {v2 , v1 } ({u1 , u2 , . . . , um } \ {uj1 , uj2 })

is a spanning set. We can continue by putting v3 , v4 , . . . at the beginning of our list, and each time we do that, we can eliminate an element of U that remains on our list. Our procedure (that is, using Lemma 2.1.17) will never eliminate one of the vs since they are linearly independent. So at each stage, we eliminate one of the us and are left with a spanning set. If at some point, all the us are gone and some vs are left, then a proper subset of the vs would be a spanning set. This contradicts the linear independence of the vs.

Exercise 2.1.30 Suppose that V is a vector space over a eld F and that S = {u1 , . . . , um } is a spanning set for V . If T is any linearly independent subset of V , then T has at most m elements. Now, we assume that our vector space V over F has a nite subset that spans. Such a vector space is called nite dimensional . The next corollary proves the existence of a basis for a nite dimensional vector space. Lemma 2.1.31 Let V be a nonzero nite dimensional vector space over a eld F . Then V has a nite basis. Proof. Let {v1 , . . . , vm } be a spanning set consisting of nonzero vectors. If {v1 , . . . , vm } is linearly dependent, then by Lemma 2.1.17 there exists an integer k such that {v1 , . . . , vk1 } is a linearly independent set and {v1 , . . . , vk } is linearly dependent. Eliminating vk from the set {v1 , . . . , vm }, we still have a spanning set. Continue this process (a nite number of times). This yields a linearly independent set that spans V , that is, a basis.

Exercise 2.1.32 Let V be a nonzero nite dimensional vector space over a eld F . Show that any basis of V has a nite number of elements. Now we get to the heart of the matter. Theorem 2.1.33 If V is a nite dimensional vector space over a eld F , then any two bases of V have the same number of elements. 57

Proof. Suppose {v1 , . . . , vn } and {u1 , . . . , um } are bases for V . Then by the Exchange Lemma, since the rst set spans and the second set is linearly independent, we must have m n. Similarly, since the second set spans and the rst set is linearly independent, we must have n m.

Now we can talk about an n-dimensional vector space over a eld F . Denition 2.1.34 Suppose V is a vector space containing a spanning set of n linearly independent vectors, n 1. The dimension of V , denoted dim V , is equal to n, that is dim V = n. If V = {0}, we set dim V = 0. Theorem 2.1.35 Suppose that V is an n-dimensional vector space over a eld F and that {v1 , . . . , vm } is a linearly independent set in V . Then m n and there exist vectors vm+1 , . . . , vn such that {v1 , . . . , vm , vm+1 , . . . , vn } is a basis for V . Proof. Let {u1 , . . . , un } be a basis for V . Applying the procedure in the proof of the Exchange Lemma

m times, we get a basis {vm , vm1 , . . . , v1 , ui1 , . . . , uinm }.

Remark 2.1.36 Informally, the above theorem says that we can take any linearly independent set, and extend it to a basis. Denition 2.1.37 If V is a vector space over a eld F , then a nonempty subset W V is a subspace if it is closed under addition and scalar multiplication. That is, if v, w W , then v + w W , and if v W and F , then v W . Exercise 2.1.38 i. Let V be a vector space over a eld F , show that {0} and V are subspaces of V . ii. When is it true that the only subspaces of V are {0} and V ? Denition 2.1.39 Let V be a vector space over a eld F . Let S be any nonempty subset of V . We dene the span of S to be the set Span(S ) of all linear combinations of elements of S . That is, Span(S ) = {v V | v = 1 v1 + 2 v2 + + m vm for some 1 , 2 , . . . , m F, v1 , v2 , . . . , vm S }. Additionally, we dene Span() to be the set {0}. Exercise 2.1.40 Let S be any subset of V . Show that Span(S ) is a subspace of V . Examples 2.1.41 i. Let V = Q[x], and let W be the collection of all polynomials in Q[x] whose degree is less than or equal to a xed non-negative integer n. Then W is a subspace of Q[x] ii. Let V = F n , and for a xed m n, let W = {v V | v = 1 e1 + 2 e2 + + m em , j F }. Then W is a subspace of V . Exercise 2.1.42 Find the dimension of W in the last two examples. Theorem 2.1.43 Let V be a nite dimensional vector space over a eld F . Suppose that W is a subspace of V . Then dim W dim V . 58

Proof. Suppose that dim V = n. Consider the set S of all positive integers that are the cardinalities of linearly independent sets in W . The set S is bounded above by n by Theorem 2.1.35. Let h be the largest element of S . Let B = {w1 , . . . , wh } be a linearly independent set in W of cardinality h. Then B must be a basis for W . Otherwise, there would be an element w W not in the span of B . Then h would not be maximal since B {w} would be linearly independent. So, dim W = h n. Most of the vector spaces which arise in analysis are not nite dimensional and thus are called innite dimensional . We will often be dealing with vector spaces of functions whose domain and range are nite dimensional vector spaces over R (or over the complex numbers C), but the spaces of functions themselves will ordinarily be innite dimensional spaces. To prove the existence of a basis for an innite dimensional space, Zorns Lemma or some other equivalent statement is needed. Exercise 2.1.44 Suppose X is a set. i. If V = F (X, F ), the set of functions from X to F , show that V is nite dimensional if and only if X is nite. ii. If X is a nite set, nd an explicit basis for V = F (X, F ). iii. Fix a subset A X and dene W (A) = {f V | f (a) = 0 a A}. Show that W (A) is a subspace of V. iv. Can you nd an innite set X and a eld F where you can write an explicit basis for V = F (X, F )? Theorem 2.1.45 Let V be a nonzero vector space over a eld F . Then V has a basis. That is, there exists a linearly independent subset B of V such that each element of V is a nite linear combination of elements of B . Proof. Let X be the collection of linearly independent sets in V . This collection can be partially ordered by inclusion. We apply Zorns Lemma to X Given any totally ordered subset of X , the union of the elements in this subset provides a maximal element in X . The conclusion of Zorns Lemma says that this maximal

element gives a basis for V .

Example 2.1.46 This last theorem means that there exists a basis B = {vi | i I } for R considered as a vector space over Q. In particular, every real number can be written as a nite linear combination of elements of this basis with coecients taken from Q. The basis is not countable. It is called a Hamel basis for R over Q.

2.2

Linear Transformations

One of the most important topics in the subject of linear algebra is the study of maps from one vector space over a eld F to another vector space over F that preserve addition and scalar multiplication. Such maps are called linear transformations and they play a vital role throughout the remainder of this text. Denition 2.2.1 Let V , W be vector spaces over a eld F . A function T : V W is called a linear transformation, linear map, or linear operator if a. T (v1 + v2 ) = T (v1 ) + T (v2 ) for all v1 , v2 V , b. T (v) = T (v) for all F and v V . Exercise 2.2.2 Let V and W be vector spaces over a eld F and T : V W a linear transformation. Show that T (0) = 0 and T (v) = T (v) for all v V . 59

Denition 2.2.3 Let V and W be vector spaces over a eld F and T : V W a linear transformation. The image of T is the set T (V ) = {w W | w = T (v) for some v V }. Exercise 2.2.4 i. Show that T (V ) is a subspace of W . ii. If T is an injection, show that T 1 : T (V ) V is a linear operator. Example 2.2.5 Consider V = F as a vector space over F and x a F . Dene Ta (x) = ax for x V . Then Ta is a linear transformation on F . Exercise 2.2.6 Consider R as a vector space over itself and x a, b R. Dene Ta,b (x) = ax + b. Show that Ta,b is a linear transformation if and only if b = 0. Example 2.2.7 Let V = F n and W = F considered as vector spaces over F . For each i {1, 2, . . . , n}, let Pi : V W be the map given by Pi (x1 , x2 , . . . , xi , . . . , xn ) = xi . This map is a linear transformation called the ith coordinate projection. Exercise 2.2.8 Let V be a vector space over a eld F , and let W be a eld V . Show that T : W V dened by T (w) = w is a linear transformation. Exercise 2.2.9 Let V be a vector space over a eld F , and let B = {v1 , v2 , . . . , vn } be a basis for V . Fix m n. Show that the function T : V V dened by T (1 v1 + 2 v2 + + m vm + + n vn ) = 1 v1 + 2 v2 + + m vm is a linear transformation. Example 2.2.10 Let V = R[x] considered as a vector space over R. Dene D : R[x] R[x] by [D(p)](x) = p (x), that is, the derivative of the polynomial p. It follows from the properties of the derivative that D is a linear transformation. Exercise 2.2.11 Show that D : R[x] R[x] as dened in the previous example is surjective but not injective. If T : V W is a linear transformation, how are the dimensions of V , W , and T (V ) related? Our rst result along these lines says that the dimension of T (V ) is less than or equal to that of V . Informally, we might say that a linear transformation cannot increase the dimension of a vector space. Theorem 2.2.12 Suppose that V and W are vector spaces over a eld F and that T : V W is a linear transformation. If V is nite dimensional with dim V = n, then dim T (V ) n. Proof. It suces to show that every subset of T (V ) consisting of n + 1 elements is linearly dependent. Let w1 , w2 , . . . , wn+1 be vectors in T (V ). Pick v1 , v2 , . . . , vn+1 V such that T (vj ) = wj for j = 1, . . . , n + 1. Since dim V = n, the set {v1 , . . . , vn+1 } is linearly dependent, so there exist scalars 1 , . . . , n+1 , not all zero, such that 1 v1 + + n+1 vn+1 = 0. It follows that 1 w1 + + n+1 wn+1 = T (1 v1 + + n+1 vn+1 ) =

T (0) = 0. Hence the set {w1 , w2 , . . . , wn+1 } is linearly dependent.

Denition 2.2.13 Let V and W be vector spaces over a eld F and T : V W a linear transformation. The transformation T is called a linear isomorphism if T is a bijection. In this case V and W are said to be linearly isomorphic. Corollary 2.2.14 Suppose that V and W are nite dimensional vector spaces over a eld F . If V and W are linearly isomorphic, then dim V = dim W. Moreover, if {v1 , . . . , vn } is a basis for V , then {T (v1 ), . . . , T (vn )} is a basis for W. 60

Exercise 2.2.15 Suppose that V and W are nite dimensional vector spaces over a eld F such that dim V = dim W . Show that V and W are linearly isomorphic. In general, if T is a linear transformation from V to W , T is neither injective nor surjective. This leads to an important idea, the kernel of a linear transformation. Denition 2.2.16 Let V and W be vector spaces over a eld F , and T : V W a linear transformation. The kernel of T is dened by ker T = {v V | T v = 0}. Exercise 2.2.17 i. Show that ker T is a subspace of V . ii. Show that T is injective if and only if ker T = {0}. iii. Let D : R[x] R[x] be dened as above, that is, [D(p)](x) = p (x). Find ker D. iv. Let Pi : F n F be the ith coordinate projection as dened in Example 2.2.7. Find ker Pi . The notions of kernel and image allow us to give a more precise answer to the question we asked earlier about the relation between the dimensions of V , W , and T (V ). Theorem 2.2.18 Suppose that V and W are vector spaces over a eld F and that dim V is nite. Let T : V W be a linear transformation. Then dim V = dim ker T + dim T (V ). Proof. Since ker T is a subspace of V , we know that k = dim ker T n. Let v1 , . . . , vk be a basis for ker T . We can extend this to a basis v1 , . . . , vk , vk+1 , . . . , vn for V by Theorem 2.1.35. We claim that {T (vk+1 ), . . . , T (vn )} is a basis for T (V ). The equation in the statement of the theorem is now obvious once we verify the claim. Let w T (V ). Then there exists v V such that T (v) = w. Since {v1 , v2 , . . . , vn } is a basis for V , there exist scalars 1 , 2 , . . . , n F such that v = 1 v1 + 2 v2 + + n vn . Then w = T (v) = T (1 v1 + 2 v2 + + n vn ) = 1 T (v1 ) + 2 T (v2 ) + + n T (vn ) = k+1 T (vk+1 ) + k+2 T (vk+2 ) + + n T (vn ). Hence the set {T (vk+1 ), T (vk+2 ), . . . , T (vn )} spans T (V ). To show that this set of vectors in linearly independent, suppose that k+1 T (vk+1 )+ k+2 T (vk+2 )+ + n T (vn ) = 0. Then T (k+1 vk+1 + k+2 vk+2 + + n vn ) = 0, hence k+1 vk+1 + k+2 vk+2 + + n vn ker T . Hence, there exist scalars 1 , 2 , . . . , k F such that k+1 vk+1 + k+2 vk+2 + + n vn = 1 v1 + 2 v2 + + k vk , or 1 v1 + 2 v2 + + k vk k+1 vk+1 k+2 vk+2 n vn = 0. Since {v1 , v2 , . . . , vn } is a basis for V , it is a linearly independent set, all of the coecients, including k+1 , k+2 , . . . , n , must be zero. Hence the set {T (vk+1 ), T (vk+2 ), . . . , T (vn )} is linearly independent, and thus a basis for T (V ). Remark 2.2.19 Other authors refer to the preceding theorem as the Rank-Nullity Theorem. This is because the rank of T is dened to be dim T (V ) and the nullity of T is dened to be dim ker T . Exercise 2.2.20 Let V and W be nite dimensional vector spaces over a eld F with dim V = dim W . Let T : V W be a linear transformation. Show that the following are equivalent. 1. T is bijective. 2. T is surjective. 3. T is injective. Denition 2.2.21 Let V and W be vector spaces over a eld F . If T , T1 , and T2 are linear transformations from V to W , we dene a. (T1 + T2 )(v) = T1 (v) + T2 (v), for v V , and 61

b. (T )(v) = T (v), for F . Theorem 2.2.22 Let V and W be vector spaces over a eld F . Let L (V, W ) denote the set of all linear transformations from V to W . Then, with the above operations, L (V, W ) is a vector space over F . Proof. Clear. Exercise 2.2.23 Show that if dim V = n and dim W = m, then L (V, W ) is a nite dimensional vector space with dim L (V, W ) = nm. The proof of this exercise is facilitated by the use of bases in V and W . This will lead us to the notion of matrices representing linear transformations in the next section. Finally, we consider the composition of linear transformations. Exercise 2.2.24 Let U , V , and W be vector spaces over a eld F . Let S, S1 , S2 L(U, V ) and T, T1 , T2 L(V, W ). i. Show that T S L(U, W ). ii. Show that T (S1 + S2 ) = (T S1 ) + (T S2 ). iii. Show that (T1 + T2 ) S = T1 S + T2 S .

2.3

Linear Transformations and Matrices

Let V be nite dimensional vector space over a eld F , and suppose that dim V = n. Let {v1 , . . . , vn } n be a basis for V . If v V then we can write v = k=1 k vk for some scalars 1 , 2 , . . . , n F . These scalars are called the coecients of v relative to the basis {v1 , v2 , . . . , vn }. Once we have chosen a basis, it is common to represent the vector v by writing these coecients in a column: 1 2 v = . . . . n

We call this expression with n scalars a n 1 column vector , where n refers to the number of rows, and 1 refers to the number of columns, i.e., a single one. We sometimes abbreviate this notation by writing v = (j ). Now take nite dimensional vector spaces V and W over a eld F and T L (V, W ). Suppose that dim V = n, dim W = m, and {v1 , . . . , vn } and {w1 , . . . , wm } are bases for V and W , respectively. For 1 k n, we can write
m

T (vk ) =
j =1

ajk wj ,

where each ajk F . That is, the particular scalar ajk is the coecient of wj when writing the vector T (vk ) in terms of the basis {w1 , w2 , . . . , wm }. These scalars are sucient to characterize T , since any vector in V can be written as a unique linear combination of v1 , . . . , vn , and any vector in W can be written as a unique linear combination of w1 , . . . , wm . We encode this information about T in a rectangular array called a matrix . That is, we write a11 a12 a1n a21 a22 a2n A= . . . . .. . . . . . . . am1 am2 amn 62

Observe that the matrix A has m rows and n columns (rows are horizontal and columns are vertical). This matrix is called an m n (m by n) matrix over F . Of course, the coecients in the matrix depend on the choice of bases in V and W . Now, for any v V , we may write v with respect to the basis {v1 , . . . , vn }. As above, we nd scalars n 1 , 2 , . . . , n F such that v = k=1 k vk . Then we write T (v) with respect to the basis {w1 , . . . , wm } as follows:
n m m n

T (v) =

k=1

Expressing this same computation with the notation introduced above, we write a11 a21 . . . am1 a12 a22 . . . am2 .. . a1n a2n . . . amn 1 2 . . . n =
n k=1 n k=1 n k=1

j =1

ajk wj =

k ajk

wj .

j =1

k=1

a1k k a2k k . . . amk k

Note that the resulting m 1 column vector is the expression for T (v) in terms of its coecients with respect to the basis {w1 , w2 , . . . , wm }. We also refer to this column vector as the product of the matrix A = (aij ) by the vector v = (j ). Incidentally, this might be a good time to formalize the denition of matrix. Denition 2.3.1 Let R be a commutative ring with 1. Let aij be elements of R, where 1 i m and 1 j n. An m n matrix over R is a rectangular array given by A= a11 a21 . . . am1 a12 a22 . . . am2 .. . a1n a2n . . . amn .

Exercise 2.3.2 Let V and W be nite dimensional vector spaces over F of dimensions n and m respectively. Let Mmn (F ) be the collection of m n matrices over F . We use the notation A = (aij ) for elements of Mmn (F ). If A = (aij ), B = (bij ), we dene A + B = (aij + bij ) and for F we dene A = (aij ). i. Show that Mmn (F ) is a vector space over F . ii. Find a basis for Mmn (F ). iii. By xing bases for V and W give an explicit linear isomorphism between L (V, W ) and Mmn (F ). In the previous section, we saw that if T1 and T2 are in L(V, W ), then T1 + T2 is in L(V, W ), and in the previous exercise, we saw how to add their respective matrices. Similarly, we saw that if F and T L(V, W ), then T is in L(V, W ), and in the previous exercise, we saw how to multiply the matrix for T by the scalar . We turn now to the question of how to write the matrix for a linear transformation that is the composition of two linear transformations. Let U , V , and W be nite dimensional vector spaces over a eld F . Let S L(U, V ), and T L(V, W ). An exercise in the previous section showed that T S is in L(U, W ). Let {u1 , . . . , ul } be a basis for U . Let {v1 , . . . , vn } be a basis for V . Let {w1 , . . . , wm } be a basis for W . How are the matrices for S , T , and T S related? Let A = (ajk ) be the n l matrix for S , and let B = (bij ) be the m n matrix for T . (Note that we have intentionally used j as an index that runs from 1 to n in both cases.) What is the matrix for T S ? We express (T S )(uk ) in terms of the basis {w1 , . . . , wm } as follows: 63

(T S )(uk ) = T (S (uk ))
n

=T
n j =1 n

j =1

ajk vj

= =
j =1 m

ajk T (vj )
m

ajk
i=1 n j =1

bij wi

=
i=1

If C = (cik ) is the matrix for T S , with respect to the two bases {u1 , . . . , ul } and {w1 , . . . , wn }, then n the preceding computation shows that cik = j =1 bij ajk . This inspires us to dene matrix multiplication formally. Denition 2.3.3 Let A = (ajk ) be an n l matrix, and let B = (bij ) be an m n matrix. Then the product BA is the m l matrix C = (cik ), where cik = n j =1 bij ajk . Remark 2.3.4 Let A = (ajk ) be an n l matrix, and let B = (bij ) be an m n matrix. The denition above shows us how to multiply the matrix B by the matrix A. In should be noted that given this setup, the product AB is not even dened unless l = m. Even if l = m, the matrices BA and AB will be dierent sizes unless l = m = n. These are severe hurdles to any thoughts of commutativity in doing matrix multiplication, but this comes back to the more fundamental idea that the composition of functions is, in general, not commutative. The astute reader will notice the similarity between the way we dened the multiplication of two matrices, and the way we dened the multiplication of a matrix by a vector. In fact, the latter is a special case of the former. Exercise 2.3.5 Let A = (aij ) be an m n matrix, and let v = (j ) be an n 1 column vector. If we think of Av as a product of matrices, what linear transformations do v and Av represent? The above development is of particular importance in the case when V = W . In this case, we use the notation L (V ) for L (V, V ). If dim V = n, then dim L (V ) = n2 , and each element of L (V ) can be represented by an n n matrix relative to a single chosen basis of V . Along with the operations of addition and scalar multiplication in L (V ), we have composition of linear transformations. Suppose that S, T L (V ). Then S T is dened in the usual way by S T (v) = S (T (v)). Exercise 2.3.6 i. If S, T L (V ), show that S T L (V ). ii. If R, S, T L (V ), then R (S T ) = (R S ) T (this actually follows from the associativity of composition of functions discussed in Appendix ??). iii. If R, S, T L (V ), show that R (S + T ) = (R S ) + (R T ) and (R + S ) T = (R T ) + (S T ). iv. Let I L (V ) be dened by I (v) = v for v V . Show that T I = I T = T for all T L (V ). v. Show that if dim V 2, then L (V ) is not commutative with respect to . That is, there exist S, T L (V ) such that S T = T S . 64

bij ajk wi .

A little vocabulary is in order here. In Chapter 0, we used the terms commutative ring with 1, integral domain, and eld. As pointed out there, the word commutative referred to the operation of multiplication. Some of the most important algebraic structures that occur in analysis are called algebras. Denition 2.3.7 Let F be a eld. An algebra over F is a set A such that A is a vector space over F and A has an internal law of composition satisfying the associative law, and left and right distributivity. That is, for a, b, c A, we have a (b c) = (a b) c, a (b + c) = (a b) + (a c)

and (a + b) c = (a c) + (b c).

For scalar multiplication we have for F , ( a) b = (a b) = a ( b). An algebra A is an algebra with identity if there is an element 1 A so that a 1 = 1 a = a for all a A. The algebra A is a commutative algebra if a b = b a for all a, b A. Example 2.3.8 If V is a nonzero vector space over a eld F , then L (V ) is an algebra with identity which is commutative if and only if dim V = 1. Exercise 2.3.9 Let R[x] be the vector space of polynomial functions in one variable over R. Dene multiplication of polynomials in the usual way. Show that R[x] is a commutative algebra with identity. Exercise 2.3.10 Let A = Mnn (F ). Show that multiplication of n n matrices over a eld F is associative and that 1 0 0 0 1 0 I= . . . . . . . . . . . . 0 0 1 is an identity for multiplication of matrices. Show further that, for n n matrices, multiplication is left and right distributive over addition and the appropriate properties hold for scalar multiplication.

Conclusion: Mnn (F ) is an algebra with identity over F which is commutative if and only if n = 1. For simplicity we will write Mn (F ) for Mnn (F ). An invertible element of Mn (F ) is one that has a multiplicative inverse. The collection of these invertible elements plays a special role, which we will investigate in several ways. First, we will nish this section by discussing how to decompose an invertible matrix into elementary matrices. Secondly, in the next section, we will investigate the determinant a matrix, which gives a precise condition under which a matrix is invertible. Thirdly, in Project 2.6.1, we will discuss the algebraic structure of GLn (F ), the collection of invertible matrices. This will be of particular importance when we get to the Change of Variables Theorem in Chapter ??. Denition 2.3.11 An elementary transformation in GLn (F ) is a linear transformation of one of the following three forms: 1. multiplication of a coordinate by a nonzero constant k : T (x1 , x2 , . . . , xa1 , xa , xa+1 , . . . , xn ) = (x1 , x2 , . . . , xa1 , kxa , xa+1 , . . . , xn ); 2. interchange of two coordinates: T (x1 , x2 , . . . , xa1 , xa , xa+1 , . . . , xb1 , xb , xb+1 , . . . xn ) = (x1 , x2 , . . . , xa1 , xb , xa+1 , . . . , xb1 , xa , xb+1 , . . . , xn ); 65

3. replacement of a coordinate by the sum of itself and another: T (x1 , x2 , . . . , xa1 , xa , xa+1 , . . . , xb , . . . , xn ) = (x1 , x2 , . . . , xa1 , xa + xb , xa+1 , . . . , xb , . . . , xn ). An elementary matrix is the matrix of an elementary transformation with respect to the standard basis. Exercise 2.3.12 i. Find the elementary matrix for each elementary transformation.

ii. Show that any matrix in GLn (F ) can be written as a product of elementary matrices.

2.4

Determinants

Exercise 2.4.1 Suppose that T L(V ) is a bijection. Then T has an inverse, T 1 : V V . Show that T 1 L(V ). (See Exercise 2.2.4.) If T L(V ), and T is invertible, then of course T T 1 = T 1 T = I , where I is the identity map. The problem that confronts us is the following: if dim V = n, and A = (aij ) is the matrix of T relative to a given basis of V , how do we nd the matrix of T 1 relative to the same basis? We seek a matrix denoted by A1 such that A A1 = A1 A = I. Well, this shouldnt present a great problem. All we do is write the matrix for A1 as (xij ). This leads to n2 linear equations in n2 unknowns: a11 x11 + a12 x21 + + a1n xn1 a21 x11 + a22 x21 + + a2n xn1 an1 x11 + an2 x21 + + ann xn1 a11 x12 + a12 x22 + + a1n xn2 an1 x1n + an2 x2n + + ann xnn = = . . . = = . . . 1 0 0 0

= 1.

This looks somewhat tedious, so maybe at this stage, we should just tell you the answer and consider it further in a project at the end of the chapter. But, that would not be true to the nature of this book. So we are led to the quest for determinants, one of the great discoveries in mathematics. To begin a discussion of determinants, we must rst consider the collection Sn of all bijections from the set [n] = {1, 2, . . . , n} to itself. These bijections are called permutations of n elements . Example 2.4.2 Let us consider two examples, namely S2 and S3 . We can represent the elements of S2 by arrays in which the top row lists domain and the bottom lists the corresponding elements in the image. There are two permutations of two elements, namely I= 1 2 1 2 , r= 1 2 2 1 .

Note that these arrays should be thought of not as matrices, but simply as a way to represent permutations as functions. Similarly, we can write the six elements of S3 as I= 1 2 1 2 1 1 3 3 2 3 3 2 , r= 1 2 2 3 3 1 , 3 1 r2 = r r = , f3 = 1 2 3 1 1 2 2 1 3 3 3 2 .

f1 = Exercise 2.4.3

f2 =

1 2 3 2

66

i. Use the fundamental counting principle (see Theorem ??) to show that the number of elements in Sn is n!. ii. Show that the composition of two elements of Sn is also an element of Sn . iii. Show that the elements of Sn satisfy the associative law under composition of functions. iv. Dene I Sn by I (x) = x for all x [n]. Show that I is an identity for Sn under composition. v. If Sn , dene 1 as one does for any bijection. That is, (x) = y i 1 (y ) = x. Show that 1 = 1 = I . The collection Sn with the internal law of composition satisfying the above properties is an example of an algebraic structure called a group . The general theory of groups is discussed in the Projects at the end of the chapter. The group Sn is called the symmetric group on n objects . Denition 2.4.4 Let be an element of Sn . We dene the sign of by sgn( ) =
{i,j }[n], i=j

(j ) (i) . ji

Exercise 2.4.5 Show that if Sn , then sgn( ) = 1. Exercise 2.4.6 Let be any function from [n] to [n]. We can dene sgn( ) as above. Show that sgn( ) = 0 if and only if is not a bijection. Proposition 2.4.7 For , Sn , we have sgn( ) = sgn( )sgn( ) Proof. We have sgn( ) = =
{i,j }[n], i=j

{i,j }[n], i=j

( (j )) ( (i)) ji ( (j )) ( (i)) (j ) (i) (j ) (i) ji ( (j )) ( (i)) (j ) (i) (l) (k ) lk (j ) (i) ji

=
{i,j }[n], i=j

{i,j }[n], i=j

=
{k,l}[n], k =l

{i,j }[n], i=j

(j ) (i) ji

= sgn( )sgn( )

Exercise 2.4.8 Find the signs of the permutations in S2 and S3 . Denition 2.4.9 a. The permutations in Sn for which sgn( ) = 1 are called even permutations . Those permutations for which sgn( ) = 1 are called odd permutations. The collection of even permutations is denoted by An . 67

b. A permutation which interchanges two distinct elements and leaves the remaining elements xed is called a transposition. The transposition which sends i to j , j to i, and leaves everything else xed is written (ij ). Exercise 2.4.10 i. Show that a transposition is an odd permutation. ii. Show that every element of Sn can be decomposed as a product of transpositions. iii. Show that Sn is an even permutation if and only if can be decomposed into an even number of transpositions. Also show that Sn is an odd permutation if and only if can be decomposed into an odd number of transpositions. The number of transpositions is not unique but the parity is always the same. iv. Show that An is a group. (See Project 2.1. An is called the Alternating Group on n objects). v. Show that the number of elements in An is n!/2. vi. Show that sgn( ) can be dened as simply the sign of the integer vii. Show that A2 = {I } and A3 = {I, r, r2 }. viii. Decompose each element in S3 as a product of transpositions. ix. Write explicitly as arrays the elements of A4 and A5 . We are now prepared to dene the determinant of an n n matrix. Denition 2.4.11 Let A = (aij ) be an n n matrix over a eld F . The determinant of A, denoted by det A, is dened as det A = sgn( )a1,(1) a2,(2) an,(n)
Sn 1i<j n ( (j )

(i)).

Example 2.4.12 Consider a 2 2 matrix A= Then det A = a11 a22 a12 a21 . Exercise 2.4.13 Write out the expression for the determinant of a 3 3 matrix a11 a12 a13 A = a21 a22 a23 . a31 a32 a33 It should have 3! = 6 terms. We have two tasks ahead. The rst is to illustrate the role of the determinant in computing the inverse of an n n matrix. The second is to nd some reasonable way to compute the determinant of a matrix. Denition 2.4.14 Suppose A = (aij ) is an n n matrix over a eld F . Then the transpose of A, denoted t A, is the matrix obtained by reecting A around the main diagonal, that is, the collection of elements a11 , a22 , . . . , ann . Thus, t A = (aji ). 68 a11 a21 a12 a22 .

Hence, if A= and, if a11 a21 a12 a22 a32 a12 a22 , then
t

A=

a11 a12

a21 a22 a21 a22 a23

, a31 a32 . a33

Exercise 2.4.15 Let A be an n n matrix over a eld F . Show that det(A) = det(t A). Exercise 2.4.16 Let A = a11 a21 a12 a22 . Suppose that det A = 0. Show that A1 exists and nd it.

a11 A = a21 a31

a13 a23 , then a33

a11 t A = a12 a13

Lemma 2.4.17 If A = (aij ) is an n n matrix over a eld F such that, for some m, k, with m = k , the m-th row is equal to the k -th row, then det A = 0. Proof. For any Sn let = (km), where (km) is the transposition dened above. Our assumption implies that a1,(1) a2,(2) an,(n) = a1, (1) a2, (2) an, (n) On the other hand, sgn( ) = sgn( ) ((km)) = sgn( ) (1) = sgn( ). This shows that det A = 0. Exercise 2.4.18 Show that, if A = (aij ) is an n n matrix over a eld F such that, for some m, k, with m = k , the m-th column is equal to the k -th column, then det A = 0. The following exercise will prove useful in our discussion of the properties of determinants. Exercise 2.4.19 Suppose that A = (aij ) and B = (bij ) are n n matrices over a eld F . Suppose further that : [n] [n] is not a bijection. Show that
Sn

sgn()a1,(1) b(1),(1) a2,(2) b(2),(2) an,(n) b(n),(n) = 0.

Denition 2.4.20 Let A = (aij ) be an n n matrix over a eld F . Let Aij be the (n 1) (n 1) matrix obtained by deleting the i-th row and the j -th column of A. The (i, j ) cofactor of A is the element Cij of F dened by Cij = (1)i+j det Aij Theorem 2.4.21 Let A = (aij ) be an n n matrix over a eld F . Then, for any xed k with 1 k n, det A = ak1 Ck1 + ak2 Ck2 + + akn Ckn . This is called the expansion of the determinant of A with respect to the k -th row. Exercise 2.4.22 Let A be a 3 3 matrix over a eld F . Show that the expansion of det A with respect to any row yields the same answer you obtained in the exercise above. Proof of the theorem. By denition, det A is the sum of products of the form sgn( )a1,(1) a2,(2) an,(n) ()

where runs through the elements of Sn . We claim that the sum of all expressions of the form () for which (k ) = j is equal to akj Ckj . If we show this for every j with 1 j n, then, summing over all j , we get the desired result. 69

We have
Sn (k)=j

sgn( )a1,(1) a2,(2) an,(n) = akj

Sn (k)=j

sgn( )a1,(1) a2,(2) ak,(k) an,(n) ,

where indicates that the factor is removed from the product. Thus, we need to check that
Sn (k)=j

sgn( )a1,(1) ak,(k) an,(n) = (1)j +k det(Akj ).

(#)

To compute det(Akj ), we must rst re-index the rows and columns such that the indices go from 1 to . . . , n}, by n 1. For this, dene : {1, 2, . . . , n 1} {1, 2, . . . , k, (j ) = j j+1 for 1 j k 1, for k j n 1.

Similarly, with k replaced by j , dene a bijection : {1, 2, . . . , n1} {1, 2, . . . j, . . . , n}. Let Sn be such that (k ) = j . Then the map : {1, 2, . . . , n 1} {1, 2, . . . n} does not contain j in its image. The map 1 : {1, 2, . . . , n 1} {1, 2, . . . , n 1} is well dened. In fact, the map { Sn | (k ) = j } Sn1 given by 1 is a bijection. Now, recalling the denition of det Akj , we see that the proof of (#) follows immediately from sgn( ) = (1)j +k sgn( 1 ). Note that and are strictly increasing maps so that sgn( 1 ) coincides with the sign (i)). Canceling the product on both sides of (##), we are left with showing that Recalling that (k ) = j , the last product is
k 1 i=1 n

(##)
1i<ln ( (l ) i,l=k

sgn

1i<ln i=k or l=k

( (l) (i)) = (1)j +k .

(j (i))

l=k+1

( (l) j ) = (1)k1

l= k

( (l) j ).

Moreover, sgn as desired.

l=k ( (l )

j ) is clearly (1)j 1 . Altogether, we obtain (1)k1 (1)j 1 = (1)j +k

Exercise 2.4.23 Suppose that A = (aij ) is a n n matrix over a eld F . Use the fact that det A = det t A to give an expansion of the determinant with respect to the k -th column. We can now assert a theorem about inverses. Theorem 2.4.24 If A is an n n matrix over a eld F and det A = 0,then A has an inverse. If det A = 0, the matrix of A1 is the transpose of the cofactor matrix multiplied by the inverse of the determinant of A. That is, C11 C21 Cn1 1 C12 C22 Cn2 A1 = . . . . . . . det A . . . . . C1n C2n Cnn 70

Proof. Let C = (Cij ) and consider the product t CA. Look at the diagonal elements in this product. The j -th diagonal element is a1j C1j + + anj Cnj = det A. For the o-diagonal elements, we take {k, m} [n], k = m and consider the (m,k )-th entry of t CA. We get a1k C1m + a2k C2m + + ank Cnm . This represents expansion of the determinant of a matrix A that is equal to A with the exception that the m-th column has been replaced by the k -th column. By the Exercise 2.4.18, this determinant is 0. Thus, if det A = 0, then
t

C/ det(A) is the left inverse of A.

Exercise 2.4.25 Show that t C/ det(A) is also a right inverse for A. We now wish to prove that an n n matrix A over a eld F has a multiplicative inverse if and only if det A = 0. One half of this fact was proved above. That is, if det A = 0, then A has an inverse. The other half depends on the following important theorem. Theorem 2.4.26 If A = (aij ) and B = (bij ) are n n matrices over a eld F , then det(AB ) = det(A) det(B ). Proof. We rst expand det(A) det(B ) =
,Sn

sgn( )sgn( )a1,(1) a2,(2) an,(n) b1, (1) b2, (2) bn, (n) .

We re-index the product of the bjk s as follows. It is clear that b1, (1) bn, (n) = b(1), ((1)) b(n), ((n)) . For xed in Sn , we see that, as runs through Sn , so does . Moreover, sgn( )sgn( ) = sgn( ). Hence, by letting = , we have det A det B =
Sn Sn

sgn()a1,(1) b(1),(1) a2,(2) b(2),(2) an,(n) b(n),(n)

By Exercise 2.4.19 this last sum will not change if we allow to run over all maps : [n] [n]. For a xed , we have

a1,(1) b(1),(1) a2,(2) b(2),(2) an,(n) b(n),(n) .

Now we let C = AB and consider det(C ). Let C = (cij ), then we have det(C ) =
Sn

sgn()c1,(1) c2,(2) cn,(n) .


n kj =1

Now, from the denition of C , we know that cj,(j ) =


n n n n

ajkj bkj ,(j ) . This gives

cj,(j ) =
j =1 k1 =1 k2 =1

kn =1

a1k1 bk1 ,(1) a2k2 bk2 ,(2) ankn bkn ,(n)

For each term (k1 , k2 , . . . , kn ) in the sum, we can dene a map : [n] [n] by 1 k1 , 2 k2 , . . . , n kn . Notice that there are exactly nn such maps. Hence, all maps : [n] [n] arise in this way. So we can index the sum by as we let run over all maps. In this way, we get det(C ) =
Sn

sgn()a1,(1) b(1),(1)) a2,(2) b(2),(2) an,(n) b(n),(n) sgn()a1,(1) b(1),(1)) a2,(2) b(2),(2) an,(n) b(n),(n)

=
Sn Sn

= det A det B.

71

Theorem 2.4.27 Let A be an n n matrix over a eld F . Show that A has a multiplicative inverse if and only if det A = 0. Proof. Exercise. The next exercise illustrates several of the important properties of determinants of n n matrices. Exercise 2.4.28 Find the determinants of each of the three types of elementary matrices. Exercise 2.4.29 Suppose that A is an n n matrix over a eld F . i. If we multiply a row or column of A by a scalar c, nd the determinant of the resulting matrix. ii. Show that if we interchange two rows or two columns of A, then the determinant is det A. iii. Show that if we add a scalar multiple of any row to any other row, or if we add a scalar multiple of any column to any other column, then the determinant remains unchanged. Exercise 2.4.30 Let A be a n n matrix over a eld F . i. If det A = 0, show that the columns of A are linearly independent and hence form a basis for F n . ii. Do the same for the rows. iii. If the columns of A are linearly independent, show that det A = 0. (Hint: consider the image of the linear transformation dened by A.) iv. Do the same for the rows.

2.5

Geometric Linear Algebra

We now wish to investigate the geometry of nite dimensional vector spaces over R. A point in Rn is represented by an n-tuple of elements of R, written p = (p1 , . . . , pn ), with each pi R. Geometrically, these are thought of as points in space, where the pi s give the coordinates of the point p. At the same time, we may consider n-tuples of real numbers as n-dimensional vectors giving the data of a direction and a magnitude, without specifying a base point from which this vector emanates. Thinking this way, we see that such vectors are elements of a vector space, En , where elements can be written as v = (v1 , . . . , vn ), with each vi R. We will consistently distinguish between the points of Rn , and vectors in En , since geometrically they are quite dierent. Observe that we are choosing the vectors e1 , e2 , . . . , en from Exercise 2.1.11 as a basis for the vector space En and further that v = (v1 , v2 , . . . , vn ) may be written as the linear combination v = v1 e1 + v2 e2 + + vn en . Moreover, the coordinates of a point p are determined by a collection of mutually perpendicular coordinate axes and a distinguished point called the origin, namely, (0, 0, . . . , 0). The idea here is that vectors are free to wander around in space, and points have to stay where they are. If we want to add two vectors v, w En , we represent v + w as the diagonal of a parallelogram as pictured below. 72

v+w

and if v = (v1 , . . . , vn ) and w = (w1 , . . . , wn ), then v + w = (v1 + w1 , . . . , vn + wn ). Continuing this idea, we specify that a direction v based at a point p = (p1 , p2 , . . . , pn ) is given by an element of Rn En . We also have a geometric operation on Rn En which is to drag the point p along v to get a new point q = (q1 , q2 , . . . , qn ). This geometric operation is algebraically encoded in the formula: Rn En Rn (p, v) p + v = q

((p1 , . . . , pn ), (v1 , . . . , vn )) (p1 + v1 , . . . , pn + vn ) = (q1 , , qn ). With this in mind, we have the statements vector + vector = vector, point + vector = point, so naturally, point point = vector! To make sense of this formally, given an ordered pair of points (p, q) Rn Rn , there is a unique vector v En such that p + v = q. This v represents q p. Of course, algebraically it is given by nothing more than v = (q1 p1 , . . . , qn pn ). We are now ready to dene some geometric objects in Rn . To describe a line, we need a point p0 Rn and a (nonzero) direction v En . Denition 2.5.1 Let p0 be a point in Rn and v be a direction in En . The line through p0 in the direction v is given by = { p R n | p = p0 + tv , t R } . Notice that we are using the formalism of vector plus point equals point. Denition 2.5.2 Suppose that p0 Rn and v, w are linearly independent vectors in En . The plane through p0 spanned by v and w is P = {p Rn | p = p0 + tv + sw, t, s R.} More generally, we can use these ideas for other subsets of Rn . 73

Denition 2.5.3 If v1 , . . . , vk are linearly independent vectors in En , then we dene the k -dimensional ane subspace through p0 Rn spanned by v1 , . . . , vk as H = {p Rn | p = p0 + t1 v1 + . . . + tk vk , where tj R, 1 j k }. Note: The collection of vectors {t1 v1 + . . . + tk vk , tj R} is actually a subspace of En . Thus, a k dimensional ane subspace is constructed by taking a k -dimensional subspace of En and adding it to a point of Rn . When k = n 1, H is called a hyperplane in Rn . Denition 2.5.4 If v1 , . . . , vk are linearly independent vectors in En , and p0 Rn , we dene the k dimensional parallelepiped with vertex p0 spanned by v1 , v2 , . . . , vk as P = {p Rn | p = p0 + t1 v1 + . . . + tk vk , with 0 tj 1}. Note that if k = n = 2 then P is just a standard parallelogram in R2 . Much of the geometry that appears in this section will arise in a more general context in later chapters. We introduce only enough here to make the reader feel comfortable in Rn . The rich interplay between Rn and the vector space En is what makes life interesting. Denition 2.5.5 Let V be a vector space over a eld F . A bilinear form , on V is a map , : V V F which satises linearity in both variables. That is, for all v, v1 , v2 , w, w1 , w2 V , and all F , v1 + v2 , w = v1 , w + v2 , w v, w = v, w v , w1 + w2 = v , w1 + v , w2 v, w = v, w . The form , is said to be symmetric if v, w = w, v for all v, w V . Denition 2.5.6 Let V be a vector space over R. The bilinear form , is said to be positive denite if v, v 0 for all v V , and v, v = 0 if and only if v = 0. Bilinear forms and their companion Hermitian forms (over C) will appear regularly throughout the book. For now, we assume F = R. The main example of a positive denite symmetric bilinear form on En is the scalar product or dot product. Denition 2.5.7 Suppose that v = (v1 , . . . , vn ) and w = (w1 , . . . , wn ) are vectors in En . The scalar product of v and w is v, w = v1 w1 + . . . + vn wn . The scalar product is sometimes called the dot product and is denoted by v w. We will try our best to be consistent and use , . Exercise 2.5.8 Prove that the scalar product is a positive denite symmetric bilinear form on En . Exercise 2.5.9 Let V = E2 . Show that the map f : E2 E2 R given by f ((v1 , v2 ), (w1 , w2 )) = v1 w1 is a symmetric bilinear form that is not positive denite. There are two concepts that arise immediately with the existence of a positive denite symmetric bilinear form. The rst is the length or norm of a vector and the second is orthogonality. Denition 2.5.10 If v = (v1 , . . . , vn ) En , then the length or norm of v is dened by ||v|| =
2 2 1/2 v, v = (v1 + + vn ) .

74

Exercise 2.5.11 Prove the following properties of the norm. If v, w En , then: i. ||v|| 0; ii. ||v|| = 0 i v = 0; iii. ||v|| = || ||v||, v. v+w
2

R;
2

iv. ||v + w|| ||v|| + ||w||; + vw =2 v


2

+ w

There is one more fundamental inequality relating the scalar product of two vectors to their norms. We will use the famous quadratic formula from high school algebra to prove it. Theorem 2.5.12 (Cauchy-Schwarz Inequality) Let v, w En . Then | v, w | v Proof. Let be a real number. Then 0 v w, v w = v, v v, w w, v + w, w = ||v||2 2 v, w + 2 ||w||2 . This is a quadratic polynomial in which is always greater than or equal to 0. For this inequality to hold, the discriminant of this quadratic must be nonpositive. That is, we must have, in the usual notation, b2 4ac 0. With a = w 2 , b = 2 v, w , and c = v 2 , we get our desired inequality immediately. Exercise 2.5.13 Prove that equality holds in the Cauchy-Schwarz Inequality i one of the vectors is a scalar multiple of the other. The denition of the norm leads to the usual denition of Euclidean distance between two points in Rn . Thus, if p1 , p2 Rn then d(p1 , p2 ) = ||p1 p2 ||. The general study of distance is carried out in Chapter 3 where we discuss metric spaces. Since we have a positive denite symmetric bilinear form, the concept of orthogonality (or perpendicularity) in En can be formalized as follows. Denition 2.5.14 Let v, w En . Then v and w are said to be orthogonal (or perpendicular ) if v, w = 0. A set {v1 , v2 , . . . , vn } of vectors in En is said to be mutually orthogonal or pairwise orthogonal if vi , vj = 0 for all pairs i, j with i = j . Exercise 2.5.15 i. Show that the vector 0 in En is orthogonal to every vector in En . ii. Show that the vectors in the set {e1 , e2 , . . . , en } are pairwise orthogonal, that is, ei , ej = 0 if i = j , and further that ei , ei = 1. iii. If v is a nonzero vector in En , show that the collection W = {w En | w, v = 0} is an (n 1)dimensional subspace of En . iv. If v1 , . . . , vk are pairwise orthogonal non-zero vectors in En , show that they form a linearly independent set in En . Now, we wish to consider the angle between two non-zero vectors v, w En . If the vectors are linearly dependent, that is, w = v for some nonzero scalar R, then the angle between them is 0 if > 0 and 180 if < 0. If v and w are linearly independent, we look at the plane through the origin spanned by v and w. In this case, there are two angles associated with v and w. One is less than 180 and one is greater than 180. We take the angle between v and w to be the angle which is less then 180 . 75 w .

Theorem 2.5.16 Let v and w be linearly independent vectors in En . The angle between v and w is the unique solution to the equation cos = v, w , ||v|| ||w|| 0 < < 180. (2.1)

(Draw a picture!)

Proof. This is just the law of cosines, which states that ||v w||2 = ||v||2 + ||w||2 2||v|| ||w|| cos .

One of the most important procedures in geometric linear algebra is producing vectors that are orthogonal to a given set of vectors. This leads us to dene one of the most important operations on vectors in En . Denition 2.5.17 Let v and w be linearly independent vectors in En . The projection of v onto w is dened by v, w w . (2.2) projw (v) = ||w|| ||w||

projw (v)

Observe that w/ w is a unit vector in the direction of w. Hence, the projection of v on w has the same direction as w if v, w > 0, and the direction of w if v, w < 0. Of course, if v, w = 0 then v and w are v ,w | orthogonal and the projection is just the zero vector. Note that the norm of projw (v) is projw (v) = | || w|| . Exercise 2.5.18 Show that v projw (v) is orthogonal to w. The exercise above allows us to construct a vector orthogonal to a single vector in En . Next, let us consider the problem of constructing a vector orthogonal to a collection of n 1 linearly independent vectors. First, lets look in 2 dimensions. Given a vector v = (v1 , v2 ) E2 , v = 0, we set v = (v2 , v1 ). Then, v, v = 0, that is, these vectors are orthogonal. Now, lets look in 3 dimensions. In E3 , we have the special notion of the cross product of two vectors. This is can be dened by using determinants, but for the moment, we dene it as follows. The cross product, , is a map that takes two vectors in E3 and produces a third vector in E3 . If v = (v1 , v2 , v3 ) and w = (w1 , w2 , w3 ) in E3 , the formula for v w is given by v w = (v2 w3 v3 w2 , v3 w1 v1 w3 , v1 w2 v2 w1 ). Observe that, unlike the scalar product, the cross product produces a vector, not a scalar. Exercise 2.5.19 i. Show that v w = 0 if and only if v and w are linearly independent. ii. Show that v, v w = w, v w = 0. iii. Show that w v = (v w). Exercise 2.5.20 76

i. Show that ||v w|| = ||v|| ||w|| sin , where is the angle between v and w. ii. Show that v w is the area of the parallelogram spanned by v and w. We have shown that given a non-zero vector v E2 , we can nd a vector v which is non-zero and orthogonal to v. Obviously, the pair {v, v } is a basis for E2 . Next, given two linearly independent vectors v1 , v2 E3 , we constructed v3 = v1 v2 orthogonal to both of the original vectors. The set {v1 , v2 , v3 } is a basis for E3 . Let us investigate how this process can be generalized to En . Determinants can be used in an interesting way in the aforegoing process. For example, given a vector v = (v1 , v2 ) E2 , we write the matrix e1 e2 . v1 v2 Without being concerned about vectors in the rst row and numbers in the second row, we can take the determinant of this matrix by expanding according to the rst row and obtain v2 e1 v1 e2 , which is the vector v = (v2 , v1 ). Similarly, if we have linearly independent vectors v = (v1 , v2 , v3 ) and w = (w1 , w2 , w3 ) in E3 , we can consider e1 e2 e3 v1 v2 v3 . w1 w2 w3 Taking the determinant of this matrix by expanding according to the rst row, we get (v2 w3 v3 w2 )e1 + (v3 w1 v1 w3 )e2 + (v1 w2 v2 w1 )e3 , which is v w. We can generalize this to n dimensions.

Let v be the vector obtained by taking the determinant of this matrix with respect to the rst row. Then v is nonzero and is orthogonal to each of the vectors v1 , v2 , . . . , vn1 . Moreover, the collection {v1 , v2 , . . . , vn1 , v} is a basis for V .

Theorem 2.5.21 Let V = En . Suppose that v1 , v2 , . . . , vn1 is a collection of linearly independent vectors in V , where vj = (vj 1 , vj 2 , . . . , vjn ). Consider the matrix e1 e2 en v11 v12 ... v1n . . . . .. . . . . . . . v(n1)1 v(n1)2 . . . v(n1)n

Proof. The vector v obtained by expanding with respect to the rst row is simply the vector of cofactors v = (C11 , C12 , . . . , C1n ). If we replace the rst row by v, we obtain the matrix C11 C12 C1n v11 v12 ... v1n A= . . . . . . . . . . . . . v(n1)1 v(n1)2 ... v(n1)n We prove rst that the vector v is orthogonal to the vectors v1 , . . . , vn1 . To see this, choose i with 1 i n 1. Then vi , v = vi1 C11 + vi2 C12 + + vin C1n . This is the determinant of the matrix obtained by replacing the rst row of A by the vector vi . By Lemma 2.4.17, this determinant is 0. Next, we establish that v is not the 0 vector. To do this, we note that if C1j = 0 for all j , and we replace v by any vector w in V , then the determinant of the resulting matrix will be 0. But since v1 , . . . , vn1 form a linearly independent set in V , we can extend this set to a basis in V with a vector vn . Replacing the vector v by vn , we get a matrix whose determinant is not 0. This is a contradiction. Given n 1 linearly independent vectors in En , the above theorem produces a vector that is orthogonal to each of these vectors. If the original set of vectors was mutually orthogonal, the new set of n vectors will 77

be mutually orthogonal. This mutually orthogonal set of vectors will be a basis for En , however, because we have norms at our disposal, we can go one step further with the following denition. Denition 2.5.22 Let V = En . We say that v V is a normalized vector or unit vector , if v 2 = v, v = 2 2 2 v1 + v2 + . . . + vn = 1. An orthonormal set in V is a collection {v1 , v2 , . . . , vk } of linearly independent vectors in V such that vi | vj = 0 if i = j , and vi | vi = 1 for all i, j = 1, . . . , k . If k = n, the orthonormal set {v1 , v2 , . . . , vn } is called an orthonormal basis for V . Example 2.5.23 The collection {e1 , e2 , . . . , en } is an orthonormal basis for En . Exercise 2.5.24 If, in the previous theorem, {v1 , v2 , . . . , vn1 } forms an orthonormal set in V , show that {v1 , v2 , . . . , vn1 , v} is an orthonormal basis for V . (Hint: Consider At A and AA1 .) Exercise 2.5.25 Suppose that v1 , v2 , . . . , vn1 are linearly independent vectors in En . Take a point p0 in Rn and consider the hyperplane H through p0 spanned by v1 , v2 , . . . , vn1 . If v is the vector determined in Theorem 2.5.21, show that H = {p Rn | p p0 , v = 0}. Specialize this to obtain formulas for a line in R2 and for a plane in R3 . Exercise 2.5.26 Let v1 , v2 , . . . , vn1 be linearly independent vectors in En . Let p0 be a point of Rn . Let H be the hyperplane through p0 spanned by v1 , v2 , . . . , vn1 . If p is any point in Rn , show that the distance from p to H, that is, inf { p q | q H}, is given by the length of the vector projv (p p0 ) where v is the vector obtained in Theorem 2.5.21. Specialize this to obtain formulas for the distance from a point to a line in R2 and from a point to a plane in R3 . Exercise 2.5.27 i. Find a formula for the distance from a point to a line in Rn . ii. Find the distance between two nonintersecting lines in R3 . iii. Find the distance between two nonintersecting planes in R5 . Continuing our theme of attempting to produce vectors that are orthogonal to each other or to other sets of vectors, we nish this section with a general procedure for turning sets of linearly independent vectors into sets of mutually orthogonal vectors. This process is known as the Gram-Schmidt Orthogonalization process. Specically, given a set of k linearly independent vectors in En , the Gram-Schmidt Orthogonalization process produces a set of k mutually orthogonal vectors that span the same subspace as the original k vectors. Moreover, the Gram-Schmidt process allows us to extend such a mutually orthogonal set to a mutually orthogonal basis for En . As a last step, dividing each vector by its norm will produce an orthonormal basis for En . 1 = v1 . We We begin with a set {v1 , . . . , vk } of linearly independent vectors and proceed as follows. Let v k by taking vk and subtracting the projections on the vectors already constructed. continue to nd vectors v More explicitly, we let 2 v 3 v = = . . . = v2 projv 1 (v2 ), (2.3) (2.4) (2.5) vk
k 1 i=1

v3 projv 2 (v3 ), 1 (v3 ) projv

k v

projv i (vk ).

(2.6)

It is easy to check that this set of vectors is pairwise orthogonal. 1 , v 2, . . . , v k span the same subspace as v1 , v2 , . . . , vk . Exercise 2.5.28 Check it, and, in addition, show that v 78

On the other hand, suppose that {v1 , v2 , . . . , vk } is a mutually orthogonal set of nonzero vectors with k < n. We can complete this to a basis {v1 , v2 , . . . , vk , vk+1 , . . . , vn } using Theorem ??. If we now apply the k+1 , . . . , v n }, Gram-Schmidt orthogonalization process to this basis, we get a mutually orthogonal basis {v1 , v2 , . . . , vk , v does not alter our original mutually orthogonal set. Exercise 2.5.29 Prove this last statement. The last step is to try to work with vectors whose norms are equal to one. Note that if v is a nonzero vector in En , then v can be converted to a vector of norm 1, that is, a unit vector, by dividing v by v . Exercise 2.5.30 Consider the vectors v1 = (1, 1, 1, 0), v2 = (1, 0, 0, 1), and v3 = (0, 1, 1, 1) in E4 . i. Use the Gram-Schmidt orthogonalization process on these three vectors to produce a set of three mutually orthogonal vectors that span the same subspace. ii. Extend the set of three vectors produced in part i to a mutually orthogonal basis for E4 . iii. Normalize your basis so that it becomes an orthonormal basis for E4 . Exercise 2.5.31 Show that, given linearly independent vectors v1 , . . . , vk in En , we can transform this col1, . . . , v k }, which spans the same subspace. In addition, the set {v 1, . . . , v k } lection into an orthonormal set {v can be completed to an orthonormal basis for En .

2.6

Independent Projects

2.6.1 GroupsIn Chapter 0, we have introduced commutative rings with 1, integral domains, and elds. In this chapter, we have introduced vector spaces and algebras over a eld, and we have dened groups in the context of the symmetric group. This project gives more details and exercises about groups. Denition 2.6.1 Let G be a set with an ILC : G G G (see Denition 0.2.21). The pair (G, ) is a group if, a. for all a, b, c G, we have (a b) c = a (b c) (Associativity); b. there is an identity element e G such that for each a G, e a = a e = a (Identity); and c. for every element a G, there is an element a1 G called the inverse of a such that a a1 = e = a1 a (Inverses). Here is a set of elementary exercises about groups. Exercise 2.6.2 Suppose that (G, ) is a group. Prove the following statements. i. Show that the identity is unique. ii. Show that inverses are unique. iii. If a, b are elements of G, show that the equations a x = b and x a = b can be solved uniquely. iv. If a, b, c G and a b = a c, show that b = c. v. If a, b G, show that (a b)1 = b1 a1 . Remark 2.6.3 It is often convenient to omit the ILC when writing a product. Thus we can write a b as ab. In particular, if (G, ) is a group and k is a positive integer, we can write ak for a a a (k times). We can also write ak = (a1 )k . Given all of this, we still occasionally write a b when we feel it is useful. 79

Examples 2.6.4 Here are some elementary examples of groups: 1. the ordinary integers (Z, +); 2. the integers {1, 1} under multiplication; 3. the rational numbers (Q, +); 4. the nonzero rational numbers (Q , ); and 5. for an integer n 2, (Zn , +). Exercise 2.6.5 Decide whether or not the following are groups: i. (N, +); ii. ({0}, +); iii. (Z, ); iv. (Q, ); v. N with ILC a b = ab ; vi. (Z, ); vii. ({1}, ); viii. (Zn \ { 0}, ). Here are some more complicated examples. Example 2.6.6 The rotations of a regular n-gon about the center form a group denoted by Cn . The ILC in this case is the composition of rotations. Let r be a counterclockwise rotation through (360/n), that is, 2/n radians. The rotation group of a regular n-gon consists of I (the identity rotation), r, r2 (the counterclockwise rotation through (2 360/n)), r3 , . . . , rn1 . Note that rn = I . Examples 2.6.7 1. The symmetric group Sn of permutations of n objects is a group whose ILC is composition of functions. 2. The alternating group An of even permutations in Sn is a group under composition of functions. Examples 2.6.8 1. The set Mn (F ) of n n matrices over a eld F with addition as the ILC is a group.

2. The set GLn (F ) = {A Mn (F ) | det A = 0} is a group under multiplication of matrices. The group GLn (F ) is called the n n general linear group over F . Remark 2.6.9 Many of the groups we encounter arise from considering a single operation in a structure that has more than one operation. For example, forgetting the multiplication in a ring or the scalar multiplication in a vector space and remembering only the addition leaves us with a group. In such situations, we will continue to denote the ILC by +, the identity element by 0, and the inverse of an element a by a. We will also write ka instead of ak for a A, k Z. Examples 2.6.10 1. If R is a ring, then (R, +) is a group. 2. If V is a vector space over a eld F , then (V, +) is a group. Denition 2.6.11 Suppose that (G, ) is a group. If the number of elements in G is nite, we write |G| = n where n is the number of elements of G, and we call n the order of G. If G has an innite number of elements, we say that G is an innite group . 80

Exercise 2.6.12

i. Find the orders of the groups Zn , Cn , Sn , and An .

ii. Find the orders of Mn (F ) and GLn (F ) when F = Zp . Many of the examples of groups above, including all of the ones in Examples 2.6.4 and 2.6.10, t a category of groups that are particularly easy to work with, that is, abelian or commutative groups. Denition 2.6.13 A group (G,) is abelian or commutative if a b = b a for all a, b G. In the examples above, the groups Sn for n 3, An for n 4, and GLn (F ) for n 2 are not abelian. The simplest groups to analyze are cyclic groups. Denition 2.6.14 If G is a group, and there exists an element a G such that G = {ak | k Z}, then G is called the cyclic group generated by a. Note that we do not assume in the denition of a cyclic group that the various powers ak are distinct. Exercise 2.6.15 i. If G is a cyclic group of order n generated by a, show that an = e. ii. Show that a cyclic group of order n generated by a is also generated by ak when k and n are relatively prime. iii. Show that cyclic groups are abelian and that in a cyclic group of order n, (ak )1 = ank for 1 k n 1. Examples 2.6.16 1. The integers, Z, under addition form an innite cyclic group generated by 1. 2. The even integers, 2Z, form an innite cyclic group under addition generated by 2. 3. Let p be a prime, and let G = {pk | k Z}. Then G is an innite cyclic group under multiplication generated by p. 4. The group Cn of rotations of a regular n-gon is a cyclic group of order n generated by r. 5. The group (Zn , +) is a cyclic group of order n generated by 1. Exercise 2.6.17 Show that the group of rational numbers under addition is not a cyclic group. Exercise 2.6.18 Can a cyclic group be uncountable? Denition 2.6.19 Let G be a group and let a G. If there exists m N such that am = e, we say that a has nite order in G. If no such m exists, we say that a has innite order in G. If a is of nite order, then n = min{m N | am = e} is called the order of a in G. Note that order of an element a is 1 if and only if a is the identity element. Exercise 2.6.20 i. If G is a nite group show that the order of an element in G is less then or equal to the order of G. ii. Find the order of the elements in C12 and S4 . iii. Does there exist a group with elements of both nite and innite order? We next need the notion of a subgroup. 81

Denition 2.6.21 Let (G, ) be a group. A non-empty subset H of G is a subgroup if the pair (H, ) is a group. If H is a subgroup of G, with H = G and H = {e}, we say that H is a proper subgroup of G. Remark 2.6.22 Thus, we require that is an ILC on H , that the identity of the group e is an element of H , and if an element a is in H , then a1 is in H . Observe that associativity is automatic for the elements of H because associativity holds for the elements of G. In this situation, we say that associativity is inherited from G. Exercise 2.6.23 Show that a subgroup of a cyclic group is a cyclic group. That is, if G is the cyclic group generated by a and H is a subgroup of G, then there exists an element k N such that ak generates H . This means that for every h H , there exists j Z such that (ak )j = h. Exercise 2.6.24 Show that every element in a group generates a cyclic subgroup. Examples 2.6.25 1. The groups {e} and G are subgroups of G. 2. The group (2Z, +), the even integers, is a subgroup of (Z, +), the additive group of integers. 3. The group (nZ, +), the integer multiples of n, is a subgroup of (Z, +). 4. The group {I, r3 , r6 } is a subgroup of C9 , the group of rotations of a regular 9-gon. Note that this is an example of Exercise 2.6.23. 5. The groups {I, (12)}, {I, (23)} and {I, (13)} are subgroups of S3 . 6. The set SLn (F ) = {A GLn (F ) | det(A) = 1} is a subgroup of GLn (F ). This is called the n n special linear group over F . 7. The group An is a subgroup of Sn . Exercise 2.6.26 i. Suppose that G is a group and H is a nonempty subset of G. Show that H is a subgroup of G i for every a, b H , a1 b H . ii. Suppose that G is a nite group and H is a nonempty subset of G. Show that H is a subgroup of G i H is closed under multiplication. Exercise 2.6.27 Let G be a group, and let H be a subgroup of G. Fix a G, and dene aHa1 = {aba1 | b H }. Show that aHa1 is a subgroup of G. Exercise 2.6.28 Find all subgroups of C12 and S4 . Denition 2.6.29 Suppose that G1 and G2 are groups. A map : G1 G2 is a homomorphism if (ab) = (a)(b) for all a, b G1 . That is, a homomorphism preserves multiplication. In general, a homomorphism of algebraic structures is a map that preserves all the operations of that structure. For example, a homomorphism of rings (or elds) preserves addition and multiplication. If a homomorphism is a surjection, then is called an epimorphism. If a homomorphism is an injection, then is called a monomorphism. If a homomorphism is a bijection, then is called an isomorphism. Group homomorphisms have many of the properties of linear transformations. Proposition 2.6.30 Let G1 and G2 be groups and : G1 G2 a homomorphism. i. If e1 is the identity in G1 , then (e1 ) = e2 , the identity in G2 . 82

ii. If a G1 , then (a1 ) = ((a))1 . Proof. You do it. Example 2.6.31 Let G1 = G2 = (Z, +). For n N with n 2, dene n (a) = na. Then n is a homomorphism, and in fact n is a monomorphism. If we let G2 = (nZ, +), then n is an isomorphism. Exercise 2.6.32 i. Let G be a nite cyclic group of order n generated by a. Dene : G G by (aj ) = a2j . Show that is a homomorphism. Determine those values of n for which is a monomorphism, epimorphism, or isomorphism. ii. Let n be a natural number with n 2. Dene : Z Zn by (k ) = k (mod n). Show that is an epimorphism. For groups, we have a situation which is analogous to the kernel of a linear transformation in vector spaces. Denition 2.6.33 Let G1 , G2 be groups and : G1 G2 a homomorphism. The kernel of is the subset of G dened by ker = {x G1 | (x) = e2 , the identity in G2 }. Exercise 2.6.34 i. Show that ker is a subgroup of G1 . ii. Show that (G1 ) is a subgroup of G2 . iii. Find ker in the homomorphisms given in Example 2.6.31 and Exercise 2.6.32 above. The kernel of a homomorphism : G1 G2 is a subgroup with special properties. In particular, if a, b G1 and b ker , then aba1 ker . Denition 2.6.35 Let G be a group and H a subgroup of G. The subgroup H is called a normal subgroup of G if, for each a G, we have aHa1 = H . Exercise 2.6.36 i. Let G1 and G2 be groups, and let : G1 G2 be a homomorphism. Show that ker is a normal subgroup of G1 . ii. Show that any subgroup of an abelian group is a normal subgroup. iii. Show that An is a normal subgroup of Sn . iv. Show that if G is a nite group and H is a subgroup of G, then H is a normal subgroup of G if and only if aHa1 H . We nish this project by considering some important subgroups of GLn (F ) = {x Mn (F ) | det x = 0}. Exercise 2.6.37 Prove that GLn (F ) is a group with the ILC given by multiplication of matrices. Exercise 2.6.38 i. Show that the following are subgroups of GLn (F ).

ii. Determine which of the following are normal subgroups of GLn (F ). 1. The n n special linear group over F , SLn (F ) = {x GLn (F ) | det x = 1}. 83

2. The upper triangular matrices in GLn (F ), B = {(bij ) GLn (F ) | bij = 0 if i > j }. 3. The upper triangular unipotent matrices in GLn (F ), N = {(bij ) B | bjj = 1, j = 1, . . . , n}. 4. The n n diagonal matrices in GLn (F ), A = {(aij | aij = 0 if i = j }. 5. The n n orthogonal group over F , O(n, F ) = {x GLn (F ) | xt x = I }.

2.6.2 Orthogonal Transformations in Euclidean SpaceThe orthogonal transformations in Rn play an important role in analysis. Denition 2.6.39 An orthogonal transformation on Rn is a linear transformation T : Rn Rn with the property that T v, T w = v, w for all v, w Rn . We say that T preserves the bilinear form , . Exercise 2.6.40 Show that an orthogonal linear transformation is distance preserving. That is, for any v, w Rn , T v T w = v w . In particular, for any v Rn , we have T v = v . Exercise 2.6.41 Let M be the matrix of a linear transformation on Rn relative to the standard basis. Show that, for any pair of vectors v, w Rn , we have M v, w = v, t M w . Show in particular that, if M is the matrix of an orthogonal linear transformation relative to the standard basis, then M t M = t M M = I . That is, M O(n, R) (see Exercise 2.6.38). Example 2.6.42 Let M= a11 a21 a12 a22

be the matrix of an orthogonal linear transformation on E2 relative to the standard basis, that is, M 2 2 2 2 2 2 2 O(2, R). From the exercise above, it follows that a2 11 + a21 = a12 + a22 = a11 + a12 = a21 + a22 = 1. Also, a11 a12 + a21 a22 = a11 a21 + a12 a22 = 0. It is now immediate that there is some with 0 < 2 , such that M= or M= cos sin cos sin sin cos sin cos .

Proposition 2.6.43 If M O(n, R), then det M = 1. det M det t M = det M t M = det I = 1. Proof. This follows from the fact that det t M = det M , so that if M O(n, R), then (det M )2 =

The collection SO(n, R) = {M O(n, R) | det M = 1} is a subgroup of O(n, R) called the special orthogonal group . The elements of SO(n, R) are called generalized rotations. 84

which may be interpreted geometrically as a reection through the hyperplane H = {(x1 , x2 , . . . , xn ) Rn | x1 = 0}. Note that R0 O(n, R) and that det R0 = 1. With this particular orthogonal transformation so identied, we may write O(n, R) = SO(n, R) R0 SO(n, R), where R0 SO(n, R) = {R0 M O(n, R) | M SO(n, R)}. The elements of the set R0 SO(n, R) are called generalized reections and have determinant equal to 1. Exercise 2.6.44 i. If M O(2, R) and M = cos sin , show that M is a counter-clockwise rotation around the sin cos origin through an angle . Obviously, det M = 1. 1 0 cos sin cos sin = , show that M is a 0 1 sin cos sin cos reection across a line through the origin. Here, det M = 1. In particular, determine the angle that the line of reection makes with the positive x-axis. sin cos

The above proposition might lead one to believe that the special orthogonal group makes up half of the orthogonal group. This is good intuition, and it can be made precise in the following manner. Let 1 0 . . . 0 0 0 1 ... 0 0 . . . . .. . . . , R0 = . . . . . . 0 0 ... 1 0 0 0 ... 0 1

ii. If M O(2, R) and M =

iii. Finally, show that cos SO(2, R) = sin parts i and ii.

0 < 2

and that O(2, R) is the union of the matrices from

Exercise 2.6.45 Let n 3 be a natural number, and consider the regular n-gon centered at the origin in R2 with one of the vertices at the point (1, 0). We dene the dihedral group D2n to be the subgroup of O(2, R) that sends this n-gon to itself. i. What is the order of D2n ? ii. Write down the elements of D2n as matrices with respect to the standard basis. iii. Identify Cn as a subgroup of D2n . It is possible to write some rotations and reections in Rn very explicitly. For example, in O(4, R) the matrix cos sin 0 0 sin cos 0 0 M = 0 0 cos sin 0 0 sin cos

is the composition of a rotation around the x1 -x2 plane through the angle with a rotation around the x3 -x4 plane through the angle . This is an example of a generalized rotation. As an example of a generalized reection, if we have n 1 pairwise orthogonal non-zero vectors in Rn , we can produce an n-th non-zero vector v orthogonal to the original vectors by Theorem 2.5.21. In this case reection through the hyperplane spanned by the original set of n 1 orthogonal vectors is given by Tv (w) = w 2 v,w v,v v. 85

Exercise 2.6.46 i. Show that Tv is an orthogonal transformation that is a generalized reection on Rn . ii. Find the matrix of Tv relative to the standard basis. When n = 3, we can be quite precise about the nature of the elements of O(3, R). If T is an element of SO(3, R), then we can nd a line through the origin such that T is a rotation around that line. Furthermore, any reection in O(3, R) can be written as a rotation around a line in R3 through the origin combined with reection through the origin, that is, multiplication by I . We conclude this project with the symmetry groups of the regular polyhedra in R3 . The regular polyhedra are the regular tetrahedron, the cube, the regular octahedron, the regular dodecahedron, and the regular icosahedron. Since the octahedron is dual to the cube and the icosahedron is dual to the dodecahedron, we need only work with the tetrahedron, the cube, and the dodecahedron. In each case, we can obtain an upper bound on the number of symmetries by proceeding as follows. Each vertex must be mapped to a vertex and the sides adjacent to a vertex must remain adjacent to that vertex, although they can be permuted after the symmetry map is applied. For a tetrahedron, this gives an upper bound of 4 6 = 24 possible symmetries. In this case, the symmetries are in one to one correspondence with the permutations of the vertices, and the symmetry group is S4 . Exercise 2.6.47 i. Write the 24 orthogonal matrices that represent the symmetries of the tetrahedron relative to the standard basis. ii. Show that the rotations in the symmetry group of the tetrahedron form the group A4 . iii. Which of the reections in the symmetry group of a regular tetrahedron can be realized as reections through a plane? Now consider the symmetry group of the cube with vertices at the eight points (1, 1, 1). The rotations of this cube can be realized as rotations around the x-axis, rotations around the y -axis, rotations around the z -axis, rotations around the four diagonals of the cube, and nally rotations around the six lines through the origin and the midpoints of the opposite edges. Exercise 2.6.48 i. Show that rotations around a coordinate axis have order 4, rotations around a diagonal have order 3, and rotations around a line through the origin connecting the midpoints of opposite edges have order 2. ii. Write the matrices of the 24 rotational symmetries of the cube. For example, rotations around the z -axis are I , 0 R= 1 0 1 0 0 0 , 0 1

1 0 R 2 = 0 1 0 0 0 R 3 = 1 0

0 0 , and 1 1 0 0 0 . 0 1 86

If we rotate around the diagonal adjoining (1, 1, 1) and (1, 1, 1), we obtain the non-identity matrices 0 0 1 0 1 0 R = 1 0 0 and R2 = 0 0 1 . 0 1 0 1 0 0 If we rotate around the line connecting (1, 0, 1) and (1, 0, 1), then we obtain the non-identity matrix 0 0 1 R = 0 1 0 . 1 0 0 We obtain, the 24 reections in the symmetry group of a cube by multiplying the 24 rotations by I . We conclude that the symmetry group of the cube has order 48. Exercise 2.6.49 i. Show that the group of rotations of a dodecahedron has order 60 and is isomorphic to A5 . Show that the full symmetry group of a regular dodecahedron has order 120. ii. Write the matrices for the 60 rotational symmetries of a regular dodecahedron.

87

88

Chapter 3

Metric Spaces
...la notion despace m etrique fut introduite en 1906 par M. Fr echet, et d evelop ee quelques ann ees plus tard par F. Hausdor dans sa Mengenlehre. Elle acquit une grande importance apr es 1920, dune part a la suite des travaus fondamentaux de S. Banach et de son ecole sur les espaces norm es et leurs applications a lAnalyse fonctionnelle, de lautre en raison de lint er et que pr esente la notion de valeur absolue en Arithm etique et en G eom etrie alg ebrique (o` u notamment la compl etion par rapport a une valeur absolue se montre tr es f econde). N. Bourbaki, Topologie G en erale, Book 3

3.1

Introduction

We have already encountered the notion of distance in n-dimensional Euclidean space. All that this involves is the repeated use of the Pythagorean theorem. If x = (x1 , . . . , xn ) and y = (y1 , . . . , yn ) are elements in Rn , then in Chapter 2 we dened
1/2 Thus, if x = ( j =1 x2 , then d(x, y ) = x y (see Denition ??). j) In this chapter, we generalize the notion of distance in Rn to that of a metric on any set S . This new notion of distance will allow us to dene open sets, closed sets, and compact sets as we did for the real and complex numbers. n

d(x, y ) = ((x1 y1 )2 + + (xn yn )2 )1/2 .

Remark 3.1.1 Note that the metric d on Rn dened above is called the usual metric on Rn . It should be mentioned that this is not the only way to dene the distance between two points in Rn . For example, in R2 , we could dene d1 (x, y ) = |x1 y1 | + |x2 y2 |. This is sometimes called the taxicab metric. This is what happens when you are driving on city streets and are not allowed to drive through buildings or across peoples lawns. The distance d1 is illustrated below.
(x 2 , y2 ) d1 (x, y ) = |x1 y1 | + |x2 y2 | |x2 y2 | (x 1 , y1 )

n |x y1 | we will introduce in short order. 1 that There are still other notions of distance in R

89

3.2

Denition and Basic Properties of Metric Spaces

Denition 3.2.1 A metric space is a pair (X, d) where X is a set and d : X X R is a map satisfying the following properties. a. For x1 , x2 X , d(x1 , x2 ) 0, and d(x1 , x2 ) = 0 if and only if x1 = x2 , b. For any x1 , x2 X , we have d(x1 , x2 ) = d(x2 , x1 ), c. For any x1 , x2 , x3 X , we have d(x1 , x2 ) d(x1 , x3 ) + d(x3 , x2 ), (triangle inequality). Exercise 3.2.2 i. Draw a triangle and gure out why the triangle inequality is so named. ii. Replace the triangle inequality by the inequality d(x1 , x2 ) d(x1 , x3 ) + d(x2 , x3 ) for any x1 , x2 , x3 X . Show that symmetry (property 3.2.1.b) follows from this version of the triangle inequality and positive deniteness (property 3.2.1.a). Example 3.2.3 Observe that we have proved in Exercise 2.5.11 that the usual metric on Rn satises this denition. Exercise 3.2.4 On Cn = {z = (z1 , z2 , . . . , zn ) | zj C}, we dene 1/2
n

(positive denite). (symmetric).

and, for z, w Cn , we dene d(z, w) = z w . Show that d is a metric on Cn . Exercise 3.2.5 Let X be any nonempty set and, for x1 , x2 X , dene d(x1 , x2 ) = 0 if x1 = x2 , 1 if x1 = x2 .

z =

j =1

|z j |2

Show that d is a metric on X . This is called the discrete metric, the pair (X, d) is referred to as a discrete metric space. It is designed to disabuse people of the notion that every metric looks like the usual metric on Rn . The discrete metric is very handy for producing counterexamples. Exercise 3.2.6 Let (X, d) be a metric space, and let Y be a proper subset of X . Show that (Y, d ) is a metric space, where we dene d (y1 , y2 ) = d(y1 , y2 ). We call d the inherited metric on Y . Expanding on Remark 3.1.1, we introduce an important collection of metrics on Rn in the next few paragraphs. Pay attention, as these are key examples for future developments. Let p be a real number such that p 1. For x = (x1 , x2 , . . . , xn ) Rn , we dene 1/p
n

As usual, if x = (x1 , x2 , . . . , xn ) Rn and y = (y1 , y2 , . . . , yn ) Rn , we dene dp (x, y ) = x y p . Note that if p = 1 and n = 2, this is the taxicab metric that we encountered in Remark 3.1.1. Note further that if p = 2, this is the usual Euclidean distance on Rn . To show that dp is a metric on Rn , we need the following inequality: 90

j =1

|xj |p

Theorem 3.2.7 (H olders Inequality) Suppose p, q are real numbers greater than 1 such that 1/p +1/q = 1. Suppose x = (x1 , x2 , . . . , xn ) Rn and y = (y1 , y2 , . . . , yn ) Rn , then
n n 1/p n 1/q

k=1

|xk yk |

k=1

|xk |

p k=1

|y k |

Proof. The proof is based on the following inequality. Suppose a and b are positive real numbers and p bq p, q are as in the hypothesis of the theorem. Then ab a p + q . This is proved with elementary calculus. Consider the function y = xp1 , for x 0. Then the inverse function is x = y q1 , for y 0. We have p a p1 b bq dx + 0 y q1 dy = a 0 x p + q . A look at the graph of the functions reveals immediately that this sum is greater than or equal to ab, where equality holds if and only if b = ap1 , which is equivalent to saying b q = ap . Using this inequality, we get
n

k=1

|xk | |yk | ||x||p ||y ||q

k=1

|xk |p + p||x||p p

k=1

| y k |q = 1/p + 1/q = 1. q ||y ||q q

Exercise 3.2.8 Prove that dp is a metric on Rn . Hint: The triangle inequality is the only hard part. The proof of the triangle inequality depends on H olders Inequality. To begin, observe that ||x + y ||p p = Now apply H older. Exercise 3.2.9 Note that H olders inequality only works for p, q > 1. Prove the triangle inequality for the d1 metric. We also dene a metric for p = . That is, if x = (x1 , x2 , . . . , xn ), we set x dene d (x, y ) = max |xj yj | = x y
1j n .

|xi + yi |p

|xi + yi |p1 |xi | +

|xi + yi |p1 |yi |

= max1j n |xj |, and

Exercise 3.2.10 Prove that d denes a metric on Rn . Denition 3.2.11 The metric space (Rn , dp ), for 1 p , is denoted by p n (R). Exercise 3.2.12 Show that everything we have just done for Rn can also be done for Cn . This yields a collection of spaces p n (C). There is a lengthy menu attached to the study of metric spaces. For example, we need to deal with such concepts as open sets, closed sets, compact sets, accumulation points, isolated points, boundary points, interior, closure, and other things. To understand metric spaces fully, the reader must deal not only with these ideas, but with the relationships among them. Most of these ideas have a setting in the context of general topological spaces. 91

3.3

Topology of metric spaces

A fundamental notion in the study of metric spaces is that of an open ball. Denition 3.3.1 Suppose that (X, d) is a metric space and x0 X . If r R, with r > 0, the open ball of radius r around x0 is the subset of X dened by Br (x0 ) = {x X | d(x, x0 ) < r}. The closed ball of radius r around x0 is the subset of X dened by B r (x0 ) = {x X | d(x, x0 ) r}. Example 3.3.2 In R2 , with the usual metric, a ball of radius 3/2 around the point (3, 4) looks like this: 10 9 8 7 6 5 4 3 2 1 0 0 1 2 3 4 5 6 7 8 9 10 Example 3.3.3 In R2 , a ball of radius 3/2 around the point (3, 4) in the d1 metric looks like this: 10 9 8 7 6 5 4 3 2 1 0 0 1 2 3 4 5 6 7 8 9 10 Denition 3.3.4 Suppose that V is a vector space with a metric d. The unit ball in V is the ball of radius 1 with center at 0, that is, B1 (0). This denition is usually only interesting when the metric arises from a norm (see Exercises 3.2.4 and 3.2.12).
n We want to analyze the nature of the unit ball in p n (R), that is, the set of all points x R such that 2 x p < 1. For the moment, lets take n = 2 and consider the cases of 1 ( R ), ( R ), and ( R ). The pictures 2 2 2 of the unit balls in these spaces are shown below. This leads to an interesting diagram.

92

q Exercise 3.3.5 If 1 p < q , show that the unit ball in p n (R) is contained in the unit ball in n (R).

Exercise 3.3.6 Choose p with 1 p , and let > 0. Show that B (0) = { x | x B1 (0)}. Exercise 3.3.7 Consider a point x R2 that lies outside the unit ball in 1 2 (R) and inside the unit ball in ( R ). Is there a p between 1 and such that x = 1? Do the same problem in Rn . p 2 Next, we look at open sets. Denition 3.3.8 Let (X, d) be a metric space and suppose that A X . The set A is an open set in X if, for each a A, there is an r > 0 such that Br (a) A. Notice that the radius r depends on the point a. Also, observe that the empty set and the whole space X are both open sets. Exercise 3.3.9 Prove that, for any x0 X and any r > 0, the open ball Br (x0 ) is open. So now we can legitimately call an open ball an open set. Exercise 3.3.10 Prove that the following are open sets. i. The rst quadrant, that is, {(x, y ) R2 | x > 0 and y > 0}, in the usual metric; ii. any subset of a discrete metric space. Example 3.3.11 Let X = [1, 1] with the metric inherited from R. What do open balls and open sets look like in this metric space? If r 1, then Br (0) = (r, r), just as in R. If r > 1, then Br (0) = [1, 1] = X . This does not look like an open set, and in fact, as a subset of R, it is not open. However, in this metric space, it is the whole space, and we have seen that X is always open as a subset of itself. Exercise 3.3.12 Let X = [1, 1] with the inherited metric as above. Describe the open balls Br (1) for various values of r. Exercise 3.3.13 Let (X, d) be a metric space, and let Y be an open set in X . Show that every open set in (Y, d ), where d is the inherited metric, is also open in X . Open sets behave nicely under certain set-theoretic operations. Theorem 3.3.14 i. If {Aj }j J is a family of open sets in a metric space (X, d), then Aj
j J

is an open set in X ; ii. if A1 , A2 , . . . , An are open sets in a metric space (X, d), then
n

Aj
j =1

is an open set in X . Proof. i. Suppose that x j J Aj . Then x Ak for some k J . Since Ak is open, there is a real number r > 0 such that Br (x) Ak . But then, Br (x) j J Aj . 93

ii. Suppose x n j =1 Aj . Then x Aj for each j = 1, 2, . . . , n. Since Aj is open, for each j , there exists a radius rj such that Brj (x) Aj . Let r = min1j n {rj }. Then r > 0 and Br (x) n j =1 Aj .

We can now say that the collection of open sets is closed under the operations of arbitrary union and nite intersection. Exercise 3.3.15 i. There can be problems with innite intersections. For example, let An = B1/n ((0, 0)) in R2 with the usual metric. Show that An
n=1

is not open. ii. Find an innite collection of distinct open sets in R2 with the usual metric whose intersection is a nonempty open set. Thus innte interesections of open sets may or may not be open. If there are open sets in a metric space, can closed sets be far behind? Denition 3.3.16 Let (X, d) be a metric space and suppose that A X . We say that A is a closed set in X if cA is open in X . (Recall that cA = X \ A is the complement of A in X .) Exercise 3.3.17 Show that the following are closed sets. i. The x-axis in R2 with the usual metric; ii. the whole space X in any metric space; iii. the empty set in any metric space; iv. a single point in any metric space; v. any subset of a discrete metric space; vi. a closed ball B r (x0 ) in any metric space. Example 3.3.18 Let X = (1, 1) with the metric inherited from R. If r < 1, then B r (0) = [r, r], just as in R. If r 1, then Br (0) = (1, 1) = X . Despite rst appearances, this is again a closed set in X . Note 1 1 ( ) = [0, 1), another unusual-looking closed set in this metric space. also that B 2 2 Exercise 3.3.19 Let (X, d) be a metric space, and let Y be a closed set in X . Show that every closed set in (Y, d ), where d is the inherited metric, is also closed in X . Exercise 3.3.20 Show that Q as a subset of R with the usual metric is neither open nor closed in R. Of course, if the metric space is simply Q with the usual metric, then Q is both open and closed in Q. Here is a basic theorem about closed sets. Theorem 3.3.21 i. Suppose that (X, d) is a metric space and that {Aj }j J is a collection of closed sets in X . Then Aj
j J

is a closed set in X ; 94

ii. if A1 , A2 , . . . , An are closed sets in X , then

Aj
j =1

is a closed set in X . Proof. Use Theorem 3.3.14 and De Morgans laws. Exercise 3.3.22 Let (X, d) be a metric space. Let A be an open set in X , and let B be a closed set in X . Show that A \ B is open and B \ A is closed. So, a set is closed i its complement is open, and a set is open i its complement is closed. However, most of time, most sets in a metric spaces are neither open nor closed. There is a dierent way to characterize closed sets. First, we need the notion of an accumulation point. From here on, we shall simply refer to a metric space X and suppress the notation d for the metric. Denition 3.3.23 Suppose that A is a subset of a metric space X . A point x0 X is an accumulation point of A if, for every r > 0, we have (Br (x0 ) \ {x0 }) A = . Thus, if x0 is an accumulation point of A, there are points of A (other than x0 ) that are arbitrarily close to x0 . Note that, x0 may or may not be an element of A. For example, for R with the usual metric, 1 and 0 are accumulation points of the open interval (0, 1) as well as all of the points in the interval itself. Denition 3.3.24 Suppose that A is a subset of a metric space X . A point x0 A is an isolated point of A if there is an r > 0 such that Br (x0 ) A = {x0 }. Denition 3.3.25 Suppose that A is a subset of a metric space X . A point x0 X is a boundary point of A if, for every r > 0, Br (x0 ) A = and Br (x0 ) cA = . The boundary of A is the set of boundary points of A, and is denoted by A. We need some examples. Examples 3.3.26 i. Let A = {(x, y, z ) R3 | x2 + y 2 + z 2 < 1}. We take the usual metric on R3 . The set of accumulation points of A is B 3 = {(x, y, z ) | x2 + y 2 + z 2 1} and is called the closed unit ball in R3 with respect to the usual metric. The set A has no isolated points, and A = S 2 = {(x, y, z ) | x2 + y 2 + z 2 = 1}. The set S 2 is called the 2-sphere in R3 with respect to the usual metric.
2 2 n ii. Let A = {(x1 , x2 , . . . , xn ) Rn | x2 1 + x2 + . . . + xn < 1}. We take the usual metric in R . The set of 2 2 n 2 accumulation points of A is B = {(x1 , x2 , . . . , xn ) | x1 + x2 + . . . + xn 1}. The set A is called the open unit ball with respect to the usual metric and the set B n is called the closed unit ball in Rn with respect to the usual metric. The set A has no isolated points and A = S n1 = {(x1 , x2 , . . . , xn ) | 2 n1 2 is called the (n 1)-sphere in Rn with respect to the usual x2 1 + x2 + . . . + xn = 1}. The set S metric.

iii. Let A = Q R with the usual metric. Then every point in R is an accumulation point of A, the set A has no isolated points, and A = R. iv. If A is any subset of a discrete metric space X , then A has no accumulation points. Every point in A is an isolated point, and A = .
1 | n N} R with the usual metric. Then every point of A is an isolated point and a v. Let A = { n boundary point, the point 0 is the only accumulation point of the set, and the set A is neither open nor closed.

95

Now, we have another means of identifying closed sets, that is, accumulation points. Theorem 3.3.27 Suppose A is a subset of a metric space X . Then A is closed i A contains all its accumulation points. Proof. If A is the empty set, then A has no accumulation points. Suppose that A is a non-empty closed set and that x0 is an accumulation point of A. If x0 / A, then x0 cA, which is open. Hence, there is an c r > 0 such that Br (x0 ) A, and this contradicts the denition of accumulation point. Conversely, suppose that A contains all its accumulation points and that x0 cA. Then x0 is not an accumulation point of A, and hence there exists r > 0 such that Br (x0 ) A = . This means that cA is open, and so A is closed. In a discrete metric space any subset is both open and closed. This is not generally the case. For example, p in the case of p n (R) and n (C), most subsets are neither open nor closed.
p Exercise 3.3.28 Find an uncountable number of subsets of p n (R) and n (C) that are neither open nor closed.

If a set A in a metric space X is not closed and we wish that it were, then we can do something about it. Denition 3.3.29 Suppose that A is a nonempty subset of a metric space X . The closure of A is the intersection of all the closed sets which contain A. The closure of any set A exists, since there are always closed sets that contain A, for example X . The closure of A is a closed set since it is the intersection of closed sets. So the closure of A is the smallest closed set that contains A. We denote the closure of a set A by A. Obviously, A A and A = A i A is closed. Examples 3.3.30 i. Let A = {(x, y, z ) R3 | x > 0, y > 0, z > 0}. If R3 has the usual metric, then A = {(x, y, z ) R3 | x 0 , y 0 , z 0 } . ii. Let Qn = {(x1 , x2 , . . . , xn ) Rn | xj Q for 1 j n}. If Rn has the usual metric, then Qn = Rn . iii. Let X be a discrete metric space and let A be any subset of X . Then A = A. It should not come as a surprise that the notions of closure and accumulation point are intimately related. Exercise 3.3.31 Suppose that A is a subset of a metric space X . Show that A = A{accumulation points of A}. Exercise 3.3.32 Suppose A is a subset of a metric space X . Prove or disprove: A = A A. Exercise 3.3.33 Suppose A is a subset of a metric space X . Prove that A = A c A. Exercise 3.3.34 Let X be a metric space and let x0 X . Suppose that r > 0. Prove or disprove: Br (x0 ) = {x X | d(x, x0 ) r}. Exercise 3.3.35 For denitions and notations for this exercise, see Project 2.1. i. Consider the set of 2 2 matrices over R, that is, M2 (R). Make this into a metric space by identifying it with R4 with the usual metric. Show that GL2 (R) is an open subset of M2 (R), and that GL2 (R) = M2 (R). ii. Show that SL2 (R) is a closed subset of GL2 (R) Exercise 3.3.36 Let A be a subset of a metric space X and let x0 be an isolated point of A. Show that x0 is in the boundary of A if and only if x0 is an accumulation point of c A. 96

Corresponding to the notion of closure is the idea of the interior of a set. Denition 3.3.37 Let A be a subset of a metric space X . The interior of A is the union of all open sets which are contained in A. The interior of A is the largest open set contained in A. We denote the interior of A by A . Obviously A A and A = A i A is open.

Examples 3.3.38 i. Let X = R3 with the usual metric and A = {(x, y, z ) | z 0}. Then A = {(x, y, z ) | z > 0}; ii. let X be a discrete metric space and let A be any subset of X . Then A = A and A = A, so that A = A = A. Exercise 3.3.39 Show that, in the usual metric on R, the interior of Q is empty, that is, Q = , but the the interior of Q is R, that is, (Q) = R. Exercise 3.3.40 Look at combinations of interior, closure, and boundary and determine how many dierent possibilities result. For this exercise only, let I stand for interior, B stand for boundary, and C stand for closure. Let X be a metric space and let A X . How many possible sets can be made from A with these operations? For example, I (I (A)) = I (A), but C (I (A)) is not necessarily A. Is it C (A)? Explore all possibilities of applying combinations of I ,C , and B . Hint: There are only a nite number. Another important concept in the theory of metric spaces is that of diameter. Denition 3.3.41 Let A be a nonempty subset of a metric space X . The diameter of A is diam(A) = sup d(x, y ).
x,y A

Note that we may have diam(A) = . Exercise 3.3.42 i. Show that the diameter of a set is 0 i the set consists of a single point. ii. Suppose A is a nonempty subset of a metric space X . Show that diam(A) = diam(A). Denition 3.3.43 Let A be a nonempty subset of Rn . We say that A is convex if, given any two points p, q A, the line segment with endpoints p and q, that is, the set {(1 t)p + tq | t R, 0 t 1}, is a subset of A. Example 3.3.44 The unit ball B n contained in Rn , in the usual metric, is a convex set.
n Exercise 3.3.45 Show that the unit ball p n (R), for 1 p , is a convex set in R .

Denition 3.3.46 Let A be a subset of Rn with the usual metric. The convex hull of A is the intersection of all convex sets containing A. The closed convex hull of A is the intersection of all closed convex sets containing A. Exercise 3.3.47 Let A be a nonempty subset of Rn and let C be the convex hull of A. i. Prove or disprove the following statement. The closed convex hull of A is C . 97

ii. Show that the diameter of A is the diameter of C . Remark 3.3.48 The concept of convex set in Rn does not involve a metric in Rn . However a particular metric is often used to dene subsets of Rn that may or may not be convex. Exercise 3.3.49 i. Describe of the closed convex hull of the unit ball in p n (R) for 1 p . ii. Suppose 0 < p < 1. For x Rn , dene,
n
1 p

x Dene Sp = {x Rn | x of Sp .

=
k=1

|xk |

1}. Determine whether Sp is convex. If not, nd the closed convex hull

Example 3.3.50 We now branch out in a slightly dierent direction. Suppose that X is a set and and F = R or C. Denote by B (X, F ) the set of all bounded functions from X to F . Thus, f B (X, F ) i there is a real number M such that |f (x)| M for all x X . For f B (X, F ), we dene f = supxX |f (x)| (the sup norm). For f, g B (X, F ), we dene d(f, g ) = supxX |f (x) g (x)| (the sup metric). It follows easily from the denition of sup that d is a metric on B (X, F ). In this example, an open ball of radius r around a function f is the collection of all functions which lie within an r-strip around f . Exercise 3.3.51 i. Let F = R or C. Show that B (X, F ), with d as dened above, is a metric space. ii. For f, g B (X, F ), dene (f + g )(x) = f (x) + g (x) and (f g )(x) = f (x)g (x). Also, for F dene (f )(x) = f (x). Show that, with these operations, B (X, F ) is a commutative algebra with 1 over F (see Denition 2.3.7). Of course, scalar multiplication is simply multiplication by a constant function. This is a step up in our examples of metric spaces. While previous examples are important, spaces of functions are the most signicant examples of metric spaces in analysis.

3.4

Convergence and Completeness

Our next big idea is convergence in a metric space. When we discussed the convergence of a sequence in R or C, we used the absolute value to measure the distance between two points in one of these elds. Here, in a general metric space, we can use the metric to accomplish the same thing. Denition 3.4.1 Suppose (an )nN is a sequence of points in a metric space X . We say that a point L X is the limit of the sequence (an )nN as n goes to innity if, for any > 0, there exists N N such that d(an , L) < whenever n N . When the limit exists, we say that (an )nN converges to L, and we write
n

lim an = L.

Sometimes, we simply say that (an )nN converges in X without mentioning L explicitly. As in Chapter 1, we have a concept of Cauchy sequences in a metric space. Denition 3.4.2 (See 1.5.3, 1.6.22.) Let X be a metric space and let (an )nN be a sequence in X . We say that (an )nN is a Cauchy sequence if, for any > 0, there exists N N such that d(an , am ) < whenever n, m N . 98

It may be that a sequence in a metric space is a Cauchy sequence even though it does not converge. For example, as we observed in Chapter 1, Cauchy sequences in Q with the usual metric do not necessarily converge in Q. This leads us to the following exercise. Exercise 3.4.3 Suppose that X is a metric space and that the sequence (an )nN converges in X . Show that for any > 0, there exists N N such that d(an , am ) < whenever n, m N . Thus, a convergent sequence is a Cauchy sequence. Exercise 3.4.4 Let (an )nN be a Cauchy sequence in a discrete metric space X . Show that there exists N N such that d(an , am ) = 0, that is, an = am for all n, m N . Hence, the sequence is convergent. Such a sequence is called eventually constant. Note that an eventually constant sequence in any metric space is convergent, and in fact, it converges to the eventual constant. There is a standard litany associated to the notions of convergent sequence and Cauchy sequence. For example, from Exercise 3.4.3, we see that in any metric space, a convergent sequence is a Cauchy sequence. In R or C with the usual metric, every Cauchy sequence converges. In Q with the usual metric, many Cauchy sequences do not converge. The best kinds of metric spaces (at least our favorite kinds) are the ones in which convergent sequence and Cauchy sequence are synonymous. Denition 3.4.5 Suppose that X is a metric space. We say that X is a complete metric space if every Cauchy sequence in X converges. Examples 3.4.6 The following metric spaces are complete. (If this seems repetitive, dont worry about it.) The proofs of (i) and (ii ) are in Chapter 1. i. R with the usual metric; ii. C with the usual metric; iii. any discrete metric space. Examples 3.4.7 The following metric spaces are not complete.

i. Q with the usual metric; ii. any proper nonempty open subset of R with the inherited metric. Exercise 3.4.8 Prove that a subset Y of a complete metric space X is also complete metric space with the inherited metric if and only if Y is closed as a subset of X .
p Exercise 3.4.9 Show that, for 1 p , the spaces p n (R) and n (C) are complete metric spaces.

We now turn to an investigation of convergence in the spaces Rn and Cn with the usual metrics. Our approach here is very similar to the one that we took in R. One big dierence is that, since there is no notion of order in Rn , the idea of monotonicity has no meaning. However, we will use it one coordinate at a time. Similarly, we need to talk about bounded sets in Rn . Denition 3.4.10 A subset A of a metric space X is bounded if X = or there exists a point x X and r > 0 such that A Br (x). Exercise 3.4.11 Let F = R or C, and let X = F n with the usual metric. Show that a subset A of X is bounded if and only if each of the n sets of the form Ai = {x F | a = (a1 , a2 , . . . , an ) A with ai = x} are bounded as subsets of F . The following lemmas and theorems for Rn and Cn will be proved for Rn and left as exercises for Cn . Lemma 3.4.12 Every bounded sequence in Rn (or Cn ) with the usual metric has a convergent subsequence. 99

Proof. Let (am )mN be a bounded sequence in Rn . Write am = (am,1 , am,2 , . . . , am,n ). We prove the Lemma by induction on n. For n = 1, this is the content of Lemma 1.6.15. Assume the lemma is true for n 1. Let a m be the (n 1)-tuple (am,1 , am,2 , . . . , am,n1 ). Then (am )mN is a bounded sequence in n1 R . By the induction hypothesis (am )mN has a convergent subsequence in Rn1 . Label this convergent subsequence (a mj )j N . Now the sequence (amj ,n )j N is a bounded sequence in R and hence has a convergent subsequence which we shall not name. Again, taking the corresponding subsequence of (amj )j N , we get a convergent subsequence of the original subsequence (am )mN . Exercise 3.4.13 i. For practice, carry out the above proof in Cn . ii. Prove the above lemma by proceeding coordinate by coordinate. You will notice that the indexing gets quite messy. Theorem 3.4.14 (Bolzano-Weierstrass) If A is a bounded innite subset of Rn or Cn , then A has an accumulation point. Proof. (Well do Rn . You do Cn .) Since A is innite there exists a sequence (xk )kN in A, where xk = xj if k = j . Then (xk )kN is a bounded sequence in Rn and by Lemma 3.4.12 has a convergent subsequence. If this subsequence converges to x0 , then x0 is an accumulation point of A. One of the most important contexts in which to discuss the convergence of sequences is when we consider sequences of functions. There is more than one notion of what it means for a sequence of functions to converge. Below, we discuss two of the most important of these notions, namely pointwise convergence and uniform convergence. We do this in the case of sequences of bounded functions from a set X to R or C, as in Example 3.3.50. The most na ve notion of convergence for a sequence of functions is pointwise convergence. Denition 3.4.15 Let X be a set, and let F = R or C. Consider a sequence of functions (fn )nN , where fn : X F is a bounded function for each n N. We say that a function f : X F is the pointwise limit of the sequence (fn )nN if, for every x X , limn fn (x) = f (x). Example 3.4.16 Let fn : [0, 1] R be given by fn (x) =
1 , 0 if 0 x 1 n 1 nx (n 1) if 1 n < x 1.

1 Note that for all x < 1, the sequence (fn (x))nN is eventually constant; namely, if n > 1 x , then fn (x) = 0. When x = 1, fn (1) = 1 for all n N. Thus we can conclude that the pointwise limit of the sequence (fn )nN is the function f : [0, 1] R given by

f (x) =

0 if 0 x < 1, 1 if x = 1.

(The astute reader will have noticed that each of the functions fn is continuous, while the pointwise limit function f is notmore on this later.) Example 3.4.17 Let fn : (0, 1) R be given by fn (x) = 0
1 x 1 , if 0 < x < n 1 if n x < 1.

100

1 1 Note that for all x (0, 1), the sequence (fn (x))nN is eventually constant; namely, if n > x , then fn (x) = x . Thus we can conclude that the poinwise limit of the sequence (fn )nN is the function f : (0, 1) R given by 1 . (The astute reader will have noticed that, similarly to the previous example, each of the functions f (x) = x fn is bounded, while the pointwise limit function f is not.)

Exercise 3.4.18 For the following sequences (fn )nN of functions, where fn : [0, 2 ] R for all n N, nd all values of x [0, 2 ] such that the sequence (fn (x))nN converges, and nd the pointwise limit function f : [0, 2 ] R if it exists. i. fn (x) = sin
x n

ii. fn (x) = sin(nx) iii. fn (x) = sinn x The other major notion of convergence for sequences of functions that we will discuss is uniform convergence. This notion of convergence utilizes the metric on B (X, F ) dened in Example 3.3.50. Denition 3.4.19 Let (fn )nN be a sequence of functions in B (X, F ), where F = R or C. The sequence is said to converge uniformly to a function f B (X, F ) provided that, given > 0, N N such that supxX |fn (x) f (x)| < for n N . Remark 3.4.20 Note that uniform convergence is convergence in the metric space B (X, F ). However, pointwise convergence is not in general given by convergence in a metric space. Exercise 3.4.21 Show that if a sequence (fn )nN converges uniformly to a function f , then it converges pointwise to the function f . The converse to the preceding idea is false, as we indicate in the following exercise. Exercise 3.4.22 Let fn (x) = xn for n N. i. Show that the sequence (fn )nN converges pointwise to the function f (x) = 0 on the interval (1, 1).
1 1 ii. Show that if we restrict to the domain [ 2 , 2 ], the sequence (fn )nN converges uniformly to the function f (x) = 0.

iii. Show that the sequence (fn )nN does not converge uniformly on the domain (1, 1). We now ask whether a Cauchy sequence (fn )nN in B (X, F ) converges uniformly to its pointwise limit f . Theorem 3.4.23 The spaces B (X, R) and B (X, C) are complete metric spaces. Proof. As above, we consider B (X, F ) where F = R or C. Suppose that (fn )nN is a Cauchy sequence in B (X, F ). Then, for every x X , the sequence (fn (x))nN is a Cauchy (and hence convergent) sequence in F because |fn (x) fm (x)| supyX |fn (y ) fm (y )|. Thus, the sequence (fn )nN has a pointwise limit f : X F . We want to show that the sequence (fn )nN converges uniformly to f . To this end, let > 0. Choose N N such that supyX |fn (y ) fm (y )| < /2 when n, m N . Fix x X and choose an integer N (x) N such that |fN (x) (x) f (x)| < /2. Then |fn (x) f (x)| |fn (x) fN (x) (x)| + |fN (x) (x) f (x)| < if n N . To complete the proof, we must show that the function f is bounded, that is, f B (X, F ). But, from

the above inequality, it follows that |f (x)| < |fn (x)| + for all x X . 101

3.5

Continuous functions

We now turn to one of the most important ideas about functions on metric spaces. Denition 3.5.1 Let (X, d) and (X , d ) be metric spaces. A function f : X X is continuous at the point x0 X if, for any > 0, there is a > 0 such that d (f (x), f (x0 )) < whenever x X and d(x, x0 ) < . This is the old familiar - denition. It is simply the statement that
xx0

lim f (x) = f (x0 ).

More generally, the limit of a function f (x) at x0 is L X , and we write


xx0

lim f (x) = L,

if, for every > 0, there exists a > 0 such that d (f (x), L) < whenever 0 < d(x, x0 ) < . Exercise 3.5.2 Suppose that X and X are metric spaces as above and that x0 X . Show that f is continuous at x0 i for every sequence (xn )nN in X which converges to x0 in X , we have
n

lim f (xn ) = f (x0 )

in X . Note that another way of saying that f is continuous at x0 is the following: given > 0, there exists > 0 such that f (B (x0 )) B (f (x0 )). Exercise 3.5.3 Let f : R R be a polynomial function, where R has the usual metric. Show that f is continuous. In discussing continuity, one must be careful about the domain of the function. For example, dene f : R R by the equation 0 if x / Q, f (x) = 1 if x Q. Then, f is not continuous at any point of R. However, suppose we restrict f to be a function from Q to Q. This means that f (x) = 1 on Q and is continuous at every point of Q. Exercise 3.5.4 Dene f : R R by f (x) = 1/q 0 if x = p/q (reduced to lowest terms, x = 0), if x = 0 or x / Q.

Show that f is continuous at 0 and any irrational point. Show that f is not continuous at any nonzero rational point. Continuity is called a pointwise property or local property of a function f , that is, as in Exercise 3.5.4, a function may be continuous at some points, but not at others. We often deal with functions f : X X which are continuous at every point of X . In this case, we simply say that f is continuous without reference to any particular point. Theorem 3.5.5 Suppose that (X, d) and (X , d ) are metric spaces. Then a function f : X X is continuous i for any open set V X , the set f 1 (V ) is an open set in X . 102

Proof. First suppose that f is continuous. Let V be an open set in X . Suppose x0 f 1 (V ). Take > 0 such that B (f (x0 )) V . Then there exists > 0 such that f (B (x0 )) B (f (x0 )), and so B (x0 ) f 1 (B (f (x0 ))) f 1 (V ). So f 1 (V ) is open. The second half of the proof is easy. You do it.

Corollary 3.5.6 Suppose that (X, d), (X , d ), and (X , d ) are metric spaces, and f : X X and g : X X are continuous. Then g f : X X is continuous. Proof. This follows immediately from the theorem. Exercise 3.5.7 Prove Corollary 3.5.6 directly from the denition, that is, without using Theorem 3.5.5. Exercise 3.5.8 i. Let X and X be metric spaces and assume that X has the discrete metric. Show that any function f : X X is continuous. ii. Let X = R with the usual metric and let X be a discrete metric space. Describe all continuous functions from X to X . Exercise 3.5.9 Suppose that (X, d) and (X , d ) are metric spaces and that f : X X is continuous. For each of the following statements, determine whether or not it is true. If the assertion is true, prove it. If it is not true, give a counterexample. i. If A is an open subset of X , then f (A) is an open subset of X ; ii. if A is a closed subset of X , then f (A) is a closed subset of X ; iii. if B is a closed subset of X , then f 1 (B ) is a closed subset of X ; iv. if A is a bounded subset of X , then f (A) is a bounded subset of X ; v. if B is a bounded subset of X , then f 1 (B ) is a bounded subset of X ; vi. if A X and x0 is an isolated point of A, then f (x0 ) is an isolated point of f (A); vii. if A X , x0 A, and f (x0 ) is an isolated point of f (A), then x0 is an isolated point of A; viii. if A X and x0 is an accumulation point of A, then f (x0 ) is an accumulation point of f (A); ix. if A X , x0 X , and f (x0 ) is an accumulation point of f (A), then x0 is an accumulation point of A. Exercise 3.5.10 Do any of your answers in the previous exercise change if we assume X and X are complete? Denition 3.5.11 Let (X, d) and (X , d ) be metric spaces. A function f : X X is a homeomorphism if a. f is a bijection, b. f is continuous, and c. f 1 is also continuous.
1 : R ( Example 3.5.12 The function tan : ( 2 , 2 ) R is a homeomorphism, with inverse tan 2 , 2 ).

103

Example 3.5.13 Let X = [0, 1) with the induced metric from R, and let X = T = {z C | |z | = 1} with the induced metric from C. The function f : X X , f (x) = e2ix is a continuous bijection whose inverse is not continuous. Exercise 3.5.14 If (X, d) is a metric space, then the function I (x) = x is a homeomorphism from X to itself. Exercise 3.5.15 Let X = R with the discrete metric, and let X = R with the usual metric. Show that the function I : X X , I (x) = x is a continuous bijection but is not a homeomorphism.
q Theorem 3.5.16 Suppose 1 p < q . Then the identity map I (x) = x from p n (R) to n (R) is a homeomorphism. q Proof. From Exercise 3.3.5, the image under I of unit ball in p n (R) is contained in the unit ball in n (R). q Furthermore, by Exercise 3.3.5, the image under I of B (0) p ( R ) is contained in B (0) ( R ). Thus, if n n > 0, choosing = shows that I is continuous at 0. Now take (x1 , x2 , . . . , xn ) Rn and suppose max1in {|xi |} 1. Then |x1 /n| + + |xn /n| 1. This 1 shows that the ball of radius 1/n in the n metric is contained in the ball of radius 1 in the n metric. In q particular, this last fact shows that if we take the unit ball in n (R) and multiply each coordinate by a factor of 1/n, then the resulting set of points is contained in the unit ball in p n (R). By Exercise 3.3.5, this means q p 1 1 (0) (R) is contained in B1 (0) (R). Similarly, the image under I that the image under I 1 of B n of n n 1 q p (0) (R) is contained in B (0) (R). Thus, if > 0, choosing = Bn is continuous n n n shows that I at 0. A similar argument to the above works at all other points.

q Exercise 3.5.17 Show that p n (C) and n (C) are homeomorphic.

Exercise 3.5.18 Let (X, d) be a metric space, and for any x, y X , let d (x, y ) = i. Show that d denes a metric on X.

d(x,y ) 1+d(x,y ) .

ii. Show that the identity map I : (X, d) (X, d ), I (x) = x, is a homeomorphism. iii. If (X, d ) is complete, is (X, d) necessarily complete? This exercise is intended to illustrate that, without additional structure, metric spaces can be twisted, expanded, or shrunken without disturbing the open sets too badly. Denition 3.5.19 Let (X, d) and (X , d ) be metric spaces. A homeomorphism f : X X is an isometry if d (f (x1 ), f (x2 )) = d(x1 , x2 ) for all x1 , x2 X . Exercise 3.5.20 Suppose that, instead, we had dene an isometry to be a bijection f : X X such that d (f (x1 ), f (x2 )) = d(x1 , x2 ) for all x1 , x2 X . Show that with this denition, any isometry is a homeomorphism. Exercise 3.5.21 Let X = R2 with the usual metric. Show that the following functions are isometries from X to itself 1. Translation by the vector (a, b) in R2 : T(a,b) (x, y ) = (x + a, y + b) for xed a, b R 2. Counterclockwise rotation about the origin by an angle : R (x, y ) = (x cos y sin , x sin + y cos ) for xed R 104

3. Reection over a line through the origin making an angle with the x-axis: S (x, y ) = (x cos 2 + y sin 2, x sin 2 y cos 2) for xed R Exercise 3.5.22 Show that the function Da : Rn Rn given by Da (x1 , x2 , . . . , xn ) = (ax1 , ax2 , . . . , axn ) for xed a R is an isometry if and only if a = 1. Exercise 3.5.23 In this exercise, we consider isometries from R to itself in the usual metric. i. Is f (x) = x3 a bijection? A homeomorphism? An isometry? ii. Is f (x) = x + sin x a bijection? A homeomorphism? An isometry? iii. Find all isometries from R to itself. Exercise 3.5.24 For denitions and notations for this exercise, see Project 2.1. Let (X, d) be a metric space. Let G be the collection of all homeomorphisms from X to X . Prove that, under composition of functions, G is a group, and the collection of all isometries is a subgroup of G. Denition 3.5.25 Suppose that (X, d) is a metric space, and let F = R or C. Dene BC (X, F ) to be the subset of B (X, F ) consisting of continuous functions from X to F . We take the metric on BC (X, F ) to be the induced metric from B (X, F ). If X is compact, then all continuous functions from X to F are bounded (see Exercise 3.6.9 below); so, when X is compact, we will sometimes write C (X, F ) in place of BC (X, F ). Theorem 3.5.26 The space BC (X, F ) is a complete metric space. Proof. Suppose that (fn )nN is a Cauchy sequence in BC (X, F ). Then by Theorem 3.4.23, (fn )nN converges to a function f B (X, F ). All we need to show is that f is a continuous function. Now, given > 0, there exists N such that supxX |fn (x) f (x)| < /3 whenever n N . Fix x0 X . Then, for any x X and n N, |f (x) f (x0 )| = |f (x) + (fn (x) + fn (x)) + (fn (x0 ) + fn (x0 )) f (x0 )| |f (x) fn (x)| + |fn (x) fn (x0 )| + |fn (x0 ) f (x0 )|

< /3 + |fn (x) fn (x0 )| + /3. then |f (x) f (x0 )| < when d(x, x0 ) < , so f is continuous.

Since fn is continuous, we can choose > 0 such that |fn (x) fn (x0 )| < /3 whenever d(x, x0 ) < . But

Remark 3.5.27 So we have proved (see Theorems 3.4.23 and 3.5.26) that the uniform limit of bounded functions is a bounded function and the uniform limit of bounded continuous functions is a bounded continuous function. We will nd these facts very useful in doing analysis. Exercise 3.5.28 Let (X, d) be a metric space, and let F = R or C. Recall that B (X, F ) is a commutative algebra with 1 over F (see Exercise 3.3.51). Show that BC (X, F ) is a subalgebra of B (X, F ), that is, BC (X, F ) is a vector subspace of B (X, F ) that is closed under pointwise multiplication. Exercise 3.5.29 In Exercise 3.4.22, we saw that the pointwise limit of the sequence of functions (fn )nN , fn (x) = xn on (1, 1), is continuous even though the convergence was not uniform. Now consider the same sequence of functions dened on [1, 1]. Find the pointwise limit of the sequence (fn )nN and show that it is not continuous. Exercise 3.5.30 Dene a sequence of functions fn : (0, 1) R by fn (x) =
1 qn

if x = p q (reduced to lowest terms, x = 0), otherwise,

for n N. Find the pointwise limit f of the sequence (fn )nN and show that (fn )nN converges to f uniformly. 105

There is an additional property of continuous functions which is important for future applications. Denition 3.5.31 Let (X, d) and (X , d ) be metric spaces, and let f be a continuous function from X to X . We say that f is uniformly continuous if, given > 0, there exists > 0 such that, for any pair x, y X , we have d (f (x), f (y )) < whenever d(x, y ) < . So, f is uniformly continuous if it is continuous at every point and, for a given > 0, we can nd a corresponding that is independent of the point. Exercise 3.5.32 deg(p(x)) < 2. i. Show that a polynomial function p(x) on R is uniformly continuous if and only if

ii. Show that f (x) = sin(x) is uniformly continuous on R. Exercise 3.5.33 Let X = (0, ) and determine whether the following functions are uniformly continuous on X :
1 i. f (x) = x ; ii. f (x) = x;

iii. f (x) = ln(x); iv. f (x) = x ln(x). Exercise 3.5.34 Show that any linear map from Rn to Rm is uniformly continuous.

3.6

Compactness and Connectedness

In Chapter 1, we considered compact subsets of R and C. Now, with the topology we have in a metric space, we can generalize this idea. Denition 3.6.1 Let A be a subset of a metric space X . A family {Uj }j J of open subsets of X is called an open covering (or open cover ) of A if A Uj .
j J

If {Uj }j J is an open cover of A, we say that this cover has a nite subcovering or (nite subcover ) if there is a nite subcollection Uj1 , Uj2 , . . . , Ujn satisfying
n

A Examples 3.6.2

Ujk .
k=1

i. Let A = (0, 1) R with the usual metric. For j N, j 2, dene Uj = ( 1 j , 1). Then A j N Uj , but there is no nite subcover. ii. Let B = [0, ) R with the usual metric. For j N, dene Uj = (1, j ). Then B j N Uj , but there is no nite subcover. iii. Let X be a discrete metric space. For any point j X , set Uj = {j }. Then {Uj }j X is an open cover of X which has a nite subcover i X is a nite set. iv. We have seen in Theorem 1.6.38 that if A is a closed and bounded set in R with the usual metric, then every open cover of A has a nite subcover. 106

Denition 3.6.3 (See Denitions 1.6.41 and 1.8.20.) Let A be a subset of a metric space X . We say that A is compact if every open covering of A has a nite subcovering. Recall that the Heine-Borel Theorem in R or C states that a subset of R or C with the usual metric is compact if and only if it is closed and bounded. The statement of the Heine-Borel theorem is certainly not true in a general metric space. For example, take R with the discrete metric. Then, R is closed and bounded in this metric. Take an open covering consisting of the individual points in R. This covering does not have a nite subcovering. For emphasis, we note that the denition insists that for every open covering, there must be a nite subcovering. For example, given any subset A of a metric space X , we have that {X } is an open covering which is already nite. So while this particular open covering has a nite subcovering, this does not necessarily imply that other open coverings have nite subcoverings. Hence, in a general metric space, the closed bounded sets are not necessarily compact. However, we do have one half of the statement of the Heine-Borel theorem in general metric spaces. Theorem 3.6.4 If a subset A of a metric space X is compact, then A is closed and bounded. Proof. Recall that a set in a metric space is bounded if and only if it is contained in a ball of nite radius with center at some point. If A is non-empty, take a point a A and consider the open covering {Bn (a) | n N}. Since A is compact, this cover has a nite subcovering, and in fact there is an integer N such that A BN (a). Hence, A is bounded. To prove that A is closed, we assume that x0 is an accumulation point of A and prove that x0 A. Suppose not. Then for each a A, let ra = d(a, x0 )/2. But the collection {Bra (a) | a A} is an open cover of A and hence has a nite subcover {Br1 (a1 ), Br2 (a2 ), . . . , Brn (an )}. Let r = min{r1 , r2 , . . . , rn }. Then Br (x0 ) Brj (aj ) = for all j . Hence, Br (x0 ) A = , which contradicts the denition of accumulation point. Hence, x0 A. Corollary 3.6.5 If A is a compact set in a metric space X , then every innite subset of A has an accumulation point in A. Proof. Suppose that A is a compact set and that C is an innite subset of A with no accumulation point in A. Then, for each a A, there is an open ball B (a) centered at a such that (B (a) \ {a}) C = . The collection {B (a) | a A} covers A. So, by compactness, we can extract a nite subcover, {B (a1 ), . . . , B (an )}. Thus, C A B (a1 ) B (an ), and each B (aj ) contains at most one element of C (at its center). This

implies that C has at most n elements, which is a contradiction.

Corollary 3.6.6 Let A be a compact set in a metric space. Then, every innite sequence in A has a subsequence that converges to a point in A. Exercise 3.6.7 Prove that the Heine-Borel theorem holds in Rn and Cn with the usual metrics. (Hint: See the proof of Theorem 1.8.21.) Exercise 3.6.8 i. Show that a nite union of compact sets is compact. ii. Give an example of a countable union of compact sets that is not compact. iii. Show that a closed subset of a compact set is compact. Exercise 3.6.9 107

i. Let f : X X be a continuous map of metric spaces. Show that if A X is compact, then f (A) X is compact. ii. Suppose that X is a compact metric space. Show that a continuous function f : X R (R with the usual metric) is bounded. iii. Suppose that X is a compact metric space. Show that a continuous function f : X R (R with the usual metric) attains a maximum and minimum value on X . Exercise 3.6.10 Suppose X and X are metric spaces with X compact. i. If f : X X is continuous on X , show that f is uniformly continuous on X . ii. If f : X X is a continuous bijection, show that f is a homeomorphism. Exercise 3.6.11 (Dinis Theorem) Let X be a compact metric space. Suppose f and (fn )nN are realvalued continuous functions on X . Suppose that, for each x X , the sequence (fn (x))nN is a monotonic sequence converging to f (x). Show that (fn )nN converges to f uniformly. Exercise 3.6.12 Suppose that A and B are nonempty subsets of a metric space X . The distance between A and B is dened by d(A, B ) = inf {d(a, b) | a A, b B }. We say that d(A, B ) is assumed if there exists a0 A and b0 B such that d(A, B ) = d(a0 , b0 ). Determine whether or not the distance between A and B is necessarily assumed in (i)(iii). i. A is closed and B is closed; ii. A is compact and B is closed; iii. A is compact and B is compact. iv. What happens in the above cases if we assume X is complete? Exercise 3.6.13 Let X be a metric space, let A X be compact, and let U X be an open set containing A. Show that there exists an open set W X containing A such that W is compact and W U . (Hint: Consider the previous exercise with A and U .) At this point we introduce an alternate notion of compactness. Denition 3.6.14 (See Denition 1.6.45.) A subset A of a metric space X is sequentially compact if every sequence in A has a subsequence that converges to an element of A. Exercise 3.6.15 If X is a metric space, and A X , we say that A is totally bounded if, for any > 0, A can be covered nite number of balls of radius . Show that a sequentially compact metric space is totally bounded. One of the most important facts about metric spaces is that compactness and sequential compactness are equivalent. We have already proved (see Corollary 3.6.6) that compactness implies sequential compactness. To prove the converse, we need the following lemma. Lemma 3.6.16 Let X be a metric space. If A X has the property that every innite subset of A has an accumulation point in X , then there exists a countable collection of open sets {Ui | i N} such that, if V is any open set in X and x A V , then there is some Ui such that x Ui V . 108

Proof. We claim that, for each n N, there is a nite set of points xn,1 , . . . , xn,N (n) in A such that the set 1 (xn,1 ), B 1 (xn,2 ), . . . , B 1 (xn,N (n) ) covers A. If A is nite, this is clearly true, so we assume of open balls B n n n A is innite. 1 Suppose our claim is false. Then, there exists n N such that no nite collection of balls of radius n centered at points of A can cover A. For each k N, dene an innite sequence of points of A inductively as 1 (y1 )} does not cover A. So choose y2 A \ B 1 (y1 ). Then {B 1 (y1 ), B 1 (y2 )} follows. Take y1 A. Then {B n n n n 1 1 (y1 ), . . . , B 1 (yk )} . Assume y1 , . . . , yk have been chosen such that {B n does not cover A and d(y1 , y2 ) n n 1 1 (y1 ) B 1 (yk )). The innite for all i = j . Choose yk+1 A \ (B n does not cover A, and d(yi , yj ) n n sequence (yk )kN does not have an accumulation point anywhere, which contradicts our assumption about A. 1 (xn,j ) | n N and 1 j N (n)} gives the required countable collection. Taking all these balls {B n Exercise 3.6.17 Verify that the above collection satises the conclusion of the lemma. Exercise 3.6.18 Let X be a metric space. If A X has the property that every innite subset of A has an accumulation point in A, show that for any open covering of A, there exists a countable subcovering. Now comes a major Theorem. Theorem 3.6.19 In any metric space, a subset A is compact if and only if it is sequentially compact. Proof. We have already proved above that compactness implies sequential compactness. For the converse, suppose that A X is sequentially compact. Then any innite subset of A contains a countable subset, which denes a sequence in A. By sequential compactness, this sequence has a subsequence that converges to a point a A. Since this point is clearly an accumulation point of A, we can apply Lemma 3.6.16 and Exercise 3.6.18 to conclude that, for any open cover U of A, we can nd a countable subcover U . From this open cover U , we wish to extract a nite subcover. Let U = {Uj | j N}. Suppose that, for each n, the collection {U1 , U2 , . . . , Un } does not cover A. Then, for each n, there exists xn A \ (U1 Un ). This denes a sequence (xn )nN in A which by sequential compactness has a convergent subsequence with limit x A. Since U covers A, x must be contained in UN for some N . But then, UN contains innitely many elements of the sequence, and hence contains some xm with m > N . This is a contradiction. Exercise 3.6.20 i. Show that a compact metric space is complete. ii. Show that a totally bounded complete metric space is compact. (See Exercise 3.6.15.) An immediate consequence of Theorem 3.6.19 is the Heine-Borel Theorem in Rn and Cn . Theorem 3.6.21 (Heine-Borel) A nonempty subset A of Rn (or Cn ) with the usual metric is compact i it is closed and bounded. Proof. Exercise. Compact sets in Rn with the usual metric have many interesting properties, some of which are illustrated in the following exercises. Exercise 3.6.22 Let B be a compact convex subset of Rn with the usual metric. Dene the nearest point function p : c B B as follows: For x c B we set p(x) to be closest point to x that lies in B . Show that i. the function p(x) is well dened; 109

ii. the point p(x) lies in the boundary of B ; iii. the function p(x) is surjective onto the boundary of B . In the next exercise, we continue with the terminology of the preceding exercise. Dene the supporting hyperplane at p(x) to be the hyperplane through p(x) orthogonal to the vector p(x) x. Dene the supporting half-space at p(x) to be the set Hp(x) = {y Rn | (y p(x)) (p(x) x) 0}. (Note that the supporting hyperplane and the supporting half-space really depend on x, not just on p(x).) Exercise 3.6.23 i. Show that, for each x c B , the set B is a subset of Hp(x) . ii. Show that B =
x c B

Hp(x) .

iii. Does the above process work when B is a closed convex unbounded subset of Rn with the usual metric? Here are a few more interesting facts and ideas about metric spaces. The rst involves the notion of separability. Denition 3.6.24 Let (X, d) be a metric space. A subset A X is said to be dense in X if A = X . Example 3.6.25 i. In the usual metric, Q is dense in R.
a 2n

ii. The dyadic numbers, that is, the set D = Exercise 3.6.26

Q | a, n Z , are dense in R in the usual metric.

i. Show that in any metric space X , X is dense in X .

ii. Show that in any discrete metric space X , the only dense subset of X is X itself. iii. Show that if the only dense subset of a metric space X is X itself, then X is discrete. Denition 3.6.27 Let (X, d) be a metric space. We say that X is separable if there exists a countable subset of X that is dense in X . Example 3.6.28 The spaces Rn and Cn with the usual metric are separable. As a countable dense subset, we can take the collection of all points in Rn whose coordinates are rational numbers, or the set of all points in Cn whose coordinates have the property that the real and imaginary parts are rational numbers. Theorem 3.6.29 If (X, d) is a compact metric space, then X is separable. Proof. For each n N, consider the collection of open balls {B1/n (x) | x X }. This is an open covering of X , and hence, there is a nite subcovering Un . Take the union over all n N of the centers of the balls

in Un . This is a countable collection of points in X that is obviously dense.

Exercise 3.6.30 Suppose X and X are metric spaces with X separable. Let f : X X be a continuous surjection. Show that X is separable. As shown in Example 3.6.28, separable metric spaces do not have to be compact. Many of the important metric spaces which occur in analysis are separable, but there are some very important examples of nonseparable metric spaces. Exercise 3.6.31 Find a metric d on R such that (R, d) is not separable. 110

Exercise 3.6.32 Determine the conditions, if they exist, for which the following metric spaces are separable. i. B (X, F ) ii. BC (X, F ) Another important idea in metric spaces is connectedness. It has a funny denition because we begin by dening a non-connected set. Denition 3.6.33 Let X be a metric space and let A X . We say that A is not connected (or disconnected ) if there exist open sets U, V X such that a. U A = and V A = , b. (U A) (V A) = , c. A = (U A) (V A). We say that A is disconnected by the open sets U and V . Denition 3.6.34 Let X be a metric space and A X . We say A is connected if A is not disconnected. Exercise 3.6.35 i. Show that a subset of a discrete metric space is connected i its cardinality is at most 1. ii. Show that a nite subset of any metric space is connected i its cardinality is at most 1. iii. Show that a subset A of R in the usual metric is connected i A is an interval. iv. Show that a convex subset of Rn with the usual metric is a connected set. The basic theorem about connected sets is the following. Theorem 3.6.36 Let X, X be metric spaces and f : X X a continuous function. If A is a connected subset of X , then f (A) is a connected subset of X . That is, the continuous image of a connected set is connected. Proof. Let U and V be open sets in X , and assume that U and V disconnect f (A). Then, f 1 (U ) and f 1 (V ) are open sets in X which disconnect A. Corollary 3.6.37 (Intermediate Value Theorem) Let X be a metric space, and take R with the usual metric. Let f : X R be a continuous function. Let A be a connected subset of X and let I = f (A). Then I is an interval in R, and if x0 I there exists a0 A such that f (a0 ) = x0 . Exercise 3.6.38 Take R with the usual metric, and let f : R R be given by f (x) = xn for n N. If b is a positive real number, show that there exists a unique positive real number a such that an = b. (Hint: Use the Corollary.) Exercise 3.6.39
i. Consider a punctured open ball, that is, a set of the from Br (a) = Br (a) \ {a}, in Rn with the usual metric. For which values of n is Br (a) connected? ii. Let Br (a) be a punctured open ball in Cn . For which values of n is Br (a) connected?

111

iii. Show that GL(2, R) with the metric inherited from M2 (R) as in Exercise 3.3.35 is not a connected set. (Hint: use the fact that the determinant is a continuous function.) iv. Show that GL(2, C) with the metric inherited from M2 (C) is a connected set. If a metric space X is not connected, then it can be decomposed into subsets called connected components. Denition 3.6.40 If X is a metric space and x0 is in X , then the connected component of x0 in X is the union of the connected sets that contain x0 . Exercise 3.6.41 i. Let X be a metric space and take x0 X . Show that the connected component of x0 is a connected set in X . ii. Show that if A is a connected subset of X that contains x0 , then A is contained in the connected component of x0 . iii. Show that if A is a connected subset of a metric space, then A is connected. Deduce that connected components are closed. Examples 3.6.42 i. Let X = R , the set of nonzero real numbers with the usual metric. This metric space has two connected components, namely, the positive real numbers and the negative real numbers. ii. The connected components of GL(2, R) with the usual metric are GL+ (2, R) = {x GL(2, R) | det x > 0} and GL (2, R) = {x GL(2, R) | det x < 0}. Exercise 3.6.43 Let O(n, R) and SO(n, R) be metric spaces with the metric inherited from GL(n, R). Show that O(n, R) is not connected and that SO(n, R) is connected. Denition 3.6.44 A metric space X is totally disconnected if the connected component of each point is the point itself. Example 3.6.45 A discrete metric space X is totally disconnected. Exercise 3.6.46 i. Find an example of a metric space which is totally disconnected but not discrete. ii. Find an example of a complete metric space which is totally disconnected but not discrete.

3.7

The Contraction Mapping Theorem and its Applications

Denition 3.7.1 Let X be a metric space and f a map from X to X . We say that f is a contraction mapping of X if there exists a real number , with 0 < < 1, such that d(f (x), f (y )) d(x, y ) for every pair x, y X . Exercise 3.7.2 Show that a contraction mapping is continuous. Exercise 3.7.3 Let f : R R be a polynomial function. Give conditions on f such that f is a contraction mapping. Exercise 3.7.4 mapping?
p Let T : p n (R) n (R), 1 p , be a linear transformation. When is T a contraction

112

Denition 3.7.5 Let X be a metric space and f a map from X to X . A point x0 X is a xed point of f if f (x0 ) = x0 . Exercise 3.7.6 i. Find a continuous function f : R R that does not have a xed point. ii. Find a continuous function f : (0, 1) (0, 1) that does not have a xed point. iii. Let f : [0, 1] [0, 1] be continuous. Show that f has a xed point. Theorem 3.7.7 (Contraction Mapping Theorem) Let X be a nonempty complete metric space and let f : X X be a contraction mapping with constant . Then f has a unique xed point x0 X . Proof. Let x1 be any element of X . Dene x2 = f (x1 ), x3 = f (x2 ) = f (f (x1 )) = f 2 (x1 ), and in general, xn = f n1 (x1 ). Then, if n > m, we have d(xm , xn ) = d(f m1 (x1 ), f n1 (x1 )) m1 d(x1 , f nm (x1 )) m1 (d(x1 , x2 ) + d(x2 , x3 ) + + d(xnm , xnm+1 )) m1 (d(x1 , x2 ) + d(x1 , x2 ) + + nm1 d(x1 , x2 )) m1 d(x1 , x2 ). 1

It follows that (xn )nN is a Cauchy sequence in X that converges since X is complete. Let x0 = limn xn . From the continuity of f , it follows that f (x0 ) = f (limn xn ) = limn f (xn ) = limn xn+1 = x0 . Exercise 3.7.8 Show that x0 is the unique xed point of f . Exercise 3.7.9 We say that a function f : Rn RN satises a Lipschitz condition if there exists a constant C such that f (x) f (y ) C x y for all x, y Rn . i. Show that f satises a Lipschitz condition with constant C for 0 < C < 1 if and only if f is a contraction mapping. ii. Show that if f satises a Lipschitz condition with any constant, then f is continuous. iii. For each C (0, ), nd an example of a function f : Rn Rn that satises the Lipschitz condition with constant C . iv. Let B = B1 (0) be the unit ball in Rn . For each C > 1, nd an example of a function f : B B that satises the Lipschitz condition with constant C . v. Find an example of a continuous function f : Rn Rn that does not satisfy any Lipschitz condition. The following theorem, sometimes called Picards Theorem, gives a direct application of the contraction mapping theorem to a problem in analysis. Theorem 3.7.10 (Picards Theorem.) Let B be a ball of radius r in R2 with center at (x0 , y0 ). Suppose that f : B R is a continuous function that satises a Lipschitz condition in the second variable, that is, there is a constant C such that |f (x, y ) f (x, y )| C |y y | for all (x, y ), (x, y ) B . Then, there exists a > 0 such that the dierential equation dy/dx = f (x, y ) has a unique solution y = (x), satisfying (x0 ) = y0 , in the interval |x x0 | < . 113

Proof. Without loss of generality, we can assume that f is bounded on B , that is, there exists a constant M such that |f (x, y )| M for all (x, y ) B . Take a > 0 such that C < 1 and {(x, y ) | |x x0 | , |y y0 | M } B . We now work inside the space X = { C ([x0 , x0 + ]) | |(x) y0 | M }. If we give X the sup metric, then by Theorem 3.5.26 and Exercise 3.6.9, X is a complete metric space. Now, x take the mapping T : X X dened by T (x) = y0 + x0 f (t, (t)) dt. It is obvious that T X and that d(T , T ) Cd(, ). Thus, T is a contraction mapping on X , and there is a unique function X such that T = . It is easy to check that the solutions to the dierential equation are precisely the xed points of T , so the proof is complete. The contraction mapping theorem can also be applied to systems of dierential equations, see for example [Kolmogorov-Fomin]. The following exercise assumes the reader is familiar with some basic concepts in calculus, especially integration theory. Exercise 3.7.11 Take C ([0, 1], R) with the sup metric, and let k (x, y ) : [0, 1] [0, 1] R be a continuous 1 function satisfying sup0x1 0 |k (x, y )| dy < 1. Given a function g (x) C ([0, 1], R), show that there is a unique solution f (x) C ([0, 1], R) to the equation
1

f (x)

k (x, y )f (y ) dy = g (x).
0

3.8

Stone-Weierstrass Theorem

In this section, we undertake a closer examination of the space C (X, F ), that is, the collection of continuous functions on a compact metric space X with values in the eld F , where F = R or C. Recall that the distance between two functions f and g in C (X, F ) is given by d(f, g ) = supxX |f (x) g (x)|, the sup metric. Our investigation will lead to a nice characterization of certain dense subsets of C (X, F ). Denition 3.8.1 Let A be a collection of functions from a set X to F . We say that the collection A separates points if, for every pair of distinct points x1 , x2 X , there is a function f A such that f (x1 ) = f (x2 ). Example 3.8.2 If X = [0, 1], then C (X, R) separates points. This is easy to see just by drawing a picture. Exercise 3.8.3 i. Show that polynomial functions in C ([0, 1], R) separate points. ii. Does the class of functions {sin(2nx) | n N} in C ([0, 1], R) separate points? Denition 3.8.4 A real polynomial function f : Rn R is a nite linear combination of expressions of mn 1 m2 where m1 , m2 , . . . , mn , are non-negative integers. The coecients of a polynomial the form xm 1 x2 xn may be taken from Z, Q, or R. The resulting set of polynomials is denoted by Z[x1 , . . . , xn ], Q[x1 , . . . , xn ], and R[x1 , . . . , xn ], respectively. 2 5 15 Example 3.8.5 A typical polynomial in R[x1 , x2 , x3 , x4 ] looks like 2x3 1 x2 x3 x4 + x1 x2 x4 11x1 x4 . Exercise 3.8.6 i. Show that R[x1 , x2 , . . . , xn ] is a commutative ring with 1 for R = Z, Q, or R. Find the units (invertible elements) in each of these rings. ii. Find the possible images of a polynomial function in R[x]. iii. Find the possible images of a polynomial function in R[x1 , x2 ]. 114

Theorem 3.8.7 (Weierstrass) Let A be a compact set in Rn . Then every continuous function f : A R is the uniform limit of a sequence of real polynomials in R[x1 , . . . , xn ]. Theorem 3.8.8 (Stone) Let X be a compact metric space. Let A be an algebra of continuous, real valued functions on X , and suppose that A separates points. Then A, the closure of A in C (X, R) under the sup metric, sometimes called the uniform closure of A, either coincides with C (X, R) or with Cx0 (X, R) = {f C (X, R) | f (x0 ) = 0}, for some point x0 X . Remark 3.8.9 When A = C (X, R), we say that A is uniformly dense in R. Exercise 3.8.10 Show that Stones theorem implies Weierstrasss theorem. (Hint: Let A = R[x1 , x2 , . . . , xn ].) Before we attempt the proof of Stones theorem, it will be helpful to gather some preliminary lemmas. Lemma 3.8.11 Let A be an algebra of real-valued, continuous functions on a compact metric space X . Then, for f A, |f | is in the uniform closure of A. That is, there is a sequence (fn )nN in A such that (fn )nN converges uniformly to |f |. Proof. Since X is compact, we know that f is bounded (see Exercise 3.6.9). Choose C R, C > 0, such 1 that |f (x)| C for all x X . Let u = C f . Then u A, and d(u, 0) 1. Now we construct a sequence (wn )nN in A converging uniformly to |u|. Let w0 = 0, and dene wn inductively by the relation wn+1 = wn +
2 u2 wn . 2

Before proceeding further, notice that if we formally take limits in n, we would have a relation of the form w = w + (u2 w2 )/2, which would imply that w2 = u2 . With a little luck, we may also show that w 0 and hence w = |u|. First notice that 0 w1 w0 = w1 = u2 /2 u2 |u|. Now suppose wk wk1 0, and wk |u| for 2 u2 wn wn = |u|+ (|u| wn ) 0. Also, keeping in mind that 1 k n. Then wn 0, and wn+1 wn = 2 2 |u| 1, we have 0 wn+1 = wn +
2 u2 wn |u| + wn = wn + ( )(|u| wn ) 2 2 wn + |u| wn = |u|.

Hence, by induction, (wn )nN is an increasing sequence of functions, and 0 wn |u| for all n. Now, as suggested in the beginning of the proof, we let w be the pointwise limit of the sequence (wn )nN . Then, w = |u|, and by Dinis Theorem (Theorem 3.6.11), we know that the sequence (wn )nN converges uniformly to |u|.

Denition 3.8.12 Let V be a vector space of real valued continuous functions on a metric space X . We say that V is a lattice if |f | V whenever f V . Exercise 3.8.13 Let V be a lattice on a metric space X . If f, g are in V , set f g = min(f, g ) and f g = max(f, g ). Show that f g , f g V . Lemma 3.8.14 Let X be a compact metric space and L a lattice of continuous functions on X . Suppose that, for any x, y X with x = y and a, b R, there is a function fxy L satisfying fxy (x) = a and fxy (y ) = b. Then, for each f C (X, R), there is a sequence (fn )nN in L such that (fn )nN converges uniformly to f . 115

Proof. Take f C (X, R) and > 0. For any x, y X , we identify the function fxy and the sets Uxy and Vxy as follows. Let a = f (x) and b = f (y ). Take fxy L such that fxy (x) = a and fxy (y ) = b. We take Uxy = {z X | fxy (z ) < f (z ) + } and Vxy = {z X | f (z ) < fxy (z )}. Notice that for any x, y X , the sets Uxy and Vxy are open and, in addition, both contain x and y . Fix y . Then by compactness, there exists a nite number of points x1 , x2 , . . . , xn such that {Ux1 y , Ux2 y , . . . , Uxn y } covers X . Set hy = min(fx1 y , fx2 y , . . . , fxn y ). By Exercise 3.8.13, we have hy L and hy (z ) < f (z ) + for n all z X . Notice that f (z ) < hy (z ) for z Vy = i=1 Vxi y . Now let y X vary, and for each y , construct hy and Vy as above. By compactness, we can select an open cover {Vy1 , Vy2 , . . . , Vym } of X . Put l = max(hy1 , hy2 , . . . , hym ). Then l L and f (z ) < l(z ) < f (z ) + . Finally, to construct (fn )nN , we let = 2n and choose fn to be the function l constructed above. We are ready to return to the proof of Stones theorem. Proof. (of Theorem 3.8.8) There are two cases to consider. First, suppose that, for each x0 X , there is an f A such that f (x0 ) = 0. Take x1 , x2 X such that x1 = x2 . Then there is a function f A so that f (x1 ) = 0 and f (x1 ) = f (x2 ). To see this, take functions h, g A such that g (x1 ) = g (x2 ) and h(x1 ) = 0. Then choose f as follows. If g (x1 ) = 0, let f = g . If g (x1 ) = 0 and h(x1 ) = h(x2 ), let f = h. If g (x1 ) = 0 and h(x1 ) = h(x2 ), let f = g + h. If f (x2 ) = 0, let u(x) = f (x)/f (x2 ) (f (x)/f (x2 )) . Then u A, u(x1 ) = 0 and u(x2 ) = 0. Hence, we can nd f1 and f2 in A such that f1 (x1 ) = 1, f1 (x2 ) = 0, f2 (x1 ) = 0, and f2 (x2 ) = 1. Now, for any a, b R, take fx1 x2 (x) = af1 (x) + bf2 (x). Then fx1 x2 (x1 ) = a and fx1 x2 (x2 ) = b. From Lemma 3.8.11, we have that A, the uniform closure of A, is a lattice. From Lemma 3.8.14, A = C (X, R). This concludes the proof in the rst case. Now we turn to the case when there is an element x0 X such that f (x0 ) = 0 for all f A. Let A = {g C (X, R) | g (x) = c + f (x) for some c R and f A}. We have that A is an algebra satisfying the conditions for the rst part of the theorem. In particular, if h(x) Cx0 (X, R) and > 0, then there is a function f A and c R such that supxX |h(x) c f (x)| < . Looking at x0 , we see that |c| < . Hence supxX |h(x) f (x)| < 2.
2

Exercise 3.8.15 Let X, Y be compact metric spaces. Let A = {(x, y ) C (Y, R), 1 i n}. i. Show that A is an algebra. ii. Show that A is uniformly dense in C (X Y, R). Exercise 3.8.16 i. Prove the complex version of the Stone-Weierstrass theorem:

n i=1

fi (x)gi (y ) | fi C (X, R), and gi

Let X be a compact metric space. Let A be an algebra of continuous complex-valued functions on X with the property that if f A then its complex conjugate f is in A. Assume that A separates points and that there is no point of x X such that f (x) = 0 for all f A. Then A = C (X, C). ii. A trigonometric polynomial from T = {z C | |z | = 1} to C is a function of the form f (ei ) = n ij , where the coecients are in C. Show that the set of trigonometric polynomials is j = n a j e uniformly dense in C (T, C). 116

3.9

The Completion of a Metric Space

Obviously, complete metric spaces play a special role among all metric spaces. We now present a procedure through which any metric space can be embedded as a dense subset of a complete metric space. d ), and an Theorem 3.9.1 Let (X, d) be a metric space. Then there exists a complete metric space (X, injection : X X , such that 1. : X (X ) is an isometry, and

. 2. (X ) is dense in X

Proof. Consider the set X of all Cauchy sequences in X . We dene an equivalence relation on X by saying that (xn )nN is equivalent to (yn )nN if limn d(xn , yn ) = 0. Exercise 3.9.2 Prove that this is an equivalence relation. be the set of equivalence classes. We denote the equivalence class of a Cauchy sequence (xn )nN Let X . Let {(xn )nN } and {(x )nN } be elements of X . We note by {(xn )nN }. We rst dene a metric on X n that (d(xn , xn ))nN is a Cauchy sequence in R. This follows from the fact that |d(xn , xn ) d(xm , x m )| d(xn , xm ) + d(x n , xm ). We set ({(xn )nN }, {(x )nN }) = lim d(xn , x ). d n n
n

This limit exists by the Cauchy criterion (see Theorem 1.6.24). is well-dened. Exercise 3.9.3 Show that d by (x) = {(xk )kN } where xk = x for all k N. It is clear that is an isometry Now dene : X X from X to (X ). ) is complete. , and second, show that (X, d There are two things left to do. First, show (X ) is dense in X Let x = {(xn )nN } X . Pick > 0. Since the sequence (xn )nN is Cauchy in X , there exists an integer N such that d(xN , xm ) < if m N . Now consider the class of the constant sequence (xN ). Then ( . d x, (xN )) = limn d(xn , xN ) and hence (X ) is dense in X is complete, take a Cauchy sequence ( . Remember, each y To show that X yn ) in X n is an equivalence 1 ( class of Cauchy sequences in X . For each n N, by density, choose z n (X ) such that d yn , z n ) < n . 1 1 yn , y m ) + m . This implies that ( zn )nN is Then d( zn , z m ) d( zn , y n ) + d( yn , y m ) + d( ym , z m ) < n + d( . Let xn = 1 ( Cauchy in X zn ). Then, since is an isometry, (xn )nN is Cauchy in X . Let y be the dened by the equivalence class of this Cauchy sequence, that is, y element of X = {(xn )nN }. Then, 1 ( ( ( ( ( d yn , y ) d yn , z n ) + d zn , y ) < n +d zn , y ). Observe that d zn , y ) = limk d(xn , xk ). Since (xn )nN is Cauchy in X , for n and k large, d(xn , xk ) can be made arbitrarily small. Thus, ( yn )nN converges to y . in the above theorem is called the completion of X . Denition 3.9.4 The metric space X ) are isometric. d Exercise 3.9.5 If (X, d) is already a complete metric space, show that (X, d) and (X, ) is unique up to isometry. That is, if (X , d ) is a complete metric space d Exercise 3.9.6 Prove that (X, d ) and (X , d ) are isometric. such that X is isometric to a dense subset of X , then (X, Remark 3.9.7 One might ask at this point, Why did we write Chapter 1 at all? Why not just take the rational numbers with the usual metric and complete them by the above process to get the real numbers? Sorry folks, but in the proof of the above theorem, we used the fact that the real numbers are complete. In Project ??, we will have a simple, yet signicant example of the completion of a metric space, namely, the p-adic completion of Q relative to a prime p. This emphasizes the fact that while R is the most familiar example of a completion of Q with respect to a metric, there are in fact innitely many other completions of Q. 117

3.10

Independent Projects

3.10.1 The p-adic completion of QThe simplest example of the completion of an incomplete metric space is called the p-adic completion of Q. The p in this case refers to a prime integer p, and the metric is that dened below. This metric plays a signicant role in analysis, number theory, theoretical physics, and other areas. Denitions and basic properties

Denition 3.10.1 Let p be a prime in Z. For r Q , we write r = pk ( a b ), where a and b are relatively prime integers not divisible by p. Dene the p-adic absolute value | |p on Q by |r|p = pk if r = 0 and |0|p = 0. Exercise 3.10.2 Show that | |p has the following properties for all r, s Q: i. |r|p 0, and |r|p = 0 if and only if r = 0; ii. |rs|p = |r|p |s|p ; iii. |r + s|p max(|r|p , |s|p ); iv. |r + s|p = max(|r|p , |s|p ) if |r|p = |s|p . Note that i and ii are familiar properties of the usual absolute value on Q, while iii, known as the nonArchimedean Triangle Inequality, is stronger than the usual triangle inequality on Q, which asserts that |r + s | | r | + |s |, The absolute value | |p gives a metric on Q dened by dp (r, s) = |r s|p , Exercise 3.10.3 i. Show that dp is a metric. ii. Find a Cauchy sequence in Q relative to dp that does not converge in Q. That is, Q is not complete with respect to dp . We denote by Qp the completion of Q with respect to the metric dp . We can dene addition and multiplication on Qp such that Qp becomes a eld. Recall that elements of Qp are equivalence classes of Cauchy sequences from Q relative to dp . The process of turning Qp into a eld proceeds exactly as in the case of the real numbers (see section 1.5). Denition 3.10.4 Addition and multiplication on Qp are dened as {(an )nN } + {(bn )nN } = {(an + bn )nN }, and {(an )nN } {(bn )nN } = {(an bn )nN }. Next, we must extend | |p to Qp . Observe that, if (an )nN is a Cauchy sequence in Q with respect to dp , then (|an |p )nN is a Cauchy sequence in R. So if {(an )nN } Qp , then the absolute value on Qp can be dened by |{(an )nN }|p = lim |an |p
n

r, s Q.

r, s Q.

118

Note that if lim |an |p = 0, then the sequence (|an |p )nN is eventually constant and hence converges to n the eventual constant. Exercise 3.10.5 i. Show that addition, multiplication, and | |p are well-dened on Qp . ii. Show that Qp is a eld with the operations given above. iii. Show that | |p on Qp satises the same properties as it does on Q (see (3.10.2)). iv. Show that the image of Qp under | |p is the same as that of Q under | |p , that is, {pk | k Z} {0}. v. Show that Qp cannot be made into an ordered eld. Denition 3.10.6 The eld Qp with | |p is called a p-adic eld. It is also called the p-adic completion of Q. 3.10.2 The additive structure of Qp We begin by dening several sets in Qp that play an important role in our study of p-adic elds. Denition 3.10.7 Dene the following subsets of Qp : a. Rp = {x Qp | |x|p 1};
1 b. = {x Rp | |x|p < 1} = {x Rp | |x|p p }; and

c. Up = {x Rp | |x|p = 1}. The set Rp is called the ring of integers in Qp . The set is called the maximal ideal in Rp . The set Up is called the group of units in Rp . Exercise 3.10.8 Show that Rp is a commutative ring with 1. Proposition 3.10.9 The set is a subgroup of Rp , and Rp =
0kp1

k + .

Proof. It follows from the non-Archimedean triangle inequality that is an additive subgroup of Rp . Let x Rp . If |x|p < 1, then x . Suppose |x|p = 1. Since Q is dense in Qp , there is some r Q such that r = a/b with (a, b) = (a, p) = (b, p) = 1 and |r x|p < 1. Hence, x + = r + . Since p and b are relatively prime, there exists an integer k with 0 < k p 1 such that p divides a kb. Hence, |a kb|p < 1, and also kb a | a b |p < 1 by Exercise 3.10.8 since p b. Thus, |k b |p < 1. It follows that k + = r + = x + so that x k + . Exercise 3.10.10 i. Show that Up is, in fact, the set of units in Rp , that is the set of elements in Rp that have multiplicative inverses in Rp . ii. Show that Up is a group under multiplication. iii. Show that is an ideal in Rp , that is, if a is in and x Rp , then ax . iv. Show that is a maximal ideal in Rp . That is, if x Up , then the smallest ideal containing x and is all of Rp . v. For n Z, dene n = pn Rp = {pn x | x Rp } = {x Qp | |x| pn }. Show that n is a subgroup of (Qp , +). 119

vi. Show that n \ n+1 = pn Up . vii. Show that, if n > 0, n is an ideal in Rp , that is, if a n and x Rp , then ax n . viii. Show that Qp = ix. Show that Q p =
nZ nZ

n . pn U p .

Denition 3.10.11 If n is an integer, the set pn Up is called a shell in Qp . 3.10.3 The topological structure of Qp We now consider the topology on Qp determined by the metric dp associated with | |p . Exercise 3.10.12 If x0 Qp and r > 0, show that there is an integer n such that Br (x0 ) = Bpn (x0 ) = x0 + n+1 = {x Qp | |x x0 |p < pn } = {x Qp | |x x0 |p pn1 }. This shows that the open balls in Qp are simply cosets of some power of . Proposition 3.10.13 For each n Z, the subsets n and pn Up are both open and closed in Qp . Proof. First, consider pn Up for some n Z. If x pn Up , then |x|p = pn . If k > n, then the ball x + k is contained in pn Up by Exercise 3.10.5.iii. This proves that pn Up is open. Now, consider n . If x n and k > n, then the ball x + k is contained in n . Hence n is open. To show that n is closed, notice that Qp \ n = k<n pk Up , which is open. Finally, pn Up is the complement of n+1 in n so that pn Up is closed by Exercise 3.3.22. Corollary 3.10.14 If n Z and x Qp , then x + n is both open and closed. Corollary 3.10.15 Any open set A in Qp can be written as a disjoint union of cosets of the subgroups n , n Z. Proof. If A is empty, then we are done, so suppose that it is not. Suppose further that A is bounded. Then the set S of integers n such that A contains some coset of n is bounded below. By the Well Ordering Principle, applied to a suitable shift of S , we see that S has a least element n0 . Let A0 = a0 + n0 be a coset of n0 contained in A. By Corollary 3.10.14, A0 is closed, so A \ A0 is open. If A \ A0 is empty, then we are done. Otherwise, repeat to get A1 , and so on. Exercise 3.10.16 Prove that this algorithm terminates, so that we have written A as a disjoint union of cosets A0 , A1 , . . . of the desired form. Exercise 3.10.17 Explain how to reduce the case of general A to the case of bounded A. (Hint: Consider the intersection of an arbitrary open set A with cosets of Rp .)

It is now easy to prove some of the basic topological properties of Qp . Exercise 3.10.18 Show that the ring of integers Rp is a maximal compact subring of Qp . Exercise 3.10.19 Show that the eld Qp has the Bolzano-Weierstrass property, that is, if A is a bounded innite subset of Qp , then A has an accumulation point in Qp . Exercise 3.10.20 Show that Qp has the Heine-Borel property, that is, if A is a closed, bounded subset of Qp , then A is compact. 120

Exercise 3.10.21 Show that the eld Qp is a locally compact eld, that is, every point in Qp has a neighborhood whose closure is compact. In fact, show that every point in Qp has a neighborhood that is both open and compact. We now introduce some exercise concerning the cosets of in Rp Exercise 3.10.22 Recall from Proposition ?? that Rp = 0kp1 (k + ). Show that these p cosets are disjoint. Exercise 3.10.23 Show that Up = 1kp1 (k + ). We introduce an algebraic structure on the collection of cosets as follows. We dene (k + ) + (j + ) = (k + j ) + , and (k + )(j + ) = kj + . Exercise 3.10.24 Show that the addition and multiplication so dened are well-dened. Exercise 3.10.25 Show that the collection F = {k + | 0 k p 1} is a eld with these two operations. Exercise 3.10.26 Show that F is isomorphic to Zp , the eld of integers modulo p. Exercise 3.10.27 For a xed n Z, show that pn U p =
1 k p1

(pn k + n+1 ).

Exercise 3.10.28 The ring of ordinary integers Z is dense in Rp relative to | |p . Denition 3.10.29 The valuation map : Qp Z {+} is dened by the following rule: p (x) = |x|p (See Exercise 3.10.5.iv. With this denition, we can now write: a. Rp = {x Qp | (x) 0}; b. = {x Rp | (x) > 0}; c. Up = {x Rp | (x) = 0}. We now consider the convergence of innite series in Qp . The situation here is simpler than that in Section 1.9. In real and complex analysis, determining whether or not an innite series converges can be a delicate matter. The p-adic case is dierent. Theorem 3.10.30 Let an Qp for all n N. Then
n=1

if x = 0,

(0) = +.

an converges in Qp if and only if lim an = 0.


n

Proof. The only if part is clear, just as in the real and complex cases. (See Theorem 1.9.9.) Now suppose that lim an = 0. This means that, given k 0, we can pick N N such that |an |p < pk n for all n > N . Thus, for all m > n > N |sm sn |p = |an+1 + + am |p 121
n+1im

max

| a i |p < p k ,

the rst inequality following from the non-Archimedean Triangle Inequality. Therefore, the sequence (sn )nN of partial sums is Cauchy, and so it must converge by the completeness of Qp . From the decomposition Q p =
nZ

pn Up into shells (see Exercise 3.10.10.ix, one can express any non-zero

x as an innite series x = ak pk , where the ak {0, 1, . . . , p 1} are uniquely determined, and there only nitely many k < 0 (possibly none) for which ak = 0. In fact, the rst non-zero term in the series is the one corresponding to the valuation of x, and one can write x in a p-adic expansion : x=
k = ( x )

ak p k ,

where a (x) = 0. It follows immediately from Theorem 3.10.30 that the p-adic expansion of x converges to x. Exercise 3.10.31 Find the 5-adic expansion of x =
1 3

in Q5 .

Exercise 3.10.32 Let x and y be the two square roots of 11 in Q5 . Find the 5-adic expansions of x and y . (What is the 5-adic expansion of x + y ?) Exercise 3.10.33 Show that there is no x in Q5 such that x2 = 7. Exercise 3.10.34 Show that a rational number has a periodic p-adic expansion and determine the length of the period. 3.10.4 Fundamental Theorem of AlgebraHere is a proof of the Fundamental Theorem of Algebra promised in Chapter ??. Exercise 3.10.35 Let P be a polynomial of positive degree with coecients in C. Show that there exists z0 C such that |P (z0 )| |P (z )| for all z C. Then show that, by considering the polynomial P (z + z0 ), we may assume z0 = 0. Theorem 3.10.36 (The Fundamental Theorem of Algebra) The eld C is algebraically closed, that is, any nonconstant polynomial with coecients in C has a root in C. Proof. Let P C[z ] be a polynomial of positive degree. By Exercise 3.10.35, we may assume P (z ) has a minimum at 0. There exists n 1 and a, b C with b = 0 such that P (z ) = a + bz n + z n+1 Q(z ), where Q C[z ]. Suppose that P (0) = a = 0, and choose an n-th root w of a/b in C. By continuity, there exists t with 0 < t < 1 such that t|wn+1 Q(tw)| < |a|. Now, we have P (tw) = a + b(tw)n + (tw)n+1 Q(tw) = (1 tn )a + (tw)n+1 Q(tw) because bwn = a. Hence, |P (tw)| (1 tn )|a| + tn+1 |wn+1 Q(tw)| < (1 tn )|a| + tn |a| = |a| = |P (0)|.

This is a contradiction, and hence we must have P (0) = a = 0.

122

Exercise 3.10.37 Find 10 other proofs of the fundamental theorem of algebra. Starting with any eld F we wish to dene an algebraic closure of F . We rst dene an algebraic extension of F . Denition 3.10.38 Let F be a eld, and let E be a eld containing F as a subeld of E . We say that E is an algebraic extension of F if given E , there exists a nonzero polynomial p(x) F [x] such that p() = 0. Denition 3.10.39 Let F a eld. A eld E containing F is an algebraic closure of F , if E is an algebraic extension of F and E is algebraically closed. The following sequence of statements leads to the existence and uniqueness, up to isomorphism, of an algebraic closure of F . Facts 3.10.40 1. The eld F is contained in an algebraically closed eld E . 2. There is an extension E of F that is both algebraically closed and algebraic over F . 3. Suppose F is a eld and E is an algebraic extension of F . Let be a monomorphism (injective homomorphism) of F into an algebraically closed eld L. Then can be extended to a monomorphism of E into L. 4. If L and L are algebraically closed elds that are algebraic over F , then there exists an isomorphism : L L such that is the identity on F. Exercise 3.10.41 Prove the above statements. Use Langs Algebra, Chapter 5, if you must. Exercise 3.10.42 Show that if F is algebraically closed, then the algebraic closure of F is F . Remark 3.10.43 The Fundamental Theorem of Algebra shows that C is algebraically closed and, in fact, that C is the algebraic closure of R. (You should prove, by writing down polynomials, that C is an algebraic extension of R.) Exercise 3.10.44 i. Show that AR , the eld of real algebraic numbers, is not algebraically closed. ii. Show that A, the eld of algebraic numbers, is the algebraic closure of Q.

123

124

Chapter 4

Dierentiation
The actual state of [innitesimal calculus] at German universities still shows the same colorful diversity. On the one hand, there is the rigorous approach that begins with a six weeks long Dedekind treatment, and then deduces as natural and necessary consequences the concrete rules of dierentiation and integration from the general concepts of number and function. On the other, there is the intuitive treatment that lets reign the magic of the dierential, and even in the last hour of a two semester course does not dissipate by the sunshine of a clear conception of limit, the fog surrounding the indivisibles. And, in between, there are the shadings of hundreds of diagonals that one can construct between the two perpendicular directions of thought. Otto Toeplitz, Jahresbericht dur Deutschen Mathematiker (1927)

4.1

Review of Dierentiation in One Variable

We assume that the reader is familiar with the standard properties of the derivative in one variable, and we will not review the computational aspects of elementary calculus. However, we shall establish rigorously those properties of the derivative in one variable that stem from the completeness of the real numbers. Many of the aspects of dierentiation which occur in several variables are motivated by, and rely on, results in one variable. We begin by dening the derivative of a real-valued function of one variable at a point. Denition 4.1.1 Let [a, b] be an interval in R, and consider f : [a, b] R. We say that f is dierentiable at a point x (a, b) if there exists L R such that
h0

lim

f (x + h) f (x) = L. h

Observe that this denition can be phrased in the following way. The function f is dierentiable at x (a, b) if there exists L R such that
h0

lim

f (x + h) f (x) Lh = 0. h

The number L is called the derivative of f at x, and is denoted by f (x) or Df (x). Exercise 4.1.2 If L exists, show that it is unique. 125

Exercise 4.1.3 Show that f is dierentiable at x (a, b) i there exists a constant L such that |f (x + h) f (x) Lh| = 0. h0 |h | lim Exercise 4.1.4 Show that f is dierentiable at c (a, b) i limxc this limit does exist, it is equal to f (c).
f (x ) f (c ) x c

exists, and show that when

The reader should be familiar with the derivative of a function at a point. Dierentiability is a pointwise property of functions, that is, it is possible for a function to be dierentiable at one point and nowhere else (see Example 4.1.6 below). Theorem 4.1.5 Suppose f : [a, b] R is dierentiable at a point x (a, b). Then f is continuous at x. Proof. Take = 1. Then there exists a > 0 such that |f (x + h) f (x) f (x)h| < |h| = |h| whenever |h| < . It follows from the triangle inequality that |f (x + h) f (x)| < |h| + |f (x)||h| = (1 + |f (x)|)|h| when |h| < . Letting h 0, we get the result. Example 4.1.6 Let f : R R be dened by f (x) = x2 if x is rational and 0 if x is irrational. This function is discontinuous at every nonzero x. On the other hand, f is continuous at x = 0, and f (0) = lim f (0 + h) f (0) . h

h0

The expression whose limit we are evaluating is equal to either h or 0 depending on whether h is rational or irrational, respectively. Thus, the limit as h approaches 0 is 0, and thus f (0) = 0. Hence f is dierentiable at x = 0 and nowhere else. Exercise 4.1.7 Generalize the function from Exercise 3.5.4. Let r 1, and set fr (x) =
1 qr

if x = p q in lowest terms, and x = 0, and if x = 0 or x is irrational.

i. Show that for any r 1, fr is continuous at 0 and the irrational numbers and is discontinuous at the nonzero rationals. ii. If 1 r 2, show that fr is not dierentiable at any irrational point. (Hint: Use Theorem 1.3.9.) iii. For which r is fr dierentiable at x = 0? To settle all discussions about the relationship between dierentiability and continuity, consider the following example. Example 4.1.8 We want to create a continuous function on the interval [0, ) that is not dierentiable at any point in that interval. Dene f1 (x) = x 1x if x 1/2 if 1/2 x 1

1 fn1 (2x). Let and extend periodically to [0, ) by f1 (x + 1) = f1 (x). Then dene for all n 2, fn (x) = 2 m Sm (x) = n=1 fn (x). Then Sm is a continuous function on [0, ).

126

Exercise 4.1.9 Show that the sequence (Sm )mN converges uniformly to a continuous function S . Exercise 4.1.10 Show that S is not dierentiable at any point in (0, ). Theorem 4.1.11 Suppose f : [a, b] R and g : [a, b] R are both dierentiable at x (a, b). Then for any , R, f + g is dierentiable at x. Also, the product f g and the quotient f g are dierentiable at x
f g

(for

we must have g (x) = 0). Then we have

i. (f + g ) (x) = f (x) + g (x); ii. (f g ) (x) = f (x)g (x) + f (x)g (x); and iii.
f g

(x) =

f (x )g (x ) f (x )g (x ) . [g(x)]2

Proof. Look in a rigorous calculus book. Before proceeding further, we want to create a setting which will prevail throughout the theory of dierentiation. We said above that dierentiability is a pointwise property. Generally speaking, we will assume that a function is not only dierentiable at a point, but at all points in a neighborhood of a given point. It is rarely the case that we deal with functions which are dierentiable only at a single point. We want to pay special attention to the derivative of a composition of functions, sometimes known as the chain rule. The proof here takes a little care and will require even more care in several variables. Theorem 4.1.12 (Chain Rule) Let f be dierentiable at a point a and let g be dierentiable at f (a). Then g f is dierentiable at a and D(g f )(a) = (Dg )(f (a))Df (a). Proof. Let b = f (a), L = Df (a) and M = Dg (b). Set F (x) = f (x) f (a) L(x a),

G(y ) = g (y ) g (b) M (y b), H (x) = (g f )(x) (g f )(a) M L(x a). By hypothesis, we know that
xa

lim

To prove the theorem, we must show that

|G(y )| |F (x)| = lim = 0. |x a| yb |y b| |H (x)| = 0. |x a |

xa

lim

Notice that H (x) = G(f (x)) + M F (x). Now, |F (x)| |M F (x)| = |M | 0 |x a | |x a | as x a. For the remaining term, it follows from above that given > 0, there exists a > 0 such that |y b| < implies |G(y )| < |y b|. The continuity of f at a implies that there exists a > 0 such that |f (x) b| < when |x a| < . Hence, if |x a| < , we have |G(f (x))| < |f (x) b|. But |f (x) b| |F (x)| + |L||x a|, so |G(f (x))| 0 |x a | as x a. 127

Exercise 4.1.13 Give a critique of the following supposed proof of the chain rule: lim g (f (x)) g (f (a)) xa = = = g (f (x)) g (f (a)) f (x) f (a) f (x) f (a) xa f (x) f (a) g (f (x)) g (f (a)) lim lim xa xa f (x) f (a) xa Dg (f (a))Df (a). lim

xa

xa

Suppose that [a, b] is a closed interval in R and f : [a, b] R is continuous. Assume f is dierentiable on (a, b). Since f is continuous, we know that f assumes a maximum and minimum value on [a, b]. This observation leads to the following familiar fact from elementary calculus. Theorem 4.1.14 Suppose that f satises the hypotheses above and f assumes a local maximum or minimum at a point c (a, b). Then f (c) = 0. Proof. Assume that f has a local maximum at c. There exists > 0 such that if |x c| < then f (x) f (c). ) f (c ) ) f (c ) It follows that if x (c , c) then f (xx 0 and if x (c, c + ) then f (xx 0. Thus, in evaluating c c

) f (c ) , the former inequality tells us that this limit, if it exists, must be greater than or f (c) = limxc f (xx c equal to 0, and the latter shows that if it exists, it must be less than or equal to 0. Since we know by hypothesis that f is dierentiable at c, the limit does exist and can only equal 0. The proof for a local

minimum is similar. We now take up various versions of the Mean Value Theorem. The Mean Value Theorem can be regarded as the most important theorem in analysis both in one and several variables. The statements in the following theorem are often called Rolles Theorem, the Mean Value Theorem (MVT), and the Generalized Mean Value Theorem. We state them all here so that the reader will be in familiar territory. We prove only the Generalized Mean Value Theorem, which immediately implies the other two. The Generalized Mean Value Theorem is often called Cauchys Mean Value Theorem. Theorem 4.1.15 (Mean Value Theorem) Let f : [a, b] R be continuous and suppose that f is dierentiable on (a, b). i. If f (a) = f (b), then there exists c (a, b) such that f (c) = 0. ii. In any case, there exists c (a, b) such that f (c)(b a) = f (b) f (a). iii. If g satises the same hypotheses as f , then there exists c (a, b) such that (f (b) f (a))g (c) = (g (b) g (a))f (c). Proof. To prove iii, set h(x) = (f (b) f (a))g (x) (g (b) g (a))f (x). Then h is continuous on [a, b] and dierentiable on (a, b). Note that h(a) = h(b). If h is constant, we are done. If not, h assumes a max or min at some point c (a, b). Theorem 4.1.14 says that h (c) = 0. The conclusion follows. The Mean Value Theorem has some serious applications. The rst is important for the Fundamental Theorem of Calculus. Corollary 4.1.16 Suppose f is continuous on [a, b] and dierentiable on (a, b). If f (x) = 0 for all x (a, b), then f is constant. Proof. Given any two points x and y in [a, b] with x < y , there exists a point c (x, y ) such that f (y ) f (x) =

(y x)f (c) = 0. Hence, f (x) = f (y ).

128

Corollary 4.1.17 Suppose f is continuous on [a, b] and dierentiable on (a, b). If f (c) > 0 for all c (a, b), then f is monotonic increasing. Proof. For x, y [a, b] with x < y , there exists c (x, y ) such that f (y ) f (x) = (y x)f (c) > 0. Exercise 4.1.18 Suppose f is continuous on [a, b] and dierentiable on (a, b), and f (c) < 0 for all c (a, b). Show that f is monotonic decreasing. Exercise 4.1.19 (LH opitals Rule) Let (a, b) be any open interval in R, and suppose f and g are dieren (x ) tiable on (a, b). Take c (a, b), and suppose limxc f (x) = limxc g (x) = 0 and limxc f g (x) exists. Show
f (x ) g (x )

that limxc

= limxc

f (x ) g (x ) .

We now pause to present one of the all time favorite examples in one variable dierentiation. Example 4.1.20 Let f (x) = Then f (x) = 2x sin(1/x) cos(1/x) when x = 0, 0 when x = 0. x2 sin(1/x) when x = 0, 0 when x = 0.

So even though f (0) exists, f is not continuous at 0. Things are not really as bad as they seem, because, although the derivative may not be continuous, it does have the intermediate value property. Theorem 4.1.21 (Intermediate Value Theorem for Derivatives) Let f : [a, b] R be continuous and let f be dierentiable on (a, b). Suppose that [c, d] (a, b), f (c) < 0, and f (d) > 0. Then there exists a point x (c, d) such that f (x) = 0. Proof. Since f is continuous on [c, d], it assumes both a maximum and minimum value. Since f (c) < 0, there exists a point x (c, d) such that f (x) < f (c) and, since f (d) > 0, there exists a point y (c, d) such

that f (y ) < f (d). Hence the minimum does not occur at either c or d. The conclusion follows.

Exercise 4.1.22 There is a point in the previous proof which requires attention. It is related to Corollary 4.1.17 and the exercise which follows it. In the theorem above, it is assumed that f (c) < 0. One might be inclined to think that this means that f is decreasing in a neighborhood of c. To show that this is not true, consider the function 2x2 sin(1/x) + x x = 0 f (x) = 0 x = 0. Show that f has a positive derivative at x = 0 but is not increasing in any neighborhood of x = 0. There is no reason to stop at one derivative. Once we get started, we can continue taking derivatives as long as the function allows us. Most of the functions encountered in elementary calculus such as polynomials, rational functions, trigonometric functions, exponential functions, logarithmic functions, hyperbolic functions, and algebraic functions, are dierentiable innitely often as long as nothing untoward happens in the domain. The above functions make up more or less the entire list of functions considered in elementary calculus. We assume that the reader knows how to dierentiate them. The following denition is useful throughout analysis. 129

Denition 4.1.23 Let f be a continuous function from (a, b) to R. If k is an integer greater than or equal to 1, we say that f C k (a, b) if f has k derivatives at each point in (a, b) and each of these derivatives is continuous on (a, b). We informally refer to such a function f as being C k on (a, b). We denote the k -th derivative of f by f (k) . By convention, we take f (0) = f . We say that f C (a, b) if f has derivatives of all orders on (a, b), and we refer to such a function as being C on (a, b). If U is any open set in R, the expressions C k (U ) and C (U ) are dened similarly. Exercise 4.1.24 Suppose f : [a, b] R is C 1 on (a, b). Let [c, d] (a, b). Then there exists a constant M such that for all x, y [c, d], we have |f (y ) f (x)| M |y x|. Exercise 4.1.25 For n N, nd the maximum value of k for which the function f (x) = |x|n is in C k (R)? Exercise 4.1.26 Let f (x) = i. Show that f C (R). ii. Using LH opitals rule, or anything you wish, show that f (k) (0) = 0 for all k 0. The higher derivatives of a function f sometimes allow us to approximate f with polynomials rather than mere linear functions. Corollary 4.1.27 (Taylors Theorem) Suppose f C k+1 (a, b) and x0 (a, b). Then, for any x (a, b), we can write f (x) = f (x0 ) + f (x0 )(x x0 ) + f (x0 ) (x x0 )2 + + 2! f (k) (x0 ) f (k+1) (c) + (x x0 )k + (x x0 )k+1 k! (k + 1)! e1/x 0
2

when x = 0 when x = 0.

where c is some point between x and x0 . Proof. Without loss of generality, we assume that x > x0 . For t [x0 , x], dene a polynomial Tk (x, t) in the variable x by f (t) f (k) (t) Tk (x, t) = f (t) + f (t)(x t) + (x t)2 + + (x t)k . 2! k! Dene Rk (x, t) = f (x) Tk (x, t). Note that Rk (x, x0 ) C k+1 (a, b), Rk (x0 , x0 ) = 0 for 0 i k , and (k+1) Rk (x, x0 ) = f (k+1) (x) for all x (a, b). For t [x0 , x], x (a, b), x > x0 , set Q(t) = (x t)k+1 Rk (x, x0 ) Rk (x, t). (x x0 )k+1
(i)

It is clear that Q is continuous on [x0 , x], dierentiable on (x0 , x), and Q(x0 ) = Q(x) = f (x). It is easy to see that Rk (x, x0 ) (x t)k (k+1) Q (t) = (k + 1)(x t)k + f (t). (x x0 )k+1 k! Hence, by Rolles Theorem, there exists c (x0 , x) such that (k + 1)(x c)k (x c)k (k+1) Rk (x, x0 ) = f (c). (x x0 )k+1 k!

130

What Taylors Theorem allows us to do is to approximate a C k+1 function in the neighborhood of a point by a polynomial of degree k . Usually, the remainder term, Rk (x, x0 ) = f (k+1) (c) (x x0 )k+1 , (k + 1)!

is reasonably small because of the (k + 1)! in the denominator. The expansion in Taylors Theorem is called the Taylor expansion of f about x0 . Exercise 4.1.28 Find the Taylor expansions of the following functions about the indicated points to at least 6 terms. i. f (x) = sin(x) about x0 = ; ii. f (x) =
1 x 1

about x0 = 1;
2

iii. f (x) = e1/x about x0 = 0; iv. f (x) = x2 + 1 about x0 = 2. Exercise 4.1.29 i. Suppose that f C 1 (a, b) and f (x0 ) = 0 for some x0 (a, b). Show that f may or may not have a local extremum at x0 . ii. Suppose that f C 2 (a, b), f (x0 ) = 0, and f (x0 ) > 0. Show that f has a local minimum at x0 . Formulate and prove an analogous statement in case f (x0 ) < 0. iii. Suppose that f C k+1 (a, b), f (i) (x0 ) = 0 for 1 i k , and f (k+1) (x0 ) = 0. Under what conditions can you say that f must have a local extremum at x0 ? As another application of the Mean Value Theorem, we present the theorem of Liouville. This theorem can sometimes be used to determine whether a given irrational real number is algebraic or transcendental. Theorem 4.1.30 (Liouvilles Theorem) Let R be algebraic of degree n 2. Then there exists C = C () depending on such that | p/q | > C/q n for all p/q Q. Proof. Let f (x) Z[x] be an irreducible polynomial of degree n such that f () = 0. Let p/q Q and assume that < p/q . By the Mean Value Theorem, there exists c with < c < p/q such that f (p/q ) f () = f (p/q ) = (p/q )f (c). By the irreducibility of f (x), we have f (p/q ) = 0, and therefore f (c) = 0. Assume that 0 < p/q < 1. Choose d > 0 such that |f (x)| < 1/d for x [, p/q ]. Then |f (p/q )| = |p/q | |f (c)| = 0. Since f has integer coecients, |q n f (p/q )| Z, so it is greater than or equal to 1. Hence, 1 q n |p/q |(1/d). It follows that | p/q | d/q n when 0 < p/q < 1. Exercise 4.1.31 i. Modify the above proof to cover the case when p/q < . ii. Show that there exists a constant C such that | p/q | > C/q n for all p/q Q. Exercise 4.1.32 Suppose n 2 and is real algebraic of degree n. Show that, if r > n, then the function fr of Exercise 4.1.7 is dierentiable at . In elementary calculus, the derivative is often motivated through a discussion of the tangent line to a curve at a point. This is accomplished using secant lines which approximate the tangent line and then taking limits. The actual denition of the tangent line to the graph of a function at a point is: If the function is dierentiable at the point, the tangent line at that point is the line through the point whose slope is the derivative of the function at that point. We can think of the tangent line as the best linear approximation at that point. This is the idea that motivates the concept of derivative in several variables. 131

4.2

Dierential Calculus in Rn

We now turn to a study of the properties of dierentiable functions from Rn to Rm . Throughout this chapter, we use the Euclidean norm in Rn . Our denition of the derivative will be stated in terms of linear transformations from Rn to Rm , and we will need the following ideas. Proposition 4.2.1 Let T : Rn Rm be a linear map. Then there exists a constant C > 0 such that for every x Rn , T x C x . Proof. Consider S n1 = {x Rn | x = 1}. This is a compact set by Heine-Borel, and thus the function S R, x T x , is continuous and hence attains a maximum C on S n1 by Exercise ??. It follows
n1 Tx x

that for any nonzero x Rn ,

= T(

x x

) C.

Denition 4.2.2 We dene the operator norm of a linear transformation T : Rn Rm to be the constant C identied in the proposition above, and we denote this by T = C . Exercise 4.2.3 If T : Rn Rm is linear, and T = 0, show that T = 0. Remark 4.2.4 We use the same symbol to indicate the operator norm for a linear transformation as for the Euclidean norm in Rn , but the interpretation will always be clear from the context. The denition of a dierentiable function is motivated by the discussion at the conclusion of the previous section about the derivative of a function f : R R. As in the case of the derivative of functions of one variable, we form a dierence quotient, which means that we must divide by something. Since division in Rn does not make sense for n > 1, we need to keep the divisor in R. Denition 4.2.5 Suppose U Rn is an open set. A function f : U Rm is dierentiable at x U if there is a linear map T : Rn Rm such that
h0

lim

f (x + h) f (x) T h = 0. h

First, note that the h 0 in Rn . Notice also that the norm sign in the numerator denotes the Euclidean norm in Rm while the norm sign in the denominator denotes the Euclidean norm in Rn . If we use the norm sign for an element of R, it indicates the usual absolute value on R. We write T = Df (x) and we call this the derivative of f at x. We say that f is dierentiable on U if f is dierentiable at each point in U . Thus the derivative of a function f : Rn Rm at a point is a linear transformation. It may be dicult to think of this as a generalization of the slope of a tangent line as it is in one variable. However, if one thinks of the tangent line in one variable as the best linear approximation to a function at a point, we can think of the derivative in Rn as a generalization of this concept; that is, f (x0 ) + Df (x0 )(x x0 ) provides the best linear approximation to the function f at the point x0 Rn . There are many theorems about derivatives of functions of several variables which are analogous to those in one variable. Theorem 4.2.6 Suppose U is an open set in Rn and f : U Rm is dierentiable at a point x0 U . Then f is continuous at x0 . 132

Proof. Take = 1. Then there exists a > 0 such that f (x0 + h) f (x0 ) Df (x0 )h < h = h whenever h < . It follows from the triangle inequality that f (x0 + h) f (x0 ) < h + Df (x0 )h h + Df (x0) h = (1 + Df (x0 ) ) h when h < . When m = 1 and f is a real-valued function, this leads to a special situation. Denition 4.2.7 Let U Rn be an open set, and let f : U R be dierentiable on U . Let x U and let v Rn be a unit vector. The directional derivative of f at x in the direction v is dened as Dv f (x) = lim
t0

f (x + tv ) f (x) . t

In the particular case when v = ej , a standard basis vector, we obtain the partial derivative in the j th direction f (x + tej ) f (x) Dj f (x) = Dej f (x) = lim . t0 t Exercise 4.2.8 i. Let U Rn be an open set and let f : U Rm be dierentiable. Write f = (f1 , f2 , . . . , fm ), where fk : U R is the k th coordinate function of f . Let v Rn be a unit vector. If x U , dene Dv f (x) = lim
t0

f (x + tv ) f (x) . t

Show that Dv f (x) exists if and only if Dv fk (x) exists for all k , 1 k m, and, in this case, Dv f (x) = (Dv f1 (x), Dv f2 (x), . . . , Dv fm (x)). ii. Explain why it is useful for us to have required v to be a unit vector. Remark 4.2.9 Note that, in Denition 4.2.7, these directional derivatives are real honest-to-goodness derivatives of functions of one variable, that is, they represent the rate of change of a function in a particular direction. The partial derivatives play a special role, as we shall see below. It is worth observing that the f (x). All sorts of theorems and properties can be stated much more easily classical notation for Dj f (x) is x j with the notation Dj f (x). Exercise 4.2.10 i. Let f : Rn R be dened by f (x1 , x2 , . . . , xn ) = xk for some k , for 1 k n. Show that f is dierentiable at any point and that Df (x) = f for all x Rn . ii. Find Dv f (x) for any unit vector v in Rn . Remark 4.2.11 The map f in the above exercise is called the projection onto the k -th coordinate and is denoted pk . More generally, if m n, we can pick indices 1 i1 < i2 < < im n and dene a projection p : Rn Rm by p(x1 , x2 , . . . , xn ) = (xi1 , xi2 , . . . , xim ). Exercise 4.2.12 Show that any such p as above is dierentiable and nd its derivative. All of the statements below are easy exercises, but we prove one or two just to show that we are working at this. 133

Proposition 4.2.13 If U is an open set in Rn and f : U Rm is dierentiable, then the derivative of f is unique. Proof. Suppose T and T are derivatives of f at x U . Then, for h Rn , T h T h h = (f (x + h) f (x) T h) (f (x + h) f (x) T h) h f (x + h) f (x) T h f (x + h) f (x) T h + . h h

Thus T h T h / h 0 as h 0. But T T is a linear map, so by the denition of the norm of an operator,


h =M

max

T h T h = T T . h

Because this is a constant independent of M for M = 0 we must have T T = 0, that is, T = T .

Exercise 4.2.14 If f : Rn Rm is a linear map and x Rn , then f is dierentiable at x and Df (x) = f . This exercise says something which is almost tautological, but it says something nonetheless. That is, if f : Rn Rm is linear and h Rn , then at any point x Rn , Df (x)h = f (h). How does this all work for a linear function f : R R? Notice rst that if f : R R is a linear map, then f (0) must be equal to 0. Moreover, there is an element a R such that f (x) = ax, for all x R. Conversely, given an element a R we can construct a linear map fa : R R dened by fa (x) = ax. In elementary calculus, we use this correspondence to treat derivatives of functions of one variable as numbers at each point, or as functions on R, rather than as linear maps R R at each point. In our present case, instead of saying that if fa (x) = ax, then fa (x) = a, we have Dfa (x) = fa for all x. Observe that this discussion tells us that if a function is already linear, then the best linear approximation is the function itself. Proposition 4.2.15 Let U be an open set in Rn and f, g : U Rm be dierentiable on U . Then f + g is dierentiable on U , and D(f + g )(x) = Df (x) + Dg (x) for x U . Proof. For x U , (f + g )(x + h) (f + g )(x) (Df (x) + Dg (x))h h g (x + h) g (x) Dg (x)h f (x + h) f (x) Df (x)h + . h h Both expressions on the right go to zero as h goes to zero. The reader might ask about the product rule. The question depends on the type of multiplication being used. Let us try the easiest case in which f and g map U to R. Proposition 4.2.16 If U is an open set in Rn , and f : U R and g : U R are real-valued functions which are dierentiable on U , then f g is dierentiable on U . For x U , we have D(f g )(x) = f (x)Dg (x) + g (x)Df (x). 134

Proof. Before starting the proof, we observe that f (x)Dg (x) really makes sense, since f (x) is a real scalar and Dg (x) is a linear map, and we can always multiply linear maps by scalars. So lets go: 1 h f (x + h)g (x + h) f (x)g (x) (f (x)Dg (x) + g (x)Df (x))h = 1 h 1 h (f (x + h)[g (x + h) g (x)] f (x)Dg (x)h) + (g (x)[f (x + h) f (x)] g (x)Df (x)h) (f (x + h) f (x))(g (x + h) g (x)) + f (x)(g (x + h) g (x) Dg (x)h) + g (x)(f (x + h) f (x) Df (x)h) . By the denition of derivative, the second and third terms vanish as h 0. For the rst term, we have
1 h

(f (x + h) f (x))(g (x + h) g (x)) =

f ( x +h ) f ( x ) h

g (x + h) g (x) .

Exercise 4.2.17 Finish the proof by showing

f ( x +h ) f ( x ) h

g (x + h) g (x) goes to zero as h 0.

We now turn to the Chain Rule which is a very important theorem in the calculus of several variables. The reader will note that the proof is pretty much the same as the proof in one variable (Theorem 4.1.12). Theorem 4.2.18 (Chain Rule) Suppose that U is an open set in Rn and f : U Rm is dierentiable on U . Let V be an open set in Rm such that f (U ) V . Suppose g : V Rp is dierentiable on V . Then g f : U Rp is dierentiable on U , and, for any a U we have D(g f )(a) = Dg (f (a)) Df (a). Proof. Let b = f (a), L = Df (a), and M = Dg (b). Set (x) = f (x) f (a) L(x a), (y ) = g (y ) g (b) M (y b), and (x) = (g f )(x) (g f )(a) M L(x a). Then, our hypotheses tell us that
xa

lim

(y ) (x) = lim = 0, y b y b xa lim (x) = 0. xa

and we need to show that


xa

But (write out the denitions). Then

(x) = (f (x)) + M (x) M (x) M xa (x) 0 xa

as x a. By the above limit and the continuity of f , for any > 0, there is > 0 such that x a < ensures (f (x)) < f (x) b . But f (x) b (x) + L x a , so (f (x)) 0 xa as x a as well. 135

4.3

The Derivative as a Matrix of Partial Derivatives

We now proceed to some of the more computational aspects of dierentiation theory in several variables. Let us consider the form of a mapping from Rn to Rm . If f : Rn Rm , we can write f (x) = (f1 (x), f2 (x), . . . , fm (x)), where for each k with 1 k m, fk : Rn R. These functions fk are the component functions of f , and there is an obvious fact about them. Corollary 4.3.1 (to the Chain Rule) If f is dierentiable, then fk is dierentiable for each k . Proof. Observe that fk = pk f and use the chain rule. Exercise 4.3.2 Show that if each fk is dierentiable on an open set U Rn , then f = (f1 , f2 , . . . , fm ) is dierentiable on U and Df = (Df1 , Df2 , . . . , Dfm ). How does one interpret (Df1 , Df2 , . . . , Dfm ) as a linear transformation from Rn to Rm ? For x U the expression (Df1 (x), Df2 (x), . . . , Dfm (x)) is the linear transformation whose value at y Rn is given by [(Df1 (x), Df2 (x), . . . , Dfm (x))](y ) = (Df1 (x)(y ), Df2 (x)(y ), . . . , Dfm (x)(y )). What do partial derivatives have to do with all this? We know that linear transformations from Rn to R can be represented by matrices with respect to any bases we choose in the respective spaces. Let us work rst with a dierentiable function f : Rn R. We know that
m h0

lim

f (x + h) f (x) Df (x)h = 0. h

If we think of Df (x) as a 1 n matrix of the form (a1 (x), a2 (x), . . . , an (x)) and write h = (h1 , h2 , . . . , hn ), then Df (x)h = (a1 (x), a2 (x), . . . , an (x)) (h1 , h2 , . . . , hn ), where the dot signies the dot product in Rn . Now consider the case h = tej . It follows from Denition 4.2.7 that aj (x) = Dj f (x). This is summarized in the following statement. Proposition 4.3.3 If U Rn is open and f : U R is dierentiable on U , then Df (x) = (D1 f (x), D2 f (x), . . . , Dn f (x)). For h Rn , Df (x)h = D1 f (x)h1 + D2 f (x)h2 + + Dn f (x)hn .

Exercise 4.3.4 Let U Rn be open, and let f = (f1 , f2 , . . . , fm ) : U Rm be dierentiable on U . If x U and (aij ) is the matrix of Df (x) with respect to the standard basis, then aij = Dj fi (x). Denition 4.3.5 Let U Rn be open, let f : U Rm be dierentiable on U , and let x U . The matrix (Dj fi (x))1im,1j n is called the Jacobian matrix of f at x. In the case that m = 1, the vector-valued function (D1 f (x), D2 f (x), . . . , Dn f (x)) is called the gradient vector and is denoted f (x). Thus, by Proposition 4.3.3, Df (x)(h) = f (x) h. Exercise 4.3.6 i. Let U Rn be open and let f : U R be dierentiable on U . If x U and v is a unit vector in Rn , then Dv f (x) = f (x) v . 136

ii. Show that the maximum value of the directional derivative of f at a point x is in the direction of f (x), and the value of this directional derivative is f (x) . The existence of the partial derivatives at a point, or even in the neighborhood of a point, does not assure the dierentiability of the function at the point. Actually, it is much worse than that. As shown in the following example, the existence of the directional derivative at a point in every direction does not assure that the function is dierentiable at the point. For dierentiability, one must approach the given point from every direction along all sorts of paths, while the directional derivative is taken along straight-line paths through the point. Example 4.3.7 Let f (x, y ) = 0
xy x 2 +y 2

when (x, y ) = (0, 0), otherwise.

Then D1 f (0, 0) = D2 f (0, 0) = 0, but nonetheless f is not continuous at the origin, and hence not dierentiable at (0, 0). Exercise 4.3.8 i. Let f (x, y ) = 0
x3 x 2 +y 2

when (x, y ) = (0, 0), otherwise.

Show that f has a directional derivative in every direction at the origin, but that f is not continuous at the origin, and hence not dierentiable. ii. Let f (x, y ) = 0 when (x, y ) = (0, 0),
2 2 1 2 x +y 2

Show that D1 f and D2 f exist at the origin but are not continuous. The previous examples and exercises illustrate a few pathologies. Nevertheless, if the partial derivatives all exist and are continuous in a neighborhood of a point, then the function is dierentiable at that point. Theorem 4.3.9 Let U be an open set in Rn and let f : U Rm be a function with the property that Dj fi is continuous on U for 1 i m, 1 j n. Then f is dierentiable on U and, as we might expect, Df (x) = (Dj fi (x))1im,1j n . Proof. We prove it rst for n = 2 and m = 1. The proof uses the Mean Value Theorem for functions of one variable. For x = (x1 , x2 ) U and h = (h1 , h2 ), write f (x + h) f (x) (Df (x))h h (f (x1 + h1 , x2 + h2 ) f (x1 , x2 ) (D1 f (x1 , x2 ), D2 f (x1 , x2 ))(h1 , h2 ) = h f (x1 + h1 , x2 + h2 ) f (x1 , x2 + h2 ) D1 f (x1 , x2 )h1 + f (x1 , x2 + h2 ) f (x1 , x2 ) D2 f (x1 , x2 )h2 = h f (x1 + h1 , x2 + h2 ) f (x1 , x2 + h2 ) D1 f (x1 , x2 )h1 h f (x1 , x2 + h2 ) f (x1 , x2 ) D2 f (x1 , x2 )h2 + h 137

(x + y ) sin

otherwise.

Now, by the Mean Value Theorem for functions of one variable, there exists 1 (x1 , x1 + h1 ) (or 1 (x1 + h1 , x1 ) if h1 < 0) and 2 (x2 , x2 + h2 ) (or 2 (x2 + h2 , x2 ) if h2 < 0) such that f (x1 + h1 , x2 + h2 ) f (x1 , x2 + h2 ) = D1 f (1 , x2 + h2 )h1 and f (x1 , x2 + h2 ) f (x1 , x2 ) = D2 f (x1 , 2 )h2 . Thus the above sequence of inequalities continues as |h 1 | |h 2 | D1 f (1 , x2 + h2 ) D1 f (x1 , x2 ) + D2 f (x1 , 2 ) D2 f (x1 , x2 ) h h D1 f (1 , x2 + h2 ) D1 f (x1 , x2 ) + D2 f (x1 , 2 ) D2 f (x1 , x2 ) , and this goes to zero as h 0 since D1 f and D2 f are continuous at (x1 , x2 ). The general case for arbitrary n and m is easy to complete by adding and subtracting enough times and using the Mean Value Theorem over and over again. So far, we have not computed a lot of derivatives because it is awkward to compute the linear transformation associated to the denition. With this last theorem, it becomes much easier to compute the derivative of a function given by explicit formulas. For example, let f : R3 R2 be dened by f (x, y, z ) = (x2 y + exz , sin(xyz )). Then, f is dierentiable on all of R3 and Df (x, y, z ) = Thus, Df (1, 0, 1) = 2xy + zexz x2 xexz yz cos(xyz ) xz cos(xyz ) xy cos(xyz ) e 1 0 1 e 1 , 1 0 .

which represents a linear transformation from R3 to R2 . The more general expression above assigns to each point (x, y, z ) R3 a linear transformation Df (x, y, z ) L(R3 , R2 ). Thus, we can think of Df as a function Df : R3 L(R3 , R2 ). This last example should totally demystify the idea of derivatives of functions from Rn to Rm . There is of course some interesting theory related to dierentiation, and we are in the midst of that exposition. However, to compute partial derivatives, and hence derivatives, requires nothing more than being able to dierentiate functions of one variable, which you learned in elementary calculus. In fact, as we have commented before, there really are not that many functions that you can dierentiate explicitly. So you will notice that all of the examples involve polynomials, rational functions, trigonometric functions, logarithmic functions, etc. Exercise 4.3.10 For any function f : C C, we can write f in terms of its real and imaginary parts as f (z ) = u(x, y ) + iv (x, y ), where u and v are functions from R2 to R, and z is written in the form z = x + iy . A function f : C C (with the usual metric on C) is complex dierentiable at z0 C if f (z0 ) = lim
z z0

f (z ) f (z0 ) z z0

exists. A function f is analytic on an open set U C if f is dierentiable at each point of U . i. Suppose f is analytic on an open set U C. Show that u and v are dierentiable on U considered as a subset of R2 . 138

ii. Suppose f is analytic on an open set U C. Show that the Cauchy-Riemann equations.

u x

v y ,

and

u y

v = x . These are called

iii. If U C is an open set, and u and v are continuously dierentiable on U and satisfy the CauchyRiemann equations, show that f (z ) = u(x, y ) + iv (x, y ) is analytic on U . iv. Find an example of a function f : C C that is dierentiable at one point but not in a neighborhood of that point. Exercise 4.3.11 Let f : C C be given by f (z ) = ez , which can be written f (x + iy ) = ex cos y + iex sin y (see Denition ??). Show that f is analytic on C, and that f (z ) = f (z ). Exercise 4.3.12 Let z0 C, and dene f : C C by f (z ) = n=0 an (z z0 )n , where an C for all n. Let r be the radius of convergence of this power series (see Section 1.9) and suppose that r > 0. i. Show that f (z ) is analytic on Br (z0 ) = {z C | |z z0 | < r}. (Hint: Show that the series can be dierentiated term-by-term inside the radius of convergence.) ii. Show that the radius of convergence of the power series for f (z ) is equal to r. Exercise 4.3.13 Let f : R2 R be dened by f (x, y ) = is dierentiable. |x| + |y |. Find those points in R2 at which f

Exercise 4.3.14 Let f : Rn R be a function such that |f (x)| x dierentiable at 0. Exercise 4.3.15 Let f : Rn Rn R be dened by f (x, y ) = x y . i. Show that f is dierentiable on Rn Rn . ii. Show that Df (a, b)(x, y ) = a y + b x.

for some > 1. Show that f is

4.4

The Mean Value Theorem

Next we consider the Mean Value Theorem for functions of several variables. As in the case of functions of one variable, the Mean Value Theorem relates the average rate of change of a function in a specied direction to the instantaneous rate of change at a particular point as measured by the derivative. In the case of a function of one variable, there is little choice about the so-called specied direction. But, when we consider functions from Rn to Rm , we nd that it is necessary to specify a direction both in the domain and in the range of the function in order to give a proper interpretation to the Mean Value Theorem. Theorem 4.4.1 Let U be an open set in Rn , and let f : U R be dierentiable on U . Let x, y be two distinct points in U such that the line segment joining x to y lies entirely in U . Then, there exists (0, 1) such that f (y ) f (x) = Df (z )(y x), where z = (1 )x + y . Proof. Dene F (t) = f ((1 t)x + ty ). Then F is continuous on [0, 1] and dierentiable on (0, 1), so there exists (0, 1) such that F (1) F (0) = F ( ). 139

The left-hand side is f (y ) f (x), and by the chain rule, F ( ) = Df ((1 )x + y )(y x).

Note that Df (z )(y x) = Du f (z ) y x , where u is the unit vector in the direction of the vector y x. Exercise 4.4.2 Let U be a convex open set in Rn , and let f : U R be dierentiable on U . Show that if Df (x) = 0 for all x U , then f is constant on U . We note again that the Mean Value Theorem for real-valued functions of several variables that we have just proved is really a one-variable theorem. That is, to make sense of the mean value property, it was essential that we move away from a point x in exactly one direction, namely, the straight-line direction from x to y . It is this idea that motivates the statement of a Mean Value Theorem for functions from Rn to Rm . To retain the one-variable nature of the Mean Value Theorem, in addition to having a straight-line direction implicitly chosen for us in the domain (namely, the direction y x in Rn ), we must also explicitly choose a direction in Rm in order to make sense of the mean value property. Theorem 4.4.3 Let U be an open subset in Rn and let f : U Rm be dierentiable on U . For any two distinct points x, y U such that the line segment joining x to y lies entirely in U , and any vector v Rm , there exists (0, 1) such that v (f (y ) f (x)) = v (Df (z )(y x)), where z = (1 )x + y . Exercise 4.4.4 Prove this. We note in connection to the discussion above that if v is a unit vector, the expression v (f (y ) f (x)) is the component of f (y ) f (x) in the direction of the vector v . A similar statement is true for the expression v (Df (z )(y x)). Exercise 4.4.5 Let U be a convex open set in Rn , and let f : U Rm be dierentiable on U . Show that if Df (x) = 0 for all x U , then f is constant on U . The signicance of the multivariable version of the Mean Value Theorem is that the direction vector v is arbitrary. However, in general, there is no single that will satisfy the conclusion of the Mean Value Theorem for all v simultaneously. Exercise 4.4.6 Show that the function f : R R2 given by f (x) = (cos x, sin x) does not satisfy the property that, given any x, y R with x < y , there exists z (x, y ) such that f (y ) f (x) = Df (z )(y x). (Hint: think about the periodic nature of f .) Exercise 4.4.7 Let U be an open set in Rn , and let f : U Rm be dierentiable on U with continuous partial derivatives. (In the next section, we call such a function C 1 .) Suppose that B is a compact, convex subset of U . Then there exists a constant M such that for any two points x, y B , we have f (y ) f (x) M y x . This is analogous to Exercise 4.1.24 for functions of one variable. 140

4.5

Higher-Order Partial Derivatives and Taylors Theorem

The next natural question is, What about second derivatives? There are two ways to look at this. If f : Rn Rm is dierentiable at a point x, then Df (x) L(Rn , Rm ). That is, it is a linear transformation from Rn to Rm . If we think of the symbol Df as a map from Rn (or a subset of Rn ) to L(Rn , Rm ), then what is the derivative of the derivative? Think about that while we move to a more tractable problem. We have seen that, for a function f : Rn R, the partial derivative Dj f can be thought of as a function from Rn to R. Consequently, it would make sense to consider the partial derivatives of the functions Dj f . For example, suppose f : R3 R is dened by f (x, y, z ) = x2 y + exz . Then, D1 f (x, y, z ) = 2xy + zexz . Then, D3 (D1 f )(x, y, z ) = exz + zxexz . Looking at this from another perspective, we next compute D3 f (x, y, z ). We get D3 f (x, y, z ) = xexz and D1 (D3 f )(x, y, z ) = exz + xzexz . So, for this f , we see that D3 (D1 f )(x, y, z ) = D1 (D3 f )(x, y, z ). The functions D3 (D1 f ) and D1 (D3 f ) are examples of mixed partial derivatives of f . For a general setting, consider a function f = (f1 , f2 , . . . , fm ) from an open set in Rn to Rm . Then, we dene the mixed partial derivative Dij fk = Di (Dj fk ) assuming that the Dj fk has partial derivatives for all j, k . The question is, Under what conditions will we have Dij fk = Dji fk for all i, j ? Theorem 4.5.1 Let U Rn be open and let f : U R be dierentiable on U . Suppose that Dij f and Dji f exist and are continuous on U . Then Dij f (x) = Dji f (x) for all x U . Proof. It is enough to prove the theorem for f : R2 R. Let x = (a, b) U , and let h be small enough that B2h (x) is contained in U . We consider the second dierences for computing the partial derivatives of f . Set A(h) = 1 (f (a + h, b + h) f (a, b + h) f (a + h, b) + f (a, b)) . h2

By the Mean Value Theorem, there exist and between a and a + h such that f (a + h, b + h) f (a, b + h) = hD1 f (, b + h) and In turn this gives between and and between b and b + h such that D1 f (, b + h) D1 f ( , b) = hD21 f ( , ) = hA(h). For the next step, we rewrite A(h) = 1 (f (a + h, b + h) f (a + h, b) f (a, b + h) + f (a, b)) h2 f (a + h, b) f (a, b) = hD1 f ( , b).

and proceed similarly. From this we get A(h) = D12 f ( , ) where and are obtained similarly. If we let h go to 0, the continuity of the mixed partials now implies the result.

Exercise 4.5.2 Let f (x, y ) = 0


x3 y xy 3 x 2 +y 2

when (x, y ) = (0, 0), otherwise.

Show that f is dierentiable everywhere. Show that D12 f (0, 0) and D21 f (0, 0) exist, but D12 f (0, 0) = D21 f (0, 0). 141

In one variable calculus, higher derivatives are dened by dierentiating the derivative considered as a function on some set. In the present situation, where we have a function f : Rn Rm , the derivative is a map from Rn to the space L(Rn , Rm ) of linear maps from Rn to Rm . From our discussion of linear algebra, we know that L(Rn , Rm ) can be identied with the space Rmn . We can then intepret the second derivative at a point as an element of L(Rn , L(Rn , Rm )) = L(Rn , Rmn ). This space has dimension mn2 , and the entries of the matrix representing the derivative are the partial derivatives Dij fk . We could continue this process and dene the th derivative, but this would not be particularly useful for this text and we conne ourselves to the rst and second derivatives. On the other hand, it will be useful to talk about all orders of dierentiation for the partial dervatives. 3 For example, if f : R5 R has partial derivatives of order 3 on an open set U Rn , that is, D1 f (x), 2 D2 D4 f (x), D2 D4 D2 f (x), etc. all exist, we can consider the situation similar to that in Theorem 4.5.1 in which two of these are equal. Denition 4.5.3 Let U Rn be an open set and f : U Rm . The function f is said to be in C k (U ) (or C k on U ) if f has all continuous partial derivatives up to and including order k on U . The function f is said to be in C (U ) or smooth on U if f has all partial derivatives of all orders on U . Exercise 4.5.4 Suppose U Rn and f : U R is C k on U . Let k and i1 , i2 , . . . , i be a collection of integers between 1 and n. Show that for any in S , the symmetric group on letters, Di1 ,i2 ,...,i f = D(i1 ),(i2 ),...,(i ) f. If f is C k on an open set U in Rn , then every partial derivative of order k can be written in the n 1 2 n form D1 D2 Dn f , where the i are nonnegative integers and i=1 i = k . There is an interesting combinatorial problem that arises here. Exercise 4.5.5 i. How many ways can one partition k into an ordered sum of n nonnegative summands? ii. How many ways can one partition k into n nonnegative summands if the order is ignored? Exercise 4.5.6 This is an exercise in high school algebra. Consider the polynomial in n variables, p(x1 , x2 , . . . , xn ) = (x1 + x2 + + xn )k .
kn 1 k2 Show that, upon expanding this in monomials, the coecient of xk 1 x2 xn , where by k! . k1 !k2 ! kn ! n i=1

ki = k , is given

This expression is called a multinomial coecient and is denoted by

k k1 ,k2 ,...,kn

Exercise 4.5.7 Recall Exercise 4.3.12. Let z0 C, and dene f : C C by f (z ) = n=0 an (z z0 )n , where an C for all n. Let r be the radius of convergence of this power series and supposed that r > 0. Show that f is innitely dierentiable on Br (z0 ). Recall that in the Mean Value Theorem for functions f : U R, where U is an open set in Rn , we stated that for points x, y U , there is an element z on the line joining x to y such that f (y ) = f (x)+ Df (z )(y x). For this to hold, the line joining x and y must lie in U . If we assume that f is C k+1 , we would like to further expand f in terms of its partial derivatives of order j where j k + 1. How would we expect such an expression to look? First, let us write y = x + th, where h = (h1 , h2 , . . . , hn ) is a unit vector in Rn and 142

t > 0. Then, the directional derivative of f at x in the direction of the vector h is Dh f (x) = f (x) h. If we iterate this directional derivative r times, we get the expression
r Dh f (x) = 1 ++n =r

r 1 n n h1 h n D1 Dn f (x). 1 , . . . , n 1

We can now write


k

f (y ) = f (x) +
r =1

r Dh f (x)tr + Rk (y ), r!

where, as one might suspect,


k

Rk (y ) = f (y ) f (x)

r =1

r Dh f (x)tr . r!

Theorem 4.5.8 (Taylors Theorem) Let U be a convex open set in Rn , let f : U R be a C k+1 function, and let x, y U with y = x. Write y = x + th, where h is a unit vector in Rn , and t > 0. Then there exists s R, 0 < s < t, such that Dk+1 f (x + sh)tk+1 . Rk (y ) = h (k + 1)! This is, of course, the multivariable analog of Corollary 4.1.27. Proof. Apply Corollary 4.1.27 to the function F (a) = f (x + ah).

Denition 4.5.9 The polynomial


k

f (x) +
r =1

r Dh f (x)tr r!

is called the Taylor polynomial of degree k for the function f at x in the direction h. The Taylor polynomial may be regarded as a reasonable approximation to f in a neighborhood of the point x in the direction h because the remainder term vanishes to order k + 1 at x and has a (k + 1)! in the denominator. Exercise 4.5.10 Compute the Taylor polynomial of degree 3 for the following functions at the specied point (in an arbitrary direction h). i. f (x, y, z ) =
1 xyz

at (1, 1, 1).

ii. f (x, y, z ) = exy+yz at (0, 0, 0). Exercise 4.5.11 Let U be a convex open set in Rn , let x U , and let f : U R be a C k+1 function. Show that the Taylor polynomial of f at x is the best polynomial approximation to f at x by proving that if P is a polynomial of degree k such that lim f (x + th) P (t) = 0, tk

t0

then P is the Taylor polynomial of degree k of f at x in the direction h. 143

4.6

Hypersurfaces and Tangent Hyperplanes in Rn

As stated in Section 4.1, the derivative of a function of one variable f : R R that is dierentiable at a point x = c is the slope of the tangent line to the curve y = f (x) at (c, f (c)). The equation of the line tangent to the curve at that point can be written y = f (c) + f (c)(x c). This situation may be regarded as a special case of a tangent line to a general curve in R2 . The curve y = f (x) may be considered as the level curve {(x, y ) R2 | F (x, y ) = 0} of the function F : R2 R given by F (x, y ) = f (x) y . We now assume that F : R2 R is any C 1 function. Consider the curve F (x, y ) = k in R2 . Let (x0 , y0 ) be a point on this curve, and assume that the gradient F (x0 , y0 ) = (0, 0). The tangent line to the curve at the point (x0 , y0 ) is the line through the point perpendicular to F (x0 , y0 ). The equation of this line is then F (x0 , y0 ) (x x0 , y y0 ) = 0, or

D1 F (x0 , y0 )(x x0 ) + D2 F (x0 , y0 )(y y0 ) = 0.

Exercise 4.6.1 Show that in the case F (x, y ) = f (x) y , the equation of the tangent line is the same as that given in the rst paragraph above. Exercise 4.6.2 i. Given the function F (x, y ) = x2 + y 2 , consider the curve F (x, y ) = 3, that is, x2 + y 2 = 3. Find the equation of the tangent line at each point of this curve. ii. Given the function F (x, y ) = x2 sin2 (1/y ) if y = 0, 0 if y = 0,

consider the curve F (x, y ) = 1. On what domain is F a C 1 function? At what points is the gradient vector nonzero? Find an equation for the tangent line at all points where the gradient vector is nonzero. How does this generalize to N -dimensional space? We consider a C 1 function F : RN R. Denition 4.6.3 A smooth hypersurface in RN is a set of points dened by the equation F (x1 , x2 , . . . , xN ) = k with the property that F = 0 at each point in this set. Examples 4.6.4
2 2 i. Let F : RN R be dened by F (x1 , x2 , . . . , xN ) = x2 1 + x2 + + xN . Then the smooth hypersurface N dened by F (x1 , x2 , . . . , xN ) = 1 is the unit sphere in R .

ii. Let F : R3 R be dened by F (x, y, z ) = x2 + y 2 z . Then the equation F (x, y, z ) = 0 gives a smooth hypersurface in R3 that is called a paraboloid . Denition 4.6.5 Let S be a smooth hypersurface in RN dened by F : RN R, and let c = (c1 , c2 , . . . , cN ) S . We dene the tangent hyperplane to S at c to be the hyperplane through the point c normal to the vector F (c), that is, the hyperplane F (c) (x c) = 0. Exercise 4.6.6 Let F : R3 R be given by F (x, y, z ) =
1 x

1 y

1 z

0 144

if x = 0, y = 0, z = 0, otherwise.

i. For what values of k does the equation F (x, y, z ) = k dene a smooth hypersurface? ii. For those values of k , nd the equation of the tangent hyperplane at every point of the hypersurface. Exercise 4.6.7 Let F : R4 R be given by F (x, y, z, w) = x2 sin y + y 2 sin z zwexy . i. For what values of k does the equation F (x, y, z, w) = k dene a smooth hypersurface? ii. For those values of k , nd the equation of the tangent hyperplane at every point of the hypersurface. In some cases the discussion above can be made simpler. Suppose that the hypersurface dened by F (x1 , x2 , . . . , xN ) = 0 may be realized as the graph of a C 1 function f : Rn R, where N = n + 1. Specically, we suppose the surface can be dened by xn+1 = f (x1 , x2 , . . . , xn ), that is, the level set {(x1 , x2 , . . . , xn , xn+1 ) Rn+1 | F (x1 , x2 , . . . , xn , xn+1 ) = 0}, where F (x1 , x2 , . . . , xn , xn+1 ) = f (x1 , x2 , . . . , xn ) xn+1 . A point C on this hypersurface can be written as C = (c, f (c)), where c = (c1 , c2 , . . . , cn ) is a point in Rn . Using the ideas developed above, we note that the tangent hyperplane to this hypersurface at the point C is dened by F (C ) (X C ) = 0, where X = (x1 , . . . , xn , xn+1 ). Writing this out in coordinates, we get D1 f (c)(x1 c1 ) + D2 f (c)(x2 c2 ) + + Dn f (c)(xn cn ) (xn+1 f (c)) = 0, which can be rearranged to give xn+1 = f (c) + D1 f (c)(x1 c1 ) + D2 f (c)(x2 c2 ) + + Dn f (c)(xn cn ). Exercise 4.6.8 Show that this tangent hyperplane is the n-dimensional ane subspace of Rn+1 through the point C = (c, f (c)) spanned by the set of vectors {v1 , v2 , . . . , vn }, where for each j with 1 j n, we have vj = (0, 0, . . . , 0, 0, 1, 0, 0, . . . , 0, 0, Dj f (c)), where the 1 in the vector vj occurs in the j th place. Exercise 4.6.9 Let f : Rn R be a C 1 function, and let S Rn+1 be the graph of f . Fix a point c Rn , and let T be the tangent hyperplane to S at the point C = (c, f (c)) S . i. For any v Rn , dene v : R Rn by v (t) = tv + c. Let v : R Rn+1 be dened by v (t) = (v (t), f (v (t))). Show that C + v (0) lies in T . ii. Show that every vector X T can be written in the form X = C + V , where V = v (0) for some v Rn . Hence, every vector in the tangent hyperplane may be realized as the tangent vector to a curve in S .

4.7

Max-Min Problems

We now consider the problem of nding maximum and/or minimum values of a function on various subsets of Rn using properties of the derivative. The rst observation is that, in order to discuss the notion of maximum or minimum, we need an order on the range of the function. Thus, we will restrict our attention to real-valued functions for the rest of this discussion. The second observation is that there are two types of maxima and minima, namely, global and local. Denition 4.7.1 Let B Rn be any set, and let f : B R be a function. If there exists p B such that f (x) f (p) for all x B , then we say that f (p) is the global maximum of f on B . Similarly, if there exists q B such that f (x) f (q ) for all x B , then we say that f (q ) is the global minimum of f on B . 145

Denition 4.7.2 Let B Rn be any set, and let f : B R be a function. We say that f assumes a local maximum at a point p B if there exists r > 0 such that f (x) f (p) for all x Br (p) B . Similarly, we say that f assumes a local minimum at a point q B if there exists r > 0 such that f (x) f (q ) for all x Br (q ) B . Note in particular that if p B is a point at which f attains a global maximum, then f automatically has a local maximum at p, and similarly for minima. This implies that our search for global maxima and minima will begin with the search for local maxima and minima, a search to which we may naturally apply the tools of analysis. Note further that global maxima and global minima need not exist. If they do exist, they need not be unique. Indeed, a global maximum or minimum may even occur at an innite number of points as in the case of a constant function. Finally, note that we do know of one very important special case in which global maxima and minima are guaranteed to exist: namely, when B is compact and f : B R is continuous. Let us begin our search for the local maxima and minima of a function f : B R. A local maximum or minimum can be found in one of two places: in the interior of B , which is an open subset of Rn , or on the boundary of B , which is typically a hypersurface or the intersection of several hypersurfaces. Let p B be a point at which f has a local maximum or minimum. If such a point p occurs in the interior of B , and f is dierentiable at p, then we will see that Df (p) = 0. If p occurs in the interior of B , but f is not dierentiable at p, we must explore the behavior of the function in a neighborhood of the point p using various estimation techniques. Finally, if p is on the boundary of B , and the boundary of B may be realized as a hypersurface or the intersection of hypersurfaces, then the theory of Lagrange multipliers (see Section 4.8) may be used to determine the point p. Note the natural correspondence with the one-variable case, where maxima and minima can occur at three types of points: critical points (f = 0), singular points (f does not exist), and endpoints. We begin the analysis by considering f : U R, where U is an open subset of Rn . Theorem 4.7.3 Let U be an open set in Rn , and f : U R. If f has a local maximum at p, and f is dierentiable at p, then Df (p) = 0. Proof. Write p = (p1 , p2 , . . . , pn ). For each j = 1, 2, . . . , n, dene fj (x) = f (p1 , p2 , . . . , pj 1 , x, pj +1 , . . . , pn ). The hypotheses of the theorem imply that, as a function of the single variable x, fj is dierentiable at pj and has a local maximum at pj . Hence Dj f (p) = fj (pj ) = 0. Since f is dierentiable, this implies Df (p) = 0.

Denition 4.7.4 As in the one-variable case, a point where Df vanishes is called a critical point of f . We note that, as in the case of one variable, f need not have either a local maximum or a local minimum at a critical point. Example 4.7.5 The function f (x, y ) = xy has vanishing derivative at the origin, but has neither a maximum nor minimum there, since f is positive in the rst and third quadrants and negative in the second and fourth quadrants. Example 4.7.6 The following example illustrates the care that we must take in identifying critical points. Let 2xy 2 if (x, y ) = (0, 0), 2 4 g (x, y ) = x +y 0 if (x, y ) = (0, 0). This function has all partial derivatives equal to zero at the origin, yet is not even continuous there. 146

Exercise 4.7.7 Let p = (x0 , y0 ) be a point in the plane, and let ax + by + c = 0 be the equation of a line in the plane. Verify that the distance from p to the this line is given by |ax0 + by0 + c| a2 + b 2 using the max-min technique discussed above. Find a similar formula for the distance from a point to a hyperplane in Rn , and verify it. Exercise 4.7.8 Let pj = (xj , yj ), j = 1, . . . , m, be m points in R2 with at least two distinct xj s. Given a m line y = mx + b, dene E (m, b) = j =1 (yj (mxj + b))2 . Find the values of m and b that minimize this sum. For the values of m and b that minimize the function E , the line y = mx + b is called the ordinary least squares approximation to the data p1 , . . . , pm . Exercise 4.7.9 Let pj = (xj , yj ), j = 1, . . . , m, be m points in R2 with at least two distinct xj s. Given a line with equation ax + by + c = 0, with a and b not both zero, denote by d(pj , ) the distance from m the point pj to the line . Consider the function E (a, b, c) = j =1 d(pj , ). Find values of a, b, and c that minimize this function. Exercise 4.7.10 Given a point p = (x0 , y0 , z0 ) with x0 , y0 , z0 > 0, nd an equation for a plane passing through this point that cuts o a tetrahedron of least volume in the rst octant. Once we have identied the critical points of a function f , we might then ask if there is a convenient way to determine whether f actually assumes a local maximum or minimum value at these points. One surere way is to check the behavior of the function in a neighborhood of the critical point directly using inequalities. The next theorem shows that there is a test, similar to the second-derivative test for functions of one variable, for determining whether a function assumes a maximum or minimum value at a critical point. Let U be an open set in Rn and let f : U R be twice dierentiable at a point x U . We dene the Hessian of f at x to be the quadratic form Hx : Rn R dened by Hx (v ) = D(Df )(x)(v )(v ). In terms of the n n matrix of second-order partial derivatives, denoted as D11 f (x) D12 f (x) D1n f (x) D21 f (x) D22 f (x) D2n f (x) Ax = , . . . .. . . . . . . . Dn1 f (x) Dn2 f (x) Dnn f (x)

we can write Hx (v ) = t vAx v .

Remark 4.7.11 We say that the quadratic form Hx is positive denite if Hx (v ) 0 for all v Rn , and Hx (v ) = 0 i v = 0. Negative deniteness is dened similarly. Lemma 4.7.12 Let U be an open set in Rn , and let f : U R be a C 2 function. If Hp is positive denite at a point p U , then Hx is positive denite for x in a neighborhood of p, and similarly when Hp is negative denite. Proof. It is enough to prove the statement about positive deniteness. Let m = inf v =1 Hp (v ); then Hp (v ) m v 2 for any vector v , because v is the product of v and a unit vector. Since the Dij f which 2 for any vector form the coecients of H are continuous in x, for x suciently close to p, Hx (v ) 1 2m v v.

147

Exercise 4.7.13 Show that m > 0 to conclude the proof of the lemma. Theorem 4.7.14 (Second derivative test for extrema) Let U be an open set in Rn , let f : U R be a C 2 function, and let p be a critical point of f . If Hp is positive denite, then f assumes a local minimum at p, and if Hp is negative denite, then f assumes a local maximum at p. Proof. It is enough to prove the statement about minima, since we can replace f by f for maxima. Given any suciently small vector h, the Taylor formula for f at p, to second order, is 1 f (p + h) = f (p) + Df (p) h + Hp+th (h)t2 2 for some t [0, 1]. Since p is a critical point, Df (p) = 0. By the lemma above, Hp+th is positive denite for h suciently small, so we have f (p + h) f (p), which proves that f assumes a local minimum at p.

Example 4.7.15 Consider the function f (x, y ) = xy . The only critical point of f is (0, 0). The Hessian matrix at this point is 0 1 1 0 and the associated quadratic form is H(0,0) (u, v ) = 2uv. Thus, the second derivative test does not apply. Nonetheless, it is easy to see that f has no local extremum at (0, 0). Exercise 4.7.16 For each of the following functions f : R2 R, nd the critical points and compute the Hessian at each such point. Use the second derivative test to determine whether the critical points are local maxima or local minima, if possible. If the test does not apply, determine the nature of the critical point by other means. i. f (x, y ) = x2 + y 2 . ii. f (x, y ) = x4 + y 4 . iii. f (x, y ) = x2 2xy + y 2 . iv. f (x, y ) = x2 y 4 . v. f (x, y ) = (1 xy )2 + x2 . Exercise 4.7.17 Let f : U R2 R be C 2 . Let p U be a critical point of f and suppose that the matrix of Hp has negative determinant. Show that f does not have a local extremum at p. The critical points of the type in the above exercise are referred to as saddle points. Exercise 4.7.18 Find the critical points of f (x, y ) = x3 + 8y 3 6xy 2. For each, determine if it is a local maximum, local minimum, or saddle point, if possible. Exercise 4.7.19 Let p0 = (x0 , y0 , z0 ) be a point in the rst octant in R3 , that is, x0 , y0 , z0 > 0. Consider planes through p0 that intersect the x, y , and z axes at p1 = (a, 0, 0), p2 = (0, b, 0), and p3 = (0, 0, c), respectively, with a, b, c > 0. Find the values of a, b, and c that minimize the area of the triangle with vertices p1 , p2 , and p3 . 148

4.8

Lagrange Multipliers

We now turn to the case where we wish to nd the extreme values of a function f : Rn R restricted to a hypersurface in Rn or the intersection of several hypersurfaces. We will rst deal with the case of a single hypersurface S . The equation for the hypersurface S is sometimes referred to as a constraint , and the overall technique we develop here is called optimization with constraints. In the case that S can be realized as the graph of a function g : Rn1 R, we are reduced to the problem of nding extreme values of the function h : Rn1 R dened by h(x1 , . . . , xn1 ) = f (x1 , . . . , xn1 , g (x1 , . . . , xn1 )) on the open set Rn1 , which we have previously done. Even if the equation dening S cannot be solved explicitly for one of the variables, it may be still possible to describe S by an unconstrained set of n 1 variables by parametrization. While we will not pursue a full discussion of the notion of parametrization here, we will give a typical example of this approach. Example 4.8.1 Let S = {(x, y, z ) | x2 + y 2 + z 2 = 1} be the unit sphere in R3 . Suppose we wish to nd the maxima and minima of the function f (x, y, z ) = x + y + z on S . We may parametrize S by using spherical coordinates as follows. Recall that, in general, spherical coordinates are given by x = cos sin , y = sin sin , z = cos . In the specic case at hand, however, we have = 1, so S is parametrized by the two unconstrained variables and . Written in spherical coordinates, the function f becomes f0 (, ) = cos sin + sin sin + cos . This parameterization causes problems at points where the Jacobian of the change of variables map above is singular, and this occurs exactly when = 0 or sin = 0. In particular, we cannot determine whether points with sin = 0 are critical with this parameterization. We have f0 f0 = sin sin + cos sin = sin (cos sin ), = cos cos + sin cos sin = cos (cos + sin ) sin .

Note that f0 / = 0 when cos = sin or sin = 0. The former occurs when cos = sin = 22 . In sin = 0, that is, when tan = 2. this case, cos + sin = 2, so f0 / = 0 when 2 cos We have thus shown that points with sin = cos = 22 and tan = 2 (where the signs are the same) are among the critical points of f . Extreme value problems of the type discussed in Example 4.8.1 can be approached in a somewhat dierent fashion, by the method of Lagrange multipliers. The basic idea works as follows. Let S be the level set of a C 1 function g : Rn R. We have already observed that g (x) is normal to the tangent hyperplane at any point x S . Now suppose that a function f : Rn R, when restricted to S , has an extreme value at the point a S . Then, for any C 1 curve : R S with (0) = a, the function f : R R has an extreme value at 0, and hence 0 = (f ) (0) = f ((0)) (0) = f (a) (0).

In other words, f (a) is normal to the vector tangent to the curve at the point a. Since this is true for any such curve , f (a) is normal to the tangent hyperplane of S at a. Since g (a) is also normal to the tangent hyperplane of S at a, this implies that f (a) is proportional to g (a), so we can write f (a) = g (a) for some R. The scalar is known as a Lagrange multiplier. Before proving a theorem about Lagrange multipliers, let us study the example above using this new idea. 149

Example 4.8.2 Let S = {(x, y, z ) | x2 + y 2 + z 2 = 1} be the unit sphere in R3 . Suppose we wish to nd the maxima and minima of the function f (x, y, z ) = x + y + z on S . We observe rst that f (x, y, z ) = (1, 1, 1). The surface S can be realized as the level set g (x, y, z ) = 0 for the function g (x, y, z ) = x2 + y 2 + z 2 1, which has gradient g (x, y, z ) = (2x, 2y, 2z ). To identify the potential points a = (x0 , y0 , z0 ) on the surface S where f attains an extreme value, we set up the Lagrange multiplier equation f (a) = g (a), which becomes (1, 1, 1) = (2x0 , 2y0 , 2z0 ). This gives us three equations in our four unknowns, namely, 2x0 = 1, 2y0 = 1, 2z0 = 1. But we must not forget that we have the original constraint, namely, that g (a) = 0. And hence, there is a fourth equation, 2 2 x2 0 + y0 + z0 = 1 . Solving this system of four equations in four unknowns gives = 23 , which leads to x0 = 33 , y0 = 33 , z0 = 33 . The point a1 = (+ 33 , + 33 , + 33 ) is the point where f attains its maximum value of 3 when restricted to S , and the point a2 = ( 33 , 33 , 33 ) is the point where f attains its minimum value of 3 when restricted to S . Finally, it is worth noting that in this case, the precise value of is irrelevant and was merely used in an auxiliary fashion to help nd the potential extrema. Let us arm this method with a theorem. Theorem 4.8.3 Let U be an open set in Rn and let f : U R be C 1 . Let g : U R be C 1 and let S be the hypersurface dened by g (x) = 0. If f |S has a local maximum or minimum at a point a S , and g (a) = 0, then there exists R such that f (a) = g (a). Proof. By Exercise 4.9.14 in the next section, every vector in the tangent hyperplane to S at a may be realized as the tangent vector (0) to some C 1 curve : R S with (0) = a. By the discussion above, f (a) and g (a) both lie in the ane space normal to the tangent hyperplane. By Exercise 4.9.13, which you will also do in the next section, this space is one-dimensional, so we are done.

Example 4.8.4 Let us use the method of Lagrange multipliers to determine the points on the ellipse x2 + 4y 2 = 4 that are closest to and farthest from the point (1, 0). The square of the distance from the point (x, y ) to this point is given by f (x, y ) = (x 1)2 + y 2 and we wish to optimize this function subject to the constraint g (x, y ) = x2 + 4y 2 4 = 0. We have f (x, y ) = (2(x 1), 2y ) and g (x, y ) = (2x, 8y ), so we consider the equations 2(x 1) = 2y = (2x) (8y )

It is easy to see that the only points on the ellipse satisfying these equations for some are (2, 0). Plugging into f we conclude that (2, 0) is the nearest point and (2, 0) is the farthest point. Exercise 4.8.5 Let S = {(x, y, z ) | x2 + y 2 + z 2 = 1} be the unit sphere in R3 . Find the maxima and minima of the function f (x, y, z ) = x3 + y 3 + z 3 on S . 150

Exercise 4.8.6 Consider the function P : (0, ) (0, ) R given by P (L, C ) = La C b , where , a, b are positive constants, and a + b = 1. Let R : (0, ) (0, ) R be given by R(L, C ) = 1 L + 2 C for positive constants 1 and 2 . i. Maximize P subject to the constraint R(L, C ) = 1 , where 1 is a positive constant. ii. Minimize R subject to the constraint P (L, C ) = 2 , where 2 is a positive constant. In economics, the function P is known as the Cobb-Douglas production function. Exercise 4.8.7 Let x1 , x2 , . . . , xn be positive real numbers. Prove the arithmetic-geometric mean inequality, 1 x1 + x2 + + xn (x1 x2 xn ) n . n Hint : Consider the function f (x1 , x2 , . . . , xn ) = constant.
x1 +x2 ++xn n

subject to the constraint x1 x2 xn = c, a

Exercise 4.8.8 If a triangle has side lengths x, y , and z , so its perimeter is 2s = x + y + z , its area A satises A2 = s(s x)(s y )(s z ). (This is Herons Formula from classical geometry.) Show that, among all triangles with given perimeter, an equilateral triangle has the largest area. There is also a max-min theorem with several constraints. Theorem 4.8.9 Let U be an open set in Rn , let f : U R be C 1 , and let g1 , g2 , . . . , gm : U R be C 1 , where m < n. Let S be the intersection of the hypersurfaces Si dened by gi (x) = 0. If f |S has a local maximum or minimum at a point a S , and the vectors g1 (a), g2 (a), . . . , gm (a) form a linearly independent set, then f (a) is a linear combination of g1 (a), g2 (a), . . . , gm (a). If f (a) = 1 g1 (a)+ 2 g2 (a)+ + m gm (a) as in the theorem, the scalars 1 , 2 , . . . , m are called Lagrange multipliers. We do not prove this theorem here, but we present an example and some exercises to illustrate the theory. Example 4.8.10 Given the line dened by P (x, y ) = y (mx + k ) = 0, and an ellipse dened by E (x, y ) = y2 x2 a2 + b2 1 = 0, we wish to nd the minimum distance between a point on the line and a point on the ellipse. Equivalently, we minimize the square distance d(x1 , y1 , x2 , y2 ) = (x1 x2 )2 + (y1 y2 )2 subject to the constraints g1 (x1 , y1 , x2 , y2 ) = P (x1 , y1 ) = 0, g2 (x1 , y1 , x2 , y2 ) = E (x2 , y2 ) = 0. We assume that the line does not intersect the ellipse. The reader should verify that this will occur when 2 b2 |k | > |b| and m2 < k a . These conditions should also emerge from the solution below. 2 We have d g1 g2 = (2(x1 x2 ), 2(y1 y2 ), 2(x1 x2 ), 2(y1 y2 )), = (m, 1, 0, 0), 2x2 2y2 = 0, 0, 2 , 2 . a b 151

We rst note that g1 and g2 are everywhere linearly independent, so by Theorem 4.8.9, if (x1 , x2 , y1 , y2 ) is a maximum or minimum value for d subject to the constraints g1 and g2 , then the following system of six equations in six unknowns must be satised. 2(x1 x2 ) = 1 m,

2(y1 y2 ) = 1 , 2x2 2(x1 x2 ) = 2 2 , a 2 y2 2(y1 y2 ) = 2 2 , b y1 = mx1 + k,


2 x2 y2 2 + 1 = 0. a2 b2

Exercise 4.8.11 Solve the above system. Exercise 4.8.12 Consider the plane dened by P (x, y, z ) = Ax + By + Cz + D = 0 and the ellipsoid 2 y2 z2 dened by E (x, y, z ) = x a2 + b2 + c2 1 = 0. i. Find conditions on A, B , C , D, a, b, c such that the plane and the ellipsoid do not intersect. ii. Find the minimum distance between the plane and the ellipsoid when they do not intersect. Exercise 4.8.13 Let v and w be vectors in Rn . Find the maximum and minimum values of f (v, w) = v w subject to the constraints v = w = 1. Exercise 4.8.14 Consider two nonparallel planes in R3 . Find the point on their line of intersection closest to the origin in R3 . Exercise 4.8.15 In the situation of Theorem 4.8.9, what happens if the number of constraints exceeds or equals the number of variables, that is, if m n?

4.9

Implicit and Inverse Function Theorems

Let n and m be positive integers, and let f : Rm Rn Rm . We attack the problem of determining a set of conditions under which we can solve f = 0 explicitly for the rst m variables, which we denote with y s, in terms of the remaining n variables, which we denote with xs. Thus, if we write f in the form f (y1 , y2 , . . . , ym ; x1 , x2 , . . . , xn ) as a function of m + n variables, we would like to produce functions 1 , 2 , . . . , m from an open set in Rn to R such that, on some open set in Rm Rn , the assignment yj = j (x1 , x2 , . . . , xn ) solves the equation f (y1 , y2 , . . . , ym ; x1 , x2 , . . . , xn ) = 0. Of course, we can expect some conditions on dierentiability, nonvanishing of the determinant of a Jacobian matrix, and other properties. We begin by considering the case when m = 1 and n is arbitrary. Here, the proof involves only simple results from basic calculus. This development can be completed by induction on m, but the techniques are somewhat tedious. Example 4.9.1 Let f : R2 R be dened by f (y, x) = y 2 + x2 1. Here, the equation f (y, x) = 0 gives us the unit circle in R2 . As is ordinarily discussed in elementary calculus courses, this analysis produces two functions that serve our y in terms of x, and this can be done with either purpose here. We wish to solve for of the equations y = 1 x2 , for 1 x 1, or y = 1 x2 , for 1 x 1. Note that, in either case, we have f (y, x) = 0. Also, f y = 2y = 0 when y = 0. This condition about the nonvanishing of the derivative is one that will be required in the next theorem. 152

Theorem 4.9.2 Let f be a C 1 function from an open set in R Rn to R. Let (y0 , x0 ) be a point in this open set such that f (y0 , x0 ) = 0, and f (y0 , x0 ) = 0. y Then, there exist open sets V R and U Rn , such that (y0 , x0 ) V U , and to every x U , there exists a unique (x) in V such that f ((x), x) = 0, and : U R is C 1 on U . Furthermore,
f x ((x), x) (x) = fj . xj y ((x), x)

Proof. We can assume that

f y (y0 , x0 )

an open interval V = (y0 , y0 + ) R such that f y (y, x) > 0 for x U , y V . Then f (y0 + , x0 ) > 0, and f (y0 , x0 ) < 0. By continuity of f , there exists an open set U U Rn containing x0 such that, for all x U , f (y0 , x) < 0 and f (y0 + , x) > 0. Fix x U . By the Intermediate Value Theorem, there exists y V such that f (y, x) = 0. The function g (y ) = f (y, x) satises g (y ) = f y (y, x) > 0 for y V , so by Rolles Theorem, the value of y for which f (y, x) = 0 is unique. We set (x) = y . The continuity of at x0 follows from the fact that we can choose in the above construction to be arbitrarily small. This same argument holds for any x U , which proves the continuity of . Since f ((x), x) = 0, formally, from the chain rule, we get

> 0. Since f is C 1 , there is an open set U Rn containing x0 and

0= Thus,

f f [f ((x), x)] = (x) + ((x), x). ((x), x) xj y xj xj


f x ((x), x) . (x) = fj xj y ((x), x)

Using the formal expression for the derivative given above, we can write the dierence quotient for the derivative of , subtract this expression, and show that the limit of the dierence goes to zero. A continuation of this proof to yield the Implicit Function Theorem for general m is outlined in Osgood []. If the reader enjoys a sti climb over rocks and across streams, they might wish to pursue this proof. As an alternative, they might wish to consult Follands Appendix []. We now make an about-face and move directly to a proof of the Inverse Function Theorem. This seems to be the more common approach in current mathematics texts. We rst review the single-variable case, to remind the reader of the nature of the result. Suppose that U is an open set in R and f : U R is C 1 . Take a point x0 U . We saw earlier in the chapter that if f (x0 ) = 0, then f is monotonic in an open interval I around x0 . This, of course, implies that f is one-to-one on I . Moreover, f (I ) is an open interval J contained in R, and f 1 : J I is C 1 and (f 1 ) (y ) = (f (f 1 (y )))1 . It is worth remarking at this point that this one-variable theorem requires the continuity of the derivative. See, for example, Exercise 4.1.22. The Inverse Function Theorem is the generalization of this result to functions f : Rn Rn . Essentially, the theorem says that if such an f is C 1 and has a nonsingular derivative at a point x0 , then, in some neighborhood of x0 , f is invertible, and f 1 is also C 1 . We approach this through a sequence of lemmas and corollaries, which can be combined to provide a proof of the Inverse Function Theorem. Lemma 4.9.3 Let U Rn be open and let f : U Rn be C 1 . Take x0 U and suppose that Df (x0 ) is nonsingular. Then there exists a neighborhood W of x0 and a constant c > 0 such that f (y ) f (x) c y x for all x, y W. 153

Proof. For any nonsingular linear transformation T : Rn Rn , we know that T (y ) T (x) T It follows immediately that T (y ) T (x) T 1 1 y x .

yx .

Take c = Df (x0 )1 1 /2. Suppose that f1 , f2 , . . . , fn are the coordinate functions of f . Of course, these are C 1 , so that there exists a convex neighborhood W of x0 such that Dfi (y ) Dfi (x0 ) c/n for y W and all i. Now the Mean Value Theorem implies that, given x, y W , there exists a point i on the line segment joining x and y such that fi (y ) fi (x) = Dfi (i )(y x). Consequently, for x, y W and each i, we have fi (y ) fi (x) Dfi (x0 )(y x) c yx . n

It follows immediately that f (y ) f (x) Df (x0 )(y x) c y x . Now, using the triangle inequality, we get f (y ) f (x) c y x for x, y W.

Corollary 4.9.4 Let U Rn be open and f : U Rn be C 1 . Take x0 U and suppose that Df (x0 ) is nonsingular. Then there exists a neighborhood W of x0 such that f |W is one-to-one. Proof. Exercise. Corollary 4.9.5 Let U Rn be open and f : U Rn be C 1 . Take x0 U and suppose that Df (x0 ) is nonsingular. Then there exists a neighborhood V of x0 such that f (V ) is open, and f |V : V f (V ) is a homeomorphism. Proof. Using the previous lemma we can pick a neighborhood W of x0 such that W U and for some constant c, f (y ) f (x) c y x for all x, y W , and nally Df (x) is nonsingular for x W . Let / f (S ), since V be any open ball contained in W and let S = V . Given a point x V with y = f (x) f (S ) is compact, the distance from y to f (S ), which we denote by d, is greater than zero. To show that f (V ) is open, we establish that f (V ) contains Bd/2 (y ). To see this, take z Bd/2 (y ). Then z y < d/2. Moreover, the distance from z to f (V ), which equals
x V

inf z f (x) ,

is less than d/2. Since the distance from y to f (S ) is equal to d, it follows from the triangle inequality that the distance from z to f (S ) is greater than d/2. For x V , we dene the function
n

g (x) = z f (x)

=
i=1

(zi fi (x))2 .

We want to minimize this function. We know that there exists x1 V such that z f (x1 )
2

= d(z, f (V ))2 .

154

From the previous inequalities, it follows that x1 / S , so that x1 V . So the minimum of g occurs at x1 . This implies that
n

0 = Dj g (x1 ) = 2

i=1

(zi fi (x1 ))Dj fi (x1 ).

It follows immediately that Df (x1 )(z f (x1 )) = 0, and since Df (x1 ) is nonsingular, z = f (x1 ). Since f is invertible on V by the previous corollary, f is a homeomorphism.

Theorem 4.9.6 (Inverse Function Theorem) Let U be an open set in Rn and let f : U Rn be C 1 . Let x0 U be such that Df (x0 ) is nonsingular. Then there exists a neighborhood V of x0 such that i. f : V f (V ) is a bijection; ii. f (V ) is an open set in Rn ; iii. f 1 : f (V ) V is C 1 and Df 1 (f (x)) = (Df (x))1 for x V . Proof. The rst two statements follow from Corollary 4.9.4 and Corollary 4.9.5. Now where are we? We have proved that f : V f (V ) is one-to-one and that f (V ) is open. We consider the map f 1 : f (V ) V . By the last corollary, this map is continuous. We want to prove that it is C 1 . A formal computation using the chain rule shows that Df 1 (f (x)), if it exists, is equal to (Df (x))1 . To complete the proof, we take y = f (x) f (V ) and consider the dierence quotient f 1 (z ) f 1 (y ) Df (x)1 (z y ) . zy We can write this as Df (x)1 [Df (x)(f 1 (z ) f 1 (y )) (z y )] , zy Df (x)1 Df (x)(f 1 (z ) f 1 (y )) (z y ) . zy Df (x)(f 1 (z ) x) (z f (x)) . f 1 (z ) x

which is less than or equal to

And using the inequality of Lemma 4.9.3, this is less than or equal to Df (x)1 c

Finally, since f 1 is continuous, this last expression goes to zero as z goes to y . The function Df 1 is continuous because f is C 1 and, by the exercise below, the map A A1 is continuous from GLn (R) GLn (R).

Exercise 4.9.7 Prove that the map f : GLn (R) GLn (R) given by f (A) = A1 is continuous. We now use the Inverse Function Theorem to outline a proof of the Implicit Function Theorem in a sequence of exercises. Theorem 4.9.8 (Implicit Function Theorem) Let f be a C 1 function from an open set in Rm Rn to Rm . Let (y0 , x0 ) be a point in this open set such that f (y0 , x0 ) = 0, and the matrix L= fi (y0 , x0 ) yj
n

i,j =1,...,m

is nonsingular. Then there exist open sets U R and V Rm such that (y0 , x0 ) V U , and to every x U , there exists a unique y = (x) in V such that f ((x), x) = 0, and is C 1 on U . 155

Proof. To begin, we construct the function F : Rm Rn Rm Rn dened by F (y, x) = (f (y, x), x). Exercise 4.9.9 Show that det DF (y0 , x0 ) = det L. By applying the Inverse Function Theorem to F , we obtain neighborhoods U Rn of x0 and V Rm of y0 such that F has a C 1 inverse on the open set W = F (V U ). Let U = {x U | (0, x) W }. Exercise 4.9.10 Show that there exists a C 1 function : W Rm such that F 1 (y, x) = ((y, x), x) on W. Exercise 4.9.11 Show that f ((y, x), x) = y . Exercise 4.9.12 Let (x) = (0, x). Show that satises the conclusion of the theorem. This concludes the proof of the implicit function theorem.

Exercise 4.9.13 Let F : Rn+1 R be a C 1 function, and let S be the hypersurface dened by F (x) = 0. Suppose that F (x) = 0 for all x S , and x x0 S . Show that the tangent hyperplane to S at x0 is spanned by n linearly independent vectors. Hint : Use the Implicit Function Theorem, and apply Exercise 4.6.8. Exercise 4.9.14 Let F : Rn+1 R be a C 1 function, and let S be the hypersurface dened by F (x) = 0. Suppose that F (x) = 0 for all x S , and x x0 S . Show that every vector in the tangent hyperplane to S at x0 is the tangent vector to some C 1 curve in S through x0 . Hint : Use the Implicit Function Theorem, and apply Exercise 4.6.9. Exercise 4.9.15 Let f : R3 \ {(0, 0, 0)} R3 \ {(0, 0, 0)} be given by f (x, y, z ) = x y z , , x2 + y 2 + z 2 x2 + y 2 + z 2 x2 + y 2 + z 2 .

Show that f is locally invertible at every point in R3 \ {(0, 0, 0)}. Find an explicit formula for f 1 . Exercise 4.9.16 Consider the equations ab2 + cde + a2 d = 3 and ace3 + 2bd b2 e2 = 2.

Determine which pairs of variables can be solved for in terms of the other three near the point (a, b, c, d, e) = (1, 1, 1, 1, 1). We now present a result, Lemma 4.9.19, related to the previous theorems that will be of assistance in the change-of-variables theorem for multiple integrals. Given a function : Rn Rn we want to dene the best linear approximation to at a point using the Taylor polynomials of its real-valued coordinate functions. Let U Rn be an open set and let : U Rn be a C 1 function. For y U , dene a function T y (x) by y y y T y (x) = (T1 (x), T2 (x), . . . , Tn (x)) where
n

Tjy (x) = j (y ) +
k=1

Dk j (y )(xk yk ).
x y x y

This is the rst-order Taylor polynomial of j in the direction of the unit vector x y in the notation of Denition 4.5.9. 156

, evaluated at t =

Lemma 4.9.17 Let K U be compact. Then


xy

lim

T y (x) (x) =0 xy

uniformly in y , for y K . Proof. It is enough to prove the analogous statement for each component function separately. By the Mean Value Theorem, there exists [0, 1] such that
n

|Tjy (x) j (x)|

=
k=1

(Dk j (y ) Dk j (y + (x y )))(xk yk ) Dj (y ) Dj (y + (x y ))

xy

by the Cauchy-Schwarz inequality. Dividing both sides by x y , we get |Tjy (x) j (x)| Dj (y ) Dj (y + (x y )) . xy The right-hand side goes to zero as x y uniformly in y , for y K , since is C 1 and K is compact. Denition 4.9.18 We dene a generalized rectangle in Rn to be a set of the form R = I1 I2 In , where Ii , i = 1, . . . , n, are bounded intervals in R. If the intervals are all open, we call R an open generalized rectangle and if the intervals are all closed, we call R a closed generalized rectangle. In the particular case when the intervals are of the form Ii = [ai , bi ), we refer to R as a half-open generalized rectangle. Lemma 4.9.19 Let U Rn be open, and suppose that : U (U ) is C 1 , one-to-one onto its image, and has C 1 inverse. Let R U Rn be a generalized rectangle with nonempty interior. For y U and > 0, we denote by Ry the generalized rectangle with center y similar to R with sides scaled by . For 0 < < 1 and h > 0, we dene R1 = R1 (h, y ) = R2 = R2 (h, y ) = R3 = R3 (h, y ) = (1 )hRy

hRy

(1 + )hRy

Then, for each compact set K U , there exists a number h0 (K ) > 0 such that, if 0 < h < h0 (K ) and y K is such that R2 (h, y ) U , then T y (R1 (h, y )) (R2 (h, y )) T y (R3 (h, y )). Proof. Since T y has a continuously dierentiable inverse, there exists a constant C > 0 such that x z C T y (x) T y (z ) for all x, y, z K . If we apply this estimate to the previous lemma, we have T y (x) (x) 0 T y (x) (y ) as x y uniformly in y . Then sup
xR2 (h,y )

T y (x) (x) 0 T y (x) (y ) 157

as h 0, so there exists a constant h0 (K ) > 0 such that, for 0 < h < h0 (K ), (R2 ) T y (R3 ) \ T y (R1 ). It follows that (R2 ) T y (R3 ). Since (R2 ) has nonempty interior and contains (y ) = T y (y ), there exists some h > 0 such that y y T (h R ) (R2 ). Let h = sup{h | T y (h Ry ) (R2 )}. Then T y (h Ry ) (R2 ) = , and hence h > (1 )h. Thus, T y (R1 ) (R2 ).

4.10

Independent Projects

4.10.1 Term-by-Term DierentiationThe question we discuss here is the following. Let U R be an open set, and let [a, b] U . If we have a series of real-valued, dierentiable functions k=1 fk on U that converges on [a, b] to f , is the series dierentiable, and if so, does f =
k=1 fk

on [a, b]? The answer is generally no. n However, if the series k=1 fk converges uniformly on [a, b], the answer is yes. Let Sn (x) = k=1 fk (x), n Rn (x) = f (x) Sn (x), and f (x) = k=1 fk (x) for x [a, b], and take > 0. Then Sn (x) = k=1 fk (x) on [a, b]. Since the series of derivatives converges uniformly, there exists N N such that, for all n > N ,
1. |Sn (x) f (x)| /8 on [a, b] ; 2. |Sn +p (x) Sn (x)| /8 on [a, b], for all p 0.

Take n > N and p 0. Exercise 4.10.1 In this exercise, you will show that for all x [a, b] and h = 0 such that x + h [a, b], we have Rn (x + h) Rn (x) . h 8 i. Assume there exists x and h as above such that Rn (x + h) Rn (x) = + , h 8 where > 0. By suitable maneuvering, show that Rn (x + h) Rn (x) h Sn+p (x + h) Sn (x + h) Sn+p (x) Sn (x) = h h f (x + h) Sn+p (x + h) f (x) Sn+p (x) + h h ii. Fixing n, show that if we take p large enough, f (x + h) Sn+p (x + h) 1 1 f (x) Sn+p (x) < , and < . h 2 h 2 158

+ . 8

This yields the inequality Sn+p (x + h) Sn (x + h) Sn+p (x) Sn (x) > . h h 8 iii. By applying the Mean Value Theorem, show that
|S n +p (x1 ) Sn (x1 )| >

for some point x1 between x and x + h, which contradicts (2). We now have f (x + h) f (x) Sn (x + h) Sn (x) Rn (x + h) Rn (x) = h h h 8 for all x [a, b] and h = 0 such that x + h [a, b]. Since Sn is dierentiable at x, there exists > 0 such that if 0 < |h| < , then Sn (x + h) Sn (x) Sn (x) . h 8

Exercise 4.10.2 Show that f (x + h) f (x) Sn (x) + = , h 8 8 4 and use this inequality to conclude that f (x) = f (x). Exercise 4.10.3 Show that the series for f must converge uniformly on [a, b]. Exercise 4.10.4 By giving a counterexample, show that if the series for f does not converge uniformly, then the conclusion can be false. (Hint: Try nding a series of dierentiable functions that converges to f (x) = |x| on [1, 1].) Exercise 4.10.5 How can one apply this result to Exercise 4.3.12? Exercise 4.10.6 i. Let f (x) = ex . Use the fact that y = f (x) satises the dierential equation y = y to nd the power series for f . ii. Let f (x) = sin x and g (x) = cos x. Using the fact that both y = f (x) and y = g (x) satisfy the dierential equation y = y , nd the power series for f and g . ( iii. Hint: f is an odd function, and g is an even function.) Exercise 4.10.7 Show that the argument is considerably simplied if you use the Fundamental Theorem of Calculus at an appropriate juncture. (If you dont remember the Fundamental Theorem, see the next chapter.) 159

4.10.2 Leibnizs RuleIn this project, we use elementary properties of the Riemann integral in one variables, including properties of improper integrals. We treat these topics in greater detail in the next chapter. The classical theorem of Leibniz about dierentiation under the integral sign states the following. Let U be an open set in R2 , and let R = [a, b] [c, d] be a closed rectangle contained in U . Let f (x, y ) be a continuous, real-valued function on U such that D2 f (x, y ) exists and is continuous on U , and let
b

F (y ) =
a

f (x, y ) dx.

Then the derivative of F exists and F (y ) =


a

D2 f (x, y ) dx.

Exercise 4.10.8 Show that the above conclusion is equivalent to the statement lim F (y + h) F (y ) h
b

h0

D2 f (x, y ) dx = 0.
a

Exercise 4.10.9 Use the Mean Value Theorem to show


b

F (y + h) F (y ) = for some 0 1.

D2 f (x, y + h) dx
a

Exercise 4.10.10 Use continuity (uniform continuity) to show that F (y ) =

b a

D2 f (x, y ) dx.

There is a more general version of Leibnizs rule using variable bounds of integration. Let f (x, t) and D2 f (x, t) be continuous on the domain [a, b] V where V R is open. Assume , : V [a, b] are C 1 functions. Dene
(t)

(t) =
(t)

f (x, t) dx.

Exercise 4.10.11 Show that is dierentiable on V , and (t) = f ( (t), t) (t) f ((t), t) (t) +
(t) (t)

f (x, t) dx. t

The next exercise presents an interesting result that can be proved in several ways. This result is
0

sin(x) dx = . x 2

Exercise 4.10.12 Let f (x, t) = ext sin(x)/x for t, x > 0, and let (t) =
0

f (x, t) dx.

i. Show that (t) 0 as t . ii. Use integration by parts twice to show that (t) = 160 1 . 1 + t2

iii. Conclude that (t) = for t > 0. iv. Use integration by parts to show that
r

tan1 (t) 2

f (x, t) dx

2 r

for all r > 0 and t > 0. v. Show that


t0 r r

lim +

f (x, t) dx =
0 0 0

sin(x) dx. x

vi. Show that

sin(x) dx = . x 2

An interesting article about Leibnizs rule by H. Flanders can be found in the American Mathematical Monthly, JunJuly 1973.

161

162

Chapter 5

Integration on Rn
The integral of a function over a region R is designed for numerous applications. The most basic of these is the determination of the area under a curve of a function of one real variable dened on a closed interval [a, b] R. The idea moves in a number of directions, such as volume in Rn , arc length of a curve, ux through a surface, and other applications. Whereas dierentiation is designed to describe the behavior of a function at a point or in the neighborhood of a point, integration is intended to study the cumulative properties of a function on a larger domain.

5.1

The Riemann Integral in One Variable: Denitions

We begin this chapter on integration in Rn with a discussion of the Riemann integral in R. We start with a closed, bounded interval [a, b] R and a bounded function f : [a, b] R. As we proceed through the denition, we will derive conditions on the function f that ensure the existence of the Riemann integral. Denition 5.1.1 A partition P of the interval [a, b] is a nite set of points P = { a0 , a1 , . . . , ak } such that a = a0 < a1 < < ak1 < ak = b. The mesh of the partition P is |P | = max (ai ai1 ).
1ik

Let f be a bounded function on [a, b] with m f (x) M for x [a, b]. Denition 5.1.2 Let [a, b] R, and let f : [a, b] R be a bounded function. If P is a partition of [a, b], we set mi = inf x[ai1 ,ai ] f (x), and Mi = supx[ai1 ,ai ] f (x). We dene the lower sum of f on [a, b] relative to the partition P to be
k

L(f, P ) =
i=1

mi (ai ai1 ),

and the upper sum of f on [a, b] relative to P to be


k

U (f, P ) =
i=1

Mi (ai ai1 ).

Exercise 5.1.3 Let f be a bounded function on [a, b] with m f (x) M for x [a, b]. Given a partition P of [a, b], show that m(b a) L(f, P ) U (f, P ) M (b a). 163

1 1 Example 5.1.4 Let [a, b] = [1, 1], and let f (x) = x2 . Consider the partition P = {1, 2 , 0, 2 , 1}. Then 1 1 1 m1 = m4 = 4 , m2 = m3 = 0, M1 = M4 = 1, M2 = M3 = 4 , so that L(f, P ) = 4 , and U (f, P ) = 5 4 . Now 1 1 2 349 853 3 , 2 , 1 , 0 , , , 1 } . Then L ( f, P ) = , and U ( f, P ) = . consider the partition P = {1, 4 4 3 3 864 864

Exercise 5.1.5 Let f (x) = sin x on [0, 2 ]. Compute the upper and lower sums for f relative to the partitions P = { k 4 | k = 0, 1, . . . , 8} and P = {0, 1, 2, 3, 4, 5, 6, 2 }. Exercise 5.1.6 Let [a, b] = [0, 2], f (x) = U (f, P ) and L(f, P ). 1 if 0 x < 1, For every partition P of [0, 2], determine 1 if 1 x 2.

Denition 5.1.7 Let P = {aj | j = 0, 1, 2, . . . , k } be a partition of [a, b]. A partition P = {a j | j = 0, 1, 2, . . . , k }, where k k , is called a renement of P if P P . Exercise 5.1.8 Let f : [a, b] R be a bounded function. Let P be a partition of [a, b], and let P be a renement of P . Show that L(f, P ) L(f, P ) U (f, P ) U (f, P ). Exercise 5.1.9 Let f : [a, b] R be a bounded function, and let P1 and P2 be any two partitions of [a, b]. Show that L(f, P1 ) U (f, P2 ). We are now ready to dene the lower and upper integrals of f on [a, b]. Denition 5.1.10 Let f : [a, b] R be a bounded function. 1. We dene the lower integral of f on [a, b] to be
b a

f = lub{L(f, P ) | P is a partition of [a, b]}.

2. We dene the upper integral of f on [a, b] to be


b a

f = glb{U (f, P ) | P is a partition of [a, b]}.

Note that the set of lower sums is bounded above by M (b a), and the set of upper sums is bounded below by m(b a), and therefore the lower and upper integrals exist for any bounded function. Exercise 5.1.11 Let f : [a, b] R be a bounded function. Show that Exercise 5.1.12 Let [a, b] = [0, 2], and let f (x) =
b af

b a f. 2 0f 2 0f

1 if 0 x < 1, Show that 1 if 1 x 2. 1 0 if x Q, Find if x / Q.


b a 1 0f

= 0.

Exercise 5.1.13 Let [a, b] = [0, 1], and let f (x) =

and

1 0 f.

Denition 5.1.14 Let f be a bounded function on an interval [a, b]. We say that f is Riemann integrable on [a, b] if on [a, b].
b f a

b f, a

and we denote this common value by

f , which we call the Riemann integral of f

We often refer to a function that is Riemann integrable as simply integrable. Exercise 5.1.15 Let f : [a, b] R be a constant function. Show that f is Riemann integrable on [a, b] and compute the value of its integral. 164

5.2

The Riemann Integral in One Variable: Properties

Theorem 5.2.1 Let f be a bounded function on an interval [a, b]. Then f is integrable on [a, b] if and only if, given > 0, there exists a partition P of [a, b] such that U (f, P ) L(f, P ) < . Proof. This useful criterion follows directly from the denitions of integrability and the upper and lower integrals. Exercise 5.2.2 Let f : [a, b] R be bounded on [a, b]. i. Show that f is integrable on [a, b] if and only if, given > 0, there exists a > 0 such that if P is a partition with |P | < , then U (f, P ) L(f, P ) < . ii. Suppose f is integrable on [a, b], and we have a sequence of partitions (Pk )kN such that limk |Pk | = b 0. Show that a f = limk L(f, Pk ) = limk U (f, Pk ). Theorem 5.2.3 Suppose f : [a, b] R is continuous. Then f is Riemann integrable on [a, b]. Proof. Fix > 0. Since f is continuous on [a, b], a compact set in R, it follows from Exercise 3.6.10.i that f is uniformly continuous on [a, b]. So there is a > 0 such that |x1 x2 | < implies |f (x1 ) f (x2 )| < b a. Choose a partition P such that |P | < . Then
k

U (f, P ) L(f, P ) =

i=1

(Mi mi )(xi xi1 )


k i=1

ba

(xi xi1 )

= .

The following exercises give the standard properties of the Riemann integral. Exercise 5.2.4 Suppose that f, g : [a, b] R are integrable on [a, b], and R. Show that i. ii.
b (f a b a

+ g) =
b a

b a

f+

b a

g;

f =

f;
b a

iii. if c [a, b], then f is integrable on [a, c] and [c, b], and iv. if f (x) g (x) for all x [a, b], then v. |f | is integrable, and
b a b a

f=

c f a

b c

f;

b a

g ; and

b a |f |.

Exercise 5.2.5 Show that the converse to part v above is false by giving an example of a function f such that |f | is integrable on some [a, b], but f is not. Exercise 5.2.6 Which of the statements in Exercise 5.2.4 are true for lower or upper integrals? 165

Example 5.2.7 Let f (x) = x2 on [1, 1]. Then f is integrable on [1, 1] by Theorem 5.2.3, and we 1 compute 1 f as follows. By Exercise 5.2.2.ii, it suces to consider partitions whose mesh goes to zero, and we make our computation easier by choosing regular partitions. That is, for each k N, we let 1) 1 2 1 Pk = {1, 1 + k , 1 + k , , 1 + 2(kk , 1}. Note that |Pk | = k . Furthermore, it suces to compute the upper integral. We have
2k

U (f, Pk ) =
i=1 k

Mi

1 k 1 k 1 k

=2
i=1 k

M k +i i k i2
2

=2
i=1

2 k3

k i=1

2 k (k + 1)(2k + 1) = 3 k 6 by a simple high school formula. Then


1

f = lim
1

2 2 k (k + 1)(2k + 1) = . 3 k 6 3

Exercise 5.2.8 i. Let f (x) = x3 on [0, 2]. Compute ii. Let f (x) =
1 q 2 0

f. when x = p q Q in lowest terms, x = 0, otherwise.


1 0

Show that f is integrable, and compute

f.

Exercise 5.2.9 Suppose f is Riemann integrable on [a, b]. Take a partition P , and choose a point a i k [ai1 , ai ]. The Riemann sum of f with respect to P and the selection of points ai is i=1 f (ai )(ai ai1 ). Show that, for any sequence of partitions (Pj )j N such that limj |Pj | = 0, any set of associated Riemann b sums converges to a f . Theorem 5.2.10 Let (fk )kN be a sequence of Riemann integrable functions dened on [a, b]. Suppose b that (fk )kN converges uniformly to a function f on [a, b]. Then f is Riemann integrable, and a f = b limk a fk . Proof. First from Theorem 3.4.23, we know that f is bounded. Given > 0, there exists k such that |fk (x) f (x)| < 3(b a) for all x [a, b]. Since fk is integrable on [a, b], there exists a partition P of [a, b] . Then such that U (fk , P ) L(fk , P ) < 3 U (f, P ) L(f, P ) |U (f, P ) U (fk , P )| + |U (fk , P ) L(fk , P )|+|L(f, P ) L(fk , P )| < . So f is integrable by Theorem 5.2.1, and it is clear that limk
b a

fk =

b a

f.

166

5.3

The Fundamental Theorem of Calculus and Its Consequences

The next theorem is often regarded as the most important theorem in integral calculus in one variable. This seems to be based on two facts: one is that it allows some interesting applications of the integral that are important both in mathematics and other sciences; the other is that it allows explicit computation of the integrals of a few functions. These few functions seem capable of providing an indenite set of questions on calculus exams. As we shall see, this type of application depends on nding a function F whose derivative is f . This is a truly special, limited class of elementary functions that are usually listed in the endpapers of calculus textbooks. On the other hand, we know from Theorem 5.2.3 that every continuous function is integrable, leaving us with the prospect of eternally dancing on a very small dance oor with a limited number of partners. Theorem 5.3.1 (Fundamental Theorem of Calculus) Let f be Riemann integrable on [a, b]. For x x [a, b], dene F (x) = a f . If f is continuous at a point x (a, b), then F is dierentiable at x, and F (x) = f (x). Proof. Fix > 0. Consider the dierence quotient F (x + h) F (x) 1 = h h
x +h a x

f
a

.
x +h f. x

1 Assume rst that h is positive. Then the dierence quotient is equal to h x, we can choose h so small that |f (x + t) f (x)| < for t [0, h]. Then x +h

Since f is continuous at

(f (x) )h < So f (x) < is similar. 1 h

f < (f (x) + )h.


x

x +h

f < f (x) + .
x

Since was arbitrary, it follows that the dierence quotient converges to f (x). For h negative, the argument

Corollary 5.3.2 Suppose f is continuous on [a, b], and G is continuous on [a, b] and dierentiable on (a, b) b with G (x) = f (x) for x (a, b). Then a f = G(b) G(a). Moreover, if F is dened as above, then F (x) = G(x) G(a). Proof. Since G = F on (a, b), it follows from Corollary 4.1.16 that there exists a constant C such that F (x) G(x) = C . It is clear that C = G(a). Remark 5.3.3 In the language of the preceding corollary, the function G is often referred to as a primitive or an antiderivative of f . A substantial portion of many single-variable calculus courses consists of a search for primitives. For the next corollary, we need to introduce the traditional notation for the integral. That is, if f is b b b integrable on [a, b], we write a f = a f (x) dx = a f (t) dt, where x, t, and any other symbol are called dummy variables. For the moment, these expressions should be thought of in terms of the bookkeeping they allow us to do. We will give a proper denition of the dierential dx in Chapter 6. 167

Corollary 5.3.4 (Change of variables in R) Let f be continuous on [a, b], and let : [c, d] [a, b] be continuous on [c, d] and C 1 on (c, d) such that (c) = a, (d) = b, and (x) > 0 for all x (c, d). Then
b d

f (u) du =
a c

f ((x)) (x) dx.

Remark 5.3.5 To emphasize the lack of a role that these dummy variables play, we could also write the b d conclusion as a f = c (f ) . Exercise 5.3.6 Prove this statement using the chain rule and the fundamental theorem of calculus. Exercise 5.3.7 Prove this statement without using the fundamental theorem of calculus. The change of variables theorem is what allows the classic integration technique known as integration by substitution. We illustrate this with the following examples.
d Example 5.3.8 Suppose we wish to evaluate c x + dx with > 0 and d > c . The function g (x) = x + is not among the small number of functions that have an obvious antiderivative. However, 1 we can write g (x) = f ((x)) (x), where f (u) = u, (x) = x + , and (x) = . The change of variables formula then tells us that d d

x + dx =
c c

f ((x)) (x) dx f (u) du,

(d )

=
(c )

d+

=
c+

1 u du.

1 u is a function one of whose The advantage of having used the change of variables formula is that f (u) = 3 2 . Corollary 5.3.2 now tells us that this integral evaluates to primitives everyone knows to be G(u) = 32 u

G(d + ) G(c + ) =
b

3 3 2 (d + ) 2 (c + ) 2 . 3

1 Example 5.3.9 Suppose we wish to evaluate 0 1+ u2 du. The thorough reader will recall this example from Project 4.10.2. Our goal here is to nd a function u = u(x) satisfying the conditions of Corollary 5.3.4 such that

can be computed directly. The technique of trigonometric substitution from a standard single-variable calculus course suggests the substitution u(x) = tan x. Then u (x) = sec2 x, and u (x) 1 1 1 1+u(x)2 = 1+tan2 x = sec2 x , so the expression 1+u(x)2 = 1. Our integral then becomes
tan1 (b) tan1 (0)

u1 (b) 1 u1 (0) 1+u(x)2 u (x) dx

1 dx = tan1 (b) tan1 (0) = tan1 (b).


1 1+x2 .

Exercise 5.3.10 Use implicit dierentiation to show that if tan y = x then y = Exercise 5.3.11 Let > 0, and let f : [a, b] R be continuous. Show that Another classic technique is that of integration by parts. Corollary 5.3.12 (Integration by parts) Suppose f and g are C 1 . Show that f (a)g (a)
b a b a

f (x) dx =

b a

dx

b a

f (x)g (x) dx. 168

f (x)g (x) dx = f (b)g (b)

Proof. Use the product rule and Corollary 5.3.2. Corollary 5.3.13 (Mean value theorem for integrals) Let f be continuous on [a, b]. Then there exists b 1 a point c (a, b) such that f (c) = b a a f. Proof. Apply the mean value theorem for derivatives to the function F (x) =
1 ba b a x a f.

Remark 5.3.14 The quantity

f is called the average value of f on [a, b].

Exercise 5.3.15 Let f : [a, b] R be a bounded integrable function. For each n N, we dene An = 1 n
n

f
i=1

a+

(2i 1)(b a) 2n

Show that limn An exists and is equal to the average value of f . (Hint : What does the formula for An represent?) Exercise 5.3.16 A train travels from Chicago to St. Louis in 4 hours. Show that at some point the speed of the train must exceed 60 miles per hour.

5.4

Principal Value Integrals

So far, in this chapter, we have worked with bounded, real-valued functions dened on a closed, bounded interval in R. For many reasons, in mathematics and its applications, it is useful and even necessary to extend the denition of integrals to both unbounded functions and unbounded intervals. Here, we present some cases of this extension, and the reader can create additional cases from these. We have already used these ideas in Project 4.10.2 in the previous chapter. First, we consider the case of a bounded, real-valued function f dened on an unbounded interval of the form [a, ). Denition 5.4.1 Suppose that f : [a, ) R is a bounded function, and suppose further that for all b b > a, the function f is integrable on [a, b]. The principal value integral, a f , is dened by limb a f if this limit exists. Example 5.4.2 Let f (x) =
1 x

on [1, ), where > 0. Then, for any b > 1, f is integrable on [1, b], and
b 1

1 dx = x

ln b

b1 1 1

if = 1, if = 1.

1 If > 1, then the principal value integral exists and is equal to 1 . If 1, the limit diverges, so the principal value integral does not exist.

Exercise 5.4.3 Decide whether the following principal value integrals exist, and if so, evaluate them. i. ii. iii. iv.
x dx 0 e 0

sin x dx dx dx 169

sin x x 0

sin x x 0

v.

1 2 x ln x

dx

Remark 5.4.4 You will notice in the above exercise that the principal value integral exists in some cases because we are simply extending the length of the interval from bounded to unbounded while the function itself is nonnegative on the entire interval, whereas in other cases the principal value integral exists due to cancelation in dierent parts of the interval of integration. Exercise 5.4.5 Dene the principal value integral
b

f including any necessary hypotheses on f .

The case of the principal value integral of a bounded real-valued function on R is more subtle. Exercise 5.4.6 i. Show that lima ii. Show that lima iii. Show that lima
a a a a a a

sin x dx = 0. cos x dx does not exist. sin(x + b) dx exists if and only if b = n for some n Z.

Since the functions we are attempting to integrate in the preceding exercise are essentially the same, that is, they are horizontal translates of each other, the values of their integrals should be the same, assuming these integrals exist. So how do we proceed? Suppose f is a bounded function on R, and suppose further that f is integrable on any closed interval [a, b]. There are two potential impediments to f being integrable on R: the rst is the behavior of f toward +, and the second is the behavior of f toward . We ensure that neither of these impediments is too great by dening the principal value integral of f on R as follows. Denition 5.4.7 Suppose that f : R R is a bounded function, and suppose further that f is integrable on 0 any closed bounded interval in R. The principal value integral of f on R is dened as f = f + 0 f , provided that the latter two principal value integrals exist. Exercise 5.4.8 Show that if the principal value integral of f on R exists, then

f = lima

a f. a

Next, we consider unbounded functions on bounded intervals. We treat an important special case rst, from which the general theory will follow. Denition 5.4.9 Let f : (a, b] R be a function. Suppose further that for all c (a, b), f is bounded b b and integrable on [c, b]. The principal value integral of f on (a, b] is a f = limca+ c f , provided this limit exists. Example 5.4.10 Let f (x) = and
1 x

on (0, 1], where > 0. Then, for any c (0, 1), f is integrable on [c, 1],
1 c

1 dx = x

ln c

1c1 1

if = 1, if = 1.
1 1 .

If < 1, then the principal value integral exists and is equal to principal value integral does not exist.

If 1, the limit diverges, so the

Exercise 5.4.11 Decide whether the following principal value integrals exist, and if so, evaluate them. i. ii. iii.
0 2
2

tan x dx
1 sin2 x

0 1 0

dx

ln x dx 170

Exercise 5.4.12 Dene the principal value integral and integrable on [a, c] for any c (a, b).

b a

f for a function f : [a, b) R, where f is bounded

Exercise 5.4.13 Dene the principal value integral a f for a function f : [a, c) (c, b] R, where f is bounded and integrable on [a, c ] and on [c + , b] for any , > 0. Exercise 5.4.14 Criticize the following computation:
1 1

x1 1 dx = 2 x 1

1 1

= 1 1 = 2.

1 1 lim0 1 x dx + x dx , then the value of this improper integral is 0. This idea of cancelation through symmetry plays a major role in some areas of advanced analysis. Finally, we note that the ideas presented above can be generalized to treat integrals of unbounded functions on unbounded intervals. This idea will be used in Project 5.11.1 on the Gamma function. We will see in the next section that there are necessary and sucient conditions on a bounded function f for it to be integrable. However, there is little point to proving it in one variable when we will need to prove it in several variables.

The class of principal value integrals we have discussed above are sometimes called improper integrals. As we have seen, there are many variations on establishing a value for an improper integral. If the procedure has a certain element of symmetry in it, then cancelation may play a role. For example, 1 1 1 1 dx has no value according to any of our previous denitions. However, if we assert that 1 x dx = 1 x
1

5.5

The Riemann Integral in Several Variables: Denitions

In developing a theory of integration, we build o of the fundamental notion of the volume of a generalized rectangle. When we did this for a function of a single variable, we considered rectangles in R, namely, intervals, whose volumes were their lengths. Such an interval formed the base for a rectangle in R2 , whose height was determined by function values on this interval. The volume of this rectangle was base times height, and the integral of the function was approximated by volumes of such rectangles. We say this here merely to emphasize the distinction between volume in the domain R (in this case, length), and volume in R2 (in this case, area). We note further that the notion of volume for the base is a nonnegative quantity, while we are allowing the height to be a signed quantity, and hence the volume in R2 is also a signed quantity. In trying to generalize this to functions of several variables, we therefore need to consider the volumes of generalized rectangles in Rn , as well as the notion of a signed volume in Rn+1 , which will be given by the volume of the base times the signed height. We begin by dening the volume of generalized rectangles in Rn . Denition 5.5.1 Let R Rn be a generalized rectangle of the form R = I1 I2 In , where Ii is a bounded interval with endpoints ai and bi , with ai bi , for each i = 1, 2, . . . , n. Then the volume of R in Rn is
n

v (R) =
i=1

(bi ai ).

For convenience, we also dene v () = 0. Denition 5.5.2 A partition P of a closed generalized rectangle R Rn is a nite collection P1 , . . . , Pk of pairwise disjoint open generalized rectangles contained in R such that
k

R=
i=1

Pi .

171

The mesh of the partition P is |P | = max diam(Pi ).


1ik

Remark 5.5.3 This new denition supersedes, but is consistent with, Denition 5.1.1. Given a partition P = {a = a0 < a1 < < ak = b} of [a, b] in the sense of Denition 5.1.1, we obtain a partition in the new sense consisting of the open generalized rectangles (which in this case are open intervals) (a0 , a1 ), (a1 , a2 ), . . . , (ak1 , ak ). Exercise 5.5.4 Let R Rn be a generalized rectangle. Show that if v (R) = 0, then there do not exist partitions of R. Exercise 5.5.5 Let R Rn be a generalized rectangle with v (R) > 0. Let Pi and Pj be two subrectangles from a partition of R. Show that v (Pi Pj ) = 0. Exercise 5.5.6 Let R Rn be a generalized rectangle with v (R) > 0, and let R1 , R2 , . . . , RN be a nite collection of subrectangles of R, each with nonzero volume. Show that there exists a partition P of R that satises the following property. For every i and every subrectangle S in P , either S is a subrectangle of Ri or S Ri = . Let R be a generalized rectangle in Rn . Let f : R R be a bounded function. We dene the integral of f over the generalized rectangle R by the same method we used in the one-variable case. Denition 5.5.7 Let R Rn be a generalized rectangle, and let f : R R be a bounded function. If P is a partition of R, we set mi = inf xPi f (x), and Mi = supxPi f (x). We dene the upper sum of f on R relative to the partition P to be
k

U (f, P ) =
i=1

Mi v (Pi ),

and the lower sum of f on R relative to P to be


k

L(f, P ) =
i=1

mi v (Pi ).

Exercise 5.5.8 Let f be a bounded function on R with m f (x) M for all x R. Given a partition P of R, show that m v (R) L(f, P ) U (f, P ) M v (R). Exercise 5.5.9 Let R R2 be the generalized rectangle R = [0, 1] [0, 2]. Let P = {P1 = (0, 1 2) 2 1 2 4 1 4 1 ) (1, 2), P3 = ( 1 , 1) (0 , ) , P = ( , 1) ( , ) , P = ( , 1) ( , 2) } , and let Q = { Q (0, 1), P2 = (0, 2 4 5 1 = 2 3 2 3 3 2 3 2 1 4 1 4 1 2 1 2 4 1 ) (0, 3 ), Q2 = (0, 2 ) (2 , ) , Q = (0 , ) ( , 2) , Q = ( , 1) (0 , ) , Q = ( , 1) ( , ) , Q = (0, 2 3 4 5 6 3 3 2 3 2 3 2 3 3 4 2 2 (1 , 1) ( , 2) } . Find the mesh of these partitions. Let f : R R , f ( x, y ) = x y . Compute the upper 2 3 and lower sums for f relative to the partitions P and Q. Denition 5.5.10 Let R Rn be a closed generalized rectangle, and let P = {Pj | j = 1, 2, . . . , k } be a partition of R. A partition P = {Pj | j = 1, 2, . . . , k } of R, where k k , is called a renement of P if every rectangle in P is contained in a rectangle in P . Exercise 5.5.11 Let f : R R be a bounded function. Let P be a partition of R, and let P a renement of P . Show that L(f, P ) L(f, P ) U (f, P ) U (f, P ). Exercise 5.5.12 Find a common renement of the partitions P and Q given in Exercise 5.5.9 and verify the inequalities in Exercise 5.5.11. Exercise 5.5.13 Let f : R R be a bounded function, and let P and Q be any two partitions of R. Show that L(f, P ) U (f, Q). 172

We are now ready to dene the lower and upper integrals of f on R. Denition 5.5.14 Let R Rn be closed generalized rectangle, and let f : R R be a bounded function. 1. We dene the lower integral of f on R to be f = lub{L(f, P ) | P is a partition of R}.

2. We dene the upper integral of f on R to be f = glb{U (f, P ) | P is a partition of R}.

Note that the set of lower sums is bounded above by M v (R), and the set of upper sums is bounded below by m v (R), and therefore the lower and upper integrals exist for any such bounded function. Exercise 5.5.15 For any bounded function f : R R, we have
R

f.

Exercise 5.5.16 Let R Rn be a closed generalized rectangle, and dene f : R R by f (x) = Find f and f. 0 1 if x = (x1 , x2 , , xn ), where xi Q for all i, otherwise.

In studying the Riemann integral in several variables, it is often useful to work with partitions that have a regularity condition. Denition 5.5.17 Let R = [a1 , b1 ] [a2 , b2 ] [an , bn ] be a closed generalized rectangle in Rn . For each i, 1 i n, we take a partition Pi = {ai = ai,0 , ai,1 , . . . , ai,ki = bi }, where ai,0 < ai,1 < < ai,ki . Note that Pi is a partition of the closed interval [a, b] in the sense of Denition 5.1.1. With this, we dene a regular partition P of R by P = P1 P2 Pn , consisting of the k1 k2 kn subrectangles of the form (a1,j1 , a1,j1 +1 ) (a2,j2 , a2,j2 +1 ) (an,jn , an,jn +1 ) with 0 ji < ki for 1 i n. Exercise 5.5.18 Show that
R

f = lub{L(f, P ) | P is a regular partition of R},

and
R

f = glb{U (f, P ) | P is a regular partition of R}.

Denition 5.5.19 Let f be a bounded function on a closed generalized rectangle R Rn . We say that f is integrable on R if R f = R f , and we denote this common value by R f , which we call the Riemann integral of f on R.

5.6

The Riemann Integral in Several Variables: Properties

The following 3 theorems are perfect analogues of the corresponding results in one variable. The careful reader should verify that the proofs carry over from before. Theorem 5.6.1 Let f be a bounded function on a closed generalized rectangle R Rn . Then f is integrable on R if and only if, given > 0, there exists a partition P of R such that U (f, P ) L(f, P ) < . 173

Theorem 5.6.2 Let R Rn be a closed generalized rectangle. Suppose f : R R is continuous. Then f is Riemann integrable on R. Theorem 5.6.3 Let (fk )kN be a sequence of Riemann integrable functions dened on a closed generalized rectangle R Rn . Suppose that (fk )kN converges uniformly to a function f on R. Then f is Riemann integrable on R, and R f = limk R fk . The results of many of the exercises from Section 5.2 carry over to several variables as well. We summarize these results in the statements below. Proposition 5.6.4 Let R Rn be a closed generalized rectangle, and let f : R R be a bounded function. 1. f is integrable on R if and only if, given > 0, there exists a > 0 such that if P is a partition of R with |P | < , then U (f, P ) L(f, P ) < . 2. Suppose f is integrable on R, and we have a sequence of partitions (Pk )kN of R such that limk |Pk | = 0. Then R f = limk L(f, Pk ) = limk U (f, Pk ). Proposition 5.6.5 Let R Rn be a closed generalized rectangle, let f, g : R R be integrable functions, and let R. Then 1. 2.
R (f R

+ g) =
R

f+

g;

f =

f;
R

3. if f (x) g (x) for all x R, then 4. |f | is integrable, and


R

g ; and

R |f |.

Exercise 5.6.6 Check that your one-variable proofs of the above statements from Exercises 5.2.2 and 5.2.4 carry over to several variables. Note that the extension of the result of Exercise 5.2.4.iii can be carried out in a number of ways. We leave these to the imagination of the reader.

5.7

The Computation of Integrals in Several Variables and Fubinis Theorem

In computing integrals in several variables, we begin by introducing a method that can use the fundamental theorem of calculus in an eective way. The basic idea is to iterate integrals one variable at a time and combine the results to obtain the integral in several variables. This is one place where our use of the traditional notation and its dummy variables will be particularly advantageous. We illustrate the basic principle with the following simple example. Example 5.7.1 Let R = [1, 3] [2, 4] R2 , and let f : R R be given by f (x, y ) = x2 y 3 . We proceed in the same manner that we did with partial derivatives. That is, we hold one variable constant, and integrate 3 4 with respect to the other variable. Symbolically, we write R f = 1 2 x2 y 3 dy dx, by which we mean
3 1 4 2

x2 y 3 dy dx. This indicates that we rst compute the integral with respect to y (the inner integral)

on the interval [2, 4] by holding x constant and applying the fundamental theorem. This inner integral will depend on x and should be thought of as a function of x, and we can compute the integral of this function (the outer integral) over the interval [1, 3] using the fundamental theorem a second time. 174

We now compute
3 1 2 4 3

x2 y 3 dy dx =
1

1 2 4 x y 4

y =4

dx =
y =2 1

60x2 dx = 520.

Note that, in computing the inner integral, if we consider g (y ) = f (x, y ) = x2 y 3 for xed x, then 2 4 G(y ) = 1 4 x y has derivative G (y ) = g (y ), which is what allows us to apply the fundamental theorem the way we did. Exercise 5.7.2 Show that
4 3 2 1

x2 y 3 dx dy = 520.
R

Exercise 5.7.3 Convince yourself that you do not want to show that the integral.

x2 y 3 = 520 using the denition of

Did we just get lucky? Or does the function f (x, y ) = x2 y 3 have properties that allow us to compute integrals by iterating one variable at a time? Observe that we knew that R x2 y 3 existed because f (x, y ) = x2 y 3 is continuous on R. Example 5.7.4 On R = [0, 1] [0, 1], we dene f (x, y ) = 1 2y if x is rational, if x is irrational.

If we attempt to iterate integrals as in the preceding example, we get


1 0 0 1 1

f (x, y ) dy dx =
0

1 dx = 1,
1 0

because of the following separate computations: if x is irrational,


1 0

2y dy = y 2

tional, 1 dy = = 1. On the other hand, if we attempt to iterate our integrals in the other order, 1 1 1 0 0 f (x, y ) dx dy , we discover that unless y = 2 , the function of one variable g (x) = 1 2y if x is rational, if x is irrational,

[y ]1 0

1 0

= 1, and if x is ra-

is not integrable on [0, 1], and thus the inner integral does not exist. Exercise 5.7.5 With R and f as in the above example, show that
R 3 f= 4 , and R 5 f= 4 .

Clearly, we need to put some restrictions on f before applying this iteration technique. It turns out that the integrability of f is sucient. Now we can state an actual theorem that says when it is possible to compute the integral of a function on a rectangle in Rn by iteration. We state the following with stronger hypotheses than necessary because the routine application of this theorem is to continuous functions, where these stronger hypotheses are met. Exercise 5.7.8 will discuss the weaker hypotheses for which the result also holds. It will be notationally convenient to reintroduce our dx notation for evaluating Riemann integrals in Rn . That is, if f is a Riemann integrable function on a closed generalized rectangle R Rn , we write R f = R f (x) dx. Theorem 5.7.6 (Fubinis Theorem) Let R1 Rn and R2 Rm be closed generalized rectangles, and let f : R1 R2 R be an integrable function such that, for any x R1 , the function gx : R2 R dened by gx (y ) = f (x, y ) is integrable. Then, the function G : R1 R dened by G(x) = R2 gx (y ) dy is integrable, and f=
R1 R2 R1

G(x) dx =
R1 R2

gx (y ) dy 175

dx =
R1 R2

f (x, y ) dy

dx.

Proof. We make use of the result of Exercise 5.5.18 on the suciency of regular partitions. Let P1 be a regular partition of R1 , and let P2 be a regular partition of R2 . We denote by P the corresponding regular partition of R1 R2 . Given a generalized rectangle S P , we can write S = S1 S2 , where S1 P1 , and S2 P2 . Then L(f, P ) =
S P

mS (f )v (S ) mS1 S2 (f )v (S2 ) v (S1 ).

=
S1 P1 S2 P2

Now, for any x S1 , we have mS1 S2 (f ) m{x}S2 (f ) = mS2 (gx ), and hence
S2 P2

mS1 S2 (f )v (S2 )
S2 P2

S2 P2

mS2 (gx )v (S2 )

gx (y ) dy = G(x).
R2

Since x was arbitrary,

mS1 S2 (f )v (S2 ) mS1 (G). This implies mS1 S2 (f )v (S2 ) v (S1 )

L(f, P ) =
S1 P1 S2 P2

mS1 (G)v (S1 )


S1 P1

= L(G, P1 ). A similar statement for upper sums gives us the inequality L(f, P ) L(G, P1 ) U (G, P1 ) U (f, P ). The assumption that f is integrable implies that the outer terms of this inequality can be made arbitrarily close by an appropriate choice of P , and hence G is integrable on R1 by Theorem 5.6.1, with R1 G =
R1 R2

f.

Exercise 5.7.7 Let R1 Rn and R2 Rm be closed generalized rectangles, and let f : R1 R2 R be an integrable function such that, for any y R2 , the function hy : R1 R dened by hy (x) = f (x, y ) is integrable. Show that the function H : R2 R dened by H (y ) = R1 hy (x) dx is integrable, and f=
R1 R2 R2

H (y ) dy =
R2 R1

hy (x) dx

dy =
R2 R1

f (x, y ) dx

dy.

Exercise 5.7.8 Show that we can weaken the hypotheses of Fubinis Theorem as follows and obtain an analogous result. Let R1 Rn and R2 Rm be closed generalized rectangles, and let f : R1 R2 R be an integrable function. For each x R1 , dene the function gx : R2 R by gx (y ) = f (x, y ). Then, the functions GL : R1 R dened by GL (x) = R2 gx (y ) dy and GR : R1 R dened by GR (x) = R2 gx (y ) dy are integrable, and GL (x) dx =
R1 R1 R2

f=
R1

GR (x) dx.

Exercise 5.7.9 Let R1 Rn and R2 Rm be closed generalized rectangles, and let R = R1 R2 Rn+m . Let g : R1 R and h : R2 R be integrable functions. Show that the function f : R R given by f (x, y ) = g (x)h(y ) is integrable on R, and f=
R R1

h .
R2

176

5.8

Sets of Measure Zero and the Riemann Integrability Criterion

We are led naturally to the following question. What are necessary and sucient conditions for a function f to be integrable? The following theorem, sometimes called the Riemann Integrability Criterion, gives a useful characterization of integrable functions in terms of their continuity. Theorem 5.8.1 Let R Rn be a closed generalized rectangle, and let f : R R be a bounded function. Then f is Riemann integrable on R if and only if f is continuous except on a set of measure zero. Obviously, we need to explain what it means for a set to have measure zero. Denition 5.8.2 Let A be a subset of Rn . We say that A has measure zero if for every > 0, there exists a countable collection {Rk }kN of open generalized rectangles in Rn such that A and
k=1 k=1

Rk ,

v (Rk ) < .

Exercise 5.8.3 i. Show that the empty set has measure zero. ii. Show that any countable set has measure zero. iii. Show that any subset of a set of measure zero has measure zero. iv. Show that a countable union of sets of measure zero has measure zero. Exercise 5.8.4 Show that the Cantor Set (see Exercise 1.6.37) has measure zero. Exercise 5.8.5 i. Show that a generalized rectangle R with v (R) = 0 has measure zero. ii. Show that a generalized rectangle R with v (R) > 0 does not have measure zero. Before returning to the proof of the Riemann Integrability Criterion, we need the idea of the oscillation of a function at a point, which is a quantitative measure of how badly discontinuous a function is at that point. Denition 5.8.6 Let R Rn be a closed generalized rectangle, and let f : R R be a bounded function. Fix x R. Given any > 0, dene M (f, x, ) = supyB (x)R {f (y )}, and m(f, x, ) = inf yB (x)R {f (y )}. Then the oscillation of f at x is dened to be o(f, x) = lim0 [M (f, x, ) m(f, x, )]. Exercise 5.8.7 Let R Rn be a closed generalized rectangle, and let f : R R be a bounded function. Show that f is continuous at a point x R if and only if o(f, x) = 0. Exercise 5.8.8 Compute the oscillations of the following functions at the point x = 0. i. f (x) = 0 if x < 0, 1 if x 0.

177

ii. f (x) =
1 sin x 0

if x = 0, if x = 0.

We are now ready to prove the Riemann Integrability Criterion. Proof. (of Theorem 5.8.1) Suppose that f is continuous except on a set B R of measure zero, and |f | M on R. Let > 0. We cover B with a countable collection {Ui }iN of open generalized rectangles in Rn such that i=1 v (Ui ) < . For each point x R \ B , we can nd an open generalized rectangle n Vx R containing x such that supyVx f (y ) inf yVx f (y ) < . The union of these two sets of open generalized rectangles covers the closed generalized rectangle R, which is compact. Hence, we can extract a nite subcovering that covers R. Because this is a cover by a nite collection of generalized rectangles, we can choose a partition P of R such that each generalized rectangle of P is a subset of the closure of one of these open generalized rectangles by Exercise 5.5.6. Then U (f, P ) L(f, P ) < 2M + v (R), where the rst term accounts for the subrectangles contained in the Ui s, and the second term accounts for those contained in the Vx s. Thus, by Theorem 5.6.1, f is integrable on R. Now suppose f is Riemann integrable on R. Again, we let B R be the set of discontinuities of f . We 1 . Since B is a countable union of sets of this note that B = k=1 B1/k , where B1/k = x R | o(f, x) k form, it suces to show that each of these sets has measure zero by 5.8.3. Fix k N, and choose > 0. By Theorem 5.6.1, we can nd a partition P of R such that U (f, P ) L(f, P ) < . Note that the interiors of the subrectangles Pi of P do not cover R, and may not cover B1/k . However, the boundaries of the Pi s form a set of measure zero, so we may cover these with a countable collection {Uj }j N of open generalized rectangles in Rn with j =1 v (Uj ) < . Let S be the subcollection of P consisting of those Pi s such that Pi B1/k = . The interiors of the elements of S along with the Ui s cover B1/k . The total volume of the generalized rectangles in the rst collection is v (Pi ) k (MPi (f ) mPi (f ))v (Pi ) (MPi (f ) mPi (f ))v (Pi )

Pi S

Pi S

Pi P

< k, and the total volume of the generalized rectangles in the second collection is, by denition, less than . So we have covered B1/k by a countable collection of generalized rectangles with total volume at most (k + 1), and thus B1/k has measure zero.

5.9

Integration over Bounded Regions in Rn

Let be a bounded region in Rn , and let f : R be a bounded function. We wish to dene f , but we will need certain restrictions, which are necessary in the Riemann theory of integration. The rst is a consideration of the boundary of . It is necessary that this boundary be nice, so that we can essentially disregard it. Denition 5.9.1 Let A Rn . We say that A has content zero if, given > 0, there is a nite collection k of open generalized rectangles R1 , R2 , . . . , Rk such that A k i=1 Ri , and i=1 v (Ri ) < . Exercise 5.9.2 1. If A Rn has content zero, then A has measure zero. 178

2. If A Rn has measure zero, and A is a compact set, then A has content zero. Thus, content zero and measure zero are equivalent for compact subsets of Rn , and in particular for the boundaries of bounded subsets of Rn . We assume that Rn is a bounded region, and the nice property referred to above is that the boundary of has content zero. By Exercise 3.3.33, we know that the boundary of is closed and hence compact, so by the exercise above, it is enough that the boundary have measure zero. We choose a closed generalized rectangle R that contains . The function f , which was originally dened as a function R, can be extended to a bounded function f : R R by setting f (x) = f (x) 0 if x , if x R \ .

Now observe that, according to Theorem 5.8.1, f is Riemann integrable on R if and only if f is continuous on R except on a set of measure zero. Where is f discontinuous? There are two potential types of points. On the one hand, if f is discontinuous at a point in , then f will be discontinuous there too. On the other hand, it is possible for f to be discontinuous on the boundary of , even if f is nicely behaved near the boundary. Thus, if the union of the boundary of and the set of discontinuities of f has measure zero, then f is integrable on R. For example, let = B1 (0) Rn be the open unit ball, and let f : R be the constant function equal to 1 on . Now let R = [1, 1] [1, 1] [1, 1]. Then, f dened as above is discontinuous on = S n1 = {x Rn | x = 1}.

Denition 5.9.3 Let Rn be a bounded region whose boundary has content zero, and let f : R be a bounded function that is continuous on except on a set of measure zero. Let R Rn be a closed generalized rectangle containing . Let f (x) = f (x) 0 if x , if x R \ .

Then we dene

f=

f.

Exercise 5.9.4 Show that the value of the integral is independent of the choice of the closed generalized rectangle R. Exercise 5.9.5 Let Rn be a bounded region whose boundary has content zero, and let f : R be a bounded function that is continuous on except on a set of measure zero. Let 1 , 2 , . . . , N be subsets N of such that = i=1 i , the sets i j have content zero if i = j , and i has content zero for each i = 1, 2, . . . , N . Show that f = N i=1 i f . Exercise 5.9.6 Let be the unit disk in R2 , and let R = [1, 1] [1, 1]. Let f : R be the constant function 1, and let f : R R be dened as above. For each n N, let Pn be the partition of R into n2 1 using these partitions. (Hint: Use congruent squares each of volume n 2 . Compute f from the denition Gausss theorem on the number of lattice points inside a circle of radius N .) Such a computation from rst principles is perhaps not the most productive approach to computing integrals in several variables. We reexamine this same example using the tool of Fubinis theorem. Example 5.9.7 Let be the closed unit disk in R2 , and let f : R the constant function 1. Again, it is convenient to choose R = [1, 1] [1, 1]. Since f is continuous on and has content zero, we know that f exists, and Fubinis theorem tells us that f=

f
R 1 1

=
1 1

f (x, y ) dx dy.

179

For each xed value of y , we compute the inner integral

function f considered as a function of the single variable x along a line segment parallel to the x-axis. For a xed value of y , we have 1 if 1 y 2 x 1 y 2 , f (x, y ) = 0 otherwise. 2 1y 1 In other words, 1 f (x, y ) dx = 2 1 dx = 2 1 y 2 . We can now use this formula to evaluate the outer integral, and get, by a standard trigonometric substitution,
1 1 1y

1 1

f (x, y ) dx. That is, we are integrating the

2 1 y 2 dy = .

The real point of this second example is that what we have eectively done is to parametrize the boundary of , and to determine the limits of integration in the inner integrals of a Fubini-type computation. In this particular example, we were fortunate in that the domain of integration was a convex set, so that we could parametrize the boundary using numerical limits of integration in one direction and simple functional expressions in the other. Exercise 5.9.8 Let be the triangle in R2 with vertices, (0, 0), (0, 1), and (1, 1), and let f : R be given by f (x, y ) = sin y 2 . Let R = [0, 1] [0, 1], and dene f : R R as above. i. Show that has measure zero. ii. Compute

f=

f using Fubinis Theorem.

iii. Observe that, though the integral exists and Fubinis theorem allows us to iterate the integrals in either order, the order of integration is nonetheless relevant to obtaining a successful computation. The ordinary treatment of multiple integration in advanced calculus books focuses principally on the determination of the boundaries of integration for various regions in the plane and in three-dimensional space. In the following set of exercises, this will be exemplied in computing multiple integrals in closed form. While this is an important aspect of studying multiple integrals, especially in computing volumes of regions in higher-dimensional space, the deeper aspects of analysis can be addressed much more eciently through the use of the Lebesgue integral. With this in mind, it is nonetheless important to recognize that in our study of vector calculus in the next chapter, we will rely almost totally on the Riemann integral. Denition 5.9.9 Let be a bounded region in Rn with having content zero. The n-volume of is vn () = 1. Exercise 5.9.10 Let and be bounded regions in Rn with and having content zero. Show that if , then vn () vn ( ). Exercise 5.9.11 Show that vn () = vn (T ()), where T is a translation on Rn . Exercise 5.9.12 Show that the n-volume of a generalized rectangle in Rn is the product of its side lengths. Exercise 5.9.13 Find the 2-volume of a right triangle whose legs are parallel to the coordinate axes. Exercise 5.9.14 Find the 2-volume of the parallelogram in R2 with vertices (0, 0), (a, b), (a + c, b + d), and (c, d). Exercise 5.9.15 In this exercise, we will show that the n-volume of the image of a generalized rectangle R Rn under an invertible linear transformation T : Rn Rn is equal to |det T | vn (R). 180

i. Show that it suces to prove this result when T is an elementary linear transformation. (See Exercise 2.3.12.) ii. Use the previous set of exercises, together with Fubinis theorem, to prove the result when T is an elementary linear transformation. Exercise 5.9.16 Let P be a plane in R3 parallel to the xy -plane. Let be a closed, bounded set in the xy -plane with 2-volume B . Pick a point Q in P and construct a pyramid by joining each point in to Q with a straight line segment. Find the 3-volume of this pyramid. Exercise 5.9.17 Find the volume of the generalized tetrahedron in Rn bounded by the coordinate hyperplanes and the hyperplane x1 + x2 + + xn = 1.

5.10

Change of Variables

We have illustrated the change of variables theorem in one variable in Corollary 5.3.4. The situation for several variables is considerably more complicated and involves not only the properties of dierentiation in Rn but also the change of volume under nonsingular linear transformations. In the one-variable case, the image of a closed interval under a monotonic C 1 function is very well understood and easily parametrized. In several variables, the notion of monotonic is meaningless, and the diculties connected to determining the volume of the image of a region (even a rectangle) under a C 1 map are already illustrated in Lemma 4.9.19. The main result of this section is the following. Theorem 5.10.1 (Change of Variables) Let V Rn be a bounded open set, and let U Rn be an open set such that V U . Let : U Rn be C 1 on U and one-to-one on V , with D(x) invertible on V . Suppose that V and (V ) have content zero. Then a bounded real-valued function f is Riemann integrable on (V ) if and only if f is Riemann integrable on V , and in this case, f (y ) dy =
(V ) V

(f )(x)|det D(x)| dx.

Our rst step is to prove the change of variables theorem when V is a rectangle. The rst step to doing that is to study the change of volume of a rectangle under a C 1 function. Lemma 5.10.2 Let R Rn be a generalized rectangle, let U Rn be an open set containing R, and let : U Rn be C 1 on U , and one-to-one on R, with D(x) invertible on R. Then vn ((R)) = R |det D(x)| dx. Proof. It is easy to see that the rectangle R can be partitioned into N congruent subrectangles R1 , R2 , . . . , RN that are similar to R, with centers y1 , y2 , . . . , yN , respectively. Recalling the notation of Lemma 4.9.19, we let 0 < < 1 and dene Ri = (1 )Ri , and Ri = (1 + )Ri , for 1 i N . By Lemma 4.9.19, if N is large enough, then
) ) (Ri ) T yi (Ri T yi (Ri

for all i, where T yi is dened as in the lemma. Hence


)). )) vn ((Ri )) vn (T yi (Ri vn (T yi (Ri

By Exercises 5.9.11 and 5.9.15, (1 )n J (yi )vn (Ri ) vn ((Ri )) (1 + )n J (yi )vn (Ri ), where J (x) = |det D(x)|. 181

Exercise 5.10.3 i. Show that (Ri ) has content 0. Hint: Use Lemma 4.9.19 and the above dissection technique. ii. Show that the boundary of a generalized rectangle is the union of nitely many closed generalized rectangles with volume zero. iii. Show that (Ri ) = (Ri ). Conclude that (Ri ) has content zero, and that vn ((R)). Hence, summing over i, we have
N N N i=1

vn ((Ri )) =

(1 )n

i=1

J (yi )vn (Ri ) vn ((R)) (1 + )n

J (yi )vn (Ri ).


i=1

Letting the mesh of our partition tend to zero, we have (1 )n J (x) dx vn ((R)) (1 + )n J (x) dx.
R

Letting 0 gives the desired result. We are now ready to prove the theorem in the case when the domain is a generalized rectangle in Rn . Theorem 5.10.4 Let R Rn be a generalized rectangle, let U Rn be an open set containing R, and let : U Rn be C 1 on U and one-to-one on R, with D(x) invertible on R. Then a bounded real-valued function f on (R) is Riemann integrable on (R) if and only if f is Riemann integrable on R, and in this case, f (y ) dy =
( R) R

(f )(x)|det D(x)| dx.

Proof. The equivalence of the two integrability conditions follows immediately from the fact that is invertible and that the Jacobian J (x) = |det D(x)| is everywhere nonzero on R. Let P = {R1 , R2 , . . . , RN } be a partition of R. Let mi = inf Ri (f ), and let Mi = supRi (f ). Then
( Ri )

mi

( Ri )

Mi .
( Ri )

By Lemma 5.10.2,
Ri

mi J (x) dx

( Ri )

Mi J (x) dx.
Ri

Let sP =

N i=1 Ri

mi J (x) dx, and let SP =

N i=1 Ri

Mi J (x) dx. Then by Exercise 5.9.5, f SP .

sP It is also clear that sP


R

( R)

(f )(x)J (x) dx SP .

So it suces to show that as |P | 0, SP sP 0. Let C be a constant bounding J (x) on R. Then


N

SP s P C

i=1

(Mi mi )v (Ri ).

182

Since f is integrable on R, the right-hand side goes to zero as |P | 0. Finally, we are ready to prove the full-scale change of variables theorem. Proof. (of Theorem 5.10.1) Let > 0. By Exercise 3.6.13, there exists an open set W such that V W and W is compact and contained in U . Let C be a bound on J (x) = |det D(x)| on (W ), and let M be a bound on |f | on this same set. Let R be a closed generalized rectangle containing V , and let P be a partition of R such that if S is a subrectangle of P , and S V = , then S W . This is possible by Exercise 3.6.12, since the distance from V to W must be positive, and so we need only choose rectangles that are suciently small. We write S1 = {S P | S V }, and S2 = {S P | S V = }. Since V has content 0, by Exercise 5.5.6, we may choose P such that S S2 v (S ) < . We may write V = A B , where A = S S1 S , and B = V \ A. Then by Exercise 5.9.5, f=
(V ) (A)

f+
(B )

f.

Since A is a nite union of generalized rectangles, (A) f = A (f )(x)J (x) dx, by Theorem 5.10.4. The second integral on the right can be bounded using Lemma 5.10.2 and our bounds on f and J as follows: f M 1
(B )

(B )

= M vn ((B )) M vn =M M (S )
S S2

vn ((S ))
S S2 S S2

C vn (S )

< M C. Similarly, (f )(x)J (x) dx = (f )(x)J (x) dx + (f )(x)J (x) dx.

The second integral on the right can be straightforwardly bounded by M C. So f (f )(x)J (x) dx = =
A

f+
(A) (B )

(V )

(f )(x)J (x) dx + f

(f )(x)J (x) dx (f )(x)J (x) dx

(f )(x)J (x) dx + f

(B )

(f )(x)J (x) dx +

=
(B )

(f )(x)J (x) dx

< 2M C. Since was arbitrary, the result follows.

183

Example 5.10.5 We compute the area of the circle C of radius a centered at the origin in R2 . (For comparison, see Example 5.9.7.) In the usual rectangular coordinates, this area is given by A = C 1 = 1 dy dx. The evaluation of the second iterated integral involves a change of variables in one variable, in particular, a trigonometric substitution. Instead, we compute this area using a change of variables known as polar coordinates. Let R be the rectangle (0, a) (0, 2 ) R2 . We dene a map : R2 R2 by (r, ) = (r cos , r sin ). Then cos sin D(r, ) = , r sin r cos so that J (r, ) = |det D(r, )| = |r|. Note that is C 1 on all of R2 , is one-to-one on R (though not on R), and J is nonzero on R (though not on R). Furthermore, R has content 0. The image of R under is C \ {(x, 0) | x 0}, which diers from C by a set of content zero, so that the integral of any integrable function on C is the same as that on (R). (See Exercise 5.9.5.) So by Theorem 5.10.1, A=
C a a2 x 2 a a2 x 2

1 1
( R)

= = =
0

J (r, ) dr d
R 2 0 a

r dr d

= a2 . Exercise 5.10.6 Compute the volume of the ball of radius a centered at the origin in R3 using the following change of variables, known as spherical coordinates. Let R = (0, a)(0, 2 )(0, ) R3 , and let : R3 R3 be dened by (r, 1 , 2 ) = (r cos 1 sin 2 , r sin 1 sin 2 , r cos 2 ). Exercise 5.10.7 1. Find the volume in R3 of the intersection of the two innite cylinders C1 = {(x, y, z ) R3 | y 2 + z 2 1} and C2 = {(x, y, z ) R3 | x2 + z 2 1}. 2. Find the volume in R3 of the intersection of the three innite cylinders C1 , C2 (as above), and C3 = {(x, y, z ) R3 | x2 + y 2 1}. Exercise 5.10.8 Find the volume in R4 of the intersection of the two innite hypercylinders H1 = {(x, y, z, w) R4 | x2 + y 2 1} and H2 = {(x, y, z, w) R4 | w2 + z 2 1}. 184

5.11

Projects

5.11.1 The Gamma Function The gamma function is in a class of functions occurring frequently in applications of mathematics. These are called Special Functions or sometimes Transcendental Functions. For a detailed and interesting exposition on this topic of Transcendental Functions, see Erd elyi et al., Higher Transcendental Functions, McGraw-Hill. If s is a real number, s > 0, we dene (s) =
0

et ts1 dt.

Of course, this is an improper integral for two reasons: 1. the domain of integration is unbounded; and 2. if 0 < s < 1, then the integrand is unbounded in a neighborhood of 0. Exercise 5.11.1 Show, by a simple use of the denition, that both parts of this improper integral converge if s > 0. Exercise 5.11.2 Show that (1) = 1. Exercise 5.11.3 Show that for any s > 0, s(s) = (s + 1). (Hint: Use and prove the validity of a generalization of integration by parts to improper integrals.) Exercise 5.11.4 Show that (n + 1) = n! for all nonnegative integers n. Remark 5.11.5 This exercise gives further justication for the surprising fact that 0! = 1. Exercise 5.11.6 Generalize Leibnizs rule to prove that is innitely dierentiable with respect to s on the interval (0, ). Exercise 5.11.7 Find the point in (0, 2) at which assumes a minimum value, and determine this minimum value. We now wish to present an exercise in which we compute one special value of the gamma function which will contribute to our computation of the volume of the unit ball in Rn in the next project. Proposition 5.11.8 Show that

ex dx =

Proof. First note that


ex dx

ex dx

ey dy =

e (x

+y 2 )

dx dy.

If we change to polar coordinates, this last integral is


2 0 0

er r dr d = 2
0

er r dr =
0

eu du = .

185

1 Exercise 5.11.9 Show that ( 2 )=

Exercise 5.11.10 Show that lims0+ (s) = +. Next, using the functional equation s(s) = (s +1), we want to extend the denition of to negative real ( 1 ) 1 numbers. For values of s with 1 < s < 0, we dene (s) = (s +1)/s. For example, ( 2 = 2 . ) = 2 1 2 Arguing inductively, we may continue this process to dene the gamma function for all negative real values of s, not including the negative integers. Exercise 5.11.11 Show that lims0 (s) = . Exercise 5.11.12 Show by induction that if n N, then we have the following limits. If n is odd, the limit as s approaches n from below of (s) is +, and the limit as s approaches n from above is . If n is even, the limit as s approaches n from below of (s) is , and the limit as s approaches n from above is +. Exercise 5.11.13 Construct a graph of the gamma function. 5.11.2 Volume of the Unit Ball Here is a lesson in developing spherical coordinates in Rn and the associated volume element in these coordinates. This is, of course, a generalization of the formulas for polar coordinates and the volume element for R2 . That is, if (x, y ) is a point in R2 \ {(0, 0)}, we can write x = r cos , where r = element y = r sin , x2 + y 2 , and is the unique solution in [0, 2 ) of the above pair of equations. The volume dx dy = r dr d can be derived from the change of variables formula or simply a geometric argument about sectors of circles in the plane. Both results that we derive are obtained by an iteration of polar coordinates. Take a point (x1 , x2 , . . . , xn ) Rn , and assume that (x1 , x2 ) = (0, 0). First, we write (x1 , x2 ) in polar coordinates. That is, x1 = r1 cos 1 and x2 = r1 sin 1 , where 0 < r1 < and 0 1 < 2 . The related volume element as stated above is dx1 dx2 = r1 dr1 d1 . For the next step, take the pair (x3 , r1 ) and treat it as a pair of Euclidean coordinates. Thus, we have x3 = r2 cos 2 , r1 = r2 sin 2 , where r2 =
2 x2 3 + r1 = 2 2 x2 3 + x2 + x1 .

Here, we have 0 < r2 < and 0 < 2 < . Two things happen here. First of all, (r2 , 1 , 2 ) are the familiar 2 2 2 spherical coordinates in R3 . That is, 2 = r2 = x2 3 + x2 + x1 , x1 = r2 sin 2 cos 1 = sin cos , and x2 = r2 sin 2 sin 1 = sin sin . Second, the volume element for (x3 , r1 ) can be written as dx3 dr1 = r2 dr2 d2 . By employing Fubinis Theorem, we can write dx3 dx2 dx1 = dx3 (r1 dr1 d1 ) = r1 r2 dr2 d2 d1 . By combining r1 and r2 , we reach the expression
2 dx1 dx2 dx3 = r2 sin 2 dr2 d2 d1 ,

and in the familiar spherical coordinates, this is dx1 dx2 dx3 = 2 sin d d d, where 0 < < , 0 < < , and 0 < 2 . 186

Exercise 5.11.14 Show that for four dimensions, we have the following change of coordinates: x1 x2 x3 x4
2 2 2 where r3 = x2 4 + x3 + x2 + x1 .

= r3 sin 3 sin 2 cos 1 = r3 sin 3 sin 2 sin 1 = r3 sin 3 cos 2 = r3 cos 3 ,

Exercise 5.11.15 Show that the four-dimensional volume element in spherical coordinates is 3 sin2 1 sin 2 d d1 d2 d3 , where the range is 0 < < , 0 1 < 2 , and 0 < 2 , 3 < . Exercise 5.11.16 Compute the surface area of the unit sphere S 3 in R4 . Exercise 5.11.17 Generalize the process developed above to prove the following formulas for spherical coordinates in Rn : x1 = sin n1 sin n2 sin 2 cos 1

x2 = sin n1 sin n2 sin 2 sin 1 x3 = sin n1 sin n2 sin 3 cos 2 . . . xn1 = sin n1 cos n2 xn = cos n1 , where 0 < < , 0 1 < 2 , and 0 < j < for 2 j n 1. We now have a formula for spherical coordinates in n dimensions.

Exercise 5.11.18 Show that the associated volume element in n dimensions is dx1 dx2 dxn = n1 sinn2 1 sinn3 2 sin n1 d d1 d2 dn1 . At this point, it is possible to compute integrals for functions f : Rn R written in spherical coordinates as f (, 1 , 2 , . . . , n1 ). In particular, the (n 1)-dimensional volume element d = sinn2 1 sinn3 2 sin n1 d1 d2 dn1 allows us to compute integrals over the hypersurface S n1 . (Note that the n-dimensional volume element can then be written as n1 d d.) So if we have f : S n1 R, we can compute such an integral as follows:
2 0

f ( ) d =
S n 1 0

f (1 , . . . , n1 ) sinn2 1 sinn3 2 sin n1 d1 d2 dn1 .

Exercise 5.11.19 Compute this integral when f is the characteristic function of S n1 . Exercise 5.11.20 Compute the volume of the unit ball in Rn by computing
1 0 S n 1

n1 d d.

To complete this project, we compute the volume of the n-ball through a process called the use of auxiliary functions in analysis. 187

Exercise 5.11.21 Take the function f : Rn R given by f (x) = e|x| . Show that e|x| dx = n/2
Rn
2

by integrating in Euclidean coordinates. Exercise 5.11.22 By integrating in spherical coordinates, show that e|x| dx =
Rn 0
2

S n 1

e n1 d d =

1 2

n d ( ). 2 S n 1

Conclude that the surface area of S n1 is

n/2 1 n . 2 2 n/2

Exercise 5.11.23 Finally, show that vol(Bn ) = Now that we have the volume of Bn , we would here is a table of these values for small n. n 1 2 3 4 . . .
n n . 2 ( 2 )

like to make some interesting observations. For instance, vol(Bn ) 2 2 4 3 2 2 . . .

Exercise 5.11.24 Continue this table and determine the value of n for which vol(Bn ) stops increasing and starts decreasing. Exercise 5.11.25 Show that after reaching its maximum, vol(Bn ) is a monotonic decreasing sequence with the property lim vol(Bn ) = 0.
n

If you want to have a nightmare, try to picture the unit ball inside a hypercube of side length 2, with the ball tangent to the hypercube at the center of each face. The volume of this hypercube is 2n , which goes to as n , while the volume of the n-ball, computed as above, goes to 0 as n .
n Exercise 5.11.26 i. Fix r > 0, and let Br (0) be the ball of radius r centered at the origin in Rn . Show n that limn vol(Br (0)) = 0. n ii. Fix r > 0, and let Cr (0) be the hypercube with side length 2r centered at the origin in Rn . Compute n limn vol(Cr (0)) for various values of r.

We have computed the volume of the unit ball in Rn in the 2 metric. One might ask what happens to the volume of the unit ball if we take the p metric for p = 2. Exercise 5.11.27 Let 1 p . i. Compute the volume in Rn of the unit ball in the p metric. ii. Determine whether this volume goes to 0 as n . For further reading, consult Lecture III in Lectures on the Geometry of Numbers by Carl Ludwig Siegel. 188

Chapter 6

Vector Calculus and the Theorems of Green, Gauss, and Stokes


The goal of this chapter is to prove the integral theorems of Green, Gauss, and Stokes. These theorems are of great use in applications of mathematics to physics and many other elds. We begin by giving an exposition that relies on restrictions of the domain of application and the use of symbols whose denition is tailored to these restrictive conditions. In fact, our presentation is sucient for most of the applications; however, the correct setting for these theorems lies in the calculus of dierential forms, which will constitute the remainder of this chapter. Our approach will allow us to breathe life into the theory through the constant use of examples. We begin with a careful treatment of curves in Rn and the study of the arc length of curves. After that, we state in order the theorems of Green, Gauss, and Stokes, and give references to what might be called the classical proofs. These proofs can be found, for example, in advanced calculus books by W. Kaplan or G. Folland.

6.1

Curves in Rn

Denition 6.1.1 Let [a, b] be a closed bounded interval in R. A smooth parametrized curve in Rn is a C 1 map : [a, b] Rn with the property that (t) = 0 for t [a, b]. Example 6.1.2 The function : [0, 2 ] R2 given by () = (cos , sin ) is a parametrization of the unit circle in R2 . Denition 6.1.3 Let [a, b] be a closed bounded interval in R. A piecewise smooth parametrized curve in Rn is a continuous map : [a, b] Rn that is continuously dierentiable except at a nite set of points and has the property that (t) = 0 wherever exists, and is injective in a neighborhood of every point where does not exist. Example 6.1.4 Suppose we wish to nd a parametrization of the square in R2 with vertices A = (0, 0), B = (1, 0), C = (1, 1), and D = (0, 1). Because of the corners of the square, we will not be able to nd such a smooth parametrized curve, but are able to give a piecewise smooth parametrized curve as follows. (t, 0) if 0 t < 1, (1, t 1) if 1 t < 2, : [0, 4] R2 , (t) = (3 t, 1) if 2 t < 3, and (0, 4 t) if 3 t 4. 189

Denition 6.1.5 A piecewise smooth parametrized curve : [a, b] Rn is closed if (a) = (b).

Denition 6.1.6 A piecewise smooth parametrized curve is simple if (s) = (t) for s = t, with the obvious exception that we allow (a) = (b) if is closed. Note that the examples above are both simple and closed. Exercise 6.1.7 i. Give an example of a smooth parametrized curve that is simple but not closed. ii. Give an example of a smooth parametrized curve that is closed but not simple. iii. Give an example of a smooth parametrized curve that is neither closed nor simple. Example 6.1.2 above shows that we can parametrize the unit circle with the function : [0, 2 ] R2 given by () = (cos , sin ), which traces out the circle at constant speed in a counterclockwise direction starting at the point (1, 0). All of these qualications should tell us that there are many ways to parametrize the unit circle. For example, the function : [0, ] R2 given by () = (cos 2, sin 2) traces out the same circle in the same direction from the same starting point but at double the speed of . Exercise 6.1.8 R2 . For each of the following sets of conditions, give a parametrization of the unit circle in

i. Counterclockwise, constant speed, starting at (0, 1). ii. Clockwise, constant speed, starting at (1, 0). iii. Counterclockwise, non-constant speed, starting at (1, 0). As functions, these parametrized curves are obviously distinct, but what they have in common is that they all trace out the same set in R2 . We formalize this idea in the following denition. Denition 6.1.9 Given a piecewise smooth parametrized curve : [a, b] Rn , the path C of is the image of in Rn , that is, C is the path from (a) to (b) dened by the (piecewise) smooth parametrization . Exercise 6.1.10 Consider the spiral : [0, 8 ] R3 , () = (cos , sin , ). Find a second parametrization of C . Find a third. We have seen in the case of the unit circle that a parametrized curve has the distinguishing characteristics of direction, speed, and in the case of a closed curve, the starting point. (Note that the path of a non-closed curve can only be parametrized starting from one end or the other, while the path of a closed curve can be parametrized starting from any point.) As we will see, it is often the case in vector calculus that important calculations are independent of the speed of a parametrization, but do depend on the direction. Denition 6.1.11 Two non-closed piecewise smooth parametrized curves : [a, b] Rn and : [c, d] Rn are equivalent if there exists a continuous, piecewise C 1 , strictly monotonic function : [a, b] [c, d] such that (a) = c and (b) = d and ( )(t) = (t) for all t [a, b]. Exercise 6.1.12 Show that if and are equivalent, then C = C . Exercise 6.1.13 Show that the property of being simple is preserved under equivalence. Denition 6.1.14 Let : [a, b] Rn be a piecewise smooth parametrized curve. The opposite parametrization of is the piecewise smooth parametrized curve : [a, b] Rn given by (t) = (a + b t) Exercise 6.1.15 Let : [a, b] Rn be a non-closed piecewise smooth parametrized curve. Show that and are not equivalent. 190

The next step in our analysis of curves is to nd a way to compute the length of a curve. This, in turn, will lead us to a means to dene the integral of a function on a curve. Denition 6.1.16 Let : [a, b] Rn be a (piecewise) smooth parametrized curve. Partition [a, b] into a nite number of points a = t0 < t1 < < tm = b, where any points at which D is discontinuous must be included. A polygonal approximation to is the polygon whose vertices are successively (t0 ), (t1 ), . . . , (tm ). Exercise 6.1.17 If : [a, b] Rn is a (piecewise) smooth parametrized curve, show that the length of the m polygonal approximation dened by the partition a = t0 < t1 < < tm = b is i=1 (ti ) (ti1 ) . We can rewrite this polygonal approximation as follows. If we write (t) = (1 (t), 2 (t), . . . , n (t)), then (ti ) (ti1 )
2

n j =1 (j (ti )

j (ti1 ))2 , and this is equal to


(j )

the mean-value theorem, there exists ti (ti1 , ti ) such that the length of the polygonal approximation to as
m i=1 n n

n 2 j (ti )j (ti1 ) j =1 (ti ti1 ) (ti ti1 ) dj (j ) j (ti )j (ti1 ) = dt (ti ). Thus we can (ti ti1 )

. By

write

(ti ti1 )

j =1

dj (j ) (t ) dt i

j If we set F (t) = (t) = j =1 dt , then the above expression is bounded by L(F, P ) and U (F, P ), where P is the partition P = (t0 , t1 , . . . , tm ). Since F is continuous (except at a nite number of points), it is integrable, and thus the lengths of the polygonal approximations converge to the value of the integral of F.

d 2

Denition 6.1.18 The length of the (piecewise) smooth parametrized curve is () = Exercise 6.1.19

b a

(t) dt.

i. Show that if and are equivalent, then () = ( ). Thus, we may speak of the length of a smooth curve. ii. Show that if and are equivalent, and f : U R is a continuous function on some open set U Rn containing C = C , then
b d

f ((t)) D(t) dt =
a c

f ( (t)) D (t) dt.

Example 6.1.20 Consider the spiral from Exercise 6.1.10. We compute () = ( sin , cos , 1), and hence
8

() =
0 8

( sin )2 + (cos )2 + 12 d 2 d

0 = 8 2

Exercise 6.1.21 Show using your parametrization from Exercise 6.1.10 that the length of the spiral is 8 2 . Exercise 6.1.22 Compute the lengths of the following curves. 191

i. : [0, 1] R3 given by (t) = (at, bt, ct), where a, b, and c are real constants, not all of which are zero. ii. : [0, 1] R2 given by (t) = (t, t3 ). iii. : [0, 1] R2 given by (t) = (t3 , t3 ). Why is your result unsurprising? Denition 6.1.23 Let : [a, b] Rn be a (piecewise) smooth parametrized curve. We say that is an arc-length parametrization if (t) = 1 wherever exists. Exercise 6.1.24 Let : [a, b] Rn be an arc-length parametrization. i. Show that () = b a. ii. Show that if [c, d] [a, b], and [c,d] : [c, d] Rn is the restriction of to [c, d], then ( ) = d c. Exercise 6.1.25 Let : [a, b] Rn be a (piecewise) smooth parametrized curve such that, for every closed subinterval [c, d] [a, b], the restriction [c,d] : [c, d] Rn of to [c, d] satises ([c,d] ) = d c. Show that is an arc-length parametrization. Exercise 6.1.26 Show that every (piecewise) smooth curve admits a unique arc-length parametrization. Exercise 6.1.27 i. Find the arc-length parametrization for the spiral from Exercise 6.1.10. ii. Show that the parametrization of the square in Example 6.1.4 is an arc-length parametrization. iii. Attempt to nd the arc-length parametrization of the curve : [0, 1] R2 given by (t) = (t, t2 ) in order to get a sense of how intractable these types of problems can be more generally.

6.2

Line Integrals in Rn and Dierential 1-Forms

Line integrals have important applications in physics and elsewhere. We dene them here and allow them to lead us to a discussion of dierential 1-forms. With these ideas in place, we will be able to approach the rst of the classical theorems of vector calculus, namely Greens theorem. In the last section (Exercise 6.1.19), we saw how to integrate a real-valued function over a parametrized curve. In this section, we will dene the line integral as the integral of a vector-valued function over a curve. In particular, if : [a, b] Rn is our parametrized curve, and if U Rn is an open set containing the path C , we will integrate continuous functions of the form F : U Rn . Such functions are often referred to as vector elds on U because they associate to each point in U a vector in Rn . Denition 6.2.1 Let : [a, b] Rn be a piecewise smooth parametrized curve, let U Rn be an open set containing C , and let F : U Rn be a continuous function which is written in terms of its coordinate functions as F = (F1 , . . . , Fn ). The line integral of F on is denoted by F and is dened to be
b

F =
a

(F )(t) (t) dt =

b a

[F1 ((t)) 1 (t) + + Fn ((t))n (t)] dt.

Remark 6.2.2 There are two important things to note here. First, our function F took values in Rn in part so that the dot product in the denition makes sense. Second, because we have taken the dot product before integrating, we are in fact integrating a single-variable function (F ) : [a, b] R, which is something we know how to do from elementary calculus. Exercise 6.2.3 If : [a, b] Rn and : [c, d] Rn are equivalent piecewise smooth parametrized curves, then F = F . 192

Denition 6.2.4 Let 1 : [a1 , b1 ] Rn and 2 : [a2 , b2 ] Rn be piecewise smooth parametrized curves such that 2 (a2 ) = 1 (b1 ). The concatenation of 1 and 2 is the piecewise smooth parametrized curve : [a1 , b1 + b2 a2 ] Rn given by (t) = 1 (t) if a1 t b1 , 2 (t b1 + a2 ) if b1 < t b1 + b2 a2 .

Example 6.2.5 Note that in Example 6.1.4, the curve is already written as the concatenation of four other curves. Exercise 6.2.6 Let 1 : [a1 , b1 ] Rn and 2 : [a2 , b2 ] Rn be piecewise smooth parametrized curves such that 2 (a2 ) = 1 (b1 ), and let be their concatenation. Show that F = 1 F + 2 F . Exercise 6.2.7 Let : [a, b] Rn be a piecewise smooth parametrized curve. Recall that the backwards parametrization of is the piecewise smooth parametrized curve : [a, b] Rn given by ()(t) = (a + b t). Show that F = F . Remark 6.2.8 The preceding exercise justies the convention from one-variable integration that if f is a b integrable on [a, b], then b f = a f . In fact, this will have important implications in the next section when we begin our discussion of the orientation of intervals and regions. Example 6.2.9 Let : [0, ] R2 be given by () = (cos , sin ). Observe that the image of this parametrized curve is the top half of the unit circle, parametrized counterclockwise. Then () = ( sin , cos ). We compute the line integrals for four dierent functions on . 1. Let F : R2 R2 be given by F (x, y ) = (1, 0). Then

F =
0

(F1 (cos , sin )( sin ) + F2 (cos , sin )(cos )) d sin d

=
0

= 2.

= cos cos 0

2. Let G : R2 R2 be given by G(x, y ) = (0, 1). Then G=


0

(G1 (cos , sin )( sin ) + G2 (cos , sin )(cos )) d cos d

=
0

= sin sin 0 = 0. 3. Let H : R2 R2 be given by H (x, y ) = (x, y ). Then

H=
0

(H1 (cos , sin )( sin ) + H2 (cos , sin )(cos )) d ( cos sin + sin cos ) d

=
0

=
0

0 d

= 0. 193

4. Let I : R2 R2 be given by I (x, y ) = (y, x). Then

I=
0

(I1 (cos , sin )( sin ) + I2 (cos , sin )(cos )) d (sin2 + cos2 ) d 1 d

=
0

=
0

= . Exercise 6.2.10 Compute the line integrals of the four functions in the example above for the following parametrized curves. i. : [1, 1] R2 given by (t) = (t, 0). ii. : [0, ] R2 given by () = (cos , sin ). iii. : [0, 2] R2 given by (t) = (t, t) iv. : [0, 2] R2 given by (t) = (t, t2 ). Exercise 6.2.11 Find a piecewise smooth parametrization of the boundary of the triangle with vertices (0, 0), (1, 0), and (1, 1) in that order. i. Let F : R2 R2 be given by F (x, y ) = (y 2 , x). Find

F.

ii. Let G : R2 R2 be given by G(x, y ) = (y 2 , 2xy ). Find

G.

The value of a line integral F obviously depends on both the curve and the function F . We explore how these two components interact. At the innitesimal level, we are taking the dot product of the vector that is the value of the function F with the vector that is the derivative of at each point. The value of this dot product, then, will tell us the extent to which these two vectors are aligned. Vectors in the same direction (that is, those that form an angle of less than 2 ) will make a positive contribution towards the value of the integral, while those in opposite direction (that is, those that form an angle of more than 2 ) will make a negative contribution. Perpendicular vectors make a contribution of zero. The integral represents the total cumulative eect of these contributions. We now consider the dierent behaviors of the four line integrals in Example 6.2.9. Recall that in all four, the curve was the upper half of the unit circle, traversed in a counterclockwise direction. What aspects of this curve are revealed by each of these four line integrals? In part (a), the function F is a constant function whose vector value is (1, 0). If we interpret this as measuring the change of in the positive x direction, the line integral tells us that the accumulated change is 2, which matches our intuition about the diameter of the unit circle. In part (b), the function G is again a constant function whose vector value is (0, 1). If we interpret this as measuring the change of in the positive y direction, the line integral tells us that the accumulated change is 0. We note that this does say that the curve has not moved at all in the y direction, but rather than from start to nish, the net change is zero. In part (c), the function H is not constant. At each point in R2 , the vector value of H is pointing radially outward from the origin. The line integral is therefore a measure of whether the curve is moving towards or away from the origin. Since our curve is always a constant distance from the origin, there is no radial change towards or away from the origin, and our dot product is identically zero. This behavior should be distinguished from part (b), where our answer was zero only by cancellation. In part (d), the function I is again not constant. At each point in R2 , the function I is producing a vector (y, x) that is orthogonal to the input (x, y ). In this sense, I is measuring change not toward or away from 194

the origin, but rather angular measure along circles of xed radii, and in particular in a counterclockwise direction. Since our curve is a half-circle parametrized in a counterclockwise direction, the line integral measures the accumulated angular change, which is, unsurprisingly, . In these examples we have measured our curve by taking the dot product of its derivative vector with a vector-valued function at each point. We can free ourselves from the implicit bonds of coordinate geometry by recognizing these dot products as evaluations of linear maps on these tangent vectors. This naturally leads us to the following denition. Denition 6.2.12 Let Rn . A dierential 1-form on is a map : L(Rn , R). The linear map in L(Rn , R) associated with the point x is denoted x . Remark 6.2.13 Thus, if is a dierential 1-form, it takes elements of and returns linear maps, while x is a linear map which takes elements of Rn and returns elements of R. Remark 6.2.14 Because L(Rn , R) is a nite dimensional real vector space of dimension n, we can pick a basis to identify it with Rn . We then give Rn the usual metric. This allows us to dene what it means for a dierential 1-form : L(Rn , R) to be continuous, dierentiable, smooth, etc. This denition should not look entirely unfamiliar. In Chapter 4, we identied the derivative of a function f : U R as a function that assigned to each point x U the linear map Df (x) : Rn R. This linear map, applied to a vector v , represented the directional derivative of f in the direction v . Thus, we see that the derivatives of dierentiable real-valued functions on Rn form a ready class of dierential 1-forms. Denition 6.2.15 Let U Rn be an open set, and let f : U R be a dierentiable function. The dierential of f is the dierential 1-form df given by dfx = Df (x). Remark 6.2.16 Note that in some sense, the preceding denition is merely a change in notation and terminology from our work in Chapter 4. Denition 6.2.17 For each i = 1, . . . , n, we identify a special dierential 1-form by the name dxi , which takes a point x = (x1 , . . . , xn ) to the linear map i : Rn R which is the ith coordinate projection, that is, if v Rn , then i (v ) = i (v1 , . . . , vn ) = vi . In the cases that n = 2 or n = 3, it is traditional to write dx = dx1 , dy = dx2 , and dz = dx3 . Exercise 6.2.18 Let i : Rn R be the ith coordinate projection function. Show that dxi = di . Remark 6.2.19 Observe that dierential 1-forms on a set Rn can be added together and be multiplied not just by real scalars but by arbitrary real-valued functions on . That is, if 1 and 2 are dierential 1-forms on , then it is obvious what we mean by 1 + 2 . Similarly, if f : R is a function, it is clear what dierential 1-form we mean by f 1 . Exercise 6.2.20 Let Rn , and let be a dierential 1-form on . Show that there exist unique functions Fi : R for 1 i n such that = F1 dx1 + + Fn dxn . Show that is continuous if and only if the functions Fi are continuous for each i. Exercise 6.2.21 Let U Rn be an open set, and let f : U R be a dierentiable function. Show that df = or in other symbols, dfx = D1 f (x)dx1 + + Dn f (x)dxn . 195 f f dx1 + + dxn , x1 xn

The implication of Exercise 6.2.20 is that there is a natural correspondence between dierential 1-forms and vector-valued functions. We restate this more precisely as follows. If is a dierential 1-form on a set Rn , then there exists a unique function F : Rn such that for all x and v Rn , we have x (v ) = F (x) v. On the other hand, given such a function F , we can dene a dierential 1-form in this way. For example, if we consider the function I : R2 R2 given by I (x, y ) = (y, x) from Example 6.2.9, we can write down the corresponding dierential 1-form as = y dx + x dy . This correspondence allows us to rewrite the denition of line integrals in the language of dierential 1-forms. Denition 6.2.22 Let Rn , and let be a continuous dierential 1-form on . Let : [a, b] be a piecewise smooth parametrized curve. The integral of over is dened to be:
b

=
a

(t) ( (t)) dt.

Remark 6.2.23 1. If = F1 dx1 + + Fn dxn , then


b

=
a

[F1 ((t)) 1 (t) + + Fn ((t))n (t)] dt

F.

2. Because we can interpret the integral of a 1-form as a line integral of a function, Exercise 6.2.3 implies that = if is equivalent to . Normally at this point, we would pose several exercises in which you would compute integrals of dierential 1-forms. Please note, however, that the remark above shows that you have already done this back in Exercises 6.2.10 and 6.2.11. We also saw in the previous section how to nd the length of a curve using integrals. Can we interpret the length of a parametrized curve in terms of dierential 1-forms? Yes, as the following important example illustrates. Example 6.2.24 Let : [a, b] Rn be a simple smooth parametrized curve, and let = C . The length 1-form : L(Rn , R) is dened by the formula (t) (v ) = The integral of this 1-form is
b

1 [D(t)](v ). (t)

=
a

1 [D(t)]( (t)) dt = (t)

b a

1 (t)

(t)

dt =
a

(t) dt,

which is precisely the length of the curve as given in Denition 6.1.18. In single-variable calculus, the fundamental theorem (in the form of Corollary 5.3.2) related the integral of the derivative of a function to the values of that function. In essence, it expressed the total change in the value of a function over an interval as the accumulation of instantaneous changes. It is not unreasonable to expect a similar result to pertain to line integrals. That is, if the 1-form whose integral we are evaluating can be recognized as the dierential of a function, we should expect that integral to represent the accumulated instantaneous changes of that function along the curve, and to be equal to the dierence of the values of the function at the endpoints of the curve. 196

Theorem 6.2.25 (Fundamental Theorem for Line Integrals) Let U Rn be an open set, and let f : U R be a C 1 function. Let : [a, b] U be a piecewise smooth curve. Then

df = f ((b)) f ((a)).

Proof. Without loss of generality, assume that is a smooth curve. Applying the single-variable fundamental theorem of calculus gives
b

df =
a b

[D1 f ((t)) 1 (t) + + Dn f ((t))n (t)] dt

=
a

d (f ) (t) dt dt

= (f )(b) (f )(a).

Corollary 6.2.26 Let U Rn be an open set, let f : U R be a C 1 function, and let = df . If and are two smooth curves in U with the same beginning and end points, then =

Proof. Exercise. Note the signicance of the previous result. The conclusion would be true for any two equivalent parametrizations and . In this corollary, though, because of our strong condition on the dierential 1-form, and are not necessarily equivalent parametrizations of the same curve. They merely need to start and end at the same points. Exercise 6.2.27 Let U Rn be a connected open set, and let be a continuous dierential 1-form on U . Suppose that, = whenever and begin and end at the same points. Show that there exists a C 1 function f : U R such that = df . (Hint: Consider how the fundamental theorem of calculus allows us to construct the antiderivative of a continuous function of a single variable as an integral.) Exercise 6.2.28 Show that the condition that = whenever and begin and end at the same points is equivalent to the condition that = 0 whenever : [a, b] U parametrizes a closed curve. (Hint: Consider the trivial curve : [0, 0] U with (0) = (a) = (b).) Example 6.2.29 We consider the vector-valued functions from Exercise 6.2.11 written in the new language of dierential 1-forms. We wish to determine for each of these dierential 1-forms whether there exists a function f : R2 R such that = df . 1. Let = y 2 dx + x dy . You showed in Exercise 6.2.11 that curve, it follows that = df for any function f .

= 0. Since parametrizes a closed

2. Let = y 2 dx + 2xy dy . You showed in Exercise 6.2.11 that = 0. This is no guarantee that the integral around every closed curve is zero, but it at least admits the possibility that such an f exists. If such an f does exist, we should be able to reconstruct it as follows. We x a point, the origin, and determine the value of f at other points by integrating along straight-line curves from the origin. 197

Let (x, y ) R2 , and let : [0, 1] R2 be the straight-line curve from the origin to (x, y ) given by (t) = (tx, ty ). Then (t) = (x, y ). So
1

=
0 1

(2 (t)2 , 21 (t)2 (t)) (x, y ) dt (t2 y 2 , 2t2 xy ) (x, y ) dt 3t2 (xy 2 ) dt

=
0 1

=
0

= xy 2 . Thus, if a function f such that = df exists, then f (x, y ) = xy 2 must be such a function. Since we can easily compute df = 2xy dx + y 2 dy = , we discover that our f is indeed such a function. Exercise 6.2.30 We consider the vector-valued functions from Example 6.2.9 written in the new language of dierential 1-forms. For each, determine whether there exists a function f : R2 R such that = df , and if so, determine such a function. (Hint: Use the computations in Example 6.2.9 and Exercise 6.2.10.) i. = dx ii. = dy iii. = x dx + y dy iv. = y dx + x dy We have shown that we have a special type of dierential 1-form, whose integral is independent of the path. Denition 6.2.31 Let U Rn be a connected open set, and let be a continuous dierential 1-form on U . We say that is exact if there exists a C 1 function f : U R such that = df . Such a function f is called a primitive for . Exercise 6.2.32 Let be an exact dierential 1-form on a connected open set U Rn with primitive f . 1. Show that for any constant C R, the function f + C is also a primitive for . 2. Show that if g is any primitive of , then g = f + C for some constant C R. Exercise 6.2.33 Let U R2 be a connected open set, and let = P (x, y ) dx + Q(x, y ) dy be a C 1 Q P = as follows: dierential 1-form on U. Show that is exact if and only if y x i. Assuming is exact with primitive f , use results about second-order partial derivatives. ii. Assuming P Q = , construct a primitive f . (Hint: First, determine the collection of functions y x g f g : U R such that = P (x, y ). Then, show that there exists such a function f with = Q(x, y ).) x y

Exercise 6.2.34 Construct a primitive for the following exact dierential 1-forms. i. = x3 y dx +
2

1 4 4x

+ y 2 dy dx + 2xyexy 2y dy 198
2

ii. = y 2 exy + xex

We nish this section with an application of dierential 1-forms to ordinary dierential equations. dy Many rst-order ordinary dierential equations can be written in the form P (x, y ) + Q(x, y ) = 0. We dx can sometimes solve such an equation by considering the dierential 1-form = P (x, y ) dx + Q(x, y ) dy on R2 . Suppose that is exact, that is, there exists a function f such that df = . If y is considered as a function of x, we have d f f dy f (x, y ) = (x, y ) + (x, y ) dx x y dx dy = P (x, y ) + Q(x, y ) . dx d Thus, y is a solution to the dierential equation if and only if f (x, y ) = 0, that is, f (x, y ) = c for some dx constant c. In other words, y is a solution to the dierential equation if and only if the graph of y is a level curve of the function f . Because of the relationship between the dierential equation and the dierential 1-form , many textbooks present the original dierential equation in the form P (x, y ) dx + Q(x, y ) dy = 0. Exercise 6.2.35 Recall that a separable dierential equation is one that can be written in the form dy q (y ) = p(x). dx 1. Show that the dierential 1-form associated with this dierential equation is exact. 2. Use the above method to solve the separable dierential equation dy = 4x3 y dx with initial condition y = 2 when x = 0.

6.3

Greens Theorem in the Plane

One of the major and most impressive theorems in vector calculus is Greens theorem, which relates the integral of a dierential 1-form along a closed curve to the integral of a related function on the region bounded by the curve. It actually produces a rather astounding result. We state it rst for curves in R2 . Lemma 6.3.1 Let U R2 be an open set, and let U be a compact region such that there exists an interval [a, b] and piecewise smooth functions f1 , f2 : [a, b] R such that = {(x, y ) | x [a, b], f1 (x) y f2 (x)}. Let be a counterclockwise parametrization of , and let P : U R be a C 1 function. Then

P = y

P dx.

Proof. By Exercise 6.2.3, we may choose a preferred parametrization of . We construct as follows. 1. Let 1 : [a, b] R be given by 1 (t) = (t, f1 (t)). 2. Let 2 : [f1 (b), f2 (b)] R be given by 2 (t) = (b, t). 3. Let 3 : [a, b] R be given by 3 (t) = (t, f2 (t)). 4. Let 4 : [f1 (a), f2 (a)] R be given by 4 (t) = (a, t). 199

Let be the piecewise smooth curve obtained by concatenating 1 , 2 , 3 , and 4 , in that order. We rst work out the left-hand side. P = y
b a f2 (x) f1 (x) b

P dy dx y

=
b

(P (x, f2 (x)) P (x, f1 (x))) dx


b

=
a

P (x, f1 (x)) dx

P (x, f2 (x)) dx.


a

The right-hand side can be broken up into the sum of four integrals. P dx =
1

P dx +
2

P dx

P dx

P dx.
4

The second and fourth integrals are zero, since there is no change in x along a vertical line. The rst integral is
b

P dx =
1 a b

P (t, f1 (t)) dt,

and the third integral is P dx =


3 a b

P (t, f2 (t)) dt. P . y

So P dx =
a

P (t, f1 (t)) dt

P (t, f2 (t)) dt =
a

Lemma 6.3.2 Let U R2 be an open set, and let U be a compact region such that there exists an interval [c, d] and piecewise smooth functions g1 , g2 : [c, d] R such that = {(x, y ) | y [c, d], g1 (y ) x g2 (y )}. Let be a counterclockwise parametrization of , and let Q : U R be a C 1 function. Then

Q = x

Q dy.

Proof. Exercise. Theorem 6.3.3 (Special Case of Greens Theorem in the Plane) Let U R2 be an open set, and let U be a compact region such that there exist intervals [a, b] and [c, d], and piecewise smooth functions f1 , f2 : [a, b] R, g1 , g2 : [c, d] R, such that = {(x, y ) | x [a, b], f1 (x) y f2 (x)} = {(x, y ) | y [c, d], g1 (y ) x g2 (y )}. Let be a counterclockwise parametrization of , and let P : U R and Q : U R be C 1 functions. Then Q P = (P dx + Q dy ). x y 200

Proof. Add the results of the previous two lemmas.

Remark 6.3.4 1. Note that there is a slight asymmetry in the two lemmas. The occurrence of the minus sign is governed by the direction in which we move along the boundary. Had we moved clockwise, the signs would have been reversed. 2. The conditions that we imposed on may seem strong. However, the types of regions for which the result of Greens Theorem applies may be broadened to include any regions that may be decomposed into a nite union of regions of the type described in the theorem. Most regions that occur in applications fall into this large class. One of the most signicant applications of Greens Theorem is the following remarkable result. If is a bounded region in the plane whose boundary can be parametrized in the fashion above, then we can Q P = 1, then the compute the area as follows. If we can nd a pair of functions P and Q such that x y left-hand integral in the statement of Greens Theorem becomes

1, which is equal to the area of . There

are many such choices of P and Q, such as P (x, y ) = y , and Q(x, y ) = 0, or, for example, P (x, y ) = 0, and 1 1 Q(x, y ) = x. A particularly nice choice is the pair of functions P (x, y ) = y , and Q(x, y ) = x. In this 2 2 case, Greens Theorem gives us Area() =

1=

Q P x y

1 1 y dx + x dy . 2 2

This particular way of evaluating the area of results in the construction of an instrument called a planimeter, an instrument used to compute area whose working principle is based on this equation. If you wheel a planimeter around the boundary of a lake, the area of the lake just pops out. Exercise 6.3.5 Use the method described above to nd the area of the region bounded by the ellipse y2 x2 + = 1. a2 b2 Exercise 6.3.6 Use the method described above to nd the area of the region bounded by the folium of Descartes x3 + y 3 = 3xy in the rst quadrant. The most important step in our proof of the special case of Greens theorem is the application of the fundamental theorem of calculus. In fact, if we take one step back, we can view the special case of Greens theorem as being a version of the fundamental theorem of calculus just one dimension higher. That is, each of them compares the behavior of a function on a region to the behavior of a closely related function on the boundary of that region. In the case of the fundamental theorem, the region is a closed interval [a, b] in R, and the boundary of that region is the set {a, b}. If the function being measured on the boundary is F , then the function being measured on the interval is f = F . In the case of Greens theorem, the region is a compact set in R2 , and the boundary is a piecewise smooth closed curve . If the function being measured on the boundary is the dierential 1-form = P dx + Q dy , then the function being measured P on the set is Q x y . Can we phrase both of these results in a common language? Yes! Unsurprisingly, this language is the language of dierential forms. Let us spend a minute rephrasing the fundamental theorem of calculus purely in the language of dierential forms. Denition 6.3.7 Let Rn . A dierential 0-form on is a function f : R. 201

Denition 6.3.8 Let U Rn , and let f be a C 1 dierential 0-form on U . The dierential of f is the dierential 1-form df on U . Remark 6.3.9 Note that in some sense, the preceding denition is merely a change in notation and terminology from our work in the previous section. (See Denition 6.2.15.) Just as dierential 1-forms can be integrated over sets that are essentially 1-dimensional, that is, piecewise smooth curves, dierential 0-forms can be integrated over sets that are fundamentally 0-dimensional in nature, that is, nite sets. How should one dene the integral of a dierential 0-form on a nite set? By evaluating the function at each point of that set and adding. Before we can dene the integral of a dierential 0-form, however, we must address the issue of orientation. Recall from Exercise 6.2.7 that if is a piecewise smooth parametrized curve and is a closed dierential 1-form, then = . This reects the fact that integration of dierential 1-forms depends not just on the underlying path C as a subset of Rn , but on the orientation of that path, which is encoded in the parametrization. Similarly, it will be important to assign an orientation to the sets on which we integrate dierential 0-forms. Denition 6.3.10 Let X Rn be a nite set. An orientation on X is a function O : X {1}. Denition 6.3.11 Let X Rn be a nite set with orientation O, and let f be a dierential 0-form on X . The integral of f on the oriented set X is dened to be X f = xX O(x) f (x). Exercise 6.3.12 Let X = {1, 2, . . . , n} have orientation O(x) = (1)x . Let f : X R be given by 1 x f (x) = 2 . Find X f . We are now ready to restate the fundamental theorem of calculus. We do so using the language of dierential 0-forms and dierential 1-forms, as well as oriented intervals and their oriented boundaries. Intervals in R are endowed with natural orientations coming from the ordering on the real numbers. If = [a, b] R with a < b, then the boundary is the set = {a, b}, and the natural orientation on this boundary is O : {1} given by O(a) = 1, O(b) = +1. Theorem 6.3.13 (Fundamental Theorem of Calculus) Let = [a, b], and let f be a continuous differential 0-form on [a, b] that is C 1 on (a, b). Then df =

f.

These same ideas immediately apply to give a reformulation of the fundamental theorem for line integrals. Let U Rn be an open set, let : [a, b] U be a piecewise smooth curve, and let = C , the path of . The parametrization of gives a natural orientation to the set , as well as to the boundary , namely O : {1} given by O((a)) = 1 and O((b)) = +1. Theorem 6.3.14 (Fundamental Theorem for Line Integrals) Let U Rn be an open set, let : [a, b] U be a piecewise smooth curve, and let = C . Let f be a C 1 dierential 0-form on U . Then df =

f.

What about Greens Theorem? The right-hand side of Greens Theorem in the plane is already expressed in the form , that is, the integral of a dierential 1-form over the boundary of a 2-dimensional region. If we are to recognize Greens Theorem as the 2-dimensional version of the Fundamental Theorem of Calculus, our new statement of the Fundamental Theorem of Calculus suggests that we should be able to view the left-hand side as the integral of d over the 2-dimensional region . This d should be a dierential 2-form, which should be the dierential of the dierential 1-form . We will explore the formalization of these ideas in the next two sections. Once we do, we will see that we can restate Greens Theorem in the Plane as follows. 202

Theorem 6.3.15 (Greens Theorem in the Plane) Let be a compact region in R2 whose boundary is the path of a piecewise smooth curve. Let be a continuous dierential 1-form on that is C 1 on the interior . Then d =

Finally, the same sort of reasoning by which we can view the fundamental theorem for line integrals as a generalization of the fundamental theorem of calculus will lead us to a version of Greens Theorem for 2-dimensional surfaces in Rn .

6.4

Surfaces in Rn

Our immediate task is to dene and characterize 2-dimensional surfaces in Rn . Traditionally, these objects are referred to simply as surfaces. When it comes time to discuss more general objects of dimension greater than 2, we will introduce new vocabulary. In the same manner that curves are essentially the images of intervals under continuous maps, with some dierentiability hypotheses, surfaces will be the images of rectangles under continuous maps with some dierentiability hypotheses. In the case of curves, we allowed for non-dierentiability on a nite set of points. In the case of surfaces, we need to allow non-dierentiability on larger sets, for examples, at the edges of a cube. To avoid the somewhat complicated theory of how to characterize these sets, we will instead simply parametrize surfaces like a cube one face at a time.
2 Denition 6.4.1 Let {Ri = [ai , bi ] [ci , di ]}m i=1 be a nite collection of closed rectangles in R . A piecewise such that each smooth parametrized surface in Rn is a nite collection of continuous map {i : Ri Rn }m i=1 i is continuously dierentiable except at a nite number of points, and for each point (s, t) Ri for which i is dierentible, Di (s, t) has rank 2 (see Remark 2.2.19), that is, D1 i (s, t) and D2 i (s, t) are linearly n m independent vectors in Rn . Two parametrized surfaces {i : Ri Rn }m i=1 and {i : Ri R }i=1 are equivalent if there exist piecewise C 1 homeomorphisms i : Ri Ri such that det Di > 0 wherever i is dierentiable, and (i i )(s, t) = i (s, t) for all (s, t) Ri . A piecewise smooth surface is an equivalence class of piecewise smooth parametrized surfaces.

Remark 6.4.2 The requirement that det Di be nonzero implies that Di is invertible. The requirement that det Di be positive implies that i is orientation-preserving. (More on this later.) Example 6.4.3 Let R = [0, 1] [0, 2 ], and let : R R2 be given by (s, t) = (s cos t, s sin t). This is the parametrization of the unit disk that we used in Example 5.10.5 to dene polar coordinates. Example 6.4.4 Let R = [0, 1] [0, 2 ], and let : R R3 be given by (s, t) = (s cos t, s sin t, 1 s). This is a parametrization of a cone. Example 6.4.5 Let R = [0, 2 ] [0, ], and let : R R3 be given by (s, t) = (cos s sin t, sin s sin t, cos t). This is the parametrization of the unit sphere coming from spherical coordinates (see Exercise 5.10.6 and Project 5.11.2). Exercise 6.4.6 Verify that D(s, t) has rank 2 for all (s, t) R in the above three examples. Exercise 6.4.7 Find a parametrization of each of the following surfaces. i. T 2 = {(x, y, z, w) R4 | x2 + y 2 = 1, z 2 + w2 = 1} ii. T = (x, y, z ) R3 | 2 x2 + y 2
2

+ z2 = 1 203

One of the most basic things one might like to know about a surface is the surface area. We rst discuss the special case of surfaces in R3 and then see how this generalizes to surfaces in Rn . The rst approach one might think to use is a sort of polyhedral approximation. Suppose we have a piecewise smooth parametrized surface dened by a single rectangle R = [a, b] [c, d] and a single function : R R3 . If we take a regular partition P of R and subdivide each subrectangle of P into a pair of triangles, then the images under of the vertices of each triangle form a triangle in R3 , and together, these triangles form a polyhedral surface in R3 with triangular faces. The sum of the triangular areas is an approximation to the area of our surface. Unfortunately, such approximations do not converge even for some extremely simple examples. Example 6.4.8 (Schwarzs Lantern) Consider the lateral surface of a right circular cylinder C in R3 with height 1 and radius 1. We triangulate C as follows. First, we subdivide C into m congruent bands, 1 each of height m . We then approximate each band with 2n congruent isosceles triangles arranged such the bases of n of them form a regular n-gon on one end of the band, and the bases of the other n form a regular 2 1 , so that the area of + 1 cos n-gon on the other end. The area of one of these triangles is sin n m2 n the lantern is 2 1 A(m, n) = 2mn sin + 1 cos . 2 n m n We might hope that as m and n tend to innity, A(m, n) 2 . Indeed, if we rst let n , and then let m , we get
m

lim

lim A(m, n) = 2.

However, for any xed n > 1,


m

lim A(m, n) = .

Even if we let m and n tend to innity together, we do not necessarily get the right answer. For instance, for any positive integer c, 4 c2 , lim A(cn2 , n) = 2 1 + n 4 which is strictly greater than 2 . We can try to understand the failure of this example by contrasting it with the case of curves. For curves, we designed a polygonal approximation and used the Mean Value Theorem to show that each segment of the polygonal approximation was parallel to the curve at some nearby point. In the above example, on the other hand, if we allow m to grow with n xed, the triangles, far from becoming closer to parallel to the surface, in fact become closer to perpendicular. This example reinforces our claim in Section 4.4 that the Mean Value Theorem is really a one-variable theorem. This rst failed approach is not without its lessons. In the modied approach that follows, we will guarantee that the approximating polygons will be parallel to the surface at at least one point by considering tangent vectors. We again assume for simplicity that we have a surface dened by a single rectangle R = [a, b] [c, d] and a piecewise smooth function : R R3 . Let P1 = {a = a0 < a1 < < ak = b} be a partition of [a, b], let P2 = {c = c0 < c1 < < c = d} be a partition of [c, d], and let P = P1 P2 be the corresponding regular partition of R (see Denition 5.5.17). For each lower-left vertex (ai , cj ), 0 i < k , 0 j < , we consider the associated point on the surface xij = (ai , cj ) and the two tangent vectors uij = D1 (ai , cj ) and vij = D2 (ai , cj ). We then consider the parallelogram spanned by the vectors uij and vij at xij , that is, the parallelogram with vertices xij , xij + uij , xij + uij + vij , and xij + vij . This parallelogram is in fact tangent to the surface at the point xij by construction. What is the area of this parallelogram? Since we have specialized to the case of surfaces in R3 , we know from Exercise 2.5.20 that the area is the norm of the cross product of the two tangent vectors, 204

Aij = uij vij . Summing over the points of our partition, we get an approximation to the surface area given by AP =
k 1 1 i=0 j =0

Aij

k 1 1 i=0 j =0

uij vij .

If we let f : R R be given by f (s, t) = D1 (s, t) D2 (s, t) , we can easily see that L(f, P ) AP U (f, P ). Since is continuously dierentiable except possibly at nitely many points, the function f is integrable on R, and hence, lim P 0 AP = R f . Thus, the area of surface is
b d c

A=
R

D1 D2 =

D1 (s, t) D2 (s, t) dt ds.

Example 6.4.9 Let R = [0, 1] [0, 2 ], and let : R R3 be given by (s, t) = (s cos t, s sin t, 1 s). This is the parametrization of a cone from Example 6.4.4. To nd the surface area, we rst compute the derivative of . D1 (s, t) = (cos t, sin t, 1), So D1 (s, t) D2 (s, t) = (s cos t, s sin t, s cos2 t + s sin2 t) = (s cos t, s sin t, s), and hence D1 (s, t) D2 (s, t) = Thus, the surface area of this cone is
1 2 0

D2 (s, t) = (s sin t, s cos t, 0).

s2 cos2 t + s2 sin2 t + s2 =

2s2 = 2s.

A=
0

2s dt ds =

2.

Exercise 6.4.10 Use the parametrization from Example 6.4.5 to compute the surface area of the unit sphere in R3 . Exercise 6.4.11 Use your parametrization from Exercise 6.4.7.ii to compute the surface area of the torus T in R3 . Exercise 6.4.12 i. Let S = {(x, y, z ) R3 | x2 + z 2 = 1, y 2 + z 2 = 1}. Find a parametrization of S and compute the surface area. ii. Let Find a parametrization of S and compute the surface area. S = {(x, y, z ) R3 | x2 + z 2 = 1, y 2 + z 2 = 1, x2 + y 2 = 1}.

Exercise 6.4.13 Let R = [a, b] [c, d] R2 be a rectangle, let f : R R be a C 1 function, and let S = {(s, t, f (s, t)) | (s, t) R} R3 . Show that the surface area of S is A=
R

1 + (D1 f (s, t))2 + (D2 f (s, t))2 . 205

Exercise 6.4.14 Let : [a, b] R3 be a curve whose path lies in the half of the xz -plane with positive x-coordinate. Let S be the surface obtained by revolving C about the z -axis. Show that the surface area of S is
b

A = 2
a

1 (t) (t) dt =

(21 (t)).

We may use these same ideas to nd a formula for the area of a surface in Rn for n > 3. However, the formula we used for the area of a parallelogram depended on the peculiar existence of the cross product in R3 . We will instead refer back to more elementary considerations and write the area of the parallelogram spanned by vectors u and v as u v sin , where is the angle between u and v. By Theorem 2.5.16, uv cos = u v , so we can rewrite the area of the parallelogram as u v sin cos1 uv u v = u = u v
2

1 v
2

uv u v

(u v)2 .

By the same considerations as in R3 , the area of a surface in Rn parametrized by a function : R Rn can be approximated by the area of parallelograms that are tangent to points at the surface and given exactly as A= D1 2 D2 2 (D1 D2 )2 .
R

Exercise 6.4.15 Use your parametrization from Exercise 6.4.7.i to compute the area of the torus T 2 in R4 .

6.5

Dierential 2-Forms

In the previous section, we dened surfaces in Rn and devised a means of computing their surface areas. In this section, we will see that, just as the length of a curve was the integral of a special dierential 1-form on the curve, the area of a surface can be viewed as the integral of a dierential 2-form on the surface. This viewpoint will lead us to an analogue of Greens Theorem for surfaces in Rn , as we suggested at the end of Section 6.3. Just as a dierential 1-form is designed to measure a single vector at each point in its domain, a dierential 2-form is designed to measure a pair of vectors. Thus, at the innitesimal level, the integration of a dierential 1-form along a curve involved applying a linear form to a single vector, namely, a tangent vector to the curve. Similarly, the integration of a dierential 2-form over a surface will involve applying an alternating bilinear form to a pair of vectors that span the tangent plane to the surface. The prototypical example of an alternating bilinear form is the one in R2 that assigns to an ordered pair of vectors the signed area of the parallelogram spanned by those two vectors. Given two vectors u = (a, b) and v = (c, d) in R2 , recall from Exercise 5.9.14 that the area of the parallelogram spanned by u and v is |ad bc|. Of course, the map that assigns the scalar |ad bc| to the pair (u, v ) is not bilinear, but the closely related map L : R2 R2 R given by L(u, v ) = ad bc is. Notice that while the original formula did not depend on the order of the two vectors, this new signed area does. If we switch u and v , then L(v, u) = bc ad = L(u, v ), and this is what we mean by alternating. We have already dened bilinear forms in Chapter 2, but we restate the denition here for convenience. Denition 6.5.1 A map L : Rn Rn R is called a bilinear form if it satises the following conditions. 1. L(a1 v1 + a2 v2 , w) = a1 L(v1 , w) + a2 L(v2 , w), 2. L(v, b1 w1 + b2 w2 ) = b1 L(v, w1 ) + b2 L(v, w2 ), a1 , a2 R, v1 , v2 , w Rn b1 , b2 R, v, w1 , w2 Rn

Exercise 6.5.2 Decide whether or not the following maps are bilinear forms. 206

1. L : R2 R2 R given by L((x1 , y1 ), (x2 , y2 )) = x1 + x2 + y1 + y2 2. L : R2 R2 R given by L((x1 , y1 ), (x2 , y2 )) = x1 x2 3. L : R2 R2 R given by L((x1 , y1 ), (x2 , y2 )) = x1 x2 + y1 y2 4. L : R2 R2 R given by L((x1 , y1 ), (x2 , y2 )) = x1 y2 x2 y1 5. L : R3 R3 R given by L((x1 , y1 , z1 ), (x2 , y2 , z2 )) = x1 y2 x2 y1 Denition 6.5.3 A bilinear form L : Rn Rn R is said to be alternating if L(v, w) = L(w, v ), v, w Rn . We denote by An the collection of alternating bilinear forms on Rn . Exercise 6.5.4 Of the maps from the previous exercise which are bilinear forms, determine which are alternating. Exercise 6.5.5 Show that the collection An of alternating bilinear forms on Rn is a real vector space. Remark 6.5.6 Note that our original bilinear forms of interest from Chapter 2, namely inner products, are not alternating, but rather symmetric. In fact, every bilinear form can be uniquely written as a sum of an alternating form and a symmetric form, so we are in some sense working with the complementary category of forms here. Exercise 6.5.7 Show that if L : R2 R2 R is an alternating bilinear form, then there exists a scalar c R such that L((x1 , y1 ), (x2 , y2 )) = c(x1 y2 x2 y1 ). Exercise 6.5.8 Show that dim(An ) = exercise as a template.) n(n 1) . (Hint: Construct a basis using the form from the previous 2

This vector space An is the 2-dimensional analogue of the vector space L(Rn , R) of linear forms on Rn , which are the innitesimal building blocks for measuring 1-dimensional aspects of Euclidean space. In other words, the elements of An will be the innitesimal building blocks for measuring 2-dimensional aspects of Euclidean space. It should not be surprising, then, that we can combine two linear forms to obtain a bilinear form. Denition 6.5.9 Let S, T L(Rn , R). The wedge product of S and T is the alternating bilinear form S T : Rn Rn R given by (S T )(v, w) = S (v )T (w) T (v )S (w). Exercise 6.5.10 i. Verify that S T is an alternating bilinear form. ii. Show that T S = S T . iii. Show that S S = 0. Exercise 6.5.11 Show that the bilinear form L : R2 R2 R from Exercise 6.5.7 given by L((x1 , y1 ), (x2 , y2 )) = c(x1 y2 x2 y1 ) is the wedge product of two linear forms. Exercise 6.5.12 Find a basis for An consisting of wedge products of linear forms. We are now ready to dene dierential 2-forms. Denition 6.5.13 Let Rn . A dierential 2-form on is a map : An , where An is the collection of alternating bilinear forms on Rn . The bilinear map in An associated with the point x is denoted x . 207

1) Remark 6.5.14 Because An is a nite dimensional real vector space of dimension k = n(n2 , we can pick k k a basis to identify it with R . We then give R the usual metric. This allows us to dene what it means for a dierential 2-form : An to be continuous, dierentiable, smooth, etc.

Example 6.5.15 Let L An . The map : Rn An given by x = L is a constant dierential 2-form. Example 6.5.16 Let : [a, b] [c, d] Rn be a simple smooth parametrized surface in Rn with image S . The area form of is the dierential 2-form : S An given by (s,t) = that is, (s,t) (v, w) = D1 (s, t)(v ) D2 (s, t)(w) D2 (s, t)(v ) D1 (s, t)(w) . D1 (s, t) D2 (s, t) D2 (s, t) D1 (s, t) D2 (s, t) D1 (s, t) , D1 (s, t) D2 (s, t)

The above example shows that just as two linear forms can be combined using the wedge product to give an alternating bilinear form, two dierential 1-forms can be combined pointwise using the wedge product to give a dierential 2-form. In fact, this is the rst of two essential methods for constructing dierential 2-forms. Denition 6.5.17 Let Rn , and let , : L(Rn , R) be dierential 1-forms. The wedge product of and , denoted , is the dierential 2-form dened by ( )x (v, w) = x (v )x (w) x (w)x (v ). Example 6.5.18 Recall that the dierential 1-forms dx and dy on R2 are given by dxa (v ) = v1 and dya (v ) = v2 for a R2 and v = (v1 , v2 ) R2 . The dierential 2-form dx dy on R2 is thus given by (dx dy )a (v, w) = v1 w2 v2 w1 , where v = (v1 , v2 ) R2 and w = (w1 , w2 ) R2 . Note that this is the constant form corresponding to the alternating bilinear form L from Exercise 6.5.7 with c = 1. Exercise 6.5.19 Use Exercise 6.5.7 to show that every bilinear form on a set R2 is of the form = f dx dy , where f : R is a function. Exercise 6.5.20 Recall that the dierential 1-form dxi on Rn is given by (dxi )a (v ) = vi , where a Rn and v = (v1 , v2 , . . . , vn ) Rn . Use Exercise ?? to show that every bilinear form on a set Rn can be written as = fij dxi dxj ,
1i<j n

where the fij : R are functions. Exercise 6.5.21 Compute the given wedge product. i. (x dx + y dy ) (y dx x dy ) ii. (y cos x dx x sin x dy ) (y sin x dx + x cos x dy ) iii. (x2 y dx + xyz dy + xy 2 dz ) (xyz dx + yz 2 dy + dz ) We are now prepared to integrate dierential 2-forms. For simplicity, we make the following denition for the integral of a dierential 2-form on a smooth surface. It is clear how to extend the denition to piecewise smooth surfaces. 208

Denition 6.5.22 Let : [a, b] [c, d] Rn be a smooth parametrized surface with image S . Let : S An be a continuous dierential 2-form on S . The integral of over is dened to be
b d

=
a c

(s,t) (D1 (s, t), D2 (s, t)) dt ds.

We rst show that this integral is independent of the parametrization of S . Proposition 6.5.23 Let R = [a, b] [c, d] and R = [a , b ] [c , d ] be rectangles in R2 . Suppose : R Rn and : R Rn are equivalent parametrizations of the surface S . Let : S An be a continuous dierential 2-form. Then = . Proof. Let : R R be the C 1 homeomorphism such that det D > 0 and = . Using the chain rule, we can write
b d

=
a b c d c b d c

(s,t) (D1 (s, t), D2 (s, t)) dt ds ( (s,t)) (D1 ( )(s, t), D2 ( )(s, t)) dt ds ( (s,t)) (D1 ( (s, t)) D1 1 (s, t) + D2 ( (s, t)) D1 2 (s, t), D1 ( (s, t)) D2 1 (s, t) + D2 ( (s, t)) D2 2 (s, t)) dt ds.

=
a

=
a

For notational simplicity, we let L = ( (s,t)) , v1 = D1 ( (s, t)), v2 = D2 ( (s, t)), and aij = Di j for 1 i, j, 2. By the denition of an alternating bilinear form, we can write the integrand as L(a11 v1 + a12 v2 , a21 v1 + a22 v2 ) = L(a11 v1 , a21 v1 ) + L(a11 v1 , a22 v2 ) + L(a12 v2 , a21 v1 ) + L(a12 v2 , a22 v2 ) = L(a11 v1 , a22 v2 ) L(a21 v1 , a12 v2 ) = (a11 a22 a21 a12 )L(v1 , v2 ) = det D (s, t)L(v1 , v2 ). Since we assumed det D > 0, it follows that det D = |det D |. Thus, by change of variables,
b d c b

=
a

( (s,t)) (D1 ( (s, t)), D2 ( (s, t))) |det D (s, t)| dt ds


d

=
a c

(s ,t ) (D1 (s , t ), D2 (s , t )) dt ds

Remark 6.5.24 Since the integral of a dierential 2-form is independent of the parametrization, we will often write S for when parametrizes S . Exercise 6.5.25 i. Let : [a, b] [c, d] Rn be a simple smooth parametrized surface in Rn . Let be the area form of dened in Example 6.5.16. Show that is the surface area of . 209

ii. Show that the area form is independent of parametrization, so that the surface area is independent of parametrization. Example 6.5.26 Let a, b, c > 0, and let : [0, a] [0, b] R3 be given by (s, t) = (s, t, c c b t). This parametrizes the rectangle S in R3 whose vertices are (0, 0, c), (a, 0, c), (a, b, 0), and (0, b, 0). We have c ). D1 (s, t) = (1, 0, 0), and D2 (s, t) = (0, 1, b We can compute
a S b 0 a b 0

dx dy = =

(dx dy ) (1, 0, 0), 0, 1, (1 1 0 0) dt ds

c b

dt ds

= ab. This is the area of the projection of the rectangle S onto the xy -plane, and this makes sense because dx dy should measure the xy -ness of the surface. Exercise 6.5.27 Let S be the rectangle from the above example. Compute the following integrals. i. ii. iii. iv.
S S S S

dx dz dy dz dy dx

As the example and exercise above show, dierent dierential 2-forms can be used to measure dierent 2-dimensional aspects of a surface. Exercise 6.5.28 Let S = {(x, y, z ) R3 | x2 + y 2 + z 2 = 1, z 0} be the upper hemisphere of the unit sphere in R3 . For each of the following integrals, rst predict what the integral will be, and then do the computation to verify your prediction. i. ii. iii. iv. v.
S S S S S

dx dy dy dx dx dz 1 x2 y 2 dx dy

210

You might also like