ARE211, Fall2013

ANALYSIS5: TUE, SEP 17, 2013 PRINTED: SEPTEMBER 19, 2013 (LEC# 5)

Contents 1. Analysis (cont) 1.8. 1.8.5. 1.8.6. 1.8.7. 1.9. Topology of Rn (cont) Closure of a Set Boundary of a Set Compact Sets The Bolzano-Weierstrass Theorem 1 1 1 1 3 6

1. Analysis (cont) 1.8. Topology of Rn (cont)

1.8.5. Closure of a Set. The analog of the interior of a set is the closure of a set. in X is the intersection of all Denition: Let A X . The closure of A in X , denoted cl(A) or A closed sets containing A. Theorem: For A X , A is closed in X i A = cl(A) in X . 1.8.6. Boundary of a Set. A point x is a boundary point of a set A X if there are points arbitrarily close to x that are in A and if there are points arbitrarily close to x that are in X but not in A.

ANALYSIS5: TUE, SEP 17, 2013

PRINTED: SEPTEMBER 19, 2013

(LEC# 5)

Example: The set A = (x, y ) R2 : x2 + y 2 < 1 . Its boundary is the circle (x, y ) R2 : x2 + y 2 = 1 That is, the boundary is the border between A and X \A. Denition: The set of boundary points of A in X , denoted bd(A), is the set cl(A) cl(X \A). Theorem: Let A X . A point x bd(A) i > 0, y, z B (x, ) such that y A and z X \A. Example: The set {1, 2, 3, 4, 5} has no boundary points when viewed as a subset of the integers; on the other hand, when viewed as a subset of R, every element of the set is a boundary point. Theorem: A set A X is closed in X i A contains all of its boundary points. Note the dierence between a boundary point and an accumulation point. Take the set A = {0} R. 0 is a boundary point of A but not an accumulation point. On the other hand, every element of the interval A = (0, 1) R is an accumulation point of A, but A contains none of the boundary points of A. Though boundary points and accumulation points are resoundingly dierent in general, there is a close connection between the two concepts. Indeed, many textbooks use one of the following two denitions of a closed set: (1) a set is closed if it contains all of its accumulations points (2) a set is closed if it contains all of its boundary points These two denitions would suggest that the two terms are synonyms for each other, but of course they arent. The following theorem explains the relationship. Theorem: Given a set A X , a point x X that does not belong to A is a boundary point of A in X i it is an accumulation point of A in X .

ARE211, Fall2013

Proof: If: Suppose x bd(A), but x / A. Fix > 0. By denition of a boundary point, y B (x, ) A. Since x / A, y = x. Weve established, then, that > 0, x = y B (x, ). Hence x is an accumulation point of A. The Only if part is parallel. Please try to do it as an exercise.

1.8.7. Compact Sets. Importance of compact sets: continuous functions dened on compact sets always attain their extrema. Denition: Given a set S X , an open covering of S is a collection C of open sets in X such that S {O : O C}.

This is a rather abstract notion because the set C can be arbitrarily large, i.e., uncountably innite. Denition: A set X is said to be compact if every open covering C of X has a nite open subcover, i.e., if there exists a nite collection of sets {O1 , ...On } C such that X
n i=1 Oi .

(Note: In some denitions, e.g., Wikipedia, youll see equality signs where I have containment signs. The reason is that the denitions in Wikipedia apply to spaces while my denitions apply to subsets of some universe. Dont worry about the distinction, its not important for us, just wanted to reassure you that Im not being inconsistent with Wikipedia (and many other sources)!) Example: Consider the closed set S = [0, 1], and for n N, let C = {B (x, 1/n) : x S }. Obviously this uncountably innite set covers S . But there are many nite subsets of C that also cover S , in particular, the set {O0 , ...O2n } C , where for arbitrary n N and i = 0, ...2n, Oi = B (i/2n, 1/n). Example: On the other hand, consider the open set S = (0, 1), and its open cover C = {B (x, x/2) : x S }. Note that this set indeed covers S : pick x S , then x B (x, x/2) C , establishing that X
xS

B (x, x/2). Now consider any n N and any nite set X = {x1 , ..., xn } S , (i.e.,

xi (0, 1), for all xi X ). Now dene {O1 , ...On } = {B (xi , xi /2) : xi X }. Well show that this

ANALYSIS5: TUE, SEP 17, 2013

PRINTED: SEPTEMBER 19, 2013

(LEC# 5)

subset doesnt cover S . Let x = min{xi X }. Now x/4 > 0 so that x/4 S . Well show that x/4 /
n i=1 Oi .

Now x / Oi if x xi /2. But x/4 < x x/2 = x/2 xi /2, for all xi X . Since

any nite subset of C is of the form we just considered, weve proved that there is no nite subset of C that covers S . Happily, the compact subsets of Euclidean spaces can be characterized in a simple way: compact sets are closed and bounded. Ive already given you a denition of boundedness for a subset of R. Heres one for subsets of Rn . Denition: For x Rn , the norm of x, written ||x||, is the Euclidean distance between x and zero, i.e., ||x|| =
n 2 i=1 xi .

Denition: A set A Rn is bounded if there exists a real number b R such that ||x|| b, for all x A. (Note that this denition of bounded is metric-specic!) Theorem: (Heine Borel) A set S Rn is compact with respect to a metric d i it is closed and bounded with respect to that metric. Its relatively straightforward to prove the only if direction of this theorem. (1) for any S is not closed, one can generalize the construction I used in the above example (for S = (0, 1)) to construct an open cover for which no nite subcover exists: (a) As in the example on pp. 3-4, rst consider C = {B (x, x/2) : x S }. (b) This may no longer work, for an arbitrary non-closed set S . Why not? Why not? Because, in that example, we used the fact that xi > 0, for all i. This implied that every ball in C contained its center. We no longer know this property. For example, we have to consider non-closed sets such as [0, 2]\{1}. And in this case, C would no longer cover S .

ARE211, Fall2013

(c) So what do we know? That y0 R\S , which is an accumulation point of S . So lets consider the following collection of open sets, which generalizes the one we picked on pp. 3-4, i.e. C = {B (x, d(x, y0 )/2) : x S }. (To see that this is a generalization of B (x, x/2) : x S }, let y0 = 0.) Note that this set indeed now covers S , because s S implies d(y0 , x)/2) > 0. Now, as before, pick an arbitrary nite set X = {x1 , ..., xn } S , and let {O1 , ...On } = {B (xi , d(y0 , xi )/2) : xi X }. We need to show that this collection doesnt cover S . For i = 1, ...n, let i = d(y0 , xi )/2 and pick = mini i . Since y0 is an accumulation point of S , there exists y S such that d(y0 , y ) < . Well show that y / d(y, xi ) d(y, xi ) +
n i=1 Oi .

For 1 i n,1 > d(xi , y0 ) d(xi , y0 ) 2i i

d(y0 , y )

d(y0 , y )

therefore y / B (xi , d(y0 , xi )/2), i.e., weve shown that

does not cover S

(2) now assume that S is not bounded above and consider the collection C =

where

On = {s S : ||s|| < n}. Clearly C covers Rn so it covers S . Now let N be an arbitrary nite set and consider the subset
nN

On . Since S is unbounded, there exists x S , such

1 This is an example of a generic proof technique. (i) You have three distances (A) one you know is big (B) one you know is small (C) one you want to prove is big (ii) you should always set up your triangle inequality this way: know big know big know small = (< ) e.g.(> 2) want to prove big want to prove big want to prove big want to prove big + know small

ANALYSIS5: TUE, SEP 17, 2013

PRINTED: SEPTEMBER 19, 2013 nN

(LEC# 5)

that ||x|| > max{n N }. Hence x / cover S .

On . Therefore, the subset we dened does not

Im not going to prove the if part, its substantially harder. To prove that closed and bounded implies compact, we would need to prove either one of the following statements (1) if there exists an open cover of S which doesnt have a nite open subcover and S is closed, then S must be unbounded. (2) if there exists an open cover of S which doesnt have a nite open subcover and S is bounded, then S cannot be closed. Proof of either statement would establish the if result. Note that as with everything else in topology, the notion of compactness depends critically on the metric Example: Consider the set [0, 1]. Its clearly closed and bounded with respect to the Euclidean metric, and so compact in the Euclidean metric. Its denitely not compact with respect to the discrete metric, however: the collection C = {{x} : x [0, 1]} is an open cover of [0, 1]; it obviously has no nite open subcover.

1.9. The Bolzano-Weierstrass Theorem

We are going to prove one of the most fundamental theorems of analysis: Theorem: (idea) Every sequence dened on a compact set A Rn contains a convergent subsequence. REVIEW:

ARE211, Fall2013

(1) recall that a sequence in A is called a convergent sequence if it converges to an element of the set A. E.g., the sequence (1/n) in the set A = (0, 2) R is not a convergent subsequence because the point in R that is the limit of the sequence does not belong to the set A that contains the sequence. (2) Recall the denition of a subsequence: a sequence {y1 , y2 , ..., yn , ...} is a subsequence of {x1 , x2 , ..., xn , ...} if there exists a strictly increasing function : N N such that for all n N, yn = x (n) . That is, you construct a subsequence by throwing out elements of the original sequence, but keeping an innite number of the original elements and preserving their order.

Intuitively, the theorem were going to prove is obvious: points in the sequence have to bunch up somewhere. Nowhere for them to escape to. The basic idea of the proof is: because the set is bounded, we can construct a Cauchy subsequence of the original sequence; because the set is closed, the point to which the subsequence is trying to converge actually belongs to the set, and so is indeed a limit of the subsequence. The proof is going to involve three inductive sequence constructions: an inductive construction of a sequence involves: (1) an initial step: dene the rst element of the sequence (2) an inductive step: assume that the nth element of the sequence has been dened; now dene the n + 1th element in terms of the nth. For example, a rst order dierence equation is dened by induction: dene x0 ; then dene xt+1 = axt + b. Back to Bolzano-Weierstrass. Heres the intuition for the proof, for the case of a sequence in R:

ANALYSIS5: TUE, SEP 17, 2013

PRINTED: SEPTEMBER 19, 2013

(LEC# 5)

(1) the initial step: dene a set A1 . (a) Because the set A is compact, its bounded. Pick b R++ such that A is contained in the interval [b/2, b/2]. (Youll see why we divide by 2 in a minute.) Now consider any sequence in A. (b) Partition2 the interval [b/2, b/2] into two halves, e.g., [b/2, 0) and [0, b/2], such that at least one half contains an innite number of elements in the sequence (possibly both halves). Pick this half (if there are two such halves, pick either). The set youve picked will be called A1 . (2) the inductive step: assume that An has been dened, now dene An+1 . (a) The inductive step: take the set youve just picked, An , split it in half, and pick again, according to the same criterion, to get An+1 . Keep going forever; note that you now have an innite sequence of nested subsets of A. (b) Dene a subsequence of the original sequence as follows: the rst element of the subsequence is the rst element of the original sequence that belongs to the rst subdivided set A1 . (c) The second element of the subsequence is the rst element in the original sequence that comes after the one youve just chosen which lies in the second subdivided set, A2 , etc. Example: Take the sequence {1, 1, 1, 1, ... } [1, 1]. Set A1 = [0, 1], A2 = [1/2, 1], A3 = [3/4, 1], etc. To construct a convergent subsequent, set (1) = 2, (2) = 4, (3) = 6, etc., and note that for all n, yn = x (n) An (d) Now you have constructed an innite (sub)sequence of points with the property that for every N , the entire tail of the subsequence (after discarding the rst N points) lies in the N th subdivision of A.
2 A partition of a set X is a collection C = (X ) of subsets of X such that X = i Xi C Xi and Xi Xj =, for i = j . Informally it is a collection of subsets of X such that every element of X is contained in exactly one element of the collection.

ARE211, Fall2013

(e) Since the subdivided intervals of A are getting smaller and smaller, the sequence is a Cauchy sequence. (f) Every Cauchy sequence dened in R converges to a point in R. (g) The point in R to which the subsequence converges is either an element of the subsequence or an accumulation point of the set A. Since A is compact, hence closed, it contains its accumulation points, hence the sequence weve constructed indeed converges to a point in A. (h) To illustrate the two possibilities mentioned in 2g, consider (i) The set A = {1, 1}, which has no accumulation points. Consider the sequence xn = (1)n . In this case, one convergent subsequence is {1, 1, ...} and the limit point 1 is an element of the subsequence. (ii) The set A = [0, 1]. Consider the sequence xn = 1/n if n is even . if n is odd

In this case, {1/n} is a convergent subsequence. This sequence converges to zero, but 0 is not an element of the sequence, but it is an accumulation point of [0, 1]. (3) Its important to understand the relationship between the three inductive constructions weve just made. (a) First, we constructed a sequence of subsets {An } of the codomain A of the sequence (b) Second, we identied a sequence of innite subsets of the domain N of the sequence. The nth subset was dened by the property that every element of it mapped into An . (c) Third, we constructed our subsequence mapping (n), such that for each n (n) belonged to the nth subset dened in 3b. The tricky part was to make sure that (n) was a strictly increasing sequence, as required by the denition of a subsequence. Theorem: (Bolzano-Weierstrass) A set A Rn is compact if and only if every sequence {xn } n=1 in
A has a convergent subsequence, i.e., there exists a subsequence {yn } n=1 of {xn }n=1 and a point

some random number

10

ANALYSIS5: TUE, SEP 17, 2013

PRINTED: SEPTEMBER 19, 2013

(LEC# 5)

y A such that {yn } converges to y. Proof: Well do the proof for A R and only in one direction, i.e., well show that A compact implies each sequence in A has a convergent subsequence. Suppose that A is bounded above by b/2 and below by b/2: that is, the interval [b/2, b/2] contains the entire (innite) number of points in the sequence. Partition the set [b/2, b/2] into two subsets of equal length,3 e.g., into [b/2, 0] and (0, b/2] and observe that at least one of these subsets contains an innite number of elements of the sequence {xn }. Denote by A1 whichever subset this is. If both have this property, then pick either. The length of the subset A1 is equal to b/2. Now argue by induction. Suppose that weve dened a subset An of length b/2n which contains an innite number of elements of the sequence {xn }. (This statement is true for n = 1 i.e., length of A1 is b/2 = b/21 .) Well construct a subset An+1 of length b/2n+1 which contains an innite number of elements of {xn }: simply divide An in half as before: at least one of the subsets contain an innite number of elements. Now weve constructed an innite sequence of subsets, {An } with the property that for each n, (a) An+1 An (i.e., the An s are nested) (b) the length of An equals b/2n (c) An contains an innite number of elements of {xn }. Now well dene a convergent subsequence {yn }. Dene the strictly increasing sequence as follows: let (1) = min {k N : xk A1 }; now assume that (n) has been dened and dene (n + 1) = min {k { (n) + 1, ...,} : xk An+1 }. Note that the sets were considering in the previous sentence are subsets of the domain of the sequence, i.e., subsets of N. Note also that by construction, () is a strictly increasing function of n. Observe that we now have constructed a subsequence {yn }, i.e., yn = x (n) , for each n, with the property that for each n, yn An . Moreover, since the An s are nested, it follows that for all N , and n, m > N , yn , ym AN . That is, {yn } is a Cauchy sequence. Now, we know that every Cauchy
3 The length or diameter of a set A, endowed with metric d, is sup({d(x, y ) : x, y A}).

ARE211, Fall2013

11

sequence in R converges to a point in R. i.e., there exists y R such that for all > 0, N N such that for all n > N , d(yn , y ) < . We need to show that y A. If y = yn , for some n, were done, since yn An A. If not, then y is an accumulation point of the set A, since for all > 0, there exists n such that yn = y and yn B (y, ). But since A is a closed set, it contains all of its accumulation points. Hence y A, and weve proved that the subsequence weve constructed converges to a point y A. Perhaps the trickiest part of this proof is the point at which we build the subsequence, i.e., construct the mapping . The method Ive proposed above is by no means the only way to build a sequence, but it has a distinct merit, which is that it works. Its important to see what goes wrong with some of the alternative constructions one might think of. For example, (1) Suppose in the inductive step, i.e., once (n) has been dened, you dened (n + 1) = min {k N : xk An+1 } rather than, as I did above, min {k N; k (n) : xk An+1 }. Why wouldnt the former alternative work? Heres an example: x1 = 0; x2 = 1; xn =
3 8 , n

3. This sequence is bounded above by b/2, for all b 2. Let b = 2 and dene

1 1 A1 = [0, 1]; A2 = [0, 1 2 ]; A3 = [ 4 , 2 ], etc. Now consider what n = min {k N : xk An+1 } 3 3 delivers: (1) = 1; (2) = 1, (3) = 3, delivering the sequence z = {0, 0, 8 , 8 , }. This 3 3 is not, however, a subsequence of {x} = {0, 1, 3 8 , 8 , 8 , }. If, instead, you proceeded as

above, and set (n + 1) = min {k N; k (n) : xk An+1 }, the resulting mapping would be (1) = 1 and for all n > 1, (n) = n + 1. This mapping would deliver the sequence
3 y = {0, 3 8 , 8 , } which is a subsequence of {xn }.

(2) There was talk in class about picking the sequence element that is the minimum of all sequence elements in An , or, more formally, setting (n) = argmin(An ) or x (n) = min{xk : xk An }. There are a couple of reasons not to do this, but the most basic one is that in general this minimum wont exist. For example, let xn = 1/n and dene An = [0, 1/n]. Now note that for each n, min{xk : xk An } doesnt exist.