This action might not be possible to undo. Are you sure you want to continue?

)

Jason Rosendale

jason.rosendale@gmail.com

December 19, 2011

This work was done as an undergraduate student: if you really don’t understand something in one of these

proofs, it is very possible that it doesn’t make sense because it’s wrong. Any questions or corrections can be

directed to jason.rosendale@gmail.com.

Exercise 1.1a

Let r be a nonzero rational number. We’re asked to show that x ,∈ Q implies that (r + x) ,∈ Q. Proof of the

contrapositive:

→ r +x is rational assumed

→ (∃p ∈ Q)(r +x = p) deﬁnition of rational

→ (∃p, q ∈ Q)(q +x = p) we’re told that r is rational

→ (∃p, q ∈ Q)(x = −q +p) existence of additive inverses in Q

Because p and q are members of the closed additive group of Q, we know that their sum is a

member of Q.

→ (∃u ∈ Q)(x = u)

→ x is rational deﬁnition of rational

By assuming that r+x is rational, we prove that x must be rational. By contrapositive, then, if x is irrational

then r +x is irrational, which is what we were asked to prove.

Exercise 1.1b

Let r be a nonzero rational number. We’re asked to show that x ,∈ Q implies that rx ,∈ Q. Proof of the

contrapositive:

→ rx is rational assumed

→ (∃p ∈ Q)(rx = p) deﬁnition of rational

→ (∃p ∈ Q)(x = r

−1

p) existence of multiplicative inverses in Q

(Note that we can assume that r

−1

exists only because we are told that r is nonzero.) Because r

−1

and p are members of the closed multiplicative group of Q, we know that their product is also a

member of Q.

→ (∃u ∈ Q)(x = u)

→ x is rational deﬁnition of rational

By assuming that rx is rational, we prove that x must be rational. By contrapositive, then, if x is irrational

then rx is irrational, which is what we were asked to prove.

1

Exercise 1.2

Proof by contradiction. If

√

12 were rational, then we could write it as a reduced-form fraction in the form of

p/q where p and q are nonzero integers with no common divisors.

→

p

q

=

√

12 assumed

→ (

p

2

q

2

= 12)

→ (p

2

= 12q

2

)

It’s clear that 3[12q

2

, which means that 3[p

2

. By some theorem I can’t remember (possibly the

deﬁnition of ‘prime’ itself), if a is a prime number and a[mn, then a[m∨ a[n. Therefore, since 3[pp

and 3 is prime,

→ 3[p

→ 9[p

2

→ (∃m ∈ N)(p

2

= 9m) deﬁnition of divisibility

→ (∃m ∈ N)(12q

2

= 9m) substitution from p

2

= 12q

2

→ (∃m ∈ N)(4q

2

= 3m) divide both sides by 3

→ (3[4q

2

) deﬁnition of divisibility

From the same property of prime divisors that we used previously, we know that 3[4 ∨ 3[q

2

: it

clearly doesn’t divide 4, so it must be the case that 3[q

2

. But if 3[qq, then 3[q ∨ 3[q. Therefore:

→ (3[q)

And this establishes a contradiction. We began by assuming that p and q had no common divisors, but we

have shown that 3[p and 3[q. So our assumption must be wrong: there is no reduced-form rational number such

that

p

q

=

√

12.

Exercise 1.3 a

If x ,= 0 and xy = xz, then

y = 1y = (x

−1

x)y = x

−1

(xy) = x

−1

(xz) = (x

−1

x)z = 1z = z

Exercise 1.3 b

If x ,= 0 and xy = x, then

y = 1y = (x

−1

x)y = x

−1

(xy) = x

−1

x = 1

Exercise 1.3 c

If x ,= 0 and xy = 1, then

y = 1y = (x

−1

x)y = x

−1

(xy) = x

−1

1 = x

−1

= 1/x

Exercise 1.3 d

If x ,= 0, then the fact that x

−1

x =1 means that x is the inverse of x

−1

: that is, x = (x

−1

)

−1

= 1/(1/x).

Exercise 1.4

We are told that E is nonempty, so there exists some e ∈ E. By the deﬁnition of lower bound, (∀x ∈ E)(α ≤ x):

so α ≤ e. By the deﬁnition of upper bound, (∀x ∈ E)(x ≤ β): so e ≤ β. Together, these two inequalities tell us

that α ≤ e ≤ β. We’re told that S is ordered, so by the transitivity of order relations this implies α ≤ β.

2

Exercise 1.5

We’re told that A is bounded below. The ﬁeld of real numbers has the greatest lower bound property, so we’re

guaranteed to have a greatest lower bound for A. Let β be this greatest lower bound. To prove that −β is

the least upper bound of −A, we must ﬁrst show that it’s an upper bound. Let −x be an arbitrary element in −A:

→ −x ∈ −A assumed

→ x ∈ A deﬁnition of membership in −A

→ β ≤ x β = inf(A)

→ −β ≥ −x consequence of 1.18(a)

We began with an arbitrary choice of −x, so this proves that (∀ −x ∈ −A)(−β ≥ −x), which by deﬁnition

tells us that −β is an upper bound for −A. To show that −β is the least such upper bound for −A, we choose

some arbitrary element less than −β:

→ α < −β assumed

→ −α > β consequence of 1.18(a)

Remember that β is the greatest lower bound of A. If −α is larger than inf(A), there must be some

element of A that is smaller than −α.

→ (∃x ∈ A)(x < −α) (see above)

→ (∃ −x ∈ −A)(−x > α) consequence of 1.18(a)

→ !(∀ −x ∈ −A)(−x ≤ α)

→ α is not an upper bound of −A deﬁnition of upper bound

This proves that any element less than −β is not an upper bound of −A. Together with the earlier proof

that −β is an upper bound of −A, this proves that −β is the least upper bound of −A.

Exercise 1.6a

The diﬃcult part of this proof is deciding which, if any, of the familiar properties of exponents are considered

axioms and which properties we need to prove. It seems impossible to make any progress on this proof unless we

can assume that (b

m

)

n

= b

mn

. On the other hand, it seems clear that we can’t simply assume that (b

m

)

1

n

= b

m/n

:

this would make the proof trivial (and is essentially assuming what we’re trying to prove).

As I understand this problem, we have deﬁned x

n

in such a way that it is trivial to prove that (x

a

)

b

= x

ab

when a and b are integers. And we’ve declared in theorem 1.21 that, by deﬁnition, the symbol x

1

n

is the element

such that (x

n

)

1

n

= x. But we haven’t deﬁned exactly what it might mean to combine an integer power like n

and some arbitrary inverse like 1/r. We are asked to prove that these two elements do, in fact, combine in the

way we would expect them to: (x

n

)

1/r

= x

n/r

.

Unless otherwise noted, every step of the following proof is justiﬁed by theorem 1.21.

3

→ b

m

= b

m

assumed

→ ((b

m

)

1

n

)

n

= b

m

deﬁnition of x

1

n

→ ((b

m

)

1

n

)

nq

= b

mq

We were told that

m

n

=

p

q

which, by the deﬁnition of the equality of rational numbers, means that

mq = np. Therefore:

→ ((b

m

)

1

n

)

nq

= b

np

→ ((b

m

)

1

n

)

qn

= b

pn

commutativity of multiplication

From theorem 12.1, we can take the n root of each side to get:

→ ((b

m

)

1

n

)

q

= b

p

From theorem 12.1, we can take the q root of each side to get:

→ (b

m

)

1

n

= (b

p

)

1

q

Exercise 1.6b

As in the last proof, we assume that b

r+s

= b

r

b

s

when r and s are integers and try to prove that the operation

works in a similar way when r and s are rational numbers. Let r =

m

n

and let s =

p

q

where m, n, p, q ∈ Z and

n, q ,= 0.

→ b

r+s

= b

m

n

+

p

q

→ b

r+s

= b

mq+pn

nq

deﬁnition of addition for rationals

→ b

r+s

= (b

mq+pn

)

1

nq

from part a

→ b

r+s

= (b

mq

b

pn

)

1

nq

legal because mq and pn are integers

→ b

r+s

= (b

mq

)

1

nq

(b

pn

)

1

nq

corollary of 1.21

→ b

r+s

= (b

mq

nq

)(b

pn

nq

) from part a

→ b

r+s

= (b

m

n

)(b

p

q

)

→ b

r+s

= (b

r

)(b

s

)

Exercise 1.6c

We’re given that b > 1. Let r be a rational number. Proof by contradiction that b

r

is an upper bound of B(r):

→ b

r

is not an upper bound of B(r) hypothesis of contradiction

→ (∃x ∈ B(r))(x > b

r

) formalization of the hypothesis

By the deﬁnition of membership in B(r), x = b

t

where t is rational and t ≤ r.

→ (∃t ∈ Q)(b

t

> b

r

∧ t ≤ r)

It can be shown that b

−t

> 0 (see theorem S1, below) so we can multiply this term against both

sides of the inequality.

→ (∃t ∈ Q)(b

t

b

−t

> b

r

b

−t

∧ t ≤ r) theorem S2

→ (∃t ∈ Q)(b

t−t

> b

r−t

∧ t ≤ r) from part b

→ (∃t ∈ Q)(1 > b

r−t

∧ r −t ≥ 0)

→ (∃t ∈ Q)(1

−(r−t)

> b ∧ r −t ≥ 0)

→ 1 > b

And this establishes our contradiction, since we were given that b > 1. Our initial assumption must have

been incorrect: b

r

is, in fact, an upper bound of B(r). We must still prove that it is the least upper bound of

B(r), though. To do so, let α represent an arbitrary rational number such that b

α

< b

r

. From this, we need to

4

prove that α < r.

→ b

α

< b

r

hypothesis of contradiction

→ b

α

b

−r

< b

r

b

−r

theorem S2

→ b

α−r

< b

r−r

from part b

→ b

α−r

< 1 from part b

Exercise 1.7 a

Proof by induction. Let S = ¦n : b

n

−1 ≥ n(b −1)¦. We can easily verify that 1 ∈ S. Now, assume that k ∈ S:

→ k ∈ S hypothesis of induction

→ b

k

−1 ≥ k(b −1) deﬁnition of membership in S

→ bb

k

−1 ≥ k(b −1) we’re told that b > 1.

→ b

k+1

−b ≥ k(b −1)

→ b

k+1

≥ k(b −1) +b

→ b

k+1

−1 ≥ k(b −1) +b −1

→ b

k+1

−1 ≥ (k + 1)(b −1)

→ k + 1 ∈ S deﬁnition of membership in S

By induction, this proves that (∀n ∈ N)(b

n

−1 ≥ n(b −1)).

Alternatively, we could prove this using the same identity that Rudin used in the proof of 1.21. From the

distributive property we can verify that b

n

−a

n

= (b −a)(b

n−1

a

0

+b

n−2

a

1

+. . . +b

0

a

n−1

). So when a = 1, this

becomes b

n

−1 = (b −1)(b

n−1

+b

n−2

+. . . +b

0

). And since b > 1, each term in the b

n−k

series is greater than

1, so b

n

−1 ≥ (b −1)(1

n−1

+ 1

n−2

+. . . + 1

0

) = (b −1)n.

Exercise 1.7 b

→ n(b

1

n

−1) = n(b

1

n

−1)

→ n(b

1

n

−1) = (1 + 1 +. . . + 1)

. ¸¸ .

n times

(b

1

n

−1)

→ n(b

1

n

−1) = (1

n−1

+ 1

n−2

+. . . + 1

0

)(b

1

n

−1)

It can be shown that b

k

> 1 when b > 1, k > 0 (see theorem S4). Replacing 1 with b

1

n

gives us the

inequality:

→ n(b

1

n

−1) ≤ ((b

1

n

)

n−1

+ (b

1

n

)

n−2

+. . . + (b

1

n

)

0

)(b

1

n

−1)

Now we can use the algebraic identity b

n

−a

n

= (b

n−1

a

0

+b

n−2

a

1

+. . . +b

0

a

n−1

)(b −a):

→ n(b

1

n

−1) ≤ ((b

1

n

)

n

−1)

→ n(b

1

n

−1) ≤ (b −1)

Exercise 1.7 c

→ n > (b −1)/(t −1) assumed

→ n(t −1) > (b −1) this holds because n, t, and b are greater than 1

→ n(t −1) > (b −1) ≥ n(b

1

n

−1) from part b

→ n(t −1) > n(b

1

n

−1) transitivity of order relations

→ (t −1) > (b

1

n

−1) n > 0 → n

−1

> 0 would be a trivial proof

→ t > b

1

n

5

Exercise 1.7 d

We’re told that b

w

< y, which means that 1 < yb

−w

. Using the substitution yb

−w

= t with part (c), we’re lead

directly to the conclusion that we can select n such that yb

−w

> b

1

n

. From this we get y > b

w+

1

n

, which is what

we were asked to prove. As a corollary, the fact that b

1

n

> 1 means that b

w+

1

n

> b

w

.

Exercise 1.7 e

We’re told that b

w

> y, which means that b

w

y

−1

> 1. Using the substitution b

w

y

−1

= t with part (c), we’re

lead directly to the conclusion that we can select n such that b

w

y

−1

> b

1

n

. Multiplying both sides by y gives us

b

w

> b

1

n

y. Multiplying this by b

−1

n

gives us b

w−

1

n

> y, which is what we were asked to prove. As a corollary,

the fact that b

1

n

> 1 > 0 means that, upon taking the reciprocals, we have b

−1

n

< 1 and therefore b

w−

1

n

< b

w

.

Exercise 1.7 f

We’ll prove that b

x

= y by showing that the assumptions b

x

> y and b

x

< y lead to contradictions.

If b

x

> y, then from part (e) we can choose n such that b

x

> b

x−

1

n

> y. From this we see that x −

1

n

is an

upper bound of A that is smaller than x. This is a contradiction, since we’ve assumed that x =sup(A).

If b

x

< y, then from part (d) we can choose n such that y > b

x+

1

n

> b

x

. From this we see that x is not an

upper bound of A. This is a contradiction, since we’ve assumed that x is the least upper bound of A.

Having ruled out these two possibilities, the trichotomy property of ordered ﬁelds forces us to conclude that

b

x

= y.

Exercise 1.7 g

Assume that there are two elements such that b

w

= y and b

x

= y. Then by the transitivity of equality relations,

b

w

= b

y

, although this seems suspiciously simple.

Exercise 1.8

In any ordered set, all elements of the set must be comparable (the trichotomy rule, deﬁnition 1.5). We will

show by contradiction that (0, 1) is not comparable to (0, 0) in any potential ordered ﬁeld containing C. First,

we assume that (0, 1) > (0, 0) :

→ (0, 1) > (0, 0) hypothesis of contradiction

→ (0, 1)(0, 1) > (0, 0) deﬁnition 1.17(ii) of ordered ﬁelds

We assumed here that (0, 0) can take the role of 0 in deﬁnition 1.17 of an ordered ﬁeld. This is

a safe assumption because the uniqueness property of the additive identity shows us immediately

that (0, 0) + (a, b) = (a, b) → (0, 0) = 0.

→ (−1, 0) > (0, 0) deﬁnition of complex multiplication

→ (−1, 0)(0, 1) > (0, 0) deﬁnition 1.17(ii) of ordered ﬁelds, since we initially assumed (0, 1) > 0

→ (0, −1) > (0, 0) deﬁnition of complex multiplication

It might seem that we have established our contradiction as soon as we concluded that (−1, 0) > 0

or (0, −1) > 0. However, we’re trying to show that the complex ﬁeld cannot be an ordered ﬁeld

under any ordered relation, even a bizarre one in which −1 > −i > 0. However, we’ve shown that

(0, 1) and (0, −1) are both greater than zero. Therefore:

→ (0, −1) + (0, 1) > (0, 0) + (0, 1) deﬁnition 1.17(i) of ordered ﬁelds

→ (0, 0) > (0, 1) deﬁnition of complex multiplication

This conclusion is in contradiction of trichotomy, since we initially assumed that (0, 0) < (0, 1). Next, we

assume that (0, 1) < (0, 0):

6

→ (0, 0) > (0, 1) hypothesis of contradiction

→ (0, 0) + (0, −1) > (0, 1) + (0, −1) deﬁnition 1.17(i) of ordered ﬁelds

→ (0, −1) > (0, 0) deﬁnition of complex addition

→ (0, −1)(0, −1) > (0, 0) deﬁnition 1.17(ii) of ordered ﬁelds

→ (−1, 0) > (0, 0) deﬁnition of complex multiplication

→ (−1, 0)(0, −1) > (0, 0) deﬁnitino 1.17(ii) of ordered ﬁelds, since we’ve established (0, −1) >

(0, 0)

→ (0, 1) > (0, 0) deﬁnition of complex multiplication

Once again trichotomy has been violated.

Proof by contradiction that (0, 1) ,= (0, 0): if we assume that (0, 1) = (0, 0) we’re led to the conclusion that

(a, b) = (0, 0) for every complex number, since (a, b) = a(0, 1)

4

+ b(0, 1) = a(0, 0) + b(0, 0) = (0, 0). By the

transitivity of equivalence relations, this would mean that every element is equal to every other. And this is

in contradiction of deﬁnition 1.12 of a ﬁeld, where we’re told that there are at least two distinct elements: the

additive identity (’0’) and the multiplicative identity (’1’).

Exercise 1.9a

To prove that this relation turns C into an ordered set, we need to show that it satisﬁes the two requirements

in deﬁnition 1.5. Proof of transitivity:

→ (a, b) < (c, d) ∧ (c, d) < (e, f) assumption

→ [a < c ∨ (a = c ∧ b < d)] ∧ [c < e ∨ (c = e ∧ d < f)] deﬁnition of this order relation

→ (a < c ∧ c < e) ∨ (a < c ∧ c = e ∧ d < f)

∨(a = c ∧ b < d ∧ c < e) ∨ (a = c ∧ b < d ∧ c = e ∧ d < f) distributivity of logical operators

→ (a < e) ∨ (a < e ∧ d < f) ∨ (a < e ∧ b < d) ∨ (a = e ∧ b < f) transitivity of order relation on R

Although we’re falling back on the the transitivity of an order relation, we are not assuming what

we’re trying to prove. We’re trying to prove the transitivity of the dictionary order relation on C,

and this relation is deﬁned in terms of the standard order relation on R. This last step is using the

transitivity of this standard order relation on R and is not assuming that transitivity holds for the

dictionary order relation.

→ (a < e) ∨ (a < e) ∨ (a < e) ∨ (a = e ∧ b < f) p ∧ q → p

→ a < e ∨ (a = e ∧ b < f) p ∨ p → p

→ (a, b) < (e, f) deﬁnition of this order relation

To prove that the trichotomy property holds for the dictionary relation on Q, we rely on the trichotomy

property of the underlying standard order relation on R. Let (a, b) and (c, d) be two elements in C. From the

standard order relation, we know that

→ (a, b) ∈ C ∧ (c, d) ∈ C assumed

→ a, b, c, d ∈ R deﬁnition of a complex number

→ (a < c) ∨ (a > c) ∨ (a = c) trichotomy of the order relation on R

→ (a < c) ∨ (a > c) ∨ (a = c ∧ (b < d ∨ b > d ∨ b = d)) trichotomy of the order relation on R

→ (a < c) ∨ (a > c) ∨ (a = c ∧ b < d) ∨ (a = c ∧ b > d)

∨(a = c ∧ b = d) distributivity of the logical operators

→ (a, b) < (c, d) ∨ (a, b) > (c, d) ∨ (a, b) < (c, d) ∨ (a, b) > (c, d)

∨(a, b) = (c, d) deﬁnition of the dictionary order relation

→ (a, b) < (c, d) ∨ (a, b) > (c, d) ∨ (a, b) = (c, d)

And this is the deﬁnition of the trichotomy law, so we have proven that the dictionary order turns the

7

complex numbers into an ordered set.

Exercise 1.9b

C does not have the least upper bound property under the dictionary order. Let E = ¦(0, a) : a ∈ R¦. This subset

is just the imaginary axis in the complex plane. This subset clearly has an upper bound, since (x, 0) > (0, a) for

any x > 0. But it does not have a least upper bound: for any proposed upper bound (x, y) with x > 0, we see

that

(x, y) < (

x

2

, y) < (0, a)

So that (

x

2

, y) is an upper bound less than our proposed least upper bound, which is a contradiction.

Exercise 1.10

This is just straightforward algebra, and is too tedious to write out.

Exercise 1.11

If we choose w =

z

|z|

and choose r = [z[, then we can easily verify that [w[ = 1 and that rw = [z[

z

|z|

= z.

Exercise 1.12

Set a

i

√

z

i

and b

i

=

√

¯ z

i

and use the Cauchy-Schwarz inequality (theorem 1.35). This gives us

¸

¸

¸

¸

¸

¸

n

j=1

√

z

j

_

¯ z

j

¸

¸

¸

¸

¸

¸

2

≤

n

j=1

[

√

z

j

[

2

n

j=1

[

_

¯ z

j

[

2

which is equivalent to

[z

1

+z

2

+. . . +z

n

[

2

≤ ([z

1

[ +[z

2

[ +. . . +[z

n

[)

2

Taking the square root of each side shows that

[z

1

+z

2

+. . . +z

n

[ ≤ [z

1

[ +[z

2

[ +. . . +[z

n

[

which is what we were asked to prove.

Exercise 1.13

[x −y[

2

= (x −y)(x −y)

= (x −y)(¯ x − ¯ y) Theorem 1.31(a)

= x¯ x −x¯ y −y¯ x +y¯ y

= x¯ x −(x¯ y +y¯ x) +y¯ y

= [x[

2

−2Re(x¯ y) +[y[

2

Theorem 1.31(c), deﬁnition 1.32

≥ [x[

2

−2[Re(x¯ y)[ +[y[

2

x ≤ [x[, so −[x[ ≥ [x[.

≥ [x[

2

−2[x¯ y[ +[y[

2

Theorem 1.33(d)

= [x[

2

−2[x[[¯ y[ +[y[

2

Theorem 1.33(c)

= [x[

2

−2[x[[y[ +[y[

2

Theorem 1.33(b)

= ([x[ −[y[)([x[ −[y[)

= ([x[ −[y[)([¯ x[ −[¯ y[) Theorem 1.33(b)

= ([x[ −[y[)([x[ −[y[) Theorem 1.31(a)

= [[x[ −[y[[

2

This chain of inequalities shows us that [[x[ − [y[[

2

≤ [x − y[

2

. Taking the square root of both sides results

in the claim we wanted to prove.

8

Exercise 1.14

[1 +z[

2

+[1 −z[

2

= (1 +z)(1 +z) + (1 −z)(1 −z)

= (1 +z)(

¯

1 + ¯ z) + (1 −z)(

¯

1 − ¯ z) Theorem 1.31(a)

= (1 +z)(1 + ¯ z) + (1 −z)(1 − ¯ z) The conjugate of 1 = 1 + 0i is just 1 −0i = 1.

= (1 + ¯ z +z +z¯ z) + (1 − ¯ z −z +z¯ z)

= (2 + 2z¯ z)

= (2 + 2) We are told that z¯ z = 1

= 4

Exercise 1.15

Using the logic and the notation from Rudin’s proof of theorem 1.35, we see that equality holds in the Schwarz

inequality when AB = [C[

2

. This occurs when have B(AB − [C[

2

) = 0, and from the given chain of equalities

we see that this occurs when

[Ba

j

− Cb

j

[

2

= 0. For this to occur we must have Ba

j

= Cb

j

for all j, which

occurs only when B = 0 or when

a

j

=

C

B

b

j

for all j

That is, each a

j

must be a constant multiple of b

j

.

Exercise 1.16

We know that [z −x[

2

= [z −y[

2

= r

2

. Expanding these terms out, we have

[z −x[

2

= (z −x) (z −x) = [z[

2

−2z x +[x[

2

[z −y[

2

= (z −y) (z −y) = [z[

2

−2z y +[y[

2

For these to be equal, we must have

−2z x +[x[

2

= −2z y +[y[

2

which happens when

z (x −y) =

1

2

[[x[

2

−[y[

2

] (1)

We also want [z − x[ = r, which occurs when z = x + rˆ u where [ˆ u[ = 1. Using this representation of z, the

requirement that r

2

= [z −y[

2

becomes

r

2

= [z −y[

2

= [x +rˆ u −y[

2

= [(x −y) +rˆ u

2

[ = [x −y[

2

+ 2rˆ u (x −y) +[rˆ u[

2

= d

2

+ 2rˆ u (x −y) +r

2

Rearranging some terms, this becomes

ˆ u (y −x) =

d

2

2r

=

d

2r

[y −x[ (2)

A quick, convincing, and informal proof would be to appeal to the relationship a b = [a[[b[ cos(θ) where θ is the

angle between the two vectors; the previous equation then becomes

[ˆ u[[y −x[ cos(θ) =

d

2r

[y −x[

Dividing by [y −x[ and remembering that ˆ u = 1, this becomes

cos(θ) =

d

2r

where θ is the angle between the ﬁxed vector (y −x) and the variable vector ˆ u. It’s easy to see that this equation

will hold for exactly one ˆ u when d = 2r; it will hold for no ˆ u when d > 2r; it will hold for two values of ˆ u when

d < 2r and n = 2; and it will hold for inﬁnitely many values of ˆ u when d < 2r and n > 2. Each value of ˆ u

corresponds with a unique value of z. More formal proofs follow.

9

part (a)

When d < 2r, equation (2) is satisﬁed for all ˆ u for which

ˆ u (y −x) =

d

2r

[y −x[ < [y −x[

By the deﬁnition of the dot product, this is equivalent to

u

1

(y

1

+x

1

) +u

2

(y

2

+x

2

) +. . . +u

k

(y

k

+x

k

) =

d

2r

[y −x[ (3)

The only other requirement for the values of u

i

is that

_

u

2

1

+u

2

2

+. . . +u

2

k

= 1 (4)

This gives us a system of two equations with k variables. As long as k ≥ 3 we have more variables than equations

and therefore the system will have inﬁnitely many solutions.

part (b)

Evaluating d

2

, we have:

d

2

= [x −y[

2

= [(x −z) + (z −y)[

2

= [(x −z) + (z −y)] [(x −z) + (z −y)] deﬁnition of inner product

= (x −z) (x −z) + 2(x −z) (z −y) + (z −y) (z −y) inner products are distributive

= [x −z[

2

+ 2(x −z) (z −y) +[z −y[

2

deﬁnition of inner product

Evaluating (2r)

2

, we have:

(2r)

2

= (r +r)

2

= ([z −x[ +[z −y[)

2

= [z −x[

2

+ 2[z −x[[z −y[ +[z −y[

2

commutativity of multiplication

If 2r = d then d

2

= (2r)

2

and therefore, by the above evaluations, we have

[x −z[

2

+ 2(x −z) (z −y) +[z −y[

2

= [z −x[

2

+ 2[z −x[[z −y[ +[z −y[

2

which occurs if and only if

2(x −z) (z −y) = 2[x −z[[z −y[

From exercise 14 we saw that this equality held only if (x − z) = c(z − y) for some constant c; we know that

[x −z[ = [z −y[ so c = ±1; we know x ,= y so c = 1. Therefore we have x −z = z −y, from which we have

z =

x +y

2

and there is clearly only one such z that satisﬁes this equation.

part (c)

If 2r < d then we have

[x −y[ > [x −z[ +[z −y[

which is equivalent to

[(x −z) + (z −y)[ > [x −z[ +[z −y[

which violates the triangle inequality (1.37e) and is therefore false for all z.

10

Exercise 1.17

First, we need to prove that a (b +c) = a b +a c and that (a +b) c = a c +b c: that is, we need to prove

that the distributive property holds between the inner product operation and addition.

a (b +c) =

a

i

(b

i

+c

i

) deﬁnition 1.36 of inner product

=

(a

i

b

i

+a

i

c

i

) distributive property of R

=

a

i

b

i

+

a

i

c

i

associative property of R

= a b +a c deﬁnition 1.36 of inner product

(a +b) c =

(a

i

+b

i

)c

i

deﬁnition 1.36 of inner product

=

(a

i

c

i

+b

i

c

i

) distributive property of R

=

a

i

c

i

+

b

i

c

i

associative property of R

= a c +b c deﬁnition 1.36 of inner product

The rest of the proof follows directly:

[x +y[

2

+[x −y[

2

= (x +y) (x +y) + (x −y) (x −y)

= (x +y) x + (x +y) y + (x −y) x −(x −y) y

= x x +y x +x y +y y +x x −y x −x y +y y

= x x +y y +x x +y y

= 2[x[

2

+ 2[y[

2

Exercise 1.18

If x = 0, then x y = 0 for any y ∈ R

k

. If x ,= 0, then at least one of the elements x

1

, . . . , x

k

must be nonzero:

let this element be represented by x

a

. Let x

b

represent any other element of x. Choose y such that:

y

i

=

_

¸

_

¸

_

x

b

x

a

i = a

−1 i = b

0 otherwise

We can now see that xy = x

a

x

b

xa

+x

b

(−1) = x

b

−x

b

= 0. We began with an arbitrary vector x and demonstrated

a method of construction for y such that x y = 0: therefore, we can always ﬁnd a nonzero y such that x y = 0.

This is not true in R

1

because the nonzero elements of R are closed with respect to multiplication.

Exercise 1.19

We need to determine the circumstances under which [x −a[ = 2[x −b[ and [x −c[ = r. To do this, we need to

manipulate these equalities until they have a common term that we can use to compare them.

[x −a[ = 2[x −b[

[x −a[

2

= 4[x −b[

2

[x[

2

−2x a +[a[

2

= 4[x[

2

−8x b + 4[b[

2

3[x[

2

= [a[

2

−2x a + 8x b −4[b[

2

[x −c[ = r

[x −c[

2

= r

2

[x[

2

−2x c +[c[

2

= r

2

[x[

2

= r

2

+ 2x c −[c[

2

3[x[

2

= 3r

2

+ 6x c −3[c[

2

Combining these last two equalities together, we have

11

→ [a[

2

−2x a + 8x b −4[b[

2

= 3r

2

+ 6x c −3[c[

2

→ [a[

2

−2x a + 8x b −4[b[

2

−3r

2

−6x c + 3[c[

2

= 0

→ [a[

2

−4[b[

2

+ 3[c[

2

−2x (a −4b −3c) −3r

2

= 0

We can zero out the dot product in this equation by letting 3c = 4b − a. Of course, this also

determines a speciﬁc value of c. This substitution gives us the new equality:

→ [a[

2

−4[b[

2

+ 3

¸

¸

¸

¸

4b

3

−

a

3

¸

¸

¸

¸

2

−3r

2

= 0

→ [a[

2

−4[b[

2

+

3 16

9

[b[

2

−

3 8

9

a b +

3

9

[a[

2

−3r

2

= 0

→ 3[a[

2

−12[b[

2

+ 16[b[

2

−8a b +[a[

2

−9r

2

= 0

→ 4[a[

2

4[b[

2

−8a b −9r

2

= 0

→ 4[a −b[

2

= 9r

2

→ 2[a −b[ = 3r

By choosing 3c = 4b −a and 3r = 2[a −b[, we guarantee that [x −a[ = 2[x −b[ iﬀ [x −c[ = r.

Exercise 1.20

We’re trying to show that R has the least upper bound property. The elements of R are certain subsets of Q,

and < is deﬁned to be “is a proper subset of”. To say that an element α ∈ R has a least upper bound, then,

is to say that the subset α has some “smallest” superset β such that α ⊂ β. We’re asked to omit property III,

which told us that each α ∈ R has no largest element. With this restriction lifted, we have a new deﬁnition of

“cut” that includes cuts such as (−∞, 1] and (−∞,

√

2].

To prove that each subset of R with an upper bound must have a least upper bound, we will follow Step 3

in the book almost exactly. We will deﬁne two subsets A ⊂ R and γ ⊂ R, show that the subset γ is also an

element of R, and then show that the subset/element γ is the least upper bound of A.

Let A be a nonempty subset of R with an upper bound of β (Note: A is a subset of R, not a subset of R.

The elements of A are cuts, which are subsets of Q). Deﬁne γ to be the union of all cuts in A. This means that

γ contains every element from every cut in A: γ consists of elements from Q.

proof that γ is a cut:

The proof of criterion (I) has two parts. (i) γ is the union of elements of A, and we were told that A is nonempty.

Therefore γ is nonempty. (ii) We are told that β is an upper bound for A, so x ∈ A → x ⊂ β (remember that

< is deﬁned as ⊂). But γ is just the union of cut elements in A, so γ is the union of proper subsets of β. This

means that γ itself is a proper subset of β. This shows that γ ⊂ β ⊆ Q, so γ ,= Q. So criterion (I) in the

deﬁnition of “cut” has been met.

To prove part (II), pick some arbitrary rational p ∈ γ. γ is the set of cuts in A, so p ∈ α for some cut α ∈ A.

Choose a second rational q such that q < p: by the deﬁnition of cut, the fact that p ∈ α and q < p means that

q ∈ α and therefore q ∈ γ. And we’re being asked to disregard part (III), so this is suﬃcient to prove that γ is

a cut.

proof that γ is the least upper bound of A

(i) Choose any arbitrary cut α ∈ A. We’ve deﬁned γ as the union of all cuts in A, so it’s clear that every rational

number in α is also contained in γ: that is, α ⊆ γ. And by the deﬁnition of < for this set, this tells us that

α ≤ γ for every α ∈ A.

(ii) Suppose that δ < γ. By the deﬁnition of < for this set, this means that δ ⊂ γ. This is a proper subset,

so there must be some rational s such that s ,∈ δ but s ∈ γ. In order for s ∈ γ, it must be the case that s ∈ α

for some cut α ∈ A. We can now show that δ ⊂ α. For every r ∈ δ,it’s also true that s ,∈ δ. By (II), this means

12

that r < s (using standard rational order). And since s ∈ α, (II) also shows that r ∈ α. This shows that δ ⊂ α,

which means δ < α ∈ A (using the subset order on R), and therefore δ is not an upper bound of A.

Together, these two facts show that γ is the least upper bound of A.

deﬁnition of addition

Following the example in the book, we deﬁne the addition of two cuts α+β to be the set of all sums r +s where

r ∈ α and s ∈ β. The book deﬁned 0

∗

to be the set of rational numbers < 0, but by omitting requirement (III)

we are forced to redeﬁne our zero. Let 0

#

be the set of all rational numbers ≤ 0.

The original deﬁnition had to omit 0 from 0

∗

in order to keep R, the set of cuts, closed under addition

(otherwise we’d have 0

∗

+ 0

∗

= (−∞, 0] which is not a cut because of criterion (III)). Our new zero 0

#

must

include 0 as an element because our newly deﬁned cuts can have a greatest element. The set 0

∗

can no longer

function as our zero since (−∞, x] +0

∗

= (−∞, x); these two cuts are not equal ((−∞, x) < (−∞, x]), so 0

∗

has

not functioned as the additive identity.

ﬁeld axioms A1,A2, and A3

The proofs from the book for closure, commutativity, and associativity are directly applicable to our new

deﬁnition of cut.

ﬁeld axiom A4: existence of an additive identity

Let α be a cut in R. Let r and s be rationals such that r ∈ α and s ∈ 0

#

. Then r + s ≤ r, which means that

r +s ∈ α. This shows us that α+0

#

⊆ α. Now let p and r be rationals such that p ∈ α, r ∈ α, and p ≤ r. This

means that p −r ≤ 0, so p −r ∈ 0

#

. Therefore r +(p −r) ∈ α +0

#

, which means that p ∈ α +0

#

. This shows

us that α ⊆ α+0

#

. Having shown that α+0

#

⊆ α and that α ⊆ α+0

#

, we conclude that α = α+0

#

for any

α ∈ R.

ﬁeld axiom A5: existence of additive inverses

The book constructed the deﬁnition of inverse so that the inverse of (−∞, x) would be (−∞, −x). This works

under the original deﬁnition of “cut”, since (−∞, x) + (−∞, −x) = (−∞, 0) = 0

∗

. It is tempting to just deﬁne

the inverse for our new deﬁnition of cut so that the inverse of (−∞, x] is (∞, −x]. This overlooks the fact that

both (−∞, x] and (−∞, x) are cuts under our new deﬁnition: both of these elements must have additive inverses.

To show that A5 is not satisﬁed, we need to demonstrate that at least one element of R has no additive in-

verse. Let α = (−∞, r) for some rational r. Assume that there was some cut β such that α+β = 0

#

= (−∞, 0]:

we will show that this assumption leads to a contradiction.

(i) Assume that β does not contain any elements greater than −r. Let p and q be arbitrary rationals such

that p ∈ α and q ∈ β. From our deﬁnitions of α and β, we know that p < r and q ≤ −r. Combining these two

inequalities tells us that p +q < −r +r, or p +q < 0. But p and q were arbitrary members of α and β, so this

shows that 0 ,∈ α +β. So it cannot be the case that α +β = 0

#

.

(ii) Assume that β contains at least one element q that is greater than −r. Let s represent the diﬀerence

q − (−r): then q = −r + s (note that s must be a positive rational). Let p be an arbitrary rational such that

p ∈ α. By the deﬁnition of α, we know that p < r. And from property (II) of cuts, we know that r −s < p < r.

Adding q = −r +s to each element in this equality gives us 0 < p +q < s. And p +q is an element of α +β, so

we see that α +β contains some positive rational: it cannot be the case that α +β = 0

#

.

Whether or not β contains an element greater than −r, we ﬁnd ourselves in contradiction with the initial

assumption that β might be the inverse of α. Therefore there can be no inverse for α. In conclusion, we see

that omitting (III) forces us to redeﬁne the additive identity, and this new deﬁnition results in the existence of

elements in R that have no inverse.

13

Exercise 2.1

This is immediately justiﬁed by noting that the deﬁnition of subset, x ∈ ∅ → x ∈ A, is satisﬁed for any set A

because of the false antecedent. A more formal proof:

→ (∃x ∈ ∅) deﬁnition of an empty set

→ (∀x)(x ,∈ ∅) negation of the ∃ quantiﬁer

→ (∀x)(x ,∈ A → x ,∈ ∅) Hilbert’s PL1

This previous step is justiﬁed by the argument p → (q → p): if something is true (like x ,∈ ∅), then

everything implies it.

→ (∀x)(x ∈ ∅ → x ∈ A) contrapositive

→ ∅ ⊆ A deﬁnition of subset

Exercise 2.2

The set of integers is countable, so by theorem 2.13 the set of all (k+1)-tuples (a

0

, a

1

, . . . , a

k

) with a

0

,= 0 is also

countable. Let this set be represented by Z

k

. For each a∈ Z

k

consider the polynomial a

0

z

k

+a

1

z

k−1

+. . .+a

k

= 0.

From the fundamental theorem of algebra, we know that there are exactly k complex roots for this polynomial.

We now have a series of nested sets that encompass every possible root for every possible polynomial with

integer coeﬃcients. More speciﬁcally, we have a countable number of Z

k

s, each containing a countable number

of (k + 1)-tuples, each of which corresponds with k roots of a k-degree polynomial. So our set of complex roots

(call it R) is a countable union of countable unions of ﬁnite sets. This only tells us that R is at most countable:

it is either countable or ﬁnite.

To show that R is not ﬁnite, consider the roots for 2-tuples in Z

1

. Each 2-tuple of the form (−1, n) corre-

sponds with the polynomial −z + n = 0 whose solution is z = n. There is clearly a unique solution for each

n ∈ Z, so R is an inﬁnite set. Because R is also at most countable, this proves that R is countable.

I did not use the hint provided in the text, which either means that this proof is invalid or that there is an

alternate (simpler?) proof.

Exercise 2.3

The set of real numbers is uncountable, so if every real number were algebraic then the set of real algebraic

numbers would be uncountable. However, exercise 2.2 showed that the algebraic complex numbers were count-

able. The real numbers are a subset of the complex numbers, so the set of algebraic real numbers is at most

countable.

Exercise 2.4

The rational real numbers are countable. If the irrational real numbers were countable, then the union of rational

and irrational reals would be countable. But this union is just R, which is not countable. So the irrational reals

are not countable.

Exercise 2.x: The Cantor Set

The idea behind the Cantor set is that each term in the series E

1

⊃ E

2

. . . ⊃ E

n

removes the middle third of

each remaining line segment, so that you get a series of increasingly smaller segments that look like:

14

We can see that E

0

has one segment of length 1, E

1

has two segments of length 1/3, and E

m

has 2

m

segments

of length 1/3

m

. Notice that for any possible segment (α, β), we can choose m to be large enough so that the

maximum segment length is less than the length of (α, β) and therefore (α, β) ,∈ E

m

. And if the segment is not

in E

m

for some m, it’s not in the union P. This is the reasoning behind Rudin’s statement that “P contains no

segment”.

To justify the claim that no point in P is an isolated point, we need to show that for every point p

1

∈ P

and any arbitrarily small radius r, we can ﬁnd some second point p

2

such that d(p

1

, p

2

) < r. P is deﬁned to

be

E

i

, so if p

1

∈ P then it is a member of every E

i

. Choose some element E

m

where m is so large that the

segments in E

m

are all shorter than r. We are assured that this is possible by the Archimedean property of the

rationals, since we need only ﬁnd m such that 3

−m

< r.

So we’ve found an E

m

with a line segment containing p

1

. Let p

2

represent one of the endpoints of this line.

The length of this segment is less than r, so d(p

1

, p

2

) < r. In each subsequent term in the series ¦E

i

¦, the

endpoint p

2

will never be “cut oﬀ”: it will always be the endpoint of a segment from which the middle third is

lost. So p

2

∈ P. And since this was true for an arbitrary point p

1

∈ P and an arbitrarily small r, this shows

that every neighborhood of every point in P is a limit point and not an isolated point.

Exercise 2.5

Let E be a subset of (0, 1] consisting only all rational numbers of the form

1

m

, m ∈ N. No point in E is a limit

point, but the limit points of E need not be members of E: we will demonstrate that the point 0 is a limit point

of E.

(i) Proof that no point in E is a limit point: let p

m

be an arbitrary member of E. Then p

m

=

1

m

for some

m ∈ N. The closest point to p

m

is therefore p

m+1

=

1

m+1

. So we just choose r such that r =

1

2

d(p

m

, p

m+1

) and

we see that there are no other points in this neighborhood of p

m

, which shows that p

m

is an isolated point.

(ii) Proof that 0 is a limit point: choose an arbitrarily small radius r. From the Archimedian property, we

can ﬁnd some m ∈ N such that

1

m

< r. This p

m

=

1

m

is a member of E, and d(0, p

m

) =

1

m

< r. This shows that

0 is a limit point.

(iii) Proof that no other points are limit points: let x be some point that is neither zero nor a member

of E. This point must be either less than zero, greater than 1, or it must lie between two sequential points

p

m

, p

m+1

∈ E. Let the d

x

represent the smallest distance from the set ¦d(x, 0), d(x, 1), d(x, p

m

), d(x, p

m+1

)¦.

Choose r such that r =

1

2

d

x

. There can be no points of E in the neighborhood N

r

(x). If there were, then this

would indicate the existence of a point of E less than zero, greater than 1, or between

1

m

and

1

m+1

. None of

these can exist in E as we deﬁned it, so no points other than 0 are limit points.

Now let F be the subset of (2, 3] containing rational numbers of the form 2 +

1

m

and let G be the subset of

(4, 5] consisting of rational numbers of the form 4 +

1

m

. Just as E was shown to have a limit point of only zero,

these sets can be shown to have limit points of only 2 and 4. Therefore the union E ∪ F ∪ G is bounded and

has three limit points ¦0, 2, 4¦.

Exercise 2.6

(i) We’re asked to prove that E

is closed. This is equivalent to proving that every limit point of E

is a point

of E. And by the deﬁnition of E

, this is equivalent to proving that every limit point of E

is a limit point of E.

To prove this, let x represent an arbitrary limit point of E

**. Choose any arbitrarily small r and let
**

s =

r

2

, t =

r

4

1

. Since x is a limit point of E

, we can ﬁnd a point y ∈ E

in the neighborhood N

s

(x). And

y, by virtue of being in E

**, is a limit point of E: so we can ﬁnd a point z ∈ E in the neighborhood N
**

t

(y).

The distance d(x, y) is less than s and d(y, z) is less than t, so deﬁnition 2.15 of metric spaces assures us

that d(x, z) ≤ d(x, y) +d(y, z) < s +t < r, so d(x, z) < r and therefore z is in the neighborhood N

r

(x) for any

1

t was chosen to be less than r to guarantee x = z.

15

arbitrarily small r. So the point x is a limit point for E. But x was an arbitrarily chosen limit point of E

, so

we have proven that every limit point of E

**is a limit point of E, which is what we were asked to prove. The
**

following image is helpful for imagining these points in R

2

.

(ii) We can use a similar technique to show that every limit point of

¯

E is a limit point E. Let x represent

an arbitrary limit point of

¯

E (which we won’t assume to be a member of

¯

E). Choose any arbitrarily small r

and let s =

r

2

, t =

r

4

. Since x is a limit point of

¯

E, theorem 2.20 tells us that the neighborhood N

s

(x) contains

inﬁnitely many points of

¯

E. Because

¯

E is deﬁned to be E

**∪ E, each of these inﬁnitely many points in N
**

s

(x) is

either a member of E

or a member of E.

(ii.a) First, assume that there exists at least one point y in N

s

(x) such that y ∈ E

. By deﬁnition, this

means that y is a limit point of E, so there is at least one point z ∈ E in the neighborhood N

t

(y). The

distance d(x, y) is less than s and d(y, z) is less than t, so deﬁnition 2.15 of metric spaces assures us that

d(x, z) ≤ d(x, y) +d(y, z) < s +t < r, so d(x, z) < r and therefore we can ﬁnd some z ∈ E in the neighborhood

N

r

(x) for any arbitrarily small r. So x is a limit point for E.

(ii.b) Next, assume that none of the points in N

s

(x) are members of E. This means that all of the in-

ﬁnitely many points of E

∪ E in N

s

(x) are members of E. From this, we see that every neighborhood of x

contains elements of E. For any neighborhood N

t

(x) with t < s, the fact that x is a limit point of

¯

E and

that N

t

(x) ⊂ N

s

(x) means that N

t

(x) contains inﬁnitely many points in E. But this means that x is a limit

point of E. For any neighborhood N

t

(x) with t > s, the fact that N

s

(x) ⊂ N

t

(x) means that N

t

(x) contains

inﬁnitely many points in E. And since every neighborhood of x contains an element of E, x is a limit point for E.

The second part of the proof is to show that every limit point of E is a limit point of

¯

E and is relatively trivial.

Every element of E is also an element of E ∪ E

=

¯

E. If x is a limit point of E, then every neighborhood of x

contains an element of E. Therefore every neighborhood of x contains an element of

¯

E. So x is a limit point of

¯

E.

We’ve shown that every limit point of

¯

E is a limit point of E and vice-versa, which is what we were asked

to prove.

(iii) We’re asked if E and E

**always have the same limit points. The answer is “no”, and a counterexample
**

can be found in the previous question. The limit points of E in exercise 2.5 were E

= ¦0, 2, 4¦. And E

clearly

has no limit points whatsoever.

Exercise 2.7a

We’ll prove the equality of these sets by showing that each is a subset of the other.

(A) Assume that x ∈

¯

B

n

. Then, x ∈ B

n

or x is a limit point of B

n

.

(A1) If x ∈ B

n

, then x ∈ A

k

for some A

k

∈

A

i

. And if x ∈ A

k

, then x ∈

¯

A

k

. And this means that x ∈

¯

A

i

.

(A2) If x is a limit point of B

n

, then it must be a limit point at least one speciﬁc A

k

∈

A

i

. Proof by

contradiction: assume that x is not a limit point for any A

k

∈

A

i

. Then for each i, there is some neighborhood

N

ri

(x) that contains no elements of A

i

. Let s =min(r

i

) (which exists because i is ﬁnite). Then N

s

(x) contains

16

no elements from any A

i

since N

s

(x) ⊆ N

ri

(x) for each i. And if this neighborhood contains no points for

any A

i

, then it contains no points of B

n

=

A

i

. And this is a contradiction, since x is a limit point of B

n

.

Therefore, by contradiction, x must be a limit point of at least one speciﬁc A

k

∈

A

i

. So x ∈

¯

A

k

, which means

that x ∈

¯

A

i

.

In either of these two cases, x ∈

¯

A

i

. This proves that x ∈

¯

B

n

→ x ∈

¯

A

i

.

(B) Assume that x ∈

¯

A

i

. Then x ∈

¯

A

k

for some k. And this means that either x ∈ A

k

or x is a limit

point of A

k

.

(B1) If x ∈ A

k

, then x ∈

A

i

, which means that x ∈ B

n

and therefore x ∈

¯

B

N

.

(B2) If x is a limit point of A

k

, then every neighborhood of x contain an element of A

k

. But every element

of A

k

is an element of

A

i

= B

n

, so this means that every neighborhood of x contains an element of B

n

. By

deﬁnition, this means that x is a limit point of B

n

, so that x ∈

¯

B

n

.

In either of these two cases, x ∈

¯

B

n

. This proves that x ∈

¯

A

i

→ x ∈

¯

B

n

. And we’ve already shown that

x ∈

¯

B

n

→ x ∈

¯

A

i

, so we have proven that

¯

B

n

=

¯

A

i

.

Exercise 2.7b

Nothing in part B of the above proof required that i be ﬁnite, so it’s still the case that

∞

i=1

¯

A

i

⊂

¯

B. But part

A of the proof assumed that we could choose a least element from the set of size i, which we can’t do for an

inﬁnite set. So we can’t assume that B ⊆

¯

A

i

. And it’s a good thing that we can’t assume this, because it’s false.

Consider the set from exercise 2.5, the set E consisting of all rational numbers of the form

1

m

, m ∈ N. We’ll

construct this set by deﬁning A

i

=

1

i

and deﬁning B =

∞

i=1

A

i

. We saw in exercise 2.5 that 0 ∈

¯

B. But

0 ,∈

¯

A

i

for any i, so 0 ,∈

∞

i=1

¯

A

i

. This shows us that

¯

B ,=

∞

i=1

A

i

for these sets. And we’ve already shown that

¯

A

i

⊆ B, so for this particular choice of sets we see that

¯

A

i

⊂ B.

Exercise 2.8a

Every point in an open set E in R

2

is a limit point of E. Proof: let p be an arbitrary point in E. We can

say two things about the neighborhoods of p. First, since we’re dealing with R

2

, every neighborhood of p con-

tains inﬁnitely many points. Second, p is an interior point (since E is open) and so there is some r such that

N

r

(p) ⊂ E. Together, these two facts tell us that N

r

(p) contains inﬁnitely many points of E.

We now show that this guarantees that every neighborhood of p contains a point of E. Choose any s such

that r < s. Because the neighborhood N

r

(p) contains some q ∈ E and N

r

(p) ⊂ N

s

(p), we know that q ∈ N

s

(p).

Now choose s such that s < r. N

s

(p) contains inﬁnitely many points, and N

s

(p) ⊂ N

r

(p) ⊂ E, so N

s

(p) contains

inﬁnitely many points of E. We’ve shown that whether r < s or r > s for any arbitrary r, N

s

(p) contains a

point of E. Thus p is a limit point of E.

Exercise 2.8b

Consider a closed set consisting of a single point, such as E = ¦(0, 0)¦. This point is clearly not a limit point of

E.

Exercise 2.9a

To prove that E

◦

is open, we will show that every point in E

◦

must be an interior point of E

◦

.

Let x be an arbitrary point in E

◦

. By the deﬁnition of E

◦

, we know that x is an interior point of E. This

means that we can choose some r such that N

r

(x) ⊂ E: i.e., every point of N

r

(x) is a point of E. To show that

x is an interior point of E

◦

, though, we will need to show that every point in the neighborhood N

r

(x) is also a

17

point of E

◦

.

Let y be an arbitrary point in N

r

(x). Clearly, 0 < d(x, y) < r. Choose a radius s such that d(x, y) + s = r

and let z be an arbitrary point in N

s

(y). Every point in N

s

(y) is a point in E, since d(x, z) ≤ d(x, y) +d(y, z) <

d(x, y) +s = r implies d(x, z) < r, and we’ve established that every point in N

r

(x) is a member of E. And since

every point in N

s

(y) is a point in E, by deﬁnition we know that y is an interior point of E. But y was an arbitrary

point in N

r

(x), so we know that every point in N

r

(x) is an interior point of E. And this means that x is itself

an interior point of the set of interior points of E: that is, x is an interior point of E

◦

. And since x was an ar-

bitrary point in E

◦

, this means that every point in E

◦

is an interior point of E

◦

. And this proves that E

◦

is open.

Exercise 2.9b

From exercise 2.9(a), we know that E

◦

is always open: so E is open if E = E

◦

. And if E is open, then every

point of E is an interior point and therefore E

◦

= E: so E = E

◦

if E is open. Together, these two statements

show that E is open if and only if E = E

◦

.

Exercise 2.9c

Let p be an arbitrary element of G. Because G is open, p is an interior point of G. This means that for every

point p ∈ G there is some neighborhood N

r

(p) such that N

r

(p) ⊂ G ⊂ E. This chain of subsets tells us that

every point in G is an interior point not only of G, but also of E. And if every point in G is an interior point

of E, then this shows us that G ⊂ E

◦

.

Exercise 2.9d

Proof that

E

◦

⊂

¯

E: let x be a member of

E

◦

, the complement of E

◦

. From the deﬁnition of E

◦

we know that x

is not an interior point of E. From the deﬁnition of “interior point”, this means that every neighborhood of x

contains some element y in the complement of E. For each neighborhood, It must be true that either x = y or

x ,= y. If x = y for one or more of these neighborhoods, then x is in

¯

E (since y is in

¯

E and x = y). If x ,= y for

every neighborhood, then by deﬁnition we know that x is a limit point of

¯

E. So we conclude that either x ∈

¯

E

or x is a limit point for

¯

E: that is, x is a member of

¯

E, the closure of

¯

E.

Proof that

¯

E ⊂

E

◦

: this proof is only a very slight modiﬁcation of the previous one. Let x be a member of

¯

E, the closure of the complement of E. From the deﬁnition of closure, either x is a limit point for

¯

E or x itself

is a member of

¯

E. In either case, every neighborhood of x contains some point of

¯

E. By deﬁnition of “interior

point”, this means that x is not an interior point of E and therefore x ∈

E

◦

, the complement of E

◦

.

Together, these two proofs show us that

E

◦

=

¯

E.

18

Exercise 2.9e

No, E and E do not always have the same interiors. Consider the set E = (−1, 0) ∪ (0, 1) in R

1

. The point 0 is

not a member of the interior of E, but it is a limit point of E. So the point 0 is an interior point of E = (−1, 1).

Exercise 2.9f

No, E and E

◦

do not always have the same closures. Let E be a subset of (0, 1] consisting only all rational

numbers of the form

1

m

, m ∈ N. As we saw in exercise 2.5, the closure of E is E ∪¦0¦. But every neighborhood

of every point in E contains a point not in E, so E has no interior points. So E

◦

= ∅. And the empty set

is already closed, so the closure of E

◦

is ∅. Clearly E ∪ ¦0¦ ,= ∅, so E and E

◦

do not always have the same

closures.

Exercise 2.10

Which sets are closed? Let E be an arbitrary subset of X and let p be an arbitrary point in E. The distance

between any two distinct points is always 1, so by choosing r = .5 we guarantee that the neighborhood N

r

(p)

contains only p itself. Because this is true for any point in E, this shows us that E contains no limit points.

It’s then vacuously true that all of the (nonexistant) limit points of E are points of E: we conclude that E is

closed. But our choice of E was arbitrary, so this shows that every subset of X is closed.

Which sets are open? Let E be an arbitrary set in X. We’ve shown that every set is closed, so it must be

the case that

¯

E is closed: and from theorem 2.23, this means that E is open. But our choice of E was arbitrary,

so this shows that every subset of X is open.

Which sets are compact? Under this metric, subsets of X are compact if and only if they are ﬁnite.

Proof: let E be an arbitrary subset of X. This set is either ﬁnite or inﬁnite.

If E is ﬁnite with cardinality k, then we can always generate a ﬁnite subcover for any open cover. For each

e

i

∈ E, we select some g

i

in our (arbitrary) open cover ¦G

α

¦ such that e

i

∈ g

i

(note that E is ﬁnite, so we don’t

need to use the axiom of choice). This gives us a collection of sets

k

i=1

g

i

such that E ⊆

k

i=1

g

i

⊆ ¦G

α

¦. By

deﬁnition, then,

k

i=1

g

i

is a ﬁnite subcover of E. Our choices for E, k, and ¦G

α

¦ were arbitrary, so this shows

that any ﬁnite subset of X is compact.

If E is inﬁnite, then we can always generate an open cover that has no ﬁnite subcover. As we saw previously,

every subset of X is open: this includes subsets that consist of a single element. So we can create an inﬁnite

open cover of E by letting each set in ¦G

α

¦ be a single element of E. Any subcover of E will need to contain one

element of ¦G

α

¦ for each element of E: and since E is inﬁnite, this means that any subcover must be inﬁnite.

Therefore there is no ﬁnite subcover for this particular open cover, and E is not compact.

Exercise 2.11

d

1

is not a metric: let x = −1, y = 0, z = 1. Then d

1

(x, y) + d

1

(y, z) = 2 while d

1

(x, z) = 4. And this tells us

that d

1

(x, z) > d

1

(x, y) +d

1

(y, z) which violates deﬁnition 2.15(c) of metric spaces.

d

2

is a metric. It’s trivial to verify that d

2

has the properties 2.15(a) and 2.15(b). To show that it has the

property 2.15(c):

→ [x −z[ ≤ [x −y[ +[y −z[ true by the triangle inequality, theorem 1.37(f)

→ [x −z[ ≤ [x −y[ +[y −z[ + 2

_

[x −y[[y −z[ this additional term is > 0

→

_

[x −z[

2

≤

_

_

[x −y[ +

_

[y −z[

_

2

→

_

[x −z[ ≤

_

[x −y[ +

_

[y −z[ legal because all terms are > 0

→ d

2

(x, z) ≤ d

2

(x, y) +d

2

(y, z)

19

d

3

is not a metric. d

3

(1, −1) = 0, which violates deﬁnition 2.15(a) of metric spaces.

d

4

is not a metric. d

4

(2, 1) = 0, which violates deﬁnition 2.15(a) of metric spaces.

d

5

is a metric. It’s trivial to verify that d

2

has the properties 2.15(a) and 2.15(b). To show that it has the

property 2.15(c):

→ [x −z[ ≤ [x −y[ +[y −z[

This ﬁrst step is always true because of the triangle inequality (theorem 1.37(f).

→ [x −z[ ≤ [x −y[ + 2[x −y[[y −z[ +[y −z[ +[x −z[[x −y[[x −z[

Adding positive terms to the right side does not change the sign of the equality.

→ [x −z[ +[x −z[[x −y[ +[x −z[[y −z[ +[x −z[[x −y[[y −z[

≤ [x −y[ + 2[x −y[[y −z[ +[y −z[ +[x −z[[x −y[ + 2[x −z[[x −y[[y −z[ +[x −z[[y −z[

It’s hard to verify from this ugly mess, but we’ve just added the same terms to each side.

→ ([x −z[)(1 +[x −y[ +[y −z[ +[x −y[[y −z[) ≤ (1 +[x −z[)([x −y[ + 2[x −y[[y −z[ +[y −z[)

It’s still hard to verify, but we’ve just factored out terms from each side.

→

[x −z[

1 +[x −z[

≤

[x −y[ + 2[x −y[[y −z[ +[y −z[

1 +[x −y[ +[y −z[ +[x −y[[y −z[

And now we’ve divided each side by two of these factored terms.

→

[x −z[

1 +[x −z[

≤

[x −y[

1 +[x −y[

+

[y −z[

1 +[y −z[

→ d

2

(x, z) ≤ d

2

(x, y) +d

2

(y, z)

The logic behind this seemingly arbitrary series of algebraic steps becomes clear if one starts at the end and

works back to the beginning (which is, of course, how I initially derived the proof).

Exercise 2.12

Let ¦G

α

¦ be any open cover of K. At least one open set in ¦G

α

¦ must contain the point 0: let G

0

be this

set, open in R, that contains 0 as an interior point. As an interior point of G

0

, there is some neighborhood

N

r

(0) ⊂ G

0

. This neighborhood contains every

1

n

∈ K such that

1

n

< r: equivalently, this neighborhood contains

one element for every natural number n such that n >

1

r

. Not only does this mean that G

0

contains an in-

ﬁnite number of elements of K, it means that it contains all but a ﬁnite number (

1

r

, to be exact) of elements of K.

Let G

i

represent an element of ¦G

α

¦ containing

1

i

. We can now see that K ⊂ G

0

∪

1/r

i=1

G

i

, which is a ﬁnite

subcover of ¦G

α

¦.

Exercise 2.13

For i ≥ 2, let A

i

represent the set of all points of the form

1

i

+

1

n

that are contained in the open interval

_

1

i

,

1

i−1

_

.

Each A

i

can be shown to have a single limit point of

1

i

(see exercise 2.5).

20

Now consider the union of sets S =

∞

i=2

A

i

. The same reasoning used in exercise 2.5 shows that the set of

limit points for S is just the collection of limit points from A

i

: that is, the rationals of the form

1

i

for i ≥ 2.

This shows us that S has one limit point for each natural number greater than 1, which is a countable number.

To show that S is compact, we return to the reasoning found in exercise 2.12. Let ¦G

α

¦ be any open cover

of S. For each i ≥ 2, Let G

i

represent the element of ¦G

α

¦ that contains

1

i

. As we saw in exercise 2.12, G

i

contains all but a ﬁnite number of elements from the interval

_

1

i

,

1

i−1

_

. If we let G

ij

represent an element of

¦G

α

¦ containing

1

i

+

1

j

, we see that all of the points of S in the interval

_

1

i

,

1

i−1

_

can be covered by the ﬁnite

union of sets G

i

∪

ri

i=1

.

Now, let G

0

represent the element of ¦G

α

¦ that contains 0 as an interior point. As an interior point, there

is some neighborhood N

r

(0) ⊂ G

0

. This neighborhood contains every limit point of the form

1

n

∈ R such that

1

n

< r. As in 2.12, this means that G

0

contains all but a ﬁnite number of the limit points of S. More than that,

though, it means that G

0

contains all but a ﬁnite number of the intervals of the form (

1

i

,

1

i−1

).

And from these sets we can construct our ﬁnite subcover. Our subcover contains G

0

. For each of the ﬁnitely

many intervals not covered by G

0

, we include the ﬁnite union of sets G

i

∪

ri

i=1

. This is a ﬁnite collection of a

ﬁnite number of elements from ¦G

α

¦, and each element of S is included in one of these elements, so we have

constructed a ﬁnite subcover for an arbitrary cover ¦G

α

¦. This proves that S is compact.

Exercise 2.14

Let G

n

represent the interval

_

1

n

, 1

_

. The union ¦G

α

¦ =

∞

i=1

G

n

is a cover for the interval (0, 1) (Proof: for

any x ∈ (0, 1), there is some n > 1/x and therefore some

1

n

< x. So x ∈ G

n+1

). But there is no ﬁnite subcover

for (0, 1). Let H be a ﬁnite subcover of ¦G

α

¦, and consider the set ¦

1

i+1

: G

i

∈ H¦. This set is ﬁnite, so there

is a least element. And this least element is in the interval (0, 1) but is not a member of any G

i

∈ H. So our

assumption that a ﬁnite subcover exists must be false.

Exercise 2.15

For each i ∈ N, deﬁne A

i

to be:

A

i

=

_

p ∈ Q :

_

2 −

1

i

≤ p ≤

_

2 +

1

i

_

Because the endpoints are irrational, each of these intervals A

i

are both bounded and closed (see exercise 2.16

for the proofs). From theorem 1.20, we also know that each interval is nonempty.

The union of any ﬁnite collection of these intervals will be nonempty, since the union

i∈∇

A

i

is equal to

A

k

, where k is the largest index in ∇. But the inﬁnite union

∞

i=1

A

i

is equal to ¦p ∈ Q :

√

2 ≤ p ≤

√

2¦, which

is obviously empty.

This proves that the collection ¦A

i

: i ∈ N¦ is a counterexample to theorem 2.36 if the word “compact” is

replaced by either “closed” or “bounded”. This same collection works as a counterexample to the corollary of

2.36, since A

n+1

⊂ A

n

.

Exercise 2.16

lemma 1: For any real number x ,∈ Q, the intervals [x, ∞) and (x, ∞) are open in Q. Proof: let x be a real

number such that x ,∈ Q. Let p be a rational number in the interval [x, ∞) or (x, ∞). It cannot be the case

that p = x (since p is rational while x is not), so p is in the interval (x, ∞) (i.e, [x, ∞) is identical to (x, ∞)

in Q). Because x < p, we can rewrite this interval as (p − d(p, x), ∞). Choose r = d(x, p). Then we see that

(p − r, p + r) ⊆ (p − d(p, x), ∞) = (x, ∞). This means that N

r

(p) is an interior point of (x, ∞): but p was an

arbitrary point chosen from [x, ∞), so every point of this interval is an interior point. This proves that [x, ∞)

and (x, ∞) are open in Q.

21

lemma 2: For any real number x ,∈ Q, the intervals (−∞, x] and (−∞, x) are open in Q. Proof: the proof

is nearly identical to that of lemma 1.

lemma 3: All of these open intervals ([x, ∞),(x, ∞),(−∞, x], and (−∞, x)) are also closed. Proof: Choose

any of these four open sets. Note that its complement is also one of these four open sets. Since its complement

is open, the set must be closed.

lemma 4: Every interval of the form (x, y) with x, y ,∈ Q is both open and closed in Q. Proof: choose an

arbitrary interval E = (x, y). It’s complement is

¯

E = (−∞, x] ∪[y, ∞). From lemma 3, we see that

¯

E is a ﬁnite

union of closed sets, so

¯

E is closed (theorem 2.24) and E is open (theorem 2.23). But

¯

E is also a ﬁnite union

of open sets (lemmas 1 and 2), so

¯

E is open (theorem 2.24) and E is closed (theorem 2.23). This proves that

(x, y) is both open and closed. The proofs for [x, y], (x, y], and [x, y) are identical to this one.

E is bounded: If [p[ ≥ ±

√

3, then p

2

> 3 and p ,∈ E. So E is bounded by the interval (−

√

3,

√

3).

E is open and closed in Q: E is the set of all rational numbers p such that 2 < p

2

< 3, which means that

E is the set of rational numbers in (−

√

3, −

√

2) ∪ (

√

2,

√

3). We know that ±

√

2 and ±

√

3 are not rational, so

by lemma 4 we know that E is both a union of closed sets and also a union of open sets. So by theorem 2.24 we

know that E is both open and closed.

E is not compact: Drawing from the example in exercise 2.14, deﬁne the interval A

i

to be:

A

i

=

_

(−

√

3, −

√

2) if i = 1

__

2 +

1

i

,

√

3

_

if i > 1

Because

_

2 +

1

i

is irrational for every i ∈ N, we know by lemma 4 that each A

i

is an open set. Deﬁne an open

cover of E to be

G = ¦A

i

: i ∈ N¦

This has no ﬁnite subcover for the same reason that the set in exercise 2.14 has no inﬁnite subcover: given any

ﬁnite collection of elements of G, we can always ﬁnd an element suﬃciently close to

√

2 that is not contained in

any of those elements of G.

Exercise 2.17

E is not countable: This is proven directly by theorem 2.14.

E is not dense: Let x = .660... and let r = .01. Then N

r

(x) = (.65, .67), and every real number in

this interval contains a number other than 4 or 7. Therefore N

r

(x) contains no points of E. This shows that x

is an element of [0, 1] that is neither a point in E nor a limit point of E, which proves that E is not dense in [0, 1].

E is compact: By theorem 2.41, we can prove that E is compact by showing that it is bounded and closed.

E ⊂ [0, 1], so E is bounded. To show that E is closed, we need to show that every limit point of E is a member

of E.

Proof by contrapositive that every limit point of E is a member of E: let x be an element of [0, 1] such that

x ,∈ E. We will show that x is not a limit point. By the deﬁnition of membership in E, the fact that x ,∈ E

means that

x =

∞

i=1

a

i

10

i

where some a

i

,∈ ¦4, 7¦. Let k represent some index such that a

k

,∈ ¦4, 7¦. Choose r = 10

−(k+1)

. Then the

neighborhood N

r

(x) does not contain any points in E:

i) If a

k+1

= 0, then the k + 1st decimal place will be either 9, 0, or 1 for every element of (x −r, x +r) and so

no element of N

r

(x) is a member of E.

22

ii) If a

k+1

= 9, then the k + 1st decimal place will be either 8, 9, or 0 for every element of (x −r, x +r) and so

no element of N

r

(x) is a member of E.

iii) If a

k+1

is neither 0 nor 9, then the kth decimal place will be be a

k

for every element of (x − r, x + r) and

so no element of N

r

(x) is a member of E.

This proves that there is some neighborhood N

r

(x) that contains no points in E. This proves that x is not

a limit point of E. But x was an arbitrary point such that x ,∈ E, so this proves that (x ,∈ E → x is not a

limit point of E). By contrapositive, this proves that every limit point of E is a member of E. And this, by

deﬁnition, means that E is closed.

E is perfect: We have already shown that E is closed in the course of proving that E is compact. To prove

that E is perfect, we need to show that every point in E is a limit point of E.

Let x be an arbitrary point in E. By the deﬁnition of membership in E, we know that

x =

∞

i=1

a

i

10

i

where each a

i

is either 4 or 7. Deﬁne a second number, x

k

, such that

x

k

=

∞

i=1

b

i

10

i

, where

_

_

_

b

i

= a

i

if i ,= k

b

k

= 4 if a

k

= 7

b

k

= 7 if a

k

= 4

From this deﬁnition, we see that x

k

diﬀers from x only in the kth decimal place, and d(x, x

k

) = 3 10

−k

.

We can use this information to show that x is a limit point of E. Let N

r

(x) be any arbitrary neighborhood

of x. From the archimedian principle, we can ﬁnd some integer k such that k > log

1

0

3

r

. And this is algebraically

equivalent to ﬁnding some integer k such that 3 10

−k

< r. This means that we can ﬁnd some x

k

∈ E (as

deﬁned above) in N

r

(x). And this was an arbitrary neighborhood of an arbitrary point in E, so we have proven

that every neighborhood of every point in E contains a second point in E: by deﬁnition, this means that every

point in E is a limit point.

Exercise 2.18

Section 2.44 describes the Cantor set. The Cantor set is a nonempty perfect subset of R

1

. Each point in the

Cantor set is an endpoint of some segment of the form

_

n

3

k

,

n+1

3

k

¸

with n, k ∈ N, so each point in the Cantor set

is rational.

Let E be the set ¦x +

√

2 : x is in the Cantor set¦ (that is, we’re shifting every element of the Cantor set

√

2

units to the right). Each element in E is irrational (exercise 1.1). The Cantor set was bounded by [0, 1] so E

is clearly bounded by [

√

2, 1 +

√

2]. The proof that E is perfect is identical to the proof that the Cantor set is

perfect (given in the book in section 2.44 and also in this document after exercise 2.4).

Exercise 2.19 a

We’re told that A and B are disjoint, so A∩B = ∅. And if A and B are closed, then by theorem 2.27 we know

that A = A and B = B. So we conclude that A ∩ B = A∩ B = A ∩ B = ∅. By deﬁnition, this means that A

and B are separated.

Exercise 2.19 b

Let A and B be disjoint open sets. Assume that A∩ B is not empty. This assumption leads to a contradiction:

23

→ A∩ B is not empty hypothesis to be contradicted

→ (∃x)(x ∈ A∩ B) deﬁnition of non-emptiness

→ (∃x)(x ∈ A∩ (B ∪ B

)) deﬁnition of closure

→ (∃x)(x ∈ (A∩ B) ∪ (A∩ B

)) distributivity (section 2.11)

→ (∃x)(x ∈ ∅∪ (A∩ B

**)) A and B are disjoint
**

→ (∃x)(x ∈ A∩ B

)

→ (∃x)(x ∈ A∧ x ∈ B

**) deﬁnition of set intersection
**

→ (∃x)(x is an interior point of A∧ x ∈ B

) A is an open set

→ (∃x)(x is an interior point of A∧ x is a limit point of B) deﬁnition of B

→ (∃x) [(∃r ∈ R)(N

r

(x) ⊆ A) ∧ (x is a limit point of B)] deﬁnition of interior point

→ (∃x) [(∃r ∈ R)(N

r

(x) ⊆ A) ∧ (∀s ∈ R)(∃y ∈ B)(y ∈ N

s

(x))] deﬁnition of limit point

If something is true for all s ∈ R, it’s true for any particular R. So choose s = r.

→ (∃x) [(∃r ∈ R)( (N

r

(x) ⊆ A) ∧ (∃y ∈ B)(y ∈ N

r

(x)) )] substitution of s = r

→ (∃x) [(∃r ∈ R)(∃y ∈ B)(y ∈ N

r

(x) ∧ N

r

(x) ⊆ A)] rearrangement of terms for clarity

→ (∃x) [(∃r ∈ R)(∃y ∈ B)(y ∈ A)] deﬁnition of subset

This last step establishes a contradiction. The sets A and B are disjoint, so there cannot be any possible

choice of variables such that y ∈ B and y ∈ A. Our assumption must have been incorrect: A ∩ B is, in fact,

empty. If we swap the roles of A and B, this same proof also shows us that A∩B is empty. By deﬁnition, then,

A and B are separated.

Exercise 2.19 c

A is open in X: The set A is, by deﬁnition 2.18(a), a neighborhood of p. By theorem 2.19, then, A is an open

subset of X.

B is open in X: Let x be an arbitrary point in B. Let r = d(p, x) − δ. Let y be an arbitrary element in

N

r

(x). Proof that y ∈ B:

→ d(p, y) +d(y, x) ≥ d(p, x) Property 2.15(c) of metric spaces

→ d(p, y) +d(y, x) −d(p, x) ≥ 0

→ d(p, y) +d(y, x) −d(p, x) ≥ 0 ∧ r > d(x, y) we know d(x, y) < r because y ∈ N

r

(x)

→ d(p, y) +d(y, x) −d(p, x) ≥ 0 ∧ d(p, x) −δ > d(x, y) deﬁnition of r

→ d(p, y) +d(y, x) −d(p, x) ≥ 0 ∧ d(p, x) −δ −d(x, y) > 0

→ (d(p, y) +d(y, x) −d(p, x)) + (d(p, x) −δ −d(x, y)) > 0 the sum of positive numbers is positive

→ d(p, y) −δ > 0 cancellation of like terms

→ d(p, y) > δ

→ y ∈ B deﬁnition of membership in B

Our choice for y was arbitrary, so every point in this neighborhood of x is a member of B: by deﬁnition,

then, x is an interior point of B. But our choice of x ∈ B was also arbitrary, so this proves that every x ∈ B is

an interior point of B. By deﬁnition, this shows that B is open in X.

A and B are disjoint: If there were some x ∈ A ∩ B, then d(x, p) > δ and d(x, p) < δ, which violates the

trichotomy rule for order relations.

A and B are separate: We’ve shown that A and B are disjoint open sets, so by exercise 2.19(b) we know

that A and B are separate.

Exercise 2.19 d

If we are given any metric space X with a countable or ﬁnite number of elements, we can always ﬁnd some

distance δ such that there are no elements x, y ∈ X with d(x, y) = δ (proof follows). This allows us to choose

24

some arbitrary p ∈ X and then use the results from part (c) to completely partition X into separated sets

A = ¦x ∈ X : d(x, p) > δ¦ and B = ¦x ∈ X : d(x, p) < δ¦. This proves that (X is at most countable → X is not

connected). By contrapositive, this proves that (X is connected → X is uncountable), which is what we were

asked to prove.

Proof that an at most countable metric space X has some distance δ such that, for all x, y ∈ X, d(x, y) ,= δ:

Let X be an at most countable metric space with elements a

1

, a

2

, . . . a

n

. Then from theorem 2.13, we know

that the set of all order pairs ¦(a

i

, a

j

) : a

i

, a

j

∈ X¦ is at most countable. And there is a clear one-to-one

correspondence between this set and the set ¦(d(a

i

, a

j

) : a

i

, a

j

∈ X¦: so the set of all distances between all

combinations of points in X is at most countable.

Distances in metric spaces are always real numbers (deﬁnition 2.15), which are of course uncountable. Be-

cause there are an at-most countable number of distances in the set ¦(d(a

i

, a

j

) : a

i

, a

j

∈ X¦, we can choose a

real number δ that is not in this set (otherwise we would have an at most countable set with R as a proper subset).

But this isn’t quite enough: if δ is so large that there are no elements in the set ¦x ∈ X : d(x, y) > δ¦, then

one of our partitions will be empty. We can avoid this problem (as long as X has at least two elements) by picking

arbitrary x, y ∈ X and then choosing delta from the interval (0, d(x, y)). This interval is still uncountable, so

we are still able to choose a δ that is not in this set. And we know that d(x, y) > δ and d(x, x) < δ, so our

partitions A and B will both be nonempty.

Exercise 2.20 a

Closures of connected sets are always connected. If A and B are connected, then there is either some p in A∩B

or some q in A∩ B. Clearly, then, we have either either p ∈ (A

∪ A) ∩ B or q ∈ A∩ (B

∪ B). Therefore A∩ B

is nonempty, so these two closed sets are connected.

Exercise 2.20 b

Consider the segment (0, 1) of the real line. Although this segment is open in R

1

, it contains no interior

points in R

2

since every neighborhood N

r

(x, 0) contains the point (x,

r

2

). So let r =

1

4

and let E be the set

N

r

(0, 0) ∪ ¦(x, 0) : 0 < x < 1¦ ∪ N

r

(1, 0).

Since the line segment ¦(x, 0) : 0 < x < 1¦ contains no interior points while every point in a neighborhood is

an interior point (theorem 2.19), the interior of E is just N

r

(0, 0) ∪ N

r

(1, 1), and it is trivial to show that this

is the union of two nonempty separated sets.

Exercise 2.21 a

The function p(t) can be thought of as the parameterization of a straight line connecting the points a and b.

For instance, consider the following sets in R

2

:

25

The set A

0

is the set of all t such that p(t) ∈ A, and the set B

0

is the set of all t such that p(t) ∈ B. We’re

told that A and B are separated and are asked to prove that A

0

and B

0

are separated. Proof by contrapositive:

• There is some element t

A

that is an element of A

0

and a limit point of B

0

.

Assume that A

0

and B

0

are not separated. By deﬁnition 2.45, either A

0

∩ B

0

or A

0

∩ B

0

is nonempty. Assume

that A

0

∩ B

0

is nonempty (the sets are interchangeable, so the proof for A

0

∩ B

0

,= ∅ is almost identical). Let

t

A

be one element of A

0

∩ B

0

: then either t

A

∈ A

0

∩ B

0

or t

A

∈ B

0

. It can’t be the case that t

A

∈ A

0

∩ B

0

,

because this would imply p(t

A

) ∈ A∩B which is impossible because A and B are separated sets. So it must be

the case that t

A

∈ A

0

∩ B

0

.

• There is a proportional relationship between d(t, t +r) and d(p(t), p(t +r)).

The distance between p(t) and p(t +r) is the vector norm of p(t +r) −p(t):

d(p(t), p(t+r)) = [p(t+r)−p(t)[ = [[(1−(t+r))a+(t+r)b]−[(1−t)a+tb][ = [r(b−a)[ = [r[[(b−a)[ = d(t, t+r)d(a, b)

• p(t

A

) is an element of A and a limit point of B.

We’ve shown that t

A

is a limit point of B

0

. So for any arbitrarily small r, there is some t

B

∈ B

0

such

that d(t

A

, t

B

) < r. And this means that for any arbitrarily small r, there is some p(t

B

) ∈ B such that

d(p(t

A

), p(t

B

)) < [r(b −a)[. And since p(t

A

) is in A and each p(t

B

) is in B, this means that there is an element

of A that is a limit point of B: So A∩ B is nonempty, which means that A and B are not separated.

This shows that if A

0

and B

0

are not separated, then A and B are not separated. By contrapositive, then,

if A and B are separated then A

0

and B

0

are separated. And this is what we were asked to prove.

Exercise 2.21 b

We are asked to ﬁnd α ∈ (0, 1) such that p(α) ,∈ A∪ B: from the deﬁnition of the function p, this is equivalent

to ﬁnding α ∈ (0, 1) such that α ,∈ A

0

∪ B

0

. Proof that such a α exists:

Let E be deﬁned as E = ¦d(0, x) : x ∈ B

0

¦. That is, E is the set of all distances between the point 0 (which

is in A

0

, since p(0) = a) and elements of B

0

. The set R has the greatest lower bound property and E is a subset

of R with a lower bound of 0, so the set E has a greatest lower bound. Let α represent this greatest lower bound.

26

• 0 < α < 1:

We know that α > 0 because α is a distance. We know that α ,= 0, because this would mean that d(0, 0) ∈ E

which means that 0 ∈ B

0

, which is false. And we know that α < 1, because 1 ∈ B

0

and B

0

is an open set: so 1

is an interior point of B

0

, which means that there is some small r such that 1 −r ∈ B

0

.

We now need to show that we can use α to construct elements that are not in A

0

∪B

0

. We know that either

α ∈ B

0

,α ∈ A

0

, or α ,∈ A

0

∪ B

0

:

• if α ∈ B

0

:

We know that α is not a limit point of A

0

(otherwise, x

α

∈ B

0

∩A

0

which contradicts the fact from part (a) that

B

0

and A

0

are separated). So there is some neighborhood N

r

(α) that contains no points in A

0

. Now consider

the points p in the range (α −r, α). Each p is in the neighborhood N

r

(α), so they aren’t members of A

0

. And

for each p we see that d(0, p) < α, so they aren’t members of B

0

(otherwise α wouldn’t have been a lower bound

of E). So we see that for every p ∈ (α −r, α), we have p ,∈ A

0

∪ B

0

.

• if α ∈ A

0

:

This assumption leads to a contradiction. We’ve shown in part (a) that A

0

and B

0

are separated, so it can’t

be the case that α is a limit point of B

0

(otherwise A

0

∩ B

0

would not be empty). So there is some neighbor-

hood N

r

(α) that contains no points of B

0

. But if the interval (0, α) contains no points of B

0

and the interval

(α−r, α+r) contains no points of B

0

, then their union (0, α+r) contains no points of B

0

. But this contradicts

our deﬁnition of α as the greatest lower bound of E. So our initial assumption must have been incorrect: it

cannot have been the case that α ∈ A

0

.

• if α ,∈ A

0

and α ,∈ B

0

:

Under this assumption, we clearly have α ,∈ A

0

∪ B

0

.

Whatever assumption we make about the set containing α, we see that there will always be at least one

element α such that 0 < α < 1 and α ,∈ A

0

∪ B

0

. And, from the deﬁnition of the function p, this means that

p(α) ,∈ A∪ B. And this is what we were asked to demonstrate.

Exercise 2.21 c

Proof by contrapositive. Assume that E is not connected: then E could be described as the union of two

separated sets (i.e, E = A ∪ B). From part (b), we could then choose a, b, and t such that (1 − t)a + (t)b ,∈

A∪ B = E. And this would mean that E is not convex. By contrapositive, if E is convex then E is connected.

Exercise 2.22

The metric space R

k

clearly contains Q

k

as a subset. We know that Q

k

is countable from theorem 2.13. To

prove that Q

k

is dense in R

k

, we need to show that every point in R

k

is a limit point of Q

k

:

Let a = (a

1

, a

2

, . . . , a

k

) be an arbitrary point in R

k

and let N

r

(a) be an arbitrary neighborhood of a. Let

b = (b

1

, b

2

, . . . , b

k

) where b

i

is chosen to be a rational number such that a

i

< b

i

< a

i

+

r

√

k

(possible via

theorem 1.20(b)). The point b is clearly in Q

k

, and

d(a, b) =

_

(a

1

−b

i

)

2

+. . . + (a

k

−b

k

)

2

<

_

r

2

k

+. . . +

r

2

k

=

_

kr

2

k

= r

This shows that every point in R

k

is a limit point of Q

k

, which completes the proof that R

k

is separable.

Exercise 2.23

Let X be a separable metric space. From the deﬁnition of separable in the previous exercise, we know that

X contains a countable dense subset E. For each e

i

∈ E, let N

q

(e

i

) be a neighborhood with rational radius

q around point e

i

. Let ¦V

α

¦ = ¦N

q

(e

i

) : q ∈ Q, i ∈ N¦ be the collection of all neighborhoods with rational

radius centered around members of E. This is a countable collection of countable sets, and therefore ¦V

α

¦ has

27

countably many elements.

Let x be an arbitrary point in X, and let G be an arbitrary open set in X such that x ∈ G. Because G is

open, we know that x is an interior point of G. So there is some neighborhood N

r

(x) such that N

r

(x) ⊆ G. But

we can choose a rational q such that 0 < q <

r

2

, so that x ∈ N

q

(x) ⊆ N

r

(x) ⊆ G.

Because E is dense in x, every neighborhood of x contains some e ∈ E. So e ∈ N

q

(x), which means that

d(e, x) < q. But, on the other hand, this also means that d(x, e) < q so that x ∈ N

q

(e). And N

q

(e) ∈ ¦V

α

¦, by

deﬁnition.

Having shown that x ∈ N

q

(e), we need to prove that N

q

(e) ⊆ G. Let y be any point in N

q

(e). We know

that d(x, e) < q and d(e, y) < q. So d(x, y) < d(x, e) + d(e, y) = 2q by deﬁnition 2.15(c) of metric spaces. But

we deﬁned q so that 0 < q <

r

2

, so we know that d(x, y) < r. This means that every y ∈ N

q

(e) → y ∈ N

r

(x), or

N

q

(e) ⊆ N

r

(x). And we chose r so that N

r

(x) ⊆ G, so by transitivity we know that N

q

(e) ⊆ G.

We started by choosing an arbitrary element x in an arbitrary open set G ⊆ X, and proved that there was an

element N

q

(e) ∈ ¦V

α

¦ such that x ∈ N

q

(e) ⊂ G. We’ve shown that ¦V

α

¦ has a countable number of elements,

and each element is an open neighborhood. This proves that ¦V

α

¦ is a base for X.

Exercise 2.24

We’re told that X is a metric space in which every inﬁnite subset has a limit point. Choose some arbitrary δ

and construct a set ¦x

i

¦ as described in the exercise.

• The set ¦x

i

¦ must be ﬁnite: if this set were inﬁnite, then it would have a limit point. And if it has a

limit point, then there would be multiple points of ¦x

i

¦ in some neighborhood of radius δ/2, which contradicts

our assumption that d(x

i

, x

j

) > δ for each pair of points in ¦x

i

¦.

• The neighborhoods of ¦x

i

¦ form a cover of X: Every x ∈ X must be contained in N

δ

(x

i

) for some

x

i

∈ ¦x

i

¦. If it weren’t, this would imply that d(x, x

i

) > δ for each x

i

∈ ¦x

i

¦, so that we could have chosen x

to be an additional element in ¦x

i

¦.

Let E

i

represent the set of ¦x

i

¦ constructed in the above way when δ =

1

i

, and consider the set E =

∞

i=1

E

i

.

We will show that E is a countable dense subset of X. This union E is a countable union of nonempty ﬁnite

sets, so E is countable. Each element of E was chosen from X, so clearly E ⊂ X. Proof that E is dense in X:

Choose an arbitrary x ∈ X. Choose an arbitrarily small radius r. From the archimedian principle, we can

ﬁnd an integer n such that

1

n

< r. We’ve shown that the neighborhoods of E

n

cover X, so there is some e ∈ E

n

such that x ∈ N

1/n

(e). But if d(e, x) <

1

n

, then d(x, e) <

1

n

and so e ∈ N

1/n

(x). And since

1

n

< r, this implies

that e ∈ N

r

(x).

So we’ve chosen an arbitrary x and an arbitrary radius r, and shown that N

r

(x) will always contain some

e ∈ E

n

⊂ E. This proves that every x is a limit point of E, which by deﬁnition means that E is dense in X.

28

This proves that X contains a countable dense subset, so by the deﬁnition in exercise 22 we have proven that

X is separable.

Exercise 2.25 a

Theorem 2.41 tells us that the set K, by virtue of being a compact space, has the property that every inﬁnite

subset of K has a limit point. By exercise 24, this means that K is separable. And exercise 23 proves that every

separable metric spaces has a countable base.

I’m not sure that this an appropriate proof, though. The question wants us to prove that K has a countable

base and that therefore K is separable, whereas this is a proof that K is separable and therefore has a countable

base. An alternate proof follows.

Exercise 2.25 b

We’re told that K is compact. Consider the set E

n

= ¦N

1/n

(k) : k ∈ K¦, the set of neighborhoods of radius

1/n around every element in K. This is clearly an open cover of K, since x ∈ K → x ∈ N

(

1/n)(x) → x ∈ E

n

.

And K is compact, so the open cover E

n

must have some ﬁnite subcover: i.e., K must be covered by some ﬁnite

number of neighborhoods from E

n

. Let ¦V

n

¦ represent a ﬁnite subcover of the open cover E

n

.

Now consider the union V =

∞

n=1

V

n

, a countable collection of ﬁnite covers of K. We will prove that V is

a base for K.

Choose an arbitrary x ∈ K, and an arbitrary open set G such that x ∈ G ⊂ K. Because G is open and

x ∈ G, x must be an interior point of G. And since x is an interior point, there is some neighborhood N

r

(x) such

that N

r

(x) ⊂ G. Now choose an integer m such that

1

m

<

r

2

. Because V

m

is an open cover of K, there is some

k ∈ K and neighborhood N

1/m

(k) in the open cover V

m

such that x ∈ N

1/m

(k). Proof that this neighborhood

N

1/m

(k) is a subset of G:

Assume that y is an element of N

1/m

(k). From the fact that x ∈ N

1/m

(k), we know that d(x, k) < 1/m.

And from the fact that y ∈ N

1/m

(k), we know that d(k, y) < 1/m. And from the deﬁnition of metric spaces, we

know that d(x, y) ≤ d(x, k) +d(k, y), which means that d(x, y) ≤ 2/m ≤ r (since we chose

1

m

<

r

2

). This proves

that every element of N

1/m

(k) is in the neighborhood N

r

(x), so N

1/m

(k) ⊆ N

r

(x) ⊆ G. And this shows that

N

1/m

(k) ⊆ G.

We’ve chosen an arbitrary element x and an arbitrary element G, and have shown that there is some

N

1/m

(k) ∈ V such that x ∈ N

1/m

(k) ⊂ G. By the deﬁnition in exercise 23, this proves that K has a base. And

this base is a countable collection of ﬁnite sets, so it’s a countable base.

This countable base V is identical to the set E constructed in exercise 24. We saw there that this base was

a dense subset, so by the deﬁnition of “separable” in exercise 23 we know that K is separable.

Exercise 2.26

Let ¦V

n

¦ be the countable base we constructed in exercises 24 and 25.

¦V

n

¦ acts as a countable subcover to any open cover of X. Proof: let G be an arbitrary open set from an

arbitrary open cover of X. Choose any x ∈ G. We can ﬁnd an element N

1/m

(k) ∈ V such that x ∈ N

1/m

(k) ⊆ G

by the same method used in the second half of exercise 2.25(b).

We must now prove that a ﬁnite subset of ¦V

n

¦ can always be chosen to act as a subcover for any open cover.

We’ll do this with a proof by contradiction.

Let ¦W

α

¦ be an open cover of X and assume that there is no ﬁnite subset of ¦V

n

¦ that acts as a subcover

of ¦W

α

¦. Construct F

n

and E as described in the exercise. Because E is an inﬁnite subset, it has a limit point:

let x be this limit point. Every neighborhood N

r

(x) contains inﬁnitely many points of E (theorem 2.20), and

we can prove that this arbitrary neighborhood N

r

(x) must contain every point of E:

29

We have constructed the F

n

sets in such a way that F

n+1

⊆ F

n

for each n. So if N

r

(x) doesn’t contain any

points from F

j

, then it doesn’t contain any points from F

j+1

, F

j+2

,etc because y ∈ F

j+1

→ y ∈ F

j

. So if N

r

(x)

doesn’t contain any points from F

j

, then it has no more than j −1 points of E: i.e, N

r

(x) ∩ E would be ﬁnite.

But x is a limit point of E, so N

r

(x) must contain inﬁnitely many points of E. So N

r

(x) ∩E contains one point

from each F

n

.

But this means that E = ¦x¦, a set with one element. If it contained a second point, we could ﬁnd a

neighborhood N

r

(x) that failed to contain this second point, which would contradict our ﬁnding that every

neighborhood of x contains E. And this means that we chose the same element from each F

n

: i.e., x was in

each F

n

. So x ∈

F

n

, which contradicts our assumption that

F

n

was empty.

Exercise 2.27

To prove that P is perfect, we must show that every limit point of P is a point of P, and vice-versa.

Proof that every limit point of P is a point of P: assume that x is a limit point of P. Choose some arbitrary

r and let s =

r

2

. By the deﬁnition of limit point, every neighborhood N

s

(x) contains some y ∈ P. And from the

fact that y ∈ P, we know that N

s

(y) contains uncountably many points of P. But for each p ∈ N

s

(y), we see

that d(x, p) ≤ d(x, y) +d(y, p) so that d(x, p) ≤ 2s = r. So the neighborhood N

r

(x) contains uncountably many

points of P. But N

r

(x) was an arbitrary neighborhood of x, so this proves that x ∈ P.

Proof that every point of P is a limit point of P: assume that x ∈ P. By the deﬁnition of P, this means that

every neighborhood of x contains uncountably many points of P. So clearly every neighborhood of x contains

at least one point of P, which means that x is a limit point of P.

This proves that P is perfect. To prove that P

c

∩E is countable, we let ¦V

n

¦ be a countable base for X and

let W be the union of all V

n

for which E ∩ V

n

is countable. We will show that W

c

= P:

Proof that P ⊆ W

c

: Assume that x ∈ P. By the deﬁnition of membership in P, we know that x is a

condensation point of E. By the deﬁnition of condensation point, then, we know that every neighborhood

N

r

(x) contains uncountably many points of E. And this means that every open set V

n

containing x contains

uncountably many points of E. So x is not a member of any countable V

n

∩ E, which means that x ∈ W

c

.

Proof that W

c

⊆ P: Assume that x ∈ W

c

. Choose any arbitrary neighborhood N

r

(x): by the deﬁnition

of the base, there is some V

n

⊂ N

r

(x) such that x ∈ V

n

. By deﬁnition of W

c

, the fact that x ∈ V

n

means

that E ∩ V

n

is uncountable, which of course means that V

n

has uncountably many elements of E. And we’ve

established that V

n

⊆ N

r

(x), which means that N

r

(x) has uncountably many elements of E. But N

r

(x) was an

arbitrary neighborhood of x, so this proves that every neighborhood of x has uncountably many elements of E:

by deﬁnition, then, x is a condensation point and therefore x ∈ P.

So we have proven that W

c

= P. Taking the complement of each of these, we see that W = P

c

. And the set

W is the union of countably many sets of the form E ∩V

n

, each of which contains countably many elements: so

W is countable. And W = P

c

, so P

c

is countable: this proves that countably many points are P

c

. And from

this, it’s trivial to show that there are countably many points in E ∩P

c

, which is what we were asked to prove.

Note that this proof assumed only that X had a countable base, so it is valid for any separable metric space.

Exercise 2.28

Let E be a closed set in a separable metric space X. Let P be the set of all condensation points for E.

Proof that E ∩P is perfect: assume that x ∈ E ∩P. Clearly x is a point of P, which means that x is a limit

point of E (from the deﬁnition of P) and a limit point of P (because P was shown to be perfect in exercise 27).

So x is a limit point of E ∩ P. Now, assume that x is a limit point of E ∩ P. From this, we know that x is a

point of E (because E is closed) and a point of P (because P is perfect). So we have shown that every point of

E ∩ P is a limit point of E ∩ P and vice-versa: this proves that E ∩ P is perfect.

30

Proof that E ∩ P

c

is at most countable: this was proven in exercise 27.

And E = (E ∩ P) ∪ (E ∩ P

c

): that is, E is the union of a perfect set and a countable set. And this is what

we were asked to prove.

Exercise 2.29

Let E be an arbitrary open set in R

1

and let ¦G

i

¦ be an arbitrary collection of disjoint segments such that

G

i

= E. Each G

i

is open and nonempty, so each G

i

contains at least one rational point. Therefore there

can’t be more G

i

elements than rational numbers, which means that there are at most a countable number of

elements in ¦G

i

¦. But ¦G

i

¦ was an arbitrary set of disjoint segements whose union is E, therefore E cannot be

the union of an uncountable number of disjoint segments.

Exercise 2.30

Proof by contradiction: Let ¦F

n

¦ be a collection of sets such that

F

n

= R and suppose that each F

n

has an

empty interior.

→ (∀n)(F

◦

n

= ∅) hypothesis of contradiction

→

F

◦

n

= ∅

→ (

F

◦

n

)

c

= R taking the complement of both sides

→

(F

◦

n

)

c

= R theorem 2.22

→

F

c

n

= R exercise 2.9d

→ (∀n)(F

c

n

= R)

This last step tells us that F

c

n

is dense in R for every n. And F

n

was closed, so F

c

n

is open. Appealing to the

proof of Baire’s theorem in exercise 3.22 (which uses only terms and concepts introduced in chapter 2), we see

that

F

c

n

is nonempty. Therefore:

→

F

c

n

,= ∅

→ (

F

n

)

c

,= ∅ theorem 2.22

→

F

n

,= R taking the complement of both sides

And this contradicts our original claim that

F

n

= R so one of our initial assumptions must be wrong. And

our only assumption was that each F

n

has an empty interior, so by contradiction we have proven that at least

one F

n

must have a nonempty interior.

Exercise 3.1

The exercise does not explicitly say that we’re operating in the metric space R

k

, but there are two reasons to

assume that we are. First, we are supposed to understand the meaning of absolute value in this metric space,

and R

k

is the only metric space we’ve encountered so far for which absolute value has been deﬁned. Second,

Rudin appears to use s

n

to represent series in R

k

and p

n

to represent series in arbitrary metric spaces. So we’ll

assume that we’re operating in R

k

for this exercise.

We’re told that the sequence ¦s

n

¦ converges to some value s: that is, for any arbitrarily small there is some

integer N such that n > N implies d(s, s

n

) < . But it’s always the case that d([s[, [s

n

[) ≤ d(s, s

n

) (exercise

1.13), so for this same and N we see that d([s[, [s

n

[) ≤ d(s, s

n

) < . By transitivity, this means that for any

choice of there is some integer N such that n > N implies d([s[, [s

n

[) < . By deﬁnition of convergence, this

means that the sequence ¦[s

n

[¦ converges to [s[.

The converse is not true. If we let s

k

= (−1)

k

, the sequence ¦[s

n

[¦ clearly converges while the sequence ¦s

n

¦

clearly does not.

31

Exercise 3.2

The exercise asks us to calculate the limit rather than prove it rigorously, so we might be able to manipulate

the expression algebraically:

_

n

2

+n −n = (

_

n

2

+n −n)

_

√

n

2

+n +n

√

n

2

+n +n

_

=

n

√

n

2

+n +n

=

n

n

_

1 + 1/n +n

=

1

_

1 + 1/n + 1

and then use basic calc 2 techniques to show that this limit is

1

2

. If we need to prove it more rigorously, we can

use theorem 3.14. We will ﬁrst need to prove that this sequence has a limit, which theorem 3.14 tells us can

be done by proving that the sequence is bounded and monotonically increasing. We can then ﬁnd the limit by

ﬁnding the least upper bound of the sequence.

The sequence is bounded

The quantitity (

√

n

2

+n − n) is bounded below: from the fact that

√

n

2

+n >

√

n

2

= n, we know that

√

n

2

+n −n > 0. And it’s bounded above by

1

2

:

→

√

n

2

+n −n ≥

1

2

hypothesis to be contradicted

→

√

n

2

+n ≥

1

2

+n

→ n

2

+n ≥

1

4

+n +n

2

both sides are positive, so the sign doesn’t change

→ 0 ≥

1

4

subtract n +n

2

from both sides

This last step is clearly false, so our initial assumption must be false: this shows that (

√

n

2

+n − n) is

bounded above by

1

2

.

The sequence is monotonically increasing

→ s

n+1

> s

n

↔

_

(n + 1)

2

+ (n + 1) −(n + 1) >

√

n

2

+n −n deﬁnition of s

n

↔

_

(n + 1)

2

+ (n + 1) −

√

n

2

+n > (n + 1) −n rearrange terms

↔

√

n

2

+ 3n + 2 −

√

n

2

+n > 1 some algebra

↔ (n

2

+ 3n + 2) + (n

2

+n) −2

√

n

2

+ 3n + 2

√

n

2

+n > 1 square both sides

↔ (2n

2

+ 4n + 2) −2

√

n

4

+ 4n

3

+ 5n

2

+ 2n > 1 simplify

↔ −2

√

n

4

+ 4n

3

+ 5n

2

+ 2n > 1 −(2n

2

+ 4n + 2) rearrange terms

↔ 2

√

n

4

+ 4n

3

+ 5n

2

+ 2n < 2n

2

+ 4n + 1 multiply both sides by −1 and simplify

↔ 4n

4

+ 16n

3

+ 20n

2

+ 8n < 4n

4

+ 16n

3

+ 20n

2

+ 8n + 1 square both sides

↔ 0 < 1 subtract 4n

4

+ 16n

3

+ 20n

2

+ 8n from each side

The sequence has a least upper bound of

1

2

We’ve already shown that

1

2

is an upper bound. To show that it is the least such upper bound, choose any

x ∈ [0,

1

2

) and assume that it is an upper bound for ¦s

n

¦.

→ (∀n ∈ N)(s

n

≤ x) hypothesis to be contradicted

→ (∀n ∈ N)(

√

n

2

+n −n ≤ x) deﬁnition of s

n

→ (∀n ∈ N)(

√

n

2

+n ≤ x +n)

→ (∀n ∈ N)(n

2

+n ≤ x

2

+ 2xn +n

2

) square both sides

→ (∀n ∈ N)(0 ≤ x

2

+ (2x −1)n) subtract n

2

+n from both sides

→ (∀n ∈ N)(−x

2

≤ (2x −1)n) subtract x

2

from both sides

→ (∀n ∈ N)(

−x

2

2x−1

≤ n) 2x −1 < 0 when x <

1

2

This last step must be false: by the Archimedian principle, we can always ﬁnd some n ∈ N greater than any

speciﬁc quantity. So, by contradiction, we can ﬁnd s

n

> x for every x ∈ [0,

1

2

): this means that the upper bound

cannot be less than

1

2

. But

1

2

is an upper bound, so we see that

1

2

is the least such upper bound.

32

We have shown that ¦s

n

¦ is a monotonically increasing bounded sequence with a least upper bound of

1

2

:

by theorem 3.14, this proves that the limit of ¦s

n

¦ is

1

2

.

Exercise 3.3

Note that we are not asked to ﬁnd the limit of this sequence. We are only asked to show that it converges and

that 2 is an upper bound. By theorem 3.14, we can prove convergence by proving that the sequence is bounded

and monotonically increasing.

The sequence is bounded above by

_

2 +

√

2

Although we’re asked to show that the sequence has an upper bound of 2, we can prove the stronger result that

it has an upper bound of

_

2 +

√

2. We can prove this by induction. We see that 0 < s

1

=

√

2 < 2. Now assume

that 0 < s

n

< 2:

→ 0 < s

n

< 2 hypothesis of induction

→ 0 <

√

s

n

<

√

2

→ 0 < 2 +

√

s

n

< 2 +

√

2

→

_

2 +

√

s

n

<

_

2 +

√

2

→ s

n+1

<

_

2 +

√

2

By induction, then, s

n

<

_

2 +

√

2 for all n.

The sequence is monotonically increasing

Proof by induction. We can immediately see that s

1

=

√

2 <

_

2 +

√

2 = s

2

. Now, assume that we have

proven s

i−1

< s

i

for i = 1 . . . n.

→ s

n

=

_

2 +

√

s

n−1

deﬁnition of s

n

→ s

n

>

_

2 +

√

s

n−2

our hypothesis of induction tells us s

n−1

> s

n−2

→ s

n

> s

n−1

deﬁnition of s

n−1

We’ve shown that s

n

is a monotonically increasing function that is bounded above. By theorem 3.14, this is

suﬃcient to prove that it converges.

Exercise 3.4

From the recursive deﬁnition we are given, we can see that s

2

= 0 and s

2m

=

1

4

+

1

2

s

2m−2

. So we can use

theorem 3.26 to ﬁnd the limit of s

2m

.

lim

m→∞

s

2m

=

1

4

+

1

8

+

1

16

+. . . =

_

∞

n=0

1

2

n

_

−1 −

1

2

=

1

1 −

1

2

−1 −

1

2

=

1

2

From the same recursive deﬁnition, we see that s

1

= 0 and s

2m+1

=

1

2

+

1

2

s

2m−1

. So the limit of s

2m+1

is given

by

lim

m→∞

s

2m+1

=

1

2

+

1

4

+

1

8

+. . . =

_

∞

n=0

1

2

n

_

−1 =

1

1 −

1

2

−1 = 1

This shows that as n increases, the terms of ¦s

n

¦ are either arbitrarily close to

1

2

or arbitrarily close to 1, so

these are our upper and lower bounds for the sequence.

Exercise 3.5

Let a

∗

= lim

n→∞

sup a

n

, let b

∗

= lim

n→∞

sup b

n

, and let s

∗

= lim

n→∞

sup(a

n

+b

n

).

If a

∗

and b

∗

are both ﬁnite:

From theorem 3.17(a) we know that there are some subsequences ¦a

j

¦, ¦b

k

¦ such that lim

j→∞

a

j

+lim

k→∞

b

k

=

33

s

∗

(that is, the supremum of E as deﬁned above is also member of E: E is a closed set). The limit of a

j

must

be less than or equal to the supremum a

∗

and the limit of b

j

must be less than or equal to the supremum b

∗

, so

s

∗

= lim

j→∞

a

j

+ lim

k→∞

b

k

≤ lim

n→∞

sup a

n

+ lim

n→∞

sup b

n

= a

∗

+b

∗

By transitivity, this shows that s

∗

≤ a

∗

+b

∗

: and, by the deﬁnitions of these terms, this proves that

lim

n→∞

sup(a

n

+b

n

) ≤ lim

n→∞

sup a

n

+ lim

n→∞

sup b

n

If either a

∗

or b

∗

is inﬁnite:

We’re asked to discount the possibility that a

∗

= ∞ and b

∗

= −∞. In all other cases when one or both of these

values is inﬁnite, the inequality can be easily shown to resolve to ∞ = ∞ or −∞ = −∞.

Exercise 3.6a

a

n

=

√

n + 1 −

√

n =

_√

n + 1 −

√

n

_

_√

n + 1 +

√

n

√

n + 1 +

√

n

_

=

1

√

n + 1 +

√

n

We can compare this last term to a known series:

a

n

=

1

√

n + 1 +

√

n

>

1

2

√

n + 1

>

1

2

_

1

n

_

1/2

We know that the series

_

1

n

_

1/2

diverges by theorem 3.28, so the terms of ¦a

n

¦ are larger than the terms of

a divergent series. By the comparison theorem 3.25, this tells us that

a

n

is divergent.

Exercise 3.6b

a

n

=

√

n + 1 −

√

n

n

=

√

n + 1 −

√

n

n

_√

n + 1 +

√

n

√

n + 1 +

√

n

_

=

1

√

n

3

+n

2

+

√

n

2

We can compare this last term to a known series:

1

√

n

3

+n

2

+

√

n

3

<

1

√

n

3

+

√

n

3

=

1

2

_

1

n

_

3/2

We know that the series

_

1

n

_

3/2

converges by theorem 3.28, so the terms of ¦a

n

¦ are smaller than the terms

of a convergent series. By the comparison theorem 3.25, this tells us that

a

n

is convergent.

Exercise 3.6c

For n ≥ 1, we can show that 0 ≤ (

n

√

n −1) < 1:

→ (∀n ∈ N)(1 ≤ n < 2

n

) the < 2

n

can easily be proven via induction

→ (∀n ∈ N)(1 ≤

n

√

n < 2) take the n

th

root of each term

→ (∀n ∈ N)(0 ≤

n

√

n −1 < 1) subtract 1 from each term

We know that the series

x

n

converges when 0 ≤ x < 1, so by the comparison theorem 3.25 we know that

a

n

is convergent.

Exercise 3.6d

This probably is more diﬃcult than it looks at ﬁrst, mainly because for complex z it’s not true that 1 +z

n

> z

n

– in fact, without absolute value signs the inequality z

1

> z

2

has no meaning whatsoever (see exercise 1.8). It’s

also not true that [1 +z

n

[ > [z

n

[. The best we can do is appeal to the triangle inequality.

34

If [z[ < 1 then by the triangle inequality we have

lim

n→∞

¸

¸

¸

¸

1

1 +z

n

¸

¸

¸

¸

≥ lim

n→∞

1

[1[ +[z

n

[

= 1

and therefore by theorem 3.23 the series doesn’t converge. If [z[ = 1 then we similarly have

lim

n→∞

¸

¸

¸

¸

1

1 +z

n

¸

¸

¸

¸

≥ lim

n→∞

1

[1[ +[z

n

[

=

1

2

and by theorem 3.23 the series again fails to converge. If [z[ > 1 then we can use the ratio test:

lim

n→∞

¸

¸

¸

¸

1 +z

n

1 +z

n+1

¸

¸

¸

¸

= lim

n→∞

¸

¸

¸

¸

z

−n

z

−n

1 +z

n

1 +z

n+1

¸

¸

¸

¸

= lim

n→∞

¸

¸

¸

¸

z

−n

+ 1

z

−n

+z

¸

¸

¸

¸

=

¸

¸

¸

¸

1

z

¸

¸

¸

¸

< 1

and by theorem 3.34 the series converges. I’m not totally satisﬁed with this proof because I can’t totally justify

the claim that lim

n→∞

z

−n

= 0 without appealing to the polar representation z

−n

= r

−n

e

−inθ

and the fact that

[e

ixθ

[ = 1 for all x.

Exercise 3.7

Deﬁne the partial sum t

n

to be

t

n

=

n

k=1

√

a

k

k

We can show that the series

√

a

n

/n converges by showing that the sequence ¦t

n

¦ converges; we can do this

by showing that ¦t

n

¦ is bounded and monotonically increasing (theorem 3.14).

the sequence is monotonically increasing:

From the deﬁnition of partial sums, we know that t

n

=

√

a

n

/n+t

n−1

for all n. And we’re told a

n

> 0 for every

n, so t

n

> t

n−1

for every n.

the sequence is bounded:

From the Cauchy-Schwartz inequality, we know that

(ab) ≤

_

a

2

b

2

. Applying this to the given series,

we see that

_

√

a

n

1

n

_

≤

_

a

n

1

n

2

We are told that

a

n

converges, and from theorem 3.28 we know that that

1/n

2

converges. Therefore (by

theorem 3.50) we know that their product converges to some α, and so the right-hand side of the above inequality

converges to

√

α. This shows us that the series

√

a

n

/n is bounded by

√

α, and therefore the sequence ¦t

n

¦ is

bounded by

√

α.

By assuming that

a

n

is convergent and that a

n

> 0, we’ve shown that t

n

is bounded and monotonically

increasing. By theorem 3.14 this is suﬃcient to show that t

n

converges. And this is what we were asked to

prove.

Exercise 3.8

We’re told that ¦b

n

¦ is bounded. Let α be the upper bound of ¦[b

n

[¦. We’re also told that

a

n

converges: so

for any arbitrarily small , we can ﬁnd an integer N such that

¸

¸

¸

¸

¸

n

k=m

a

k

¸

¸

¸

¸

¸

≤

α

for all n, m such that n ≥ m ≥ N

which is algebraically equivalent to

¸

¸

¸

¸

¸

n

k=m

a

k

α

¸

¸

¸

¸

¸

≤ for all n, m such that n ≥ m ≥ N

35

which, since [b

k

[ ≤ α for every k, means that

¸

¸

¸

¸

¸

n

k=m

a

k

b

k

¸

¸

¸

¸

¸

≤

¸

¸

¸

¸

¸

n

k=m

a

k

α

¸

¸

¸

¸

¸

≤ for all n, m such that n ≥ m ≥ N

By theorem 3.22, this is suﬃcient to prove that

a

n

b

n

converges.

Exercise 3.9a

α = limsup

n

_

[n

3

[ = limsup [n

3/n

[ = [n

0

[ = 1

So the radius of convergence is R = 1/α = 1.

Exercise 3.9b

Example 3.40(b) mentions that “the ratio test is easier to apply than the root test”, although we’re never

explicitly told what the ratio test is, how it might relate to the ratio test for series convergence, or how to apply

the ratio test to example (b). According to some course handouts I found online

2

, we can’t simply apply the

ratio test in the same way we use the root test. That is, we can’t use the fact that

β = limsup

¸

¸

¸

¸

2

n+1

(n + 1)!

n!

2

n

¸

¸

¸

¸

= limsup

¸

¸

¸

¸

2

n + 1

¸

¸

¸

¸

= 0

to assume that the radius of convergence is 1/β = ∞. Instead, we look at theorem 3.37, which tells us that

α ≤ β, so that

R =

1

α

≥

1

β

= ∞

Exercise 3.9c

α = limsup

n

¸

¸

¸

¸

¸

2

n

n

2

¸

¸

¸

¸

= limsup

2

¸

¸

¸

n

√

n

2

¸

¸

¸

=

2

(limsup

√

n)

2

The last step of this equality is justiﬁed by theorem 3.3c. From theorem 3.20c, we know the limit of this last

term:

2

(limsup

√

n)

2

=

2

1

So α = 2, which gives us a radius of convergence of 1/2.

Exercise 3.9d

α = limsup

n

_

[n

3

3

n

[ = limsup

n

√

n

3

3

=

(limsup

n

√

n)

3

3

The last step of this equality is justiﬁed by theorem 3.3c. From theorem 3.20c, we know the limit of this last

term:

(limsup

n

√

n)

3

3

=

1

3

So α = 1/3, which gives us a radius of convergence of 3.

Exercise 3.10

We’re told that inﬁnitely many of the coeﬃcients of

a

n

z

n

are positive integers. This means that for every N,

there is some n > N such that a

n

≥ 1. Therefore lim sup

n→∞

[a

n

[ ≥ 1

→ lim sup

n→∞

n

_

[a

n

[ ≥ 1 from the fact thatk > 1 → k

1/n

> 1

→

1

R

≥ 1 theorem 3.39

→ 1 ≥ R

This ﬁnal step indicates that the radius of convergence is at most 1.

2

http://math.berkeley.edu/∼gbergman/

36

Exercise 3.11a

Proof by contrapositive: we will assume that

a

n

/(1 + a

n

) converges and show that this implies that

a

n

converges.

To simplify the form of

a

n

/(1 +a

n

), we can multiply the numerator and denominator of each term of the

series by 1/a

n

to get

1

1

an

+ 1

For this to converge, it is necessary that

lim

n→∞

1

1

an

+ 1

= 0

which can only happen if the limit of the denominator is ∞; and this can only happen if the limit of 1/a

n

is ∞.

That is, it must be the case that

lim

n→∞

a

n

= 0

This alone is not suﬃcient to prove that

a

n

converges. It does, however, allow us to make the terms of a

n

as

arbitrarily small as we like. For this proof, we it’s helpful to recognize that there is some integer N

1

such that

a

n

< 1 for all n > N

1

.

From the assumption that

a

n

/(1 +a

n

) converges, we know that for every there is some N

such that

n

k=m

a

k

1 +a

k

< for every n ≥ m ≥ N

**Let N be the larger of ¦N
**

1

, N

**¦. For any and any n ≥ m ≥ N, we can produce two inequalities:
**

n

k=m

a

k

1 + 1

<

n

k=m

a

k

1 +a

k

n

k=m

a

k

1 +a

k

< for every n ≥ m ≥ N

The ﬁrst is true because each k is larger than N

1

(so that a

k

< 1); the second is true because each k is larger

than N

**. Together, these inequalities tell us that
**

n

k=m

a

k

2

< for every n ≥ m ≥ N

or, equivalently, that

n

k=m

a

k

< 2 for every n ≥ m ≥ N

And our choice of was arbitrary, so we have shown that for every there is some N that makes the last

statement true: and this is the deﬁnition of convergence for the series

a

n

.

We’ve shown that the convergence of

an

1+an

implies the convergence of

a

n

. By contrapositive, the fact

that

a

n

diverges is proof that

an

1+an

diverges.

Exercise 3.11b

From the deﬁnition of s

n

, we can see that

s

n+1

= a

1

+. . . +a

n+1

= s

n

+a

n+1

37

and we’re told that every a

n

> 0, so we know that s

n+1

> s

n

. By induction, we also know that s

n

≥ s

m

whenever n ≥ m. And from this, we know that

1

sm

≥

1

sn

whenever n ≥ m. Therefore:

a

N+1

s

N+1

+. . . +

a

N+k

s

N+k

≥

a

N+1

s

N+k

+. . . +

a

N+k

s

N+k

=

a

N+1

+. . . +a

N+k

s

N+k

A bit of algebraic manipulation shows us that a

N+1

+. . .+a

N+k

= s

N+k

−s

N

, so this last inequality is equivalent

to

a

N+1

s

N+1

+. . . +

a

N+k

s

N+k

≥

s

N+k

−s

N

s

N+k

= 1 −

s

N

s

N+k

which is what we were asked to prove.

To show that

an

sn

is divergent, we once again look at the sequence ¦s

n

¦. We’ve already determined that

¦s

n

¦ is an increasing sequence, and we know that’s it’s not convergent (otherwise, by the deﬁnition 3.21 of

“convergent series”,

a

n

would be convergent). Therefore, we know that ¦s

n

¦ is not bounded (from theorem

3.14, which says that a bounded monotonic series is convergent).

Now let s

N

be an arbitrary element of ¦s

n

¦ and let be an arbitrarily small real. From the fact that ¦s

n

¦

is not bounded, we can make s

N+k

arbitrarily large by choosing a suﬃciently large k. This means that, for any

N and any arbitrarily small , we can make

s

N

s

N+k

< by choosing a suﬃciently large k. And we’ve established

that

a

N+1

s

N+1

+. . . +

a

N+k

s

N+k

≥ 1 −

s

N

s

N+k

which means that, by choosing suﬃciently large k, we get the inequality

N+k

k=N+1

a

k

s

k

≥ 1 −

s

N

s

N+k

≥ 1 −

We’ve now established everything we need to show that

an

sn

is divergent. Choose any such that 0 < <

1

2

.

From theorem 3.22, in order for

an

sn

to be convergent we must be able to ﬁnd some integer N such that

n

k=m

a

k

s

k

< for all n ≥ m ≥ N

But there can be no such N, because for every N we’ve shown that there is some k such that

N+k

k=N+1

a

k

s

k

> 1 − >

(note that 1 - > because 0 < <

1

2

). And this is suﬃcient to show that the series does not converge.

Exercise 3.11c

To prove the inequality:

s

n

≥ s

n−1

¦s

n

¦ is an increasing sequence (see part b)

→

1

sn

≤

1

sn−1

→

an

s

2

n

≤

an

snsn−1

multiply both sides by a

n

/s

n

→

an

s

2

n

≤

sn−sn−1

snsn−1

s

n

−s

n−1

= a

n

(see part b)

→

an

s

2

n

≤

1

sn−1

−

1

sn

algebra

Now consider the summation

n

k=2

1

s

k−1

−

1

s

k

=

_

1

s

1

−

1

s

2

_

+

_

1

s

2

−

1

s

3

_

+. . . +

_

1

s

n−1

−

1

s

n

_

38

Most of the terms in this summation cancel one another out: the summation “telescopes down” and simpliﬁes

to

n

k=2

1

s

k−1

−

1

s

k

=

1

s

1

−

1

s

k

As we saw in part (b), the terms of ¦s

n

¦ increase without bound as n → ∞, so that

∞

k=2

1

s

k−1

−

1

s

k

=

1

s

1

Now, by the inequality we proved above, we know that

∞

k=2

a

n

s

2

n

≤

∞

k=2

1

s

k−1

−

1

s

k

=

1

s

1

We can add one term to each side so that our summation starts at 1 instead of at 2 to get

∞

k=1

a

n

s

2

n

≤

a

1

s

2

1

+

1

s

1

We can now show that the series converges. Deﬁne the partial sum of this series to be

¦t

n

¦ =

n

k=1

1

s

k−1

−

1

s

k

We’ve shown that ¦t

n

¦ is bounded above by a

1

/s

2

1

+ 1/s

1

, and we know that ¦t

n

¦ is monotonically increasing

because each a

n

is positive. Therefore, by theorem 3.14 we know that ¦t

n

¦ converges; and so by deﬁnition 3.21

we know that its associated series converges. And this is what we were asked to prove.

Exercise 3.11d

The series

a

n

/(1 + n

2

a

n

) always converges. From the fact that a

n

> 0, we can establish the following chain

of inequalities:

a

n

1 +n

2

a

n

=

1/a

n

1/a

n

a

n

1 +n

2

a

n

=

1

1

an

+n

2

<

1

n

2

From this, we see that

∞

n=0

a

n

1 +n

2

a

n

<

∞

n=0

1

n

2

We know that

1/n

2

converges (theorem 3.28), and therefore

a

n

/(1 + n

2

a

n

) converges by the comparison

test of theorem 3.25.

The series

a

n

/(1 +na

n

) may or may not converge. If a

n

= 1/n, for instance, the summation becomes

a

n

1 +na

n

=

1

n

2

=

1

2

1

n

which is divergent by theorem 3.28.

To construct a convergent series, let a

n

be deﬁned as

a

n

=

_

1 if n = 2

m

−1 (m ∈ Z)

0 otherwise

The series

a

n

is divergent, since there are inﬁnitely many integers of the form 2

m

− 1. But the series

a

n

/(1 +na

n

) is convergent:

∞

n=0

a

n

1 +na

n

=

∞

m=0

1

2

m

=

_

1

2

_

m

This series is convergent to 2 by theorem 3.26.

39

Exercise 3.12a

establishing the inequality

From the deﬁnition of r

n

, we see that

r

k

=

∞

m=k

a

m

= a

k

+

∞

m=k+1

a

m

= a

k

+r

k+1

so that a

k

= r

k

−r

k+1

. And because each a

k

is positive, we see that r

k

> r

k+1

which (from transitivity) means

that r

m

> r

n

when n > m (that is, ¦r

n

¦ is a decreasing sequence). With these equalities, we can form our proof.

Take m ≤ n and consider the sum

n

k=m

a

k

r

k

=

a

m

r

m

+. . . +

a

n

r

n

From the fact that r

m

≥ r

n

, we know that a

k

/r

m

≤ a

k

/r

n

for all k, so that

a

m

+. . . +a

n

r

m

≤

a

m

r

m

+. . . +

a

n

r

n

≤

a

m

+. . . +a

n

r

n

Now, from the fact that a

k

= r

k

−r

k+1

, we see that this inequality is equivalent to

(r

m

−r

m+1

) + (r

m+1

−r

m+2

) +. . . + (r

n

−r

n+1

)

r

m

≤

a

m

r

m

+. . .+

a

n

r

n

≤

(r

m

−r

m+1

) + (r

m+1

−r

m+2

) +. . . + (r

n

−r

n+1

)

r

n

Notice that most of the terms of these numerators cancel one another out, leaving us with

r

m

−r

n+1

r

m

≤

a

m

r

m

+. . . +

a

n

r

n

≤

r

m

−r

n+1

r

n

Taking the two leftmost terms of this inequality and performing some simple algebra then gives us

1 −

r

n+1

r

m

≤

a

m

r

m

+. . . +

a

n

r

n

This is close to what we want to prove, but our index is oﬀ by one. But each a

n

is positive and each r

n

is

the sum of positive terms, so a

n+1

/r

n+1

is strictly positive. By adding this term to the right-hand side of the

inequality, we accomplish two things: we correct our index and we make this a strict inequality (<) instead of

a non-strict inequality (≤).

1 −

r

n+1

r

m

<

a

m

r

m

+. . . +

a

n

r

n

+

a

n+1

r

n+1

This last statement is true whenever n ≥ m, which means that it’s true whenever n + 1 > m. By a simple

replacement of variables, then, we know that

1 −

r

n

r

m

<

a

m

r

m

+. . . +

a

n

r

n

whenever n

**> m, which is what we were asked to prove.
**

proof of divergence

We’re told that

a

n

converges, so

a

n

= α for some α ∈ R. From the deﬁnition of r

n

, we see that there is a

relationship between ¦r

n

¦ and α:

r

n

=

∞

k=n

a

k

=

∞

k=1

a

k

−

n−1

k=1

a

k

= α −

n−1

k=1

a

k

As n → ∞, this last term approaches α −α = 0, and so we see that lim

n→∞

r

n

= 0.

40

Now choose any arbitrarly small from the interval (0,

1

2

) and choose any arbitrary integer N. From the

inequality we veriﬁed in part (a1), we see that

a

N

r

N

+. . . +

a

N+n

r

N+n

> 1 −

r

N+n

r

N

We can choose n to be arbitrarily large; in doing so, r

N+n

approaches zero while r

N

remains ﬁxed. So, by

choosing a suﬃciently large n, we can guarantee that r

N+n

/r

N

< . This choice of n gives us the inequality

a

N

r

N

+. . . +

a

N+n

r

N+n

> 1 −

r

N+n

r

N

> 1 − >

where the last step of this chain of inequalities is justiﬁed by our choice of 0 < <

1

2

.

We’ve shown that there is some such that, for every N, we can ﬁnd

n

k=N

a

n

r

n

>

From theorem 3.22, this is suﬃcient to show that

a

n

/r

n

diverges (since we’ve proven the negation of the “for

every > 0 there is some N . . . ” statement of the theorem).

Exercise 3.12b

r

n

> r

n+1

> 0 established in part (a)

→

√

r

n

>

√

r

n+1

take square root of each side

→ 2

√

r

n

>

√

r

n

+

√

r

n+1

add

√

r

n

to each side

→ 2 >

√

rn+

√

rn+1

√

rn

divide by

√

r

n

→ 2(

√

r

n

−

√

r

n+1

) >

(

√

rn+

√

rn+1)(

√

rn−

√

rn+1)

√

rn

multiply each side by a positive term

→ 2(

√

r

n

−

√

r

n+1

) >

rn−rn+1

√

rn

simply the right-hand side

→ 2(

√

r

n

−

√

r

n+1

) >

an

√

rn

we established a

n

= r

n

−r

n+1

in part (a)

Having established this inequality, we see that

n

k=1

a

k

√

r

k

<

n

k=1

2(

√

r

k

−

√

r

k+1

)

and many of the terms in the right-hand summation cancel one another out, so that the series “telescopes down”

and simpliﬁes to

n

k=1

a

k

√

r

k

< 2

_

√

r

1

−

√

r

k+1

_

so that

lim

n→∞

n

k=1

a

k

√

r

k

< lim

n→∞

2

_

√

r

1

−

√

r

k+1

_

= 2

√

r

1

This shows that this sum of nonnegative terms is bounded, which is suﬃcient to show that it converges by

theorem 3.24.

Exercise 3.13

Let

a

n

be a series that converges absolutely to α and let

b

n

be a series that converges absolutely to β. We

can prove that the Cauchy product of these two series (deﬁnition 3.48) is bounded.

41

→

n

k=0

[c

k

[

=

n

k=0

¸

¸

¸

k

j=0

a

j

b

k−j

¸

¸

¸ deﬁnition 3.48 of c

k

≤

n

k=0

k

j=0

[a

j

b

k−j

[ triangle inequality

=

n

k=0

k

j=0

[a

j

[[b

k−j

[ 1.33(c)

We can expand this summation out to get:

= [a

0

[[b

0

[ + ([a

0

[[b

1

[ +[a

1

[[b

0

[) +. . . + ([a

0

[[b

n

[ +[a

1

[[b

n−1

[ +. . . +[a

n

[[b

0

[)

Deﬁne B

n

to be

n

k=0

[b

k

[ and A

n

to be

n

k=0

[a

k

[. We can rearrange these terms to get:

= [a

0

[B

n

+[a

1

[B

n−1

+[a

2

[B

n−2

+. . . +[a

n

[B

0

Now we can add several nonnegative terms to get

≤ [a

0

[B

n

+[a

1

[(B

n−1

+[b

n

[) +[a

2

[(B

n−2

+[b

n−1

[ +[b

n

[) +. . . +[a

n

[(B

0

+[b

1

[ +. . . +[b

n

[)

= [a

0

[B

n

+[a

1

[B

n

+[a

2

[B

n

+. . . +[a

n

[B

n

= A

n

B

n

= (

n

k=0

[a

k

[) (

n

k=0

[b

k

[)

= αβ

This shows that each partial sum of

[c

k

[ is bounded above by αβ and below by 0 (since it’s a series of

nonnegative terms). By theorem 3.24, this is suﬃcient to prove that

[c

k

[ converges; and since

c

k

is the

Cauchy product, we have proved that the Cauchy product

c

k

converges absolutely.

Exercise 3.14a

Choose any arbitrarily small > 0. We’re told that lim¦s

n

¦ = s, which means that there is some N such that

d(s, s

n

) < whenever n > N. So we’ll rearrange the terms of σ

n

a bit:

σ

n

=

n

k=0

s

k

n + 1

=

N

k=0

s

k

+

n

k=N+1

s

k

n + 1

Whenever n > N we know that d(s, s

n

) < , which means that − < s

n

−s < , or that s − < s

n

< s +. This

gives us the inequality

N

k=0

s

k

+

n

k=N+1

(s −)

n + 1

< σ

n

<

N

k=0

s

k

+

n

k=N+1

(s +)

n + 1

The terms in some of these summations don’t depend on k, so we can further rewrite this as

N

k=0

s

k

+ (n −(N + 1))(s −)

n + 1

< σ

n

<

N

k=0

s

k

+ (n −(N + 1))(s +)

n + 1

N

k=0

s

k

n + 1

+

n(s −)

n + 1

−

(N + 1)(s −)

n + 1

< σ

n

<

N

k=0

s

k

n + 1

+

n(s +)

n + 1

−

(N + 1)(s +)

n + 1

For many of these terms, the numerators are constant with respect to n. Therefore many of these terms will

approach zero as n → ∞.

lim

n→∞

n(s −)

n + 1

< lim

n→∞

σ

n

< lim

n→∞

n(s +)

n + 1

s − < lim

n→∞

σ

n

< s +

And this is just another way of saying d(s, limσ

n

) < . And was arbitrarily small, so we have shown that

lim

n→∞

σ

n

= s.

42

Exercise 3.14b

Let s

n

= (−1)

n

. Then

s

n

is either equal to 0 or 1, depending on n, and

0

n + 1

≤ σ

n

≤

1

n + 1

Taking the limit as n → ∞ gives us

0 ≤ lim

n→∞

σ

n

≤ 0

which can only be true if

lim

n→∞

σ

n

= 0

Exercise 3.14c

Deﬁne s

n

to be

s

n

=

_ _

1

2

_

n

+k if n = k

3

(k ∈ Z)

_

1

2

_

n

otherwise

There are no more than

3

√

n perfect cubes within the ﬁrst n integers, so ¦s

n

¦ will contain no more than ¸

3

√

n|

terms of the form (1/2)

n

+k. This gives us the inequality

n

k=0

s

n

=

n

k=0

_

1

2

_

n

+

3

√

n

k=0

k ≤ 2 +

3

√

n(

3

√

n + 1)

2

where the last step in this chain of inequalities is justiﬁed by the common summation

n

k=1

k = k(k + 1)/2.

Continuing, we see that

n

k=0

s

n

≤ 2 +

3

√

n(

3

√

n + 1)

2

≤ 2 +

3

√

n(

3

√

n +

3

√

n)

2

= 2 +

3

√

n

2

We can now analyze the value of σ

n

.

σ

n

=

n

k=0

s

n

n + 1

≤

2 +

3

√

n

2

n + 1

=

2

3

√

n

2

+ 1

3

√

n +

1

3

√

n

2

Taking the limit of each side as n → ∞ gives us

lim

n→∞

σ

n

≤ lim

n→∞

2

3

√

n

2

+ 1

3

√

n +

1

3

√

n

2

= lim

n→∞

1

3

√

n

= 0

Every term of ¦s

n

¦ was greater than zero, so we know that the arithmetic average σ

n

is greater than zero.

Therefore 0 ≤ limσ

n

≤ 0, and therefore limσ

n

= 0.

Exercise 3.14

proving the equality

n

k=1

ka

k

= (s

1

−s

0

) + 2(s

2

−s

1

) +. . . +n(s

n

−s

n−1

)

= −s

0

+ (1 −2)s

1

+ (2 −3)s

2

+. . . + ((n −1) −n)s

n−1

+s

n

(n)

= −s

0

−s

1

−. . . −s

n

+ (n + 1)s

n

Therefore, if we divide this by n + 1, we get

1

n + 1

n

k=1

ka

k

=

−s

0

−s

1

−. . . −s

n

n + 1

+s

n

= −σ

n

+s

n

43

establishing convergence

We’re told that lim

n→∞

na

n

= 0, so by part a we know that

lim

n→∞

n

k=1

ka

k

n + 1

= 0

We’re also told that ¦σ

n

¦ converges to some value σ, so by theorem 3.3(a) we know that

lim

n→∞

_

n

k=1

ka

k

n + 1

+σ

n

_

= (0 +σ) = σ

And since s

n

=

n

k=1

ka

k

n+1

+σ

n

, this is suﬃcient to prove that lim

n→∞

s

n

= σ.

Exercise 3.15

If you think I’m going to go through this tedious exercise, you can lick me where I shit.

Exercise 3.16a

¦x

n

¦ is monotonically decreasing

From the fact that 0 <

√

α < x

n

we know that α < x

2

n

, and therefore

x

n+1

=

1

2

_

x

n

+

α

x

n

_

<

1

2

_

x

n

+

x

2

n

x

n

_

= x

n

This shows that x

n

> x

n+1

for all n, which proves that ¦x

n

¦ is monotonically decreasing.

The limit of ¦x

n

¦ exists and is not less than

√

a

First, we show that ¦x

n

¦ is bounded below by

√

a. We know that x

0

>

√

a because we chose it to be. And if

x

n

>

√

a for any n, we have

→ x

n

,=

√

a assumed

→ x

n

−

√

a ,= 0

→ (x

n

−

√

a)

2

≥ 0 1.18(d)

→ x

2

n

−2x

n

√

a +a ≥ 0

→ x

2

n

+a ≥ 2x

n

√

a

→

x

2

n

+a

2xn

≥

√

a

→

1

2

_

x

n

+

a

xn

_

≥

√

a

→ x

n+1

≥

√

a deﬁnition of x

n+1

So by induction, we know that x

n

≥

√

a for all n. We’ve now demonstrated that ¦x

n

¦ is monotonically

decreasing and is bounded below by

√

a, so we’re guaranteed that the limit of ¦x

n

¦ exists and that lim¦x

n

¦ ≥

√

a.

The limit of ¦x

n

¦ is not greater than

√

a

Now we show that lim¦x

n

¦ ≤

√

a. Assume that lim¦x

n

¦ =

√

b for some

√

b >

√

a. By the deﬁnition of “limit”,

we can ﬁnd N such that n > N implies d(x

n

,

√

b) < for any arbitrarily small . We’ll choose =

√

b −a (which

is positive since

√

b >

√

a ≥ 1 implies b > a).

→ d(x

n

,

√

b) <

→ d(x

n

,

√

b) <

√

b −a chosen value of

→ x

n

−

√

b <

√

b −a metric on R

1

→ x

n

<

√

b −a +

√

b

44

We can then calculate x

n+1

in terms of x

n

to get a chain of inequalities:

x

n+1

=

x

2

n

+a

2x

n

<

(

√

b −a +

√

b)

2

+a

2(

√

b −a +

√

b)

=

(b −a) + 2

√

b

√

b −a +b +a

2(

√

b −a +

√

b)

=

2b + 2

√

b

√

b −a

2(

√

b −a +

√

b)

=

√

b(

√

b +

√

b −a)

√

b +

√

b −a

=

√

b

This shows us that if the distance between x

n

and

√

b becomes less than

√

b −a (which it is guaranteed to

do eventually because lim¦x

n

¦ =

√

b), then the next term in the sequence ¦x

n

¦ will be less than

√

b. And since

we’ve already shown that the sequence is monotonically decreasing, this means that every subsequent term will

be even farther from

√

b, which contradicts our assumption that lim¦x

n

¦ =

√

b. Therefore the limit is not

√

b:

but

√

b was an arbitrary value greater than

√

a, so we’ve shown that the limit cannot be any value greater than

√

a.

Having shown that lim¦x

n

¦ exists, that the limit is not less than

√

a, and that the limit is not greater than

√

a, we have therefore proven that lim¦x

n

¦ =

√

a.

Exercise 3.16b

From the deﬁnition of x

n+1

and e

n

, we have

e

n+1

= x

n+1

−

√

a =

x

2

n

+

√

a

2x

n

−

√

a =

x

2

n

−2x

n

√

a +a

2x

n

=

(x

n

−

√

a)

2

2x

n

=

e

2

n

2x

n

<

e

2

n

2

√

a

where the last step is justiﬁed by the fact that x

n

>

√

a (see part (a)).

The next part can be proven by induction. Setting n = 1 in the above inequality, we have

e

2

<

e

2

1

2

√

a

=

e

2

1

β

= β

_

e

1

β

_

2

1

And, if the statement is true for n, then

e

n+1

<

e

2

n

2

√

a

=

1

β

(e

2

n

) <

1

β

_

β

_

e

1

β

_

2

n−1

_

2

=

1

β

_

β

2

_

e

1

β

_

2

n

_

= β

_

e

1

β

_

2

n

which shows that it is therefore true for n + 1. By induction, this is suﬃcient to show that it is true for all

n ∈ N.

Exercise 3.16c

When a = and x

1

= 2, we have

e

1

β

=

x

1

−

√

a

2

√

a

=

2 −

√

3

2

√

3

which can be shown to be less than 1/10 through simple algebra. Then, by part (b), we have

e

n

< β

_

e

1

β

_

2

n

< 2

√

3

_

1

10

_

2

n

< 4

_

1

10

_

2

n

45

Exercise 3.17

Lemma 1: x

n

>

√

a → x

n+1

<

√

a, and x

n

<

√

a → x

n+1

>

√

a

→ x

n

>

√

a

↔ x

n

(

√

a −1) >

√

a(

√

a −1)

↔ x

n

√

a −x

n

> a −

√

a

↔ x

n

√

a +

√

a > a +x

n

↔

√

a(x

n

+ 1) > a +x

n

↔

√

a >

a+xn

xn+1

↔

√

a > x

n+1

deﬁnition of x

n+1

This shows that x

n

>

√

a → x

n+1

<

√

a. If “>” is replaced with “<” in each of the above steps, we will

have also constructed a proof that x

n

<

√

a → x

n+1

>

√

a.

Lemma 2: x

n

>

√

a → x

n+2

< x

n

and x

n

<

√

a → x

n+2

> x

n

→ x

n

>

√

a

↔ x

n

−

√

a > 0

↔ 2(x

n

−

√

a)(x

n

+

√

a) > 0 multiplied by positive terms, so it’s still > 0

↔ 2(x

2

n

−a) > 0

At this point, the algebraic steps become bizarre and seemingly nonsensical. But bear with me.

↔ 2x

2

n

+ 0x

n

−2a) > 0

↔ 2x

2

n

+ (1 +a −1 −a)x

n

−2a) > 0

↔ x

n

+x

2

n

+ax

n

+x

2

n

> a +ax

n

+a +x

n

↔ x

n

(1 +x

n

) +x

n

(a +x

n

) > a(1 +x

n

) +a +x

n

↔

xn(1+xn)+xn(a+xn)

1+xn

>

a(1+xn)+a+xn

1+xn

division by a positive term

↔ x

n

_

1 +

a+xn

1+xn

_

> a +

a+xn

1+xn

↔ x

n

(1 +x

n+1

) > a +x

n+1

↔ x

n

>

a+xn+1

1+xn+1

↔ x

n

> x

n+2

deﬁnition of x

n+2

This shows that x

n

>

√

a → x

n+2

< x

n

. If “>” is replaced with “<” in each of the above steps, we will have

also constructed a proof that x

n

<

√

a → x

n+2

> x

n

.

Exercise 3.17a

We’re forced to choose x

1

such that x

1

>

√

a, so we can use induction with lemma 2 to show than ¦x

2k+1

¦ (the

subsequence of ¦x

n

¦ consisting of elements with odd indices) is a monotonically decreasing sequence.

Exercise 3.17b

Because we chose x

1

such that x

1

>

√

a, lemma 1 tells us that x

2

<

√

a. We can then use induction with lemma

2 to show that ¦x

2k

¦ (the subsequence of ¦x

n

¦ consisting of elements with even indices) is a monotonically

increasing sequence.

Exercise 3.17c

We can show than ¦x

n

¦ converges to

√

a by showing that we can make d(x

n

,

√

a) arbitrarily small. We do this

by demonstrating the relationship between d(x

n

,

√

a) and d(x

1

,

√

a).

46

e

n+2

= x

n+2

−

√

a =

a +x

n+1

1 +x

n+1

−

√

a =

a +x

n+1

−

√

a −x

n+1

√

a

1 +x

n+1

→ e

n+2

= x

n+2

−

√

a

→ e

n+2

=

a+xn+1

1+xn+1

−

√

a

→ e

n+2

=

a+xn+1−

√

a−xn+1

√

a

1+xn+1

→ e

n+2

(1 +x

n+1

) = a +x

n+1

−

√

a −x

n+1

√

a

→ e

n+2

(1 +x

n+1

) = x

n+1

(1 −

√

a) −

√

a(1 −

√

a)

→ e

n+2

(1 +x

n+1

) = (x

n+1

−

√

a)(1 −

√

a)

→ e

n+2

(1 +x

n+1

) = e

n+1

(1 −

√

a)

→ e

n+2

= e

n+1

_

1−

√

a

1+xn+1

_

This tells us how to express e

n+2

in terms of e

n+1

. We can use this same equality to express e

n+1

in terms

of e

n

, giving us

e

n+2

= e

n+1

_

1 −

√

a

1 +x

n+1

_

= e

n

_

1 −

√

a

1 +x

n

__

1 −

√

a

1 +x

n+1

_

= e

n

_

1 −

√

a

1 +x

n

_

_

1 −

√

a

1 +

a+xn

1+xn

_

= e

n

_

1 −

√

a

1 +x

n

_

_

1 −

√

a

1+xn+a+xn

1+xn

_

= e

n

_

1 −

√

a

1 +x

n

__

(1 +x

n

)(1 −

√

a)

1 + 2x

n

+a

_

= e

n

_

(1 −

√

a)

2

1 + 2x

n

+a

_

< e

n

_

(

√

a −1)

2

a −1

_

The last step in this chain of inequalities is justiﬁed by the fact that a + 2x

n

+ 1 > a − 1 > 1. Continuing, we

have

e

n+2

< e

n

_

(

√

a −1)

2

a −1

_

= e

n

_√

a −1

√

a + 1

_

= ce

n

, 0 < c < 1

This tells us how to express e

n+2

in terms of e

n

. We can use this same inequality to express e

2n+2

in terms e

2

and e

2n+1

in terms of e

1

:

e

2n+2

< ce

2n

< c(ce

2n−2

) < . . . < c

n

e

2

e

2n+1

< ce

2n−1

< c(ce

2n−3

) < . . . < c

n

e

1

And since 0 < c < 1, this means that we can make c

n

arbitrarily small by taking suﬃciently large n; and

e

n

< c

2

[maxe

1

, e

2

], so we can make e

n

arbitrarily small by taking suﬃciently large n; and this shows that

lim

n→∞

e

n

= 0. Finally, remember that we deﬁned e

n

to be d(x

n

,

√

a). So lim

n→∞

d(x

n

,

√

a) = 0 which is

suﬃcient to prove that lim

n→∞

¦x

n

¦ =

√

a.

Exercise 3.17d

As shown above, every two iterations reduces the error term by a factor of (

√

a−1)/(

√

a+1) (linear convergence).

For the algorithm in exercise 16, the nth iteration reduced the error term by a factor of 10

−2

n

(quadratic

convergence).

Exercise 3.18

Lemma 1 : ¦x

n

¦ is decreasing.

We’re asked to choose that x

1

>

p

√

a. And if x

n

>

p

√

a, we have

47

→ x

n

>

p

√

a

→ x

p

n

> a

→ 0 > a −x

p

n

→ px

p

n

> px

p

n

+a −x

p

n

→ px

p

n

> (p −1)x

p

n

+a

→

px

p

n

px

p−1

n

>

(p−1)x

p

n

+a

px

p−1

n

→ x

n

>

(p−1)xn

p

+

a

p

x

p−1

n

→ x

n

> x

n+1

Lemma 2 : 0 < k < 1 → p(1 −k) > 1 −k

p

Let k be a positive number less than 1 and let p be a positive integer.

→ p = 1

0

+ 1

2

+ 1

3

+. . . + 1

p−1

→ p > k +k

2

+k

3

+. . . +k

p−1

→ p >

(k−1)(k+k

2

+k

3

+...+k

p−1

)

k−1

→ p >

k

p

−1

k−1

→ p(k −1) > k

p

−1

Lemma 3 : x

n

>

p

√

a

We know that x

1

>

p

√

a because we chose it. And if x

n

>

p

√

a, we have:

→ x

n

>

p

√

a > 0

→ 1 >

p

√

a

xn

> 0

→ p

_

1 −

p

√

a

xn

_

> 1 −

_

p

√

a

xn

_

p

from lemma 2

→ p −1 > p

p

√

a

xn

−

a

x

p

n

expand and rearrange the terms

→ x

p

n

(p −1) > x

p

n

_

p

p

√

a

xn

−

a

x

p

n

_

multiply both sides by x

p

n

→ px

p

n

−x

p

n

> p

p

√

ax

p−1

n

−a

→ (p −1)x

p

n

+a > p

p

√

ax

p−1

n

rearrange the terms and simplify

→

(p−1)x

p

n

+a

px

p−1

n

>

p

p

√

ax

p−1

n

px

p−1

n

divide both sides by px

p−1

n

→

(p−1)xn

p

+

a

px

p−1

n

>

p

√

a simplify

→ x

n+1

>

p

√

a deﬁnition of x

n+1

Finally : lim

n→∞

¦x

n

¦ =

p

√

a

We’ve shown that ¦x

n

¦ is decreasing (lemma 1) and that it’s bounded below (lemma 3), which is suﬃcient

to show that it has some limit x

∗

with x∗ ≥

p

√

a. From the deﬁnition of limit, we can ﬁnd N such that

n > N → d(x

n

, x

∗

) <

2p

. From this, we see that we can ﬁnd x

n

and x

n+1

such that:

48

→ d(x

n

, x

∗

) < ∧ d(x

n+1

, x

∗

) <

2p

→ d(x

n

, x

n+1

) < d(x

n

, x

∗

) +d(x

∗

, x

n+1

) =

p

triangle inequality

→ x

n

−x

n+1

<

p

lemma 1: x

n

> x

n+1

→ x

n

−

_

x

n

−

xn

p

+

a

px

p−1

n

_

<

p

deﬁnition of x

n+1

→

xn

p

−

a

px

p−1

n

<

p

→

x

p

n

−a

px

p−1

n

<

p

→ x

n

−

a

x

p−1

n

<

This last statement tells us that lim

n→∞

x

n

−

a

x

p−1

n

= 0, or that

→ lim

n→∞

x

n

−

a

x

p−1

n

= 0 assumed

→ lim

n→∞

x

n

(1 −a/x

p

n

) = 0 theorem 3.3c

→ x

∗

_

1 −

_

p

√

a

x

∗

_

p

_

= 0 theorem 3.3c

→

p

√

a/x

∗

= 1

→

p

√

a = x

∗

From the deﬁnition of x

∗

as the limit of ¦x

n

¦, this tells us that lim

n→∞

¦x

n

¦ =

p

√

a which is what we wanted

to prove.

Exercise 3.19

The idea behind this proof is to consider the elements of the line segment [0, 1] in the form of their base-3

expansion. When we do this, we notice that the mth iteration of the Cantor set eliminates every number with

a 1 as the mth digit of its ternary decimal expansion.

From equation 3.24 in the book, we know that a real number r is not in the Cantor set iﬀ it is in any interval

of the form

_

3k + 1

3

m

,

3k + 2

3

m

_

m, k ∈ N

Let ¦α

n

¦ represent an arbitrary ternary sequence (that is, each digit is either 0,1, or 2). We can determine the

necessary and suﬃcient conditions for x(α) to fall into such an interval:

x(α) ,∈ Cantor iﬀ x(α) ∈

_

3k + 1

3

m

,

3k + 2

3

m

_

iﬀ (∃m)

∞

n=1

α

n

3

n

∈

_

3k + 1

3

m

,

3k + 2

3

m

_

iﬀ (∃m)

∞

n=1

α

n

3

n−m

∈ (3k + 1, 3k + 2)

We then split up the summation into three distinct parts.

iﬀ (∃m)

m−1

n=1

α

n

3

n−m

+

m

n=m

α

n

3

n−m

+

∞

n=m+1

α

n

3

n−m

∈ (3k + 1, 3k + 2)

iﬀ (∃m)

m−1

n=1

α

n

3

m−n

+α

m

+

∞

n=m+1

α

n

3

n−m

∈ (3k + 1, 3k + 2)

Every term in the leftmost sum is divisible by three, so the leftmost sum is itself divisible by three. This gives

us, for some j ∈ N,

iﬀ (∃m) 3j +α

m

+

∞

n=m+1

α

n

3

n−m

∈ (3k + 1, 3k + 2), j ∈ N

49

Without specifying a certain sequence ¦α

n

¦, we can’t evaluate the rightmost summation. But we can establish

bounds for it. Each α

n

is either 0,1, or 2. So the summation is largest whenever α

n

= 2 for all n, and it’s

smallest when α

n

= 0 for all n. This gives us the bound

0 =

∞

n=m+1

0

3

n−m

<

∞

n=m+1

α

n

3

n−m

<

∞

n=m+1

2

3

n−m

= 2

∞

n=1

1

3

n

= 1

Note that the upper bound is 1, not 3, since the index is from 1 instead of 0. Continuing with our chain of “iﬀ”

statements, we conclude that x(a) is not member of the Cantor set iﬀ:

x(a) ,∈ Cantor iﬀ (∃m) 3j +α

m

+δ ∈ (3k + 1, 3k + 2), j ∈ N, 0 < δ < 1

From the bounds on α

m

and δ, we know that 0 < α

m

+ δ < 3: their sum is never a multiple of 3. So the only

way that the set membership in the previous statement can be true is

iﬀ (∃m) α

m

+δ ∈ (1, 2), 0 < δ < 1

iﬀ (∃m) α

m

∈ (0, 2)

We know that α

m

is an integer, and there’s only one integer in the open interval (0, 2):

iﬀ (∃m) α

m

= 1

iﬀ x(α) ,∈ x(a)

where x(a) is as deﬁned in the exercise. This shows that any real number x(α) is a member of the Cantor set if

and only if it is a member of x(a): this is suﬃcient to prove that the Cantor set is equal to x(a).

Exercise 3.20

Choose some arbitrarily small . From the deﬁnition of Cauchy convergence, there is some N such that i, n > N

implies d(p

i

, p

n

) < /2. From the convergence of the subsequence to p

∗

, there is some M such that i > M

implies d(p

∗

, pi) < /2.

So, for i, n > max(M, N) we have

d(p

∗

, p

n

) ≤ d(p

∗

, p

i

) +d(p

i

, p

n

) =

which, because is arbitrarily small, is suﬃcient to prove that ¦p

n

¦ converges to p

∗

.

Exercise 3.21

We know that each E

i

∈

E

n

is nonempty (see note 1), so we can construct at least one sequence whose ith

element is an arbitrary element of E

i

. Let ¦s

n

¦ be an arbitrary sequence constructed in this way. We can

immediately say three things about this sequence.

1) The sequence is Cauchy. This comes from the fact that lim diam E

n

= 0: see the text below deﬁnition 3.9.

2) The sequence is convergent. This comes from the fact that it’s a sequence in a complete metric space X: see

deﬁnition 3.12.

3) The sequence converges to some s∗ ∈

E

n

. The set

E

n

is a intersection of closed sets, and is therefore

closed (thm 2.24b), so any limit point of

E

n

is a point of

E

n

, and therefore s

∗

(being a limit point of

E

n

) is an element of

E

n

.

This shows that

E

n

is nonempty: it contains at least one element s

∗

. We can then follow the proof of theorem

3.10b to conclude that

E

n

contains only one point.

50

note 1

We’re told that E

i

⊃ E

i+1

for all i. If E

i

were empty, then E

k

would be empty for all k > i. This would mean

that

limdiam E

n

= diam∅ = diam sup ¦d(p, q) : p, q ∈ ∅¦ = sup ∅

To ﬁnd the supremum of the empty set in R, we need to rely on the deﬁnition of supremum:

sup ∅ = least upper bound of ∅ in R = min ¦x ∈ R : a ∈ ∅ → a ≤ x¦

The set for which we’re seeking a minimum contains every x ∈ R (because of false antecedent a ∈ ∅), so the

supremum of the empty set in R is the minimum of R itself. Regardless of how we deﬁne this minimum (or if

it’s deﬁned at all), it certainly isn’t equal to zero. But we’re told that

limdiam E

n

= 0

so our initial assumption must have been wrong: there is no empty E

i

∈

E

n

.

Exercise 3.22

The set G

1

is an open set. Choose p

1

∈ G

1

and ﬁnd some open neighborhood N

r1

(p

1

) ⊆ G

1

. Choose a smaller

neighborhood N

s1

(p

1

). Not only do we have N

s1

⊂ N

r1

⊆ G

1

, we can take the closure of N

s1

and still have

N

s1

⊂ N

r1

⊆ G

1

(since δ ≤ s

1

< r

1

). Deﬁne E

1

to be the neighborhood N

s1

(not its closure). At this point, we

have

E

1

⊂ E

1

⊆ G

1

The set G

2

is dense in X, so p

1

is either an element of G

2

or is a limit point for G

2

. In either case, every

neighborhood of p

1

contains some point p

2

∈ G

2

. More speciﬁcally, the neighborhood E

1

contains some point

p

2

∈ G

2

. Since p

2

is in both E

1

and G

2

, both of which are open sets, there is some neighborhood N

r2

(p

2

) that’s a

subset of both E

1

and G

2

. We now choose an even smaller radius N

s2

(p

2

). Not only do we have N

s2

⊂ N

r2

⊆ G

2

and N

s2

⊂ N

r2

⊆ E

1

, we could take the closure N

s2

. Deﬁne E

2

to be the neighborhood N

s2

(not its closure).

At this point, we have

E

2

⊂ E

2

⊂ G

2

, E

2

⊂ E

2

⊂ E

1

⊂ E

1

We can continue in this same way, constructing a series of nested sets

E

1

⊃ E

2

⊃ ⊃ E

n

This is a series of closed, bounded, nonempty, nested sets. So (from exercise 21) we know that

E

n

contains a

single point x. This single point x must be in every E

i

∈

E

n

, and E

i

⊂ G

i

, and x must be in every G

i

∈

G

n

.

Therefore

G

n

is nonempty.

Exercise 3.23

We’re aren’t given enough information about the metric space X to assume that the Cauchy sequences converge

to elements in X. All we can say is that, for any arbitrarily small , there exists some M, N ∈ R such that

n, m > M → d(p

n

, p

m

) ≤ /2

n, m > N → d(q

n

, q

m

) ≤ /2

From multiple applications of the triangle inequality, we have

d(p

n

, q

n

) ≤ d(p

n

, p

m

) +d(p

m

, q

n

) ≤ d(p

n

, p

m

) +d(p

m

, q

m

) +d(q

m

, q

n

)

which, for n, m > max¦N, M¦, becomes

d(p

n

, q

n

) ≤ +d(p

m

, q

m

)

or

d(p

n

, q

n

) −d(p

m

, q

m

) ≤

51

Exercise 3.24a

reﬂexivity

For all sequences ¦p

n

¦, we have

lim

n→∞

d(p

n

, p

n

) = lim

n→∞

[p

n

−p

n

[ = lim

n→∞

0 = 0

so ¦p

n

¦ = ¦p

n

¦.

symmetry

If ¦p

n

¦ = ¦q

n

¦, we have

lim

n→∞

d(p

n

, q

n

) = lim

n→∞

[p

n

−q

n

[ = lim

n→∞

[q

n

−p

n

[ = lim

n→∞

d(q

n

, p

n

)

so ¦q

n

¦ = ¦p

n

¦.

transitivity

If ¦p

n

¦ = ¦q

n

¦ and ¦q

n

¦ = ¦r

n

¦, the triangle inequality gives us

lim

n→∞

d(p

n

, r

n

) ≤ lim

n→∞

d(p

n

, q

n

) +d(q

n

, r

n

) = 0 + 0

This tells us that lim

n→∞

d(p

n

, r

n

) = 0, so ¦p

n

¦ = ¦r

n

¦.

Exercise 3.24b

Let ¦a

n

¦ = ¦p

n

¦ and let ¦b

n

¦ = ¦q

n

¦. From the triangle inequality, we have

∆(P, Q) = lim

n→∞

d(p

n

, q

n

) ≤ lim

n→∞

d(p

n

, a

n

) +d(a

n

, b

n

) +d(b

n

, q

n

)

Which, from the deﬁnition of equality established in part (a), gives us

∆(P, Q) = lim

n→∞

d(p

n

, q

n

) ≤ lim

n→∞

d(a

n

, b

n

) (5)

A similar application of the triangle inequality gives us

lim

n→∞

d(a

n

, b

n

) ≤ lim

n→∞

d(a

n

, p

n

) +d(p

n

, q

n

) +d(q

n

, b

n

)

Which, from the deﬁnition of equality established in part (a), gives us

lim

n→∞

d(a

n

, b

n

) ≤ lim

n→∞

d(p

n

, q

n

) (6)

Combining equations (5) and (6), we have

∆(P, Q) = lim

n→∞

d(p

n

, q

n

) = lim

n→∞

d(a

n

, b

n

)

Exercise 3.24c

Deﬁne X

∗

to be the set of equivalence classes from part (b). Let ¦P

n

¦ be a Cauchy sequence in (X

∗

, ∆). This is

an unusual metric space, so to be clear: The sequence ¦P

n

¦ is a sequence ¦P

0

, P

1

, P

2

. . .¦ of equivalence classes

in X

∗

that get “closer” to one another with respect to distance function ∆. Each P

i

∈ ¦P

n

¦ is an equivalence

class of Cauchy sequences of the form ¦p

i0

, p

i1

, p

i2

, . . .¦ containing elements of X that get “closer together” with

respect to distance function d.

Each P

i

∈ ¦P

n

¦ is a set of equivalent Cauchy sequences in X. From each of these, choose some sequence

¦p

in

¦ ∈ P

i

. For each i we have

(∃N

i

∈ R)

_

m, n > N

i

→ d(p

im

, p

in

) <

i

_

52

We’ll deﬁne a new sequence ¦q

n

¦ by letting q

i

= p

im

for some m > N

i

. Let Q be the equivalence class containing

¦q

n

¦. We can show that lim

n→∞

P

n

= Q.

0 ≤ lim

n→∞

∆(P

n

, Q) = lim

n→∞

_

lim

k→∞

d(p

nk

, q

k

)

_

≤ lim

n→∞

n

So, by the squeeze theorem, we have

lim

n→∞

∆(P

n

, Q) = 0 (7)

which, by the deﬁnition of equality in X

∗

, means that lim

n→∞

P

n

= Q.

Proof that Q ∈ X

∗

:

Choose any arbitrarily small values

1

,

2

,

3

. From (7) we have

∃X : n, k > X → d(p

nk

, q

n

) <

1

And ¦P

n

¦ is Cauchy, so we have

∃Y : n > Y → ∆(P

n

, P

n+m

) <

2

which, after choosing appropriately large n, gives us

∃Z : k > Z → d(p

nk

, p

(n+m)k

) <

3

Therefore, by the triangle inequality:

d(q

n

, q

n+m

) ≤ d(q

n

, p

nk

) +d(p

nk

, p

(n+m)k

) +d(p

(n+m)k

, q

n+m

)

Taking the limits of each side gives us

lim

n→∞

d(q

n

, q

n+m

) ≤

1

+

3

+

1

But these epsilon values were arbitrarily small and m was an arbitrary integer. This shows that for every there

exists some integer N such that n, n +m > N implies d(q

n

, q

n+m

) < . And this is the deﬁnition of ¦q

n

¦ being

a Cauchy sequence. Therefore Q ∈ X

∗

.

Exercise 3.24d

Let ¦p

n

¦ represent the sequence whose terms are all p, and let ¦q

n

¦ represent the sequence whose terms are all

q.

∆(P

p

, P

q

) = lim

n→∞

d(p

n

, q

n

) = lim

n→∞

d(p, q) = d(p, q)

Exercise 3.24e : proof of density

Let ¦a

n

¦ be an arbitrary Cauchy sequence in X and let A ∈ X

∗

be the equivalence class containing ¦a

n

¦. If

the sequence ¦a

n

¦ converges to some element x ∈ X, then A = P

x

= ϕ(x) and therefore A ∈ ϕ(X) (see exercise

3.24 for the deﬁnition of P

x

).

If the sequence ¦a

n

¦ does not converge to some element x ∈ X, then choose an arbitrarily small . The

sequence ¦a

n

¦ is Cauchy, so we’re guaranteed the existence of K such that

j, k > K → d(a

j

, a

k

) <

From this, we can consider the sequence whose terms are all a

k

. This sequence is a member of the equivalence

class P

a

k

= ϕ(a

k

) and

∆(A, P

a

k

) = lim

n→∞

d(a

n

, a

k

) <

This shows that we can ﬁnd some element of ϕ(X) to be arbitrarily close to A, which means that A is a limit

point of ϕ(X).

We have shown that an arbitrary Cauchy sequence A ∈ X

∗

is either an element of ϕ(X) or a limit point of

ϕ(X). By deﬁnition, this means that ϕ(X) is dense in X

∗

.

53

Exercise 3.24e : if X is complete

If X is complete then every arbitrary Cauchy sequence A ∈ X

∗

converges to some point a ∈ X, so that

A = P

a

= ϕ(a) ∈ ϕ(X). This shows that X

∗

⊆ ϕ(X). And for every ϕ(b) ∈ ϕ(X) there is some Cauchy

sequence in X

∗

whose every element is b, so that ϕ(b) = P

b

∈ X

∗

. This shows that ϕ(X) ⊆ X

∗

.

This shows that X

∗

⊆ ϕ(X) ⊆ X

∗

, or that ϕ(X) = X

∗

.

Exercise 3.25

The completion of the set of rational numbers is a set that’s isomorphic to R.

Exercise 4.1

Consider the function

f(x) =

_

1, x = 0

0, x ,= 0

This function satisﬁes the condition that

lim

h→∞

[f(x +h) −f(x −h)] = 0

for all x, but the function is not continuous at x = 0: We can choose < 1, and every neighborhood N

δ

(0) will

contain a point p for which

d(f(p), f(0)) = 1

Therefore we can’t pick δ such that d(p, 0) < δ → d(f(p), f(0)) < 1, which means that f is not continuous by

deﬁnition 4.5.

Exercise 4.2

Let X be a metric space, let E be an arbitrary subset of X, and let E represent the closure of E. We want to

prove that f(E) ⊆ f(E). To do this, assume y ∈ f(E). This means that y = f(e) for some e ∈ (E ∪ E

).

case 1: e ∈ E

If e ∈ E, then y ∈ f(E) and therefore y ∈ f(E).

case 2: e ∈ E

If e ∈ E

**, then every neighborhood of e contains inﬁnitely many points of E. Choose an arbitrarily small
**

neighborhood N

**(y). We’re told that f is continuous so, by deﬁnition 4.1, we’re guaranteed the existence of δ
**

such that f(x) ∈ N

(y) whenever x ∈ N

δ

(e). But there are inﬁnitely many elements of E in the neighborhood

N

δ

(e), so there are inﬁnitely many elements of f(E) in N

**(y). This means that y is a limit point of f(E).
**

We’ve shown that every arbitrary element y ∈ f(E) is either a member of f(E) or a limit point of f(E),

which means that y ∈ (f(E) ∪ f(E)

**) = f(E). This proves that f(E) ⊆ f(E).
**

A function f for which f(E) is a proper subset f(E)

Let X be the metric space consisting of the interval (0, 1) with the standard distance metric. Let Y be the

metric space R

1

. Deﬁne the function f : X → Y as f(x) = x. The interval (0, 1) is closed in X but open in Y ,

so we have

f(X) = f(X) = (0, 1) ,= (0, 1)

Exercise 4.3

If we consider the image of Z(f) under f, we have f(Z(f)) = ¦0¦. This range is a ﬁnite set, and is therefore a

closed set. By the corollary of theorem 4.8, we know that f

−1

(¦0¦) = Z(f) must also be a closed set.

54

Exercise 4.4: f(E) is dense in f(X)

To show that f(E) is dense in f(X) we must show that every element of f(X) is either an element of f(E) or

a limit point of f(E).

Assume y ∈ f(X). Then p = f

−1

(y) ∈ X. We’re told that E is dense in X, so either p ∈ E or p ∈ E

.

case 1: p ∈ E

If p ∈ E, then y = f(p) ∈ f(E).

case 2: f

−1

(y) ∈ E

**If p is a limit point of E, then there is a sequence ¦e
**

n

¦ of elements of E such that e

n

,= p and lim

n→∞

e

n

= p.

We’re told that f is continuous, so by theorem 4.2 we know that lim

n→∞

f(e

n

) = f(p) = y. Using deﬁnition

4.2 again, we know that there is a sequence ¦f(e

n

)¦ of elements of f(E) From theorem 4.2, this tells us that

lim

x→p

f(x) = f(p) = y. Therefore y is a limit point of f(E).

We’ve shown that every element y ∈ f(X) is either an element of f(E) or a limit point of f(E). By deﬁnition,

this means that f(X) is dense in f(E).

Exercise 4.4b

Choose an arbitrary p ∈ X. We’re told E is dense in X, so p is either an element of E or a limit point of E.

Case 1: p ∈ E

If p ∈ E, then we’re told that f(p) = g(p).

Case 2: p ∈ E

**If p is a limit point of E, then there is a sequence ¦e
**

n

¦ of elements of E such that e

n

,= p and lim

n→∞

e

n

=

p. We’re told that f and g are continuous, so by theorem 4.2 we know that lim

n→∞

f(e

n

) = f(p) and

lim

n→∞

g(e

n

) = g(p). But each e

n

is an element of E, so we have f(e

n

) = g(e

n

) for all n. This tells us

that

g(p) = lim

n→∞

g(e

n

) = lim

n→∞

f(e

n

) = f(p)

We see that f(p) = g(p) in either case. This proves that f(p) = g(p) for all p ∈ X.

Exercise 4.5

The set E

C

can be formed from an at most countable number of disjoint open intervals

We’re told that E is closed, so E

C

is open. Exercise 2.29 tells us that E

C

contains an at-most countable number

of disjoint segments. Each of these segments must be open (if any of them contained a non-interior point, it

would be a non-interior point of E

C

, but open sets have no non-interior points). Let ¦(a

n

, b

n

)¦ be the at-most

countable collection of disjoint open segments.

constructing the function

We must separately consider the cases for x ∈ E and x ,∈ E; and if x ,∈ E we must consider the possibility that

E contains an interval of the form (−∞, b) or (a, ∞). We’ll deﬁne the function to be

g(x) =

_

¸

¸

_

¸

¸

_

f(x), x ∈ E

f(b

i

), x ∈ (a

i

, b

i

) ∧ a

i

= −∞

f(a

i

), x ∈ (a

i

, b

i

) ∧ b

i

= ∞

f(a

i

) +

f(bi)−f(ai)

bi−ai

(x −a

i

), x ∈ (a

i

, b

i

) ∧ −∞ < a

i

< b

i

< ∞

(8)

This function is the one mentioned in the hint: the graph of g is a straight line on each closed interval [b

i

, a

i+1

] ∈

E

C

. This function can easily (albeit tediously) be shown to be continuous on R

1

.

55

failure if “closed” is omitted

Consider the following function deﬁned on the open set (−∞, 0) ∩ (0, ∞):

f(x) =

_

1, x > 0

−1 x < 0

This function will be discontinuous at x = 0 no matter how we deﬁne the function at this point.

Extending this to vector-valued functions

Let E be a closed subset of R and let f : E →R

k

be a vector-valued function deﬁned by

f(x) = (f

1

(x), f

2

(x), . . . , f

k

(x))

For each f

i

we can deﬁne a function g

i

as in equation (8) to create a new vector-valued function g : R → R

deﬁned by

g(x) = (g

1

(x), g

2

(x), . . . , g

k

(x))

We’ve shown that each of these g

i

functions are continuous, therefore by theorem 4.10 the function g(x) is

continuous.

Exercise 4.6

Let G represent the graph of f. Let g : E → G be deﬁned as as g(x) = (x, f(x)). We can make G into a metric

space by deﬁning a distance function: the set G is a subset of R R, so the natural choice is to use the metric

d(g(x), g(y)) = d((x, f(x)), (y, f(y)) =

_

(x −y)

2

+ (f(x) −f(y))

2

. This choice allows us to treat G as a subset

of R

2

.

If f is continuous

Both x → x and x → f(x) are continuous mappings, so g(x, f(x)) is a continuous mapping by theorem 4.10.

The domain of g is a compact metric space so by theorem 4.14 (or theorem 4.15) G is compact.

If the graph is compact

Deﬁne g as above. The inverse of g is g

−1

(x, f(x)) = x. It’s clear that g

−1

is a one-to-one and onto function

from G to E. Therefore by theorem 4.17 the inverse of g

−1

– that is, g itself – is continuous. We can then appeal

to theorem 4.10 once again to conclude that f is continuous.

Exercise 4.7

f is bounded

If x = 0, then f(x, y) = 0 for any value of y. If x > 0:

(x −y

2

)

2

≥ 0 squares are positive

→ x

2

−2xy

2

+y

4

≥ 0 expanding the squared term

→ x

2

−xy

2

+y

4

≥ 0 LHS remains positive after adding the nonnegative term (xy

2

)

→ x

2

+y

4

≥ xy

2

add xy

2

to both sides

→ 1 ≥

xy

2

x

2

+y

4

divide both sides by positive term x

2

+y

4

If x < 0:

(x +y

2

)

2

≥ 0 squares are positive

→ x

2

+ 2xy

2

+y

4

≥ 0 expanding the squared term

→ x

2

−xy

2

+y

4

≥ 0 LHS remains positive after adding the nonnegative term (−3xy

2

)

→ x

2

+y

4

≥ xy

2

add xy

2

to both sides

→ 1 ≥

xy

2

x

2

+y

4

divide both sides by positive term x

2

+y

4

56

g is unbounded near (0,0)

If we let x = n

α

and let y = n

β

, we have

g(n

α

, n

β

) =

n

α+2β

n

2α

+n

6β

We can divide the numerator and denominator by n

α+2β

to get

g(n

α

, n

β

) =

1

n

α−2β

+n

4β−α

If we let α = −3, β = −1 this becomes

g

_

1

n

3

,

1

n

_

=

1

n

−1

+n

−1

=

n

2

Taking limits, we have

g(0, 0) = lim

n→∞

g

_

1

n

3

,

1

n

_

= lim

n→∞

n

2

The rightmost limit is +∞.

f is not continuous at (0,0)

Choose 0 < δ < 1/2. For f to be continuous we must be able to choose some > 0 such that

d ((0, 0), (x, y)) < → d (f(0, 0), f(x, y)) < δ

But there can be no such epsilon. To see this, choose an arbitrary > 0, choose x > 0 such that x

2

+ x <

2

,

and then choose y =

√

x. This gives us

d ((0, 0), (x, y)) =

_

x

2

+y

2

=

_

x

2

+x <

but

d (f(0, 0), f(x, y)) =

xy

2

x

2

+y

4

=

x

2

2x

2

=

1

2

The restriction of f to a straight line is continuous

Any straight line that doesn’t pass through (0, 0) doesn’t encounter any of the irregularities that occur at the

origin; it’s trivial but tedious to show that the restriction of f to such a line is continuous. So we need only

consider lines that pass through the origin: that is, lines of the form y = cx or x = 0 for some constant c.

For the line y = cx: let > 0 be given and let δ = /c

2

. Choose x such that 0 < x < δ. Then:

d(f(0, 0), f(x, y)) =

¸

¸

¸

¸

xy

2

x

2

+y

4

¸

¸

¸

¸

=

¸

¸

¸

¸

c

2

x

3

x

2

+c

4

x

4

¸

¸

¸

¸

=

¸

¸

¸

¸

c

2

x

1 +c

4

x

2

¸

¸

¸

¸

<

¸

¸

¸

¸

c

2

x

1

¸

¸

¸

¸

<

¸

¸

c

2

δ

¸

¸

=

And was arbitrary, so this proves that f is continuous on this line.

For the line x = 0: let > 0 be given choose any y ,= 0. Regardless of our choice of δ, we have

d(f(0, 0), f(x, y)) =

¸

¸

¸

¸

xy

2

x

2

+y

4

¸

¸

¸

¸

=

¸

¸

¸

¸

0

y

4

¸

¸

¸

¸

= 0 <

And was arbitrary, so this proves that f is continuous on this line.

The restriction of g to a straight line is continuous

The proof for the continuity of the restriction of g is almost identical to that of the continuity of the restriction

of f.

57

Exercise 4.8

To show that f is bounded, we need to show that there exists some M such that [f(p)[ < M for every p ∈ R

1

.

Let p be an arbitrary element of R

1

. We’re told that E is a bounded subset of R

1

, so we know that E has a

lower bound α. Choose any > 0. From the deﬁnition of uniform continuity there exists some δ > 0 such that,

for all p, q ∈ R

1

,

d(p, q) < δ → d(f(p), f(q)) <

Let n be the smallest integer such that nδ > p − α (which exists from the Archimedean property of the reals)

and deﬁne γ = (p −α)/n. This allows us to divide the interval (α, p) into n intervals of length γ, each of which

is smaller than δ. We can then apply the triangle inequality multiple times:

d(f(α), f(p)) ≤ d(f(α), f(α +γ)) +d(f(α +γ), f(α + 2γ)) +. . . +d(f(α + (n −1)γ), f(α +nγ))

< + +. . . + (n terms)

= n

This shows us that [f(p)[ is bounded by [f(α) ± n[ for all p ∈ R

1

, so by deﬁnition 4.13 we know that f is

bounded.

if E is not bounded

If E is not bounded below, then we can revise the previous proof using the upper bound β in place of the lower

bound α. If E is not bounded above or below, then we will not be able to use the Archimedean property to ﬁnd

n such that nδ > p −(−∞) and the proof fails. For an example of a real uniformly continuous function that is

not bounded, we can simply look to f(x) = x.

Exercise 4.9

Deﬁnition 4.18 says that f is uniformly continuous on X if for every > 0 there exists δ > 0 such that

d

X

(p, q) < δ → d

Y

(f(p), f(q)) < (9)

We want to show that this conditional statement is true iﬀ for every > 0 there exists δ > 0 such that

diam E < δ → diam f(E) < (10)

(10) implies (9)

Let equation (10) hold and suppose d(p, q) < δ.

d(p, q) < δ assumed

→ diam ¦p, q¦ < δ we’re letting E be the set containing only p and q.

→ diam ¦f(p), f(q)¦ < from (10)

→ d(f(p), f(q)) < deﬁnition of diameter

Therefore (9) holds.

(9) implies (10)

Let equation (9) hold, let E be a nonempty subset of X, and suppose diam E < δ.

diam E < δ assumed

→ (∀p, q ∈ E)(d(p, q) < δ) deﬁnition of diameter

→ (∀p, q ∈ E)(d(f(p), f(q)) < ) deﬁnition of diameter

→ (∀f(p), f(q) ∈ f(E))(d(f(p), f(q)) < ) p ∈ E → f(p) ∈ f(E)

→ diam f(E) < deﬁnition of diameter

Therefore (10) holds.

58

Exercise 4.10

We want to prove the converse of theorem 4.19: that is, we want to prove that if f is not uniformly continuous

on X then either X is not a compact metric space or f is not a continuous function.

If f is not uniformly continuous, then the converse of deﬁnition 4.18 tells us that there exists some such

that, for every δ > 0, we can ﬁnd p, q ∈ X such that d(p, q) < δ but d(f(p), f(q)) ≥ . We’ll be contradicting

this claim, so we’ll restate it formally in a numbered equation:

(∃ > 0)(∀δ > 0) d(p, q) < δ ∧ d(f(p), f(q)) ≥ (11)

From the fact that we can ﬁnd such p, q for every value of δ, we can set δ

n

=

1

n

and construct two sequences

¦p

n

¦ and ¦q

n

¦ where d(p

n

, q

n

) <

1

n

but d(f(p

n

), f(q

n

)) ≥ .

If the set X is compact (supposition 1) the sequence ¦p

n

¦ must have at least one subsequential limit p

(theorem 3.6a or 2.37). And, since d(p

n

, q

n

) <

1

n

, the sequence ¦q

n

¦ must also have p as a subsequential limit

since

d(q

n

, p) ≤ d(q

n

, p

n

) +d(p

n

, p) <

1

n

+

But then, from theorem 4.2, it must be the case that both ¦f(p

n

)¦ and ¦f(q

n

)¦ have subsequential limits of

f(p). If the function f is continuous without being uniformly continuous (supposition 2) this would imply

that, for some suﬃciently large n,

d(f(p

n

), f(q

n

)) ≤ d(f(p

n

), f(p)) +d(f(p), f(q

n

)) <

2

+

2

=

which contradicts (11).

We’ve established a contradiction from our initial assumption that f is uniformly continuous, so one of our

suppositions must be incorrect: either X is not compact or f is not continuous. By converse, if X is compact

and f is continuous then f is uniformly continuous. And this is what we were asked to prove.

Exercise 4.11

Choose an arbitrary > 0. We’re told that f is uniformly continuous, so

(∃δ > 0) : d(x

m

, x

n

) < δ → d(f(x

m

), f(x

n

)) <

And we’re told that ¦x

n

¦ is Cauchy, so

(∃N) : n, m > N → d(x

m

, x

n

) < δ

Together, these two conditional statements tell us that, for any ,

(∃N) : n, m > N → d(f(x

m

), f(x

n

)) <

which, by deﬁnition, shows that ¦f(x

n

)¦ is Cauchy. See exercise 4.13 for the proof that uses this result.

Exercise 4.12

We’re asked to prove that the composition of uniformly continuous functions is uniformly continuous. Let

f : X → Y and g : Y → Z be uniformly continuous functions. Choose an arbitrary > 0. Because g is

uniformly continuous, we have

(∃α > 0) : d(y

1

, y

2

) < α → d(g(y

1

), g(y

2

)) <

Because f is uniformly continuous, we have

(∃δ > 0) : d(x

1

, x

2

) < δ → d(f(x

1

), f(x

2

)) < α

Since the elements f(x

1

), f(x

2

) are elements of Y these two conditional statements together tell us that, for

arbitrary > 0,

(∃δ > 0) : d(x

1

, x

2

) < δ → d(f(x

1

), f(x

2

)) < α → d(g(f(x

1

)), g(f(x

2

))) <

which, by deﬁnition means that the composite function (g ◦ f) : X → Z is uniformly continuous.

59

Exercise 4.13 (Proof using exercise 4.11)

For every p ∈ X it is either the case that p ∈ E or p ,∈ E. If p ,∈ E, then p is a limit point of E (because

of density) and therefore we can construct a sequence ¦e

n

¦ that converges to p. We now deﬁne the function

g : X → Y as

g(p) =

_

lim

n→∞

f(p), if p ∈ E

lim

n→∞

f(e

n

), if p ,∈ E

We need to prove that this is actually a well-deﬁned function and that it’s continuous.

The function g is well-deﬁned

It’s clear that every element of X is mapped to at least one element of Y , but it’s not immediately clear that

each element of X is mapped to only one element of Y . We need to show that if ¦p

n

¦ and ¦q

n

¦ are sequences

in E that converge to p, then ¦f(p

n

)¦ and ¦f(q

n

)¦ both converge to the same element.

Let ¦p

n

¦ and ¦q

n

¦ be arbitrary sequences in E that converge to p. From the deﬁnition of convergence (or

from exercise 4.11), we know that

(∀δ > 0)(∃N) : n > N → d(p

n

, p) <

δ

2

(∀δ > 0)(∃M) : m > M → d(q

m

, p) <

δ

2

We can then use the triangle inequality to show that

(∀δ > 0)(∃M, N) : n > max¦M, N¦ → d(p

n

, q

n

) ≤ d(p

n

, p) +d(p, q

n

) < δ (12)

We know that the function f is continuous on E, so for any > 0 there exists some δ > 0 such that

d(p

n

, q

n

) < δ → d(f(p

n

), f(q

n

)) < (13)

Combining equations (12) and (13) we see that for any > 0 we can ﬁnd some integer N such that

n > N → (d(p

n

, q

n

) < δ) → (d(f(p

n

), f(q

n

)) < )

This tells us that ¦f(p

n

)¦ and ¦f(q

n

)¦ don’t converge to diﬀerent points, but without being able to conclude

that Y is compact it’s entirely possible that ¦f(p

n

)¦ and ¦f(q

n

)¦ are merely Cauchy without actually converging

to any point in Y at all. The exercise asks us to assume that Y is R

1

, but we’ll try for a more general solution

and assume only that Y is a complete metric space (see deﬁnition 3.12). This allows us to conclude that ¦f(p

n

)¦

and ¦f(q

n

)¦ both converge to the same point of Y , and we call this point g(p). Therefore the function g(p) has

a unique value for every p ∈ X.

60

The function g is continuous

d(p, q) < δ/2 assumed

Let ¦p

n

¦ be a sequence that converges to p and let ¦q

n

¦ be a sequence that converges to q.

→ d(p

n

, q

n

) ≤ d(p

n

, p) +d(p, q) +d(q, q

n

)

→ d(p

n

, q

n

) < d(p

n

, p) +δ/2 +d(q, q

n

) from our initial assumption

This holds true for any n, even when n is suﬃciently large to make d(q, q

n

) < δ/4, d(p, p

n

) < δ/4 in

which case we have

→ d(p

n

, q

n

) < δ

We’re told that f is continuous on E, so we can make d(f(p

n

), f(q

n

)) arbitrarily small if we can

make d(p

n

, q

n

) arbitrarily small. And we can make δ arbitrarily small, so for any we can choose

p

n

, q

n

such that

→ d(f(p

n

), f(q

n

)) < /2

By deﬁnition, we have g(p) = f(p) for p ∈ E, so

→ d(g(p

n

), g(q

n

)) < /2

→ d(g(p), g(q)) ≤ d(g(p), g(p

n

)) +d(g(p

n

), g(q

n

)) +d(g(q

n

), g(q)) triangle inequality

→ d(g(p), g(q)) < d(g(p), g(p

n

)) +/2 +d(g(q

n

), g(q)) established bound for d(g(p

n

), g(q

n

))

This holds true for any n, even when n is suﬃciently large to make d(q, q

n

) < /4, d(p, p

n

) < /4

(we know that ¦g(p

n

)¦ converges to g(p) and that ¦g(q

n

)¦ converges to g(q) because we deﬁned

g(p), g(q) speciﬁcally so that this would be true). For such values of n we have

d(g(p), g(q)) <

This shows that we can constrain the range of g by constraining its domain, and therefore g is continuous.

Could we replace the range with R

k

or any compact metric space? By any complete metric space?

Yes. Deﬁnition 3.12 tells us that all compact metric spaces and all Euclidean metric spaces are also complete

metric spaces. Our proof didn’t depend on the range of g being R

1

: we assumed only that the range was a

complete metric space.

Could we replace the range with any metric space?

No. Consider the function f(x) =

1

x

with a domain of E =

_

1

n

: n ∈ N

_

. The set E is dense in X = E ∪ ¦0¦.

This function is continuous at every x ∈ E (we can ﬁnd a neighborhood N

δ

(x) that contains only x, which

guarantees that d(f(y), f(x)) = 0 for any y in N

δ

(x)). However, there is no way to deﬁne f(0) to make this a

continuous function. This can be seen intuitively by recognizing that the range of f(x) = 1/x is unbounded in

every neighborhood of x = 0. If you’re satisﬁed with an “intuitive proof”, we’re done. Otherwise, buckle up

and get ready for a few more paragraphs of inequalities and Greek letters.

To prove that f is discontinous at x = 0, choose any arbitrary neighborhood around 0 with radius δ. We

can ﬁnd an integer n such that n >

1

δ

(Archimedian property of the reals) so that d(

1

n

, 0) < δ. For any positive

integer k we also have d(

1

n+k

, 0) < δ. Now, by the triangle inequality, we have

d

_

f

_

1

n

_

, f

_

1

n +k

__

≤ d

_

f

_

1

n

_

, f (0)

_

+d

_

f (0) , f

_

1

n +k

__

We haven’t speciﬁcied a value for f(0) yet, but we do know that [f(1/n) −f(1/n+k)[ = [n−(n+k)[ = k. This

last inequality is therefore equivalent to

k ≤ d

_

f

_

1

n

_

, f (0)

_

+d

_

f (0) , f

_

1

n +k

__

61

Every term of this inequality is nonnegative, so one of the two terms on the right-hand side of this inequality

must be ≥ k/2. But this is true for arbitrarily large k and arbitrarily small δ.

We can now show that f is not continuous at x = 0. Choose any > 0. For all possible choice of δ > 0 we can

ﬁnd integers n >

1

δ

and k > 2, so that by the previous method we can be sure that either d(f(

1

n

), f(0)) ≥ k/2 =

or d(f(

1

n+k

), f(0)) ≥ k/2 > . By deﬁnition 4.5 of “continuous”, this proves that f is not continous at 0. And

we never speciﬁed the value of f(0), so we’ve proven that f will be discontinuous however we deﬁne f(0). This

means that there is no continuous extension from E to X.

Exercise 4.14 proof 1

If f(0) = 0 or f(1) = 1, we’re done. Otherwise, we know that f(0) > 0 and f(1) < 1.

We’ll now inductively deﬁne a sequence of intervals ¦[a

i

, b

i

]¦ as follows

3

: Let [a

0

, b

0

] = [0, 1]. Now assume

that we have deﬁned ¦[a

i

, b

i

]¦ for i = 0, 1, . . . , n. Let m = (a

n

+b

n

)/2. We deﬁne [a

n+1

, b

n+1

] as

[a

n

, m], if f(m) ≤ m

[m, b

n

], if f(m) ≥ m

(This procedure is underdeﬁned for the case of f(m) = m, but when f(m) = m we’ve found the point we’re

looking for anyway). The interval [a

n+1

, b

n+1

] has diameter 2

−(n+1)

with f(a

n+1

) ≥ a

n+1

, f(b

n+1

) ≤ b

n+1

. We

therefore have a nested sequence of compact sets ¦[a

n

, b

n

]¦ whose diameter converges to zero, so by exercise 3.21

we know that

¦[a

n

, b

n

]¦ consists of just one point i ∈ I. Our method of constructing ¦[a

n

, b

n

]¦ guarantees that

f(i) ≤ i and f(i) ≥ i, therefore f(i) = i and this is the point we were asked to ﬁnd.

Exercise 4.14 proof 2

This proof, which is much clearer and more concise than mine, was taken from the homework of Men-Gen Tsai

(b89902089@ntu.edu.tw).

Deﬁne a new function g(x) = f(x) − x. Both f(x) and x are continuous, so g(x) is continuous by theorem

4.9. If g(0) = 0 or g(1) = 0, then we’re done. Otherwise, we have g(0) > 0 and g(1) < 0. Therefore by theorem

4.23 there must be some intermediate point in the interval [0, 1] for which g(x) = 0: at this point, f(x) = x.

Exercise 4.15

Although the details of this proof might be ugly, the general idea is simple. If the function f is continuous

but not monotonic then we can ﬁnd some open interval (x

1

, x

2

) on which f(x) obtains a local maximum or

minimum. This point f(x) is not an interior point of f( (x

1

, x

2

) ), so f( (x

1

, x

2

) ) is not an open set.

Let f : R

1

→R

1

be a continuous mapping and assume that f is not monotonically increasing or decreasing.

Then we can ﬁnd x

1

< x

2

< x

3

such that f(x

2

) > f(x

1

) and f(x

2

) > f(x

3

), or such that f(x

2

) < f(x

1

) and

f(x

2

) < f(x

3

). The proofs for either case are analogous, so we’ll focus only on the case that f(x

2

) > f(x

1

) and

f(x

2

) > f(x

3

).

We can’t assume that the supremum of f( (x

1

, x

3

) ) occurs at f(x

2

): there could be some x

2.5

for which

f(x

2.5

) > f(x

2

). We want to construct a closed subinterval [x

a

, x

b

] ⊂ (x

1

, x

3

) containing the x for which f(x)

obtains its local maximum. To do this, we deﬁne to be

= min¦d( f(x

2

), f(x

1

) ), d( f(x

2

), f(x

3

) )¦

Because f is continuous, we know that there exists some δ

a

and δ

b

such that

d(x

1

, x) < δ

a

→ d(f(x

1

), f(x)) <

d(x

3

, x) < δ

b

→ d(f(x

3

), f(x)) <

3

Our method of deﬁning this sequence is similar to Newton’s method or the Bisection method of root ﬁnding.

62

We can now let x

a

= x

1

+

1

2

δ

a

and x

b

= x

3

−

1

2

δ

b

. We know that f doesn’t obtain its local maximum for any x

in the intervals (x

1

, x

a

) or (x

b

, x

3

) because all of these xs are suﬃciently close (within δ) to x

1

or x

3

so that f(x)

is less (by at least

1

2

δ) than f(x

2

). And [x

a

, x

b

] is a closed subset of R

1

and is therefore a compact set, so we

know that f([x

a

, x

b

]) is a closed and bounded subset of R

1

(theorem 4.15), and therefore f([x

a

, x

b

]) contains its

supremum (which may or may not occur at x

2

, but we only care that the supremum exists for some x ∈ [x

a

, x

b

]).

This is all getting a bit muddled, so to recap our results so far:

1) [x

a

, x

b

] ⊂ (x

1

, x

3

)

2) f([x

a

, x

b

]) ⊂ f((x

1

, x

3

))

3) (x ∈ (x

1

, x

a

) ∪ (x

b

, x

3

)) → f(x) < f(x

2

)

4) f(x) has a local maximum for some x ∈ [x

a

, x

b

]

Together these four facts tell us that f( (x

1

, x

3

) ) contains its own supremum for some x ∈ [x

a

, x

b

]. But the

supremum of a set can’t be an interior point of the set, so f( (x

1

, x

3

) ) is not open even though (x

1

, x

3

) is

open, so f is not an open mapping (theorem 4.8) We’ve shown that “f is not monotonic” implies “f is not a

continuous open mapping from R

1

to R

1

”. By converse, then, we have shown that if f is a continuous open

mapping from R

1

to R

1

then f is monotonic. And this is what we were asked to prove.

Exercise 4.16

The functions [x] and (x) are both discontinuous for each x ∈ N. If x ∈ N then, for every 0 < δ <

1

2

, we have

d([x −δ], [x]) = [(x −1) −x[ = 1

d((x −δ), (x)) = [(1 −δ) −0[ = 1 −δ >

1

2

This shows that we can’t constrain the range of these functions by restraining the domain around x ∈ N, which

by deﬁnition means that the functions are discontinuous for integer values of x.

Note that the the sum of these discontinuous functions, [x] + (x), is the continuous function f(x) = x.

Exercise 4.17

Deﬁne E as in the hint. To the three criteria listed in the exercise, we add a fourth:

(d) a < q < x < r < b

Lemma 1: For each x ∈ E we can ﬁnd p, q, r such that the four criteria are met

We’re assuming that f(x−) < f(x+), so by the density of Q in R (theorem 1.20b) we can ﬁnd some rational p

such that f(x−) < p < f(x+). From our deﬁnition of the set E we know that f(x−) exists for all x ∈ E, so we

can ﬁnd some neighborhood N

δ

(x) containing a rational q that fulﬁlls criteria (b). More formally,

(∃q ∈ N

δ

(x) ∩ (a, x))(a < q < x ∧ [q < t < x → f(t) < p]

Similarly, the existence of f(x+) tells us that there is some neighborhood N

δ2

(x) containing a rational r such

that

(∃r ∈ N

δ2

(x) ∩ (x, b))(x < r < b ∧ [x < t < r → f(t) > p]

Therefore each x ∈ E can be associated with at least one (p, q, r) rational triple with q < x < r.

Lemma 2: If a given (p, q, r) triple fulﬁlls the criteria for x and y, then x = y

Suppose x, y are two elements of E that both meet the four criteria. Assume, without loss of generality, that

x < y. The density of the reals guarantees that there exists some rational w such that x < w < y. But this

means that f(w) < p (by criteria (b) since q < w < y) and that f(w) > p (by criteria (c) since x < w < r).

Reversing the roles of x and y establishes the same results for the case of x > y. This is clearly a contradiction,

so one of our assumptions must be wrong: for any given triple (p, q, r) it’s not possible to ﬁnd two distinct

elements of E that fulﬁll all four criteria. Note that we haven’t proven that each triple can be associated with

a unique x ∈ E: we’ve only shown that each triple can be associated with at most one x ∈ E.

63

Lemma 3: E is at most countable

The set of rational triples is countable (theorem 2.13). Lemma 2 tells us that we can create a function from

the set of triples onto E, so the cardinality of E is not more than the cardinality of the set of rational triples.

Therefore the cardinality of E is at most countable.

Lemma 4: F is at most countable

If we deﬁne F to be the set of x for which f(x+) < f(x−) we can prove that F is at most countable with only

trivial modiﬁcations to the previous lemmas.

Lemma 5: G is at most countable

Deﬁne G to be the set of x for which f(x+) = f(x−) = α but f(x) ,= α. Let x be an arbitrary element of

G. If every neighborhood of x contained another element of G, we could construct a sequence ¦t

n

¦ for which

¦t

n

¦ → x but ¦f(t

n

)¦ ,→ α. But this would mean that either f(x−) or f(x+) wouldn’t exist, which contradicts

our deﬁnition of G. Therefore each x is an isolated point of G: we can ﬁnd some radius δ such that N

δ

(x)∩G = x

and N

δ

(x) contains rational numbers q, r such that q < x < r.

Now consider all of the (q, r) pairs of rational numbers. Each of these pairs will either have one, more than

one, or zero elements of G in the associated open interval (q, r). Let ( be the set of rational pairs (q, r) for

which there is exactly one element of g in the open interval (q, r). The set ( is a subset of QQ, so it’s at most

countable (theorem 2.13). And each x ∈ G can be associated with at least one (p, q) ∈ (, so we can create a

function from ( onto G. This proves that G is at most countable.

The sets E,F,and G exhaust all the diﬀerent types of simple discontinuities and all of these sets are countable.

Therefore their union is countable (theorem 2.12). And this is what we were asked to prove.

Exercise 4.18

f is continuous at every irrational point

Let x be irrational. Choose any > 0 and let n be the smallest integer such that n >

1

. If we want to constrain

the range of f to f(x) ±, we will need to ﬁnd a neighborhood of x that contains no rational numbers in the set

_

p

q

: p, q ∈ Q, q ≤ n

_

For each i ≤ n the product xi is irrational, so there exists some integer m

i

such that

m < xi < m+ 1

which means that

m

i

< x <

m+ 1

i

So we can deﬁne a neighborhood around x with radius δ

i

where

δ

i

= min

_

d

_

x,

m

i

_

, d

_

x,

m+ 1

i

__

The neighborhood (x −δ

i

, x +δ

i

) contains no rationals of the form

k

i

. So if we have constructed neighborhoods

δ

1

, δ

2

, . . . , δ

n

we can let δ = min¦δ

i

¦ (this minimum exists because n is ﬁnite), and we will have constructed a

neighborhood (x−δ, x+δ) that contains no rational numbers with denominators ≤ n. Therefore, for all rational

numbers y, we have

d(x, y) < δ → d(f(x), f(y)) <

1

n

<

and for irrational numbers y we have

d(x, y) < δ → d(f(x), f(y)) = 0 <

Which, by deﬁnition, means that f is continuous at x.

64

f has a simple discontinuity at every rational point

Let x be a rational point. If we follow the previous method of constructing δ we immediately see that f(x+) =

f(x−) = 0. When x is rational, though, we no longer have f(x) = 0 and therefore f has a simple discontinuity

at x.

Exercise 4.19

Suppose that f is discontinous at some point p. By deﬁnition 4.5 of “continuity”, we know that there exists

some > 0 such that, for every δ > 0, we can ﬁnd some x such that d(x, p) < δ but d(f(x), f(p)) ≥ . Therefore

we are able construct a sequence ¦x

n

¦ where each term of ¦x

n

¦ gets closer to p even though every term of

¦f(x

n

)¦ diﬀers from f(p) by at least . More formally, we choose each element of ¦x

n

¦ so that d(p, x

n

) <

1

n

but

d(f(p), f(x

n

)) ≥ . By constructing the sequence in this way we see that limx

n

→ p but d(f(x

n

), f(p)) ≥ for

all n.

Assume, without loss of generality, that f(x

n

) < f(p) for every term of ¦x

n

¦ (see note 1 for justiﬁcation of

WLOG). From the density property of the reals we can ﬁnd a rational number r such that f(x

n

) < r < f(p).

We’re told that f has the intermediate value property, so for each x

n

there is some y

n

between x

n

and p such

that f(y

n

) = r. This sequence ¦y

n

¦ converges to p (since the terms of ¦y

n

¦ are squeezed between the terms of

¦x

n

¦ and p) and ¦f(y

n

)¦ clearly converges to r (since f(y

n

) = r for all n). But the fact that ¦y

n

¦ converges to

p means that p is a limit point of ¦y

n

¦, which means that p is a limit point of the closed set described in the

exercise (the set of all x with f(x) = r). Therefore p is an element of this set, which means that f(p) = r. And

this is a contradiction since we speciﬁcally chose r so that f(p) ,= r.

We’ve established a contradiction, so our initial supposition must be false: f is not discontinous at any point

p. Therefore f is continuous.

Note 1: justiﬁcation for WLOG claim

For each term of the sequence ¦f(x

n

)¦ either f(p) < f(x

n

) or f(x

n

) < f(p). Therefore ¦f(x

n

)¦ either contains

an inﬁnite subsequence where f(x

n

) < f(p) for all n or f(x

n

) > f(p) for all n (or both). If it contains an inﬁnite

subsequence where f(x

n

) < f(p) for all n then we use this subsequence in place of ¦x

n

¦ in the proof. If it does

not contain such a subsequence, then it contains an inﬁnite subsequence where f(x

n

) > f(p) for all n: the proof

for this case requires only the trivial twiddling of a few inequality signs. Therefore we can assume, without loss

of generality, that f(x

n

) < f(p) for each term of ¦x

n

¦.

Exercise 4.20a

If x ,∈ E then x is an element of the open set E

C

, therefore there is some radius such that d(x, y) < → y ,∈ E,

and therefore inf

z∈E

d(x, z) ≥ > 0.

If x ∈ E then for every > 0 we can ﬁnd some z ∈ E such that d(x, z) < and therefore inf

z∈E

d(x, z) = 0.

Exercise 4.20b

Case 1: x, y ∈ E

If x and y are both elements of E then from part (a) we have ρ

E

(x) = ρ

E

(y) = 0 and therefore

[ρ

E

(x) −ρ

E

(y)[ = 0 ≤ d(x, y)

Case 2: x ∈ E, y ,∈ E

If y ∈ E but x ,∈ E then ρ

E

(y) = 0 and d(x, y) ∈ ¦d(x, z) : z ∈ E)¦. Therefore

[ρ

E

(x) −ρ

E

(y)[ = [ρ

E

(x)[ = inf

z∈E

d(x, z) ≤ d(x, y)

65

Case 3: x, y ,∈ E

If neither x nor y are elements of E then, for arbitrary z ∈ E we have

ρ

E

(x) ≤ d(x, z) ≤ d(x, y) +d(y, z)

This must hold for any choice of z. By choosing z to make d(y, z) arbitrarily close to inf

z∈E

d(x, z) we have

ρ

E

(x) ≤ d(x, y) +ρ

E

(y) +

Our choice of can be made arbitrarily small (possibly even zero, if y is a limit point of E), so this is equivalent

to

ρ

E

(x) ≤ d(x, y) +ρ

E

(y) → ρ

E

(x) −ρ

E

(y) ≤ d(x, y)

By changing the roles of x and y we can similarly show that

ρ

E

(y) −ρ

E

(x) ≤ d(x, y)

Together, these last two inequalities show us that

[ρ

E

(y) −ρ

E

(x)[ ≤ d(x, y)

These three cases exhaust the possibilities for x, y. This shows us that for every > 0 we have d(ρ

E

(x), ρ

E

(y)) <

whenever d(x, y) < . And this is just deﬁnition 4.18 for uniform continuity with δ = .

Exercise 4.21

K is compact and ρ

F

is continuous on K, so ρ

F

(K) is compact (and is therefore both closed and bounded from

theorem 2.41). The results of exercise 20 tell us that 0 is not an element of ρ

F

(K), and therefore 0 is not a limit

point of ρ

F

(K) (since ρ

F

(K) is closed). Therefore we can ﬁnd some neighborhood around 0 with radius δ such

that none of the elements in the interval (0 −δ, 0 +δ) are in ρ

F

(K).

Choose p ∈ K, q ∈ F. The distance d(p, q) is an element of ρ

F

(K) and therefore d(p, q) ,∈ (0 − δ, 0 + δ).

Therefore d(p, q) > δ.

The conclusion fails if neither set is compact

Consider the sets

4

K =

_

n +

1

n

: n ∈ N

_

F = ¦n : n ∈ N¦ −¦2¦

The set F is closed, neither K nor F is compact, and we can choose p ∈ K, q ∈ F such that d(p, q) <

1

n

for any

n.

Exercise 4.22

Exercise 20a shows us that ρ

A

(p) = 0 iﬀ p ∈ A and ρ

B

(p) = 0 iﬀ p ∈ B, so it’s clear that f(p) = 0 iﬀ p ∈ A and

f(p) = 1 iﬀ p ∈ B. This means that the ρ

A

(p) +ρ

B

(p) is never zero, so f is continuous from theorem 4.9.

The intervals [0,

1

2

) and (

1

2

, 1] are open in [0, 1]; therefore V and W are open by theorem 4.8. The sets V

and W are clearly disjoint since f(p) can have only a single value for any given p. From the deﬁnition of A and

B we have

A = f

−1

(0) ⊆ f

−1

__

0,

1

2

__

= V

B = f

−1

(1) ⊆ f

−1

__

1

2

, 1

__

= W

4

This counterexample was taken from the homework of Men-Gen Tsai, b89902089@ntu.edu.tw

66

Exercise 4.23a: proof of continuity

Choose x ∈ (a, b). Assume without loss of generality that f(x) ,= f(a) and f(x) ,= f(b) (see below for justiﬁcation

of WLOG). Choose some > 0, and choose

such that

0 <

< min¦, [f(b) −f(x)[, [f(a) −f(x)[¦

Choose δ such that

0 < δ < min

_

(b −x)

[f(b) −f(x)[

,

(x −a)

[f(a) −f(x)[

_

Deﬁne λ

a

, λ

b

as

λ

a

= 1 −

δ

x −a

λ

b

= 1 −

δ

b −x

The method we used to select

**and δ ensure that both of these λ values are in the interval (0, 1). Notice that
**

we can express δ in terms of λ

a

or λ

b

as

δ = (1 −λ

a

)(x −a) = (1 −λ

b

)(b −x)

If we restrict the domain of f to (x −δ, x +δ) we have

[f(x +δ) −f(x)[

= [f(x + (1 −λ

b

)(b −x)) −f(x)[ from our deﬁnition of δ

= [f(λ

b

x + (1 −λ

b

)b) −f(x)[ algebraic rearrangement

≤ [λ

b

f(x) + (1 −λ

b

)f(b) −f(x)[ from the convexity of f

= [(1 −λ

b

)(f(b) −f(x))[ algebraic rearrangement

= [

δ

b−x

(f(b) −f(x))[ from our deﬁnition of λ

<

(b−x)

(b−x)|f(b)−f(x)|

[(f(b) −f(x))[ from our deﬁnition of δ

=

algebra

< from our deﬁnition of

and also

[f(x −δ) −f(x)[

= [f(x + (1 −λ

a

)(x −a)) −f(x)[ from our deﬁnition of δ

= [f(λ

a

x + (1 −λ

a

)a) −f(x)[ algebraic rearrangement

≤ [λ

a

f(x) + (1 −λ

a

)f(a) −f(x)[ from the convexity of f

= [(1 −λ

a

)(f(a) −f(x))[ algebraic rearrangement

= [

δ

x−a

(f(a) −f(x))[ from our deﬁnition of λ

<

(x−a)

(x−a)|f(a)−f(x)|

[(f(a) −f(x))[ from our deﬁnition of δ

=

algebra

< from our deﬁnition of

We chose an arbitrarily small and showed how to ﬁnd a nonzero upper bound for δ such that d(x, y) < δ →

d(f(x), f(y)) < . By deﬁnition, this means that f is continuous.

Justifying “without loss of generality”: special case of f(a) = f(x) or f(b) = f(x)

Let x be an arbitrary point in the interval (a, b). Choose α ∈ (a, x) such that f(α) ,= f(x): if this is not possible,

then choose α = x. Choose β ∈ (x, b) such that f(β) ,= f(x): if this is not possible, then choose β = x.

67

If α < x < β, we can use the previous method to prove that f is continuous on an interval (α, β) containing x

such that (α, β) ⊆ (a, b). Since x was an arbitrary element of (a, b) this is suﬃcient to prove that f is continuous

on (a, b).

If α = x < β, we can use the previous method to prove that f is continuous on an interval (x, β) containing

x such that (x, β) ⊆ (a, b). We only have α = x if there was no α ∈ (a, x) such that f(α) ,= f(x): this can

only happen if f is constant (and therefore continuous) on the interval (a, x]. Therefore f is continuous on the

interval (a, β) containing our arbitrary x; this is suﬃcient to prove that f is continuous on (a, b).

If α < x = β, we can use the previous method to prove that f is continuous on an interval (α, x) containing

x such that (α, x) ⊆ (a, b). We only have β = x if there was no β ∈ (x, b) such that f(β) ,= f(x): this can

only happen if f is constant (and therefore continuous) on the interval [x, b). Therefore f is continuous on the

interval (α, b) containing our arbitrary x; this is suﬃcient to prove that f is continuous on (a, b).

If α = x = β then f is a constant function on the entire interval (a, b) and is therefore trivially continuous.

Exercise 4.23b: increasing convex functions of convex functions are convex

Let g be an increasing convex function, let f be a convex function, and deﬁne their composite to be h = g ◦ f.

f(λx + (1 −λ)y) ≤ λf(x) + (1 −λ)f(y) from convexity of f

→ g(f(λx + (1 −λ)y)) ≤ g(λf(x) + (1 −λ)f(y)) g is an increasing function

→ g(f(λx + (1 −λ)y)) ≤ g(λf(x) + (1 −λ)f(y)) ≤ λg(f(x)) + (1 −λ)g(f(y)) from convexity of g

→ g(f(λx + (1 −λ)y)) ≤ λg(f(x)) + (1 −λ)g(f(y)) transitivity

→ h(λx + (1 −λ)y) ≤ λh(x) + (1 −λ)h(y) deﬁnition of h

x and y were arbitrary, so this shows that h is a convex function.

Exercise 4.23c: the ugly-looking inequality

We’re given that s < t < u, so we know that 0 < t −s < u −s. Deﬁne λ to be

λ =

t −s

u −s

From this deﬁnition we immediately see that 0 < λ < 1. We can also see that

(1 −λ) = 1 −

t −s

u −s

=

u −t

u −s

From this, we can establish an inequality:

λf(u) + (1 −λ)f(s) ≥ f(λu + (1 −λ)s) from convexity of f

→ λf(u) + (1 −λ)f(s) ≥ f

_

t−s

u−s

u +

u−t

u−s

s

_

deﬁnition of λ, (1 −λ)

→ λf(u) + (1 −λ)f(s) ≥ f

_

ut−us+su−st

u−s

_

algebra

→ λf(u) + (1 −λ)f(s) ≥ f

_

t(u−s)

u−s

_

algebra

→ λf(u) + (1 −λ)f(s) ≥ f(t) algebra

From this last inequality, we can derive two results.

λf(u) + (1 −λ)f(s) ≥ f(t) previously derived

→ λf(u) −λf(s) ≥ f(t) −f(s) subtract f(s) from each side

→

t−s

u−s

(f(u) −λf(s)) ≥ f(t) −f(s) deﬁnition of λ

68

Dividing both sides of this last inequality by t −s gives us

f(u) −f(s)

u −s

≥

f(t) −f(s)

t −s

(14)

Similarly, we have:

λf(u) + (1 −λ)f(s) ≥ f(t) previously derived

→ −λf(u) −(1 −λ)f(s) ≤ −f(t) multiply both sides by −1

→ (1 −λ)f(u) −(1 −λ)f(s) ≤ f(u) −f(t) add f(u) to both sides

→ (1 −λ)(f(u) −f(s)) ≤ f(u) −f(t) add f(u) to both sides

→

u−t

u−s

(f(u) −f(s)) ≤ f(u) −f(t) deﬁnition of λ

Dividing both sides of this last equation by u −t gives us

f(u) −f(s)

u −s

≤

f(u) −f(t)

u −t

(15)

Combining (14) and (15) gives us

f(t) −f(s)

t −s

≤

f(u) −f(s)

u −s

≤

f(u) −f(t)

u −t

which is what we were asked to prove.

Exercise 4.24

To prove that f is convex on (a, b) we must show that

(∀x, y ∈ (a, b), λ ∈ [0, 1]) : f(λx + (1 −λ)y) ≤ λf(x) + (1 −λ)f(y) (16)

To do this, choose an arbitrary x, y ∈ (a, b) (this choice of x and y will remain ﬁxed for the majority of this

proof). Let Λ represent the set of values for λ for which (16) holds. It’s immediately clear that 0 ∈ Λ and 1 ∈ Λ.

We’re also told that f has the property

f

_

x +y

2

_

≤

f(x) +f(y)

2

for all x, y ∈ (a, b) (17)

which is just (16) with λ =

1

2

; therefore

1

2

∈ Λ. To prove that f is convex we must show that [0, 1] ⊂ Λ.

Lemma 1: If j, k ∈ Λ then (j +k)/2 ∈ Λ

Assume that j, k ∈ Λ and let m =

j+k

2

. Proof that m ∈ Λ:

f(mx + (1 −m)y) = f

_

j+k

2

x +

2−j−k

2

y

_

deﬁnition of m

= f

_

jx+kx+2y−jy−ky

2

_

algebra

= f

_

jx−jy+y

2

+

kx−ky+y

2

_

algebra

= f

_

jx+(1−j)y

2

+

kx+(1−k)y

2

_

algebra

j is in Λ, so j ∈ [0, 1], so (jx + (1 −j)y) ∈ (a, b). The same is true for k. So we can apply (17).

≤

f(jx+(1−j)y)+f(kx+(1−k)y)

2

≤

jf(x)+(1−j)f(y)+kf(x)+(1−k)f(y)

2

We can apply (16) because j, k ∈ Λ

=

j+k

2

f(x) +

2−j−k

2

f(y) algebra

= mf(x) + (1 −m)f(y) deﬁnition of m

This shows that f(mx + (1 −m)y) ≤ mf(x) + (1 −m)f(y), therefore m ∈ Λ.

69

Lemma 2: all rationals of the form m/2

n

with 0 ≤ m ≤ 2

n

are members of Λ

This can be proven by induction. Let E be the set of all n for which the lemma is true. We know that

¦0,

1

2

, 1¦ ⊂ Λ so we have 0 ∈ E, 1 ∈ E. Now assume that E contains 1, 2, . . . , k. Choose an arbitrary m such

that 0 ≤ m ≤ 2

k+1

.

case 1: m is even

If m is even then m = 2α for some α ∈ Z, 0 ≤ α ≤ 2

k

and therefore m/2

k+1

= α/2k. By the hypothesis of

induction α/2k ∈ Λ therefore m/2

k+1

∈ Λ.

case 2: m is odd

If m is odd then (m−1)/2 = α and (m+ 1)/2 = β for some α, β ∈ Z with α, β ∈ [0, 2

k

]. From this we have

m

2

k+1

=

m+ 1

2

k+2

+

m−1

2

k+2

=

1

2

_

α

2

k

+

β

2

k

_

By the hypothesis of induction α/2

k

∈ Λ and β/2

k

∈ Λ. Therefore, by lemma 1, m/2

k+1

∈ Λ.

This covers all possible cases for m, so k + 1 ∈ E. This completes the inductive step, therefore E = N.

Lemma 3: every element of [0, 1] is a member of Λ

Choose any p ∈ [0, 1]. From lemma 2, we see that Λ is dense in [0, 1] so we can construct a sequence ¦λ

n

¦ in Λ

such that lim

n→∞

λ

n

= p. Each λ

n

is an element of Λ, so for each n we have

f(λ

n

x + (1 −λ

n

)y) ≤ λ

n

f(x) + (1 −λ

n

)f(y)

The function f is continuous, so by theorem 4.2 we can take the limit of both sides as n → ∞ to get

f(px + (1 −p)y) ≤ pf(x) + (1 −p)f(y)

which means that p ∈ Λ.

By lemma 3 we know that

(∀λ ∈ [0, 1])f(λx + (1 −λ)y) ≤ λf(x) + (1 −λ)f(y)

But we chose x and y arbitrarily from the interval (a, b), so we have proven that

(∀x, y ∈ (a, b), λ ∈ [0, 1])f(λx + (1 −λ)y) ≤ λf(x) + (1 −λ)f(y)

which is (16), the deﬁnition of convexity. This shows that f is convex, which is what we were asked to prove.

Exercise 4.25a

The “hint” given for this problem is actually a proof. The set F = z − C is clearly closed because z is a ﬁxed

element and C is closed. The sets K and F are disjoint:

k ∈ K ∩ F assumed

→ k ∈ K ∧ k ∈ F def. of intersection

→ k ∈ K ∧ (∃c ∈ C)(k = z −c) def. of F

→ k ∈ K ∧ (∃c ∈ C)(k +c = z)

→ (∃k ∈ K, c ∈ C)k +c = z rearrangement of quantiﬁers

→ z ∈ K +C def. of K +C

From the deﬁnition of the set F, we see that (∀c ∈ C)(z − c ∈ F). And we’ve established that F is closed

and that and K is compact, so by exercise 21 we know that there exists some δ > 0 such that

70

(∀c ∈ C, k ∈ K)(d(z −c, k) > δ) by exercise 21

→ (∀c ∈ C, k ∈ K)([z −c −k[ > δ) def. of distance function in R

k

→ (∀c ∈ C, k ∈ K)([z −(c +k)[ > δ) algebra

→ (∀c ∈ C, k ∈ K)(d(z, c +k) > δ) def. of distance function in R

k

The set of c +k for all c ∈ C, k ∈ K is simply the set C +K: so we have

→ (∀y ∈ C +K)(d(z, y) > δ) def. of distance function in R

k

This shows us that z is an interior point of C +K. But z was an arbitrary point of C +K, so we have shown

that C +K is open. By theorem 2.23 the complement of an open set is closed, so C +K is closed. And this is

what we were asked to prove.

Exercise 4.25b

Let α be an irrational number and let C

1

+C

2

be deﬁned as the set

C

1

+C

2

= ¦m+nα : m, n ∈ N¦ (18)

We’ll ﬁrst prove that C

1

+C

2

is dense in [0, 1) and then extend this proof out to prove that C

1

+C

2

is dense in

R

1

. Deﬁne the set ∆ to be the set of radii δ such that, for every x ∈ [0, 1), every neighborhood of radius > δ

contains a point of C

1

+C

2

. More formally, we deﬁne it to be

∆ = ¦δ : (∀x ∈ [0, 1))(N

δ

(x) ∩ C

1

+C

2

,= ∅)

We want to prove that C

1

+C

2

is dense in [0, 1): we can do this by proving that ∆ has a greatest lower bound

of 0.

Lemma 1: Each element of C

1

+C

2

has a unique representation of the form m+nα

Assume that m+nα and p +qα are two ways of describing the same element of C

1

+C

2

.

m+nα = p +qα assumption of equality

→ (m−p) + (n −q)α = 0 algebra

→ (n −q)α = (p −m) algebra

If both sides of this equation are zero, then n = q and p = m so that our two representations are

not unique. If both sides are nonzero, we can divide by p −m:

→ α =

p−m

n−q

algebra

But m, n, p, q are integers so this last statement implies that α is a rational number. By contradiction, each

element of C

1

+C

2

has a unique representation of the form m+nα.

Lemma 2: For each n ∈ N, there is exactly one m ∈ N such that m+nα ∈ [0, 1).

Assume that p +nα and q +nα are both in the interval [0, 1) Then [(p +nα) −(q +nα[ = [p −q[ ∈ [0, 1). And

p and q are both integers, so it must be the case that p = q.

Lemma 3: If d(x, y) = δ for any x, y ∈ [0, 1) ∩ C

1

+C

2

then δ ∈ ∆

Assume d(x, y) = δ for some 0 ≤ x < y < 1 with x, y ∈ C

1

+C

2

. Then we have

d(x, y) = y −x = p +mα −q +nα = (p −q) + (m−n)α = δ

So that δ is itself an element of C

1

+C

2

. It’s also clear that any integer multiple of δ is an element of C

1

+C

2

.

Now, choose an arbitrary p ∈ [0, 1) that is not a multiple of δ. Every real number lies between two integers, so

there exists some a such that

a <

p

δ

< a + 1

71

which implies

aδ < p < (a + 1)δ

which shows that p is in a neighborhood of radius δ of some element of C

1

+C

2

. But p was an arbitrary element

of [0, 1), so every element of [0, 1) lies in such a neighborhood of radius δ, and therefore δ ∈ ∆.

Lemma 4: if δ ∈ ∆, then

1

2

δ ∈ ∆

Proof by contradiction. Assume that δ ∈ ∆ but

1

2

δ ,∈ ∆. By lemma 3 we know that d(x, y) > δ for all

x, y ∈ [0, 1) ∩ C

1

+C

2

. This gives us a maximum size for the set [0, 1) ∩ C

1

+C

2

:

[[0, 1) ∩ C

1

+C

2

[ ≤

1

δ

But this contradicts lemmas 1 and 2, which tell us that there is one unique element of [0, 1) ∩ C

1

+C

2

for each

n ∈ N. By contradiction, then, we must have

1

2

δ ∈ ∆ whenever δ ∈ ∆.

The set C

1

+C

2

is dense in [0, 1)

We can now use induction to show that inf ∆ = 0. We know that 1 ∈ ∆ because a neighborhood of radius δ = 1

around any x ∈ [0, 1) will contain 0 ∈ C

1

+ C

2

. Therefore 1 ∈ ∆. Using induction with lemma 4 tells us that

(∀n ∈ N)(2

−n

∈ ∆). This is suﬃcient to prove that inf ∆ ≤ 0. Each element of ∆ is a distance, so inf ∆ ≥ 0.

Therefore inf ∆ = 0.

By the deﬁnition of ∆, this means that every neighborhood of every element of [0, 1) has an element of

C

1

+ C

2

. This allows us to conclude that every element of [0, 1) is a limit point of C

1

+ C

2

, which means that

C

1

+C

2

is dense in [0, 1).

The set C

1

+C

2

is dense in R

1

Choose an arbitrary p ∈ R

1

. Let m be the integer such that m ≤ p < m+1. This means that 0 ≤ p−m < 1, and

we know that C

1

+C

2

is dense in [0, 1). Therefore we can construct a sequence ¦c

n

¦ of elements of [0, 1)∩C

1

+C

2

that converges to p − m. The deﬁnition of C

1

+ C

2

guarantees that ¦c

n

+ m¦ is also a sequence in C

1

+ C

2

,

and theorem 3.3 tells us that ¦c

n

+ m¦ converges to p. Therefore p is a limit point of C

1

+ C

2

. But p was an

arbitrary element of R

1

, so every element of R

1

is a limit point of C

1

+ C

2

. And this proves that C

1

+ C

2

is

dense in R

1

.

The sets C

1

and C

2

are closed, but C

1

+C

2

is not closed.

The sets C

1

and C

2

are closed because they have no limit points, so it’s trivially true that they contain all of

their limit points. The set C

1

+C

2

doesn’t contain any non-integer rational numbers, but every real number is

a limit point of C

1

+C

2

. Therefore C

1

+C

2

doesn’t contain all of its limit points which means it is not closed.

Exercise 4.26

We’re told that Y is compact and that g is continuous and one-to-one. We conclude that g(Y ) is a compact

subset of Z (theorem 4.14) and therefore g is uniformly continuous on Y (theorem 4.19). The fact that g is

one-to-one tells us that g is one-to-one and onto g(Y ), so by theorem 4.17 we conclude that g

−1

(g(Y )) is a

continuous mapping from compact space g(Y ) to compact space X. So by theorem 4.19 we see that g

−1

is

uniformly continuous on g(Y ).

In exercise 12 we proved that the composition of uniformly continuous functions is uniformly continuous,

therefore f(x) = g

−1

(h(x)) is uniformly continuous if h is uniformly continuous. Theorem 4.7 tells us that the

composition of continuous functions is continuous, therefore f(x) = g

−1

(h(x)) is continuous if h is continuous.

To construct the counter example, deﬁne X = Z = [0, 1] and Y = [0,

1

2

) ∪ [1, 2]. Deﬁne the functions

f : X → Y and g : Y → Z as

f(x) =

_

x if x <

1

2

2x if x ≥

1

2

72

g(y) =

_

y if y <

1

2

y

2

if y ≥

1

2

We can easily demonstrate that f fails to be continuous at x =

1

2

and that g is continous at every point. The

composite function h : X → Z is just h(x) = x, which is clearly continuous.

Exercise 5.1

Choose arbitrary elements x, y ∈ R

1

. We’re told that f is continuous and that

[f(x) −f(y)[ ≤ (x −y)

2

= [x −y[

2

Dividing the leftmost and rightmost terms by [x −y[ we have

¸

¸

¸

¸

f(x) −f(y)

x −y

¸

¸

¸

¸

≤ [x −y[

Taking the limit of each side as y → x gives us

[f

(x)[ ≤ 0

But our choice of x was arbitrary, so the derivative of f is zero at every point. By theorem 5.11b, this proves

that f is a constant function.

Exercise 5.2

Lemma 1: g(f(x)) = x

We’re told that f

**(x) > 0 for all x, so f is strictly increasing in (a, b) (this result is a trivial extension of the
**

proofs for theorem 5.11). By deﬁnition, this means that x < y ↔ f(x) < f(y) and x > y ↔ f(x) > f(y): from

this we conclude

x ,= y ↔ f(x) ,= f(y)

By contrapositive this is equivalent to

x = y iﬀ f(x) = f(y)

which, by deﬁnition, means that f is one-to-one. Therefore g(f(x)) = x (theorem 4.17 ).

g

**(x) exists for all x ∈ f( (a, b) )
**

Using lemma 1 we see that f

(x) is inversely proportional to g

(x):

f

(x) = lim

t→x

f(x) −f(t)

x −t

=

f(x) −f(t)

g(f(x)) −g(f(t))

= limt → x

1

g(f(x))−g(f(t))

f(x)−f(t)

=

1

g

(x)

We’re told that the derivative f

(x) is deﬁned for all x ∈ (a, b), so 1/g

**(x) must also be deﬁned.
**

bonus proofs: a big wad of properties for f and g

We can show that both f and g are uniformly continous, injective, diﬀerentiable, strictly increasing functions

whose domains and ranges are compact.

From lemma 1, we know that f is a one-to-one function that is strictly increasing on (a, b). We’re told

that f is diﬀerentiable on (a, b), therefore f is continuous on [a, b] (theorem 5.2). Because the domain of f is

a compact metric space we can conclude that f is uniformly continuous (theorem 4.19). From the continuity

and injectiveness of f we can conclude that g is continuous on f( [a, b] ) (theorem 4.17). The domain of g is

the range of f, so the domain of g is a compact space (thoerem 4.14) and therefore g is uniformly continuous

(theorem 4.19). From lemma 1 we know that g(f(x)) = x for all x, therefore the inequalities in lemma 1 are

equivalent to

g(f(x)) > g(f(y)) ↔ f(x) > f(y)

g(f(x)) < g(f(y)) ↔ f(x) < f(y)

g(f(x)) = g(f(y)) ↔ f(x) = f(y)

Which means that g is a one-to-one function and is strictly increasing on f( (a, b) ).

73

Exercise 5.3

Choose such that [[ <

1

M

. The function f is the sum of two diﬀerentiable functions, so by theorem 5.3 f is

itself diﬀerentiable:

f

(x) = lim

t→x

x −t

x −t

+ lim

t→x

g(x) −g(t)

x −t

= 1 +g

(x)

From our bounds on and g

(x) this gives us

1 −

_

1

M

M

_

< f

(x) < 1 +

_

1

M

M

_

These are strict inequalities, so we conclude that f

**(x) is always positive for this choice of . Therefore f is an
**

increasing function and is one-to-one (see lemma 1 of the previous exercise).

Exercise 5.4

Consider the function

f(x) = C

0

x +

C

1

x

2

2

+ +

C

n

x

n+1

n + 1

When x = 1 this evaluates to the function given to us in the exercise, so f(1) = 0. When x = 0 every term

evaluates to zero, so f(0) = 0. From example 5.4 we know that f(x) is diﬀerentiable and its derivative is given

by

f

(x) = C

0

+C

1

x +C

2

x

2

+ +C

n

x

n

By theorem 5.10, the fact that f(0) = f(1) = 0 means that f

(x) = 0 for some x ∈ (0, 1). Therefore

C

n

x

n

has a real root in (0, 1), and this is what we were asked to prove.

Exercise 5.5

Choose an arbitrary > 0. We’re told that f

**(t) → 0 as t → ∞, so there exists some N such that
**

t > N → [f

(t)[ <

With some algebraic manipulation and the use of the the mean value theorem (theorem 5.10) we can express g

as

g(x) = f(x + 1) −f(x) =

f(x + 1) −f(x)

(x + 1) −(x)

= f

(t) for some t ∈ (x, x + 1)

This must be true for all possible values of x, so choose x > N. We now have t > x > N, so the f

(t) term in

the previous equation is now less than .

[g(x)[ = [f

(t)[ ≤

which means that [g(x)[ < for all x > N. And was an arbitrary positive real, which by deﬁnition means that

g(x) → 0 as x → ∞. And this is what we were asked to prove.

Exercise 5.6

The function g is diﬀerentiable (theorem 5.3c) and its derivative is

g

(x) =

xf

(x) −f(x)

x

2

We want to prove that g is monotonically increasing. This is true iﬀ g

**(x) > 0 for all x, which is true iﬀ
**

xf

(x) −f(x)

x

2

> 0

which is true iﬀ

xf

(x) > f(x)

which is true iﬀ

f

(x) >

f(x)

x

(19)

74

To show that (19) holds for all x, choose an arbitrary x ∈ R. From the fact that f(0) = 0 we know from

theorem 5.10 that

f(x)

x

=

f(x) −f(0)

x −0

= f

(c) for some c ∈ (0, x)

We’re told that f

is monotonically increasing, so f

(x) > f

(c). Therefore:

f

(x) > f

(c) =

f(x)

x

Therefore (19) holds, and we’ve shown that this occurs iﬀ g

**(x) > 0, which means that g is monotonically
**

increasing. And this is what we were asked to prove.

Exercise 5.7

From the fact that f(x) = g(x) = 0 we see that

lim

t→x

f(t)

g(t)

= lim

t→x

f(t) −f(x)

g(t) −g(x)

= lim

t→x

f(t) −f(x)

t −x

t −x

g(t) −g(x)

=

f

(x)

g

(x)

Exercise 5.8

We’re told that f

is a continuous function on the compact space [a, b] therefore f

is uniformly continuous

(theorem 4.19). Choose any > 0: by the deﬁnition of uniform continuity there exists some δ such that

0 < [t − x[ < δ → [f

(t) − f

**(x)[ < . Choose t, x ∈ [a, b]: by the mean value theorem there exists c ∈ (t, x)
**

such that

f

(c) =

f(t) −f(x)

t −x

From the fact that c ∈ (t, x) we know that [c −x[ < [t −x[δ. Therefore [f

(c) −f

**(x)[ < . Therefore, from our
**

deﬁnition of c, we have

[f

(c) −f

(x)[ =

¸

¸

¸

¸

f(t) −f(x)

t −x

−f

(x)

¸

¸

¸

¸

<

Our initial choice of t, x, and was arbitrary, some δ > 0 must exist so that this previous inequality is true for

all t, x, and . And this is what we were asked to prove.

Does this hold for vector-valued functions?

Yes. Choose an arbitrary > 0 and deﬁne the vector-valued function f to be

f(x) = (f

1

(x), f

2

(x), . . . , f

n

(x))

Assume that f is diﬀerentiable on [a, b]. Then each f

i

is diﬀerentiable on [a, b] (remark 5.16) and [a, b] is compact

so by the preceeding proof we know that for each f

i

there exists some δ

i

> 0 such that

[t −x[ < δ

i

→

¸

¸

¸

¸

f

i

(t) −f

i

(x)

t −x

−f

i

(x)

¸

¸

¸

¸

<

n

Deﬁne δ = min¦δ

1

, δ

2

, . . . , δ

n

¦. For [t −x[ < δ we now have

¸

¸

¸

¸

f(t) −f(x)

t −x

−f

(x)

¸

¸

¸

¸

=

¸

¸

¸

¸

(f

1

(t) +f

2

(t) + +f

n

(t)) −(f

1

(x) +f

2

(x) + +f

n

(x))

t −x

−(f

1

(x) +f

2

(x) + +f

n

(x))

¸

¸

¸

¸

=

¸

¸

¸

¸

_

f

1

(t) −f

1

(x)

t −x

−f

1

(x)

_

+

_

f

2

(t) −f

2

(x)

t −x

−f

2

(x)

_

+ +

_

f

n

(t) −f

n

(x)

t −x

−f

n

(x)

_¸

¸

¸

¸

≤

¸

¸

¸

¸

_

f

1

(t) −f

1

(x)

t −x

−f

1

(x)

_¸

¸

¸

¸

+

¸

¸

¸

¸

_

f

2

(t) −f

2

(x)

t −x

−f

2

(x)

_¸

¸

¸

¸

+ +

¸

¸

¸

¸

_

f

n

(t) −f

n

(x)

t −x

−f

n

(x)

_¸

¸

¸

¸

<

n

+

n

+ +

n

=

75

Exercise 5.9

We’re asked to show that f

**(0) exists. From the deﬁnition of the derivative, we know that
**

f

(0) = lim

x→0

f(x) −f(0)

x −0

The function f is continuous, so lim

x→0

f(x) − f(0) = 0 and lim

x→0

x = 0. Therefore we can use L’Hopital’s

rule.

f

(0) = lim

x→0

f(x) −f(0)

x

= lim

x→0

f

(x) −01 −0 = lim

x→0

f

(x)

We’re told that the right-hand limit exists and is equal to 3, therefore the leftmost term (f

(0)) exists and is

equal to 3. And this is what we were asked to prove.

Exercise 5.9 : Alternate proof

If lim

x→0

f

(0) = 3 and f

(0) ,= 3, then f

would have a simple discontinuity at x = 0. Therefore f

(0) = 3 as

an immediate consequence of the corollary to theorem 5.12.

Exercise 5.10

Let f

1

, g

1

represent the real parts of the functions f, g and let f

2

, g

2

represent their imaginary parts: that is,

f(x) = f

1

(x) + if

2

(x) and g(x) = g

1

(x) + ig

2

(x). We’re told that f and g are diﬀerentiable, therefore each of

these dependent functions is diﬀerentiable (see Rudin’s remark 5.16). Applying the hint given in the exercise,

we have

lim

x→0

f(x)

g(x)

= lim

x→0

_

f

1

(x) +if

2

(x)

x

−A

_

x

g

1

(x) +ig

2

(x)

+A

x

g

1

(x) +ig

2

(x)

Each of these functions is diﬀerentiable and the denominators all tend to 0, so we can apply L’Hopital’s rule.

lim

x→0

f(x)

g(x)

= lim

x→0

_

f

1

(x) +if

2

(x)

1

−A

_

1

g

1

(x) +ig

2

(x)

+A

1

g

1

(x) +ig

2

(x)

= lim

x→0

_

f

(x)

1

−A

_

1

g

(x)

+A

1

g

(x)

= ¦A−A¦

1

B

+A

1

B

=

A

B

Exercise 5.11

The denominator of the given ratio tends to 0 as h → 0, so we can use L’Hopital’s rule (diﬀerentiating with

respect to h):

lim

h→0

f(x +h) +f(x −h) −2f(x)

h

2

= lim

h→0

f

(x +h) −f

(x −h)

2h

= lim

h→0

1

2

_

f

(x +h) −f

(x)

h

+

f

(x) −f

(x −h)

h

_

We’re told that f

**(x) exists, so this limit exists and is equal to
**

= lim

h→0

1

2

(f

(x) +f

(x)) = f

(x)

A function for which this limit exists although f

(x) does not

For a counterexample, we need only ﬁnd a diﬀerentiable function for which f

(x) = 1 when x > 0 and f

(x) = −1

when x < 0. These criteria are met by

f(x) =

_

_

_

x

2

, if x > 0

−(x

2

), if x < 0

0, if x = 0

This function is continuous and diﬀerentiable, but f

**(x) does not exist at x = 0.
**

76

Exercise 5.12

From the deﬁnition of f

(x), we have

f

(x) = lim

h→0

[x +h[

3

−[x[

3

h

(20)

If x > 0 then the terms in the numerator are positive and (20) resolves to

f

(x) = lim

h→0

(x +h)

3

−(x)

3

h

= lim

h→0

3x

2

+ 3xh +h

2

= 3x

2

If x < 0 then the terms in the numerator are negative and (20) resolves to

f

(x) = lim

h→0

−([x[ +h)

3

+ ([x[)

3

h

= lim

h→0

−(3[x[

2

+ 3[x[h +h

2

) = −(3[x

2

[)

It’s clear from the above results that f

(x) → 0 as x → 0, and this agrees with f

(0):

f

(0) = lim

h→0

[h

3

[

h

= 0

So f

(x) = 3x[x[ for all x.

From the deﬁnition of f

(x), we have

f

(x) = lim

h→0

[3(x +h)

2

[ −[3x

2

[

h

(21)

If x > 0 then the terms in the numerator are positive and (21) resolves to

f

(x) = lim

h→0

3(x +h)

2

−3x

2

h

= lim

h→0

6x + 3h = 6x

If x < 0 then the terms in the numerator are negative and (21) resolves to

f

(x) = lim

h→0

−3([x[ +h)

2

−3[x

2

[

h

= lim

h→0

6[x[ + 3h = 6[x[

It’s clear from the above results that f

(x) → 0 as x → 0, and this agrees with f

(0):

f

(0) = lim

h→0

[3h

2

[

h

= 0

So f

(x) = [6x[ for all x.

From the deﬁnition of f

(x), we have

f

(x) = lim

h→0

[6(x +h)[ −[6x[

h

(22)

If x = 0, then when h > 0 we have

lim

h→0

[6h[

h

= 6

and when h < 0 we have

lim

h→0

[6h[

h

= −6

So the limit in (22) (which is f

(3)

(0)) doesn’t exist.

Exercise 5.13a

f is continuous when x ,= 0

The proof that x

a

is continuous when x ,= 0 is trivial. The sin function hasn’t been well-deﬁned yet, but we can

assume that it’s a continuous function

5

. Therefore their product x

a

sin(x

−c

) is continuous wherever it’s deﬁned

(theorem 4.9), which is everywhere but x = 0.

5

Example 5.6 says that we should assume without proof that sin is diﬀerentiable, so we can also assume that it’s continuous

(theorem 5.2).

77

f is continous at x = 0 if a > 0

We have f(0) = 0 by deﬁnition. For f to be continuous at x = 0 it must be the case that lim

x→0

f(x) = 0. The

range of the sin function is [−1, 1], so if a > 0 we have

x

a

(−1) ≤ x

a

sin(x

−c

) ≤ x

a

(1)

Taking the limit of each of these terms as x → 0 gives us

0 ≤ lim

x→0

x

a

sin(x

−c

) ≤ 0

which shows that lim

x→0

f(x) = 0, and therefore f is continuous at x = 0.

f is discontinous at x = 0 if a = 0

To show that f is not continuous at x = 0, it’s suﬃcient to construct a sequence ¦x

n

¦ such that limx

n

= 0 but

limf(x

n

) ,= f(0) (theorem 4.2). Deﬁne the terms of ¦x

n

¦ to be

x

n

=

_

1

2nπ +π/2

_1

c

This sequence clearly has a limit of 0, but

f(x

n

) = x

0

n

sin(x

−c

n

) = sin(2nπ +π/2) = 1

so that lim¦f(x

n

)¦ = 1. Note that we’re making lots of unjustiﬁed assumptions about the sin function and the

properties of the as-of-yet undeﬁned symbol π.

f is discontinous at x = 0 if a < 0

Deﬁne the terms of ¦x

n

¦ to be

x

n

=

_

1

2nπ +π/2

_1

c

This sequence clearly has a limit of 0, but

f(x

n

) = x

a

n

sin(x

−c

n

) =

_

1

2nπ +π/2

_a

c

sin(2nπ +π/2) =

_

1

2nπ +π/2

_a

c

We’re told that c > 0, therefore a/c < 0 and we have

f(x

n

) =

_

1

2nπ +π/2

_a

c

= (2nπ +π/2)

−a

c

where −a/c > 0. By theorem 3.20a we see that limf(x

n

) = ∞, which means that limx

n

= 0 but limf(x

n

) ,= 0

and therefore f is not continuous at x = 0.

These cases show that f is continous iﬀ a > 0.

Exercise 5.13b

From the deﬁnition of limit we have

f

(0) = lim

h→0

f(0 +h) −f(0)

h

= lim

h→0

h

a

sin(h

−c

) −0

h

= lim

h→0

h

a−1

sin(h

−c

) (23)

We can evaluate the rightmost term by noting that sin is bounded by [−1, 1] so that

[h

a−1

[(−1) ≤ h

a−1

sin(h

−c

) ≤ [h

a−1

[(1) (24)

78

Lemma 1: f is diﬀerentiable when x ,= 0

The proof that x

a

and x

−c

are diﬀerentiable when x ,= 0 is trivial. The sin function hasn’t been well-deﬁned yet,

but Rudin asks us in example 5.6 to assume that it’s diﬀerentiable. Therefore sin(x

−c

) is diﬀerentiable when

x ,= 0 (theorem 5.5, the chain rule) and therefore x

a

sin(x

−c

) is diﬀerentiable when x = 0 (theorem 5.3b, the

product rule).

case 1: f

(0) exists when a > 1

When a > 1 we have a −1 > 0 and therefore taking the limits of (24) as h → 0 gives us

0 ≤ lim

h→0

[h

a−1

sin(h

−c

)[ ≤ 0

which means that (23) becomes

f

(0) = lim

h→0

h

a−1

sin(h

−c

) = 0

which shows that f

(0) is deﬁned.

case 2: f

**(0) does not exist when a = 1
**

Deﬁne the sequences ¦h

n

¦ and ¦j

n

¦ such that

h

n

=

_

1

2nπ +

π

2

_

1/c

j

n

=

_

1

(2n + 1)π +

π

2

_

1/c

When a = 1 we have a −1 = 0 and therefore equation (23) gives us

lim

h→0

h

a−1

sin(h

−c

) = lim

n→∞

sin(h

−c

n

) = 1

lim

j→0

j

a−1

sin(j

−c

) = lim

n→∞

sin(j

−c

n

) = −1

We know the sequences ¦f

(h

n

)¦ and ¦j

(h

n

)¦ are well-deﬁned because of lemma 1, therefore we have conﬂicting

deﬁnitions of f

(0). This means that the limit in (23) (and therefore f

**(0) itself) does not exist.
**

case 3: f

**(0) does not exist when a < 1
**

Deﬁne the sequences ¦h

n

¦ and ¦j

n

¦ such that

h

n

=

_

1

2nπ +

π

2

_

1/c

j

n

=

_

1

(2n + 1)π +

π

2

_

1/c

When a < 1 we have a −1 < 0 and therefore equation (23) gives us

lim

h→0

h

a−1

sin(h

−c

) = lim

n→∞

h

a−1

n

= ∞

lim

j→0

j

a−1

sin(j

−c

) = lim

n→∞

−j

a−1

n

= −∞

We know the sequences ¦f

(h

n

)¦ and ¦j

(h

n

)¦ are well-deﬁned because of lemma 1, therefore we have conﬂicting

deﬁnitions of f

(0). This means that the limit in (23) (and therefore f

**(0) itself) does not exist.
**

These cases show that f

(0) exists iﬀ a > 1.

79

Exercise 5.13c

Note that we’ve only deﬁned f on the domain [−1, 1], so we only need to show that f is bounded on this domain.

case 1: f

is unbounded when a < 1

We saw in case (3) of part (b) that f

**is unbounded near 0 when a < 1.
**

case 2: f

is unbounded when 1 ≤ a < c + 1

By the lemma of part (b) we know that f

**(x) is deﬁned for all x ∈ [1, 1] except for possibly x = 0. By the chain
**

rule and product rule we know that the derivative of f when x ,= 0 is

f

(x) = ax

a−1

sin(x

−c

) −cx

a−(c+1)

cos(x

−c

) (25)

Deﬁne the sequence ¦h

n

¦ such that

h

n

= (2nπ)

−1/c

Evaluating the derivative in (25) at x = h

n

gives us

f

(h

n

) = (2nπ)

1−a

c

sin(2nπ) −c(2nπ)

(1+c)−a

c

cos(2nπ) = −c(2nπ)

(1+c)−a

c

We’re assuming in this case that a < c + 1, so taking the limits of this last equation as n → ∞ gives us

lim

n→∞

f

(h

n

) = lim

n→∞

−c(2nπ)

(1+c)−a

c

= −∞

This doesn’t prove anything about f

(0) itself, but it does show that f

(x) is unbounded near 0.

case 3: f

is bounded when a ≥ c + 1

If a ≥ c + 1 then clearly a > 1, so by the lemma of part (b) we know that f

**(x) is deﬁned for all x ∈ [1, 1]
**

including x = 0. By the chain rule and product rule we know that the derivative of f is

f

(x) = ax

a−1

sin(x

−c

) −cx

a−(c+1)

cos(x

−c

) (26)

Since f

**(x) is deﬁned for x = 0, we can take the limit of (26) as x → 0:
**

f

(0) = lim

x→0

f

(x) = lim

x→0

ax

a−1

sin(x

−c

) −cx

a−(c+1)

cos(x

−c

)

Because x is bounded by [−1, 1] we know that x

a−1

, x

a−(c+1)

, sin, and cos are all bounded by [−1, 1]. Therefore

the rightmost limit of the previous equation is bounded by

−(a +c) ≤ lim

x→0

ax

a−1

sin(x

−c

) −cx

a−(c+1)

cos(x

−c

) ≤ a +c

Which, of course, means that f

**(x) is also bounded by [−(a + c), a + c] for x ∈ [−1, 1]. We could ﬁnd stricter
**

bounds for f

**(x), but it’s not necessary.
**

These three cases show that f

is bounded iﬀ a ≥ c + 1.

Exercise 5.13d

lemma: f

is continuous when x ,= 0

From lemma 1 of part (b) we know that f

**exists for all x ,= 0 and its derivative is given by
**

f

(x) = ax

a−1

sin(x

−c

) −cx

a−(c+1)

cos(x

−c

) (27)

Rudin asks us to assume that sin and cos are continuous functions and it’s trivial to show that x

±α

is continuous

when x ,= 0 for any α, so we can use the chain rule (theorem 5.5) and product rule (theorem 5.3b) to show that

f

is continuous when x ,= 0.

80

case 1: f

is continuous at x = 0 when a > 1 +c

We’ve shown that f

(0) = 0 when a > 1 (case 1 of part (b)). For f

**to be continuous at x = 0 it must be the
**

case that lim

x→0

f

**(x) = 0. We can algebraically rearrange (27) to obtain
**

f

(x) = x

a−(c+1)

ax

c

sin(x

−c

) −c cos(x

−c

)

The range of the cosine and sine functions are [−1, 1], so we can establish a bound on this function.

[f

(x)[ = [x

a−(c+1)

ax

c

sin(x

−c

) −c cos(x

−c

)[ ≤ [x

a−(c+1)

[ [ax

c

+c[ ≤ [x

a−(c+1)

[ ([ax

c

[ +[c[)

Because a > c + 1, we have [x

a−(c+1)

[ → 0 and [ax

c

[ → 0 as x → 0. Taking the limits of this last inequality as

x → 0 therefore gives us

lim

x→0

[f

(x)[ ≤ 0 (0 +c) = 0

This shows that lim

x→0

f

(x) = f

(0), therefore f

is continuous at x = 0.

case 2: f

**is not continuous at x = 0 when a = 1 +c
**

To show that f

**is not continuous at x = 0 it’s suﬃcient to construct a sequence ¦x
**

n

¦ such that limx

n

= 0 but

limf

(x

n

) ,= f

**(0) = 0 (theorem 4.2). Deﬁne the terms of x
**

n

to be

x

n

=

_

1

2nπ

_1

c

This sequence clearly has a limit of 0, but sin(x

n

) = 0 and cos(x

n

) = 1 so that the terms of ¦f

(x

n

)¦ are

f

(x

n

) = ax

n

(0) −cx

0

(1) = −c

so that lim¦f

(x

n

)¦ = −c ,= f

(0).

case 3: f

**is not continuous at x = 0 when a < 1 +c
**

If a < 1+c we know that f

is not bounded on [−1, 1] (part (c)) therefore f

**is not continuous on [−1, 1] (theorem
**

4.15). From lemma 1, we know that the discontinuity must occur at the point x = 0.

For an alternative proof we could use the sequence ¦x

n

¦ established in case 2 and show that f

(x

n

) → ∞ , =

f

(0) as x

n

→ 0 when a < 1 +c.

Exercise 5.13e

Lemma 1: f

is diﬀerentiable when x ,= 0

We established in part (b) that f

**exists for x ,= 0 and is given by
**

f

(x) = ax

a−1

sin(x

−c

) −cx

a−(c+1)

cos(x

−c

) (28)

We know that all of the exponential powers of x are diﬀerentiable when x ,= 0 and Rudin asks us to assume that

sin and cos are diﬀerentiable, so we can use theorem 5.5 (the chain rule) and theorem 5.3(the product rule) to

show that f

is diﬀerentiable when x ,= 0.

case 1: f

(0) exists when a > 2 +c

From the deﬁnition of limit we know that

f

(0) = lim

h→0

f

(0 +h) −f

(0)

h

= lim

h→0

ah

a−1

sin(h

−c

) −ch

a−(c+1)

cos(h

−c

)

h

= lim

h→0

_

(ah

a−2

) sin(h

−c

) −(ch

a−(c+2)

) cos(h

−c

)

_

(29)

81

The range of the sin and cos functions is [−1, 1] so we can establish bounds for the limited term.

−([ah

a−2

[ +[ch

a−(c+2)

[) ≤ ah

a−2

sin(h

−c

) −ch

a−(c+2)

cos(h

−c

) ≤ [ah

a−2

[ +[ch

a−(c+2)

[

When a > (2 +c) > 2, the powers of h tend tend to zero as h → 0. Taking the limits of the previous inequality

as h → 0, we have

0 ≤ lim

h→0

ah

a−2

sin(h) −ch

a−(c+2)

cos(h

−c

) ≤ 0

This shows that the limit in (29) (and therefore f

(0)) exists.

case 2: f

**(0) does not exist when a = 2 +c
**

Deﬁne the sequences ¦h

n

¦ and ¦j

n

¦ such that

h

n

=

_

1

2nπ

_

1/c

j

n

=

_

1

(2n + 1)π

_

1/c

When a = 2 +c we have a −(c + 2) = 0 and therefore equation (28) gives us

lim

h→0

f

(h) = lim

n→∞

f

(h

n

) =

_

0 −(ch

0

n

)(1)

¸

= −c

lim

j→0

f

(j) = lim

n→∞

f

(j

n

) =

_

0 −(cj

0

n

)(−1)

¸

= c

We know the sequences ¦f

(h

n

)¦ and ¦j

(h

n

)¦ are well-deﬁned because of lemma 1, therefore we have conﬂicting

deﬁnitions of f

(0). This means that the limit in (28) (and therefore f

**(0) itself) does not exist.
**

case 2: f

**(0) does not exist when a < 2 +c
**

Deﬁne the sequences ¦h

n

¦ and ¦j

n

¦ such that

h

n

=

_

1

2nπ

_

1/c

j

n

=

_

1

(2n + 1)π

_

1/c

When a < 2 +c we have a −(c + 2) < 0 and therefore h

a−(c+2)

n

→ ∞. This means that equation (28) gives us

lim

h→0

f

(h) = lim

n→∞

f

(h

n

) =

_

0 −(ch

a−(c+2)

n

)(1)

_

= −∞

lim

j→0

f

(j) = lim

n→∞

f

(j

n

) =

_

0 −(cj

a−(c+2)

n

)(−1)

_

= ∞

We know the sequences ¦f

(h

n

)¦ and ¦j

(h

n

)¦ are well-deﬁned because of lemma 1, therefore we have conﬂicting

deﬁnitions of f

(0). This means that the limit in (28) (and therefore f

**(0) itself) does not exist.
**

These three cases show that f

(0) exists iﬀ a > 2 +c.

Exercise 5.13f

Lemma 1: to hell with this

By the lemma of part (e) we know that f

**(x) is deﬁned for all x ∈ [1, 1] except for possibly x = 0. By the chain
**

rule and product rule we know that the derivative of f when x ,= 0 is

f

(x) =

_

(a

2

−a)x

a−2

−c

2

x

a−(2+2c)

_

sin(x

−c

) +

_

(c

2

+c −ca)x

a−(2+c)

−cax

a−(1+c)

_

cos(x

−c

) (30)

And I’m not going to screw around with limits and absolute values of something so annoying to type out, so I’ll

conclude with “the proof is similar to that of part(c)”.

82

Exercise 5.13g

The proof is similar to that of part (d), and the sentiment is similar to that of lemma (1) of part (f).

Exercise 5.14

Lemma 1: If f is not convex then the convexity condition fails for all λ for some s, t ∈ (a, b)

This could also be stated as “if f is not convex on (a, b) then it is strictly concave on some interval (s, t) ⊂ (a, b)”.

Assume that f is continuous on the interval (a, b) and that f is not convex. By deﬁnition, this means that we

can choose c, d ∈ (a, b) and λ ∈ (0, 1) such that

f(λc + (1 −λ)d) > λf(c) + (1 −λ)f(d) (31)

Having ﬁxed our choice of c, d, and λ so that the previous equation is true, we deﬁne the function g(x) to be

g(x) = f((x)c + (1 −x)d) −(x)f(c) + (1 −x)f(d)

We know that g is continuous on [0, 1] because f is continuous on [c, d] ⊂ (a, b) (theorem 4.9). And we know

that there is at least one p ∈ [0, 1] such that g(p) > 0 because we can choose p = λ which causes the previous

equation to simplify to

g(p) = f(λc + (1 −λ)d) −[λf(c) + (1 −λ)f(d)]

which is > 0 from (31). Let Z

1

be the set of all p ∈ [0, λ) for which g(p) = 0 and let Z

2

be the set of all

p ∈ (λ, 1] for which g(p) = 0 It’s immediately clear that these sets are nonempty since g(0) = g(1) = 0. These

sets are closed (exercise 4.3) and therefore contain their supremums and inﬁmums. Let α = sup¦Z

1

¦ and let

β = inf¦Z

2

¦. We can now claim that equation (31) holds for all λ ∈ (α, β). And this is the same as saying that

f(λs + (1 −λ)t) > λf(s) + (1 −λ)f(t)

for all λ ∈ (0, 1) for

s = αc + (1 −α)d, t = βc + (1 −β)d

which, by deﬁnition, means that the convexity conditions fails on the inteval (s, t) ∈ (a, b) for all λ ∈ (0, 1).

The “if ” case: f

**is monotonically increasing if f is convex
**

Let f be a function that is convex on the interval (a, b) (see exercise 4.23) and choose x, y ∈ (a, b) with y ≥ x.

By deﬁnition this means that for all x, y ∈ (a, b) and for all λ ∈ (0, 1) it must be the case that

f(λx + (1 −λ)y) ≤ λf(x) + (1 −λ)f(y) (32)

We want to express the left-hand side of this inequality as f(x +h): we can do this by deﬁning h such that

h = (λ −1)(x −y)

Rearranging this algebraically allows us to express λ and 1 −λ as

λ = 1 −

h

y −x

, (1 −λ) =

h

x −y

Substituting these values of λ and λ −1 into (32) results in

f(x +h) ≤ f(x) −

hf(x)

y −x

+

hf(y)

y −x

which is algebraically equivalent to

f(x +h) −f(x)

h

≤

f(y) −f(x)

y −x

83

Equation (32) had to be true for any value of λ ∈ (0, 1). As λ → 1 we see that h → 0. Taking the limit of both

sides of the previous equation as h → 0 gives us

f

(x) ≤

f(y) −f(x)

y −x

(33)

Having established (33), we now want to express the left-hand side of (32) as f(y − h): we can do this by

redeﬁning h such that

h = −(λ)(x −y)

Rearranging this algebraically allows us to express λ and 1 −λ as

λ =

h

y −x

, (1 −λ) = 1 −

h

x −y

Substituting these values of λ and 1 −λ into (32) results in

f(y −h) ≤

hf(x)

y −x

+f(y) −

hf(y)

y −x

which is algebraically equivalent to

f(y) −f(y −h)

h

≥

f(y) −f(x)

y −x

Equation (32) had to be true for any value of λ ∈ (0, 1). As λ → 0 we see that h → 0. Taking the limit of both

sides of the previous equation as h → 0 gives us

f

(y) ≥

f(y) −f(x)

y −x

(34)

Combining equations (33) and (34), we have have shown that

f

(y) ≥ f

(x)

We assumed only that f was convex and that y > x; we concluded that f

(y) ≥ f

**(x). By deﬁnition, this means
**

that f

**is monotonically increasing if f is convex.
**

The “only if ” case: f

**is monotonically increasing only if f is convex
**

Assume that f is not convex. By lemma 1, we can ﬁnd some subinterval (s, t) ∈ (a, b) such that

f(λs + (1 −λ)t) > λf(s) + (1 −λ)f(t) (35)

is true for all λ ∈ (0, 1). We can now follow the logic of the “if” case and deﬁne

h = (λ −1)(s −t)

Rearranging this algebraically allows us to express λ and 1 −λ as

λ = 1 −

h

t −s

, (1 −λ) =

h

s −t

Substituting these values of λ and λ −1 into (35) results in

f(s +h) > f(s) −

hf(s)

t −s

+

hf(t)

t −s

which is algebraically equivalent to

f(s +h) −f(s)

h

>

f(t) −f(t)

t −s

Equation (35) had to be true for any value of λ ∈ (0, 1). As λ → 1 we see that h → 0. Taking the limit of both

sides of the previous equation as h → 0 gives us

f

(s) >

f(t) −f(s)

t −s

(36)

84

Having established (36), we now redeﬁne h such that

h = −(λ)(s −t)

Rearranging this algebraically allows us to express λ and 1 −λ as

λ =

h

t −s

, (1 −λ) = 1 −

h

s −t

Substituting these values of λ and 1 −λ into (35) results in

f(t −h) >

hf(s)

t −s

+f(t) −

hf(t)

t −s

which is algebraically equivalent to

f(t) −f(t −h)

h

<

f(t) −f(s)

t −s

Equation (35) had to be true for any value of λ ∈ (0, 1). As λ → 0 we see that h → 0. Taking the limit of both

sides of the previous equation as h → 0 gives us

f

(t) <

f(t) −f(s)

t −s

(37)

Combining equations (36) and (37), we have have shown that

f

(t) < f

(s)

for some t > s. We assumed only that f was not convex; we concluded that f

**was not monotonically increasing.
**

By contrapositive, this means that f

**is monotonically increasing only if f is convex.
**

f is convex iﬀ f

(x) ≥ 0 for all x ∈ (a, b)

We’ve shown that f is convex iﬀ f

is monotonically inreasing, and theorem 5.11 tells us that f

is monotonically

increasing iﬀ f

(x) ≥ 0 for all x ∈ (a, b).

Exercise 5.15

Note on the bounds of f and its derivatives

When Rudin asks us to assume that [f[ and its derivatives have upper bounds of M

0

, M

1

, and M

2

it appears

that we must assume that these bounds are ﬁnite. Otherwise the function f(x) = x would be a counterexample

to the claim that M

2

1

≤ 4M

0

M

2

.

Proof that M

2

1

≤ 4M

0

M

2

for real-valued functions

Choose any h > 0. Using Taylor’s theorem (theorem 5.15) we can express f(x + 2h) as

f(x + 2h) = f(x) + 2hf

(x) +

4h

2

f

(ξ)

2

, ξ > a (38)

which can be algebraically arranged to give us

f

(x) =

f(x + 2h)

2h

−

f(x)

2h

−hf

(ξ)

[f

(x)[ = [

f(x + 2h)

2h

−

f(x)

2h

−hf

(ξ)[ ≤ [

f(x + 2h)

2h

[ +[

f(x)

2h

[ +[hf

(ξ)[

We’re given upper bounds for [f(x)[ and [f

**(x)[; these bounds give us the inequality
**

[f

(x)[ ≤

2M

0

h

+hM

2

85

This inequality must hold for all x, even when [f(x)[ approaches its upper bound, so we have

M

1

≤

2M

0

h

+hM

2

We can multiply both sides by h and then algebraically rearrange this into a quadratic equation in h.

h

2

M

2

−hM

1

+M

0

≥ 0 (39)

The quadratic solution to this equation is

h =

M

1

±

_

M

2

1

−4M

0

M

2

2M

2

(40)

We want to make sure that there are not two solutions to (40): we want (39) to hold for all values of h, and if

(40) had two solutions then we would have h

2

M

2

− hM

1

+ M

0

< 0 on some interval of h. To make sure that

there is at most a single real solution we need to make sure that the discriminant of (40) is either zero (one

solution) or negative (zero solutions). This occurs exactly when

M

2

1

≤ 4M

0

M

2

Does this apply to vector-valued functions?

Yes. Let f be a vector-valued function that is continuous on (a, ∞) and is deﬁned by

f(x) = ( f

1

(x), f

2

(x), . . . , f

n

(x) )

Assume that [f[, [f

[, and [f

**[ have ﬁnite upper bounds of (respectively) M
**

0

, M

1

, and M

2

. If we evaluate f at

x + 2h we have

f(x + 2h) = ( f

1

(x + 2h), f

2

(x + 2h), . . . , f

n

(x + 2h) )

Taking the Taylor expansion of each of these terms, we have

f(x + 2h) =

_

[f

1

(x) + 2hf

1

(x) + 2h

2

f

1

(ξ

1

)], [f

2

(x) + 2hf

2

(x) + 2h

2

f

2

(ξ

1

)], . . . , [f

n

(x) + 2hf

n

(x) + 2h

2

f

n

(ξ

1

)]

_

= ( f

1

(x), f

2

(x), . . . , f

n

(x) ) + ( 2hf

1

(x), 2hf

2

(x), . . . , 2hf

n

(x) ) +

_

2h

2

f

1

(ξ), 2h

2

f

2

(ξ), . . . , 2h

2

f

n

(ξ)

_

= f(x) + 2hf

(x) + 2h

2

f

(ξ)

This tells us that equation (38) holds for vector-valued functions. But nothing in the proof following equation

(38) requires us to assume that f is real-valued instead of vector-valued. So the proof following (38) still suﬃces

to prove that

M

2

1

≤ 4M

0

M

2

Exercise 5.16 proof 1

In exercise 5.15 we established the inequality

[f

(x)[ ≤ [

f(x + 2h)

2h

[ +[

f(x)

2h

[ +[hf

(ξ)[

We are given an upper bound of M

2

for [f

**(ξ)[, so this inequality can be expressed as
**

[f

(x)[ ≤ [

f(x + 2h)

2h

[ +[

f(x)

2h

[ +hM

2

We’re told that f is twice-diﬀerentiable, so we know that both f and f

**are continuous (theorem 5.2), so we can
**

take the limit of both sides of this equation as x → 0 (theorem 4.2).

lim

x→∞

[f

(x)[ ≤ lim

x→∞

_

[

f(x + 2h)

2h

[ +[

f(x)

2h

[ +hM

2

_

86

We’re told that f(x) → 0 as x → 0 so this becomes

lim

x→∞

[f

(x)[ ≤ lim

x→∞

hM

2

This must be true for all h, so we can take the limit of both sides of this as h → 0 (we can do this because both

[f

(x)[ and hM

2

are continuous functions with respect to the variable h).

lim

h→0

lim

x→∞

[f

(x)[ ≤ lim

h→0

lim

x→∞

hM

2

The left-hand side of the previous inequality is independent of h and therefore doesn’t change; the right-hand

side becomes 0.

lim

x→∞

[f

(x)[ = lim

h→0

lim

x→∞

[f

(x)[ ≤ lim

h→0

lim

x→∞

hM

2

= 0

lim

x→∞

[f

(x)[ ≤ 0

This show that f

**(x) → 0 as x → ∞, which is what we were asked to prove.
**

Exercise 5.16 proof 2

In exercise 5.15 we established the inequality

M

2

1

≤ 4M

0

M

2

where each of the M terms represented a supremum on the interval (a, ∞). This inequality was proven to hold

for all a such that f was twice-diﬀerentiable on the interval (a, ∞). To show more explicitly that the value of

these M terms depends on a, we might express the previous inequality as

sup

x>a

[f

(x)[

2

≤ sup

x>a

4[f(x)[[f

(x)[

Each of these terms is continuous with respect to a (the proof of this claim is trivial but tedious) so we can take

the limit of both sides as a → ∞.

lim

a→∞

sup

x>a

[f

(x)[

2

≤ lim

a→∞

sup

x>a

4[f(x)[[f

(x)[

We’re told that [f

**(x)[ has an upper bound of M
**

2

. We’re also told that f(x) → 0 as x → ∞. The last inequality

therefore allows us to conclude that

lim

a→∞

sup

x>a

[f

(x)[

2

≤ lim

a→∞

sup

x>a

4[0[[M

2

[ = 0

It’s clear that [f

(x)[ must be less than or equal to the supremum of a set containing [f

(x)[, so we have

lim

x→∞

[f

(x)[ ≤ lim

a→∞

sup

x>a

[f

(x)[

2

≤ 0

lim

x→∞

[f

(x)[ ≤ 0

This shows that f

**(x) → 0 as x → ∞, which is what we were asked to prove.
**

Exercise 5.17

We’re told that f is three-times diﬀerentiable on the interval (−1, 1) so we can take the Taylor expansions of

f(−1) and f(1) around x = 0:

f(−1) = f(0) +f

(0)(−1) +

f

(0)(−1)

2

2

+

f

(ξ

1

)(−1)

3

6

, ξ

1

∈ (−1, 0)

f(1) = f(0) +f

(0)(1) +

f

(0)(1)

2

2

+

f

(ξ

2

)(1)

3

6

, ξ

2

∈ (0, 1)

87

When we evaluate f(1) −f(−1), many of these terms cancel out and we’re left with

f(1) −f(−1) = 2f

(0) +

f

(ξ

1

) +f

(ξ

2

)

6

, ξ

1

∈ (−1, 0)ξ

2

∈ (0, 1)

We’re given the values of f(1), f(−1), and f

**(0) so this last equation becomes
**

1 =

f

(ξ

1

) +f

(ξ

2

)

6

, ξ

1

∈ (−1, 0)ξ

2

∈ (0, 1)

which is algebraically equivalent to

f

(ξ

2

) = 6 −f

(ξ

1

), ξ

1

∈ (−1, 0)ξ

2

∈ (0, 1)

If f

(ξ

1

) ≤ 3 then f

(ξ

2

) ≥ 3, and vice-versa: so f

(x) ≥ 3 for either ξ

1

or ξ

2

∈ (−1, 1). And this is what we

were asked to prove.

Exercise 5.18

The nth derivative of f(t)

The exercise tells gives us the following formula for f(t):

f(t) = f(β) −(β −t)Q(t)

Since β is a ﬁxed constant, the ﬁrst two derivatives of f with respect to t are

f

(t) = Q(t) −(β −t)Q

(t)

f

(t) = 2Q

(t) −(β −t)Q

(t)

And, in general, we have

f

(n)

(t) = nQ

(n−1)

(t) −(β −t)Q

(n)

(t)

which, after multiplying by (β −α)

n

/n! and setting t = α, becomes

f

(n)

(t)

n!

(β −α)

n

=

Q

(n−1)

(t)

(n −1)!

(β −α)

n

−

Q

(n)

(t)

n!

(β −α)

n+1

(41)

Modifying the Taylor formula

The formula for the Taylor expansion of f(β) around the point f(α) (theorem 5.15) includes a P(β) term deﬁned

as

P(β) =

n−1

k=0

_

f

(k)

(α)

k!

(β −α)

k

_

If we isolate the k = 0 case this becomes

P(β) = f(α) +

n−1

k=1

_

f

(k)

(α)

k!

(β −α)

k

_

We can now use equation (41) to express the terms of the summation as functions of Q.

P(β) = f(α) +

n−1

k=1

_

Q

(k−1)

(α)

(k −1)!

(β −α)

k

_

−

n−1

k=1

_

Q

(k)

(α)

k!

(β −α)

k+1

_

If we extract the k = 1 term from the leftmost summation and the k = n−1 term from the rightmost summation,

we have

P(β) = f(α) +Q(α)(β −α) +

n−1

k=2

_

Q

(k−1)

(α)

(k −1)!

(β −α)

k

_

−

n−2

k=1

_

Q

(k)

(α)

k!

(β −α)

k+1

_

−

Q

(n−1)

(α)

(n −1)!

(β −α)

n

88

We can then re-index the leftmost summation to obtain

P(β) = f(α) +Q(α)(β −α) +

n−2

k=1

_

Q

(k)

(α)

(k)!

(β −α)

k+1

_

−

n−2

k=1

_

Q

(k)

(α)

k!

(β −α)

k+1

_

−

Q

(n−1)

(α)

(n −1)!

(β −α)

n

The two summations cancel one another, leaving us with

P(β) = f(α) +Q(α)(β −α) −

Q

(n−1)

(α)

(n −1)!

(β −α)

n

If we replace Q(α) with the deﬁnition of Q given in the exercise, we see that this previous equation evaluates to

P(β) = f(α) +

f(α) −f(β)

α −β

(β −α) −

Q

(n−1)

(α)

(n −1)!

(β −α)

n

This simpliﬁes to

P(β) = f(α) + (f(β) −f(α)) −

Q

(n−1)

(α)

(n −1)!

(β −α)

n

A simple algebraic rearrangement of these terms gives us

f(β) = P(β) −

Q

(n−1)

(α)

(n −1)!

(β −α)

n

which is the equation we were asked to derive.

Exercise 5.19a

The given expression for D

n

is algebraically equivalent to

D

n

=

_

f(β

n

) −f(0)

β

n

−0

β

n

β

n

−α

n

_

+

_

f(0) −f(α

n

)

0 −α

n

−α

n

β

n

−α

n

_

We really want to be able to evaluate this by taking the limit of each side of this previous equation in the

following manner:

lim

n→∞

D

n

= lim

n→∞

_

f(β

n

) −f(0)

β

n

−0

_

lim

n→∞

_

β

n

β

n

−α

n

_

+ lim

n→∞

_

f(0) −f(α

n

)

0 −α

n

_

lim

n→∞

_

−α

n

β

n

−α

n

_

(42)

There two conditions that must be met in order for this last step to be justiﬁed:

condition 1: theorem 3.3 tells us that each of the limits must actually exist (and must not be ±∞)

condition 2: and theorem 4.2 tells us that we must have limf(α

n

) = limf(β

n

) = f(0).

The fact that α

n

< 0 < β

n

guarantees that 0 < β

n

/(β

n

− α

n

) < 1 and 0 < −α

n

/(β

n

− α

n

) < 1, which tells us

that at 2 of the 4 limits in (42) exist. The other two limits exist because they’re equal to f

**(0), and we’re told
**

that f

(0) exists. Therefore condition 1 is met. The fact that f

**(0) exists tells us that f is continuous at x = 0
**

(theorem 5.2) and therefore condition 2 is met (theorem 4.2). Therefore we’re justiﬁed in taking the limits in

(42), giving us

lim

n→∞

D

n

= f

(0) lim

n→∞

_

β

n

β

n

−α

n

_

+f

(0) lim

n→∞

_

−α

n

β

n

−α

n

_

= lim

n→∞

f

(0)

β

n

−α

n

β

n

−α

n

= f

(0)

89

Exercise 5.19b

The given expression for D

n

is algebraically equivalent to

D

n

=

_

f(β

n

) −f(0)

β

n

−0

β

n

β

n

−α

n

_

+

_

f(α

n

) −f(0)

α

n

−0

_

1 −

β

n

β

n

−α

n

__

We want to evaluate this by taking the limits of the individual terms as we did in part (a):

lim

n→∞

D

n

= lim

n→∞

_

f(β

n

) −f(0)

β

n

−0

_

lim

n→∞

_

β

n

β

n

−α

n

_

+ lim

n→∞

_

f(α

n

) −f(0)

α

n

−0

_

lim

n→∞

__

1 −

β

n

β

n

−α

n

__

(43)

In order for this to be justiﬁed we must once again meet the two conditions mentioned in part (a). We’re told

that 0 < β

n

/(β

n

− α

n

) < M for some M, which tells us that at 2 of the 4 limits in (43) exist. The other

two limits exist because they’re equal to f

(0), and we’re told that f

**(0) exists. Therefore condition 1 is met.
**

The fact that f

**(0) exists tells us that f is continuous at x = 0 (theorem 5.2) and therefore condition 2 is met
**

(theorem 4.2). Therefore we’re justiﬁed in taking the limits in (43), giving us

lim

n→∞

D

n

= f

(0) lim

n→∞

_

β

n

β

n

−α

n

_

+f

(0) lim

n→∞

_

1 −

β

n

β

n

−α

n

_

= f

(0) lim

n→∞

_

β

n

β

n

−α

n

+ 1 −

β

n

β

n

−α

n

_

= f

(0)

Exercise 5.19c proof 1

Deﬁne the sequence ¦h

n

¦ where h

n

= β

n

−α

n

. We can now express D

n

as

D

n

=

f(α

n

+h

n

) −f(α

n

)

h

n

We know that α

n

→ 0 and β

n

→ 0 as n → ∞, so clearly h

n

→ 0 as n → ∞. We’re also told that f

is continuous

on (−1, 1), so we know that f

**is deﬁned on this interval. Therefore we have
**

lim

n→∞

D

n

= lim

n→∞

lim

h→0

f(α

n

+h

n

) −f(α

n

)

h

n

= lim

n→∞

f

(α

n

)

We’re told that f

**is continuous on the interval (−1, 1) and that limα
**

n

= 0 so by theorem 4.2 we have

lim

n→∞

D

n

= lim

n→∞

f

(α

n

) = f

(0)

Exercise 5.19c proof 2

The mean value theorem (theorem 5.10) allows us to construct a sequence ¦γ

n

¦ as follows: for each n, choose

some γ

n

∈ (α

n

, β

n

) such that

D

n

=

f(β

n

) −f(α

n

)

β

n

−α

n

= f

(γ

n

) (44)

Each γ

n

is in the interval (α

n

, β

n

). We’re told that limα

n

= limβ

n

= 0. Therefore we can use the squeeze

theorem to determine limγ

n

.

0 = lim

n→∞

α

n

≤ lim

n→∞

γ

n

≤ lim

n→∞

β

n

= 0

which means that limγ

n

→ 0. We’re told that f

**is continuous so by theorem 5.2 we can take the limit of
**

equation (44).

lim

n→∞

D

n

= lim

n→∞

f

(γ

n

) = f

(0)

90

Exercise 5.20

Exercise 5.21

We saw in exercise 2.29 that any open set in R

1

can be represented as a countable number of disjoint open

segments. Let ¦(a

n

, b

n

)¦ represent such a countable collection of disjoint sets such that

(a

i

, b

i

) = E

C

. Deﬁne

f : R

1

→R

1

as

f(x) =

_

¸

¸

_

¸

¸

_

0, x ∈ E

(x −b

i

)

2

, x ∈ (a

i

, b

i

) ∧ a

i

= −∞

(x −a

i

)

2

, x ∈ (a

i

, b

i

) ∧ b

i

= ∞

(x −a

i

)

2

(x −b

i

)

2

, x ∈ (a

i

, b

i

) ∧ −∞ < a

i

< b

i

< ∞

It’s easy but tedious to verify that this a continuous function which is diﬀerentiable and that f(x) = 0 iﬀ

x ∈ E. Note that the derivative also has the property that f

**(x) = 0 iﬀ x ∈ E. For the second part of the
**

exercise we can deﬁne a function f to be

f(x) =

_

¸

¸

_

¸

¸

_

0, x ∈ E

(x −b

i

)

n+1

, x ∈ (a

i

, b

i

) ∧ a

i

= −∞

(x −a

i

)

n+1

, x ∈ (a

i

, b

i

) ∧ b

i

= ∞

(x −a

i

)

n+1

(x −b

i

)

n+1

, x ∈ (a

i

, b

i

) ∧ −∞ < a

i

< b

i

< ∞

It’s similarly easy but tedious to verify that this is a continuous function that is n times diﬀerentiable and

that f(x) = 0 iﬀ x ∈ E. It can also be seen that when k ≤ n we have f

(k)

(x) = 0 iﬀ x ∈ E. Finally

6

, we can

pretend that we’ve deﬁned the exponential function and deﬁne f to be

f(x) =

_

¸

¸

¸

¸

¸

_

¸

¸

¸

¸

¸

_

0, x ∈ E

exp

_

−1

(x−bi)

2

_

, x ∈ (a

i

, b

i

) ∧ a

i

= −∞

exp

_

−1

(x−ai)

2

_

, x ∈ (a

i

, b

i

) ∧ b

i

= ∞

exp

_

−1

(x−ai)

2

(x−bi)

2

_

, x ∈ (a

i

, b

i

) ∧ −∞ < a

i

< b

i

< ∞

It’s fairly easy to conﬁrm that this function is continuous and that f(x) = 0 iﬀ x ∈ E. Looking at the

derivative with respect to x we have

f

(x) =

_

¸

¸

¸

¸

¸

_

¸

¸

¸

¸

¸

_

0, x ∈ E

−2

(x−bi)

3

exp

_

−1

(x−bi)

2

_

, x ∈ (a

i

, b

i

) ∧ a

i

= −∞

−2

(x−ai)

3

exp

_

−1

(x−ai)

2

_

, x ∈ (a

i

, b

i

) ∧ b

i

= ∞

−2(bi−ai)

(x−ai)

2

(x−bi)

2

exp

_

−1

(x−ai)

2

(x−bi)

2

_

, x ∈ (a

i

, b

i

) ∧ −∞ < a

i

< b

i

< ∞

To calculate the limit of f

**(x) in the ﬁrst case, we use L’Hopital’s rule.
**

lim

x→a

f

(x) = lim

x→a

exp

_

−1

(x−bi)

2

_

(x −b

i

)

3

The numerator and denominator of this last term both tend to ±∞, so L’Hopital’s rule is applicable.

Repeated applications of L’Hopital’s rule will eventually give us a constant term in the numerator and a term

that tends to ±∞in the denominator, so we see that limf

**(x) = 0. Similar results hold for the limits in the other
**

two cases. The general idea is that, for any polynomial term p(n), the exponential limit lim

n→∞

exp(−1/n)

will tend towards zero faster than lim

n→∞

p(n) tends towards ∞. Therefore p(n)exp(−1/n) will tend to zero

as n → ∞. Every term of every derivative of f(x) will consist only of polynomial multiples of the exponential

term, so it will hold that for all k:

lim

x→ai

f

(k)

(x) = lim

x→bi

f

(k)

(x) = 0

It seems clear in a vague, poorly-deﬁned Calc II-ish way that f is inﬁnitely diﬀerentiable and that, for all n,

f

(n)

(x) = 0 iﬀ x ∈ E. I have no idea how to prove this.

6

this example was provided by Boris Shekhtman, University of South Florida

91

Exercise 5.22a

Assume that f(x

1

) = x

1

and f(x

2

) = x

2

with x

1

,= x

2

. Then

f(x

2

) −f(x

1

)

x

2

−x

1

=

x

2

−x

1

x

2

−x

1

= 1

By theorem 5.10 this means that f

**(x) = 1 for some x between x
**

1

and x

2

. This contradicts the presumption

that f

**(x) ,= 1, so our initial assumption must be wrong: there cannot be two ﬁxed points.
**

Exercise 5.22b

For f(t) to have a ﬁxed point it would be necessary that f(t) = t, in which case we would have

t = t + (1 −e

t

)

−1

−→

1

1 −e

t

= 0

This statement is not true for any t.

Exercise 5.22c

Choose an arbitrary value for x

0

and let ¦x

n

¦ be the sequence recursively deﬁned by x

n+1

= f(x

n

). We’re told

that [f

**(t)[ ≤ A for all real t. So for any n we have
**

¸

¸

¸

¸

f(x

n−1

) −f(x

n−2

)

x

n−1

−x

n−2

¸

¸

¸

¸

≤ A −→ [f(x

n−1

) −f(x

n−2

)[ ≤ A[x

n−1

−x

n−2

[

which, since f(x

n

) = x

n+1

, gives us

[x

n

−x

n−1

[ ≤ A[x

n−1

−x

n−2

[

But this holds for all n. So:

[x

n

−x

n−1

[ ≤ A[x

n−1

−x

n−2

[ ≤ A

2

[x

n−2

−x

n−3

[ ≤ A

3

[x

n−3

−x

n−4

[ ≤ ≤ A

n−2

[x

2

−x

1

[

We can use this general formula to determine the diﬀerence between x

n+k

and x

n

.

[x

n+k

−x

n

[ = [(x

n+k

−x

n+k−1

) + (x

n+k−1

−x

n+k−2

) + + (x

n+2

−x

n+1

) + (x

n+1

−x

n

)[

≤ [x

n+k

−x

n+k−1

[ +[x

n+k−1

−x

n+k−2

[ + +[x

n+2

−x

n+1

[ +[x

n+1

−x

n

)[

≤ A

n+k−2

[x

2

−x

1

[ +A

n+k−3

[x

2

−x

1

[ + +A

n

[x

2

−x

1

[ +A

n−1

[x

2

−x

1

)[

≤ A

n

[x

2

−x

1

[

This shows us that

[x

n+k

−x

n

[ ≤ A

n

[x

2

−x

1

[

We’re told that A < 1, so taking the limit of eachs ide of this inequality as n → ∞ gives us

lim

n→∞

[x

n+k

−x

n

[ ≤ lim

n→∞

A

n

[x

2

−x

1

[ = 0

And this is just the Cauchy criterion for convergence for the sequence ¦x

n

¦. And we’re converging in R

1

so by

theorem 3.11 we know that the sequence converges to some element x ∈ R

1

.

∃x ∈ R : lim

n→∞

x

n

= x

But the elements of ¦x

n

¦ are just the elements of ¦f(x

n

)¦, so we can also conclude that

∃x ∈ R : lim

n→∞

f(x

n

) = x

And we’re told that f is continuous, so by theorem 4.6 we see that

∃x ∈ R : f(x) = x

92

Exercise 5.23a

If x < α, then we can express x as x = α −δ for some δ > 0. This gives us

f(x) = f(α −δ) deﬁnition of x

=

(α−δ)

3

−1

3

deﬁnition of f

=

α

3

−3α

2

δ+3αδ

2

−δ

3

+1

3

algebra

=

α

3

+1

3

+

−3α

2

δ+3αδ

2

3

−

δ

3

3

algebra

= f(α) −αδ(α −δ) −

1

3

δ

3

algebra

= f(α) −αδx −

1

3

δ

3

deﬁnition of x

= α −αδx −

1

3

δ

3

α is a ﬁxed point

We know that x < α < −1, so αx > 1 and therefore αδx > δ. From this we have an inequality:

< α −δ −

1

3

δ

3

< α −δ

= x deﬁnition of x

This establishes that x < α → f(x) < x. Therefore if x

1

< α then the sequence ¦x

n

¦ will be a nonincreasing

sequence. We know that this sequence doesn’t converge because we’re told that f has no ﬁxed points less than

α.

Exercise 5.23b

The ﬁrst derivative of f is f

(x) = x

2

. This is nonnegative, so we know that f is monotonically increasing

(theorem 5.11). The second derivative is f

**(x) = 2x. This is negative for x < 0 and positive for x > 0: therefore
**

f is strictly convex on (0, γ) and is strictly concave on (α, 0). Now let x

k

be chosen from the interval (α, γ).

Case 1: x

k

∈ (α, 0)

Let x

k

∈ (α, 0) be chosen. We can express this x as x

k

= λα+(1−λ)0 for some λ ∈ (0, 1). The second derivative

of f

**(x) = 2x is negative on the interval (α, 0) and so the function is concave on this interval (corollary of exercise
**

5.14). Therefore we have

x

k+1

= f(x

k

) deﬁnition of x

k+1

= f(λα + (1 −λ)0) deﬁnition of x

k

> λf(α) + (1 −λ)f(0) f is concave on (α, 0)

= λα + (1 −λ)(1/3) α is a ﬁxed point, f(0) = 1/3

= x

k

+ (1 −λ)(1/3) deﬁnition of x

k

> x

k

We see that x

k

< x

k+1

; from our choice of x

k

we have α < x

k

; from the monotonic nature of f we have

x

k

< β → x

k+1

= f(x

k

) < f(β) = β. Combining these inequalities yields

α < x

k

< x

k+1

< β

Therefore our initial choice of x

k

gives us an increasing sequence ¦x

n

¦ with an upper bound of β. Therefore it

converges to a some ﬁxed point in the interval (x

1

, beta], and this point must be β.

Case 2: x ∈ (β, γ)

Let x

k

∈ (β, γ) be chosen. We can express this as x

k

= λβ +(1 −λ)γ for some λ ∈ (0, 1). The second derivative

of f

**(x) = 2x is positive on the interval (β, γ) and so the function is convex on this interval (exercise 5.14). There-
**

93

fore we have x

k+1

= f(x

k

) deﬁnition of x

k+1

= f(λβ + (1 −λ)γ) deﬁnition of x

k

< λf(β) + (1 −λ)f(γ) f is convex on (β, γ)

= λβ + (1 −λ)γ β and γ are ﬁxed points

= x

k

deﬁnition of x

k

We see that x

k+1

< x

k

; from our choice of x

k

we have x

k

< γ; from the monotonic nature of f we have

β < x

k

→ β = f(β) < f(x

k

) = x

k+1

. Combining these inequalities yields

β < x

k+1

< x

k

< γ

Therefore our initial choice of x

k

gives us an decreasing sequence ¦x

n

¦ with an lower bound of β. Therefore it

converges to a some ﬁxed point in the interval [β, x

1

), and this point must be β.

Case 3: x ∈ (0, β)

Let x

k

∈ (0, β) be chosen. We can express this as x

k

= λ0 +(1 −λ)β for some λ ∈ (0, 1). The second derivative

of f

**(x) = 2x is positive on the interval (0, β) and so the function is convex on this interval (exercise 5.14).
**

Therefore we have

x

k+1

= f(x

k

) deﬁnition of x

k+1

= f(λ(0) + (1 −λ)β) deﬁnition of x

k

< λf(0) + (1 −λ)f(β) deﬁnition of convexity

= λ(1/3) + (1 −λ)β β is a ﬁxed point, f(0) = 1/3

= λ(1/3) +x deﬁnition of x

k

> x

k

We see that x

k

< x

k+1

; from our choice of x

k

we have α < x

k

; from the monotonic nature of f we have

x

k

< β → x

k+1

= f(x

k

) < f(β) = β. Combining these inequalities yields

α < x

k

< x

k+1

< β

Therefore our initial choice of x

k

gives us an increasing sequence ¦x

n

¦ with an upper bound of β. Therefore it

converges to a some ﬁxed point in the interval (x

1

, beta], and this point must be β.

Case 4: x

k

= β or x

k

= 0

If x

k

= β then every term of ¦x

n

¦ is β, so this sequence clearly converges to β. If x

k

= 0 then x

k+1

= f(0) = (1/3)

and the remainder of the sequence ¦x

n

¦ converges to β by one of the previous cases.

Exercise 5.23c

If x > γ then we can express x as x = γ +δ for some δ > 0. This gives us

f(x) = f(γ +δ) deﬁnition of x

=

(γ+δ)

3

−1

3

deﬁnition of f

=

γ

3

+3γ

2

δ+3γδ

2

+δ

3

+1

3

algebra

=

γ

3

+1

3

+

3γ

2

δ+3γδ

2

3

+

δ

3

3

algebra

= f(γ) +γδ(γ +δ) +

1

3

δ

3

algebra

= f(γ) −γδx +

1

3

δ

3

deﬁnition of x

= γ +γδx +

1

3

δ

3

γ is a ﬁxed point

We know that γ > 1 and x > 1, so γxδ > δ. From this we have an inequality :

> γ +δ +

1

3

δ

3

> γ +δ

= x deﬁnition of x

94

This establishes that x > γ → f(x) > x. Therefore if x

1

> γ then the sequence ¦x

n

¦ will be a nonincreasing

sequence. We know that this sequence doesn’t converge because we’re told that f has no ﬁxed points greater

than γ.

Exercise 5.24

The function f(x) has a derivative of zero at its ﬁxed point, so when x

k

and x

k+1

are both close to

√

a the mean

value theorem guarantees us that f(x

k

) and f(x

k+1

) will be very near one another:

[f(x

k

) −f(x

k+1

)[ = [(x

k

−x

k+1

)f

(ξ)[ ≈ 0

The lefthand term of this equation converges very rapidly to 0 because both [x

k

−x

k+1

[ and f

(x) are converging

toward zero. The function g(x) does not have a derivative of zero at its ﬁxed point and therefore does not have

this property (although, as we saw in exercise 3.17, it still converges albeit more slowly).

Exercise 5.25a

Each x

n+1

is chosen to be the point where the line tangent tangent to f(x

n

) crosses the x-axis.

Exercise 5.25b

Lemma 1: x

n+1

< x

n

if x

n

> ξ

We’re told that f(b) > 0 and that f(x) = 0 only at x = ξ. We know that f is continuous since it is diﬀerentiable

(theorem 5.2), so by the intermediate value theorem (theorem 4.23) we know that f(x) > 0 when x > ξ (otherwise

we’d have f(x) = 0 for some second x ∈ (ξ, b). We’re also told that f

**(x) > δ > 0 for all x ∈ (a, b) (without the
**

δ it might be the case that f

(x) ,= 0 but limf

**(x) = 0). Therefore the ratio f(x
**

n

)/f

(x

n

) is deﬁned at each x

and is positive when x > ξ. Therefore we have

x

n

−

f(x

n

)

f

(x

n

)

< x

n

, x

n

> ξ

This of course means that x

n+1

< x

n

when x

n

> ξ.

Lemma 2: x

n+1

> ξ if x

n

> ξ

We’re told that f

(x) ≥ 0, which means that f

**(x) is monotonically increasing, which means that c < x
**

n

implies

f

(c) ≤ f

(x

n

), which means that

f

(x

n

) −f(ξ)

x

n

−ξ

≤ f

(x

n

)

because the LHS of this inequality is equal to f

(c) for some c < x

n

. Using the fact that f(ξ) = 0 and a bit of

algebraic manipulation, this equation is equivalent to

x

n

−

f(x

n

)

f

(x

n

)

≥ ξ

which of course means that x

n+1

≥ ξ.

Lemma 3: if ¦x

n

¦ → κ then f(κ) = 0

Suppose that lim

n→∞

x

n+1

= κ. Then by the Cauchy criterion we have lim

n→∞

x

n

− x

n+1

= 0 which is

equivalent to

lim

n→∞

x

n

−

_

x

n

−

f(x

n

)

f

(x

n

)

_

= 0

or

lim

n→∞

f(x

n

)

f

(x

n

)

= 0

For this to hold it must either be the case that f(x

n

) → 0 or f

(x

n

) → ±∞: but we know that f

(x

n

)

is bounded from the fact that f

(x

n

) is bounded (mean value theorem : if f

(x

n

) were unbounded, then

95

(f

(x

n

) −f

(x

1

))/x

n

−x

1

= f

**(c) would be unbounded). Therefore it must be the case that f(x
**

n

) → 0. So we

have lim

n→∞

x

n

= κ and lim

n→∞

f(x

n

) = 0. Therefore by theorem 4.2 we have

f(κ) = lim

x→κ

f(x) = 0

The actual proof

Choose x

1

> ξ. By induction, using lemma 2, we know that every element of ¦x

n

¦ will be > ξ. Therefore,

by lemma 1, x

n+1

< x

n

for all n. This means that ¦x

n

¦ is a decreasing sequence with a lower bound of ξ.

Therefore, by theorem 3.14, the sequence converges to some point κ. By lemma 3 we have f(κ) = 0. But we’re

told that f has only one zero, so it must be the case that κ = ξ. This means that ¦x

n

¦ converges to ξ.

Exercise 5.25c

Using Taylor’s theorem to expand f(ξ) around f(x

n

), we have

f(ξ) = f(x

n

) +f

(x

n

)(ξ −x

n

) +

f

(t

n

)(ξ −x

n

)

2

2

Subtracting f(x

n

) from both sides then dividing by f

(x

n

) (we’re told f

(x) ≥ δ > 0) gives us

f(ξ) −f(x

n

)

f

(x

n

)

= (ξ −x

n

) +

f

(t

n

)(ξ −x

n

)

2

2f

(x

n

)

Rearranging some terms and recognizing that f(ξ) = 0, we have

x

n

−

f(x

n

)

f

(x

n

)

−ξ =

f

(t

n

)(x

n

−ξ)

2

2f

(x

n

)

Which, by the deﬁnition of x

n+1

, is equivalent to

x

n+1

−ξ =

f

(t

n

)(x

n

−ξ)

2

2f

(x

n

)

Exercise 5.25d

We’re told that f

(x) ≥ δ and f

**(x) ≤ M, so the inequality in part (c) guarantees us that
**

x

n+1

−ξ ≤

M(x

n

−ξ)

2

2δ

= A(x

n

−ξ)

2

This allows us to recursively construct a chain of inequalities.

x

n+1

−ξ ≤ A

1

(x

n

−ξ)

2

≤ A

1

A

2

(x

n−1

−ξ)

4

≤ A

1

A

2

A

4

(xn −2 −ξ)

8

≤ ≤ A

1

A

2

. . . A

2

n−1

(x

1

−ξ)

2

n

Collapsing the exponents of the rightmost term, we have

x

n+1

−ξ ≤ A

2

n

−1

(x

1

−ξ)

2

n

=

1

A

[A(x

1

−ξ)]

2

n

Exercise 5.25e

How does g behave near ξ?

We’re told f and f

**are diﬀerentiable, and clearly x is diﬀerentiable, therefore g is diﬀerentiable (theorem 5.3).
**

Using the quotient rule (theorem 5.3 again), the derivative of g is given by

g

(x) = 1 −

f

(x)

2

−f(x)f

(x)

f

(x)

2

=

f(x)f

(x)

f

(x)

96

Taking the absolute value of each side, we have

[g

(x)[ =

¸

¸

¸

¸

f(x)f

(x)

f

(x)

¸

¸

¸

¸

Using the inequality we established in part (d) we have

[g

(x)[ ≤ A[f(x)[

When we take the limit of each side of this equation as x → ξ, the RHS reduces to f(ξ) = 0 because f is

continuous.

lim

x→ξ

[g

(x)[ ≤ lim

x→ξ

A[f(x)[ = 0

Which immediately implies

lim

x→ξ

g

(x) = 0

and this describes the behavior of g

**as x → ξ, which is what we were asked to describe.
**

Show that Newton’s method involves ﬁnding a ﬁxed point of g

This is a slightly modiﬁed version of lemma 3 for part (b). Suppose that we have found a ﬁxed point κ such

that g(κ) = κ. From the deﬁnition of g this would mean that

g(κ) −κ = −

f(κ)

f

(κ)

= 0

For this to hold it must either be the case that f(κ) = 0 or f

(κ) = ±∞: but we know that f

is bounded from

the fact that f

is bounded (mean value theorem : if f

**were unbounded, then
**

f

(x)−f

(y)

x−y

= f

(c) would be

unbounded for some x, y). Therefore it must be the case that f(κ) = 0 and therefore κ must be the unique

point for which f(κ) = 0.

Exercise 5.25f

We’ll consider the more general case in which f(x) = x

m

. In this case the single real zero occurs at x = 0.

Newton’s formula gives us the step function

x

n+1

= x

n

−

f(x

n

)

f

(x

n

)

= x

n

−

x

m

n

mx

m−1

n

= x

n

−

x

n

m

=

_

m−1

m

_

x

n

This gives us a recursive deﬁnition of x

n

.

x

n+1

=

_

m−1

m

_

x

n

=

_

m−1

m

_

2

x

n−1

=

_

m−1

m

_

3

x

n−2

= =

_

m−1

m

_

n

x

1

Taking the limit of each side as n → ∞, we have

lim

n→∞

x

n+1

= lim

n→∞

_

m−1

m

_

n

x

1

(45)

In order to have ¦x

n

¦ converge to 0 we must either have x

1

= 0 or [m−1/m[ < 1. The latter case occurs when

[m−1[ < [m[

This inequality holds iﬀ m >

1

2

. If m is larger than

1

2

the limit in equation (45) exists and is equal to zero and

so ¦x

n

¦ → 0; if m is smaller than

1

2

then the limit in equation (45) is unbounded and so ¦x

n

¦ → ±∞; if m =

1

2

then equation (45) becomes

lim

n→∞

x

n+1

= lim

n→∞

(−1)

n

x

1

This limit clearly doesn’t exist when x

1

,= 0.

In the speciﬁc case given in the exercise we have m =

1

3

and therefore the sequence ¦x

n

¦ fails to converge.

97

Exercise 5.26

Following the hint given in the exercise, let M

0

= sup [f(x)[ and let M

1

= sup [f

(x)[ for x ∈ [a, x

0

]. We know

that

[f(x)[ ≤ M

1

(x

0

−a) (46)

because otherwise we’d have

¸

¸

¸

¸

f(x)

x

0

−a

¸

¸

¸

¸

=

¸

¸

¸

¸

f(x) −f(a)

x

0

−a

¸

¸

¸

¸

= f

(c) for some c ∈ (a, b) > M

1

= sup [f

(x)[

which is clearly contradictory. Additionally, we know that

M

1

= sup [f

(x)[ ≤ sup [Af(x)[ = Asup [f(x)[ = AM

0

(47)

Combining (46) and (47) gives us

[f(x)[ ≤ M

1

(x

0

−a) ≤ AM

0

(x

0

−a) , x ∈ [a, x

0

] (48)

From the fact that f(a) = 0 we know that M

1

(x

0

− a) is the maximum possible value that f could possibly

obtain at f(x

0

) (otherwise, if f(x

0

) > M

1

(x

0

−a), then the mean value theorem would give us some f

(c) > M

1

).

In comparison, M

0

is the maximum value that f actually does obtain at some x ∈ [a, b]. Therefore we have

M

0

≤ M

1

(x

0

−a) (49)

Suppose 0 < A(x

0

−a) < δ < 1. Then (48) would give us

[f(x)[ ≤ M

1

(x

0

−a) ≤ AM

0

(x

0

−a) ≤ δM

0

which contradicts (49) unless M

1

= M

0

= 0. And since we can force 0 < A(x

0

− a) < 1 by choosing an

appropriate x

0

, it must be the case that M

1

= M

0

= 0. This turns (48) into

[f(x)[ ≤ 0 , x ∈ [a, x

0

]

which shows that f(x) = 0 on the interval [a, x

0

]. We can now repeat these steps using the interval [x

0

, b]. We

need only perform this a ﬁnite number of times (speciﬁcally, a maximum of of [(b −a)/(x

0

−a)] +1 times) before

we have covered the entire interval. Therefore f(x) = 0 on the entire interval [a, b].

Exercise 5.27

The given hint is pretty much the entire solution. Let y

1

and y

2

be two solutions to the initial-value problem

and deﬁne the function f(x) = y

2

(x)−y

1

(x). The function f meets all of the prerequisites spelled out in exercise

5.26: it’s diﬀerentiable, f(a) = 0, and we’re told that there is a real number A such that

[f

(x)[ = [y

2

−y

1

[ = [φ(x, y

2

) −φ(x, y

1

)[ ≤ A[y

2

−y

1

[ = A[f(x)[

Therefore, by exericse 5.26, we have y

2

(x) − y

1

(x) = f(x) = 0 for all x; therefore y

1

= y

2

. But y

1

and y

2

were

arbitrary solutions for the intial-value problem, so we have proven that there is only one unique solution.

Exercise 5.28

Let y

a

and y

b

be two solution vectors to the initial-value problem and deﬁne the vector-valued function f(x) =

y

b

(x) − y

a

(x). As in the previous problem, we see that f(x) is diﬀerentiable and that f(a) = a. Therefore, if

there is a real number A such that

[f

(x)[ = [y

b

−y

a

[ = [φ(x, y

b

) −φ(x, y

a

)[ ≤ A[y

2

−y

1

[ = A[f(x)[

then, by exercises 5.26 and 5.27 the initial-value problem has a unique solution.

98

Exercise 5.29

Note: there’s probably a more subtle answer here, probably involving Taylor’s theorem.

Let v and v be two solution vectors to the initial value problem and deﬁne the vector valued function

f(x) = w(x) −v(x):

f(x) = [w

1

(x) −v

1

(x), w

2

(x) −v

2

(x), . . . , w

k−1

(x) −v

k−1

(x), w

k

(x) −v

k

(x)] (50)

The derivative of this is given by

f

(x) =

_

w

1

(x) −v

1

(x), w

2

(x) −v

2

(x), . . . , w

k−1

(x) −v

k−1

(x), w

k

(x) −v

k

(x)

¸

which, by the equivalences given in the exercise, becomes

f

(x) =

_

_

w

2

(x) −v

2

(x), w

3

(x) −v

3

(x), . . . , w

k

(x) −v

k

(x),

k

j=1

g

j

(w

j

−v

j

)

_

_

(51)

From exercises 5.26-5.28 we know that we want an inequality of the form [f

(x)[ ≤ A[f

(x)[. Using (50) and

(51) we see that this inequality will hold if there exists some A such that

[w

1

(x) −v

1

(x), w

2

(x) −v

2

(x), . . . , w

k−1

(x) −v

k−1

(x), w

k

(x) −v

k

(x)[

≤ A

¸

¸

¸

¸

¸

¸

w

2

(x) −v

2

(x), w

3

(x) −v

3

(x), . . . , w

k

(x) −v

k

(x),

k

j=1

g

j

(w

j

−v

j

)

¸

¸

¸

¸

¸

¸

The order of the components don’t matter when we’re taking the norm, and the rightmost term is a dot product,

so this is equivalent to

[w

2

(x) −v

2

(x), w

3

(x) −v

3

(x), . . . , w

k

(x) −v

k

(x), w

1

(x) −v

1

(x)[ ≤ A[w

2

(x) −v

2

(x), w

3

(x) −v

3

(x), . . . , w

k

(x) −v

k

(x), g(x) (w(x) −v(x))[

Exercise 6.1 (Proof 1)

Outline of the proof

We’ll see that L(P, f) = 0 for every partition; therefore sup L(P, f) = 0. By choosing an appropriate partition

we can make U(P, f) arbitrarily small; therefore inf U(P, f) = 0. We conclude that

_

f = 0.

sup L(P,f ) = 0

Although it’s easy to construct a partition such that L(P, f) = 0, we have to show that 0 is the supremum of the

set of all L(P, f). To do this, let P be an arbitrary partition of [a, b] and let m

i

be an arbitrary interval of this

arbitrary partition. If m

i

contains any point other than x

0

then inf f(x) = 0 on this interval, so m

i

∆α

i

= 0. If

m

i

contains only the point x

0

(that is, if the interval is [x

0

, x

0

]) then ∆α

i

= [α(x

0

) −α(x

0

)] = 0 and therefore

m

i

∆α

i

= 0. Therefore m

i

∆α

i

= 0 for all i, which means L(P, f) = 0. But P was an arbitrary partition, so

L(P, f) = 0 for all P. Therefore

0 = sup¦L(P, f) : P is a partition of [a, b]¦ =

_

b

a

f dx

inf U(P,f )=0

We need to show that for any > 0 we can ﬁnd some partition P such that 0 ≤ U(P, f) < . So let > 0 be given.

We’re told that α is continuous at x

0

, so by the deﬁnition of continuity we can ﬁnd some δ > 0 such that

d(x, x

0

) < δ → d(α(x), α(x

0

)) <

2

(52)

99

Let 0 < µ <

δ

2

and let P be the partition ¦a, x

0

−µ, x

0

+µ, b¦. We now calculate each M

i

:

M

1

∆α

1

= M

1

[α(x

0

−µ) −α(a)] = 0 [α(x

0

−µ) −α(a)] = 0

M

2

∆α

2

= M

2

[α(x

0

+µ) −α(x

0

−µ)] = 1 [α(x

0

+µ) −α(x

0

−µ)]

M

3

∆α

3

= M

3

[α(b) −α(x

0

+µ)] = 0 [α(b) −α(x

0

−µ)] = 0

To determine a bound for M

2

∆α

2

we apply (52) which allows us to conclude that α(x

0

+µ) −α(x

0

−µ) <

from the fact that d(x

0

−µ, x

0

+µ) = 2µ < δ.

M

2

∆α

2

= α(x

0

+µ) −α(x

0

−µ) ≤

From this we have

U(P, f)

i=1

3

M

i

∆α

i

= 0 + + 0 =

But this epsilon is arbitrarily small, so

0 = inf¦U(P, f) : P is a partition of [a, b]¦ =

_

b

a

f dx

Exercise 6.1 (Proof 2)

Thanks to Helen Barclay (hbarcla2@mail.usf.edu) for this proof. The function f is discontinuous at only one

point and α is continous at that point, so f ∈ R by theorem 6.10. Every partition of [a, b] will contain a point

other than x

0

so m

i

= 0 for every interval; therefore the inﬁmum of L(P, f) is zero; therefore

_

b

a

f dα =

_

b

a

f dα = 0

Exercise 6.2

Outline of the proof

If we assume that f(x) ,= 0 for some x ∈ [a, b], then we can construct a partition P for which L(P, f) > 0. From

this we conclude that sup L(P, f) > 0 and therefore

_

f ,= 0.

The proof

Suppose, for purposes of contradiction, that f(x

0

) = κ for some x

0

∈ [a, b] where κ is some arbitrary nonzero

number. We’re told that f is continuous on a closed set, therefore f is uniformly continuous (theorem 4.19).

Therefore there exists some δ > 0 such that [x

0

−x[ < δ → [f(x

0

) −f(x)[ <

κ

2

. Now let 0 < µ < δ and let P be

the four-element partition ¦a, x

0

−µ, x

0

+µ, b¦. For clarity, let X = (x

0

−µ, x

0

+µ).

Since [x

0

− x[ ≤ µ < δ for all x ∈ X, we have [f(x

0

) − f(x)[ = [κ − f(x)[ <

κ

2

for all x ∈ X. For this

inequality to hold for all f(x) it must be the case that min f(X) >

κ

2

. This means that

L(P, f) =

m

i

∆x

i

≥ m

2

∆x

2

= min f(X)[(x

0

+µ) −(x

0

−µ)] = min f(X)(2µ) >

κ

2

(2µ) > 0

Since L(P, f) > 0 for this particular P, it must be the case that sup L(P, f) > 0 and therefore

_

f dx ,= 0.

We assumed that f(x) ,= 0 for some x ∈ [a, b] and concluded that

_

f dx ,= 0: by contrapositive, if

_

f dx = 0

then f(x) = 0 for all x ∈ [a, b].

Exercise 6.3

Outline of the proof

We’ll see that U(P, f) − L(P, f) is equivalent to M

i

− m

i

on a single arbitrarily small interval around 0. To

make this diﬀerence arbitrarily small we will need to make sup f(x) −inf f(x) arbitrarily small by restricting x

to a suﬃciently small neighborhood of 0: this is possible iﬀ f is continuous at 0.

100

Lemma 1: For each of the β functions, the upper and lower Riemann sums depend entirely on

the intervals containing 0

Let P be an arbitrary partition of [−1, 1]. By theorem 6.4 we can assume, without loss of generality, that 0 is an

element of this partition. We deﬁne α to be the partition element immediately preceding 0 and deﬁne ω to be

the partition element immediately following 0 (so our partition has the form P = ¦x

0

< x

1

< . . . < α < 0 < ω <

. . . < x

n

< x

n+1

¦). Note that α and ω must both exist since P is a ﬁnite set, although it’s possible that α = −1

or ω = 1. For every x

i−1

< x

i

< 0 we have β(x

i−1

) = β(x

i

) = 0, therefore ∆β = 0, therefore M

i

∆β = m

i

∆β = 0.

Similarly, for every 0 < x

i−1

< x

i

we have β(x

i−1

) = β(x

i

) = 1, therefore ∆β = 0, therefore M

i

∆β = m

i

∆β = 0.

This shows that M

i

∆β = m

i

∆β = 0 for every interval that doesn’t contain 0. Therefore the values of U(P, f)

and L(P, f) depend entirely on the intervals [α, 0] and [0, ω]. This holds for arbitrary P.

The general form of M

i

∆β and m

i

∆β on [α, 0] and [0, ω]

Let P be an arbitrary partition. Let M

α

denote the supremum of f(x) on the interval [α, 0]; similarly deﬁne

m

α

, M

ω

, m

ω

, etc. For all three of the β functions we have

M

α

∆β

α

= M

α

[β(0) −β(α)] = M

α

β(0) (53)

M

ω

∆β

ω

= M

ω

[β(ω) −β(0)] = M

ω

[1 −β(0)] (54)

m

α

∆β

α

= m

α

[β(0) −β(α)] = m

α

β(0) (55)

m

ω

∆β

ω

= m

ω

[β(ω) −β(0)] = m

ω

[1 −β(0)] (56)

(57)

Case 1: β(0) = 0

When β(0) = 0 we have

M

i

∆β = M

α

β(0) +M

ω

[1 −β(0)] = M

ω

m

i

∆β = m

α

β(0) +m

ω

[1 −β(0)] = m

ω

Proof that continuity of f implies integrability: let > 0 be given. We’re told that f(0+) = f(0), so by deﬁnition

of continuity at this point there exists some δ > 0 such that

d(0, x) < δ → d(f(0), f(x)) <

2

Construct a partition P = ¦−1, 0, δ, 1¦. The interval [0, δ] is compact, therefore by theorem 4.16 the points

f

−1

(M

ω

) and f

−1

(m

ω

) both exist in [0, δ] and therefore:

d(f

−1

(M

ω

), 0) < δ → d(M

ω

, f(0)) <

2

d(f

−1

(m

ω

), 0) < δ → d(m

ω

, f(0)) <

2

and therefore, by the triangle inequality,

d(M

ω

, m

ω

) ≤ d(M

ω

, f(0)) +d(f(0), m

ω

) <

and therefore

U(P, f) −L(P, f) = M

ω

−m

ω

<

And since was arbitrary, this is suﬃcient to show that

_

f dx exists.

Proof by contrapositive that integrability of f implies right-hand continuity: Assume that f(0+) ,= f(0).

From the negation of the deﬁnition of continuity we therefore know that there exists some > 0 such that for

all δ we can ﬁnd some 0 < x < δ such that [f(x) − f(0)[ ≥ . So for any partition P let a be a point at which

f(a) = M

ω

and let b be a point for which [f(b) − f(0)[ ≥ . This gives us f(b) ≤ f(a) = M

ω

and m

ω

≤ f(0),

therefore

U(P, f) −L(P, f) = M

ω

−m

ω

= f(a) −m

ω

≥ f(a) −f(0) ≥ f(b) −f(0) ≥

101

The partition P was arbitrary so this inequality must be true for all possible partitions, so we can never ﬁnd P

such that

inf U(P, f) −sup L(P, f) ≤

and therefore

_

f dx does not exist.

Proof that this integral, if it exists, is equal to f(0): from the fact that 0 is an element of [0, δ] we know that

M

ω

≥ f(0). We also know that L(P, f) ≤

_

f (by theorem 6.4 and/or the deﬁnition of

_

f). This gives us the

inequalities

M

ω

= U(P, f) ≥ f(0)

L(P, f) ≤

_

f dβ

from which we have

U(P, f) −L(P, f) ≥ f(0) −

_

f dβ (58)

But we also know that m

ω

≤ f(0) and that U(P, f) ≥

_

f. This gives us

U(P, f) ≥

_

f dβ

m

ω

= L(P, f) ≤ f(0)

from which we have

U(P, f) −L(P, f) ≥

_

f dβ −f(0) (59)

Combining (58) and (59) gives us

¸

¸

¸

¸

_

f dβ −f(0)

¸

¸

¸

¸

≤ U(P, f) −L(P, f) =

If this is to be true for all > 0 we must have

_

f dβ = f(0).

Case 2: β(0) = 1

When β(0) = 1 we have

M

i

∆β = M

α

β(0) +M

ω

[1 −β(0)] = M

α

m

i

∆β = m

α

β(0) +m

ω

[1 −β(0)] = m

α

From here the proof that

_

f dβ = f(0) iﬀ f(0) = f(0−) is almost identical to the previous proof with α in the

place of ω.

Case 3: β(0) =

1

2

When β(0) = 1/2 we have

M

i

∆β = M

α

β(0) +M

ω

[1 −β(0)] =

1

2

[M

α

+M

ω

]

m

i

∆β = m

α

β(0) +m

ω

[1 −β(0)] =

1

2

[m

α

+m

ω

]

Therefore

U(P, f) −L(P, f) =

1

2

[M

α

−m

α

] +

1

2

[M

ω

−m

ω

]

The function f is integrable iﬀ this term can be made arbitrarily small; we saw in case (1) that [M

ω

−m

ω

] can

be made arbitrarily small iﬀ f(0) = f(0+), and we saw in case (2) that [M

α

−m

α

] can be made arbitrarily small

iﬀ f(0) = f(0−). Therefore we can minimize U(P, f) − L(P, f) in this case iﬀ f(0+) = f(0−) = f(0); that is,

iﬀ f is continuous at 0.

102

Proof that this integral, if it exists, is equal to f(0): from the fact that 0 is an element of [0, δ] we know that

M

ω

≥ f(0) and M

α

≥ f(0). We also know that L(P, f) ≤

_

f (by theorem 6.4 and/or the deﬁnition of

_

f).

This gives us the inequalities

1

2

[M

α

+M

ω

] = U(P, f) ≥

1

2

[f(0) +f(0)] = f(0)

L(P, f) ≤

_

f dβ

from which we have

U(P, f) −L(P, f) ≥ f(0) −

_

f dβ (60)

But we also know that m

ω

≤ f(0) and that m

α

≤ f(0) and that U(P, f) ≥

_

f. This gives us

U(P, f) ≥

_

f dβ

1

2

[m

α

+m

ω

] = L(P, f) ≤

1

2

[f(0) +f(0)] = f(0)

from which we have

U(P, f) −L(P, f) ≥

_

f dβ −f(0) (61)

Combining (58) and (59) gives us

¸

¸

¸

¸

_

f dβ −f(0)

¸

¸

¸

¸

≤ U(P, f) −L(P, f) =

If this is to be true for all > 0 we must have

_

f dβ = f(0).

Part (d) of the exercise

If f is continuous at 0 then we have f(0) = f(0+) = f(0−), so by case (1) we have

_

f dβ

1

= f(0) and by case

(2) we have

_

f dβ

2

= f(0) and by case (3) we have f dβ

3

= f(0).

Exercise 6.4

Let P be an arbitrary partition of [a, b]. Let n = [P[ − 1, so that n represents the number of intervals in the

partition P. Every interval will contain at least one rational number (theorem 1.20b) and at least one irrational

number (by pigeonhole principle, since [(x

i

, x

i+1

)[ = [R[ > [Q[). Therefore we have M

i

= 1, m

i

= 0 for all i.

This gives us

M

i

∆x

i

=

n

1

1 (x

i+1

−x

i

) = x

n+1

−x

0

= b −a

m

i

∆x

i

=

n

1

0 (x

i+1

−x

i

) = 0

But P was an arbitrary partition, so this holds for all partitions, and therefore

sup L(P, f) = 0, inf U(P, f) = 1

and therefore f ,∈ R.

Exercise 6.5

Is f ∈ R if f

2

∈ R?

No. Consider the function

f(x) =

_

-1, x ∈ Q

1, x ,∈ Q

We saw that f ,∈ R in exercise 6.4, but clearly f

2

(x) = 1 ∈ R.

103

Is f ∈ R if f

2

∈ R?

Yes. The function φ(x) =

3

√

x is a continuous function, so we let h(x) = φ(f

3

(x)) = f(x) and appeal to theorem

6.11 to claim that

f

3

∈ R → h ∈ R → f ∈ R

Note that the claim that h(x) = φ(f

3

(x)) = f(x) relied on the fact that x → x

3

is a one-to-one mapping, so

that

3

√

x

3

= x. In contrast, the mapping x → x

2

is not one-to-one as can be seen by the fact that

_

(−1)

2

,= −1.

Exercise 6.6

Outline of the proof

We can deﬁne a partition for which f is discontinuous on 2

n

intervals each of which has length 3

−n

, so the

Riemann sum across these intervals is proportional to (2/3)

n

. This sum approaches 0 as n → ∞.

The proof

Let E

i

be deﬁned as in sec. 2.44. Note that E

0

consists of one interval of length 1; E

1

consists of two inter-

vals of length

1

3

; E

2

consists of 4 intervals of length

1

9

; and, in general, E

n

will consist of 2

n

intervals of length 3

−n

.

With each E

n

we can associate a set F

n

that contains the endpoints of the intervals contained in E

n

. That

is, we deﬁne

F

n

= ¦a

1

< b

1

< a

2

< b

2

< . . . < a

2

n < b

2

n¦

where each [a

i

, b

i

] is an interval in E

n

. Since every point of the Cantor set – and therefore every point at which

f might be discontinuous – is in an interval of the form [a

i

, b

i

] we’ll choose a partition that lets us isolate these

intervals.

Let m and M represent the lower and upper bounds of f on [0, 1]. Choose an arbitrary > 0. Choose n

large enough that

3

n−1

>

2

n+1

(M −m)

**and choose δ such that
**

0 < δ <

1

3

n+1

These choice will be justiﬁed later. Deﬁne P

n

to be

P

n

= ¦0 = a

1

< (b

1

+δ) < (a

2

−δ) < (b

2

+δ) < (a

3

−δ) < (b

3

+δ) < . . . < (a

2

n −δ) < b

2

n = 1¦

This partition contains 2

n+1

points and therefore contains 2

n+1

−1 segments. We must now show that we can

make U(P

n

, f) −L(P

n

, f) arbitrarily small.

U(P

n

, f) −L(P

n

, f) =

2

n+1

−1

i=1

(M

i

−m

i

)∆x

i

We’ll separate this sum into the intervals on which f is continuous (i = 2, 4, 6, . . .) and the intervals on which f

might contain discontinuities (i = 1, 3, 5, . . .).

=

2

n

i=1

(M

2i

−m

2i

)∆x

2i

+

2

n

i=0

(M

2i+1

−m

2i+1

)∆x

2i+1

The function f is continuous on every interval of the form [b

i

+δ, a

i+1

−δ], and these intervals are represented

by the lefthand summation. We can therefore reﬁne P

n

such that

≤

2

+

2

n

i=1

(M

2i+1

−m

2i+1

)∆x

2i+1

104

We know the exact value of the ∆x terms. For P

n

, we have d(a

i

, b

i

) = 3

−n

and therefore ∆x

2i+1

= 3

−n

+ 2δ.

Similarly, we know that M

i

≤ M and m

i

≥ m, so we have

≤

2

+

2

n

i=1

(M −m)

_

1

3

n

+ 2δ

_

From our choice of δ <

1

3

n+1

<

1

3n

this becomes

≤

2

+

2

n

i=1

(M −m)

_

1

3

n−1

_

This summation is constant with respect to i so it becomes simply

=

2

+ 2

n

(M −m)

_

1

3

n−1

_

From our choice of n we have 3

n−1

>

2

n+1

(M−m)

and therefore 3

−(n−1)

< /2

n+1

(M −m).

≤

2

+

2

n

(M −m)

2

n+1

(M −m)

=≤

2

+

2

=

Exercise 6.7a

If f ∈ R, then from theorem 6.12 we have

_

1

0

f dx =

_

c

0

f dx +

_

1

c

f dx (62)

Now consider the partition P of [0, c] deﬁned by P = ¦0, c¦. This partition has only a single interval, so we have

inf

0<x<c

f(x) c = L(P, f) ≤

_

c

0

f dx ≤ U(P, f) = sup

0<x<c

f(x) c

From (62), adding

_

1

c

f dx to each term of this inequality gives us

_

1

c

f dx + inf

0<x<c

f(x) c ≤

_

1

0

f dx ≤

_

1

c

f dx + sup

0<x<c

f(x) c

Taking the limit of each term of this inequality as c → 0, we see that f(x) c → 0 while the center term remains

constant. The resulting inequality is

lim

c→0

_

1

c

f dx ≤

_

1

0

f dx ≤ lim

c→0

_

1

c

f dx

which of course implies that

lim

c→0

_

1

c

f dx =

_

1

0

f dx

which is what we wanted to prove.

Exercise 6.7b

Outline of the proof

We construct a function f for which

_

f =

n

i=1

(−1)

i

i

: this is the alternating harmonic series, which converges

(see Rudin’s remark 3.46). We then see that

_

[f[ =

n

i=1

1

i

: this is the harmonic series, which diverges (theorem

3.28).

105

The proof

Choose any c ∈ (0, 1) and consider the following function deﬁned on [c, 1].

f(x) =

_

(−1)

n

(n + 1),

1

n+1

< x <

1

n

for some n ∈ N

0, otherwise

Claim 1 : f is integrable on any interval [c, 1]

Choose an arbitrarily small > 0. Let

1

N

be the smallest harmonic number greater than c. We want to choose

δ such that each harmonic number in [c, 1] is contained in a distinct interval of radius δ: so choose δ such that

0 < δ <

1

2

[

1

N

− c]. We must also make sure that δ <

2(N+1)(N+2)

for reasons that will become clear later. Let

H

c

represent the partition containing the elements ¦c¦ ∪

_

1

n

±δ : n ∈ N, n <

1

c

_

∩ [0, 1].

• The partition H

c

contains one interval of the form [c,

1

N

−δ].

M

1

∆x

1

= sup f(x)∆x

1

= (−1)

N

(N + 1)∆x

1

= (−1)

N

(N + 1)

_

1

N

−δ −

1

N + 1

_

= (−1)

N

(N + 1)

_

1

N(N + 1)

−δ

_

= (−1)

N

_

1

N

−(N + 1)δ

_

The function is constant over this interval, so we have

m

1

∆x

1

= inf f(x)∆x

1

= sup f(x)∆x

1

= M

1

therefore

[M

1

−m

1

[∆x

1

= 0 (63)

• The partition H

c

contains one interval of the form [1 −δ, 1] for which

M

n

∆x

n

= sup f(x)∆x

n

= 0∆x = 0

m

n

∆x

n

= inf f(x)∆x

n

= −2∆x = −2δ

therefore

[M

n

−m

n

[∆x

n

= 2δ (64)

• The partition contains N −1 intervals of the form

_

1

i

−δ,

1

i

+δ

¸

for which

M

i

∆x

i

= sup f(x)∆x

i

≤ sup [f(x)[2δ ≤ [i + 1[2δ

m

i

∆x

i

= inf f(x)∆x

i

≥ inf −[f(x)[2δ ≥ −[i + 1[2δ

therefore

[M

i

−m

i

[∆x

i

≤ [M

i

[ +[m

i

[ ≤ 4δ[i + 1[ (65)

• The partition contains N −2 intervals of the form

_

1

j+1

+δ,

1

j

−δ

_

for which

M

j

∆x

j

= sup f(x)∆x

j

= (−1)

j

(j+1)

_

1

j

−δ −

1

j + 1

−δ

_

= (−1)

j

(j+1)

_

1

j(j + 1)

−2δ

_

= (−1)

j

_

1

j

−2δ(j + 1)

_

The function is constant over this interval, so we have

m

j

∆x

j

= inf f(x)∆x

j

= sup f(x)∆x = M = (−1)

j

_

1

j

−2δ(j + 1)

_

therefore

(M

j

−m

j

)∆x

j

= 0 (66)

106

Summing equations (63)-(66), we have

[U(P, f) −L(P, f)[ = [(M

1

−m

1

)∆x

1

+ (M

n

−m

n

)∆x

n

+

N

i=2

(M

i

−m

i

)∆x

i

+

N−1

j=2

(M

j

−m

j

)∆x

j

[

≤ [(M

1

−m

1

)[∆x

1

+[(M

n

−m

n

)[∆x

n

+

N

i=2

[M

i

−m

i

[∆x

i

+

N−1

j=2

[M

j

−M

j

[∆x

j

≤ 0 + 2δ +

N

i=2

4δ[i + 1[ +

N−1

j=2

0

= 2δ + 4δ

(N + 1)(N + 2) −1

2

= 2δ(N + 1)(N + 2)

Earlier in the proof we chose δ such that δ <

2(N+1)(N+2)

, so we obtain the inequality

[U(P, f) −L(P, f)[ ≤

And was arbitrary, so this proves that f ∈ R.

Claim 2 : The value of lim

c→0

_

1

c

f is deﬁned

Adding the values for the M terms, we have

U(P, f) = M

1

+M

n

+

N

i=2

M

i

+

N−1

i=2

M

j

≤ (−1)

N

_

1

N

−(N + 1)δ

_

+ 0 +

N

i=2

[i + 1[2δ +

N−1

j=2

(−1)

j

_

1

j

−2δ(j + 1)

_

= (−1)

N

_

1

N

−(N + 1)δ

_

δ[(N + 1)(N + 2) −6] +

N−1

j=2

(−1)

j

_

1

j

−2δ(j + 1)

_

Remember that we ﬁrst chose a value for c, which forced us to use a particular value for N; our choice of δ came

afterward, so we’re still free to select δ small enough to make this last inequality become:

U(P

c

, f) ≤

(−1)

N

N

+

N−1

j=2

(−1)

j

1

j

+

We deﬁned

1

N

to be the smallest harmonic number greater than c, so

1

N

→ 0 and N → ∞ as c → 0. Therefore

taking the limit of both sides as c → 0 gives us

lim

c→0

U(P

c

, f) =

∞

j=2

(−1)

j

1

j

+ = 1 −ln(2) +

which proves that

lim

c→0

_

1

c

f dx ≥ 1 −ln(2)

The values for the m terms are almost identical to the M terms, and adding them gives us the inequality

lim

c→0

_

1

c

f dx ≤ 1 −ln(2)

Together, of course, these two inequalities prove that

lim

c→0

_

1

c

f dx = 1 −ln(2)

107

Claim 3 : The value of lim

c→0

_

1

c

[f[ is not deﬁned

If we follow the logic from claims 1 and 2, we ﬁnd that adding the values for the M terms gives us

U(P, f) = M

1

+M

n

+

N

i=2

M

i

+

N−1

i=2

M

j

≤

_

1

N

−(N + 1)δ

_

+

1

2

+

N

i=2

[i + 1[2δ +

N−1

j=2

_

1

j

−2δ(j + 1)

_

=

_

1

N

−(N + 1)δ

_

+

1

2

δ[(N + 1)(N + 2) −6] +

N−1

j=2

_

1

j

−2δ(j + 1)

_

Again following the logic of part 2, we select δ small enough to gives us

lim

c→0

_

1

c

f dx =

∞

j=2

1

j

+

But we know from chapter 3 that this series doesn’t converge, so the lefthand limit doesn’t exist and therefore

lim

c→0

_

1

c

f dx does not exist.

Exercise 6.8

Choose an arbitrary real number c > 1. Let n be the greatest integer such that n < c.Deﬁne the partition P

n

of

[0, n] to be P

n

= ¦x

0

= 1, 2, 3, 4, . . . , n = x

n−1

¦.

First, assume that

f(n) diverges. Because f(x) > 0 this means that

f(n) is unbounded (theorem 3.24).

The fact that f is monotonically decreasing allows us to derive the following chain of inequalities.

_

c

1

f(x) dx ≥

_

n

1

f(x) dx

≥ L(P

n

, f)

=

n−2

i=1

m

i

∆x

i

=

n−2

i=1

f(i + 1)

(Note: the index is going to n − 2 instead of n − 1 because of the awkward numbering of the partition: note

that the partition only goes to x

n−1

). Taking the limit as c → ∞, we have

_

∞

1

f(x) dx = lim

c→∞

_

c

1

f(x) dx ≥ lim

n→∞

n

i=2

f(i + 1)

The integral on the left-hand side of this chain of inequalities is greater than the unbounded sum on the right-

hand side, so we conclude that the integral does not converge.

Next assume that

**f(n) converges to some κ ∈ R. Choose c, n, P
**

n+1

as deﬁned above. The fact that f is

monotonically decreasing allows us to derive the following chain of inequalities.

_

c

1

f(x) dx ≤

_

n+1

1

f(x) dx

≤ U(P

n+1

, f)

=

n−1

i=1

M

i

∆x

i

=

n−1

i=1

f(i)

108

Taking the limit as c → ∞, we have

_

∞

1

f(x) dx = lim

c→∞

_

c

1

f(x) dx ≤ lim

n

→ ∞

n−1

i=1

f(i)

We’re told that f(x) > 0, so we know that

_

c

0

f(x) is a monotonically increasing function of c. The previous

inequality tells us that

_

c

1

is bounded above. Therefore by theorem 3.14 we know that lim

c→∞

_

c

1

f dx is deﬁned.

Exercise 6.9

By the deﬁnition given in exercise 6.8:

_

∞

0

f(x)g

(x) dx = lim

c→∞

_

c

0

f(x)g

(x) dx

The integral from 0 to c is ﬁnite, so we can use integration by parts:

_

∞

0

f(x)g

(x) dx = lim

c→∞

f(c)g(c) −f(0)g(0) −

_

c

0

f

(x)g(x) dx

Applying this to the given function with f(x) = (1 + x)

−1

and g(x) = sin x, we have f

(x) = −(1 + x)

−2

and

g

**(x) = cos x. Integration by parts therefore yields
**

_

∞

0

cos x

1 +x

dx = lim

c→∞

sin c

1 +c

−

sin 0

1

+

_

c

0

sin x

(1 +x)

2

dx

Since −1 ≤ sin c ≤ 1, the non-integral terms both tend to zero as c → ∞. This leaves us with

_

∞

0

cos x

1 +x

dx = lim

c→∞

_

c

0

sin x

(1 +x)

2

dx

By the deﬁnition in exercise 6.8 this is equivalent to

_

∞

0

cos x

1 +x

dx =

_

∞

0

sin x

(1 +x)

2

dx

Exercise 6.10 a: method 1

This proof was due to Helen Barclay (hbarcla2@mail.usf.edu). We rewrite uv as the exponential of a natural

log:

uv = (u

p

)

1/p

(v

q

)

1/q

= exp(ln((u

p

)

1/p

(v

q

)

1/q

)) = exp

_

1

p

ln u

p

+

1

q

ln v

q

_

The exponential function f(x) = e

x

has a strictly positive second derivative; therefore its ﬁrst derivative is

monotonically increasing (theorem 5.11); therefore f(x) is a convex function (proof in exercise 5.14, deﬁnition

of convex in exercise 4.23). By deﬁnition of convexity with λ = 1/p, 1 −λ = 1/q we have

exp

_

1

p

ln u

p

+

1

q

ln v

q

_

≤

1

p

exp (ln u

p

) +

1

q

exp (v

q

) =

u

p

p

+

v

q

q

Combining these last two inequalities, we have

uv ≤

u

p

p

+

v

q

q

which is what we were asked to prove.

109

Exercise 6.10 a: method 2

Using the variable substitutions s = u

p

, t = v

q

, a =

1

p

, (1 −a) =

1

q

we can rewrite

uv ≤

u

p

p

+

v

q

q

(67)

as

s

a

t

1−a

≤ as + (1 −a)t

Multiplying both sides by s

−a

t

a−1

gives us

1 ≤ as

1−a

t

a−1

+ (1 −a)s

−a

t

a

We can rewrite this previous inequality in two equivalent ways:

1 ≤ a

_

t

s

_

a−1

+ (1 −a)

_

t

s

_

a

(68)

1 ≤ a

_

s

t

_

1−a

+ (1 −a)

_

s

t

_

−a

(69)

If s ≤ t then t/s ≥ 1 and therefore

a

_

t

s

_

a−1

+ (1 −a)

_

t

s

_

a

≥ a(1)

a−1

+ (1 −a)(1)

a

= 1

so the inequality in (68) holds. If s ≥ t then s/t ≥ 1, and therefore

a

_

s

t

_

1−a

+ (1 −a)

_

s

t

_

−a

≥ a(1)

1−a

+ (1 −a)(1)

−a

= 1

so the inequality in (69) holds. But both of these inequalities are just (67) with a variable change, so we see

that (67) holds when s ≥ t or when t ≥ s – that is, it always holds. And this is what we were asked to prove.

Exercise 6.10 a: method 3

This proof was due to Boris Shektman (don’t email him). Deﬁne the function f(u) to be

f(u) =

u

p

p

+

v

q

q

−uv

This function has the derivative

f

(u) = u

p−1

−v

Note that f

**(u) = 0 has only one positive real solution, which occurs at u = v
**

1/(p−1)

. We can show algebraically

that

1

p−1

=

q

p

. For u < v

q/p

we have f

(u) < 0 and for u > v

q/p

we have f

**(u) > 0, so the the point u = v
**

q/p

is

the unique global minimum of f(u). Evaluating the function at this value of u gives us

f(v

q/p

) =

v

q

p

+

v

q

q

−v

q/p

v

=

_

1

p

+

1

q

_

v

q

−(v

(p+q)/p

)

We’re given that 1/p + 1/q = 1 and, from this, we also know that p +q = pq. Making these substitutions gives

us

= v

q

−v

pq/p

= v

q

−v

q

0

Therefore the unique minimum of f is f(v

q/p

) = 0, and f(u) ≥ 0 for all other u.

110

Exercise 6.10b

From part (a), we have

_

b

a

fg dα ≤

_

b

a

f

p

p

+

g

q

q

dα =

1

p

_

b

a

f

p

dα +

1

q

_

b

a

g

q

dα =

1

p

+

1

q

= 1

Exercise 6.10c

Deﬁne κ and λ to be

κ =

_

b

a

[f[

p

dα, λ =

_

b

a

[g[

q

dα

Deﬁne

ˆ

f and ˆ g to be

ˆ

f(x) =

f(x)

κ

1/p

, ˆ g(x) =

g(x)

λ

1/q

These two functions are “normalized” in the sense that

_

b

a

[

ˆ

f[

p

dα =

_

b

a

[f[

p

κ

dα =

1

κ

_

b

a

[f[

p

dα =

κ

κ

= 1

_

b

a

[ˆ g[

q

dα =

_

b

a

[g[

q

λ

dα =

1

λ

_

b

a

[g[

q

dα =

λ

λ

= 1

We know that

ˆ

f and ˆ g are Riemann integrable by theorem 6.11, and therefore by theorem 6.13 and part (b) of

this exercise we know that [

ˆ

f[ and [ˆ g[ are Riemann integrable and that

¸

¸

¸

¸

¸

_

b

a

ˆ

fˆ g dα

¸

¸

¸

¸

¸

≤

_

b

a

[

ˆ

f[[ˆ g[ dα ≤ 1

By deﬁnition of

ˆ

f and ˆ g this becomes:

¸

¸

¸

¸

¸

_

b

a

f

κ

1/p

g

λ

1/q

dα

¸

¸

¸

¸

¸

≤

_

b

a

[f[

κ

1/p

[g[

λ

1/q

dα ≤ 1

Multiplying both sides by the constant κ

1/p

λ

1/q

gives us

¸

¸

¸

¸

¸

_

b

a

fg dα

¸

¸

¸

¸

¸

≤

_

b

a

[fg[ dα ≤ κ

1/p

λ

1/q

which, by the deﬁnition of κ and λ, is equivalent to

¸

¸

¸

¸

¸

_

b

a

fg dα

¸

¸

¸

¸

¸

≤

_

b

a

[fg[ dα ≤

_

_

b

a

[f[

p

dα

_

1/p

_

_

b

a

[g[

q

dα

_

1/q

which is what we wanted to prove.

Exercise 6.10d

Proof by contrapositive. If the inequality were false for the improper integrals, we would be able to ﬁnd some

function f such that

lim

c→∞

¸

¸

¸

¸

_

c

a

fg dα

¸

¸

¸

¸

> lim

c→∞

__

c

a

[f[

p

dα

_

1/p

__

c

a

[g[

q

dα

_

1/q

or

lim

c→x0

¸

¸

¸

¸

¸

_

b

x0

fg dα

¸

¸

¸

¸

¸

> lim

c→x0

_

_

b

x0

[f[

p

dα

_

1/p

_

_

b

x0

[g[

q

dα

_

1/q

But this is a strict inequality, so in either of these cases we would have to be able to ﬁnd a neighborhood of ∞

or x

0

for which this inequality still held. And this would give us a proper integral for which Holder’s inequality

doesn’t hold. By contrapositive, the fact that Holder’s inequality is valid for proper integrals shows that it’s

true for improper integrals.

111

Exercise 6.11

If we can prove that [[f+g[[ ≤ [[f[[+[[g[[, then we can prove the given inequality by letting f = f−g and g = g−h.

[[f +g[[

2

=

_

b

a

[f +g[

2

dα deﬁnition of [[ [[

≤

_

b

a

([f[ +[g[)

2

dα The triangle inequality is established

for [ [

=

_

b

a

[f[

2

+ 2[fg[ +[g[

2

dα

=

_

b

a

[f[

2

dα + 2

_

b

a

[fg[ dα +

_

b

a

[g[

2

dα properties of integrals: theorem 6.12

≤

_

b

a

[f[

2

dα + 2

_

_

b

a

[f[

2

dα

_

1/2

_

_

b

a

[g[

2

dα

_

1/2

+

_

b

a

[g[

2

dα Holder’s inequality (exercise 6.10)

= [[f[[

2

+ 2[[f[[ [[g[[ +[[g[[

2

deﬁnition of [[ [[

= ([[f[[ +[[g[[)

2

Taking the square root of the ﬁrst and last terms of this chain of inequalities, we have

[[f +g[[ ≤ [[f[[ +[[g[[

This inequality must hold for any f, g ∈ R. Letting f = f −g and g = g −h this becomes

[[f −h[[ ≤ [[f −g[[ +[[g −h[[

which is what we were asked to prove.

Exercise 6.12

We’re told that f is continuous, so it has upper and lower bounds m ≤ f(x) ≤ M. Choose any > 0. Since

f ∈ R(α), we can ﬁnd some partition P such that

U(P, f, α) −L(P, f, α) <

2

M −m

Having now determined a particular partition P we can deﬁne g as suggested in the hint. This function

is simply a series of straight lines connecting f(x

i−1

) to f(x

i

). This becomes clearer if we rewrite g in an

algebraically equivalent form:

g(t) = f(x

i−1

) +

f(x

i

) −f(x

i−1

)

x

i

−x

i−1

(t −x

i−1

)

We see that on any interval [x

i−1

, x

i

] the function g(t) is bounded between f(x

i−1

) and f(x

i

). That is,

m

i

min¦f(x

i−1

), f(x

i

)¦ ≤ g(t), t ∈ [x

i−1

, x

i

] ≤ max¦f(x

i−1

), f(x

i

)¦ ≤ M

i

And of course we have similar bounds on f:

m

i

≤ f(t) ≤ M

i

and therefore, since f(t), g(t) ≤ M

i

and −f(t), −g(t) ≤ −m

i

, we have

[f(t) −g(t)[ ≤ [M

i

−m

i

[ = M

i

−m

i

(70)

Similarly, since M

i

≤ M and −m

i

≤ −m we have

M

i

−m

i

≤ M −m (71)

We’ve now established all of the inequalities we need to complete our proof.

112

[[f −g[[

2

=

_

b

a

[f(x) −g(x)[

2

dα deﬁnition of [[ [[

=

n

i=0

_

xi+1

xi

[f(x) −g(x)[

2

dα integral property 6.12c

≤

n

i=0

_

xi+1

xi

[M

i

−m

i

[

2

dα from (70)

=

n

i=0

[M

i

−m

i

[

2

_

xi+1

xi

dα integral property 6.12a

=

n

i=0

[M

i

−m

i

[

2

∆α(x

i

)

≤

n

i=0

(M −m)(M

i

−m

i

) ∆α(x

i

) from (71)

= (M −m)

n

i=0

(M

i

−m

i

) ∆α(x

i

) integral property 6.12a

= (M −m)[U(P, f, α) −L(P, f, α)] Deﬁnition of U and L

≤ (M −m)

2

M−m

we chose P so that this would hold

=

2

we chose so that this would hold

Taking the square root of the ﬁrst and last terms gives us

[[f −g[[ ≤

which is what we were asked to prove.

Exercise 6.13a

Letting u = t

2

we have du/dt = 2t and therefore

dt =

du

2t

=

du

2

√

u

Using the change of variables theorem (6.19) we see that the given integral is equivalent to

f(x) =

_

(x+1)

2

x

2

sin u

2

√

u

du (72)

Integration by parts (theorem 6.22) with F = 1/2

√

u and g = sin udu gives us

f(x) =

−cos u

2

√

u

¸

¸

¸

¸

(x+1)

2

x

2

−

_

(x+1)

2

x

2

cos u

4u

3/2

du

which expands to

f(x) =

−cos(x + 1)

2

2(x + 1)

+

cos x

2

2x

−

_

(x+1)

2

x

2

cos u

4u

3/2

du

By the triangle inequality this gives us

[f(x)[ =

¸

¸

¸

¸

¸

−cos(x + 1)

2

2(x + 1)

+

cos x

2

2x

−

_

(x+1)

2

x

2

cos u

4u

3/2

du

¸

¸

¸

¸

¸

≤

¸

¸

¸

¸

−cos(x + 1)

2

2(x + 1)

¸

¸

¸

¸

+

¸

¸

¸

¸

cos x

2

2x

¸

¸

¸

¸

+

¸

¸

¸

¸

¸

_

(x+1)

2

x

2

cos u

4u

3/2

du

¸

¸

¸

¸

¸

<

¸

¸

¸

¸

1

2(x + 1)

¸

¸

¸

¸

+

¸

¸

¸

¸

1

2x

¸

¸

¸

¸

+

¸

¸

¸

¸

¸

_

(x+1)

2

x

2

1

4u

3/2

du

¸

¸

¸

¸

¸

=

¸

¸

¸

¸

1

2(x + 1)

¸

¸

¸

¸

+

¸

¸

¸

¸

1

2x

¸

¸

¸

¸

+

¸

¸

¸

¸

1

2

_

1

x + 1

−

1

x

_¸

¸

¸

¸

=

1

2(x + 1)

+

1

2x

+

1

2x(x + 1)

=

x + (x + 1) + 1

2x(x + 1)

=

2(x + 1)

2x(x + 1)

=

1

x

113

Note that the strict inequality is justiﬁed by the fact that cos u ≤ 1 for all u ∈ [x

2

, (x + 1)

2

] but cos is not a

constant function so cos u < 1 for some u ∈ [x

2

, (x + 1)

2

].

Exercise 6.13b

In the previous problem we used integration by parts to determine that

f(x) =

−cos(x + 1)

2

2(x + 1)

+

cos x

2

2x

−

_

(x+1)

2

x

2

cos u

4u

3/2

Multiplying by 2x gives us

2xf(x) =

−xcos(x + 1)

2

x + 1

+ cos x

2

−2x

_

(x+1)

2

x

2

cos u

4u

3/2

du

which is algebraically equivalent to

2xf(x) = cos x

2

−cos[(x + 1)

2

] +

cos(x + 1)

2

x + 1

−2x

_

(x+1)

2

x

2

cos u

4u

3/2

du

Letting r(x) be deﬁned as

r(x) =

cos(x + 1)

2

x + 1

−2x

_

(x+1)

2

x

2

cos u

4u

3/2

du

we have, by the triangle inequality,

[r(x)[ ≤

¸

¸

¸

¸

cos(x + 1)

2

x + 1

¸

¸

¸

¸

+

¸

¸

¸

¸

¸

2x

_

(x+1)

2

x

2

cos u

4u

3/2

du

¸

¸

¸

¸

¸

≤

¸

¸

¸

¸

1

x + 1

¸

¸

¸

¸

+

¸

¸

¸

¸

¸

2x

_

(x+1)

2

x

2

1

4u

3/2

du

¸

¸

¸

¸

¸

=

¸

¸

¸

¸

1

x + 1

¸

¸

¸

¸

+

¸

¸

¸

¸

2x

1

2

_

1

x + 1

−

1

x

_¸

¸

¸

¸

=

1

x + 1

+x

1

x(x + 1)

=

1

x + 1

+

1

x + 1

=

2

x + 1

<

2

x

So we see that

2xf(x) = cos x

2

−cos[(x + 1)

2

] +r(x)

where [r(x)[ < 2/x.

Exercise 6.13c

We’re asked to ﬁnd the limsup and liminf of the function

xf(x) =

1

2

_

cos(x

2

) −cos([x + 1]

2

) +r(x)

¸

We established in part (b) that r(x) → 0 as x → ∞. It’s also clear that this function is bounded above by 1

and bounded below by −1, but it’s not immediately clear that these bounds are the limsup and the liminf.

114

Remark: The supremum and inﬁmum of cos(x

2

) −cos([x + 1]

2

) are never obtained

The supremum would be obtained if we could ﬁnd x such that

cos(x

2

) −cos([x + 1]

2

) = 2 (73)

which would occur precisely when

x

2

= 2kπ, [x + 1]

2

= (2j + 1)π, j, k ∈ N (74)

After some tedious algebra, we can see that these two equalities hold when

_

π[2(j −k) + 1] −1

2

√

2π

_

2

= k (75)

But this equation can’t hold. If it did, we could rearrange it algebraically to give us

π

2

[2(j −k) + 1]

2

−2π[2j + 3k + 1] + 1 = 0

This would allow us to use the quadratic formula to give an algebraic expression for π; but this is impossible as

π is a transcendental number. The inﬁmum of cos(x

2

) −cos([x + 1]

2

) is not obtained for the same reason.

Proof: The limsup of xf(x) is 1

Let 1 > > 0 and N ∈ N be given. Let δ be chosen so that

0 < δ <

cos

−1

(1 −)

2π

Deﬁne the function p(m) to be

p(m) =

_

π[2m+ 1] −1

2

√

2π

_

2

Its derivative with respect to m is

p

(m) =

√

2π(π[2m+ 1] −1)

which is strictly increasing. Therefore we can make the derivative as large as we want by choosing a suﬃciently

large m. Speciﬁcally, we are able to choose M ∈ N such that p

**(M)δ > 1 and M > N. By the mean value
**

theorem we have

p(M +δ) −p(M) = p

(ξ)δ, ξ ∈ (M, M +δ)

From the strictly increasing nature of p

**and our choice of M we have
**

p(M +δ) −p(M) = p

(ξ)δ > p

(M)δ > 1

Therefore p(x) must take an integer value for at least one x ∈ [M, M + δ]. Let κ represent one such x. The

two important properties of κ are that p(κ) is an integer and that κ − M < δ (so that κ itself is “almost” an

integer). We now have

p(κ) = p(M +δ) =

_

π[2(M +δ) + 1] −1

2

√

2π

_

2

= k ∈ N

If we deﬁne j to be j = k +M this becomes

_

π[2(j −k +δ) + 1] −1

2

√

2π

_

2

= k ∈ N

Reversing the algebraic steps that led from (74) to (75) tells us that there exists some x ∈ R such that

x

2

= 2kπ, [x + 1]

2

= (2(j +δ) + 1)π, j, k ∈ N

Using this value of x in our original function, we have

xf(x) =

1

2

_

cos(x

2

) −cos([x + 1]

2

) +r(x)

¸

=

1

2

[cos(2kπ) −cos(2(j + 1)π + 2δπ) +r(x)]

115

Using trig identities, this becomes

xf(x) =

1

2

[1 −¦cos(2(j + 1)π) cos(2π) −sin(2(j + 1)π) sin(2π)¦ +r(x)]

=

1

2

[1 −(−1) cos(2π) −0 +r(x)]

Finally, from our original choice of δ as an inverse cosine, this becomes

xf(x) =

1

2

[1 + (1 −) +r(x)]

= 1 +

1

2

[r(x) −]

From part (b), we know that r(x) → 0 as x → ∞:

lim

x→∞

xf(x) = 1 −

2

And was arbitrary, so the supremum is

lim

x→∞

sup xf(x) = 1

The proof that liminf xf(x) = −1 is similar.

Exercise 6.13d

_

∞

0

sin(t

2

) dt =

∞

0

_

√

(n+1)π

√

nπ

sin(t

2

) dt (76)

=

∞

0

_

√

(n+1)π

√

nπ

(−1)

n

[ sin(t

2

)[ dt

=

∞

0

(−1)

n

_

√

(n+1)π

√

nπ

[ sin(t

2

)[ dt

≤

∞

0

(−1)

n

_

√

(n+1)π

√

nπ

1 dt (and is similarly bounded below)

=

√

π

∞

0

(−1)

n

_√

n + 1 −

√

n

_

(77)

(78)

We saw in exercise 3.6.a that the sequence ¦

√

n + 1 −

√

n¦ is decreasing. Therefore, by the alternating series

theorem (3.43) the series in (77) converges. Therefore, by the comparison test (3.25) the series in (76) converges

and therefore the integral in (76) converges.

Exercise 6.14a

Letting u = e

t

we have du/dt = e

t

= u and therefore

dt =

du

e

t

=

du

u

Using the change of variables theorem (6.19) we see that the given integral is equivalent to

f(x) =

_

e

x+1

e

x

sin(u)

u

du (79)

116

We use integration by parts as we did in 6.13(a).

f(x) =

−cos u

u

¸

¸

¸

¸

e

x+1

e

x

−

_

e

x+1

e

x

cos u

u

2

du

which expands to

f(x) =

cos e

x

e

x

−

cos e

x+1

e

x+1

−

_

e

x+1

e

x

cos u

u

2

du

By the triangle inequality this gives us

[f(x)[ =

¸

¸

¸

¸

¸

cos e

x

e

x

−

cos e

x+1

e

x+1

−

_

e

x+1

e

x

cos u

u

2

du

¸

¸

¸

¸

¸

≤

¸

¸

¸

¸

cos e

x

e

x

¸

¸

¸

¸

+

¸

¸

¸

¸

cos e

x+1

e

x+1

¸

¸

¸

¸

+

¸

¸

¸

¸

¸

_

e

x+1

e

x

cos u

u

2

du

¸

¸

¸

¸

¸

<

¸

¸

¸

¸

1

e

x

¸

¸

¸

¸

+

¸

¸

¸

¸

1

e

x+1

¸

¸

¸

¸

+

¸

¸

¸

¸

¸

_

e

x+1

e

x

1

u

2

du

¸

¸

¸

¸

¸

<

¸

¸

¸

¸

1

e

x

¸

¸

¸

¸

+

¸

¸

¸

¸

1

e

x+1

¸

¸

¸

¸

+

¸

¸

¸

¸

1

e

x

−

1

e

x+1

¸

¸

¸

¸

=

1

e

x

+

1

e

x+1

+

e −1

e

x+1

=

2e

e

x+1

=

2

e

x

Therefore e

x

[f(x)[ < 2. Note that the strictness of the inequality is justiﬁed by the fact that cos u ≤ 1 for all

u ∈ [e

x

, e

x+1

] but cos is not a constant function so cos u < 1 for some u ∈ [e

x

, e

x+1

].

Exercise 6.14b

In the previous exercise we used integration by parts to determine that

f(x) =

cos e

x

e

x

−

cos e

x+1

e

x+1

−

_

e

x+1

e

x

cos u

u

2

du

Multiplying by e

x

gives us

e

x

f(x) = cos e

x

−e

−1

cos e

x+1

−e

x

_

e

x+1

e

x

cos u

u

2

du

Letting r(x) be deﬁned as

r(x) = −e

x

_

e

x+1

e

x

cos u

u

2

du

we have, by the triangle inequality,

117

[r(x)[ ≤

¸

¸

¸

¸

¸

−e

x

_

e

x+1

e

x

cos u

u

2

du

¸

¸

¸

¸

¸

=

¸

¸

¸

¸

¸

−e

x

_

e

x+1

e

x

1

u

2

du

¸

¸

¸

¸

¸

= [e

x

[

¸

¸

¸

¸

¸

_

e

x+1

e

x

1

u

2

du

¸

¸

¸

¸

¸

= [e

x

[

¸

¸

¸

¸

1

e

x

−

1

e

x+1

¸

¸

¸

¸

=

e −1

e

<

2

e

So we see that

e

x

f(x) = cos e

x

−e

−1

cos e

x+1

+r(x)

where [r(x)[ < 2e

−1

.

Exercise 6.15a

Using integration by parts with F = f

2

(x) and g = dx gives us

1 =

_

b

a

f

2

(x) dx = xf

2

(x)

¸

¸

b

a

−

_

b

a

2f(x)f

(x)xdx

which evaluates to

1 = bf

2

(b) −af

2

(a) −

_

b

a

2f(x)f

(x)xdx

which, since f(a) = f(b) = 0, becomes

1 = −

_

b

a

2f(x)f

(x)xdx

Dividing both sides by −2 gives us

−1

2

=

_

b

a

f(x)f

(x)xdx

which is the desired equality.

Exercise 6.15b

Applying Holder’s inequality to the last equation in part (a) gives us

¸

¸

¸

¸

−1

2

¸

¸

¸

¸

=

¸

¸

¸

¸

¸

_

b

a

f(x)f

(x)xdx

¸

¸

¸

¸

¸

≤

_

_

b

a

[x

2

f

2

(x)[ dx

_

1/2

_

_

b

a

[[f

(x)]

2

[ dx

_

1/2

Since f(x)f

**(x) must be negative at some point (by mean value theorem, since f(a) = f(b) = 0) while x
**

2

f

2

and

[f

(x)]

2

are strictly positive, this inequality must be strict:

¸

¸

¸

¸

−1

2

¸

¸

¸

¸

=

¸

¸

¸

¸

¸

_

b

a

f(x)f

(x)xdx

¸

¸

¸

¸

¸

<

_

_

b

a

[x

2

f

2

(x)[ dx

_

1/2

_

_

b

a

[[f

(x)]

2

[ dx

_

1/2

Squaring the ﬁrst and last term in this chain of inequalities, we have

1

4

<

_

b

a

[x

2

f

2

(x)[ dx

_

b

a

[[f

(x)]

2

[ dx

118

All of the normed values are squares, so the norms are redundant:

1

4

<

_

b

a

x

2

f

2

(x) dx

_

b

a

[f

(x)]

2

dx

Exercise 6.16a

We can integrate the given function separately over inﬁnitely many intervals of length 1:

s

_

∞

1

[x]

x

s+1

dx =

∞

n=1

s

_

n+1

n

[x]

x

s+1

dx

We have [x] = n on the interval [n, n + 1) so this becomes

=

∞

n=1

s

_

n+1

n

n

x

s+1

dx

=

∞

n=1

sn

_

−1

sx

s

¸

¸

¸

¸

n+1

x=n

=

∞

n=1

n

_

1

n

s

−

1

(n + 1)

s

_

We then split up the summation into three parts.

=

1

n=1

n

1

n

s

+

∞

n=2

n

1

n

s

−

∞

n=1

n

1

(n + 1)

s

Evaluating the ﬁrst summation and changing the index of the third gives us

= 1 +

∞

n=2

n

1

n

s

−

∞

n=2

(n −1)

1

(n)

s

= 1 +

∞

n=2

n −(n −1)

n

s

= 1 +

∞

n=2

1

n

s

And since 1 is clearly equal to 1/n

s

when n = 1, this is equivalent to

=

∞

n=1

1

n

s

Exercise 6.16b

We’re asked to evaluate the integral

s

s −1

−s

_

∞

1

x −[x]

x

s+1

dx

Having determined that [x] was integrable in part (a) we can split up the integral as follows:

=

s

s −1

−s

_

∞

1

1

x

s

dx +s

_

∞

1

[x]

x

s+1

dx

Elementary calculus allows us to calculate the left integral:

=

s

s −1

−

s

s −1

+s

_

∞

1

[x]

x

s+1

dx

= s

_

∞

1

[x]

x

s+1

dx

This, as we saw in part (a), is equivalent to

= ζ(s)

119

Exercise 6.17

I’m going to change the notation of this problem a bit to make it clearer (to me, at least) by letting f = G and

f

**= g. We’re told that α is a monotonically increasing function on [a, b] and that f is continuous. We’re asked
**

to prove that

_

b

a

α(x)f

(x) dx = f(b)α(b) −f(a)α(a) −

_

b

a

f(x) dα

Most of the work is done for us by the theorem 6.22 (integration by parts) which tells us that

_

b

a

α(x)f

(x) dx = f(b)α(b) −f(a)α(a) −

_

b

a

f(x)α

(x) dx

So our proof is reduced to simply proving that

_

b

a

f(x) dα =

_

b

a

f(x)α

(x) dx (80)

It’s tempting to appeal to elementary calculus to show that dα = α

**(x) dx, but we can prove this more formally.
**

Proving this more formally

Let > 0 be given. We’re told that f is continuous and that α is monotonically inreasing; therefore f ∈ R(α).

Let P be an arbitrary partition of [a, b] such that U(P, f, α) −L(P, f, α) <

2

.

On any interval [x

i−1

, x

i

] the mean value theorem tells us that there is some t

i

∈ [x

i−1

, x

i

] such that

α(x

i−1

) −α(x

i

) = α

(t

i

)[x

i−1

−x

i

]

or, to express the same equation with diﬀerent notation,

∆α

i

= α

(t

i

)∆x

i

(81)

By theorem 6.7 (c) we have

¸

¸

¸

¸

¸

n

i=1

f(t

i

)∆α

i

−

_

b

a

f(x) dα

¸

¸

¸

¸

¸

<

2

(82)

and also ¸

¸

¸

¸

¸

n

i=1

f(t

i

)α

(t

i

)∆x

i

−

_

b

a

f(x)α

(x) dx

¸

¸

¸

¸

¸

<

2

(83)

By the inequality (81) we have

n

i=1

f(t

i

)∆α

i

=

n

i=1

f(t

i

)α

(t

i

)∆x

i

Therefore, by (82) and (83) and the triangle inequality, we have

¸

¸

¸

¸

¸

_

b

a

f(x) dα −

_

b

a

f(x)α

(x) dx

¸

¸

¸

¸

¸

<

which, in order to be true for any > 0, requires that

_

b

a

f(x) dα =

_

b

a

f(x)α

(x) dx

This proves (80), and this completes the proof.

120

Exercise 6.18

The derivative γ

1

(t) = ie

it

is continuous, therefore γ

1

(t) is rectiﬁable (theorem 6.27) and its length is given by

_

2π

0

[ie

it

[ =

_

2π

0

1 = 2π

The derivative γ

2

(t) = 2ie

it

is continuous, therefore γ

2

(t) is rectiﬁable and its length is given by

_

2π

0

[2ie

it

[ =

_

2π

0

2 = 4π

The arc γ

3

is not rectiﬁable

Proof by contradiction. If γ

3

were rectiﬁable then its length would be given by the integral

_

2π

0

[γ

(t)[ dt =

_

2π

0

¸

¸

¸

¸

_

2πi sin

_

1

t

_

+

−1

t

2

2πit cos

_

1

t

__

e

2πit sin(1/t)

¸

¸

¸

¸

dt

=

_

2π

0

2π

¸

¸

¸

¸

sin

_

1

t

_

−

1

t

cos

_

1

t

_¸

¸

¸

¸

dt (84)

But this integral isn’t deﬁned, since Riemann integrals are deﬁned only for bounded functions and this function

is unbounded:

lim

t→0

[γ

(t) dt[ = lim

k→∞

¸

¸

¸

¸

γ

_

1

2kπ

_¸

¸

¸

¸

= lim

k→∞

2π [sin (2kπ) −2kπ cos (2kπ)[

= lim

k→∞

2π [±2kπ[

= lim

k→∞

4π

2

k = ∞

Therefore the integral in (84) doesn’t exist, which means the arc length of γ

3

is not deﬁned, which means that

γ

3

is not rectiﬁable.

Exercise 6.19

Lemma 1: If f : A → B and G : B → C are both one-to-one, then g ◦ f : A → C is one-to-one

By deﬁnition of “one-to-one”, we have

f(x) = f(y) → x = y, g(s) = g(t) → s = t

Therefore

g(f(x)) = g(f(y)) → f(x) = f(y) → x = y

and so g ◦ f is one-to-one.

Lemma 2: If g ◦ f : A → C is one-to-one and f : A → B is one-to-one and onto, then g : B → C is

one-to-one.

Proof by contrapositive. Suppose that g is not one-to-one. Then we could ﬁnd x, y ∈ B such that g(x) = g(y)

but x ,= y. But f is one-to-one and onto, so there exist unique s, t ∈ A such that f(s) = x, f(t) = y. So we have

g(f(s)) = g(f(t)) but f(s) ,= f(t) and therefore s ,= t. So g ◦ f is not one-to-one. By contrapositive, the lemma

is proven.

Lemma 3: If f : [a, b] → [c, d] is a continuous, one-to-one, real function then f is either strictly

decreasing or strictly increasing

Proof by contrapositive. Let f be a continuous real function. If f were not strictly increasing and not strictly

decreasing then we could ﬁnd some x < y < z such that either f(y) ≤ f(x) and f(y) ≤ f(z) or such that

f(y) ≥ f(x) and f(y) ≥ f(z). From the intermediate value property of continuous functions we know that we

must then be able to ﬁnd x

, y

such that

x ≤ x

< y < z

< z, f(x

) = f(z

)

so that f is not one-to-one. By contrapositive, the lemma is proven.

121

Lemma 4: If P = ¦x

i

¦ is a partition of γ

2

, then P

= ¦φ(x

i

)¦ is a partition of γ

1

From lemma 3 we know that φ is strictly increasing or decreasing, and since φ(c) = a it must be strictly

increasing. And φ is onto, so φ(d) = b. Let P be an arbitrary partition of [c, d]:

P = ¦x

i

¦ = ¦c = x

0

, x

1

, . . . , x

n

, x

n+1

= d¦

From the properties of φ we have

a = φ(c) = φ(x

0

) < φ(x

1

) < φ(x

2

) < . . . < φ(x

n

) < φ(x

n+1

) = φ(d) = b

and therefore

¦φ(x

i

)¦ = ¦a = φ(x

0

), φ(x

1

), φ(x

2

), . . . , φ(x

n

), φ(x

n+1

) = b¦

which is a partition of [a, b].

Lemma 5: If P = ¦x

i

¦ is a partition of γ

1

, then P

= φ

−1

(x

i

) is a partition of γ

2

We’re told that φ is a continuous one-to-one function from [a, b] to [c, d], so its inverse exists and is a continuous

mapping from [a, b] onto [c, d] (theorem 4.17). By lemma 3, this means that φ

−1

is either strictly increasing or

strictly decreasing, and since φ

−1

(a) = c it must be strictly increasing. And φ

−1

is onto, so φ

−1

(b) = d. Let P

be an arbitrary partion of [a, b]:

P = ¦x

i

¦ = ¦a = x

0

, x

1

, . . . , x

n

, x

n+1

= b¦

From the properties of φ

−1

we have

c = φ

−1

(a) = φ

−1

(x

0

) < φ

−1

(x

1

) < φ

−1

(x

2

) < . . . < φ

−1

(x

n

) < φ

−1

(x

n+1

) = φ

−1

(b) = d

and therefore

¦φ

−1

(x

i

)¦ = ¦c = φ

−1

(x

0

), φ

−1

(x

1

), φ

−1

(x

2

), . . . , φ

−1

(x

n

), φ

−1

(x

n+1

) = d¦

which is a partition of [c, d].

Proof 1: γ

1

is an arc iﬀ γ

2

is an arc

If γ

1

is an arc then γ

1

is a one-to-one function. We’re told that φ is one-to-one, therefore by lemma 1 γ

1

◦ φ is

one-to-one. And γ

2

= γ

1

◦ φ so γ

2

is one-to-one. There γ

2

is an arc.

If γ

2

is an arc then γ

2

= γ

1

◦ φ is one-to-one. We’re told that φ is a one-to-one and onto function, therefore

by lemma 2 we know that γ

1

is one-to-one. Therefore γ

1

is an arc.

Proof 2: γ

1

is a closed curve iﬀ γ

2

is a closed curve

Assume that γ

1

is a closed curve so that γ

1

(a) = γ

1

(b). We saw in lemma 4 that φ(c) = a and φ(d) = b, so we

have

γ

2

(c) ≡ γ

1

(φ(c)) = γ

1

(a) = γ

1

(b) = γ

1

(φ(d)) ≡ γ

2

(d)

Therefore γ

2

(c) = γ

2

(d), which means that γ

2

is a closed curve.

Now assume that γ

2

is a closed curve so that γ

2

(c) = γ

2

(d). We saw in lemma 5 that φ

−1

(a) = c and

φ

−1

(b) = d so we have

γ

1

(a) = γ

1

(φ(φ

−1

(a))) ≡ γ

2

(φ

−1

(a)) = γ

2

(c) = γ

2

(d) = γ

2

(φ

−1

(b)) ≡ γ

1

(φ(φ

−1

(b))) = γ

1

(b)

Therefore γ

1

(a) = γ

1

(b), which means that γ

1

is a closed curve.

122

Proof 3: γ

1

and γ

2

have the same length

Let P be an arbitrary partitioning of [a, b]. Using the notation from deﬁnition 6.26, we have

Λ(P, γ

2

) =

n

i=1

[(γ

2

(x

i

) −γ

2

(x

i−1

)[ < M

From the deﬁnition of γ

2

, this becomes

Λ(P, γ

2

) =

n

i=1

[(γ

1

(φ(x

i

)) −γ

1

(φ(x

i−1

))[ < M

From lemma 4 we see that ¦φ(x

i

)¦ describes a partitioning of [c, d]:

Λ(P, γ

2

) = Λ(φ(P), γ

1

)

Similarly, if we let P

**be an arbitrary partition of [a, b] we can use lemma 5 to show that
**

Λ(P

, γ

1

) = Λ(φ

−1

(P), γ

2

)

This means that the set ¦Λ(P, γ

1

)¦ is identical to the set of ¦Λ(P, γ

2

)¦ so clearly these sets have the same

supremums, which means that Λ(γ

1

) = Λ(γ

2

).

Proof 4: γ

1

is rectiﬁable iﬀ γ

2

is rectiﬁable

Having shown previously that Λ(γ

1

) = Λ(γ

2

), it’s clear that Λ(γ

1

) is ﬁnite iﬀ Λ(γ

2

) is ﬁnite and so γ

1

is rectiﬁable

iﬀ γ

2

is rectiﬁable.

Exercise 7.1

Let ¦f

n

¦ be a sequence of functions that converges uniformly to f. Let M

n

= sup [f

n

[ and let M = sup [f[. Let

> 0 be given. Because of the uniform convergence of ¦f

n

¦ → f we can ﬁnd some N such that

n > N → [f

n

(x)[ = [f

n

(x) −f(x) +f(x)[ ≤ [f

n

(x) −f(x)[ +[f(x)[ ≤ +M

This tells us that for all n > N the function f

n

is bounded above by M +. There are only ﬁnitely many n ≤ N,

each of which is bounded by some M

n

. So by choosing

max¦M +, M

1

, M

2

, . . . , M

N

¦

we have found a bound for all f

n

.

Exercise 7.2a

Let > 0 be given. We’re told that ¦f

n

¦ and ¦g

n

¦ converge uniformly on E so we can ﬁnd N, M such that

n, m > N → [f

n

(x) −f

m

(x)[ <

2

, n, m > M → [g

n

(x) −g

m

(x)[ <

2

By choosing N

∗

> max¦N, M¦ we have

n, m > N

∗

→ [(f

n

(x) +g

n

(x)) −(f

m

(x) −g

m

(x))[ ≤ [f

n

(x) −f

m

(x)[ +[g

n

(x) −g

m

(x)[ <

which shows that f

n

+g

n

converges uniformly on E.

123

Exercise 7.2b

If ¦f

n

¦ and ¦g

n

¦ are bounded functions then by exercise 7.1 they are uniformly bounded, say by F and G. Let

> 0 be given and choose δ such that

δ < min

_

,

_

3

,

3F

,

3G

_

We’re told that ¦f

n

¦ and ¦g

n

¦ converge uniformly on E so we can ﬁnd N, M such that

n, m > N → [f

n

(x) −f

m

(x)[ < δ, n, m > M → [g

n

(x) −g

m

(x)[ < δ

[f

n

g

n

−f

m

g

m

[ = [(f

n

−f

m

)(g

n

−g

m

) +f

m

(g

n

−g

m

) +g

m

(f

n

−f

m

)[

≤ [(f

n

−f

m

)[[(g

n

−g

m

)[ +[f

m

[[(g

n

−g

m

)[ +[g

m

[[(f

n

−f

m

)[ ≤ δ

2

+[f

m

[δ +[g

m

[δ ≤ δ

2

Fδ +Gδ

≤

__

3

_

2

+F

_

3F

_

+G

_

3G

_

=

Exercise 7.3

Let ¦f

n

¦ be a sequence such that f

n

(x) = x. This function obviously converges uniformly to the function

f(x) = x on the set R. Let ¦g

n

¦ be a sequence of constant functions such that g

n

(x) =

1

n

. This sequence obviously

converges uniformly to the function g(x) = 0. Their product is the sequence ¦f

n

g

n

¦ where f

n

(x)g

n

(x) = x/n.

It’s clear that this sequence converges pointwise to f

n

(x)g

n

(x) = 0.

To show that f

n

g

n

is not uniformly convergent, let > 0 be given. Choose an arbitrarily large n ∈ Z and

choose t ∈ R such that t > (n(n + 1)). We now have

[f

n

(t)g

n

(t) −f

n+1

(t)g

n+1

(t)[ =

¸

¸

¸

¸

t

n

−

t

n + 1

¸

¸

¸

¸

=

¸

¸

¸

¸

t

n(n + 1)

¸

¸

¸

¸

>

¸

¸

¸

¸

(n(n + 1))

n(n + 1)

¸

¸

¸

¸

=

This shows that the necessary requirements for uniform convergence given in theorem 7.8 do not hold.

Exercise 7.4

Holy shit this exercise is a mess.

For what values of x does the series converge absolutely? (Incomplete)

For values of x > 0 we have

¸

¸

¸

¸

1

1 +n

2

x

¸

¸

¸

¸

=

¸

¸

¸

¸

1

x

¸

¸

¸

¸

¸

¸

¸

¸

1

1/x +n

2

¸

¸

¸

¸

≤

¸

¸

¸

¸

1

x

¸

¸

¸

¸

¸

¸

¸

¸

1

n

2

¸

¸

¸

¸

By the comparison test this shows that f(x) converges absolutely. If x = 0 then we have

¸

¸

¸

¸

1

1 +n

2

x

¸

¸

¸

¸

=

[1[ = ∞

This series clearly doesn’t converge, absolutely or otherwise. If x < 0 things get more complicated: If x = −1/n

2

for any n ∈ N then the nth term of the series is undeﬁned and therefore f(x) is undeﬁned. If x ,= −1/n

2

for any

n ∈ N then we can use the fact that

For what intervals does the function converge uniformly?

If E is any interval of the form [a, b] with a > 0 then we have

sup [f

n

(x)[ = f

n

(a) =

1

1 +n

2

a

124

And therefore we have

sup [f

n

(x)[ =

1

1 +n

2

a

≤

1

a

1

n

2

This shows that

sup [f

n

(x)[ converges by the comparison test, and so by theorem 7.10 we see that the series

f

n

(x) converges uniformly on E.

If E is any interval of the form [a, b] with b < 0 that does not contain any elements of the form −1/n

2

, n ∈ N

For what intervals does the function fail to converge uniformly?

Deﬁne the set X = ¦x

n

¦ with x

n

= −1/n

2

. The function f will fail to converge uniformly on any interval that

contains an element of X ∪ 0 or has an element of X ∪ 0 as a limit point of E.

Proof: Let E be an arbitrary interval. If E contains any x

n

∈ X then f(x

n

) undeﬁned and so f fails to

converge uniformly on E. If E contains 0 then f(0) =

**1 = ∞, but we will never ﬁnd some ﬁnite N such that
**

¸

¸

¸

N

1

1 −f(0)

¸

¸

¸ < so it’s clear that f fails to converge uniformly on E.

Now suppose that some x

n

∈ X is a limit point of E. The nth term of f is unbounded near x

n

, so f is

unbounded near x

n

, and therefore lim

t→xn

f(t) = ∞. From this we have

lim

n→∞

lim

t→xn

f(t) = lim

n→∞

∞ = ∞

On the other hand, if we ﬁrst ﬁx a value of t and take the limit of f as n → ∞ we have

lim

n→∞

f(t) = lim

n→∞

1

1 +n

2

t

= 0

and therefore

lim

t→xn

lim

n→∞

f(t) = lim

t→xn

0 = 0

Exercise 7.5

If x ≤ 0 or x > 1 then x = 0 for all n and therefore in these cases lim

n→∞

f

n

(x) = 0. For any other x, choose

an integer N large enough so that N > 1/x. For all n > N we now have x > 1/n and therefore f

n

(x) = 0, so

for this case we have lim

n→∞

f

n

(x) = 0. This exhausts all possible values of x, so ¦f

n

¦ converges pointwise to

the continuous function f(x) = 0.

To show that this function doesn’t converge uniformly to f(x) = 0 let 1 > > 0 be given and let n be an

arbitrarily large integer. We can easily verify that

f

n

_

1

n + 1/2

_

= sin

2

(nπ +π/2) = 1

and therefore the deﬁnition of uniform convergence is not satisﬁed.

The last part of the question asks us to use the series

f

n

to show that absolute convergence for all x does

not imply uniform convergence. The proof of this is simple: we’ve already shown that ¦f

n

¦ is not uniformly

convergent but

[f

n

(x)[ converges to f(x) = [ sin

2

(π/x)[ so

[f

n

[ is absolutely convergent.

Exercise 7.6

To show that the series doesn’t converge absolutely for any x:

¸

¸

¸

¸

(−1)

n

x

2

+n

n

2

¸

¸

¸

¸

=

¸

¸

¸

¸

x

2

n

2

+

1

n

¸

¸

¸

¸

≥

¸

¸

¸

¸

1

n

¸

¸

¸

¸

The rightmost sum is the harmonic series, which is known to diverge. By comparison test the leftmost series

must also diverge.

125

To show that the series converges uniformly in every bounded interval [a, b], let > 0 be given. Deﬁne X to

be sup¦[a[, [b[¦. Deﬁne the partial sum f

m

to be

f

m

=

m

n=1

(−1)

n

x

2

+n

n

2

We can rearrange this algebraically to form

f

m

=

m

n=1

(−1)

n

1

n

+x

2

m

n=1

(−1)

n

1

n

2

We know that the alternating harmonic series converges (theorem 3.44, or example 3.40(d)) and that

1/n

2

converges (theorem 3.28). Therefore by the Cauchy criterion for convergence we can ﬁnd N such that

p > q > N → [

p

n=q

(−1)

n

1

n

[ <

2

and we can also ﬁnd M such that

p > q > M → [

p

n=q

(−1)

n

1

n

2

[ <

2[X[

2

So, by choosing p > q > max¦N, M¦ we have (for all x ∈ [a, b]):

[f

p

(x) −f

q

(x)[ =

¸

¸

¸

¸

¸

p

n=q

(−1)

n

1

n

+x

2

p

n=q

(−1)

n

1

n

2

¸

¸

¸

¸

¸

≤

¸

¸

¸

¸

¸

p

n=q

(−1)

n

1

n

¸

¸

¸

¸

¸

+

¸

¸

¸

¸

¸

x

2

p

n=q

(−1)

n

1

n

2

¸

¸

¸

¸

¸

≤

2

+

x

2

2X

2

≤

2

+

2

=

Exercise 7.7

We can establish a global maximum for [f

n

(x)[

The derivative of f

n

(x) is

f

n

(x) =

1 +nx

2

−2nx

2

(1 +nx

2

)

2

=

1 −nx

2

(1 +nx

2

)

2

This derivative is zero only when nx

2

= 1, which occurs only when x = ±1/

√

n, at which point we have

f

n

_

±1

√

n

_

=

±1

2

√

n

These extrema must be the global extrema for the function, since f

n

has no asymptotes and f

n

(x) → 0 as

x → ±∞. Therefore [f

n

(x)[ ≤ 1/(2

√

n).

f

n

converges uniformly to f(x) = 0

Clearly f

n

converges pointwise to f(x) = 0, since for any x we have

lim

n→∞

f

n

(x) = lim

n→∞

x

1 +nx

2

= 0

126

Now let > 0 be given and choose n suﬃciently large so that 1/(2

√

n) < : this can be done by choosing

n >

1

4

2

. From the previously established bounds, we now have

[f

n

(x) −f(x)[ = [f

n

(x) −0[ = [f

n

(x)[ ≤

1

2

√

n

<

By theorem 7.9 this is suﬃcient to prove that f

n

converges uniformly to f(x) = 0.

When does f

(x) = limf

n

(x)?

We’ve established that f(x) = 0, so clearly f

(x) = 0 for all x. The limit of f

n

is given by:

lim

n→∞

f

n

(x) = lim

n→∞

1 −nx

2

n

2

x

4

+ 2nx

2

+ 1

=

_

0 x ,= 0

1 x = 0

This shows that it’s not necessarily true that [limf

n

]

= lim[f

n

] even if f

n

converges uniformly.

Exercise 7.8

Proof of uniform convergence

Let > 0 be given. Let f

m

represent the partial sum

f

m

=

m

n=1

c

n

I(x −x

i

)

We’re told that

[c

n

[ converges, so

[c

n

[ satisﬁes the Cauchy convergence criterion, so we can ﬁnd N such

that p > q > N implies

p

n=q

[c

n

[ <

For the same values of p > q > N we also have, by the triangle inequality and the fact that I(x −x

n

) ≤ 1,

[f

p

−f

q

[ =

¸

¸

¸

¸

¸

p

n=q

c

n

I(x −x

n

)

¸

¸

¸

¸

¸

≤

p

n=q

[c

n

I(x −x

n

)[ ≤

p

n=q

[c

n

[ <

so ¦f

n

¦ converges uniformly by theorem 7.8

Proof of continuity

Let t be an arbitrary point (not necessarily in the interval (a, b)). This t either is or isn’t a limit of some

subsequence of ¦x

n

¦.

If t is not a limit point of some subsequence then we can ﬁnd some neighborhood around t that contains no

points of ¦x

n

¦; the function I(t −x

n

) is constant on this interval for all n and therefore f(x) =

c

n

I(x −x

n

)

is constant on this interval for all n. (see below for a a more thorough justiﬁcation of this claim). And if f is

constant on an interval around t then it is clearly continuous at t.

Now suppose that t a limit point of ¦x

n

¦. By the Cauchy convergence of

[c

n

[ we can ﬁnd N such that

∞

n=N

[c

n

[ < . Choose a neighborhood around t small enough that it does not contain the ﬁrst N terms of ¦x

n

¦;

let δ represent the radius of this neighborhood. If we choose s such that [s −t[ < δ, then I(s −x

n

) = I(t −x

n

)

for n = 1, 2, . . . , N. From this we have

[s −t[ < δ → f(s) −f(t) ≤

¸

¸

¸

¸

¸

∞

n=N+1

(c

n

I(s −x

n

) −c

n

I(t −x

n

))

¸

¸

¸

¸

¸

≤

∞

n=N+1

[c

n

[I(s −x

n

) −I(t −x

n

)][

127

Each I(x −x

n

) is either 0 or 1 so [I(s −x

n

) −I(t −x

n

)[ ≤ 1, so we have

[s −t[ < δ → f(s) −f(t) ≤

∞

n=N+1

[c

n

[I(s −x

n

) −I(t −x

n

)][ ≤

∞

n=N+1

[c

n

[ <

That is, we’ve found δ such that

[s −t[ < δ → f(s) −f(t) <

which, by deﬁnition, means that f is continuous at t.

We have therefore shown that f is continuous at t under all cases. But t was an arbitrary point (not

necessarily conﬁned to (a, b)) and therefore f is continuous at all points.

Justiﬁcation of I(x −x

n

) being constant on some interval

As mentioned above, if t is not a limit point of some subsequence then we can ﬁnd some neighborhood around

t that contains no points of ¦x

n

¦. Let A be the set of elements of ¦x

n

¦ that are greater than t, and let B be

the set of elements of ¦x

n

¦ that are smaller than t. If [s −t[ < δ then there are no elements of ¦x

n

¦ between s

and t and so every element of A is greater than both t and s and every element of B is smaller than both t and

s. From this we see that I(s − x

n

) = I(t − x

n

) = 0 if x

n

∈ A and I(s − x

n

) = I(t − x

n

) = 1 if x

n

∈ B. This

means that I(x − x

n

) has the same value for every x ∈ N

δ

(t); so

c

n

I(x − x

n

) has the same value for every

x ∈ N

δ

(t); and therefore f(x) has the same value for every x ∈ N

δ

(t).

Exercise 7.9

We’re told that ¦f

n

¦ → f uniformly on E, so we can ﬁnd N such that

n > N → [f

n

(t) −f(t)[ <

2

for all t ∈ E

This holds for all t ∈ E, so it must also hold for the elements of ¦x

n

¦, which means that

n > N → [f

n

(x

n

) −f(x

n

)[ <

2

(85)

We’re also told that ¦f

n

¦ is a uniformly convergent sequence of continuous functions, so f is continuous; and

since ¦x

n

¦ → x we can ﬁnd M such that

n > M → [f(x

n

) −f(x)[ <

2

(86)

Using the triangle inequality and equations (85) and (86), we have

[f

n

(x

n

) −f(x)[ ≤ [f

n

(x

n

) −f(x

n

)[ +[f(x

n

) −f(x)[ <

2

+

2

=

Converse statement 1

Suppose the converse is “Let ¦f

n

¦ be a sequence of continuous functions that converges uniformly to f. Is it

true that if limf

n

(x

n

) = f(x) then ¦x

n

¦ → x?”. The answer to this question is “no”. Consider the function

f

n

(x) = x/n and the sequence ¦x

n

¦ = ¦1¦ (an inﬁnite sequence of 1s). The sequence ¦f

n

¦ converges uniformly

to f(x) = 0 on the set [0, 1], so we have

lim

n→∞

f

n

(x

n

) = lim

n→∞

f

n

(1) = 0 = f(0)

It’s clear that ¦x

n

¦ ,→ 0, so the converse does not hold.

128

Converse statement 2

Suppose the converse is “Let ¦f

n

¦ be a sequence such that limf

n

(x

n

) = f(x) for some function f and for all

sequences of points x

n

∈ E such that ¦x

n

¦ → x ∈ E. Is it true that ¦f

n

¦ converges uniformly to f on E?

The answer to this question is “no”. Let f

n

(x) = 1/(nx). Let E be the harmonic set ¦1/n¦. Then f

n

converges pointwise on E to f(x) = 0. Although 0 is a limit point of E, 0 is not an element of E and this

converse statement is only concerned with sequences ¦x

n

¦ that converge to some x ∈ E. So the only sequences

of points x

n

∈ E such that ¦x

n

¦ → x ∈ E are sequences where every term of the sequence is eventually just

x itself (that is, sequences for which there is some N such that n > N → x

n

= x). So clearly for every such

sequence we must have limf

n

(x

n

) = limf

n

(x) = f(x). But, despite the fact that limf

n

(x

n

) = f(x) whenever

¦x

n

¦ → x ∈ E, it’s clear that f

n

(x) does not converge uniformly to f(x) = 0 on E (to prove this, choose any

and any n and then choose x < (n)

−1

).

Exercise 7.10a

f is continuous at every irrational number

We know that

1/n

2

converges, so by the Cauchy criterion we can choose N such that

b > a > N →

b

a

1

n

2

<

4

The fractional part of (nx) will never be zero because x is irrational, so we can choose δ such that

0 < δ <

1

2

min¦(nx), 1 −(nx) : n < N¦

This guarantees that 0 < (n[x −δ]) < (nx) for all n < N and (nx) < (n[x +δ]) < 1 for all n. More importantly,

this choice of δ guarantees that (n[x − δ]) = (nx) − (nδ) for n < N. We now derive the following chain of

inequalities:

[f(x) −f(x ±δ)[ =

¸

¸

¸

¸

¸

N

1

_

(nx)

n

2

−

(nx ±nδ)

n

2

_

+

_

∞

N+1

(nx)

n

2

−

(nx ±nδ)

n

2

_¸

¸

¸

¸

¸

≤

¸

¸

¸

¸

¸

N

1

_

(nx)

n

2

−

(nx ±nδ)

n

2

_

¸

¸

¸

¸

¸

+

¸

¸

¸

¸

¸

∞

N+1

(nx)

n

2

¸

¸

¸

¸

¸

+

¸

¸

¸

¸

¸

∞

N+1

(nx ±nδ)

n

2

¸

¸

¸

¸

¸

<

¸

¸

¸

¸

¸

N

1

_

(nx)

n

2

−

(nx ±nδ)

n

2

_

¸

¸

¸

¸

¸

+

4

+

4

<

¸

¸

¸

¸

¸

N

1

_

(nx)

n

2

−

(nx)

n

2

±

(nδ)

n

2

_

¸

¸

¸

¸

¸

+

2

=

¸

¸

¸

¸

¸

N

1

_

±(nδ)

n

2

_

¸

¸

¸

¸

¸

+

2

We chose our value of δ after we ﬁxed a particular value of N, so we could have chosen δ small enough to make

this sum less than /2. In doing so, we have

[f(x) −f(x ±δ)[ <

2

+

2

=

And this would hold not just for x ±δ, but for all y such that [x −y[ < δ. That is:

[x −y[ < δ → [f(x) −f(y)[ <

which means that f is continuous at x.

129

Lemma 1:

(nx)/n

2

converges uniformly to f

This is an immediate consequence of the fact that (nx) < 1 for all n, x and that

1/n

2

converges. Let > 0

be given. By the Cauchy criterion we can choose N such that

b > a > N →

b

a

1

n

2

<

and therefore we have

¸

¸

¸

¸

¸

f(x) −

N

n=1

(nx)

n

2

¸

¸

¸

¸

¸

=

¸

¸

¸

¸

¸

∞

n=1

(nx)

n

2

−

N

n=1

(nx)

n

2

¸

¸

¸

¸

¸

=

¸

¸

¸

¸

¸

∞

n=N+1

(nx)

n

2

¸

¸

¸

¸

¸

≤

∞

n=N+1

¸

¸

¸

¸

(nx)

n

2

¸

¸

¸

¸

<

∞

n=N+1

¸

¸

¸

¸

1

n

2

¸

¸

¸

¸

<

f is discontinuous at every rational number

When nx is not an integer the following limits are identical:

lim

t→x+

(nt)

n

2

=

(nx)

n

2

= lim

t→x−

(nt)

n

2

When nx is an integer then this equality no longer holds. Instead, we have

lim

t→x+

(nt)

n

2

=

0

n

2

, lim

t→x−

(nt)

n

2

=

1

n

2

We know by lemma 1 that

(nx)/n

2

converges uniformly to f, so we also have (by theorem 7.11):

∞

n=1

f

n

(x+) = f(x+),

∞

n=1

f

n

(x−) = f(x−)

We know that

1/n

2

converges, so by the Cauchy criterion we can choose N such that

b > a > N →

b

a

1

n

2

<

4

If f were continuous at our rational point x, then we would have [f(x+) − f(x−)[ = 0. But when we actually

calculate this diﬀerence we ﬁnd:

[f(x+) −f(x−)[ =

¸

¸

¸

¸

¸

∞

n=1

f

n

(x+) −

∞

n=1

f

n

(x−)

¸

¸

¸

¸

¸

We determined that f

n

(x+) = f

n

(x−) unless nx is an integer; this occurs when n is a multiple of q, at which

point nx = [mq]x = mq[p/q] = mp. Most of the terms of the summations cancel out, leaving us with

[f(x+)−f(x−)[ =

¸

¸

¸

¸

¸

∞

m=1

f

mq

(x+) −

∞

m=1

f

mq

(x−)

¸

¸

¸

¸

¸

=

¸

¸

¸

¸

¸

∞

m=1

lim

t→x+

(mq)

[mq]

2

− lim

t→x−

(mq)

[mq]

2

¸

¸

¸

¸

¸

=

¸

¸

¸

¸

¸

∞

m=1

(0)

n

2

−

1

[mq]

2

¸

¸

¸

¸

¸

=

¸

¸

¸

¸

¸

∞

m=1

1

[mq]

2

¸

¸

¸

¸

¸

This is clearly not equal to zero and it can’t be made arbitrarily small because we have no freedom to select

particular values for m or q. Therefore f(x+) ,= f(x−) and therefore f is not rational at x. But x was an

arbitrary rational number, and therefore f is not continuous at any rational number.

Exercise 7.10b

We’ve shown that the discontinuities of f are the rational numbers, and these are clearly countable and clearly

dense in R.

130

Exercise 7.10c

To prove that f ∈ R we need only show that (nx)/n

2

∈ R for any ﬁxed value of n. We can then use the fact that

(nx)/n

2

converges uniformly to f (lemma 1) and theorem 7.16 to show that

_

(nx)/n

2

=

_

(nx)/n

2

=

_

f.

Let n ∈ N be given. Let [a, b] be an arbitrary interval. Assume without loss of generality that a and b are

integers.

_

b

a

(nx)

n

2

dx =

b−1

k=a

_

k+1

k

(nx)

n

2

dx

For 0 ≤ δ < 1 we have (n(k +δ)) = (nk +nδ) = (nδ), so we can integrate over (0, 1) instead of (k, k +1) without

changing the value of the integral.

_

b

a

(nx)

n

2

dx =

b−1

k=a

_

1

0

(nx)

n

2

dx

As x ranges from 0 to 1, nx ranges from 0 to n. So we split up the interval (0, 1) into n intervals of length 1/n:

_

b

a

(nx)

n

2

dx =

b−1

k=a

n−1

j=0

_

[j+1]/n

j/n

(nx)

n

2

dx

To make this integral a bit more manageable we make the variable substitution u = nx. When x = j/n we have

u = j; when x = [j + 1]/n we have u = j + 1; and we have dx =

1

n

du. Using this variable substitution in the

previous integral:

_

b

a

(nx)

n

2

dx =

b−1

k=a

n−1

j=0

_

j+1

j

(u)

n

3

du

Again, there is no diﬀerence between the value of (u) on the intervals (j, j + 1) and the value of (u) on the

interval (0, 1):

_

b

a

(nx)

n

2

dx =

b−1

k=a

n−1

j=0

_

1

0

(u)

n

3

du

On the interval (0, 1) we have (u) = u, so we can ﬁnally start integrating this thing. After a bunch of trivial

calculus, we have

_

b

a

(nx)

n

2

dx = [b −a −1][n]

u

2

2n

3

=

b −a −1

n

2

And therefore

_

b

a

f(x) =

_

b

a

∞

n=1

(nx)

n

2

=

∞

n=1

_

b

a

(nx)

n

2

=

∞

n=1

b −a −1

n

2

We know this rightmost sum converges (speciﬁcally, it converges to π

2

[b −a−1]/6), therefore

_

b

a

f(x) exists and

therefore f ∈ R.

Exercise 7.11

Let > 0 be given. We’re told that

f

n

has uniformly bounded partial sums, so let M represent the upper

bound of the partial sums of [

f

n

[. We’re told that g

n

→ 0 uniformly, so we can ﬁnd N such that

n > N → g

n

(x) <

M

for all x

Let A

n

represent the partial sum

n

k=1

f

k

g

k

. Our result is proven if we can prove that ¦A

n

¦ satisﬁes the

Cauchy converge criterion. To do this, we choose q > p > N. Following the logic of theorem 3.42, we have

[A

q

−A

p−1

[ =

¸

¸

¸

¸

¸

¸

q

k=p

f

k

g

k

¸

¸

¸

¸

¸

¸

=

¸

¸

¸

¸

¸

¸

_

_

q−1

k=p

A

n

(g

n

−g

n+1

)

_

_

+A

q

g

q

−A

p−1

g

p

¸

¸

¸

¸

¸

¸

≤ M

¸

¸

¸

¸

¸

¸

_

_

q−1

k=p

(g

n

−g

n+1

)

_

_

+g

q

−g

p

¸

¸

¸

¸

¸

¸

131

The (g

n

−g

n+1

) terms telescope down, leaving us with

[A

q

−A

p−1

[ ≤ M

¸

¸

¸

¸

¸

¸

_

_

g

p

−

q−1

k=p

(g

n

−g

n+1

)

_

_

+g

q

+g

p

¸

¸

¸

¸

¸

¸

= 2M[g

p

[ ≤ 2M[g

N

[ =

Exercise 7.12

Let > 0 be given. We’re told that 0 ≤ f

n

, f ≤ g and that

_

g < ∞. Let M =

_

∞

0

g. Deﬁne G(t) to be

G(t) =

_

c

0

g(x) dx

Since G(t) is continuous (theorem 6.30) and is strictly increasing (because g > 0) and has an upper bound of

M, we know that

lim

t→∞

G(t) = M

we can ﬁnd some N such that

n > N → M −G(n) <

2

which is equivalent to saying that

n > N →

_

∞

0

g −

_

c

0

g =

_

∞

c

g <

2

(87)

We’re given that f

n

→ f uniformly, so for all x and any speciﬁed value for c we can ﬁnd some M such that

m > M → [f

m

(x) −f(x)[ <

2c

(88)

from which we can conclude that

m > M →

¸

¸

¸

¸

_

c

0

f

m

−f

¸

¸

¸

¸

≤

_

c

0

[f

m

−f[ ≤ c

2c

=

2

(89)

So, for the given value of > 0 we choose c large enough that (88) holds and then, based on this choice of c, we

choose n to be large enough that (89) holds. By the triangle inequality we then have

¸

¸

¸

¸

_

∞

0

f

n

−f

¸

¸

¸

¸

≤

¸

¸

¸

¸

_

c

0

f

n

−f

¸

¸

¸

¸

+

¸

¸

¸

¸

_

∞

c

f

n

−f

¸

¸

¸

¸

≤

_

c

0

[f

n

−f[ +

_

∞

c

[f

n

−f[

≤

2

+

2

=

We can make this hold for arbitrary by taking n suﬃciently large, which is simply saying that

lim

n→∞

¸

¸

¸

¸

_

∞

0

f

n

−f

¸

¸

¸

¸

= 0

from which we conclude

lim

n→∞

_

∞

0

f

n

=

_

∞

0

f

132

Exercise 7.13a

For any choice of x

1

∈ [0, 1], the sequence ¦f

(1)

n

(x

1

)¦ is a bounded sequence of real numbers in the compact

domain [0, 1]. Therefore there exists some subsequence of functions ¦f

(1)

n

¦ for which the subsequence of real

numbers ¦f

(1)

n

(x

1

)¦ converges to some point in [0, 1] (theorem 3.6a). Call this subsequence of functions ¦f

(2)

n

¦

and choose any x

2

∈ [0, 1]. The sequence ¦f

(2)

(x

2

) is still a bounded sequence of real numbers in the compact

domain [0, 1], so we can ﬁnd some subsequence of functions ¦f

(3)

n

¦ for which ¦f

n

(x

1

)¦ and ¦f

n

(x

2

)¦ both con-

verge. This can be repeated a countable number of times to construct a sequence of functions ¦F

n

¦ for which

¦F

n

(x)¦ converges at all rational numbers and for all points of discontinuity for each f

n

(this set is countable

by theorems 4.30,2.12, and 2.13).

Now let t be a rational number or a point of discontinuity for some f

n

. We have constructed ¦F

n

¦ so that

¦F

n

(t)¦ converges to some point, so we simply deﬁne f(t) for rational t to be

f(t) = lim

n→∞

F

n

(t), if t is rational or a point of discontinuity

Now t be a rational number and a point at which f

n

is continuous for all n. The rational numbers are dense in

R so we can construct a sequence ¦r

n

¦ of rational numbers such that ¦r

n

¦ → t. Each F

n

is continuous at t, so

lim

k→∞

F

n

(r

k

) = F

n

(t) for every n. Therefore we simply deﬁne the value of f(t) to be

f(t) = lim

n→∞

F

n

(t), if t is irrational and each F

n

is continuous at t

Exercise 7.13: example

An example where a sequence ¦f

n

¦ of monotonically increasing functions converges pointwise to f and f is not

continuous:

f

n

(x) =

_

_

_

0 x ≤ 0

1 −

1

1+n

x ≥ 1

1 −

1

1+nx

0 < x < 1

lim

n→∞

f

n

(x) =

_

0 x ≤ 0

1 x > 0

Exercise 7.13b

Let α = inf f(x) and let β = sup f(x) (these values may be ±∞). Let > 0 be given. The function f must

come arbitrarily close to its supremum and inﬁmum on R. Let a be a point at which [f(a) − α[ < and let b

be a point at which [f(b) −β[ < . From the monotonicity of f we have

x < a → [f(x) −α[ < , x > b → [f(x) −β[ < (90)

Proving uniform convergence is now a matter of proving uniform convergence on [a, b].

We’re given that f is continuous on [a, b]; therefore it’s uniformly continuous on [a, b] (theorem 4.19); therefore

we can ﬁnd some δ such that

[x −y[ < δ → [f(x) −f(y) < (91)

We can cover [a, b] with ﬁnitely many intervals of length δ/2 ([a, b] is ﬁnite, so we don’t even need to rely on

compactness for this). From each interval we choose a rational number q

i

. We know that f converges pointwise

at each rational number, and the set of q

i

s is a ﬁnite set, so we can ﬁnd some integer N such that

n > N → [F

n

(q

i

) −f(q

i

)[ < for each q

i

(92)

Now choose any x ∈ [a, b]. If x = q

i

for some i then by (92) we have n > N → [F

n

(q

i

) − f(q

i

)[ < . If x ,= q

i

then we can ﬁnd q

i

, q

i+1

such that q

i

< x < q

i+1

and [q

i+1

−q

i

[ < δ. By the triangle inequality:

[F

n

(x) −f(x)[ ≤ [F

n

(x) −F

n

(q

i+1

)[ +[F

n

(q

i+1

) −f(q

i+1

)[ +[f(q

i+1

) −f(x)[

Each F

n

is monotonic, so we have [F

n

(x) −F

n

(q

i+1

)[ ≤ [F

n

(q

i

) −F

n

(q

i+1

)[ so this last inequality becomes

[F

n

(x) −f(x)[ ≤ [F

n

(q

i

) −F

n

(q

i+1

)[ +[F

n

(q

i+1

) −f(q

i+1

)[ +[f(q

i+1

) −f(x)[

133

The term [F

n

(q

i+1

) −f(q

i+1

)[ is < by (92). The term [f(q

i+1

) −f(x)[ is < by (91). This leaves us with

[F

n

(x) −f(x)[ ≤ [F

n

(q

i

) −F

n

(q

i+1

)[ + 2

Additional applications of the triangle inequality gives us

[F

n

(x) −f(x)[ ≤ [F

n

(q

i

) −f(q

i

)[ +[f(q

i

) −f(q

i+1

)[[f(q

i

) −F

n

(q

i+1

)[ + 2

The terms [F

n

(q

i

) − f(q

i

)[ and [f(q

i

) − F

n

(q

i+1

)[ are both < by (92). The term [f(q

i

) − f(q

i+1

)[ is < by

(91). This leave us with

[F

n

(x) −f(x)[ ≤ 5

Exercise 7.14

Exercise 7.15

Let > 0 be given. Each f

n

is equicontinous on [0, 1], which means that there exists some δ such that, for all n,

[0 −y[ < δ → [f

n

(0) −f

n

(y)[ <

By the deﬁnition of f this means that, for all y ∈ (0, 1) and all n ∈ N,

[y[ < δ → [f(0) −f(ny)[ <

By taking n suﬃciently large and choosing y appropriately we can cause ny to take on any value in (0, ∞). So

we conclude that f is a constant function on [0, ∞).

Exercise 7.16

See part (b) of exercise 7.13

Exercise 7.17

Exercise 7.18

Lemma 1: ¦F

n

¦ is uniformly bounded

We’re told that ¦f

n

¦ is a uniformly bounded sequence: let M be the upper bound of ¦[f

n

[¦. For each n, we have

[F

n

(x)[ =

¸

¸

¸

¸

_

x

a

f

n

(t) dt

¸

¸

¸

¸

≤

_

x

a

[f

n

(t)[ dt ≤

_

b

a

[f

n

(t)[ dt ≤ (b −a)M

and therefore ¦[F

n

[¦ is uniformly bounded.

Lemma 2: ¦F

n

¦ is equicontinuous

Let > 0 be given. Let δ = /M.

[y −x[ < δ → [F

n

(y) −F

n

(x)[ =

¸

¸

¸

¸

_

y

x

f

n

(t) dt

¸

¸

¸

¸

≤

_

y

x

[f

n

(t) dt[ ≤ (y −x)M < δM <

Therefore, by theorem 7.25, we know that ¦F

n

¦ contains a uniformly convergent subsequence.

Exercise 7.19

This was solved during class

Exercise 7.21

We know that A separates points because f(e

iθ

) = e

iθ

∈ A. For this same function, [f(e

iθ

)[ = [e

iθ

[ = 1 and so

A vanishes at no point of K. But the function f(e

iθ

) = −e

iθ

is a continuous function that is not in the closure

of A.

134

Exercise 1.2

√ Proof by contradiction. If 12 were rational, then we could write it as a reduced-form fraction in the form of p/q where p and q are nonzero integers with no common divisors. →

p q

=

2

√

12

assumed

→ ( p2 = 12) q → (p2 = 12q 2 ) It’s clear that 3|12q 2 , which means that 3|p2 . By some theorem I can’t remember (possibly the deﬁnition of ‘prime’ itself), if a is a prime number and a|mn, then a|m ∨ a|n. Therefore, since 3|pp and 3 is prime, → 3|p → 9|p2 → (∃m ∈ N)(p2 = 9m) → (∃m ∈ N)(12q = 9m) → (∃m ∈ N)(4q = 3m) → (3|4q )

2 2 2

deﬁnition of divisibility substitution from p2 = 12q 2 divide both sides by 3 deﬁnition of divisibility

From the same property of prime divisors that we used previously, we know that 3|4 ∨ 3|q 2 : it clearly doesn’t divide 4, so it must be the case that 3|q 2 . But if 3|qq, then 3|q ∨ 3|q. Therefore: → (3|q) And this establishes a contradiction. We began by assuming that p and q had no common divisors, but we have shown that 3|p and 3|q. So our assumption must be wrong: there is no reduced-form rational number such √ that p = 12. q

Exercise 1.3 a

If x = 0 and xy = xz, then y = 1y = (x−1 x)y = x−1 (xy) = x−1 (xz) = (x−1 x)z = 1z = z

Exercise 1.3 b

If x = 0 and xy = x, then y = 1y = (x−1 x)y = x−1 (xy) = x−1 x = 1

Exercise 1.3 c

If x = 0 and xy = 1, then y = 1y = (x−1 x)y = x−1 (xy) = x−1 1 = x−1 = 1/x

Exercise 1.3 d

If x = 0, then the fact that x−1 x =1 means that x is the inverse of x−1 : that is, x = (x−1 )−1 = 1/(1/x).

Exercise 1.4

We are told that E is nonempty, so there exists some e ∈ E. By the deﬁnition of lower bound, (∀x ∈ E)(α ≤ x): so α ≤ e. By the deﬁnition of upper bound, (∀x ∈ E)(x ≤ β): so e ≤ β. Together, these two inequalities tell us that α ≤ e ≤ β. We’re told that S is ordered, so by the transitivity of order relations this implies α ≤ β.

2

Exercise 1.5

We’re told that A is bounded below. The ﬁeld of real numbers has the greatest lower bound property, so we’re guaranteed to have a greatest lower bound for A. Let β be this greatest lower bound. To prove that −β is the least upper bound of −A, we must ﬁrst show that it’s an upper bound. Let −x be an arbitrary element in −A: → −x ∈ −A →x∈A →β≤x → −β ≥ −x assumed deﬁnition of membership in −A β = inf(A) consequence of 1.18(a)

We began with an arbitrary choice of −x, so this proves that (∀ − x ∈ −A)(−β ≥ −x), which by deﬁnition tells us that −β is an upper bound for −A. To show that −β is the least such upper bound for −A, we choose some arbitrary element less than −β: → α < −β → −α > β assumed consequence of 1.18(a)

Remember that β is the greatest lower bound of A. If −α is larger than inf(A), there must be some element of A that is smaller than −α. → (∃x ∈ A)(x < −α) → (∃ − x ∈ −A)(−x > α) → !(∀ − x ∈ −A)(−x ≤ α) → α is not an upper bound of −A (see above) consequence of 1.18(a) deﬁnition of upper bound

This proves that any element less than −β is not an upper bound of −A. Together with the earlier proof that −β is an upper bound of −A, this proves that −β is the least upper bound of −A.

Exercise 1.6a

The diﬃcult part of this proof is deciding which, if any, of the familiar properties of exponents are considered axioms and which properties we need to prove. It seems impossible to make any progress on this proof unless we 1 can assume that (bm )n = bmn . On the other hand, it seems clear that we can’t simply assume that (bm ) n = bm/n : this would make the proof trivial (and is essentially assuming what we’re trying to prove). As I understand this problem, we have deﬁned xn in such a way that it is trivial to prove that (xa )b = xab 1 when a and b are integers. And we’ve declared in theorem 1.21 that, by deﬁnition, the symbol x n is the element 1 n n such that (x ) = x. But we haven’t deﬁned exactly what it might mean to combine an integer power like n and some arbitrary inverse like 1/r. We are asked to prove that these two elements do, in fact, combine in the way we would expect them to: (xn )1/r = xn/r . Unless otherwise noted, every step of the following proof is justiﬁed by theorem 1.21.

3

Let r = m and let s = p where m. q = 0. p. → (∃t ∈ Q)(bt > br ∧ t ≤ r) It can be shown that b−t > 0 (see theorem S1. We must still prove that it is the least upper bound of B(r). Proof by contradiction that br is an upper bound of B(r): → br is not an upper bound of B(r) → (∃x ∈ B(r))(x > br ) hypothesis of contradiction formalization of the hypothesis By the deﬁnition of membership in B(r). by the deﬁnition of the equality of rational numbers. → (∃t ∈ Q)(bt b−t > br b−t ∧ t ≤ r) → (∃t ∈ Q)(b t−t theorem S2 from part b >b r−t ∧ t ≤ r) → (∃t ∈ Q)(1 > br−t ∧ r − t ≥ 0) → (∃t ∈ Q)(1−(r−t) > b ∧ r − t ≥ 0) →1>b And this establishes our contradiction.6b As in the last proof.21 from part a )(b ) →b = (b )(b ) r+s →b = (br )(bs ) r+s p q Exercise 1. q ∈ Z and n q n. we assume that br+s = br bs when r and s are integers and try to prove that the operation works in a similar way when r and s are rational numbers. To do so.→ bm = bm → ((bm ) )n = bm → ((bm ) n )nq = bmq 1 1 n assumed deﬁnition of x n p q 1 We were told that m = n mq = np. means that commutativity of multiplication From theorem 12.6c We’re given that b > 1. though. Therefore: → ((bm ) n )nq = bnp → ((bm ) n )qn = bpn 1 1 which. p m → br+s = b n + q → br+s = b →b →b →b r+s mq+pn nq deﬁnition of addition for rationals ) 1 nq = (b mq+pn from part a legal because mq and pn are integers 1 nq → br+s = (bmq bpn ) r+s r+s 1 nq = (b = (b mq mq nq m n ) 1 nq (b ) pn nq pn corollary of 1. we need to 4 . below) so we can multiply this term against both sides of the inequality.1. since we were given that b > 1. Our initial assumption must have been incorrect: br is. we can take the n root of each side to get: → ((bm ) n )q = bp From theorem 12. n. Let r be a rational number. x = bt where t is rational and t ≤ r. From this.1. we can take the q root of each side to get: → (bm ) n = (bp ) q 1 1 1 Exercise 1. let α represent an arbitrary rational number such that bα < br . an upper bound of B(r). in fact.

.prove that α < r. . so bn − 1 ≥ (b − 1)(1n−1 + 1n−2 + . this proves that (∀n ∈ N)(bn − 1 ≥ n(b − 1)). and b are greater than 1 from part b transitivity of order relations n > 0 → n−1 > 0 would be a trivial proof 5 . Replacing 1 with b n gives us the inequality: → n(b n − 1) ≤ ((b n )n−1 + (b n )n−2 + . Alternatively. Now. + b0 an−1 ). → bα < br →b b →b →b α −r α−r hypothesis of contradiction r −r <b b theorem S2 from part b from part b <b r−r α−r <1 Exercise 1. + b0 an−1 )(b − a): → n(b n − 1) ≤ ((b n )n − 1) → n(b n − 1) ≤ (b − 1) 1 1 1 1 1 1 1 1 Exercise 1. And since b > 1. .7 a Proof by induction. + (b n )0 )(b n − 1) Now we can use the algebraic identity bn − an = (bn−1 a0 + bn−2 a1 + . . . . each term in the bn−k series is greater than 1. this becomes bn − 1 = (b − 1)(bn−1 + bn−2 + . Let S = {n : bn − 1 ≥ n(b − 1)}. we could prove this using the same identity that Rudin used in the proof of 1. Exercise 1.21. . t. k > 0 (see theorem S4). . + 10 ) = (b − 1)n. assume that k ∈ S: →k∈S → bk − 1 ≥ k(b − 1) → bbk − 1 ≥ k(b − 1) →b →b k+1 k+1 hypothesis of induction deﬁnition of membership in S we’re told that b > 1. . So when a = 1. . We can easily verify that 1 ∈ S. . .7 c → n > (b − 1)/(t − 1) → n(t − 1) > (b − 1) → n(t − 1) > (b − 1) ≥ n(b − 1) → n(t − 1) > n(b − 1) → (t − 1) > (b − 1) →t>b 1 n 1 n 1 n 1 n assumed this holds because n. . + b0 ). . + 1)(b n − 1) → n(b − 1) = (1 1 n 1 1 1 1 n times n−1 n−2 +1 + . From the distributive property we can verify that bn − an = (b − a)(bn−1 a0 + bn−2 a1 + . + 10 )(b n − 1) 1 1 It can be shown that bk > 1 when b > 1.7 b → n(b n − 1) = n(b n − 1) → n(b n − 1) = (1 + 1 + . − b ≥ k(b − 1) ≥ k(b − 1) + b → bk+1 − 1 ≥ k(b − 1) + b − 1 → bk+1 − 1 ≥ (k + 1)(b − 1) → k+1∈S deﬁnition of membership in S By induction.

Exercise 1. then from part (d) we can choose n such that y > bx+ n > bx . From this we see that x − n is an upper bound of A that is smaller than x. −1) > 0. 1) > (0. 0) : → (0. 0) = 0. However. −1 1 1 n > 1 > 0 means that. 1) > (0. 0) + (0. which means that 1 < yb−w . 0) → (0. −1) are both greater than zero. 1 1 If bx > y. However. Exercise 1. −1) + (0. which means that bw y −1 > 1.17(i) of ordered ﬁelds deﬁnition of complex multiplication This conclusion is in contradiction of trichotomy. bw = by . 0) can take the role of 0 in deﬁnition 1. From this we see that x is not an upper bound of A. 1) is not comparable to (0. all elements of the set must be comparable (the trichotomy rule. Multiplying this by b n gives us b > y. Having ruled out these two possibilities. We will show by contradiction that (0. 0) > (0. upon taking the reciprocals. 0) → (0. From this we get y > bw+ n . 0) > 0 or (0.7 e We’re told that bw > y. 0) → (−1. 0)(0. deﬁnition 1.7 d We’re told that bw < y. b) = (a. we’re lead 1 1 directly to the conclusion that we can select n such that yb−w > b n . 0) < (0.8 In any ordered set. 0) + (a. 1) > (0. Next. Then by the transitivity of equality relations. Therefore: → (0.5). the fact that b Exercise 1. 1) < (0. 0): 6 . 1) > (0. → (−1. −1) > (0.17 of an ordered ﬁeld. 1) deﬁnition 1. 1) → (0. 0) in any potential ordered ﬁeld containing C. we’re 1 lead directly to the conclusion that we can select n such that bw y −1 > b n .7 f We’ll prove that bx = y by showing that the assumptions bx > y and bx < y lead to contradictions.7 g Assume that there are two elements such that bw = y and bx = y. 1) and (0. b) → (0. 0) > (0. although this seems suspiciously simple. Exercise 1. the trichotomy property of ordered ﬁelds forces us to conclude that bx = y. which is what we were asked to prove. This is a contradiction. As a corollary. since we’ve assumed that x =sup(A). 1). 0) hypothesis of contradiction deﬁnition 1. since we initially assumed that (0. we assume that (0. since we initially assumed (0. 1) > (0. since we’ve assumed that x is the least upper bound of A. Using the substitution bw y −1 = t with part (c).Exercise 1. This is a contradiction. even a bizarre one in which −1 > −i > 0.17(ii) of ordered ﬁelds We assumed here that (0. This is a safe assumption because the uniqueness property of the additive identity shows us immediately that (0. 0) deﬁnition of complex multiplication deﬁnition 1. we assume that (0. we’re trying to show that the complex ﬁeld cannot be an ordered ﬁeld under any ordered relation. Using the substitution yb−w = t with part (c). 1)(0. we’ve shown that (0. then from part (e) we can choose n such that bx > bx− n > y. As a corollary. 1) > 0 deﬁnition of complex multiplication It might seem that we have established our contradiction as soon as we concluded that (−1. we have b n < 1 and therefore bw− n < bw .17(ii) of ordered ﬁelds. First. Multiplying both sides by y gives us −1 1 1 w w− n b > b n y. 1 If bx < y. which is what 1 1 w+ n w we were asked to prove. the fact that b n > 1 means that b >b .

0) for every complex number. −1) > (0. b) > (c. d) ∨ (a. 0) = (0. 0) hypothesis of contradiction deﬁnition 1. → (a < e) ∨ (a < e) ∨ (a < e) ∨ (a = e ∧ b < f ) → a < e ∨ (a = e ∧ b < f ) → (a. this would mean that every element is equal to every other. −1) > (0. d) ∨ (a. 0) → (−1. Exercise 1. d) ∈ C → a.→ (0. b) < (c. 0) > (0. 0) → (0. 0) > (0. We’re trying to prove the transitivity of the dictionary order relation on C. −1) → (0. we know that → (a. 1) = a(0. 0) → (−1. b. 0) + (0. b) and (c. And this is in contradiction of deﬁnition 1. Proof by contradiction that (0.17(ii) of ordered ﬁelds. f ) p∧q →p p∨p→p deﬁnition of this order relation To prove that the trichotomy property holds for the dictionary relation on Q. 0) + b(0. we are not assuming what we’re trying to prove.9a To prove that this relation turns C into an ordered set. b) > (c. Proof of transitivity: → (a. 0) we’re led to the conclusion that (a. b) > (c. This last step is using the transitivity of this standard order relation on R and is not assuming that transitivity holds for the dictionary order relation. d) And this is the deﬁnition of the trichotomy law. b) ∈ C ∧ (c. 1) > (0.17(i) of ordered ﬁelds deﬁnition of complex addition deﬁnition 1. since we’ve established (0. 0). −1) > (0. d) < (e. 1)4 + b(0. Let (a. d) ∧ (c. d) ∨ (a. b) = (0. From the standard order relation. b) = (c. By the transitivity of equivalence relations. 1) + (0. so we have proven that the dictionary order turns the deﬁnition of the dictionary order relation distributivity of the logical operators assumed deﬁnition of a complex number trichotomy of the order relation on R trichotomy of the order relation on R 7 . 0) deﬁnition of complex multiplication Once again trichotomy has been violated. 1) → (0. and this relation is deﬁned in terms of the standard order relation on R. b) < (c. −1) > (0. 0) → (0. 0)(0. −1) > (0. b) < (c.5. d) ∨ (a. −1)(0. f ) → [a < c ∨ (a = c ∧ b < d)] ∧ [c < e ∨ (c = e ∧ d < f )] → (a < c ∧ c < e) ∨ (a < c ∧ c = e ∧ d < f ) ∨(a = c ∧ b < d ∧ c < e) ∨ (a = c ∧ b < d ∧ c = e ∧ d < f ) → (a < e) ∨ (a < e ∧ d < f ) ∨ (a < e ∧ b < d) ∨ (a = e ∧ b < f ) distributivity of logical operators transitivity of order relation on R assumption deﬁnition of this order relation Although we’re falling back on the the transitivity of an order relation. d ∈ R → (a < c) ∨ (a > c) ∨ (a = c) → (a < c) ∨ (a > c) ∨ (a = c ∧ (b < d ∨ b > d ∨ b = d)) → (a < c) ∨ (a > c) ∨ (a = c ∧ b < d) ∨ (a = c ∧ b > d) ∨(a = c ∧ b = d) → (a. c. d) → (a.17(ii) of ordered ﬁelds deﬁnition of complex multiplication deﬁnitino 1. b) = a(0. b) < (c. 0): if we assume that (0. we need to show that it satisﬁes the two requirements in deﬁnition 1.12 of a ﬁeld. b) = (c. d) be two elements in C. we rely on the trichotomy property of the underlying standard order relation on R. since (a. d) ∨ (a. d) ∨(a. where we’re told that there are at least two distinct elements: the additive identity (’0’) and the multiplicative identity (’1’). b) < (e. 1) = (0. 1) = (0.

9b C does not have the least upper bound property under the dictionary order. 8 . This gives us ¯ n 2 √ n zj zj ¯ ≤ j=1 | zj | √ n 2 j=1 ¯ | zj |2 j=1 which is equivalent to |z1 + z2 + . Taking the square root of both sides results in the claim we wanted to prove. then we can easily verify that |w| = 1 and that rw = |z| |z| = z. This subset is just the imaginary axis in the complex plane.31(a) = |x|2 − 2|x||y| + |y|2 = (|x| − |y|)(|x| − |y|) = (|x| − |y|)(|¯| − |¯|) x y = (|x| − |y|)(|x| − |y|) = ||x| − |y|| 2 This chain of inequalities shows us that ||x| − |y||2 ≤ |x − y|2 . .35). so −|x| ≥ |x|. + zn | ≤ |z1 | + |z2 | + . y) < ( . . Exercise 1. This subset clearly has an upper bound. 2 Exercise 1. deﬁnition 1.33(d) Theorem 1.33(c) Theorem 1.33(b) Theorem 1. y) with x > 0.31(a) Theorem 1. and is too tedious to write out.31(c). + |zn |) Taking the square root of each side shows that |z1 + z2 + . 0) > (0. . Let E = {(0. we see that x (x. a) 2 So that ( x . . since (x. 2 Exercise 1. Exercise 1. . which is a contradiction. Theorem 1.32 x ≤ |x|.12 √ √ Set ai zi and bi = zi and use the Cauchy-Schwarz inequality (theorem 1. .33(b) Theorem 1.13 |x − y|2 = (x − y)(x − y) = (x − y)(¯ − y ) x ¯ = x¯ − x¯ − y¯ + y y x y x ¯ = x¯ − (x¯ + y¯) + y y x y x ¯ = |x|2 − 2Re(x¯) + |y|2 y ≥ |x|2 − 2|Re(x¯)| + |y|2 y ≥ |x|2 − 2|x¯| + |y|2 y = |x| − 2|x||¯| + |y| y 2 2 Theorem 1. a) : a ∈ R}.10 This is just straightforward algebra.complex numbers into an ordered set. a) for any x > 0. y) is an upper bound less than our proposed least upper bound. .11 If we choose w = z |z| z and choose r = |z|. Exercise 1. But it does not have a least upper bound: for any proposed upper bound (x. + zn |2 ≤ (|z1 | + |z2 | + . + |zn | which is what we were asked to prove. . y) < (0.

15 Using the logic and the notation from Rudin’s proof of theorem 1. and from the given chain of equalities we see that this occurs when |Baj − Cbj |2 = 0.16 We know that |z − x|2 = |z − y|2 = r2 . this becomes ˆ cos(θ) = d 2r where θ is the angle between the ﬁxed vector (y − x) and the variable vector u.31(a) The conjugate of 1 = 1 + 0i is just 1 − 0i = 1. and informal proof would be to appeal to the relationship a · b = |a||b| cos(θ) where θ is the angle between the two vectors. which occurs when z = x + rˆ where |ˆ| = 1. it will hold for no u when d > 2r. the u u requirement that r2 = |z − y|2 becomes z · (x − y) = r2 = |z − y|2 = |x + rˆ − y|2 = |(x − y) + rˆ2 | = |x − y|2 + 2rˆ · (x − y) + |rˆ|2 = d2 + 2rˆ · (x − y) + r2 u u u u u Rearranging some terms. we see that equality holds in the Schwarz inequality when AB = |C|2 . this becomes u · (y − x) = ˆ d2 d = |y − x| 2r 2r (2) A quick. 9 . each aj must be a constant multiple of bj . Using this representation of z. Exercise 1. the previous equation then becomes |ˆ||y − x| cos(θ) = u d |y − x| 2r Dividing by |y − x| and remembering that u = 1. It’s easy to see that this equation ˆ will hold for exactly one u when d = 2r.Exercise 1. and it will hold for inﬁnitely many values of u when d < 2r and n > 2.14 |1 + z|2 + |1 − z|2 = (1 + z)(1 + z) + (1 − z)(1 − z) = (1 + z)(¯ + z ) + (1 − z)(¯ − z ) 1 ¯ 1 ¯ = (1 + z)(1 + z ) + (1 − z)(1 − z ) ¯ ¯ = (1 + z + z + z z ) + (1 − z − z + z z ) ¯ ¯ ¯ ¯ = (2 + 2z z ) ¯ = (2 + 2) =4 We are told that z z = 1 ¯ Theorem 1. More formal proofs follow. For this to occur we must have Baj = Cbj for all j.35. This occurs when have B(AB − |C|2 ) = 0. we have |z − x|2 = (z − x) · (z − x) = |z|2 − 2z · x + |x|2 |z − y|2 = (z − y) · (z − y) = |z|2 − 2z · y + |y|2 For these to be equal. Expanding these terms out. convincing. Each value of u ˆ ˆ corresponds with a unique value of z. we must have −2z · x + |x|2 = −2z · y + |y|2 which happens when 1 [|x|2 − |y|2 ] (1) 2 We also want |z − x| = r. Exercise 1. which occurs only when B = 0 or when C aj = bj for all j B That is. it will hold for two values of u when ˆ ˆ ˆ d < 2r and n = 2.

. . equation (2) is satisﬁed for all u for which ˆ u · (y − x) = ˆ d |y − x| < |y − x| 2r By the deﬁnition of the dot product.37e) and is therefore false for all z. . .part (a) When d < 2r. this is equivalent to u1 (y1 + x1 ) + u2 (y2 + x2 ) + . we have |x − z|2 + 2(x − z) · (z − y) + |z − y|2 = |z − x|2 + 2|z − x||z − y| + |z − y|2 which occurs if and only if 2(x − z) · (z − y) = 2|x − z||z − y| From exercise 14 we saw that this equality held only if (x − z) = c(z − y) for some constant c. As long as k ≥ 3 we have more variables than equations and therefore the system will have inﬁnitely many solutions. Therefore we have x − z = z − y. we know x = y so c = 1. part (c) If 2r < d then we have |x − y| > |x − z| + |z − y| which is equivalent to |(x − z) + (z − y)| > |x − z| + |z − y| which violates the triangle inequality (1. + u2 = 1 1 2 k (4) d |y − x| 2r (3) This gives us a system of two equations with k variables. we have: d2 = |x − y|2 = |(x − z) + (z − y)|2 = [(x − z) + (z − y)] · [(x − z) + (z − y)] = (x − z) · (x − z) + 2(x − z) · (z − y) + (z − y) · (z − y) = |x − z| + 2(x − z) · (z − y) + |z − y| Evaluating (2r)2 . we have: (2r)2 = (r + r)2 = (|z − x| + |z − y|)2 = |z − x|2 + 2|z − x||z − y| + |z − y|2 commutativity of multiplication 2 2 deﬁnition of inner product · inner products are distributive deﬁnition of inner product · If 2r = d then d2 = (2r)2 and therefore. part (b) Evaluating d2 . we know that |x − z| = |z − y| so c = ±1. by the above evaluations. from which we have z= x+y 2 and there is clearly only one such z that satisﬁes this equation. + uk (yk + xk ) = The only other requirement for the values of ui is that u2 + u2 + . 10 .

. . If x = 0. we can always ﬁnd a nonzero y such that x · y = 0.36 of inner product (ai bi + ai ci ) ai bi + =a·b+a·c (a + b) · c = = = (ai + bi )ci bi c i (ai ci + bi ci ) ai ci + =a·c+b·c The rest of the proof follows directly: |x + y|2 + |x − y|2 = (x + y) · (x + y) + (x − y) · (x − y) = (x + y) · x + (x + y) · y + (x − y) · x − (x − y) · y =x·x+y·x+x·y+y·y+x·x−y·x−x·y+y·y =x·x+y·y+x·x+y·y = 2|x|2 + 2|y|2 Exercise 1.36 of inner product distributive property of R associative property of R deﬁnition 1.36 of inner product distributive property of R associative property of R deﬁnition 1. We began with an arbitrary vector x and demonstrated a method of construction for y such that x · y = 0: therefore.19 We need to determine the circumstances under which |x − a| = 2|x − b| and |x − c| = r. we need to manipulate these equalities until they have a common term that we can use to compare them. This is not true in R1 because the nonzero elements of R are closed with respect to multiplication. Choose y such that: x b i=a xa yi = −1 i = b 0 otherwise xb We can now see that x·y = xa xa +xb (−1) = xb −xb = 0.36 of inner product deﬁnition 1. a · (b + c) = = = ai (bi + ci ) ai ci deﬁnition 1. Let xb represent any other element of x. . . then x · y = 0 for any y ∈ Rk . we have 11 . we need to prove that a · (b + c) = a · b + a · c and that (a + b) · c = a · c + b · c: that is. xk must be nonzero: let this element be represented by xa . we need to prove that the distributive property holds between the inner product operation and addition. Exercise 1.Exercise 1.17 First.18 If x = 0. |x − a| = 2|x − b| |x − a|2 = 4|x − b|2 |x|2 − 2x · a + |a|2 = 4|x|2 − 8x · b + 4|b|2 3|x|2 = |a|2 − 2x · a + 8x · b − 4|b|2 |x − c| = r |x − c|2 = r2 |x|2 − 2x · c + |c|2 = r2 |x|2 = r2 + 2x · c − |c|2 3|x|2 = 3r2 + 6x · c − 3|c|2 Combining these last two equalities together. To do this. then at least one of the elements x1 .

we have a new deﬁnition of √ “cut” that includes cuts such as (−∞. so γ = Q. so p ∈ α for some cut α ∈ A. This means that γ itself is a proper subset of β. pick some arbitrary rational p ∈ γ. And we’re being asked to disregard part (III). To say that an element α ∈ R has a least upper bound. we guarantee that |x − a| = 2|x − b| iﬀ |x − c| = r. show that the subset γ is also an element of R. To prove part (II). so x ∈ A → x ⊂ β (remember that < is deﬁned as ⊂). We’ve deﬁned γ as the union of all cuts in A. γ is the set of cuts in A. By (II). But γ is just the union of cut elements in A. To prove that each subset of R with an upper bound must have a least upper bound.it’s also true that s ∈ δ. this means 12 . then. By the deﬁnition of < for this set.→ |a|2 − 2x · a + 8x · b − 4|b|2 = 3r2 + 6x · c − 3|c|2 → |a|2 − 2x · a + 8x · b − 4|b|2 − 3r2 − 6x · c + 3|c|2 = 0 → |a|2 − 4|b|2 + 3|c|2 − 2x · (a − 4b − 3c) − 3r2 = 0 We can zero out the dot product in this equation by letting 3c = 4b − a. This is a proper subset. and then show that the subset/element γ is the least upper bound of A. this tells us that α ≤ γ for every α ∈ A. proof that γ is the least upper bound of A (i) Choose any arbitrary cut α ∈ A. Exercise 1. so γ is the union of proper subsets of β. we will follow Step 3 in the book almost exactly. Let A be a nonempty subset of R with an upper bound of β (Note: A is a subset of R. For every r ∈ δ. the fact that p ∈ α and q < p means that q ∈ α and therefore q ∈ γ. so there must be some rational s such that s ∈ δ but s ∈ γ. Choose a second rational q such that q < p: by the deﬁnition of cut. The elements of A are cuts.20 We’re trying to show that R has the least upper bound property. is to say that the subset α has some “smallest” superset β such that α ⊂ β. this also determines a speciﬁc value of c. this means that δ ⊂ γ. So criterion (I) in the deﬁnition of “cut” has been met. α ⊆ γ. which are subsets of Q). and we were told that A is nonempty. And by the deﬁnition of < for this set. (ii) Suppose that δ < γ. We’re asked to omit property III. and < is deﬁned to be “is a proper subset of”. We will deﬁne two subsets A ⊂ R and γ ⊂ R. This shows that γ ⊂ β ⊆ Q. We can now show that δ ⊂ α. 1] and (−∞. Deﬁne γ to be the union of all cuts in A. This means that γ contains every element from every cut in A: γ consists of elements from Q. it must be the case that s ∈ α for some cut α ∈ A. (ii) We are told that β is an upper bound for A. not a subset of R. 2]. Of course. With this restriction lifted. so it’s clear that every rational number in α is also contained in γ: that is. so this is suﬃcient to prove that γ is a cut. Therefore γ is nonempty. The elements of R are certain subsets of Q. proof that γ is a cut: The proof of criterion (I) has two parts. which told us that each α ∈ R has no largest element. In order for s ∈ γ. This substitution gives us the new equality: → |a|2 − 4|b|2 + 3 4b a − 3r2 = 0 − 3 3 3 · 16 2 3 · 8 3 → |a|2 − 4|b|2 + |b| − a · b + |a|2 − 3r2 = 0 9 9 9 → 3|a|2 − 12|b|2 + 16|b|2 − 8a · b + |a|2 − 9r2 = 0 2 → 4|a|2 4|b|2 − 8a · b − 9r2 = 0 → 4|a − b|2 = 9r2 → 2|a − b| = 3r By choosing 3c = 4b − a and 3r = 2|a − b|. (i) γ is the union of elements of A.

we conclude that α = α + 0# for any α ∈ R. Then r + s ≤ r. By the deﬁnition of α. Together. 0) = 0∗ . So it cannot be the case that α + β = 0# . Let s represent the diﬀerence q − (−r): then q = −r + s (note that s must be a positive rational). or p + q < 0. To show that A5 is not satisﬁed. 0]: we will show that this assumption leads to a contradiction. r ∈ α. Whether or not β contains an element greater than −r. ﬁeld axiom A5: existence of additive inverses The book constructed the deﬁnition of inverse so that the inverse of (−∞. 0] which is not a cut because of criterion (III)). and p ≤ r. x) are cuts under our new deﬁnition: both of these elements must have additive inverses. these two facts show that γ is the least upper bound of A. these two cuts are not equal ((−∞. Let α = (−∞. −x). r) for some rational r. x] is (∞. This overlooks the fact that both (−∞. closed under addition (otherwise we’d have 0∗ + 0∗ = (−∞. and this new deﬁnition results in the existence of elements in R that have no inverse. x]). In conclusion. so we see that α + β contains some positive rational: it cannot be the case that α + β = 0# . x] + 0∗ = (−∞. so this shows that 0 ∈ α + β. the set of cuts. Let 0# be the set of all rational numbers ≤ 0.that r < s (using standard rational order). ﬁeld axiom A4: existence of an additive identity Let α be a cut in R. x) would be (−∞.A2. but by omitting requirement (III) we are forced to redeﬁne our zero. x) + (−∞. (II) also shows that r ∈ α. The set 0∗ can no longer function as our zero since (−∞. and associativity are directly applicable to our new deﬁnition of cut. Now let p and r be rationals such that p ∈ α. And since s ∈ α. This shows us that α + 0# ⊆ α. commutativity. (ii) Assume that β contains at least one element q that is greater than −r. we know that p < r and q ≤ −r. Our new zero 0# must include 0 as an element because our newly deﬁned cuts can have a greatest element. Combining these two inequalities tells us that p + q < −r + r. and therefore δ is not an upper bound of A. Let p be an arbitrary rational such that p ∈ α. (i) Assume that β does not contain any elements greater than −r. Adding q = −r + s to each element in this equality gives us 0 < p + q < s. since (−∞. deﬁnition of addition Following the example in the book. This shows that δ ⊂ α. which means δ < α ∈ A (using the subset order on R). Having shown that α + 0# ⊆ α and that α ⊆ α + 0# . so 0∗ has not functioned as the additive identity. so p − r ∈ 0# . we ﬁnd ourselves in contradiction with the initial assumption that β might be the inverse of α. we deﬁne the addition of two cuts α + β to be the set of all sums r + s where r ∈ α and s ∈ β. It is tempting to just deﬁne the inverse for our new deﬁnition of cut so that the inverse of (−∞. −x]. And from property (II) of cuts. And p + q is an element of α + β. which means that p ∈ α + 0# . This means that p − r ≤ 0. we know that p < r. x] and (−∞. we see that omitting (III) forces us to redeﬁne the additive identity. 13 . Let p and q be arbitrary rationals such that p ∈ α and q ∈ β. which means that r + s ∈ α. This shows us that α ⊆ α + 0# . Therefore r + (p − r) ∈ α + 0# . ﬁeld axioms A1. The original deﬁnition had to omit 0 from 0∗ in order to keep R. From our deﬁnitions of α and β. Therefore there can be no inverse for α. we need to demonstrate that at least one element of R has no additive inverse. x). This works under the original deﬁnition of “cut”. But p and q were arbitrary members of α and β. −x) = (−∞. and A3 The proofs from the book for closure. Assume that there was some cut β such that α + β = 0# = (−∞. we know that r − s < p < r. The book deﬁned 0∗ to be the set of rational numbers < 0. Let r and s be rationals such that r ∈ α and s ∈ 0# . x) < (−∞.

which is not countable. x ∈ ∅ → x ∈ A. More speciﬁcally.2 The set of integers is countable. If the irrational real numbers were countable.x: The Cantor Set The idea behind the Cantor set is that each term in the series E1 ⊃ E2 . . . ⊃ En removes the middle third of each remaining line segment. For each a∈ Zk consider the polynomial a0 z k +a1 z k−1 +. so if every real number were algebraic then the set of real algebraic numbers would be uncountable.1 This is immediately justiﬁed by noting that the deﬁnition of subset. Exercise 2. Exercise 2. There is clearly a unique solution for each n ∈ Z. Exercise 2. then the union of rational and irrational reals would be countable. consider the roots for 2-tuples in Z1 . This only tells us that R is at most countable: it is either countable or ﬁnite. I did not use the hint provided in the text. . . each of which corresponds with k roots of a k-degree polynomial. So the irrational reals are not countable. which either means that this proof is invalid or that there is an alternate (simpler?) proof. a1 . we have a countable number of Zk s. Let this set be represented by Zk .3 The set of real numbers is uncountable. Because R is also at most countable. each containing a countable number of (k + 1)-tuples. But this union is just R.2 showed that the algebraic complex numbers were countable. exercise 2.Exercise 2. so the set of algebraic real numbers is at most countable. . . so R is an inﬁnite set.13 the set of all (k + 1)-tuples (a0 . To show that R is not ﬁnite. → (∀x)(x ∈ ∅ → x ∈ A) →∅⊆A contrapositive deﬁnition of subset Exercise 2. then everything implies it. n) corresponds with the polynomial −z + n = 0 whose solution is z = n. .4 The rational real numbers are countable. However. so by theorem 2. From the fundamental theorem of algebra. The real numbers are a subset of the complex numbers. A more formal proof: → ¬(∃x ∈ ∅) → (∀x)(x ∈ ∅) → (∀x)(x ∈ A → x ∈ ∅) deﬁnition of an empty set negation of the ∃ quantiﬁer Hilbert’s PL1 This previous step is justiﬁed by the argument p → (q → p): if something is true (like x ∈ ∅). So our set of complex roots (call it R) is a countable union of countable unions of ﬁnite sets. . We now have a series of nested sets that encompass every possible root for every possible polynomial with integer coeﬃcients. ak ) with a0 = 0 is also countable. this proves that R is countable. is satisﬁed for any set A because of the false antecedent. Each 2-tuple of the form (−1.+ak = 0. so that you get a series of increasingly smaller segments that look like: 14 . we know that there are exactly k complex roots for this polynomial.

We are assured that this is possible by the Archimedean property of the rationals. 1 Now let F be the subset of (2. And y. by virtue of being in E . This is equivalent to proving that every limit point of E is a point of E. we can ﬁnd a point y ∈ E in the neighborhood Ns (x). If there were. β). z) is less than t. (ii) Proof that 0 is a limit point: choose an arbitrarily small radius r.6 (i) We’re asked to prove that E is closed. 1] consisting only all rational numbers of the form m .15 of metric spaces assures us that d(x. From the Archimedian property. m ∈ N. z) < s + t < r. 4}. To justify the claim that no point in P is an isolated point. z) < r and therefore z is in the neighborhood Nr (x) for any 1t was chosen to be less than r to guarantee x = z. And since this was true for an arbitrary point p1 ∈ P and an arbitrarily small r. So we just choose r such that r = 2 d(pm . The distance d(x. In each subsequent term in the series {Ei }. 1 (i) Proof that no point in E is a limit point: let pm be an arbitrary member of E. Choose some element Em where m is so large that the segments in Em are all shorter than r. Exercise 2. is a limit point of E: so we can ﬁnd a point z ∈ E in the neighborhood Nt (y). pm ). There can be no points of E in the neighborhood Nr (x). Choose any arbitrarily small r and let r r s = 2 . then this 1 1 would indicate the existence of a point of E less than zero. 3] containing rational numbers of the form 2 + m and let G be the subset of 1 (4. and d(0. This is the reasoning behind Rudin’s statement that “P contains no segment”. the endpoint p2 will never be “cut oﬀ”: it will always be the endpoint of a segment from which the middle third is lost. Since x is a limit point of E . So p2 ∈ P . None of these can exist in E as we deﬁned it. Therefore the union E ∪ F ∪ G is bounded and has three limit points {0. Let the dx represent the smallest distance from the set {d(x. pm ) = m < r. E1 has two segments of length 1/3. d(x. 0). t = 4 1 . d(x. β) and therefore (α. we can choose m to be large enough so that the maximum segment length is less than the length of (α. so d(p1 . This shows that 0 is a limit point. so no points other than 0 are limit points. 2. we can ﬁnd some second point p2 such that d(p1 . so deﬁnition 2. 5] consisting of rational numbers of the form 4 + m . so d(x. The length of this segment is less than r. and Em has 2m segments of length 1/3m . this shows that every neighborhood of every point in P is a limit point and not an isolated point. but the limit points of E need not be members of E: we will demonstrate that the point 0 is a limit point of E. z) ≤ d(x. y) is less than s and d(y. Then pm = m for some 1 1 m ∈ N. And by the deﬁnition of E . The closest point to pm is therefore pm+1 = m+1 . greater than 1. Notice that for any possible segment (α. let x represent an arbitrary limit point of E . greater than 1. 15 . p2 ) < r. we 1 1 1 can ﬁnd some m ∈ N such that m < r. (iii) Proof that no other points are limit points: let x be some point that is neither zero nor a member of E.5 1 Let E be a subset of (0. so if p1 ∈ P then it is a member of every Ei . pm+1 ∈ E. pm+1 )}. or between m and m+1 . P is deﬁned to be Ei . This pm = m is a member of E. Let p2 represent one of the endpoints of this line. 1 Choose r such that r = 2 dx . it’s not in the union P . β) ∈ Em .We can see that E0 has one segment of length 1. pm+1 ) and we see that there are no other points in this neighborhood of pm . 1). So we’ve found an Em with a line segment containing p1 . or it must lie between two sequential points pm . y) + d(y. Exercise 2. And if the segment is not in Em for some m. this is equivalent to proving that every limit point of E is a limit point of E. we need to show that for every point p1 ∈ P and any arbitrarily small radius r. these sets can be shown to have limit points of only 2 and 4. To prove this. d(x. since we need only ﬁnd m such that 3−m < r. This point must be either less than zero. which shows that pm is an isolated point. Just as E was shown to have a limit point of only zero. No point in E is a limit point. p2 ) < r.

z) < r and therefore we can ﬁnd some z ∈ E in the neighborhood Nr (x) for any arbitrarily small r. z) ≤ d(x. Let x represent ¯ ¯ an arbitrary limit point of E (which we won’t assume to be a member of E). Therefore every neighborhood of x contains an element of E. ¯ We’ve shown that every limit point of E is a limit point of E and vice-versa. y) + d(y. So x is a limit point for E. so we have proven that every limit point of E is a limit point of E. And if x ∈ Ak . Then Ns (x) contains 16 .7a We’ll prove the equality of these sets by showing that each is a subset of the other. The distance d(x. 2. then x ∈ Ak for some Ak ∈ ¯ Ai . The following image is helpful for imagining these points in R2 . (ii. From this. t = 4 . ¯ If x is a limit point of E. (ii. ¯ ¯ contains an element of E. For any neighborhood Nt (x) with t > s. For any neighborhood Nt (x) with t < s. And since every neighborhood of x contains an element of E. each of these inﬁnitely many points in Ns (x) is either a member of E or a member of E. (iii) We’re asked if E and E always have the same limit points. then x ∈ Ak . Choose any arbitrarily small r r r ¯ theorem 2. By deﬁnition. ¯ (ii) We can use a similar technique to show that every limit point of E is a limit point E. The answer is “no”. and a counterexample can be found in the previous question. Then. the fact that Ns (x) ⊂ Nt (x) means that Nt (x) contains inﬁnitely many points in E. But x was an arbitrarily chosen limit point of E . 4}. which is what we were asked to prove. Because E is deﬁned to be E ∪ E. assume that there exists at least one point y in Ns (x) such that y ∈ E . So x is a limit point of E. Then for each i. Exercise 2. The limit points of E in exercise 2. Let s =min(ri ) (which exists because i is ﬁnite).15 of metric spaces assures us that d(x. So the point x is a limit point for E. then it must be a limit point at least one speciﬁc Ak ∈ Ai . so there is at least one point z ∈ E in the neighborhood Nt (y).b) Next. so d(x. But this means that x is a limit point of E. Since x is a limit point of E. assume that none of the points in Ns (x) are members of E. Proof by contradiction: assume that x is not a limit point for any Ak ∈ Ai . z) is less than t.5 were E = {0. then every neighborhood of x Every element of E is also an element of E ∪ E = E.a) First. z) < s + t < r. ¯ The second part of the proof is to show that every limit point of E is a limit point of E and is relatively trivial. (A2) If x is a limit point of Bn . x is a limit point for E. so deﬁnition 2. And this means that x ∈ ¯ Ai . ¯ (A) Assume that x ∈ Bn . we see that every neighborhood of x ¯ contains elements of E. (A1) If x ∈ Bn . x ∈ Bn or x is a limit point of Bn . This means that all of the inﬁnitely many points of E ∪ E in Ns (x) are members of E. And E clearly has no limit points whatsoever.20 tells us that the neighborhood Ns (x) contains and let s = 2 . there is some neighborhood Nri (x) that contains no elements of Ai . y) is less than s and d(y. the fact that x is a limit point of E and that Nt (x) ⊂ Ns (x) means that Nt (x) contains inﬁnitely many points in E. ¯ ¯ inﬁnitely many points of E. which is what we were asked to prove. this means that y is a limit point of E.arbitrarily small r.

we will need to show that every point in the neighborhood Nr (x) is also a 17 . then it contains no points of Bn = Ai . But part A of the proof assumed that we could choose a least element from the set of size i.. so we have proven that Bn = Ai . Together. First. But ¯i for any i.8b Consider a closed set consisting of a single point. By ¯ deﬁnition.e. then x ∈ (B2) If x is a limit point of Ak . every point of Nr (x) is a point of E.5. This point is clearly not a limit point of E. and Ns (p) ⊂ Nr (p) ⊂ E. since x is a limit point of Bn . But every element of Ak is an element of Ai = Bn . We saw in exercise 2. we know that q ∈ Ns (p). x must be a limit point of at least one speciﬁc Ak ∈ Ai . which we can’t do for an ¯ inﬁnite set. Now choose s such that s < r. by contradiction. This means that we can choose some r such that Nr (x) ⊂ E: i. so it’s still the case that i=1 Ai ⊂ B. ¯ ¯ Ai . (B1) If x ∈ Ak . We can say two things about the neighborhoods of p. And this is a contradiction. x ∈ (B) Assume that x ∈ point of Ak . And it’s a good thing that we can’t assume this. so this means that every neighborhood of x contains an element of Bn . We now show that this guarantees that every neighborhood of p contains a point of E. so Ns (p) contains inﬁnitely many points of E. these two facts tell us that Nr (p) contains inﬁnitely many points of E.7b ∞ ¯ ¯ Nothing in part B of the above proof required that i be ﬁnite. the set E consisting of all rational numbers of the form m . so 0 ∈ ∞ Ai .5 that 0 ∈ B. every neighborhood of p contains inﬁnitely many points. such as E = {(0. ¯ Therefore. Exercise 2. And if this neighborhood contains no points for any Ai . By the deﬁnition of E ◦ . which means that x ∈ Bn and therefore x ∈ BN .no elements from any Ai since Ns (x) ⊆ Nri (x) for each i. We’ll ∞ 1 ¯ construct this set by deﬁning Ai = i and deﬁning B = i=1 Ai . This shows us that B = ∞ Ai for these sets. ¯ In either of these two cases. In either of these two cases. This proves that x ∈ ¯ ¯ ¯ ¯ x ∈ Bn → x ∈ Ai . So x ∈ Ak . Exercise 2. To show that x is an interior point of E ◦ . Thus p is a limit point of E. m ∈ N. we will show that every point in E ◦ must be an interior point of E ◦ . Choose any s such that r < s. so for this particular choice of sets we see that Ai ⊂ B. Because the neighborhood Nr (p) contains some q ∈ E and Nr (p) ⊂ Ns (p). Ns (p) contains inﬁnitely many points. And this means that either x ∈ Ak or x is a limit ¯ Ai . Proof: let p be an arbitrary point in E. 1 Consider the set from exercise 2. which means ¯ that x ∈ Ai . we know that x is an interior point of E. because it’s false. though. ¯ ¯ Ai → x ∈ Bn . 0)}. Ns (p) contains a point of E. And we’ve already shown that ¯ ¯ 0∈A i=1 i=1 ¯ ¯ Ai ⊆ B. since we’re dealing with R2 . We’ve shown that whether r < s or r > s for any arbitrary r. And we’ve already shown that Exercise 2. so that x ∈ Bn . So we can’t assume that B ⊆ Ai . This proves that x ∈ Bn → x ∈ ¯ Ai . Second. p is an interior point (since E is open) and so there is some r such that Nr (p) ⊂ E. this means that x is a limit point of Bn .9a To prove that E ◦ is open. Let x be an arbitrary point in E ◦ . then every neighborhood of x contain an element of Ak . ¯ ¯ Ai .8a Every point in an open set E in R2 is a limit point of E. x ∈ Bn . Then x ∈ Ak for some k. Exercise 2.

Exercise 2. And since x was an arbitrary point in E ◦ .9d Proof that E ◦ ⊂ E: let x be a member of E ◦ . the complement of E ◦ . then by deﬁnition we know that x is a limit point of E. the complement of E ◦ . If x = y for every neighborhood. x is a member of E. Clearly. either x is a limit point for E or x itself is a member of E. by deﬁnition we know that y is an interior point of E. then this shows us that G ⊂ E ◦ . y) + s = r and let z be an arbitrary point in Ns (y). It must be true that either x = y or x = y. and we’ve established that every point in Nr (x) is a member of E. For each neighborhood. the closure of E. And this proves that E ◦ is open. By deﬁnition of “interior point”. so we know that every point in Nr (x) is an interior point of E. these two statements show that E is open if and only if E = E ◦ . this means that x is not an interior point of E and therefore x ∈ E ◦ . So we conclude that either x ∈ E or x is a limit point for E: that is. this means that every point in E ◦ is an interior point of E ◦ . From the deﬁnition of closure. 18 . Let x be a member of E. And if every point in G is an interior point of E. z) ≤ d(x. we know that E ◦ is always open: so E is open if E = E ◦ . Proof that E ⊂ E ◦ : this proof is only a very slight modiﬁcation of the previous one. Together. Exercise 2. then every point of E is an interior point and therefore E ◦ = E: so E = E ◦ if E is open. And if E is open. p is an interior point of G. since d(x. This means that for every point p ∈ G there is some neighborhood Nr (p) such that Nr (p) ⊂ G ⊂ E. z) < r. this means that every neighborhood of x contains some element y in the complement of E. And since every point in Ns (y) is a point in E. 0 < d(x. From the deﬁnition of E ◦ we know that x is not an interior point of E. Let y be an arbitrary point in Nr (x).9c Let p be an arbitrary element of G. Together. but also of E. From the deﬁnition of “interior point”.9(a).9b From exercise 2. y) + s = r implies d(x. Exercise 2. Because G is open. If x = y for one or more of these neighborhoods. these two proofs show us that E ◦ = E. But y was an arbitrary point in Nr (x). x is an interior point of E ◦ . the closure of the complement of E.point of E ◦ . Choose a radius s such that d(x. y) + d(y. y) < r. then x is in E (since y is in E and x = y). And this means that x is itself an interior point of the set of interior points of E: that is. Every point in Ns (y) is a point in E. This chain of subsets tells us that every point in G is an interior point not only of G. In either case. every neighborhood of x contains some point of E. z) < d(x.

y) + d1 (y. By k deﬁnition. But our choice of E was arbitrary. so the closure of E ◦ is ∅. 1).11 d1 is not a metric: let x = −1. y = 0. this shows us that E contains no limit points. we select some gi in our (arbitrary) open cover {Gα } such that ei ∈ gi (note that E is ﬁnite. Therefore there is no ﬁnite subcover for this particular open cover. As we saw in exercise 2. z = 1.23. y) + d2 (y. z) > d1 (x. As we saw previously. If E is ﬁnite with cardinality k. so we don’t k k need to use the axiom of choice). So we can create an inﬁnite open cover of E by letting each set in {Gα } be a single element of E. then we can always generate an open cover that has no ﬁnite subcover. but it is a limit point of E.37(f) |x − y||y − z| 2 this additional term is > 0 |x − y| + |x − y| + |y − z| |y − z| legal because all terms are > 0 → d2 (x. z) ≤ d2 (x. Clearly E ∪ {0} = ∅. It’s trivial to verify that d2 has the properties 2. so by choosing r = . z) which violates deﬁnition 2. And this tells us that d1 (x. Any subcover of E will need to contain one element of {Gα } for each element of E: and since E is inﬁnite. The distance between any two distinct points is always 1.9f No. But our choice of E was arbitrary.15(c): → |x − z| ≤ |x − y| + |y − z| → |x − z| ≤ |x − y| + |y − z| + 2 → → |x − z| ≤ |x − z| ≤ 2 true by the triangle inequality. The point 0 is not a member of the interior of E. this means that E is open. y) + d1 (y. And the empty set is already closed. so this shows that every subset of X is closed. This gives us a collection of sets i=1 gi such that E ⊆ i=1 gi ⊆ {Gα }. m ∈ N. z) 19 . Because this is true for any point in E. Which sets are open? Let E be an arbitrary set in X.5.15(a) and 2. E and E ◦ do not always have the same closures. then. z) = 2 while d1 (x.5 we guarantee that the neighborhood Nr (p) contains only p itself. i=1 gi is a ﬁnite subcover of E. then we can always generate a ﬁnite subcover for any open cover. If E is inﬁnite. this means that any subcover must be inﬁnite. Exercise 2. Which sets are compact? Under this metric. the closure of E is E ∪ {0}. subsets of X are compact if and only if they are ﬁnite.10 Which sets are closed? Let E be an arbitrary subset of X and let p be an arbitrary point in E. every subset of X is open: this includes subsets that consist of a single element. so it must be the case that E is closed: and from theorem 2. z) = 4. 1] consisting only all rational 1 numbers of the form m . 1) in R1 . So E ◦ = ∅. For each ei ∈ E. Then d1 (x. To show that it has the property 2. and {Gα } were arbitrary. theorem 1. Proof: let E be an arbitrary subset of X. E and E do not always have the same interiors. d2 is a metric. Our choices for E. Exercise 2. so E and E ◦ do not always have the same closures. so this shows that every subset of X is open. So the point 0 is an interior point of E = (−1. Exercise 2. so E has no interior points. 0) ∪ (0. and E is not compact. k.Exercise 2. Consider the set E = (−1. We’ve shown that every set is closed.15(b). But every neighborhood of every point in E contains a point not in E. It’s then vacuously true that all of the (nonexistant) limit points of E are points of E: we conclude that E is closed. This set is either ﬁnite or inﬁnite.9e No.15(c) of metric spaces. Let E be a subset of (0. so this shows that any ﬁnite subset of X is compact.

which is a ﬁnite Exercise 2. this neighborhood contains 1 one element for every natural number n such that n > r . This neighborhood contains every n ∈ K such that n < r: equivalently. → |x − z| ≤ |x − y| + 2|x − y||y − z| + |y − z| + |x − z||x − y||x − z| Adding positive terms to the right side does not change the sign of the equality. which violates deﬁnition 2. To show that it has the property 2. d4 (2. → |x − z| |x − y| |y − z| ≤ + 1 + |x − z| 1 + |x − y| 1 + |y − z| → d2 (x. there is some neighborhood 1 1 Nr (0) ⊂ G0 .15(a) of metric spaces.15(a) of metric spaces. which violates deﬁnition 2.15(c): → |x − z| ≤ |x − y| + |y − z| This ﬁrst step is always true because of the triangle inequality (theorem 1. r Let Gi represent an element of {Gα } containing 1 . → |x − z| |x − y| + 2|x − y||y − z| + |y − z| ≤ 1 + |x − z| 1 + |x − y| + |y − z| + |x − y||y − z| And now we’ve divided each side by two of these factored terms. It’s trivial to verify that d2 has the properties 2. z) ≤ d2 (x. but we’ve just factored out terms from each side. −1) = 0. Not only does this mean that G0 contains an inﬁnite number of elements of K. We can now see that K ⊂ G0 ∪ i subcover of {Gα }. of course. 1/r i=1 Gi . z) The logic behind this seemingly arbitrary series of algebraic steps becomes clear if one starts at the end and works back to the beginning (which is. that contains 0 as an interior point.5). As an interior point of G0 . d3 (1.12 Let {Gα } be any open cover of K. 20 . At least one open set in {Gα } must contain the point 0: let G0 be this set. open in R. y) + d2 (y. let Ai represent the set of all points of the form Each Ai can be shown to have a single limit point of 1 i 1 i 1 + n that are contained in the open interval 1 1 i .15(a) and 2.15(b). → |x − z| + |x − z||x − y| + |x − z||y − z| + |x − z||x − y||y − z| ≤ |x − y| + 2|x − y||y − z| + |y − z| + |x − z||x − y| + 2|x − z||x − y||y − z| + |x − z||y − z| It’s hard to verify from this ugly mess. i−1 .37(f).13 For i ≥ 2. it means that it contains all but a ﬁnite number ( 1 .d3 is not a metric. (see exercise 2. d5 is a metric. Exercise 2. 1) = 0. how I initially derived the proof). but we’ve just added the same terms to each side. → (|x − z|)(1 + |x − y| + |y − z| + |x − y||y − z|) ≤ (1 + |x − z|)(|x − y| + 2|x − y||y − z| + |y − z|) It’s still hard to verify. to be exact) of elements of K. d4 is not a metric.

From theorem 1. since An+1 ⊂ An . As we saw in exercise 2.15 For each i ∈ N.14 1 Let Gn represent the interval n . ∞) are open in Q. Gi i contains all but a ﬁnite number of elements from the interval {Gα } containing 1 i ri i=1 . ∞). there 1 is some neighborhood Nr (0) ⊂ G0 . Let p be a rational number in the interval [x.12.20. let G0 represent the element of {Gα } that contains 0 as an interior point. i−1 + 1 . the rationals of the form 1 for i ≥ 2. the intervals [x.Now consider the union of sets S = i=2 Ai . The same reasoning used in exercise 2. this means that G0 contains all but a ﬁnite number of the limit points of S. Exercise 2. which is obviously empty. As an interior point. there is some n > 1/x and therefore some n < x. since the union i∈ Ai is equal to √ √ ∞ Ak .12. 1 1 i . For each of the ﬁnitely ri many intervals not covered by G0 . we see that all of the points of S in the interval j can be covered by the ﬁnite union of sets Gi ∪ Now. i−1 ). 21 . 1 . so every point of this interval is an interior point. This proves that the collection {Ai : i ∈ N} is a counterexample to theorem 2. which is a countable number. The union of any ﬁnite collection of these intervals will be nonempty. To show that S is compact. 1) but is not a member of any Gi ∈ H.16 lemma 1: For any real number x ∈ Q. 1) (Proof: for 1 any x ∈ (0. It cannot be the case that p = x (since p is rational while x is not). As in 2. Our subcover contains G0 . The union {Gα } = i=1 Gn is a cover for the interval (0. ∞) is identical to (x. ∞) or (x.36. So x ∈ Gn+1 ). So our assumption that a ﬁnite subcover exists must be false. ∞) (i. deﬁne Ai to be: Ai = p∈Q: 2− 1 ≤p≤ i 2+ 1 i Because the endpoints are irrational. so we have constructed a ﬁnite subcover for an arbitrary cover {Gα }. ∞). ∞) and (x. If we let Gij represent an element of 1 1 i . Let H be a ﬁnite subcover of {Gα }. where k is the largest index in . This means that Nr (p) is an interior point of (x. Proof: let x be a real number such that x ∈ Q. each of these intervals Ai are both bounded and closed (see exercise 2. For each i ≥ 2. so p is in the interval (x.16 for the proofs). ∞) = (x. ∞) in Q). Exercise 2. we return to the reasoning found in exercise 2.5 shows that the set of limit points for S is just the collection of limit points from Ai : that is. This is a ﬁnite collection of a ﬁnite number of elements from {Gα }. x). 1 though. i And from these sets we can construct our ﬁnite subcover. Then we see that (p − r. i−1 ∞ . we also know that each interval is nonempty. 1). p). so there is a least element.e. p + r) ⊆ (p − d(p. Let Gi represent the element of {Gα } that contains 1 . [x. i This shows us that S has one limit point for each natural number greater than 1. ∞). we include the ﬁnite union of sets Gi ∪ i=1 . ∞) are open in Q. This set is ﬁnite. And this least element is in the interval (0. and each element of S is included in one of these elements. it means that G0 contains all but a ﬁnite number of the intervals of the form ( 1 . This neighborhood contains every limit point of the form n ∈ R such that 1 n < r.36 if the word “compact” is replaced by either “closed” or “bounded”. ∞). we can rewrite this interval as (p − d(p. x). ∞): but p was an arbitrary point chosen from [x. Choose r = d(x. But the inﬁnite union i=1 Ai is equal to {p ∈ Q : 2 ≤ p ≤ 2}. Let {Gα } be any open cover of S. ∞ Exercise 2.12. Because x < p. More than that. and consider the set { i+1 : Gi ∈ H}. But there is no ﬁnite subcover 1 for (0. This same collection works as a counterexample to the corollary of 2. 1). This proves that [x. This proves that S is compact. ∞) and (x.

which means that the √ rational < √ E is the set of rational numbers in (− 3. . From lemma 3. lemma 4: Every interval of the form (x. Note that its complement is also one of these four open sets. Proof by contrapositive that every limit point of E is a member of E: let x be an element of [0. lemma 3: All of these open intervals ([x.24) and E is open (theorem 2. then p2 > 3 and p ∈ E. x) are open in Q. x]. The proofs for [x.23). ∞). So E is bounded by the interval (− 3. Therefore Nr (x) contains no points of E. we see that E is a ﬁnite union of closed sets.24) and E is closed (theorem 2.14.01. and [x. y ∈ Q is both open and closed in Q.lemma 2: For any real number x ∈ Q. This shows that x is an element of [0. Then Nr (x) = (. y]. so E is open (theorem 2. 7}.(−∞. E is not compact: Drawing from the example in exercise 2. To show that E is closed. − 2) √ Ai = 1 2+ i. then the k + 1st decimal place will be either 9.24 we know that E is both open and closed. the set must be closed. Deﬁne an open i cover of E to be G = {Ai : i ∈ N} This has no ﬁnite subcover for the same reason that the set in exercise 2. Then the neighborhood Nr (x) does not contain any points in E: i) If ak+1 = 0. which proves that E is not dense in [0. y).14. But E is also a ﬁnite union of open sets (lemmas 1 and 2).14 has no inﬁnite subcover: given any √ ﬁnite collection of elements of G. E is not dense: Let x = . E ⊂ [0. So by theorem 2. 1]. ∞). 22 . We will show that x is not a limit point. so E is closed (theorem 2. we know by lemma 4 that each Ai is an open set. x)) are also closed. the intervals (−∞. we need to show that every limit point of E is a member of E. 1] such that x ∈ E. 1]. y) with x. x] ∪ [y.660. This proves that (x.. 3). √ √ √ E is bounded: If |p| ≥ ± 3. we can always ﬁnd an element suﬃciently close to 2 that is not contained in any of those elements of G. x + r) and so no element of Nr (x) is a member of E. By the deﬁnition of membership in E. Exercise 2. ∞). deﬁne the interval Ai to be: √ √ if i = 1 (− 3.17 E is not countable: This is proven directly by theorem 2. Choose r = 10−(k+1) . or 1 for every element of (x − r. We know that ± 2 and ± 3 are not rational. 3). E is open and closed in Q: E is√ set of all √ √ numbers p such that 2 < p2 √ 3. and (−∞. the fact that x ∈ E means that ∞ ai x= 10i i=1 where some ai ∈ {4.(x. 7}.. we can prove that E is compact by showing that it is bounded and closed.41.67). Let k represent some index such that ak ∈ {4. 0. 1] that is neither a point in E nor a limit point of E. E is compact: By theorem 2. Proof: Choose any of these four open sets. Since its complement is open.65. y) is both open and closed.23). so by lemma 4 we know that E is both a union of closed sets and also a union of open sets. (x. It’s complement is E = (−∞. and let r = . − 2) ∪ ( 2. y]. Proof: the proof is nearly identical to that of lemma 1. and every real number in this interval contains a number other than 4 or 7. y) are identical to this one. x] and (−∞. 3 if i > 1 Because 2 + 1 is irrational for every i ∈ N. Proof: choose an arbitrary interval E = (x. so E is bounded.

27 we know that A = A and B = B. E is perfect: We have already shown that E is closed in the course of proving that E is compact. so we have proven that every neighborhood of every point in E contains a second point in E: by deﬁnition. we see that xk diﬀers from x only in the kth decimal place. This assumption leads to a contradiction: 23 . such that ∞ bi = ai if i = k bi bk = 4 if ak = 7 xk = . Exercise 2. so each point in the Cantor set 3k is rational. k ∈ N. Exercise 2. and d(x. Deﬁne a second number. so A ∩ B = ∅. this proves that every limit point of E is a member of E. where 10i i=1 bk = 7 if ak = 4 From this deﬁnition. Exercise 2. The Cantor set was bounded by [0. To prove that E is perfect. Each point in the n Cantor set is an endpoint of some segment of the form 3k . then by theorem 2. This proves that x is not a limit point of E. We can use this information to show that x is a limit point of E. so this proves that (x ∈ E → x is not a limit point of E).19 b Let A and B be disjoint open sets. we need to show that every point in E is a limit point of E. Each element in E is irrational (exercise 1. or 0 for every element of (x − r.44 and also in this document after exercise 2. And this was an arbitrary neighborhood of an arbitrary point in E. 9.1). x + r) and so no element of Nr (x) is a member of E. this means that A and B are separated. The proof that E is perfect is identical to the proof that the Cantor set is perfect (given in the book in section 2. So we conclude that A ∩ B = A ∩ B = A ∩ B = ∅. xk .4). we’re shifting every element of the Cantor set 2 units to the right). this means that every point in E is a limit point. we know that ∞ x= i=1 ai 10i where each ai is either 4 or 7.ii) If ak+1 = 9. The Cantor set is a nonempty perfect subset of R1 . means that E is closed. And this is algebraically r equivalent to ﬁnding some integer k such that 3 × 10−k < r. Assume that A ∩ B is not empty. iii) If ak+1 is neither 0 nor 9. we can ﬁnd some integer k such that k > log1 0 3 . by deﬁnition. By the deﬁnition of membership in E. xk ) = 3 × 10−k .18 Section 2. then the kth decimal place will be be ak for every element of (x − r.44 describes the Cantor set. This means that we can ﬁnd some xk ∈ E (as deﬁned above) in Nr (x). x + r) and so no element of Nr (x) is a member of E. And if A and B are closed. 1 + 2]. But x was an arbitrary point such that x ∈ E. And this. Let x be an arbitrary point in E. This proves that there is some neighborhood Nr (x) that contains no points in E. By deﬁnition. n+1 with n. Let Nr (x) be any arbitrary neighborhood of x. √ √ Let E be the set {x + 2 : x is in the Cantor set} (that is. By contrapositive. From the archimedian principle.19 a We’re told that A and B are disjoint. then the k + 1st decimal place will be either 8. 1] so E √ √ is clearly bounded by [ 2.

y) + d(y. it’s true for any particular R. in fact. empty. so there cannot be any possible choice of variables such that y ∈ B and y ∈ A. we can always ﬁnd some distance δ such that there are no elements x.11) A and B are disjoint deﬁnition of set intersection A is an open set deﬁnition of B deﬁnition of interior point deﬁnition of limit point If something is true for all s ∈ R. Let y be an arbitrary element in Nr (x). y) + d(y. p) < δ. so this proves that every x ∈ B is an interior point of B. x) − δ.19(b) we know that A and B are separate. If we swap the roles of A and B. x) − d(p. x)) + (d(p. x) − δ > d(x. The sets A and B are disjoint. x) − δ − d(x. x) ≥ 0 ∧ d(p.19. x) − d(p. y) > 0 → (d(p. x is an interior point of B. y ∈ X with d(x. Exercise 2. → (∃x) [(∃r ∈ R)( (Nr (x) ⊆ A) ∧ (∃y ∈ B)(y ∈ Nr (x)) )] → (∃x) [(∃r ∈ R)(∃y ∈ B)(y ∈ Nr (x) ∧ Nr (x) ⊆ A)] → (∃x) [(∃r ∈ R)(∃y ∈ B)(y ∈ A)] substitution of s = r rearrangement of terms for clarity deﬁnition of subset This last step establishes a contradiction. x) ≥ 0 ∧ r > d(x. x) − d(p.15(c) of metric spaces we know d(x. y) < r because y ∈ Nr (x) deﬁnition of r Our choice for y was arbitrary. y)) > 0 → d(p. this shows that B is open in X. x) ≥ 0 ∧ d(p. A and B are separated. so by exercise 2. x) ≥ d(p.19 c A is open in X: The set A is. x) − δ − d(x. So choose s = r. x) − d(p. then. By deﬁnition. Proof that y ∈ B: → d(p. y) = δ (proof follows). p) > δ and d(x. x) → d(p. But our choice of x ∈ B was also arbitrary. y) + d(y. by deﬁnition 2. so every point in this neighborhood of x is a member of B: by deﬁnition. then d(x. B is open in X: Let x be an arbitrary point in B. y) − δ > 0 → d(p. Let r = d(p. This allows us to choose 24 . Exercise 2. A is an open subset of X. A and B are disjoint: If there were some x ∈ A ∩ B.19 d If we are given any metric space X with a countable or ﬁnite number of elements. x) − d(p. then. By theorem 2. y) + d(y. Our assumption must have been incorrect: A ∩ B is. y) → d(p. y) + d(y. then. which violates the trichotomy rule for order relations. y) > δ →y∈B deﬁnition of membership in B the sum of positive numbers is positive cancellation of like terms Property 2.18(a). a neighborhood of p. this same proof also shows us that A∩ B is empty. By deﬁnition. y) + d(y.→ A ∩ B is not empty → (∃x)(x ∈ A ∩ B) → (∃x)(x ∈ A ∩ (B ∪ B )) → (∃x)(x ∈ (A ∩ B) ∪ (A ∩ B )) → (∃x)(x ∈ ∅ ∪ (A ∩ B )) → (∃x)(x ∈ A ∩ B ) → (∃x)(x ∈ A ∧ x ∈ B ) → (∃x)(x is an interior point of A ∧ x ∈ B ) → (∃x)(x is an interior point of A ∧ x is a limit point of B) → (∃x) [(∃r ∈ R)(Nr (x) ⊆ A) ∧ (x is a limit point of B)] → (∃x) [(∃r ∈ R)(Nr (x) ⊆ A) ∧ (∀s ∈ R)(∃y ∈ B)(y ∈ Ns (x))] hypothesis to be contradicted deﬁnition of non-emptiness deﬁnition of closure distributivity (section 2. y) → d(p. x) ≥ 0 → d(p. A and B are separate: We’ve shown that A and B are disjoint open sets.

And there is a clear one-to-one correspondence between this set and the set {(d(ai . . so our partitions A and B will both be nonempty. By contrapositive. And we know that d(x. aj ) : ai . d(x. 2 ).20 b Consider the segment (0. Exercise 2. then. We can avoid this problem (as long as X has at least two elements) by picking arbitrary x. For instance. This interval is still uncountable. so these two closed sets are connected. which are of course uncountable. for all x. we can choose a real number δ that is not in this set (otherwise we would have an at most countable set with R as a proper subset).20 a Closures of connected sets are always connected. d(x. x) < δ. 0) : 0 < x < 1} contains no interior points while every point in a neighborhood is an interior point (theorem 2. If A and B are connected. an .21 a The function p(t) can be thought of as the parameterization of a straight line connecting the points a and b. 0) contains the point (x. aj ∈ X}: so the set of all distances between all combinations of points in X is at most countable. So let r = 1 and let E be the set 4 Nr (0. Distances in metric spaces are always real numbers (deﬁnition 2. it contains no interior r points in R2 since every neighborhood Nr (x. Although this segment is open in R1 . Exercise 2.13. then there is either some p in A ∩ B or some q in A ∩ B. But this isn’t quite enough: if δ is so large that there are no elements in the set {x ∈ X : d(x. y) > δ}. so we are still able to choose a δ that is not in this set. y) = δ: Let X be an at most countable metric space with elements a1 . aj ∈ X} is at most countable. Therefore A ∩ B is nonempty. This proves that (X is at most countable → X is not connected). Exercise 2. y ∈ X. Proof that an at most countable metric space X has some distance δ such that. 0) ∪ {(x. consider the following sets in R2 : 25 . 0) : 0 < x < 1} ∪ Nr (1. this proves that (X is connected → X is uncountable). aj ∈ X}. and it is trivial to show that this is the union of two nonempty separated sets. a2 .some arbitrary p ∈ X and then use the results from part (c) to completely partition X into separated sets A = {x ∈ X : d(x. 1). Clearly.15). y ∈ X and then choosing delta from the interval (0.19). p) > δ} and B = {x ∈ X : d(x. aj ) : ai . . Since the line segment {(x. Then from theorem 2. we have either either p ∈ (A ∪ A) ∩ B or q ∈ A ∩ (B ∪ B). p) < δ}. 0). aj ) : ai . 1) of the real line. we know that the set of all order pairs {(ai . . the interior of E is just Nr (0. which is what we were asked to prove. y)). then one of our partitions will be empty. 0) ∪ Nr (1. y) > δ and d(x. Because there are an at-most countable number of distances in the set {(d(ai .

1) such that α ∈ A0 ∪ B0 . then A and B are not separated.The set A0 is the set of all t such that p(t) ∈ A. And since p(tA ) is in A and each p(tB ) is in B. so the proof for A0 ∩ B0 = ∅ is almost identical). and the set B0 is the set of all t such that p(t) ∈ B. p(tB )) < |r(b − a)|. because this would imply p(tA ) ∈ A ∩ B which is impossible because A and B are separated sets. x) : x ∈ B0 }. So for any arbitrarily small r. t+r)d(a. t + r) and d(p(t). That is. By contrapositive. We’ve shown that tA is a limit point of B0 . so the set E has a greatest lower bound. E is the set of all distances between the point 0 (which is in A0 . By deﬁnition 2. So it must be the case that tA ∈ A0 ∩ B0 . And this means that for any arbitrarily small r. The distance between p(t) and p(t + r) is the vector norm of p(t + r) − p(t): d(p(t). It can’t be the case that tA ∈ A0 ∩ B0 . there is some p(tB ) ∈ B such that d(p(tA ). this is equivalent to ﬁnding α ∈ (0. And this is what we were asked to prove. This shows that if A0 and B0 are not separated. b) • p(tA ) is an element of A and a limit point of B. 1) such that p(α) ∈ A ∪ B: from the deﬁnition of the function p. Proof that such a α exists: Let E be deﬁned as E = {d(0. Exercise 2. Let tA be one element of A0 ∩ B0: then either tA ∈ A0 ∩ B0 or tA ∈ B0 . if A and B are separated then A0 and B0 are separated. 26 . this means that there is an element of A that is a limit point of B: So A ∩ B is nonempty. there is some tB ∈ B0 such that d(tA . which means that A and B are not separated. Proof by contrapositive: • There is some element tA that is an element of A0 and a limit point of B0 . p(t + r)). Assume that A0 and B0 are not separated. Assume that A0 ∩ B0 is nonempty (the sets are interchangeable. We’re told that A and B are separated and are asked to prove that A0 and B0 are separated. p(t+r)) = |p(t+r)−p(t)| = |[(1−(t+r))a+(t+r)b]−[(1−t)a+tb]| = |r(b−a)| = |r||(b−a)| = d(t.21 b We are asked to ﬁnd α ∈ (0. then. • There is a proportional relationship between d(t. Let α represent this greatest lower bound. either A0 ∩ B0 or A0 ∩ B0 is nonempty. tB ) < r. The set R has the greatest lower bound property and E is a subset of R with a lower bound of 0. since p(0) = a) and elements of B0 .45.

. We know that α = 0. For each ei ∈ E. and therefore {Vα } has 27 . E = A ∪ B). Exercise 2. But if the interval (0. then their union (0. α). Let r b = (b1 . So there is some neighborhood Nr (α) that contains no points in A0 . And for each p we see that d(0. 0) ∈ E which means that 0 ∈ B0 . . Now consider the points p in the range (α − r. And this would mean that E is not convex. . b) = (a1 − bi )2 + . b2 . if E is convex then E is connected. we have p ∈ A0 ∪ B0 . we know that X contains a countable dense subset E. Each p is in the neighborhood Nr (α). b. . We’ve shown in part (a) that A0 and B0 are separated. So our initial assumption must have been incorrect: it cannot have been the case that α ∈ A0 . And. . .e. or α ∈ A0 ∪ B0 : • if α ∈ B0 : We know that α is not a limit point of A0 (otherwise. so it can’t be the case that α is a limit point of B0 (otherwise A0 ∩ B0 would not be empty). ak ) be an arbitrary point in Rk and let Nr (a) be an arbitrary neighborhood of a. To prove that Qk is dense in Rk . we need to show that every point in Rk is a limit point of Qk : Let a = (a1 . We now need to show that we can use α to construct elements that are not in A0 ∪ B0 . α + r) contains no points of B0 . From part (b). xα ∈ B0 ∩ A0 which contradicts the fact from part (a) that B0 and A0 are separated). From the deﬁnition of separable in the previous exercise. which completes the proof that Rk is separable. We know that either α ∈ B0 . .22 The metric space Rk clearly contains Qk as a subset. This is a countable collection of countable sets. . Assume that E is not connected: then E could be described as the union of two separated sets (i. i ∈ N} be the collection of all neighborhoods with rational radius centered around members of E. + (ak − bk )2 < r2 r2 + . Exercise 2.. we clearly have α ∈ A0 ∪ B0 . . because 1 ∈ B0 and B0 is an open set: so 1 is an interior point of B0 . and d(a. So we see that for every p ∈ (α − r. • if α ∈ A0 and α ∈ B0 : Under this assumption.20(b)).• 0 < α < 1: We know that α > 0 because α is a distance. which is false. By contrapositive. And we know that α < 1. + = k k kr2 =r k This shows that every point in Rk is a limit point of Qk . We know that Qk is countable from theorem 2. . bk ) where bi is chosen to be a rational number such that ai < bi < ai + √k (possible via theorem 1. α + r) contains no points of B0 . α). And this is what we were asked to demonstrate. Whatever assumption we make about the set containing α. we could then choose a. p) < α. Let {Vα } = {Nq (ei ) : q ∈ Q. we see that there will always be at least one element α such that 0 < α < 1 and α ∈ A0 ∪ B0 . because this would mean that d(0. so they aren’t members of A0 . But this contradicts our deﬁnition of α as the greatest lower bound of E. from the deﬁnition of the function p. and t such that (1 − t)a + (t)b ∈ A ∪ B = E. The point b is clearly in Qk . • if α ∈ A0 : This assumption leads to a contradiction.α ∈ A0 . a2 ..21 c Proof by contrapositive.23 Let X be a separable metric space. which means that there is some small r such that 1 − r ∈ B0 . so they aren’t members of B0 (otherwise α wouldn’t have been a lower bound of E).13. let Nq (ei ) be a neighborhood with rational radius q around point ei . this means that p(α) ∈ A ∪ B. α) contains no points of B0 and the interval (α − r. So there is some neighborhood Nr (α) that contains no points of B0 . Exercise 2.

But r we deﬁned q so that 0 < q < 2 . And Nq (e) ∈ {Vα }. e) + d(e. We started by choosing an arbitrary element x in an arbitrary open set G ⊆ X. 28 ∞ . Let y be any point in Nq (e). We’ve shown that the neighborhoods of En cover X. then it would have a limit point. xi ) > δ for each xi ∈ {xi }. we can 1 ﬁnd an integer n such that n < r. From the archimedian principle. and proved that there was an element Nq (e) ∈ {Vα } such that x ∈ Nq (e) ⊂ G. e) < n and so e ∈ N1/n (x). If it weren’t. this also means that d(x.countably many elements. and let G be an arbitrary open set in X such that x ∈ G. this implies that e ∈ Nr (x). so by transitivity we know that Nq (e) ⊆ G. So e ∈ Nq (x). which by deﬁnition means that E is dense in X. y) = 2q by deﬁnition 2. we know that x is an interior point of G. i We will show that E is a countable dense subset of X.15(c) of metric spaces. • The set {xi } must be ﬁnite: if this set were inﬁnite. Proof that E is dense in X: Choose an arbitrary x ∈ X. on the other hand. and each element is an open neighborhood. by deﬁnition. which contradicts our assumption that d(xi . Because G is open. then there would be multiple points of {xi } in some neighborhood of radius δ/2. so clearly E ⊂ X. Let Ei represent the set of {xi } constructed in the above way when δ = 1 . And we chose r so that Nr (x) ⊆ G. so that we could have chosen x to be an additional element in {xi }. so E is countable. We’ve shown that {Vα } has a countable number of elements. every neighborhood of x contains some e ∈ E. And since n < r. So d(x.24 We’re told that X is a metric space in which every inﬁnite subset has a limit point. This proves that every x is a limit point of E. So we’ve chosen an arbitrary x and an arbitrary radius r. we need to prove that Nq (e) ⊆ G. y) < q. or Nq (e) ⊆ Nr (x). y) < d(x. so we know that d(x. x) < q. so that x ∈ Nq (x) ⊆ Nr (x) ⊆ G. this would imply that d(x. Choose an arbitrarily small radius r. Let x be an arbitrary point in X. Having shown that x ∈ Nq (e). This union E is a countable union of nonempty ﬁnite sets. So there is some neighborhood Nr (x) such that Nr (x) ⊆ G. • The neighborhoods of {xi } form a cover of X: Every x ∈ X must be contained in Nδ (xi ) for some xi ∈ {xi }. But. e) < q so that x ∈ Nq (e). And if it has a limit point. This proves that {Vα } is a base for X. Because E is dense in x. Exercise 2. Each element of E was chosen from X. which means that d(e. This means that every y ∈ Nq (e) → y ∈ Nr (x). then d(x. e) < q and d(e. xj ) > δ for each pair of points in {xi }. We know that d(x. so there is some e ∈ En 1 1 1 such that x ∈ N1/n (e). and shown that Nr (x) will always contain some e ∈ En ⊂ E. x) < n . and consider the set E = i=1 Ei . But if d(e. Choose some arbitrary δ and construct a set {xi } as described in the exercise. But r we can choose a rational q such that 0 < q < 2 . y) < r.

e. Every neighborhood Nr (x) contains inﬁnitely many points of E (theorem 2. Consider the set En = {N1/n (k) : k ∈ K}. Construct Fn and E as described in the exercise. Because G is open and x ∈ G. An alternate proof follows. x must be an interior point of G. this means that K is separable. And K is compact.25(b). by virtue of being a compact space. Now choose an integer m such that m < 2 . y) ≤ 2/m ≤ r (since we chose m < 2 ). The question wants us to prove that K has a countable base and that therefore K is separable. We will prove that V is Choose an arbitrary x ∈ K. k) + d(k. This countable base V is identical to the set E constructed in exercise 24. y) < 1/m.. there is some k ∈ K and neighborhood N1/m (k) in the open cover Vm such that x ∈ N1/m (k). we 1 r know that d(x. We’ll do this with a proof by contradiction. Proof: let G be an arbitrary open set from an arbitrary open cover of X. there is some neighborhood Nr (x) such r 1 that Nr (x) ⊂ G. ∞ n=1 Vn . and an arbitrary open set G such that x ∈ G ⊂ K. has the property that every inﬁnite subset of K has a limit point. We saw there that this base was a dense subset. it has a limit point: let x be this limit point. we know that d(x. By the deﬁnition in exercise 23. y). We must now prove that a ﬁnite subset of {Vn } can always be chosen to act as a subcover for any open cover. And from the deﬁnition of metric spaces. And this shows that N1/m (k) ⊆ G. We can ﬁnd an element N1/m (k) ∈ V such that x ∈ N1/m (k) ⊆ G by the same method used in the second half of exercise 2. which means that d(x. Exercise 2.26 Let {Vn } be the countable base we constructed in exercises 24 and 25. so by the deﬁnition of “separable” in exercise 23 we know that K is separable. Because E is an inﬁnite subset. Choose any x ∈ G. {Vn } acts as a countable subcover to any open cover of X. Let {Vn } represent a ﬁnite subcover of the open cover En .25 b We’re told that K is compact. And exercise 23 proves that every separable metric spaces has a countable base. a countable collection of ﬁnite covers of K. we know that d(k. This proves that every element of N1/m (k) is in the neighborhood Nr (x). so it’s a countable base.25 a Theorem 2. this proves that K has a base. Because Vm is an open cover of K. Let {Wα } be an open cover of X and assume that there is no ﬁnite subset of {Vn } that acts as a subcover of {Wα }. since x ∈ K → x ∈ N( 1/n)(x) → x ∈ En .This proves that X contains a countable dense subset. We’ve chosen an arbitrary element x and an arbitrary element G. whereas this is a proof that K is separable and therefore has a countable base. k) < 1/m. so by the deﬁnition in exercise 22 we have proven that X is separable. I’m not sure that this an appropriate proof. Proof that this neighborhood N1/m (k) is a subset of G: Assume that y is an element of N1/m (k). From the fact that x ∈ N1/m (k). Now consider the union V = a base for K. y) ≤ d(x. Exercise 2. And this base is a countable collection of ﬁnite sets. Exercise 2. K must be covered by some ﬁnite number of neighborhoods from En . the set of neighborhoods of radius 1/n around every element in K. By exercise 24. so N1/m (k) ⊆ Nr (x) ⊆ G. And since x is an interior point.20). And from the fact that y ∈ N1/m (k). and we can prove that this arbitrary neighborhood Nr (x) must contain every point of E: 29 . so the open cover En must have some ﬁnite subcover: i.41 tells us that the set K. This is clearly an open cover of K. though. and have shown that there is some N1/m (k) ∈ V such that x ∈ N1/m (k) ⊂ G.

the fact that x ∈ Vn means that E ∩ Vn is uncountable. And from the fact that y ∈ P . This proves that P is perfect. so this proves that every neighborhood of x has uncountably many elements of E: by deﬁnition. Taking the complement of each of these. Exercise 2. so this proves that x ∈ P . Note that this proof assumed only that X had a countable base. p) ≤ d(x. And this means that we chose the same element from each Fn : i. which means that Nr (x) has uncountably many elements of E. So we have shown that every point of E ∩ P is a limit point of E ∩ P and vice-versa: this proves that E ∩ P is perfect. Choose some arbitrary r r and let s = 2 . Proof that every limit point of P is a point of P : assume that x is a limit point of P . we see that W = P c . And we’ve established that Vn ⊆ Nr (x).e. By the deﬁnition of limit point. a set with one element. every neighborhood Ns (x) contains some y ∈ P . Proof that every point of P is a limit point of P : assume that x ∈ P . We will show that W c = P : Proof that P ⊆ W c : Assume that x ∈ P . we know that every neighborhood Nr (x) contains uncountably many points of E. y) + d(y. we see that d(x. which of course means that Vn has uncountably many elements of E. it’s trivial to show that there are countably many points in E ∩ P c . But this means that E = {x}. p) so that d(x. But Nr (x) was an arbitrary neighborhood of x.. then it has no more than j − 1 points of E: i. So clearly every neighborhood of x contains at least one point of P . By deﬁnition of W c . x was in each Fn . this means that every neighborhood of x contains uncountably many points of P . So x is a limit point of E ∩ P . So x ∈ Fn . By the deﬁnition of condensation point. So the neighborhood Nr (x) contains uncountably many points of P . So if Nr (x) doesn’t contain any points from Fj .27 To prove that P is perfect. we must show that every limit point of P is a point of P . But Nr (x) was an arbitrary neighborhood of x. Proof that E ∩ P is perfect: assume that x ∈ E ∩ P . Clearly x is a point of P . and vice-versa. 30 . By the deﬁnition of P . then. Exercise 2. assume that x is a limit point of E ∩ P . And from this. So Nr (x) ∩ E contains one point from each Fn . But for each p ∈ Ns (y). which contradicts our assumption that Fn was empty. then. so it is valid for any separable metric space. p) ≤ 2s = r. Now. which is what we were asked to prove. so Nr (x) must contain inﬁnitely many points of E. Let P be the set of all condensation points for E. So if Nr (x) doesn’t contain any points from Fj .We have constructed the Fn sets in such a way that Fn+1 ⊆ Fn for each n. we let {Vn } be a countable base for X and let W be the union of all Vn for which E ∩ Vn is countable. which means that x is a limit point of P . we know that x is a condensation point of E. Nr (x) ∩ E would be ﬁnite. Fj+2 . there is some Vn ⊂ Nr (x) such that x ∈ Vn .28 Let E be a closed set in a separable metric space X. which means that x is a limit point of E (from the deﬁnition of P ) and a limit point of P (because P was shown to be perfect in exercise 27). And W = P c . By the deﬁnition of membership in P . So x is not a member of any countable Vn ∩ E. which would contradict our ﬁnding that every neighborhood of x contains E. which means that x ∈ W c . so P c is countable: this proves that countably many points are P c . To prove that P c ∩ E is countable. we know that Ns (y) contains uncountably many points of P .etc because y ∈ Fj+1 → y ∈ Fj . So we have proven that W c = P . each of which contains countably many elements: so W is countable. And this means that every open set Vn containing x contains uncountably many points of E. x is a condensation point and therefore x ∈ P . From this. Proof that W c ⊆ P : Assume that x ∈ W c . If it contained a second point. But x is a limit point of E. Choose any arbitrary neighborhood Nr (x): by the deﬁnition of the base. we know that x is a point of E (because E is closed) and a point of P (because P is perfect).e. And the set W is the union of countably many sets of the form E ∩ Vn . we could ﬁnd a neighborhood Nr (x) that failed to contain this second point. then it doesn’t contain any points from Fj+1 .

but there are two reasons to assume that we are. sn ) < .29 Let E be an arbitrary open set in R1 and let {Gi } be an arbitrary collection of disjoint segments such that Gi = E. so for this same and N we see that d(|s|. |sn |) < . we see c that Fn is nonempty. Each Gi is open and nonempty. So we’ll assume that we’re operating in Rk for this exercise.9d → → → → → ◦ Fn = ∅ ◦ c ( Fn ) = ◦ c (Fn ) = c Fn = R c (∀n)(Fn = R R R) c c This last step tells us that Fn is dense in R for every n. and Rk is the only metric space we’ve encountered so far for which absolute value has been deﬁned. Therefore there can’t be more Gi elements than rational numbers. But it’s always the case that d(|s|. And Fn was closed. Exercise 2. If we let sk = (−1)k . we are supposed to understand the meaning of absolute value in this metric space. We’re told that the sequence {sn } converges to some value s: that is. |sn |) ≤ d(s.13). Appealing to the proof of Baire’s theorem in exercise 3. which means that there are at most a countable number of elements in {Gi }. By transitivity.22 taking the complement of both sides And this contradicts our original claim that Fn = R so one of our initial assumptions must be wrong.22 (which uses only terms and concepts introduced in chapter 2). Exercise 3. But {Gi } was an arbitrary set of disjoint segements whose union is E. so Fn is open. The converse is not true. so by contradiction we have proven that at least one Fn must have a nonempty interior. And our only assumption was that each Fn has an empty interior. sn ) < . And this is what we were asked to prove. Exercise 2. Therefore: → → c Fn = ∅ c → ( Fn ) = ∅ Fn = R theorem 2. the sequence {|sn |} clearly converges while the sequence {sn } clearly does not. for any arbitrarily small there is some integer N such that n > N implies d(s. ◦ → (∀n)(Fn = ∅) Fn = R and suppose that each Fn has an hypothesis of contradiction taking the complement of both sides theorem 2. First. E is the union of a perfect set and a countable set. 31 . this means that the sequence {|sn |} converges to |s|. sn ) (exercise 1. By deﬁnition of convergence.22 exercise 2. Rudin appears to use sn to represent series in Rk and pn to represent series in arbitrary metric spaces. so each Gi contains at least one rational point.Proof that E ∩ P c is at most countable: this was proven in exercise 27. And E = (E ∩ P ) ∪ (E ∩ P c ): that is.30 Proof by contradiction: Let {Fn } be a collection of sets such that empty interior. |sn |) ≤ d(s. therefore E cannot be the union of an uncountable number of disjoint segments.1 The exercise does not explicitly say that we’re operating in the metric space Rk . Second. this means that for any choice of there is some integer N such that n > N implies d(|s|.

14. so our initial assumption must be false: this shows that ( n2 + n − n) is 1 bounded above by 2 . by contradiction. So. The sequence is monotonically increasing → sn+1 > sn √ ↔ (n + 1)2 + (n + 1) − (n + 1) > n2 + n − n √ ↔ (n + 1)2 + (n + 1) − n2 + n > (n + 1) − n √ √ ↔ n2 + 3n + 2 − n2 + n > 1 √ √ ↔ (n2 + 3n + 2) + (n2 + n) − 2 n2 + 3n + 2 n2 + n > 1 √ ↔ (2n2 + 4n + 2) − 2 n4 + 4n3 + 5n2 + 2n > 1 √ ↔ −2 n4 + 4n3 + 5n2 + 2n > 1 − (2n2 + 4n + 2) √ ↔ 2 n4 + 4n3 + 5n2 + 2n < 2n2 + 4n + 1 ↔ 4n + 16n + 20n + 8n < 4n + 16n + 20n + 8n + 1 ↔0<1 4 3 2 4 3 2 deﬁnition of sn rearrange terms some algebra square both sides simplify rearrange terms multiply both sides by −1 and simplify square both sides subtract 4n4 + 16n3 + 20n2 + 8n from each side 1 The sequence has a least upper bound of 2 1 We’ve already shown that 2 is an upper bound. we can 2 use theorem 3. we can always ﬁnd some n ∈ N greater than any 1 speciﬁc quantity. 2 ): this means that the upper bound 1 1 1 cannot be less than 2 . And it’s bounded above by 1 : 2 → → √ √ n2 + n − n ≥ n2 + n ≥ 1 4 1 4 1 2 1 2 hypothesis to be contradicted both sides are positive. The sequence is bounded √ √ √ The quantitity ( n2 + n − n) is bounded below: from the fact that n2 + n > n2 = n.14 tells us can be done by proving that the sequence is bounded and monotonically increasing. 32 .Exercise 3. We can then ﬁnd the limit by ﬁnding the least upper bound of the sequence.2 The exercise asks us to calculate the limit rather than prove it rigorously. 1 ) and assume that it is an upper bound for {sn }. choose any x ∈ [0. If we need to prove it more rigorously. 2 → (∀n ∈ N)(sn ≤ x) hypothesis to be contradicted √ → (∀n ∈ N)( n2 + n − n ≤ x) deﬁnition of sn √ → (∀n ∈ N)( n2 + n ≤ x + n) → (∀n ∈ N)(n2 + n ≤ x2 + 2xn + n2 ) → (∀n ∈ N)(0 ≤ x + (2x − 1)n) → (∀n ∈ N)(−x2 ≤ (2x − 1)n) −x → (∀n ∈ N)( 2x−1 ≤ n) 2 square both sides subtract n2 + n from both sides subtract x2 from both sides 2x − 1 < 0 when x < 1 2 2 This last step must be false: by the Archimedian principle. so we see that 2 is the least such upper bound. But 2 is an upper bound. we can ﬁnd sn > x for every x ∈ [0. To show that it is the least such upper bound. we know that √ n2 + n − n > 0. We will ﬁrst need to prove that this sequence has a limit. so the sign doesn’t change subtract n + n2 from both sides +n → n +n≥ →0≥ 2 + n + n2 √ This last step is clearly false. which theorem 3. so we might be able to manipulate the expression algebraically: n2 +n−n=( n2 + n − n) √ n2 + n + n √ n2 + n + n =√ n n = = +n+n n 1 + 1/n + n 1 1 + 1/n + 1 n2 and then use basic calc 2 techniques to show that this limit is 1 .

. we can prove the stronger result that √ √ it has an upper bound of 2 + 2.3 Note that we are not asked to ﬁnd the limit of this sequence. So the limit of s2m+1 is given 2 by ∞ 1 1 1 1 1 lim s2m+1 = + + + . we can see that s2 = 0 and s2m = theorem 3. We can prove this by induction. 1 2: Exercise 3. 1 2 or arbitrarily close to 1. we can prove convergence by proving that the sequence is bounded and monotonically increasing. . sn < hypothesis of induction 2+ √ 2 for all n. Exercise 3. and let s∗ = limn→∞ sup(an + bn ). = 4 8 16 1 2n n=0 ∞ 1 4 + 1 s2m−2 . Now assume that 0 < sn < 2: → 0 < sn < 2 √ √ → 0 < sn < 2 √ √ → 0 < 2 + sn < 2 + 2 √ √ → 2 + sn < 2 + 2 √ → sn+1 < 2 + 2 By induction. {bk } such that limj→∞ aj + limk→∞ bk = 33 . so Exercise 3. .14. If a∗ and b∗ are both ﬁnite: From theorem 3. By theorem 3.We have shown that {sn } is a monotonically increasing bounded sequence with a least upper bound of 1 by theorem 3. The sequence is monotonically increasing √ √ Proof by induction. this proves that the limit of {sn } is 2 .. n.14. then.26 to ﬁnd the limit of s2m . We see that 0 < s1 = 2 < 2. Now. We are only asked to show that it converges and that 2 is an upper bound. this is suﬃcient to prove that it converges. By theorem 3. lim s2m = 1 1 1 + + + .. let b∗ = limn→∞ sup bn . √ → sn = 2 + sn−1 deﬁnition of sn √ → sn > 2 + sn−2 our hypothesis of induction tells us sn−1 > sn−2 → sn > sn−1 deﬁnition of sn−1 We’ve shown that sn is a monotonically increasing function that is bounded above. So we can use 2 1 1 = 2 2 m→∞ −1− 1 1 = 2 1− 1 2 −1− 1 From the same recursive deﬁnition. assume that we have proven si−1 < si for i = 1 . the terms of {sn } are either arbitrarily close to these are our upper and lower bounds for the sequence.14. we see that s1 = 0 and s2m+1 = 1 + 2 s2m−1 .5 Let a∗ = limn→∞ sup an . = −1= 1 −1=1 m→∞ 2 4 8 2n 1− 2 n=0 This shows that as n increases. . We can immediately see that s1 = 2 < 2 + 2 = s2 .4 From the recursive deﬁnition we are given. √ The sequence is bounded above by 2 + 2 Although we’re asked to show that the sequence has an upper bound of 2.17(a) we know that there are some subsequences {aj }.

so s∗ = lim aj + lim bk ≤ lim sup an + lim sup bn = a∗ + b∗ j→∞ k→∞ n→∞ n→∞ By transitivity. It’s also not true that |1 + z n | > |z n |. so the terms of {an } are smaller than the terms We know that the series n of a convergent series. mainly because for complex z it’s not true that 1 + z n > z n – in fact. 34 . Exercise 3. Exercise 3.25. In all other cases when one or both of these values is inﬁnite. n the < 2n can easily be proven via induction take the nth root of each term subtract 1 from each term x converges when 0 ≤ x < 1. so by the comparison theorem 3.25 we know that Exercise 3. By the comparison theorem 3. the inequality can be easily shown to resolve to ∞ = ∞ or −∞ = −∞.28. by the deﬁnitions of these terms. this tells us that an is convergent.6a an = √ n+1− √ n= √ n+1− √ n √ √ n+1+ n √ √ n+1+ n =√ 1 √ n+1+ n We can compare this last term to a known series: an = √ 1/2 1 1 1 > √ > √ 2 n+1+ n 2 n+1 1 n 1/2 1 We know that the series diverges by theorem 3. this shows that s∗ ≤ a∗ + b∗ : and.8). so the terms of {an } are larger than the terms of n a divergent series. Exercise 3.6d This probably is more diﬃcult than it looks at ﬁrst. without absolute value signs the inequality z1 > z2 has no meaning whatsoever (see exercise 1. The limit of aj must be less than or equal to the supremum a∗ and the limit of bj must be less than or equal to the supremum b∗ . this tells us that an is divergent. The best we can do is appeal to the triangle inequality.25.s∗ (that is.6b an = √ n+1− n √ n √ = n+1− n √ n √ √ n+1+ n √ √ n+1+ n =√ 1 n3 + n2 + √ n2 We can compare this last term to a known series: √ 3/2 1 n3 + n2 + √ n3 <√ 1 n3 + √ n3 = 1 2 1 n 3/2 1 converges by theorem 3.6c √ For n ≥ 1.28. By the comparison theorem 3. this proves that n→∞ ∗ ∗ lim sup(an + bn ) ≤ lim sup an + lim sup bn n→∞ n→∞ If either a or b is inﬁnite: We’re asked to discount the possibility that a∗ = ∞ and b∗ = −∞. the supremum of E as deﬁned above is also member of E: E is a closed set). we can show that 0 ≤ ( n n − 1) < 1: → (∀n ∈ N)(1 ≤ n < 2n ) √ → (∀n ∈ N)(1 ≤ n n < 2) √ → (∀n ∈ N)(0 ≤ n n − 1 < 1) We know that the series an is convergent.

28 we know that that 1/n2 converges. Let α be the upper bound of {|bn |}. We’re also told that for any arbitrarily small .If |z| < 1 then by the triangle inequality we have lim 1 1 ≥ lim =1 n n→∞ |1| + |z n | 1+z n→∞ and therefore by theorem 3. the sequence is monotonically increasing: √ From the deﬁnition of partial sums.23 the series doesn’t converge.7 Deﬁne the partial sum tn to be tn = √ k=1 n √ ak k We can show that the series an /n converges by showing that the sequence {tn } converges. If |z| = 1 then we similarly have lim 1 1 1 ≥ lim = n→∞ |1| + |z n | 1 + zn 2 n→∞ and by theorem 3. By assuming that an is convergent and that an > 0. And we’re told an > 0 for every n. we know that tn = an /n + tn−1 for all n. Exercise 3. I’m not totally satisﬁed with this proof because I can’t totally justify the claim that limn→∞ z −n = 0 without appealing to the polar representation z −n = r−n e−inθ and the fact that |eixθ | = 1 for all x. and so the√ right-hand side of the above inequality √ √ an /n is bounded by α.14). we can ﬁnd an integer N such that n an converges: so ak ≤ k=m α for all n. and from theorem 3.8 We’re told that {bn } is bounded. This shows us that the series bounded by α. we’ve shown that tn is bounded and monotonically increasing.23 the series again fails to converge. m such that n ≥ m ≥ N which is algebraically equivalent to n ak α ≤ k=m for all n. so tn > tn−1 for every n. Therefore (by theorem 3. If |z| > 1 then we can use the ratio test: lim z −n 1 + z n z −n + 1 1 + zn 1 = lim −n = lim −n = <1 n+1 n+1 n→∞ z n→∞ z 1+z 1+z +z z n→∞ and by theorem 3. Exercise 3. m such that n ≥ m ≥ N 35 .14 this is suﬃcient to show that tn converges. By theorem 3. and therefore the sequence {tn } is converges to √ α. And this is what we were asked to prove. Applying this to the given series.34 the series converges. we can do this by showing that {tn } is bounded and monotonically increasing (theorem 3. the sequence is bounded: From the Cauchy-Schwartz inequality.50) we know that their product converges to some α. We are told that an converges. we know that (ab) ≤ we see that √ 1 an an ≤ n a2 1 n2 b2 .

10 We’re told that inﬁnitely many of the coeﬃcients of an z n are positive integers. we know the limit of this last term: √ (lim sup n n)3 1 = 3 3 So α = 1/3.20c.37. which gives us a radius of convergence of 3.9a α = lim sup So the radius of convergence is R = 1/α = 1. α = lim sup n Exercise 3. we look at theorem 3.9c α = lim sup n 2 2n 2 √ = lim sup √ = n n2 (lim sup n)2 2 n The last step of this equality is justiﬁed by theorem 3. this is suﬃcient to prove that Exercise 3.22. n |n3 | = lim sup |n3/n | = |n0 | = 1 Exercise 3. means that n n ak bk ≤ k=m k=m ak α ≤ for all n. so that 1 1 R= ≥ =∞ α β Exercise 3. By theorem 3.3c. m such that n ≥ m ≥ N an bn converges. which tells us that α ≤ β. since |bk | ≤ α for every k.40(b) mentions that “the ratio test is easier to apply than the root test”. how it might relate to the ratio test for series convergence. According to some course handouts I found online 2 . although we’re never explicitly told what the ratio test is. From theorem 3. we know the limit of this last term: 2 2 √ 2 = 1 (lim sup n) So α = 2. √ n3 (lim sup n n)3 = 3 3 The last step of this equality is justiﬁed by theorem 3. That is. Therefore lim supn→∞ |an | ≥ 1 → lim supn→∞ → 1 R n |an | ≥ 1 from the fact thatk > 1 → k 1/n > 1 theorem 3. we can’t simply apply the ratio test in the same way we use the root test. or how to apply the ratio test to example (b). 2 http://math. This means that for every N .9d √ n |n3 3n | = lim sup Exercise 3.berkeley. we can’t use the fact that β = lim sup 2 2n+1 n! = lim sup =0 (n + 1)! 2n n+1 to assume that the radius of convergence is 1/β = ∞.20c. From theorem 3.39 ≥1 →1≥R This ﬁnal step indicates that the radius of convergence is at most 1.3c. there is some n > N such that an ≥ 1.9b Example 3. which gives us a radius of convergence of 1/2. Instead.edu/∼gbergman/ 36 .which.

it is necessary that lim n→∞ 1 an 1 =0 +1 which can only happen if the limit of the denominator is ∞. From the assumption that an /(1 + an ) converges. that n ak < 2 k=m for every n ≥ m ≥ N there is some N that makes the last an . so we have shown that for every statement true: and this is the deﬁnition of convergence for the series an We’ve shown that the convergence of 1+an implies the convergence of an that an diverges is proof that 1+an diverges. Together. we it’s helpful to recognize that there is some integer N1 such that an < 1 for all n > N1 . N }. we can multiply the numerator and denominator of each term of the 1 an 1 +1 For this to converge. + an+1 = sn + an+1 37 . It does. we can see that sn+1 = a1 + . Exercise 3. the fact And our choice of was arbitrary. For this proof. By contrapositive. we know that for every n there is some N such that k=m ak < 1 + ak for every n ≥ m ≥ N Let N be the larger of {N1 . we can produce two inequalities: n k=m n ak < 1+1 n k=m ak 1 + ak k=m ak < 1 + ak for every n ≥ m ≥ N The ﬁrst is true because each k is larger than N1 (so that ak < 1). it must be the case that lim an = 0 n→∞ This alone is not suﬃcient to prove that an converges. an . For any and any n ≥ m ≥ N . equivalently. That is. .11b From the deﬁnition of sn . however. and this can only happen if the limit of 1/an is ∞. these inequalities tell us that n k=m ak < 2 for every n ≥ m ≥ N or. allow us to make the terms of an as arbitrarily small as we like. To simplify the form of series by 1/an to get an /(1 + an ) converges and show that this implies that an an /(1 + an ). .11a Proof by contrapositive: we will assume that converges. the second is true because each k is larger than N .Exercise 3.

.. + 1 1 − sn−1 sn . we once again look at the sequence {sn }. because for every N we’ve shown that there is some k such that N +k k=N +1 ak >1− > sk (note that 1 - > 1 because 0 < < 2 ). for any s N and any arbitrarily small .. By induction.11c To prove the inequality: sn ≥ sn−1 → → → → 1 sn an s2 n an s2 n an s2 n {sn } is an increasing sequence (see part b) multiply both sides by an /sn sn − sn−1 = an (see part b) algebra ≤ ≤ ≤ ≤ 1 sn−1 an sn sn−1 sn −sn−1 sn sn−1 1 1 sn−1 − sn Now consider the summation n k=2 1 1 − = sk−1 sk 1 1 − s1 s2 + 1 1 − s2 s3 38 + . we can make sN +k arbitrarily large by choosing a suﬃciently large k. an To show that sn is divergent.and we’re told that every an > 0.. + aN +k + . + = sN +1 sN +k sN +k sN +k sN +k A bit of algebraic manipulation shows us that aN +1 +. . an From theorem 3.. Exercise 3.. + ≥ + . we can make sNN < by choosing a suﬃciently large k.22. + ≥1− sN +1 sN +k sN +k which means that.14. in order for sn to be convergent we must be able to ﬁnd some integer N such that n k=m ak < sk for all n ≥ m ≥ N But there can be no such N .. And this is suﬃcient to show that the series does not converge. This means that. And from this. by choosing suﬃciently large k. .. we also know that sn ≥ sm whenever n ≥ m. so this last inequality is equivalent to aN +k sN +k − sN sN aN +1 + . an would be convergent). Choose any such that 0 < < 2 . + ≥ =1− sN +1 sN +k sN +k sN +k which is what we were asked to prove.+aN +k = sN +k −sN ..21 of “convergent series”. by the deﬁnition 3.. Therefore: m n aN +1 aN +k aN +1 aN +k aN +1 + . Now let sN be an arbitrary element of {sn } and let be an arbitrarily small real. And we’ve established +k that aN +1 aN +k sN + . Therefore. We’ve already determined that {sn } is an increasing sequence. and we know that’s it’s not convergent (otherwise. From the fact that {sn } is not bounded.. we know that {sn } is not bounded (from theorem 3. we know that s1 ≥ s1 whenever n ≥ m. so we know that sn+1 > sn . . which says that a bounded monotonic series is convergent). we get the inequality N +k k=N +1 ak sN ≥1− ≥1− sk sN +k an 1 We’ve now established everything we need to show that sn is divergent.

25. If an = 1/n. the summation becomes an = 1 + nan 1 n 2 = 1 2 1 n which is divergent by theorem 3.28). But the series an /(1 + nan ) is convergent: ∞ ∞ m an 1 1 = = 1 + nan 2m 2 n=0 m=0 This series is convergent to 2 by theorem 3. Exercise 3. and so by deﬁnition 3. the terms of {sn } increase without bound as n → ∞.11d The series an /(1 + n2 an ) always converges. so that ∞ 1 sk−1 − k=2 1 1 = sk s1 Now. To construct a convergent series. and we know that {tn } is monotonically increasing 1 because each an is positive.14 we know that {tn } converges. Therefore. we know that ∞ k=2 an ≤ s2 n ∞ k=2 1 1 1 − = sk−1 sk s1 We can add one term to each side so that our summation starts at 1 instead of at 2 to get ∞ k=1 a1 1 an ≤ 2 + s2 s1 s1 n We can now show that the series converges.Most of the terms in this summation cancel one another out: the summation “telescopes down” and simpliﬁes to n 1 1 1 1 − = − sk−1 sk s1 sk k=2 As we saw in part (b). we see that an 1 < 1 + n2 an n2 n=0 n=0 We know that 1/n2 converges (theorem 3. and therefore test of theorem 3. Deﬁne the partial sum of this series to be n {tn } = k=1 1 sk−1 − 1 sk We’ve shown that {tn } is bounded above by a1 /s2 + 1/s1 .28. From the fact that an > 0. we can establish the following chain of inequalities: an 1/an an 1 1 = = 1 < 2 2a 2a 2 1+n n 1/an 1 + n n n an + n From this. The series an /(1 + n2 an ) converges by the comparison ∞ ∞ an /(1 + nan ) may or may not converge. 39 . for instance.21 we know that its associated series converges. by the inequality we proved above. let an be deﬁned as an = 1 0 if n = 2m − 1 (m ∈ Z) otherwise The series an is divergent. by theorem 3.26. since there are inﬁnitely many integers of the form 2m − 1. And this is what we were asked to prove.

. + an ≤ + . we know that 1− am a rn < + . so relationship between {rn } and α: rn = k=n an = α for some α ∈ R. Take m ≤ n and consider the sum n k=m am an ak = + . . + (rn − rn+1 ) ≤ +. + n rm rm rn whenever n > m.. . which is what we were asked to prove. we can form our proof. + ≤ rm rm rn rn Taking the two leftmost terms of this inequality and performing some simple algebra then gives us 1− rn+1 am an ≤ + . then. which means that it’s true whenever n + 1 > m. 40 .. . + ≤ rm rm rn rn Now. leaving us with rm − rn+1 am an rm − rn+1 ≤ + . + + rm rm rn rn+1 This last statement is true whenever n ≥ m... + (rn − rn+1 ) am an (rm − rm+1 ) + (rm+1 − rm+2 ) + .. But each an is positive and each rn is the sum of positive terms. By adding this term to the right-hand side of the inequality. we see that ∞ ∞ rk = m=k am = ak + m=k+1 am = ak + rk+1 so that ak = rk − rk+1 . By a simple replacement of variables. {rn } is a decreasing sequence). . . we see that rk > rk+1 which (from transitivity) means that rm > rn when n > m (that is. but our index is oﬀ by one. so an+1 /rn+1 is strictly positive. .. we know that ak /rm ≤ ak /rn for all k.. this last term approaches α − α = 0. we accomplish two things: we correct our index and we make this a strict inequality (<) instead of a non-strict inequality (≤). With these equalities. + an am an am + .+ ≤ rm rm rn rn Notice that most of the terms of these numerators cancel one another out. + rk rm rn From the fact that rm ≥ rn ..Exercise 3. we see that this inequality is equivalent to (rm − rm+1 ) + (rm+1 − rm+2 ) + . rn+1 am an an+1 1− < + . from the fact that ak = rk − rk+1 . . + rm rm rn This is close to what we want to prove.. ... we see that there is a ∞ ∞ n−1 n−1 ak = k=1 ak − k=1 ak = α − k=1 ak As n → ∞. so that am + . proof of divergence We’re told that an converges. . And because each ak is positive..12a establishing the inequality From the deﬁnition of rn . and so we see that limn→∞ rn = 0. From the deﬁnition of rn .

13 Let an be a series that converges absolutely to α and let bn be a series that converges absolutely to β. + >1− >1− > rN rN +n rN where the last step of this chain of inequalities is justiﬁed by our choice of 0 < < 1 .. Exercise 3. in doing so. . this is suﬃcient to show that every > 0 there is some N . + >1− rN rN +n rN We can choose n to be arbitrarily large. rN +n approaches zero while rN remains ﬁxed. for every N . We can prove that the Cauchy product of these two series (deﬁnition 3.22. which is suﬃcient to show that it converges by theorem 3. we see that aN aN +n rN +n + . 1 ) and choose any arbitrary integer N . So. we see that n k=1 ak √ < rk n √ √ 2( rk − rk+1 ) k=1 and many of the terms in the right-hand summation cancel one another out. From the 2 inequality we veriﬁed in part (a1).24. This choice of n gives us the inequality aN +n rN +n aN + . . 2 We’ve shown that there is some such that.Now choose any arbitrarly small from the interval (0. we can guarantee that rN +n /rN < ... we can ﬁnd n k=N an > rn an /rn diverges (since we’ve proven the negation of the “for From theorem 3. 41 .48) is bounded.12b rn > rn+1 > 0 √ √ → rn > rn+1 √ √ √ → 2 rn > rn + rn+1 √ → 2 > n√rn n+1 √ √ → 2( rn − rn+1 ) > √ √ → 2( rn − rn+1 ) > √ √ → 2( rn − rn+1 ) > √ r + r established in part (a) take square root of each side √ add rn to each side √ divide by rn multiply each side by a positive term simply the right-hand side we established an = rn − rn+1 in part (a) √ √ √ √ ( rn + rn+1 )( rn − rn+1 ) √ rn rn −rn+1 √ rn a √n rn Having established this inequality. ” statement of the theorem). so that the series “telescopes down” and simpliﬁes to n √ ak √ √ < 2 r1 − rk+1 rk k=1 so that n→∞ n lim k=1 √ √ ak √ √ < lim 2 r1 − rk+1 = 2 r1 n→∞ rk This shows that this sum of nonnegative terms is bounded.. Exercise 3. by choosing a suﬃciently large n.

→ = ≤ =

n k=0 |ck | n k=0 n k=0 n k=0

k j=0 k j=0 k j=0

aj bk−j

deﬁnition 3.48 of ck triangle inequality 1.33(c)

|aj bk−j | |aj ||bk−j |

**We can expand this summation out to get: = |a0 ||b0 | + (|a0 ||b1 | + |a1 ||b0 |) + . . . + (|a0 ||bn | + |a1 ||bn−1 | + . . . + |an ||b0 |) Deﬁne Bn to be
**

n k=0

|bk | and An to be

n k=0

|ak |. We can rearrange these terms to get:

= |a0 |Bn + |a1 |Bn−1 + |a2 |Bn−2 + . . . + |an |B0 Now we can add several nonnegative terms to get ≤ |a0 |Bn + |a1 |(Bn−1 + |bn |) + |a2 |(Bn−2 + |bn−1 | + |bn |) + . . . + |an |(B0 + |b1 | + . . . + |bn |) = |a0 |Bn + |a1 |Bn + |a2 |Bn + . . . + |an |Bn = An B n =( = αβ This shows that each partial sum of |ck | is bounded above by αβ and below by 0 (since it’s a series of nonnegative terms). By theorem 3.24, this is suﬃcient to prove that |ck | converges; and since ck is the Cauchy product, we have proved that the Cauchy product ck converges absolutely.

n k=0

|ak |) (

n k=0

|bk |)

Exercise 3.14a

Choose any arbitrarily small > 0. We’re told that lim{sn } = s, which means that there is some N such that d(s, sn ) < whenever n > N . So we’ll rearrange the terms of σn a bit: σn =

n k=0 sk

n+1

=

N k=0 sk

+ k=N +1 sk n+1

n

**Whenever n > N we know that d(s, sn ) < , which means that − < sn − s < , or that s − < sn < s + . This gives us the inequality
**

N k=0 sk

+ k=N +1 (s − ) < σn < n+1

n

N k=0 sk

+ k=N +1 (s + ) n+1

n

**The terms in some of these summations don’t depend on k, so we can further rewrite this as
**

N k=0 sk

+ (n − (N + 1))(s − ) < σn < n+1

N k=0 sk

+ (n − (N + 1))(s + ) n+1

sk n(s − ) (N + 1)(s − ) n(s + ) (N + 1)(s + ) − < σn < k=0 + − n+1 n+1 n+1 n+1 n+1 n+1 For many of these terms, the numerators are constant with respect to n. Therefore many of these terms will approach zero as n → ∞. n(s − ) n(s + ) lim < lim σn < lim n→∞ n + 1 n→∞ n→∞ n + 1 s − < lim σn < s + +

n→∞

N k=0 sk

N

And this is just another way of saying d(s, lim σn ) < . And limn→∞ σn = s.

was arbitrarily small, so we have shown that

42

Exercise 3.14b

Let sn = (−1)n . Then sn is either equal to 0 or 1, depending on n, and 1 0 ≤ σn ≤ n+1 n+1 Taking the limit as n → ∞ gives us 0 ≤ lim σn ≤ 0

n→∞

**which can only be true if
**

n→∞

lim σn = 0

Exercise 3.14c

Deﬁne sn to be sn =

1 n 2 n 1 2

+k

if n = k 3 (k ∈ Z) otherwise

√ √ There are no more than 3 n perfect cubes within the ﬁrst n integers, so {sn } will contain no more than 3 n terms of the form (1/2)n + k. This gives us the inequality

n n

sn =

k=0 k=0

1 2

n

√ 3

n

√ 3 k ≤2+

+

k=0

√ n( 3 n + 1) 2

n k=1

where the last step in this chain of inequalities is justiﬁed by the common summation Continuing, we see that

n

k = k(k + 1)/2.

√ 3 sn ≤ 2 +

k=0

√ √ √ √ 3 √ n( 3 n + 1) n( 3 n + 3 n) 3 ≤2+ = 2 + n2 2 2

**We can now analyze the value of σn . σn =
**

n k=0 sn

n+1

√ 2 3 √ 3 2 + 1 2 + n2 ≤ = √ n 3 n+1 n + √1 2 3

n

**Taking the limit of each side as n → ∞ gives us lim σn ≤ lim √ 3
**

n→∞ 2 √ 3 2 n

+1

1 √ 3 2 n

n→∞

n+

1 = lim √ = 0 n→∞ 3 n

Every term of {sn } was greater than zero, so we know that the arithmetic average σn is greater than zero. Therefore 0 ≤ lim σn ≤ 0, and therefore lim σn = 0.

Exercise 3.14

proving the equality

n

kak = (s1 − s0 ) + 2(s2 − s1 ) + . . . + n(sn − sn−1 )

k=1

**= −s0 + (1 − 2)s1 + (2 − 3)s2 + . . . + ((n − 1) − n)sn−1 + sn (n) = −s0 − s1 − . . . − sn + (n + 1)sn Therefore, if we divide this by n + 1, we get 1 n+1
**

n

kak =

k=1

−s0 − s1 − . . . − sn + sn = −σn + sn n+1

43

**establishing convergence We’re told that limn→∞ nan = 0, so by part a we know that
**

n→∞

lim

kak =0 n+1

n k=1

**We’re also told that {σn } converges to some value σ, so by theorem 3.3(a) we know that lim kak + σn n+1
**

n k=1

n→∞

= (0 + σ) = σ

And since sn =

n k=1

kak n+1

+ σn , this is suﬃcient to prove that limn→∞ sn = σ.

Exercise 3.15

If you think I’m going to go through this tedious exercise, you can lick me where I shit.

Exercise 3.16a

{xn } is monotonically decreasing √ From the fact that 0 < α < xn we know that α < x2 , and therefore n xn+1 = 1 2 xn + α xn < 1 2 xn + x2 n xn = xn

This shows that xn > xn+1 for all n, which proves that {xn } is monotonically decreasing. √ The limit of {xn } exists and is not less than a √ √ First, √ show that {xn } is bounded below by a. We know that x0 > a because we chose it to be. And if we xn > a for any n, we have → xn = → xn − → → → → → → √ √ a assumed 1.18(d)

deﬁnition of xn+1 √ So by induction, we know that √n ≥ a for all n. We’ve now demonstrated that {xn } is monotonically x √ decreasing and is bounded below by a, so we’re guaranteed that the limit of {xn } exists and that lim{xn } ≥ a. √ The limit of {xn } is not greater than a √ √ √ √ Now we show that lim{xn } ≤ a. Assume that lim{xn } = b for some b > a. By the deﬁnition of “limit”, √ √ we can ﬁnd N such that n > N implies d(xn , b) < for any arbitrarily small . We’ll choose = b − a (which √ √ is positive since b > a ≥ 1 implies b > a). √ → d(xn , b) < √ √ → d(xn , b) < b − a √ √ → xn − b < b − a √ √ → xn < b − a + b

a=0 √ 2 (xn − a) ≥ 0 √ x2 − 2xn a + a ≥ 0 n √ x2 + a ≥ 2xn a n √ x2 +a n a 2xn ≥ √ 1 a a 2 xn + xn ≥ √ xn+1 ≥ a

chosen value of metric on R1

44

by part (b). we have en < β e1 β 2n √ <2 3 1 10 2n <4 1 10 2n 45 . that the limit is not less than a. then en+1 1 1 e2 n β < √ = (e2 ) < n β β 2 a e1 β 2n−1 2 e1 β 21 1 β2 = β e1 β 2n =β e1 β 2n which shows that it is therefore true for n + 1. Then.16c When a = and x1 = 2. √ a. The next part can be proven by induction. we have therefore proven that lim{xn } = a. we have √ √ x1 − a 2− 3 e1 √ √ = = β 2 a 2 3 which can be shown to be less than 1/10 through simple algebra. then the next term in the sequence {xn } will be less than b. And since we’ve already shown that the sequence is monotonically decreasing. √ √ Having shown that lim{xn } exists. we have e2 e2 1 e2 < √ = 1 = β β 2 a And. so we’ve shown that the limit cannot be any value greater than √ a. if the statement is true for n. this is suﬃcient to show that it is true for all n ∈ N. and that the limit is not greater than Exercise 3. which contradicts √ assumption that lim{xn } = b. Setting n = 1 in the above inequality. Exercise 3. this means that every subsequent term will √ √ √ our be even farther from b. By induction. Therefore the limit is not b: √ but b was an arbitrary value greater than a. we have en+1 = xn+1 − √ a= √ √ √ x2 + a √ x2 − 2xn a + a (xn − a)2 n − a= n = 2xn 2xn 2xn = e2 e2 n n < √ 2xn 2 a √ where the last step is justiﬁed by the fact that xn > a (see part (a)).We can then calculate xn+1 in terms of xn to get a chain of inequalities: √ √ √ √ (b − a) + 2 b b − a + b + a x2 + a ( b − a + b)2 + a n √ √ = xn+1 = < √ √ 2xn 2( b − a + b) 2( b − a + b) √ √ √ √ √ 2b + 2 b b − a b( b + b − a) √ √ = √ = √ = b √ 2( b − a + b) b+ b−a √ √ This shows us that if the distance between xn and b becomes less than b − a (which it is √ guaranteed to √ do eventually because lim{xn } = b).16b From the deﬁnition of xn+1 and en .

17c √ √ We can show than {xn } converges to a by showing that we can make d(xn . Exercise 3. √ √ Lemma 2: xn > a → xn+2 < xn and xn < a → xn+2 > xn √ → xn > a √ ↔ xn − a > 0 √ √ multiplied by positive terms. so it’s still > 0 ↔ 2(xn − a)(xn + a) > 0 ↔ 2(x2 − a) > 0 n At this point. We do this √ √ by demonstrating the relationship between d(xn . Exercise 3. If “>” is replaced with “<” in each of the above steps.17 √ √ √ √ Lemma 1: xn > a → xn+1 < a. Exercise 3.Exercise 3. a). If “>” is replaced with “<” in each of the above steps.17b √ √ Because we chose x1 such that x1 > a. ↔ 2x2 + 0xn − 2a) > 0 n ↔ 2x2 + (1 + a − 1 − a)xn − 2a) > 0 n ↔ xn + x2 + axn + x2 > a + axn + a + xn n n ↔ xn (1 + xn ) + xn (a + xn ) > a(1 + xn ) + a + xn ↔ xn (1+xn )+xn (a+xn ) 1+xn a+xn 1+xn > a(1+xn )+a+xn 1+xn a+xn 1+xn division by a positive term ↔ xn 1 + ↔ xn > >a+ ↔ xn (1 + xn+1 ) > a + xn+1 a+xn+1 1+xn+1 ↔ xn > xn+2 √ deﬁnition of xn+2 This shows that xn > a → xn+2 < xn . so we can use induction with lemma 2 to show than {x2k+1 } (the subsequence of {xn } consisting of elements with odd indices) is a monotonically decreasing sequence. and xn < a → xn+1 > a √ → xn > a √ √ √ ↔ xn ( a − 1) > a( a − 1) √ √ ↔ xn a − xn > a − a √ √ ↔ xn a + a > a + xn √ ↔ a(xn + 1) > a + xn √ ↔ a > a+xn xn +1 √ ↔ a > xn+1 deﬁnition of xn+1 √ √ This shows that xn > a → xn+1 <√ a. a) arbitrarily small. 46 . we will √ have also constructed a proof that xn < a → xn+1 > a. a) and d(x1 .17a √ We’re forced to choose x1 such that x1 > a. we will have √ also constructed a proof that xn < a → xn+2 > xn . But bear with me. the algebraic steps become bizarre and seemingly nonsensical. lemma 1 tells us that x2 < a. We can then use induction with lemma 2 to show that {x2k } (the subsequence of {xn } consisting of elements with even indices) is a monotonically increasing sequence.

So limn→∞ d(xn . 0 < c < 1 = en √ a−1 a+1 This tells us how to express en+2 in terms of en . e2 ]. giving us √ √ √ 1− a 1− a 1− a en+2 = en+1 = en 1 + xn+1 1 + xn 1 + xn+1 √ √ √ √ 1− a 1− a 1− a 1− a = en = en 1+xn +a+xn n 1 + xn 1 + xn 1 + a+xn 1+x 1+xn √ √ √ (1 + xn )(1 − a) (1 − a)2 1− a = en = en 1 + xn 1 + 2xn + a 1 + 2xn + a √ ( a − 1)2 < en a−1 The last step in this chain of inequalities is justiﬁed by the fact that a + 2xn + 1 > a − 1 > 1. every two iterations reduces the error term by a factor of ( a−1)/( a+1) (linear convergence). and this shows that √ limn→∞ en = 0. Exercise 3. the nth iteration reduced the error term by a factor of 10−2 (quadratic convergence). we have 47 . a) = 0 which is √ suﬃcient to prove that limn→∞ {xn } = a. . And if xn > p a. Exercise 3. a). We can use this same equality to express en+1 in terms of en . We can use this same inequality to express e2n+2 in terms e2 and e2n+1 in terms of e1 : e2n+2 < ce2n < c(ce2n−2 ) < . . < cn e2 e2n+1 < ce2n−1 < c(ce2n−3 ) < . .18 Lemma 1 : {xn } is decreasing. . Finally. this means that we can make cn arbitrarily small by taking suﬃciently large n.17d √ √ As shown above. we have √ √ ( a − 1)2 a−1 en+2 < en = cen . n For the algorithm in exercise 16.en+2 = xn+2 − → en+2 = xn+2 − → en+2 = → en+2 = a+xn+1 1+xn+1 √ a= √ √ a + xn+1 √ a + xn+1 − a − xn+1 a − a= 1 + xn+1 1 + xn+1 √ − a √ a √ a √ √ → en+2 (1 + xn+1 ) = a + xn+1 − a − xn+1 a √ √ √ → en+2 (1 + xn+1 ) = xn+1 (1 − a) − a(1 − a) √ √ → en+2 (1 + xn+1 ) = (xn+1 − a)(1 − a) √ → en+2 (1 + xn+1 ) = en+1 (1 − a) → en+2 = en+1 √ 1− a 1+xn+1 a+xn+1 − a−xn+1 1+xn+1 √ This tells us how to express en+2 in terms of en+1 . Continuing. < cn e1 And since 0 < c < 1. so we can make en arbitrarily small by taking √ suﬃciently large n. √ √ We’re asked to choose that x1 > p a. and en < c2 [maxe1 . remember that we deﬁned en to be d(xn .

.. . x∗ ) < 2p . which is suﬃcient √ to show that it has some limit x∗ with x∗ ≥ p a. + 1p−1 → p > k + k 2 + k 3 + . we see that we can ﬁnd xn and xn+1 such that: 48 . And if xn > √ p a. .→ xn > → xp n √ p a >a → 0 > a − xp n → pxp > pxp + a − xp n n n → pxp > (p − 1)xp + a n n → pxp n pxp−1 n → xn > (p−1)xp +a n pxp−1 n (p−1)xn + a xp−1 p p n > → xn > xn+1 Lemma 2 : 0 < k < 1 → p(1 − k) > 1 − k p Let k be a positive number less than 1 and let p be a positive integer.. From the deﬁnition of limit. we can ﬁnd N such that n > N → d(xn . → p = 10 + 12 + 13 + . + k p−1 →p> →p> (k−1)(k+k2 +k3 +. we have: a>0 >0 √ p p a xn → p 1− √ p → p − 1 > p xn − → xp (p n − 1) > xp n a xn √ p >1− a a xp n √ p a xn from lemma 2 expand and rearrange the terms √ → pxp − xp > p p axp−1 − a n n n √ p−1 p → (p − 1)xn + a > p p axn → → √ (p−1)xp +a p p axp−1 n n > pxp−1 pxp−1 n n √ (p−1)xn a + pxp−1 > p a p n p xn − a a xp n multiply both sides by xp n rearrange the terms and simplify divide both sides by pxp−1 n simplify deﬁnition of xn+1 → xn+1 > √ p a √ p a Finally : limn→∞ {xn } = We’ve shown that {xn } is decreasing (lemma 1) and that it’s bounded below (lemma 3). From this. .+kp−1 ) k−1 kp −1 k−1 p → p(k − 1) > k − 1 Lemma 3 : xn > √ p √ p a √ p We know that x1 > → xn > →1> √ p a because we chose it.

3k + 2) 3n−m n=1 m ∞ We then split up the summation into three distinct parts. xn+1 ) = p xn p a pxp−1 n triangle inequality lemma 1: xn > xn+1 + p < p deﬁnition of xn+1 − a p−1 pxn < xp −a n pxp−1 n < a p → xn − xp−1 n < a xp−1 n This last statement tells us that limn→∞ xn − → limn→∞ xn − ∗ √ p a a xp−1 n p = 0. x∗ ) < ∧ d(xn+1 . 3k + 2) Every term in the leftmost sum is divisible by three. each digit is either 0.→ d(xn .1. for some j ∈ N.3c → limn→∞ xn (1 − a/xp ) = 0 n → x 1 − x∗ √ → p a/x∗ = 1 √ → p a = x∗ =0 From the deﬁnition of x∗ as the limit of {xn }. ∞ αn iﬀ (∃m) 3j + αm + ∈ (3k + 1. This gives us. we know that a real number r is not in the Cantor set iﬀ it is in any interval of the form 3k + 1 3k + 2 . From equation 3. this tells us that limn→∞ {xn } = to prove. We can determine the necessary and suﬃcient conditions for x(α) to fall into such an interval: x(α) ∈ Cantor iﬀ x(α) ∈ αn ∈ 3n n=1 ∞ ∞ 3k + 1 3k + 2 .19 The idea behind this proof is to consider the elements of the line segment [0. k ∈ N 3m 3m Let {αn } represent an arbitrary ternary sequence (that is. When we do this. 3m 3m iﬀ (∃m) 3k + 1 3k + 2 . so the leftmost sum is itself divisible by three. m. or 2). 3m 3m iﬀ (∃m) αn ∈ (3k + 1.3c theorem 3. 3k + 2). x∗ ) + d(x∗ . √ p a which is what we wanted Exercise 3. m−1 iﬀ (∃m) n=1 αn 3n−m + n=m αn 3n−m + n=m+1 ∞ αn n−m 3 ∈ (3k + 1. 3k + 2) m−1 iﬀ (∃m) n=1 αn 3m−n + αm + n=m+1 αn n−m 3 ∈ (3k + 1. we notice that the mth iteration of the Cantor set eliminates every number with a 1 as the mth digit of its ternary decimal expansion.24 in the book. 1] in the form of their base-3 expansion. x∗ ) < → xn − xn+1 < → xn − xn − → → xn p 2p p → d(xn . xn+1 ) < d(xn . or that =0 assumed theorem 3. j ∈ N 3n−m n=m+1 49 .

The set En is a intersection of closed sets. So the only way that the set membership in the previous statement can be true is iﬀ (∃m) αm + δ ∈ (1. N ) we have d(p∗ .1. we know that 0 < αm + δ < 3: their sum is never a multiple of 3. pn ) < /2. Exercise 3. and it’s smallest when αn = 0 for all n. Let {sn } be an arbitrary sequence constructed in this way. Continuing with our chain of “iﬀ” statements.9. Each αn is either 0. 0 < δ < 1 From the bounds on αm and δ. so any limit point of En is a point of En . 2). and there’s only one integer in the open interval (0. This comes from the fact that lim diam En = 0: see the text below deﬁnition 3. pn ) ≤ d(p∗ . So. 0<δ<1 iﬀ (∃m) αm ∈ (0. pi) < /2. From the deﬁnition of Cauchy convergence. 2): iﬀ (∃m) αm = 1 iﬀ x(α) ∈ x(a) where x(a) is as deﬁned in the exercise. and is therefore closed (thm 2. So the summation is largest whenever αn = 2 for all n.10b to conclude that En contains only one point. 3k + 2). j ∈ N. n > N implies d(pi . there is some M such that i > M implies d(p∗ . 2) We know that αm is an integer. there is some N such that i.21 We know that each Ei ∈ En is nonempty (see note 1). is suﬃcient to prove that {pn } converges to p∗ . pi ) + d(pi . This shows that any real number x(α) is a member of the Cantor set if and only if it is a member of x(a): this is suﬃcient to prove that the Cantor set is equal to x(a). We can then follow the proof of theorem 3. 3) The sequence converges to some s∗ ∈ En . and therefore s∗ (being a limit point of En ) is an element of En . 50 . Exercise 3. This shows that En is nonempty: it contains at least one element s∗ .12. This comes from the fact that it’s a sequence in a complete metric space X: see deﬁnition 3. pn ) = which.Without specifying a certain sequence {αn }. 1) The sequence is Cauchy. we conclude that x(a) is not member of the Cantor set iﬀ: x(a) ∈ Cantor iﬀ (∃m) 3j + αm + δ ∈ (3k + 1. for i. or 2. so we can construct at least one sequence whose ith element is an arbitrary element of Ei . But we can establish bounds for it.24b). 2) The sequence is convergent. because is arbitrarily small. From the convergence of the subsequence to p∗ . We can immediately say three things about this sequence. since the index is from 1 instead of 0. This gives us the bound ∞ 0= 0 ∞ 3n−m n=m+1 < αn ∞ 3n−m n=m+1 < 2 3n−m n=m+1 =2 1 =1 3n n=1 ∞ Note that the upper bound is 1. not 3. we can’t evaluate the rightmost summation. n > max(M.20 Choose some arbitrarily small .

there is some neighborhood Nr2 (p2 ) that’s a subset of both E1 and G2 .22 The set G1 is an open set. q) : p. bounded. q ∈ ∅} = sup ∅ To ﬁnd the supremum of the empty set in R. constructing a series of nested sets E1 ⊃ E2 ⊃ · · · ⊃ En This is a series of closed. qn ) ≤ d(pn . we can take the closure of Ns1 and still have Ns1 ⊂ Nr1 ⊆ G1 (since δ ≤ s1 < r1 ). qm ) or d(pn . and Ei ⊂ Gi . Not only do we have Ns1 ⊂ Nr1 ⊆ G1 . At this point. Therefore Gn is nonempty. Exercise 3. Deﬁne E1 to be the neighborhood Ns1 (not its closure). and x must be in every Gi ∈ Gn . every neighborhood of p1 contains some point p2 ∈ G2 . we could take the closure Ns2. pm ) + d(pm . E2 ⊂ E2 ⊂ E1 ⊂ E1 We can continue in this same way. pm ) ≤ /2 n. the neighborhood E1 contains some point p2 ∈ G2 . Choose a smaller neighborhood Ns1 (p1 ). we have E2 ⊂ E2 ⊂ G2 . In either case. qn ) ≤ d(pn . Not only do we have Ns2 ⊂ Nr2 ⊆ G2 and Ns2 ⊂ Nr2 ⊆ E1 . N ∈ R such that n. Regardless of how we deﬁne this minimum (or if it’s deﬁned at all). So (from exercise 21) we know that En contains a single point x. becomes d(pn . But we’re told that lim diam En = 0 so our initial assumption must have been wrong: there is no empty Ei ∈ En . we need to rely on the deﬁnition of supremum: sup ∅ = least upper bound of ∅ in R = min {x ∈ R : a ∈ ∅ → a ≤ x} The set for which we’re seeking a minimum contains every x ∈ R (because of false antecedent a ∈ ∅).23 We’re aren’t given enough information about the metric space X to assume that the Cauchy sequences converge to elements in X. qn ) ≤ + d(pm . m > M → d(pn . More speciﬁcally. Choose p1 ∈ G1 and ﬁnd some open neighborhood Nr1 (p1 ) ⊆ G1 . Exercise 3. both of which are open sets. nested sets. we have E1 ⊂ E1 ⊆ G1 The set G2 is dense in X. We now choose an even smaller radius Ns2 (p2 ). pm ) + d(pm . Deﬁne E2 to be the neighborhood Ns2 (not its closure). qm ) ≤ 51 . we have d(pn . This would mean that lim diam En = diam∅ = diam sup {d(p. M }. At this point. qm ) ≤ /2 From multiple applications of the triangle inequality. Since p2 is in both E1 and G2 . qn ) which. so the supremum of the empty set in R is the minimum of R itself.note 1 We’re told that Ei ⊃ Ei+1 for all i. nonempty. there exists some M. If Ei were empty. then Ek would be empty for all k > i. for any arbitrarily small . qn ) − d(pm . m > N → d(qn . so p1 is either an element of G2 or is a limit point for G2 . it certainly isn’t equal to zero. All we can say is that. m > max{N. for n. qm ) + d(qm . This single point x must be in every Ei ∈ En.

. pi1 . n > Ni → d(pim .} containing elements of X that get “closer together” with respect to distance function d. From each of these. gives us ∆(P. Q) = lim d(pn . from the deﬁnition of equality established in part (a). qn ) = lim |pn − qn | = lim |qn − pn | = lim d(qn . This is an unusual metric space. we have n→∞ lim d(pn . bn ) n→∞ Which. P1 . Each Pi ∈ {Pn } is a set of equivalent Cauchy sequences in X. the triangle inequality gives us n→∞ lim d(pn . . qn ) = lim d(an .Exercise 3. Let {Pn } be a Cauchy sequence in (X ∗ . qn ) n→∞ (6) Combining equations (5) and (6). pn ) = lim |pn − pn | = lim 0 = 0 n→∞ n→∞ so {pn } = {pn }. From the triangle inequality. pn ) n→∞ n→∞ n→∞ so {qn } = {pn }. For each i we have (∃Ni ∈ R) m. ∆). qn ) ≤ lim d(an . rn ) = 0. transitivity If {pn } = {qn } and {qn } = {rn }. bn ) ≤ lim d(an . choose some sequence {pin } ∈ Pi . from the deﬁnition of equality established in part (a). rn ) ≤ lim d(pn . we have ∆(P. we have n→∞ lim d(pn . Q) = lim d(pn . pn ) + d(pn . qn ) + d(qn .24a reﬂexivity For all sequences {pn }. . so {pn } = {rn }. Exercise 3.24c Deﬁne X ∗ to be the set of equivalence classes from part (b). Each Pi ∈ {Pn } is an equivalence class of Cauchy sequences of the form {pi0 . Q) = lim d(pn . qn ) n→∞ n→∞ Which. . bn ) ≤ lim d(pn . so to be clear: The sequence {Pn } is a sequence {P0 . we have ∆(P. P2 .} of equivalence classes in X ∗ that get “closer” to one another with respect to distance function ∆. gives us n→∞ lim d(an . bn ) n→∞ n→∞ (5) A similar application of the triangle inequality gives us n→∞ lim d(an . an ) + d(an . bn ) + d(bn . bn ) n→∞ n→∞ Exercise 3. rn ) = 0 + 0 n→∞ This tells us that limn→∞ d(pn . pi2 . qn ) + d(qn . qn ) ≤ lim d(pn . .24b Let {an } = {pn } and let {bn } = {qn }. pin ) < 52 i . symmetry If {pn } = {qn }.

Let Q be the equivalence class containing {qn }. p(n+m)k ) + d(p(n+m)k . then A = Px = ϕ(x) and therefore A ∈ ϕ(X) (see exercise 3. after choosing appropriately large n. ak ) < From this. we have n→∞ lim ∆(Pn .24d Let {pn } represent the sequence whose terms are all p. We have shown that an arbitrary Cauchy sequence A ∈ X ∗ is either an element of ϕ(X) or a limit point of ϕ(X). From (7) we have 1 ∃X : n. this means that ϕ(X) is dense in X ∗ . means that limn→∞ Pn = Q. ak ) < n→∞ This shows that we can ﬁnd some element of ϕ(X) to be arbitrarily close to A. k > X → d(pnk . qn+m ) Taking the limits of each side gives us n→∞ lim d(qn . pnk ) + d(pnk . gives us ∃Z : k > Z → d(pnk . Pq ) = lim d(pn . ∆(Pp . 3. 53 . And this is the deﬁnition of {qn } being a Cauchy sequence. The sequence {an } is Cauchy. so we’re guaranteed the existence of K such that j. so we have ∃Y : n > Y → ∆(Pn . We can show that limn→∞ Pn = Q. 0 ≤ lim ∆(Pn . which means that A is a limit point of ϕ(X). by the deﬁnition of equality in X ∗ . Pn+m ) < which. qn+m ) < . Therefore Q ∈ X ∗ . k > K → d(aj . 2. Q) = lim n→∞ n→∞ k→∞ lim d(pnk .24e : proof of density Let {an } be an arbitrary Cauchy sequence in X and let A ∈ X ∗ be the equivalence class containing {an }. and let {qn } represent the sequence whose terms are all q. p(n+m)k ) < Therefore. If the sequence {an } converges to some element x ∈ X. then choose an arbitrarily small . qn ) = lim d(p. By deﬁnition.24 for the deﬁnition of Px ). Proof that Q ∈ X ∗ : Choose any arbitrarily small values 1. by the triangle inequality: 2 3 d(qn . Pak ) = lim d(an . Exercise 3. If the sequence {an } does not converge to some element x ∈ X. n + m > N implies d(qn . by the squeeze theorem. This shows that for every there exists some integer N such that n. qk ) ≤ lim n→∞ n (7) So.We’ll deﬁne a new sequence {qn } by letting qi = pim for some m > Ni . qn+m ) ≤ 1 + 3 + 1 But these epsilon values were arbitrarily small and m was an arbitrary integer. qn+m ) ≤ d(qn . q) = d(p. Q) = 0 which. qn ) < And {Pn } is Cauchy. we can consider the sequence whose terms are all ak . q) n→∞ n→∞ Exercise 3. This sequence is a member of the equivalence class Pak = ϕ(ak ) and ∆(A.

but the function is not continuous at x = 0: We can choose contain a point p for which d(f (p). This proves that f (E) ⊆ f (E). so that A = Pa = ϕ(a) ∈ ϕ(X). The interval (0. case 2: e ∈ E If e ∈ E . let E be an arbitrary subset of X. so there are inﬁnitely many elements of f (E) in N (y). which means that f is not continuous by deﬁnition 4.25 The completion of the set of rational numbers is a set that’s isomorphic to R. f (0)) < 1. then y ∈ f (E) and therefore y ∈ f (E). We want to prove that f (E) ⊆ f (E).8. Deﬁne the function f : X → Y as f (x) = x. 0) < δ → d(f (p). we have f (Z(f )) = {0}. 1) with the standard distance metric. we know that f −1 ({0}) = Z(f ) must also be a closed set. and every neighborhood Nδ (0) will 1. or that ϕ(X) = X ∗ . Let Y be the metric space R1 . 1) = (0. case 1: e ∈ E If e ∈ E. then every neighborhood of e contains inﬁnitely many points of E. Exercise 4. f (0)) = 1 < 1. so that ϕ(b) = Pb ∈ X ∗ .5. This means that y is a limit point of f (E). This shows that X ∗ ⊆ ϕ(X).1. so we have f (X) = f (X) = (0. We’re told that f is continuous so. Choose an arbitrarily small neighborhood N (y). To do this. This means that y = f (e) for some e ∈ (E ∪ E ). But there are inﬁnitely many elements of E in the neighborhood Nδ (e). we’re guaranteed the existence of δ such that f (x) ∈ N (y) whenever x ∈ Nδ (e).3 If we consider the image of Z(f ) under f . which means that y ∈ (f (E) ∪ f (E) ) = f (E). 54 . By the corollary of theorem 4. and let E represent the closure of E.2 Let X be a metric space. and is therefore a closed set.Exercise 3. assume y ∈ f (E). And for every ϕ(b) ∈ ϕ(X) there is some Cauchy sequence in X ∗ whose every element is b. This shows that ϕ(X) ⊆ X ∗ . This shows that X ∗ ⊆ ϕ(X) ⊆ X ∗ . Exercise 3.24e : if X is complete If X is complete then every arbitrary Cauchy sequence A ∈ X ∗ converges to some point a ∈ X.1 Consider the function f (x) = This function satisﬁes the condition that limh→∞ [f (x + h) − f (x − h)] = 0 for all x. We’ve shown that every arbitrary element y ∈ f (E) is either a member of f (E) or a limit point of f (E). Exercise 4. 0. A function f for which f (E) is a proper subset f (E) Let X be the metric space consisting of the interval (0. This range is a ﬁnite set. by deﬁnition 4. 1) Exercise 4. 1) is closed in X but open in Y . x=0 x=0 Therefore we can’t pick δ such that d(p.

so either p ∈ E or p ∈ E . so we have f (en ) = g(en ) for all n. Case 1: p ∈ E If p ∈ E. x ∈ (ai . Assume y ∈ f (X).2 again. 55 . But each en is an element of E. bi ) ∧ bi = ∞ f (ai ).29 tells us that E C contains an at-most countable number of disjoint segments. We’re told that f and g are continuous. Using deﬁnition 4. bi ) ∧ −∞ < ai < bi < ∞ This function is the one mentioned in the hint: the graph of g is a straight line on each closed interval [bi . Exercise 4. Therefore y is a limit point of f (E). then there is a sequence {en } of elements of E such that en = p and limn→∞ en = p. it would be a non-interior point of E C . By deﬁnition. We’ll deﬁne the function to be x∈E f (x). so by theorem 4. ∞). Case 2: p ∈ E If p is a limit point of E. so E C is open. b) or (a. We’re told that E is dense in X. then there is a sequence {en } of elements of E such that en = p and limn→∞ en = p.4: f (E) is dense in f (X) To show that f (E) is dense in f (X) we must show that every element of f (X) is either an element of f (E) or a limit point of f (E). This function can easily (albeit tediously) be shown to be continuous on R1 . Exercise 2. f (b ). but open sets have no non-interior points). case 1: p ∈ E If p ∈ E. bn )} be the at-most countable collection of disjoint open segments. so by theorem 4.2 we know that limn→∞ f (en ) = f (p) and limn→∞ g(en ) = g(p). Exercise 4. Then p = f −1 (y) ∈ X. constructing the function We must separately consider the cases for x ∈ E and x ∈ E. x ∈ (ai . and if x ∈ E we must consider the possibility that E contains an interval of the form (−∞.Exercise 4. then we’re told that f (p) = g(p). We’re told that f is continuous. this tells us that limx→p f (x) = f (p) = y. This proves that f (p) = g(p) for all p ∈ X.4b Choose an arbitrary p ∈ X. so p is either an element of E or a limit point of E. f (bi )−f (ai ) f (ai ) + bi −ai (x − ai ). this means that f (X) is dense in f (E). ai+1 ] ∈ E C . Let {(an . case 2: f −1 (y) ∈ E If p is a limit point of E. We’ve shown that every element y ∈ f (X) is either an element of f (E) or a limit point of f (E).2 we know that limn→∞ f (en ) = f (p) = y. Each of these segments must be open (if any of them contained a non-interior point. This tells us that g(p) = lim g(en ) = lim f (en ) = f (p) n→∞ n→∞ We see that f (p) = g(p) in either case. We’re told E is dense in X. we know that there is a sequence {f (en )} of elements of f (E) From theorem 4.5 The set E C can be formed from an at most countable number of disjoint open intervals We’re told that E is closed. then y = f (p) ∈ f (E).2. bi ) ∧ ai = −∞ i g(x) = (8) x ∈ (ai .

.7 f is bounded If x = 0. The inverse of g is g −1 (x.15) G is compact. so the natural choice is to use the metric d(g(x). g(y)) = d((x. . gk (x)) We’ve shown that each of these gi functions are continuous. . so g(x. The domain of g is a compact metric space so by theorem 4. Exercise 4. . It’s clear that g −1 is a one-to-one and onto function from G to E. ∞): f (x) = 1. fk (x)) For each fi we can deﬁne a function gi as in equation (8) to create a new vector-valued function g : R → R deﬁned by g(x) = (g1 (x). 0) ∩ (0.14 (or theorem 4.10 the function g(x) is continuous.17 the inverse of g −1 – that is. f2 (x). f (x)) is a continuous mapping by theorem 4.10. . therefore by theorem 4.failure if “closed” is omitted Consider the following function deﬁned on the open set (−∞. If the graph is compact Deﬁne g as above. then f (x. . Exercise 4. Therefore by theorem 4.6 Let G represent the graph of f . We can make G into a metric space by deﬁning a distance function: the set G is a subset of R × R. We can then appeal to theorem 4. If f is continuous Both x → x and x → f (x) are continuous mappings. f (x)). −1 x>0 x<0 This function will be discontinuous at x = 0 no matter how we deﬁne the function at this point. g itself – is continuous. . g2 (x).10 once again to conclude that f is continuous. If x > 0: (x − y 2 )2 ≥ 0 → x − 2xy + y ≥ 0 → x − xy + y ≥ 0 → x + y ≥ xy →1≥ If x < 0: (x + y 2 )2 ≥ 0 → x2 + 2xy 2 + y 4 ≥ 0 → x2 − xy 2 + y 4 ≥ 0 → x2 + y 4 ≥ xy 2 →1≥ xy x2 +y 4 2 squares are positive 4 2 2 expanding the squared term LHS remains positive after adding the nonnegative term (xy 2 ) add xy 2 to both sides divide both sides by positive term x2 + y 4 2 2 4 2 4 2 xy 2 x2 +y 4 squares are positive expanding the squared term LHS remains positive after adding the nonnegative term (−3xy 2 ) add xy 2 to both sides divide both sides by positive term x2 + y 4 56 . (y. f (x)) = x. . Let g : E → G be deﬁned as as g(x) = (x. y) = 0 for any value of y. This choice allows us to treat G as a subset of R2 . Extending this to vector-valued functions Let E be a closed subset of R and let f : E → Rk be a vector-valued function deﬁned by f (x) = (f1 (x). f (x)). f (y)) = (x − y)2 + (f (x) − f (y))2 .

0). x2 + x < xy 2 x2 1 = 2 = 2 + y4 x 2x 2 The restriction of f to a straight line is continuous Any straight line that doesn’t pass through (0. The restriction of g to a straight line is continuous The proof for the continuity of the restriction of g is almost identical to that of the continuity of the restriction of f . 0) doesn’t encounter any of the irregularities that occur at the origin. (x. 0). f (x. choose x > 0 such that x2 + x < 2 > 0 such that . we have g(nα .g is unbounded near (0. 0). lines of the form y = cx or x = 0 for some constant c. y)) = 0 xy 2 = 4 =0< 2 + y4 x y For the line x = 0: let And was arbitrary. Then: xy 2 c2 x3 c2 x c2 x = 2 = < < c2 δ = x2 + y 4 x + c4 x 4 1 + c4 x2 1 d(f (0. nβ ) = nα+2β n2α + n6β We can divide the numerator and denominator by nα+2β to get g(nα . 0) = lim g n→∞ 1 nα−2β + n4β−α 1 1 . f (x. choose an arbitrary such and then choose y = x. f is not continuous at (0. 0). This gives us d ((0. so this proves that f is continuous on this line. For the line y = cx: let > 0 be given and let δ = /c2 . Choose x such that 0 < x < δ.0) Choose 0 < δ < 1/2. n3 n = lim n→∞ The rightmost limit is +∞. Regardless of our choice of δ. f (x. > 0 be given choose any y = 0. y)) < δ But there can be no √ epsilon. To see this. it’s trivial but tedious to show that the restriction of f to such a line is continuous. y)) < → d (f (0. f (x. For f to be continuous we must be able to choose some d ((0. we have d(f (0.0) If we let x = nα and let y = nβ . β = −1 this becomes g Taking limits. 57 . So we need only consider lines that pass through the origin: that is. (x. n3 n = n−1 n 1 = −1 +n 2 n 2 1 1 . 0). we have g(0. y)) = but d (f (0. so this proves that f is continuous on this line. 0). y)) = x2 + y 2 = > 0. y)) = And was arbitrary. nβ ) = If we let α = −3.

and suppose diam E < δ. f (q)) < Therefore (9) holds. 58 assumed deﬁnition of diameter deﬁnition of diameter p ∈ E → f (p) ∈ f (E) deﬁnition of diameter assumed we’re letting E be the set containing only p and q. d(p. + + . q) < δ → diam {p. each of which is smaller than δ. q ∈ E)(d(f (p). let E be a nonempty subset of X. for all p. f (α + 2γ)) + . (9) implies (10) Let equation (9) hold. f (q) ∈ f (E))(d(f (p). From the deﬁnition of uniform continuity there exists some δ > 0 such that. we need to show that there exists some M such that |f (p)| < M for every p ∈ R1 . f (q)) < We want to show that this conditional statement is true iﬀ for every diam E < δ → diam f (E) < (10) implies (9) Let equation (10) hold and suppose d(p. f (p)) ≤ d(f (α). so by deﬁnition 4. + (n terms) Exercise 4. Let p be an arbitrary element of R1 .. from (10) deﬁnition of diameter > 0 there exists δ > 0 such that (10) . f (q)) < Let n be the smallest integer such that nδ > p − α (which exists from the Archimedean property of the reals) and deﬁne γ = (p − α)/n.9 Deﬁnition 4.13 we know that f is bounded. q ∈ R1 . p) into n intervals of length γ. f (α + γ)) + d(f (α + γ). This allows us to divide the interval (α. If E is not bounded above or below. q) < δ → d(f (p).18 says that f is uniformly continuous on X if for every > 0 there exists δ > 0 such that (9) dX (p. .Exercise 4.8 To show that f is bounded. + d(f (α + (n − 1)γ). we can simply look to f (x) = x. We’re told that E is a bounded subset of R1 . q) < δ. f (q)} < → d(f (p). Choose any > 0. f (q)) < ) → diam f (E) < Therefore (10) holds. so we know that E has a lower bound α. f (q)) < ) → (∀f (p). q} < δ → diam {f (p). diam E < δ → (∀p. then we can revise the previous proof using the upper bound β in place of the lower bound α. then we will not be able to use the Archimedean property to ﬁnd n such that nδ > p − (−∞) and the proof fails. . d(p. q ∈ E)(d(p. q) < δ) → (∀p.. q) < δ → dY (f (p). We can then apply the triangle inequality multiple times: d(f (α). if E is not bounded If E is not bounded below. For an example of a real uniformly continuous function that is not bounded. f (α + nγ)) < = n This shows us that |f (p)| is bounded by |f (α) ± n | for all p ∈ R1 .

the sequence {qn } must also have p as a subsequential limit since 1 d(qn .19: that is. from theorem 4. (∃N ) : n. q for every value of δ. f (qn )) ≥ . qn ) < n but d(f (pn ). g(y2 )) < Because f is uniformly continuous.2. f (qn )) ≤ d(f (pn ). 59 .37).6a or 2. for arbitrary > 0. See exercise 4. we can ﬁnd p. f (q)) ≥ From the fact that we can ﬁnd such p. m > N → d(f (xm ). these two conditional statements tell us that. f (x2 ) are elements of Y these two conditional statements together tell us that. Let f : X → Y and g : Y → Z be uniformly continuous functions. By converse.11 Choose an arbitrary > 0. shows that {f (xn )} is Cauchy. 2 + 2 = Exercise 4. so (∃δ > 0) : d(xm . And this is what we were asked to prove. If f is not uniformly continuous. it must be the case that both {f (pn )} and {f (qn )} have subsequential limits of f (p). q) < δ but d(f (p). And. Choose an arbitrary > 0. p) ≤ d(qn . f (p)) + d(f (p). Exercise 4. 1 n (11) and construct two sequences If the set X is compact (supposition 1) the sequence {pn } must have at least one subsequential limit p 1 (theorem 3. x2 ) < δ → d(f (x1 ). (∃δ > 0) : d(x1 . y2 ) < α → d(g(y1 ). for some suﬃciently large n. qn ) < n . by deﬁnition.10 We want to prove the converse of theorem 4. so one of our suppositions must be incorrect: either X is not compact or f is not continuous. p) < + n But then. f (xn )) < which. xn ) < δ Together. d(f (pn ). by deﬁnition means that the composite function (g ◦ f ) : X → Z is uniformly continuous. we can set δn = 1 {pn } and {qn } where d(pn . so (∃N ) : n. pn ) + d(pn . we want to prove that if f is not uniformly continuous on X then either X is not a compact metric space or f is not a continuous function. for any . then the converse of deﬁnition 4. We’ll be contradicting this claim. since d(pn . m > N → d(xm . q) < δ ∧ d(f (p). We’ve established a contradiction from our initial assumption that f is uniformly continuous. Because g is uniformly continuous.Exercise 4. f (x2 )) < α → d(g(f (x1 )). g(f (x2 ))) < which. f (x2 )) < α Since the elements f (x1 ). We’re told that f is uniformly continuous. for every δ > 0. If the function f is continuous without being uniformly continuous (supposition 2) this would imply that. we have (∃α > 0) : d(y1 . so we’ll restate it formally in a numbered equation: (∃ > 0)(∀δ > 0) d(p. xn ) < δ → d(f (xm ).12 We’re asked to prove that the composition of uniformly continuous functions is uniformly continuous.18 tells us that there exists some such that.13 for the proof that uses this result. f (xn )) < And we’re told that {xn } is Cauchy. f (q)) ≥ . q ∈ X such that d(p. f (qn )) < which contradicts (11). if X is compact and f is continuous then f is uniformly continuous. we have (∃δ > 0) : d(x1 . x2 ) < δ → d(f (x1 ).

then p is a limit point of E (because of density) and therefore we can construct a sequence {en } that converges to p. if p ∈ E We need to prove that this is actually a well-deﬁned function and that it’s continuous. N ) : n > max{M. This allows us to conclude that {f (pn )} and {f (qn )} both converge to the same point of Y . If p ∈ E.12). qn ) < δ We know that the function f is continuous on E. then {f (pn )} and {f (qn )} both converge to the same element. and we call this point g(p). Let {pn } and {qn } be arbitrary sequences in E that converge to p. but without being able to conclude that Y is compact it’s entirely possible that {f (pn )} and {f (qn )} are merely Cauchy without actually converging to any point in Y at all. From the deﬁnition of convergence (or from exercise 4. f (qn )) < ) This tells us that {f (pn )} and {f (qn )} don’t converge to diﬀerent points. We now deﬁne the function g : X → Y as limn→∞ f (p). so for any > 0 there exists some δ > 0 such that (13) (12) δ 2 δ 2 d(pn .13 (Proof using exercise 4. qn ) ≤ d(pn . The function g is well-deﬁned It’s clear that every element of X is mapped to at least one element of Y .11) For every p ∈ X it is either the case that p ∈ E or p ∈ E. but we’ll try for a more general solution and assume only that Y is a complete metric space (see deﬁnition 3. N } → d(pn . 60 . p) < (∀δ > 0)(∃M ) : m > M → d(qm . if p ∈ E g(p) = limn→∞ f (en ). We need to show that if {pn } and {qn } are sequences in E that converge to p.Exercise 4. The exercise asks us to assume that Y is R1 . p) < We can then use the triangle inequality to show that (∀δ > 0)(∃M.11). qn ) < δ → d(f (pn ). f (qn )) < Combining equations (12) and (13) we see that for any > 0 we can ﬁnd some integer N such that n > N → (d(pn . p) + d(p. qn ) < δ) → (d(f (pn ). we know that (∀δ > 0)(∃N ) : n > N → d(pn . Therefore the function g(p) has a unique value for every p ∈ X. but it’s not immediately clear that each element of X is mapped to only one element of Y .

To prove that f is discontinous at x = 0. If you’re satisﬁed with an “intuitive proof”. g(q)) triangle inequality established bound for d(g(pn ). g(q)) < This shows that we can constrain the range of g by constraining its domain. Deﬁnition 3. even when n is suﬃciently large to make d(q. so for any we can choose pn . qn ) arbitrarily small. qn ) from our initial assumption This holds true for any n. q) < δ/2 assumed Let {pn } be a sequence that converges to p and let {qn } be a sequence that converges to q. g(q)) → d(g(p). Consider the function f (x) = x with a domain of E = n : n ∈ N . qn ) < δ/4. g(q)) ≤ d(g(p). qn ) → d(pn . f 1 n+k We haven’t speciﬁcied a value for f (0) yet. f 1 n+k 61 . This can be seen intuitively by recognizing that the range of f (x) = 1/x is unbounded in every neighborhood of x = 0. choose any arbitrary neighborhood around 0 with radius δ. but we do know that |f (1/n) − f (1/n + k)| = |n − (n + k)| = k. However. by the triangle inequality. qn ) ≤ d(pn . which guarantees that d(f (y). buckle up and get ready for a few more paragraphs of inequalities and Greek letters. → d(pn . g(q)) < d(g(p). Otherwise.f 1 n+k ≤d f 1 n . For such values of n we have d(g(p). g(qn )) < /2 → d(g(p).12 tells us that all compact metric spaces and all Euclidean metric spaces are also complete metric spaces. even when n is suﬃciently large to make d(q. f (qn )) < /2 By deﬁnition. Could we replace the range with any metric space? 1 1 No. p) + d(p. g(q) speciﬁcally so that this would be true). we have g(p) = f (p) for p ∈ E. d(p. we have d f 1 n . And we can make δ arbitrarily small. p) + δ/2 + d(q. g(pn )) + /2 + d(g(qn ). d(p. f (x)) = 0 for any y in Nδ (x)). For any positive δ 1 integer k we also have d( n+k . 0) < δ. qn ) < δ We’re told that f is continuous on E. f (qn )) arbitrarily small if we can make d(pn . The set E is dense in X = E ∪ {0}. we’re done. qn such that → d(f (pn ). g(qn )) This holds true for any n. Our proof didn’t depend on the range of g being R1 : we assumed only that the range was a complete metric space.The function g is continuous d(p. This last inequality is therefore equivalent to k≤d f 1 n . Could we replace the range with Rk or any compact metric space? By any complete metric space? Yes. g(pn )) + d(g(pn ). qn ) < d(pn . f (0) + d f (0) . 0) < δ. so we can make d(f (pn ). pn ) < /4 (we know that {g(pn )} converges to g(p) and that {g(qn )} converges to g(q) because we deﬁned g(p). g(qn )) + d(g(qn ). qn ) < /4. This function is continuous at every x ∈ E (we can ﬁnd a neighborhood Nδ (x) that contains only x. We 1 can ﬁnd an integer n such that n > 1 (Archimedian property of the reals) so that d( n . and therefore g is continuous. so → d(g(pn ). f (0) + d f (0) . q) + d(q. pn ) < δ/4 in which case we have → d(pn . Now. there is no way to deﬁne f (0) to make this a continuous function.

But this is true for arbitrarily large k and arbitrarily small δ. the general idea is simple. This means that there is no continuous extension from E to X. so one of the two terms on the right-hand side of this inequality must be ≥ k/2. f (x) = x. x2 ) ). we know that there exists some δa and δb such that d(x1 . x2 ) on which f (x) obtains a local maximum or minimum. We therefore have a nested sequence of compact sets {[an . Then we can ﬁnd x1 < x2 < x3 such that f (x2 ) > f (x1 ) and f (x2 ) > f (x3 ). Exercise 4. . x3 ) containing the x for which f (x) obtains its local maximum. bn ]} consists of just one point i ∈ I.9. bn ]} guarantees that f (i) ≤ i and f (i) ≥ i. bn+1 ] as [an .14 proof 2 This proof. Exercise 4. which is much clearer and more concise than mine. If g(0) = 0 or g(1) = 0. so that by the previous method we can be sure that either d(f ( n ). Otherwise. Therefore by theorem 4. f (x)) < d(x3 . We deﬁne [an+1 . 62 . The proofs for either case are analogous. x) < δa → d(f (x1 ). n. We’ll now inductively deﬁne a sequence of intervals {[ai . f (bn+1 ) ≤ bn+1 . [m.edu. we have g(0) > 0 and g(1) < 0. so g(x) is continuous by theorem 4. so by exercise 3. Both f (x) and x are continuous. 1]. We want to construct a closed subinterval [xa . bi ]} as follows3 : Let [a0 . Let f : R1 → R1 be a continuous mapping and assume that f is not monotonically increasing or decreasing. we deﬁne to be = min{d( f (x2 ). was taken from the homework of Men-Gen Tsai (b89902089@ntu. we know that f (0) > 0 and f (1) < 1. x) < δb → d(f (x3 ). Now assume that we have deﬁned {[ai .5 of “continuous”. Exercise 4. . so we’ve proven that f will be discontinuous however we deﬁne f (0). m]. Let m = (an + bn )/2. f (x1 ) ). bi ]} for i = 0. f (0)) ≥ k/2 > . . this proves that f is not continous at 0. We can now show that f is not continuous at x = 0.14 proof 1 If f (0) = 0 or f (1) = 1. x2 ) ) is not an open set. bn ]} whose diameter converges to zero. If the function f is continuous but not monotonic then we can ﬁnd some open interval (x1 . We can’t assume that the supremum of f ( (x1 . so f ( (x1 .5 ) > f (x2 ). 1] for which g(x) = 0: at this point. f (x)) < 3 Our method of deﬁning this sequence is similar to Newton’s method or the Bisection method of root ﬁnding. f (x3 ) )} Because f is continuous. bn+1 ] has diameter 2−(n+1) with f (an+1 ) ≥ an+1 . Choose any > 0. xb ] ⊂ (x1 . if f (m) ≤ m if f (m) ≥ m (This procedure is underdeﬁned for the case of f (m) = m. The interval [an+1 . bn ]. This point f (x) is not an interior point of f ( (x1 . Our method of constructing {[an .5 for which f (x2. b0 ] = [0. Deﬁne a new function g(x) = f (x) − x. or such that f (x2 ) < f (x1 ) and f (x2 ) < f (x3 ). And we never speciﬁed the value of f (0). therefore f (i) = i and this is the point we were asked to ﬁnd.tw). . 1. d( f (x2 ). For all possible choice of δ > 0 we can 1 ﬁnd integers n > 1 and k > 2 .Every term of this inequality is nonnegative. x3 ) ) occurs at f (x2 ): there could be some x2.15 Although the details of this proof might be ugly. By deﬁnition 4. Otherwise. but when f (m) = m we’ve found the point we’re looking for anyway). we’re done. To do this.23 there must be some intermediate point in the interval [0. so we’ll focus only on the case that f (x2 ) > f (x1 ) and f (x2 ) > f (x3 ). f (0)) ≥ k/2 = δ 1 or d(f ( n+k ). then we’re done.21 we know that {[an .

which by deﬁnition means that the functions are discontinuous for integer values of x. so one of our assumptions must be wrong: for any given triple (p. xa ) ∪ (xb . is the continuous function f (x) = x. If x ∈ N then.We can now let xa = x1 + 1 δa and xb = x3 − 1 δb . q. q. we have 2 d([x − δ]. xb ]). without loss of generality. r) rational triple with q < x < r. d((x − δ). q. This is all getting a bit muddled. Exercise 4. From our deﬁnition of the set E we know that f (x−) exists for all x ∈ E.15). But the supremum of a set can’t be an interior point of the set. x3 )) 3) (x ∈ (x1 . This is clearly a contradiction. xb ] Together these four facts tell us that f ( (x1 . we have shown that if f is a continuous open mapping from R1 to R1 then f is monotonic. r) it’s not possible to ﬁnd two distinct elements of E that fulﬁll all four criteria. x3 ) ) is not open even though (x1 . r such that the four criteria are met We’re assuming that f (x−) < f (x+). that x < y. and therefore f ([xa . so to recap our results so far: 1) [xa . xb ]) ⊂ f ((x1 . so we can ﬁnd some neighborhood Nδ (x) containing a rational q that fulﬁlls criteria (b). x3 )) → f (x) < f (x2 ) 4) f (x) has a local maximum for some x ∈ [xa . We know that f doesn’t obtain its local maximum for any x 2 2 in the intervals (x1 . (∃q ∈ Nδ (x) ∩ (a. x))(a < q < x ∧ [q < t < x → f (t) < p] Similarly. xb ]. [x]) = |(x − 1) − x| = 1 1 2 This shows that we can’t constrain the range of these functions by restraining the domain around x ∈ N. Lemma 2: If a given (p. xb ] is a closed subset of R1 and is therefore a compact set. Reversing the roles of x and y establishes the same results for the case of x > y. we add a fourth: (d) a<q<x<r<b Lemma 1: For each x ∈ E we can ﬁnd p. then. x3 ) is open. the existence of f (x+) tells us that there is some neighborhood Nδ2 (x) containing a rational r such that (∃r ∈ Nδ2 (x) ∩ (x. xb ]) contains its supremum (which may or may not occur at x2 . And this is what we were asked to prove. y are two elements of E that both meet the four criteria. To the three criteria listed in the exercise. But this means that f (w) < p (by criteria (b) since q < w < y) and that f (w) > p (by criteria (c) since x < w < r). The density of the reals guarantees that there exists some rational w such that x < w < y. x3 ) because all of these xs are suﬃciently close (within δ) to x1 or x3 so that f (x) 1 is less (by at least 2 δ) than f (x2 ). xb ]) is a closed and bounded subset of R1 (theorem 4. r) triple fulﬁlls the criteria for x and y. And [xa . but we only care that the supremum exists for some x ∈ [xa .20b) we can ﬁnd some rational p such that f (x−) < p < f (x+). so by the density of Q in R (theorem 1.17 Deﬁne E as in the hint. so we know that f ([xa . q. [x] + (x). xa ) or (xb . x3 ) ) contains its own supremum for some x ∈ [xa . (x)) = |(1 − δ) − 0| = 1 − δ > Note that the the sum of these discontinuous functions. so f ( (x1 . Exercise 4. Assume. x3 ) 2) f ([xa . b))(x < r < b ∧ [x < t < r → f (t) > p] Therefore each x ∈ E can be associated with at least one (p. By converse. xb ] ⊂ (x1 . for every 0 < δ < 1 .16 The functions [x] and (x) are both discontinuous for each x ∈ N. so f is not an open mapping (theorem 4. then x = y Suppose x.8) We’ve shown that “f is not monotonic” implies “f is not a continuous open mapping from R1 to R1 ”. More formally. 63 . Note that we haven’t proven that each triple can be associated with a unique x ∈ E: we’ve only shown that each triple can be associated with at most one x ∈ E.

64 . Now consider all of the (q. Lemma 4: F is at most countable If we deﬁne F to be the set of x for which f (x+) < f (x−) we can prove that F is at most countable with only trivial modiﬁcations to the previous lemmas. r) pairs of rational numbers. q ≤ n q For each i ≤ n the product xi is irrational.Lemma 3: E is at most countable The set of rational triples is countable (theorem 2. Choose any > 0 and let n be the smallest integer such that n > 1 . x + δi ) contains no rationals of the form k . So if we have constructed neighborhoods i δ1 .13). But this would mean that either f (x−) or f (x+) wouldn’t exist. r such that q < x < r. And this is what we were asked to prove. Lemma 2 tells us that we can create a function from the set of triples onto E.12). . y) < δ → d(f (x). Therefore. y) < δ → d(f (x). Let G be the set of rational pairs (q. The sets E. Let x be an arbitrary element of G. d x. q ∈ Q. f (y)) < < n and for irrational numbers y we have d(x. we could construct a sequence {tn } for which {tn } → x but {f (tn )} → α. Therefore each x is an isolated point of G: we can ﬁnd some radius δ such that Nδ (x)∩G = x and Nδ (x) contains rational numbers q. q) ∈ G. and we will have constructed a neighborhood (x − δ. so it’s at most countable (theorem 2. which contradicts our deﬁnition of G. Each of these pairs will either have one. Exercise 4. by deﬁnition. for all rational numbers y.F . r). r) for which there is exactly one element of g in the open interval (q. .13). i i which means that The neighborhood (x − δi . If every neighborhood of x contained another element of G. x + δ) that contains no rational numbers with denominators ≤ n. so the cardinality of E is not more than the cardinality of the set of rational triples.18 f is continuous at every irrational point Let x be irrational. means that f is continuous at x. Lemma 5: G is at most countable Deﬁne G to be the set of x for which f (x+) = f (x−) = α but f (x) = α. This proves that G is at most countable. The set G is a subset of Q × Q. we will need to ﬁnd a neighborhood of x that contains no rational numbers in the set p : p. If we want to constrain the range of f to f (x) ± . more than one. so we can create a function from G onto G. f (y)) = 0 < Which. we have 1 d(x. δ2 . so there exists some integer mi such that m < xi < m + 1 m m+1 <x< i i So we can deﬁne a neighborhood around x with radius δi where δi = min d x. . Therefore their union is countable (theorem 2. Therefore the cardinality of E is at most countable. or zero elements of G in the associated open interval (q.and G exhaust all the diﬀerent types of simple discontinuities and all of these sets are countable. . And each x ∈ G can be associated with at least one (p. r). δn we can let δ = min{δi } (this minimum exists because n is ﬁnite). m+1 m .

Exercise 4. By constructing the sequence in this way we see that lim xn → p but d(f (xn ). without loss of generality. If it does not contain such a subsequence. But the fact that {yn } converges to p means that p is a limit point of {yn }. Exercise 4. y) z∈E 65 . that f (xn ) < f (p) for every term of {xn } (see note 1 for justiﬁcation of WLOG). for every δ > 0. y) < → y ∈ E.20a If x ∈ E then x is an element of the open set E . therefore there is some radius such that d(x. By deﬁnition 4. Assume. which means that p is a limit point of the closed set described in the exercise (the set of all x with f (x) = r). y) Case 2: x ∈ E. Therefore f is continuous. though. we no longer have f (x) = 0 and therefore f has a simple discontinuity at x. we can ﬁnd some x such that d(x. z) : z ∈ E)}. z) = 0. Therefore we are able construct a sequence {xn } where each term of {xn } gets closer to p even though every term of 1 {f (xn )} diﬀers from f (p) by at least . Therefore p is an element of this set. From the density property of the reals we can ﬁnd a rational number r such that f (xn ) < r < f (p). so for each xn there is some yn between xn and p such that f (yn ) = r. When x is rational.f has a simple discontinuity at every rational point Let x be a rational point. z) < and therefore inf z∈E d(x. we know that there exists some > 0 such that. Therefore we can assume. without loss of generality. If x ∈ E then for every > 0 we can ﬁnd some z ∈ E such that d(x. z) ≥ > 0. xn ) < n but d(f (p). If we follow the previous method of constructing δ we immediately see that f (x+) = f (x−) = 0. Therefore |ρE (x) − ρE (y)| = |ρE (x)| = inf d(x. Therefore {f (xn )} either contains an inﬁnite subsequence where f (xn ) < f (p) for all n or f (xn ) > f (p) for all n (or both). then it contains an inﬁnite subsequence where f (xn ) > f (p) for all n: the proof for this case requires only the trivial twiddling of a few inequality signs. We’re told that f has the intermediate value property. y ∈ E If x and y are both elements of E then from part (a) we have ρE (x) = ρE (y) = 0 and therefore |ρE (x) − ρE (y)| = 0 ≤ d(x. Note 1: justiﬁcation for WLOG claim For each term of the sequence {f (xn )} either f (p) < f (xn ) or f (xn ) < f (p). so our initial supposition must be false: f is not discontinous at any point p. y ∈ E If y ∈ E but x ∈ E then ρE (y) = 0 and d(x. f (p)) ≥ . that f (xn ) < f (p) for each term of {xn }. p) < δ but d(f (x). we choose each element of {xn } so that d(p.5 of “continuity”. f (xn )) ≥ . We’ve established a contradiction. This sequence {yn } converges to p (since the terms of {yn } are squeezed between the terms of {xn } and p) and {f (yn )} clearly converges to r (since f (yn ) = r for all n). And this is a contradiction since we speciﬁcally chose r so that f (p) = r. which means that f (p) = r.20b Case 1: x. If it contains an inﬁnite subsequence where f (xn ) < f (p) for all n then we use this subsequence in place of {xn } in the proof. f (p)) ≥ for all n. More formally. C Exercise 4.19 Suppose that f is discontinous at some point p. y) ∈ {d(x. z) ≤ d(x. and therefore inf z∈E d(x.

q) is an element of ρF (K) and therefore d(p. b89902089@ntu. 1] are open in [0. q ∈ F such that d(p.tw 66 . z) This must hold for any choice of z.41). The conclusion fails if neither set is compact Consider the sets4 K= n+ 1 :n∈N n 1 n F = {n : n ∈ N} − {2} The set F is closed. The distance d(p. z) ≤ d(x. y. 0 + δ). q) < n.9. 0 + δ) are in ρF (K). y) By changing the roles of x and y we can similarly show that ρE (y) − ρE (x) ≤ d(x. so it’s clear that f (p) = 0 iﬀ p ∈ A and f (p) = 1 iﬀ p ∈ B. q ∈ F . By choosing z to make d(y. y) < . and we can choose p ∈ K. z) we have ρE (x) ≤ d(x. for any Exercise 4. z) arbitrarily close to inf z∈E d(x.8. y) These three cases exhaust the possibilities for x.18 for uniform continuity with δ = . Therefore d(p. q) ∈ (0 − δ.21 K is compact and ρF is continuous on K. Exercise 4. so this is equivalent ρE (x) ≤ d(x. Therefore we can ﬁnd some neighborhood around 0 with radius δ such that none of the elements in the interval (0 − δ. 1]. so ρF (K) is compact (and is therefore both closed and bounded from theorem 2.edu. Choose p ∈ K. q) > δ. y) + ρE (y) + Our choice of to can be made arbitrarily small (possibly even zero. so f is continuous from theorem 4. From the deﬁnition of A and B we have 1 A = f −1 (0) ⊆ f −1 0. therefore V and W are open by theorem 4. y) Together. ρE (y)) < whenever d(x. these last two inequalities show us that |ρE (y) − ρE (x)| ≤ d(x. and therefore 0 is not a limit point of ρF (K) (since ρF (K) is closed). This means that the ρA (p) + ρB (p) is never zero. y) + d(y. The sets V 2 and W are clearly disjoint since f (p) can have only a single value for any given p. if y is a limit point of E). neither K nor F is compact. =V 2 B = f −1 (1) ⊆ f −1 4 This 1 .1 2 =W counterexample was taken from the homework of Men-Gen Tsai. And this is just deﬁnition 4. y ∈ E If neither x nor y are elements of E then. The results of exercise 20 tell us that 0 is not an element of ρF (K). y) + ρE (y) → ρE (x) − ρE (y) ≤ d(x. This shows us that for every > 0 we have d(ρE (x).Case 3: x.22 Exercise 20a shows us that ρA (p) = 0 iﬀ p ∈ A and ρB (p) = 0 iﬀ p ∈ B. 2 ) and ( 1 . 1 The intervals [0. for arbitrary z ∈ E we have ρE (x) ≤ d(x.

Choose α ∈ (a. |f (b) − f (x)| |f (a) − f (x)| The method we used to select and δ ensure that both of these λ values are in the interval (0. then choose α = x. b) such that f (β) = f (x): if this is not possible.23a: proof of continuity Choose x ∈ (a. b). 1). |f (a) − f (x)|} (x − a) (b − x) . and choose such that 0< Choose δ such that 0 < δ < min Deﬁne λa . x) such that f (α) = f (x): if this is not possible. Notice that we can express δ in terms of λa or λb as δ = (1 − λa )(x − a) = (1 − λb )(b − x) If we restrict the domain of f to (x − δ.Exercise 4. y) < δ → d(f (x). Choose some > 0. this means that f is continuous. then choose β = x. x + δ) we have |f (x + δ) − f (x)| = |f (x + (1 − λb )(b − x)) − f (x)| = |f (λb x + (1 − λb )b) − f (x)| ≤ |λb f (x) + (1 − λb )f (b) − f (x)| = |(1 − λb )(f (b) − f (x))| = < = < and also |f (x − δ) − f (x)| = |f (x + (1 − λa )(x − a)) − f (x)| = |f (λa x + (1 − λa )a) − f (x)| ≤ |λa f (x) + (1 − λa )f (a) − f (x)| = |(1 − λa )(f (a) − f (x))| δ = | x−a (f (a) − f (x))| δ | b−x from our deﬁnition of δ algebraic rearrangement from the convexity of f algebraic rearrangement from our deﬁnition of λ from our deﬁnition of δ algebra from our deﬁnition of (f (b) − f (x))| |(f (b) − f (x))| (b−x) (b−x)|f (b)−f (x)| from our deﬁnition of δ algebraic rearrangement from the convexity of f algebraic rearrangement from our deﬁnition of λ from our deﬁnition of δ algebra from our deﬁnition of < = < (x−a) (x−a)|f (a)−f (x)| |(f (a) − f (x))| We chose an arbitrarily small and showed how to ﬁnd a nonzero upper bound for δ such that d(x. |f (b) − f (x)|. f (y)) < . By deﬁnition. Justifying “without loss of generality”: special case of f (a) = f (x) or f (b) = f (x) Let x be an arbitrary point in the interval (a. Choose β ∈ (x. b). λb as λa = 1 − δ x−a δ λb = 1 − b−x < min{ . 67 . Assume without loss of generality that f (x) = f (a) and f (x) = f (b) (see below for justiﬁcation of WLOG).

let f be a convex function. Exercise 4. β) containing x such that (α. b). β) containing our arbitrary x. b). we can derive two results. we can establish an inequality: λf (u) + (1 − λ)f (s) ≥ f (λu + (1 − λ)s) → λf (u) + (1 − λ)f (s) ≥ f → λf (u) + (1 − λ)f (s) ≥ f → λf (u) + (1 − λ)f (s) ≥ f t−s u−s t−s u−t = u−s u−s from convexity of f deﬁnition of λ. b). We only have α = x if there was no α ∈ (a. b).23b: increasing convex functions of convex functions are convex Let g be an increasing convex function. so this shows that h is a convex function. We can also see that (1 − λ) = 1 − From this. and deﬁne their composite to be h = g ◦ f . Since x was an arbitrary element of (a. Deﬁne λ to be λ= t−s u−s From this deﬁnition we immediately see that 0 < λ < 1. we can use the previous method to prove that f is continuous on an interval (α. b). β) ⊆ (a. (1 − λ) algebra algebra algebra u+ u−t u−s s ut−us+su−st u−s t(u−s) u−s → λf (u) + (1 − λ)f (s) ≥ f (t) From this last inequality. Therefore f is continuous on the interval (α. x) such that f (α) = f (x): this can only happen if f is constant (and therefore continuous) on the interval (a. β) ⊆ (a. b) such that f (β) = f (x): this can only happen if f is constant (and therefore continuous) on the interval [x. λf (u) + (1 − λ)f (s) ≥ f (t) → λf (u) − λf (s) ≥ f (t) − f (s) → t−s u−s previously derived subtract f (s) from each side deﬁnition of λ (f (u) − λf (s)) ≥ f (t) − f (s) 68 . x]. If α = x < β. f (λx + (1 − λ)y) ≤ λf (x) + (1 − λ)f (y) → g(f (λx + (1 − λ)y)) ≤ g(λf (x) + (1 − λ)f (y)) → g(f (λx + (1 − λ)y)) ≤ g(λf (x) + (1 − λ)f (y)) ≤ λg(f (x)) + (1 − λ)g(f (y)) → g(f (λx + (1 − λ)y)) ≤ λg(f (x)) + (1 − λ)g(f (y)) → h(λx + (1 − λ)y) ≤ λh(x) + (1 − λ)h(y) x and y were arbitrary. If α = x = β then f is a constant function on the entire interval (a. We only have β = x if there was no β ∈ (x. so we know that 0 < t − s < u − s. this is suﬃcient to prove that f is continuous on (a. b) and is therefore trivially continuous. we can use the previous method to prove that f is continuous on an interval (α. this is suﬃcient to prove that f is continuous on (a. If α < x = β. b) this is suﬃcient to prove that f is continuous on (a. b). x) ⊆ (a.23c: the ugly-looking inequality We’re given that s < t < u. x) containing x such that (α. b) containing our arbitrary x. Therefore f is continuous on the interval (a. from convexity of f g is an increasing function from convexity of g transitivity deﬁnition of h Exercise 4. b). we can use the previous method to prove that f is continuous on an interval (x.If α < x < β. β) containing x such that (x.

k ∈ Λ and let m = f (mx + (1 − m)y) = f =f =f =f j+k 2 j+k 2 . therefore m ∈ Λ. 69 . 1]. λ ∈ [0. k ∈ Λ algebra deﬁnition of m = mf (x) + (1 − m)f (y) This shows that f (mx + (1 − m)y) ≤ mf (x) + (1 − m)f (y). It’s immediately clear that 0 ∈ Λ and 1 ∈ Λ. Let Λ represent the set of values for λ for which (16) holds. y ∈ (a. y ∈ (a. We’re also told that f has the property f x+y 2 1 2 ≤ f (x) + f (y) for all x. b) 2 (17) 1 which is just (16) with λ = 2 . so j ∈ [0. (15) Exercise 4. b). so (jx + (1 − j)y) ∈ (a. choose an arbitrary x.24 To prove that f is convex on (a. Proof + 2−j−k y 2 that m ∈ Λ: deﬁnition of m algebra algebra algebra x jx+kx+2y−jy−ky 2 jx−jy+y 2 + kx−ky+y 2 kx+(1−k)y 2 jx+(1−j)y 2 + j is in Λ. b) (this choice of x and y will remain ﬁxed for the majority of this proof). To prove that f is convex we must show that [0. The same is true for k. ≤ ≤ = f (jx+(1−j)y)+f (kx+(1−k)y) 2 jf (x)+(1−j)f (y)+kf (x)+(1−k)f (y) 2 j+k f (x) + 2−j−k f (y) 2 2 We can apply (16) because j. Lemma 1: If j. 1]) : f (λx + (1 − λ)y) ≤ λf (x) + (1 − λ)f (y) (16) To do this. b). we have: λf (u) + (1 − λ)f (s) ≥ f (t) → −λf (u) − (1 − λ)f (s) ≤ −f (t) → (1 − λ)f (u) − (1 − λ)f (s) ≤ f (u) − f (t) → (1 − λ)(f (u) − f (s)) ≤ f (u) − f (t) → u−t u−s (14) previously derived multiply both sides by −1 add f (u) to both sides add f (u) to both sides deﬁnition of λ (f (u) − f (s)) ≤ f (u) − f (t) Dividing both sides of this last equation by u − t gives us f (u) − f (s) f (u) − f (t) ≤ u−s u−t Combining (14) and (15) gives us f (u) − f (s) f (u) − f (t) f (t) − f (s) ≤ ≤ t−s u−s u−t which is what we were asked to prove. b) we must show that (∀x. 1] ⊂ Λ. k ∈ Λ then (j + k)/2 ∈ Λ Assume that j.Dividing both sides of this last inequality by t − s gives us f (t) − f (s) f (u) − f (s) ≥ u−s t−s Similarly. therefore ∈ Λ. y ∈ (a. So we can apply (17).

1} ⊂ Λ so we have 0 ∈ E. By the hypothesis of induction α/2k ∈ Λ therefore m/2k+1 ∈ Λ. . 2k ]. From this we have m 2k+1 = m+1 m−1 1 + k+2 = k+2 2 2 2 β α + k k 2 2 By the hypothesis of induction α/2k ∈ Λ and β/2k ∈ Λ. Now assume that E contains 1. by lemma 1. 1] so we can construct a sequence {λn } in Λ such that limn→∞ λn = p. the deﬁnition of convexity. 1 ∈ E. we see that (∀c ∈ C)(z − c ∈ F ). β ∈ Z with α. case 1: m is even If m is even then m = 2α for some α ∈ Z. And we’ve established that F is closed and that and K is compact. of intersection def. From lemma 2. The set F = z − C is clearly closed because z is a ﬁxed element and C is closed. Each λn is an element of Λ. Therefore. so for each n we have f (λn x + (1 − λn )y) ≤ λn f (x) + (1 − λn )f (y) The function f is continuous. 1])f (λx + (1 − λ)y) ≤ λf (x) + (1 − λ)f (y) But we chose x and y arbitrarily from the interval (a. we see that Λ is dense in [0.Lemma 2: all rationals of the form m/2n with 0 ≤ m ≤ 2n are members of Λ This can be proven by induction. . This shows that f is convex. This completes the inductive step. This covers all possible cases for m. We know that 1 {0. so by theorem 4. 1]. 2 . . which is what we were asked to prove.25a The “hint” given for this problem is actually a proof. of F From the deﬁnition of the set F . k. therefore E = N. . 0 ≤ α ≤ 2k and therefore m/2k+1 = α/2k. so by exercise 21 we know that there exists some δ > 0 such that 70 . λ ∈ [0. c ∈ C)k + c = z → z ∈K +C rearrangement of quantiﬁers def. The sets K and F are disjoint: k ∈K ∩F → k ∈K ∧k ∈F → k ∈ K ∧ (∃c ∈ C)(k = z − c) → k ∈ K ∧ (∃c ∈ C)(k + c = z) → (∃k ∈ K. By lemma 3 we know that (∀λ ∈ [0. b). case 2: m is odd If m is odd then (m − 1)/2 = α and (m + 1)/2 = β for some α. Lemma 3: every element of [0. 1] is a member of Λ Choose any p ∈ [0. so k + 1 ∈ E. Let E be the set of all n for which the lemma is true. 1])f (λx + (1 − λ)y) ≤ λf (x) + (1 − λ)f (y) which is (16). so we have proven that (∀x. m/2k+1 ∈ Λ. y ∈ (a. β ∈ [0. of K + C assumed def. b). Exercise 4.2 we can take the limit of both sides as n → ∞ to get f (px + (1 − p)y) ≤ pf (x) + (1 − p)f (y) which means that p ∈ Λ. 2. Choose an arbitrary m such that 0 ≤ m ≤ 2k+1 .

q are integers so this last statement implies that α is a rational number. of distance function in Rk algebra def. k ∈ K)(|z − (c + k)| > δ) → (∀c ∈ C. so C + K is closed. there is exactly one m ∈ N such that m + nα ∈ [0. y ∈ C1 + C2 . p. But z was an arbitrary point of C + K. we deﬁne it to be ∆ = {δ : (∀x ∈ [0. k) > δ) → (∀c ∈ C. 1) and then extend this proof out to prove that C1 + C2 is dense in R1 . Lemma 1: Each element of C1 + C2 has a unique representation of the form m + nα Assume that m + nα and p + qα are two ways of describing the same element of C1 + C2 . m + nα = p + qα → (m − p) + (n − q)α = 0 → (n − q)α = (p − m) assumption of equality algebra algebra If both sides of this equation are zero. Then we have d(x. Assume that p + nα and q + nα are both in the interval [0. for every x ∈ [0. Now. 1) Then |(p + nα) − (q + nα| = |p − q| ∈ [0.(∀c ∈ C. then n = q and p = m so that our two representations are not unique. 1) ∩ C1 + C2 then δ ∈ ∆ Assume d(x. k ∈ K is simply the set C + K: so we have → (∀y ∈ C + K)(d(z. It’s also clear that any integer multiple of δ is an element of C1 + C2 . choose an arbitrary p ∈ [0. k ∈ K)(d(z. And this is what we were asked to prove. n ∈ N} (18) We’ll ﬁrst prove that C1 + C2 is dense in [0. y) = y − x = p + mα − q + nα = (p − q) + (m − n)α = δ So that δ is itself an element of C1 + C2 . c + k) > δ) by exercise 21 def. By contradiction. 1). And p and q are both integers. k ∈ K)(|z − c − k| > δ) → (∀c ∈ C. 1) that is not a multiple of δ. 1). If both sides are nonzero. More formally. of distance function in Rk This shows us that z is an interior point of C + K. so there exists some a such that p a< <a+1 δ 71 . each element of C1 + C2 has a unique representation of the form m + nα. so we have shown that C + K is open. every neighborhood of radius > δ contains a point of C1 + C2 . y) = δ for some 0 ≤ x < y < 1 with x. we can divide by p − m: →α= p−m n−q algebra But m.25b Let α be an irrational number and let C1 + C2 be deﬁned as the set C1 + C2 = {m + nα : m. n. Deﬁne the set ∆ to be the set of radii δ such that. so it must be the case that p = q. y) > δ) def. y) = δ for any x. Lemma 2: For each n ∈ N. By theorem 2. Lemma 3: If d(x. of distance function in Rk The set of c + k for all c ∈ C. 1))(Nδ (x) ∩ C1 + C2 = ∅) We want to prove that C1 + C2 is dense in [0. 1): we can do this by proving that ∆ has a greatest lower bound of 0. y ∈ [0. Exercise 4. 1). Every real number lies between two integers.23 the complement of an open set is closed. k ∈ K)(d(z − c.

Therefore p is a limit point of C1 + C2 . Deﬁne the functions f : X → Y and g : Y → Z as x if x < 1 2 f (x) = 2x if x ≥ 1 2 72 .7 tells us that the composition of continuous functions is continuous. y) > δ for all x. 1).14) and therefore g is uniformly continuous on Y (theorem 4. By lemma 3 we know that d(x. 1] and Y = [0. so inf ∆ ≥ 0. then. Let m be the integer such that m ≤ p < m + 1. but every real number is a limit point of C1 + C2 . But p was an arbitrary element of [0. So by theorem 4. The sets C1 and C2 are closed because they have no limit points. we must have 2 δ ∈ ∆ whenever δ ∈ ∆. so every element of [0. 1 To construct the counter example. deﬁne X = Z = [0. The deﬁnition of C1 + C2 guarantees that {cn + m} is also a sequence in C1 + C2 . then 1 2δ ∈∆ 1 Proof by contradiction. This gives us a maximum size for the set [0. Therefore we can construct a sequence {cn } of elements of [0.26 We’re told that Y is compact and that g is continuous and one-to-one.17 we conclude that g −1 (g(Y )) is a continuous mapping from compact space g(Y ) to compact space X. 1). Exercise 4. y ∈ [0. We conclude that g(Y ) is a compact subset of Z (theorem 4. But p was an arbitrary element of R1 . The set C1 + C2 is dense in R1 Choose an arbitrary p ∈ R1 . This is suﬃcient to prove that inf ∆ ≤ 0. 1) is a limit point of C1 + C2 . therefore f (x) = g −1 (h(x)) is uniformly continuous if h is uniformly continuous. 1) ∩ C1 + C2 . Each element of ∆ is a distance. We know that 1 ∈ ∆ because a neighborhood of radius δ = 1 around any x ∈ [0. 1)∩C1 +C2 that converges to p − m. Using induction with lemma 4 tells us that (∀n ∈ N)(2−n ∈ ∆). 1) ∩ C1 + C2 : |[0. but C1 + C2 is not closed.19). 1) has an element of C1 + C2 . In exercise 12 we proved that the composition of uniformly continuous functions is uniformly continuous. The set C1 + C2 doesn’t contain any non-integer rational numbers. Assume that δ ∈ ∆ but 2 δ ∈ ∆. 1) lies in such a neighborhood of radius δ. and theorem 3.19 we see that g −1 is uniformly continuous on g(Y ). Therefore C1 + C2 doesn’t contain all of its limit points which means it is not closed. The set C1 + C2 is dense in [0. Therefore inf ∆ = 0. Therefore 1 ∈ ∆. 2 ) ∪ [1. Lemma 4: if δ ∈ ∆. which means that C1 + C2 is dense in [0. which tell us that there is one unique element of [0. 1). so it’s trivially true that they contain all of their limit points. By contradiction.which implies aδ < p < (a + 1)δ which shows that p is in a neighborhood of radius δ of some element of C1 + C2 . so every element of R1 is a limit point of C1 + C2 . The fact that g is one-to-one tells us that g is one-to-one and onto g(Y ). and therefore δ ∈ ∆. this means that every neighborhood of every element of [0. 2]. 1) will contain 0 ∈ C1 + C2 . This means that 0 ≤ p − m < 1. and we know that C1 +C2 is dense in [0. This allows us to conclude that every element of [0. By the deﬁnition of ∆. so by theorem 4. 1) ∩ C1 + C2 for each 1 n ∈ N. The sets C1 and C2 are closed. And this proves that C1 + C2 is dense in R1 . 1) We can now use induction to show that inf ∆ = 0. therefore f (x) = g −1 (h(x)) is continuous if h is continuous. Theorem 4. 1) ∩ C1 + C2 | ≤ 1 δ But this contradicts lemmas 1 and 2.3 tells us that {cn + m} converges to p.

2). From lemma 1 we know that g(f (x)) = x for all x. strictly increasing functions whose domains and ranges are compact. b] ) (theorem 4. so the derivative of f is zero at every point. The domain of g is the range of f . 1 2 1 2 1 2 Exercise 5. b) (this result is a trivial extension of the proofs for theorem 5.1 Choose arbitrary elements x. b). so f is strictly increasing in (a. so 1/g (x) must also be deﬁned.17 ). The composite function h : X → Z is just h(x) = x. so the domain of g is a compact space (thoerem 4.2 Lemma 1: g(f (x)) = x We’re told that f (x) > 0 for all x.19). b). Because the domain of f is a compact metric space we can conclude that f is uniformly continuous (theorem 4. b). Therefore g(f (x)) = x (theorem 4. 73 . by deﬁnition. therefore f is continuous on [a. diﬀerentiable. injective.11). b) ). means that f is one-to-one. From the continuity and injectiveness of f we can conclude that g is continuous on f ( [a. We’re told that f is continuous and that |f (x) − f (y)| ≤ (x − y)2 = |x − y|2 Dividing the leftmost and rightmost terms by |x − y| we have f (x) − f (y) ≤ |x − y| x−y Taking the limit of each side as y → x gives us |f (x)| ≤ 0 But our choice of x was arbitrary. therefore the inequalities in lemma 1 are equivalent to g(f (x)) > g(f (y)) ↔ f (x) > f (y) g(f (x)) < g(f (y)) ↔ f (x) < f (y) g(f (x)) = g(f (y)) ↔ f (x) = f (y) Which means that g is a one-to-one function and is strictly increasing on f ( (a. We’re told that f is diﬀerentiable on (a. y ∈ R1 .14) and therefore g is uniformly continuous (theorem 4.17). which is clearly continuous. From lemma 1. bonus proofs: a big wad of properties for f and g We can show that both f and g are uniformly continous.19). b] (theorem 5. By deﬁnition. this proves that f is a constant function. Exercise 5. b) ) Using lemma 1 we see that f (x) is inversely proportional to g (x): f (x) = lim t→x f (x) − f (t) f (x) − f (t) 1 1 = = lim t → x g(f (x))−g(f (t)) = x−t g(f (x)) − g(f (t)) g (x) f (x)−f (t) We’re told that the derivative f (x) is deﬁned for all x ∈ (a. By theorem 5. this means that x < y ↔ f (x) < f (y) and x > y ↔ f (x) > f (y): from this we conclude x = y ↔ f (x) = f (y) By contrapositive this is equivalent to x = y iﬀ f (x) = f (y) which.11b. we know that f is a one-to-one function that is strictly increasing on (a. g (x) exists for all x ∈ f ( (a.g(y) = y y 2 if y < if y ≥ We can easily demonstrate that f fails to be continuous at x = and that g is continous at every point.

5 Choose an arbitrary > 0.3 f is f (x) = lim From our bounds on x−t g(x) − g(t) + lim = 1 + g (x) t→x x−t x−t and g (x) this gives us t→x 1− 1 M M < f (x) < 1 + 1 M M These are strict inequalities.3c) and its derivative is g (x) = xf (x) − f (x) x2 We want to prove that g is monotonically increasing.4 we know that f (x) is diﬀerentiable and its derivative is given by f (x) = C0 + C1 x + C2 x2 + · · · + Cn xn f (x) = C0 x + By theorem 5. x + 1) g(x) = f (x + 1) − f (x) = (x + 1) − (x) This must be true for all possible values of x. Therefore f is an increasing function and is one-to-one (see lemma 1 of the previous exercise). We now have t > x > N . C n xn Consider the function Exercise 5. and this is what we were asked to prove. 1). Therefore has a real root in (0.Exercise 5. the fact that f (0) = f (1) = 0 means that f (x) = 0 for some x ∈ (0. 1).10. When x = 0 every term evaluates to zero. which is true iﬀ xf (x) − f (x) >0 x2 which is true iﬀ xf (x) > f (x) which is true iﬀ f (x) > 74 f (x) x (19) . The function f is the sum of two diﬀerentiable functions. Exercise 5. so f (1) = 0.3 Choose such that | | < itself diﬀerentiable: 1 M. so f (0) = 0. And was an arbitrary positive real. so there exists some N such that t > N → |f (t)| < With some algebraic manipulation and the use of the the mean value theorem (theorem 5. |g(x)| = |f (t)| ≤ which means that |g(x)| < for all x > N . so by theorem 5.6 The function g is diﬀerentiable (theorem 5. And this is what we were asked to prove.10) we can express g as f (x + 1) − f (x) = f (t) for some t ∈ (x. so the f (t) term in the previous equation is now less than . From example 5. so choose x > N . so we conclude that f (x) is always positive for this choice of . Exercise 5. We’re told that f (t) → 0 as t → ∞. This is true iﬀ g (x) > 0 for all x. which by deﬁnition means that g(x) → 0 as x → ∞.4 C 1 x2 Cn xn+1 + ··· + 2 n+1 When x = 1 this evaluates to the function given to us in the exercise.

x ∈ [a.7 From the fact that f (x) = g(x) = 0 we see that lim f (t) f (t) − f (x) f (t) − f (x) t−x f (x) = lim = lim = t→x g(t) t→x g(t) − g(x) t−x g(t) − g(x) g (x) t→x Exercise 5. x. x) x x−0 We’re told that f is monotonically increasing. . and we’ve shown that this occurs iﬀ g (x) > 0. Therefore: f (x) > f (c) = f (x) x Therefore (19) holds. Exercise 5. .16) and [a. b] therefore f is uniformly continuous (theorem 4. Choose t. . f2 (x).8 We’re told that f is a continuous function on the compact space [a. fn (x)) Assume that f is diﬀerentiable on [a. b] is compact so by the preceeding proof we know that for each fi there exists some δi > 0 such that |t − x| < δi → fi (t) − fi (x) − fi (x) < t−x n Deﬁne δ = min{δ1 . b]. .19). x) we know that |c − x| < |t − x|δ. . And this is what we were asked to prove.10 that f (x) f (x) − f (0) = = f (c) for some c ∈ (0. Choose any > 0: by the deﬁnition of uniform continuity there exists some δ such that 0 < |t − x| < δ → |f (t) − f (x)| < . x. we have f (t) − f (x) |f (c) − f (x)| = − f (x) < t−x Our initial choice of t. . δn }. b]: by the mean value theorem there exists c ∈ (t. Choose an arbitrary > 0 and deﬁne the vector-valued function f to be f (x) = (f1 (x). from our deﬁnition of c. . From the fact that f (0) = 0 we know from theorem 5. so f (x) > f (c). Therefore |f (c) − f (x)| < .To show that (19) holds for all x. and . b] (remark 5. Does this hold for vector-valued functions? Yes. δ2 . Then each fi is diﬀerentiable on [a. For |t − x| < δ we now have f (t) − f (x) − f (x) t−x = = ≤ < (f1 (t) + f2 (t) + · · · + fn (t)) − (f1 (x) + f2 (x) + · · · + fn (x)) − (f1 (x) + f2 (x) + · · · + fn (x)) t−x f1 (t) − f1 (x) f2 (t) − f2 (x) fn (t) − fn (x) − f1 (x) + − f2 (x) + · · · + − fn (x) t−x t−x t−x f1 (t) − f1 (x) f2 (t) − f2 (x) fn (t) − fn (x) − f1 (x) + − f2 (x) + · · · + − fn (x) t−x t−x t−x n + n + ··· + n = 75 . Therefore. x) such that f (t) − f (x) f (c) = t−x From the fact that c ∈ (t. And this is what we were asked to prove. choose an arbitrary x ∈ R. . and was arbitrary. some δ > 0 must exist so that this previous inequality is true for all t. which means that g is monotonically increasing.

f (x) x→0 g(x) lim = f1 (x) + if2 (x) 1 1 −A · +A· x→0 1 g1 (x) + ig2 (x) g1 (x) + ig2 (x) f (x) 1 1 = lim −A · +A· x→0 1 g (x) g (x) 1 1 A = {A − A} · + A · = B B B lim Exercise 5. therefore each of these dependent functions is diﬀerentiable (see Rudin’s remark 5. We’re told that f and g are diﬀerentiable. These criteria are met by if x > 0 x2 . we know that f (0) = lim x→0 f (x) − f (0) x−0 The function f is continuous. Therefore f (0) = 3 as an immediate consequence of the corollary to theorem 5.Exercise 5.9 We’re asked to show that f (0) exists.9 : Alternate proof If limx→0 f (0) = 3 and f (0) = 3. And this is what we were asked to prove.12. so we can use L’Hopital’s rule (diﬀerentiating with respect to h): f (x + h) + f (x − h) − 2f (x) f (x + h) − f (x − h) lim = lim h→0 h→0 h2 2h 1 f (x + h) − f (x) f (x) − f (x − h) = lim + h→0 2 h h We’re told that f (x) exists. Applying the hint given in the exercise. so this limit exists and is equal to = lim h→0 1 (f (x) + f (x)) = f (x) 2 A function for which this limit exists although f (x) does not For a counterexample. if x < 0 f (x) = 0. so we can apply L’Hopital’s rule. g1 represent the real parts of the functions f. we have f1 (x) + if2 (x) x x f (x) = lim −A · +A· lim x→0 x→0 g(x) x g1 (x) + ig2 (x) g1 (x) + ig2 (x) Each of these functions is diﬀerentiable and the denominators all tend to 0. Exercise 5. Exercise 5.11 The denominator of the given ratio tends to 0 as h → 0. −(x2 ). but f (x) does not exist at x = 0. so limx→0 f (x) − f (0) = 0 and limx→0 x = 0. g2 represent their imaginary parts: that is.10 Let f1 . 76 . then f would have a simple discontinuity at x = 0. if x = 0 This function is continuous and diﬀerentiable. we need only ﬁnd a diﬀerentiable function for which f (x) = 1 when x > 0 and f (x) = −1 when x < 0. From the deﬁnition of the derivative. Therefore we can use L’Hopital’s rule. f (x) − f (0) f (0) = lim = lim f (x) − 01 − 0 = lim f (x) x→0 x→0 x→0 x We’re told that the right-hand limit exists and is equal to 3. therefore the leftmost term (f (0)) exists and is equal to 3. g and let f2 . f (x) = f1 (x) + if2 (x) and g(x) = g1 (x) + ig2 (x).16).

we have f (x) = lim |3(x + h)2 | − |3x2 | h→0 h (21) If x > 0 then the terms in the numerator are positive and (21) resolves to f (x) = lim 3(x + h)2 − 3x2 = lim 6x + 3h = 6x h→0 h→0 h If x < 0 then the terms in the numerator are negative and (21) resolves to f (x) = lim −3(|x| + h)2 − 3|x2 | = lim 6|x| + 3h = 6|x| h→0 h→0 h |3h2 | =0 h→0 h It’s clear from the above results that f (x) → 0 as x → 0. Therefore their product xa sin(x−c ) is continuous wherever it’s deﬁned (theorem 4. The sin function hasn’t been well-deﬁned yet.9).12 From the deﬁnition of f (x). we have |x + h|3 − |x|3 h→0 h If x > 0 then the terms in the numerator are positive and (20) resolves to f (x) = lim f (x) = lim (x + h)3 − (x)3 = lim 3x2 + 3xh + h2 = 3x2 h→0 h→0 h (20) If x < 0 then the terms in the numerator are negative and (20) resolves to f (x) = lim −(|x| + h)3 + (|x|)3 = lim −(3|x|2 + 3|x|h + h2 ) = −(3|x2 |) h→0 h→0 h |h3 | =0 h→0 h It’s clear from the above results that f (x) → 0 as x → 0.13a f is continuous when x = 0 The proof that xa is continuous when x = 0 is trivial. and this agrees with f (0): f (0) = lim So f (x) = 3x|x| for all x. 5 Example 5. From the deﬁnition of f (x). we have f (x) = lim If x = 0. 77 . which is everywhere but x = 0. Exercise 5. and this agrees with f (0): f (0) = lim So f (x) = |6x| for all x. From the deﬁnition of f (x). then when h > 0 we have |6h| =6 h→0 h lim and when h < 0 we have h→0 h→0 |6(x + h)| − |6x| h (22) lim |6h| = −6 h So the limit in (22) (which is f (3) (0)) doesn’t exist.2).Exercise 5. so we can also assume that it’s continuous (theorem 5. but we can assume that it’s a continuous function5 .6 says that we should assume without proof that sin is diﬀerentiable.

Exercise 5. therefore a/c < 0 and we have f (xn ) = 1 2nπ + π/2 a c = (2nπ + π/2) −a c where −a/c > 0. which means that lim xn = 0 but lim f (xn ) = 0 and therefore f is not continuous at x = 0. Deﬁne the terms of {xn } to be xn = This sequence clearly has a limit of 0. so if a > 0 we have xa (−1) ≤ xa sin(x−c ) ≤ xa (1) Taking the limit of each of these terms as x → 0 gives us 0 ≤ lim xa sin(x−c ) ≤ 0 x→0 which shows that limx→0 f (x) = 0. it’s suﬃcient to construct a sequence {xn } such that lim xn = 0 but lim f (xn ) = f (0) (theorem 4. By theorem 3. The range of the sin function is [−1. For f to be continuous at x = 0 it must be the case that limx→0 f (x) = 0. and therefore f is continuous at x = 0.2). Note that we’re making lots of unjustiﬁed assumptions about the sin function and the properties of the as-of-yet undeﬁned symbol π.f is continous at x = 0 if a > 0 We have f (0) = 0 by deﬁnition. but f (xn ) = xa sin(x−c ) = n n 1 2nπ + π/2 a c 1 2nπ + π/2 1 c 1 2nπ + π/2 1 c sin(2nπ + π/2) = 1 2nπ + π/2 a c We’re told that c > 0. 1] so that |ha−1 |(−1) ≤ ha−1 sin(h−c ) ≤ |ha−1 |(1) (24) 78 . f is discontinous at x = 0 if a = 0 To show that f is not continuous at x = 0. These cases show that f is continous iﬀ a > 0. but f (xn ) = x0 sin(x−c ) = sin(2nπ + π/2) = 1 n n so that lim{f (xn )} = 1. f is discontinous at x = 0 if a < 0 Deﬁne the terms of {xn } to be xn = This sequence clearly has a limit of 0.13b From the deﬁnition of limit we have f (0) = lim f (0 + h) − f (0) ha sin(h−c ) − 0 = lim = lim ha−1 sin(h−c ) h→0 h→0 h→0 h h (23) We can evaluate the rightmost term by noting that sin is bounded by [−1. 1].20a we see that lim f (xn ) = ∞.

This means that the limit in (23) (and therefore f (0) itself) does not exist.3b. This means that the limit in (23) (and therefore f (0) itself) does not exist.Lemma 1: f is diﬀerentiable when x = 0 The proof that xa and x−c are diﬀerentiable when x = 0 is trivial. the product rule). therefore we have conﬂicting deﬁnitions of f (0). but Rudin asks us in example 5. case 1: f (0) exists when a > 1 When a > 1 we have a − 1 > 0 and therefore taking the limits of (24) as h → 0 gives us 0 ≤ lim |ha−1 sin(h−c )| ≤ 0 h→0 which means that (23) becomes f (0) = lim ha−1 sin(h−c ) = 0 h→0 which shows that f (0) is deﬁned. case 3: f (0) does not exist when a < 1 Deﬁne the sequences {hn } and {jn } such that hn = 1 2nπ + 1/c π 2 1/c π 2 jn = 1 (2n + 1)π + When a < 1 we have a − 1 < 0 and therefore equation (23) gives us h→0 lim ha−1 sin(h−c ) = lim ha−1 = ∞ n n→∞ n→∞ j→0 a−1 lim j a−1 sin(j −c ) = lim −jn = −∞ We know the sequences {f (hn )} and {j (hn )} are well-deﬁned because of lemma 1. These cases show that f (0) exists iﬀ a > 1. The sin function hasn’t been well-deﬁned yet.5.6 to assume that it’s diﬀerentiable. case 2: f (0) does not exist when a = 1 Deﬁne the sequences {hn } and {jn } such that hn = 1 2nπ + 1/c π 2 1/c π 2 jn = 1 (2n + 1)π + When a = 1 we have a − 1 = 0 and therefore equation (23) gives us h→0 lim ha−1 sin(h−c ) = lim sin(h−c ) = 1 n n→∞ n→∞ j→0 −c lim j a−1 sin(j −c ) = lim sin(jn ) = −1 We know the sequences {f (hn )} and {j (hn )} are well-deﬁned because of lemma 1. therefore we have conﬂicting deﬁnitions of f (0). 79 . the chain rule) and therefore xa sin(x−c ) is diﬀerentiable when x = 0 (theorem 5. Therefore sin(x−c ) is diﬀerentiable when x = 0 (theorem 5.

so by the lemma of part (b) we know that f (x) is deﬁned for all x ∈ [1. so we only need to show that f is bounded on this domain. 1] including x = 0. 80 . but it does show that f (x) is unbounded near 0. We could ﬁnd stricter bounds for f (x). By the chain rule and product rule we know that the derivative of f is f (x) = axa−1 sin(x−c ) − cxa−(c+1) cos(x−c ) Since f (x) is deﬁned for x = 0. 1] we know that xa−1 . Therefore the rightmost limit of the previous equation is bounded by −(a + c) ≤ lim axa−1 sin(x−c ) − cxa−(c+1) cos(x−c ) ≤ a + c x→0 Which. xa−(c+1) . These three cases show that f is bounded iﬀ a ≥ c + 1. sin. and cos are all bounded by [−1.13d lemma: f is continuous when x = 0 From lemma 1 of part (b) we know that f exists for all x = 0 and its derivative is given by f (x) = axa−1 sin(x−c ) − cxa−(c+1) cos(x−c ) (27) Rudin asks us to assume that sin and cos are continuous functions and it’s trivial to show that x±α is continuous when x = 0 for any α. a + c] for x ∈ [−1. 1]. 1] except for possibly x = 0.13c Note that we’ve only deﬁned f on the domain [−1. but it’s not necessary. Exercise 5.5) and product rule (theorem 5. we can take the limit of (26) as x → 0: f (0) = lim f (x) = lim axa−1 sin(x−c ) − cxa−(c+1) cos(x−c ) x→0 x→0 (26) Because x is bounded by [−1. 1]. of course. 1]. so we can use the chain rule (theorem 5. case 3: f is bounded when a ≥ c + 1 If a ≥ c + 1 then clearly a > 1. so taking the limits of this last equation as n → ∞ gives us n→∞ lim f (hn ) = lim −c(2nπ) n→∞ (1+c)−a c = −∞ This doesn’t prove anything about f (0) itself.3b) to show that f is continuous when x = 0. means that f (x) is also bounded by [−(a + c). By the chain rule and product rule we know that the derivative of f when x = 0 is f (x) = axa−1 sin(x−c ) − cxa−(c+1) cos(x−c ) Deﬁne the sequence {hn } such that hn = (2nπ)−1/c (25) Evaluating the derivative in (25) at x = hn gives us f (hn ) = (2nπ) 1−a c sin(2nπ) − c(2nπ) (1+c)−a c cos(2nπ) = −c(2nπ) (1+c)−a c We’re assuming in this case that a < c + 1.Exercise 5. case 2: f is unbounded when 1 ≤ a < c + 1 By the lemma of part (b) we know that f (x) is deﬁned for all x ∈ [1. case 1: f is unbounded when a < 1 We saw in case (3) of part (b) that f is unbounded near 0 when a < 1.

5 (the chain rule) and theorem 5. For an alternative proof we could use the sequence {xn } established in case 2 and show that f (xn ) → ∞ = f (0) as xn → 0 when a < 1 + c. case 1: f (0) exists when a > 2 + c From the deﬁnition of limit we know that f (0) = lim aha−1 sin(h−c ) − cha−(c+1) cos(h−c ) f (0 + h) − f (0) = lim h→0 h→0 h h = lim (aha−2 ) sin(h−c ) − (cha−(c+2) ) cos(h−c ) h→0 (29) 81 . |f (x)| = |xa−(c+1) axc sin(x−c ) − c cos(x−c )| ≤ |xa−(c+1) | · |axc + c| ≤ |xa−(c+1) | · (|axc | + |c|) Because a > c + 1. we have |xa−(c+1) | → 0 and |axc | → 0 as x → 0. Taking the limits of this last inequality as x → 0 therefore gives us lim |f (x)| ≤ 0 · (0 + c) = 0 x→0 This shows that limx→0 f (x) = f (0). For f to be continuous at x = 0 it must be the case that limx→0 f (x) = 0.15).3(the product rule) to show that f is diﬀerentiable when x = 0. but sin(xn ) = 0 and cos(xn ) = 1 so that the terms of {f (xn )} are f (xn ) = axn (0) − cx0 (1) = −c so that lim{f (xn )} = −c = f (0). 1] (part (c)) therefore f is not continuous on [−1. therefore f is continuous at x = 0. From lemma 1. case 3: f is not continuous at x = 0 when a < 1 + c If a < 1+c we know that f is not bounded on [−1. Exercise 5.case 1: f is continuous at x = 0 when a > 1 + c We’ve shown that f (0) = 0 when a > 1 (case 1 of part (b)). case 2: f is not continuous at x = 0 when a = 1 + c To show that f is not continuous at x = 0 it’s suﬃcient to construct a sequence {xn } such that lim xn = 0 but lim f (xn ) = f (0) = 0 (theorem 4. 1]. Deﬁne the terms of xn to be xn = 1 2nπ 1 c This sequence clearly has a limit of 0. so we can use theorem 5. so we can establish a bound on this function.2).13e Lemma 1: f is diﬀerentiable when x = 0 We established in part (b) that f exists for x = 0 and is given by f (x) = axa−1 sin(x−c ) − cxa−(c+1) cos(x−c ) (28) We know that all of the exponential powers of x are diﬀerentiable when x = 0 and Rudin asks us to assume that sin and cos are diﬀerentiable. We can algebraically rearrange (27) to obtain f (x) = xa−(c+1) axc sin(x−c ) − c cos(x−c ) The range of the cosine and sine functions are [−1. we know that the discontinuity must occur at the point x = 0. 1] (theorem 4.

82 . Exercise 5. therefore we have conﬂicting deﬁnitions of f (0). the powers of h tend tend to zero as h → 0. case 2: f (0) does not exist when a = 2 + c Deﬁne the sequences {hn } and {jn } such that hn = 1 2nπ 1/c jn = 1 (2n + 1)π 1/c When a = 2 + c we have a − (c + 2) = 0 and therefore equation (28) gives us h→0 lim f (h) = lim f (hn ) = 0 − (ch0 )(1) = −c n n→∞ j→0 0 lim f (j) = lim f (jn ) = 0 − (cjn )(−1) = c n→∞ We know the sequences {f (hn )} and {j (hn )} are well-deﬁned because of lemma 1. so I’ll conclude with “the proof is similar to that of part(c)”. we have 0 ≤ lim aha−2 sin(h) − cha−(c+2) cos(h−c ) ≤ 0 h→0 This shows that the limit in (29) (and therefore f (0)) exists.13f Lemma 1: to hell with this By the lemma of part (e) we know that f (x) is deﬁned for all x ∈ [1. This means that equation (28) gives us lim f (h) = lim f (hn ) = 0 − (cha−(c+2) )(1) = −∞ n a−(c+2) lim f (j) = lim f (jn ) = 0 − (cjn )(−1) = ∞ n→∞ j→0 We know the sequences {f (hn )} and {j (hn )} are well-deﬁned because of lemma 1. These three cases show that f (0) exists iﬀ a > 2 + c. therefore we have conﬂicting deﬁnitions of f (0). 1] except for possibly x = 0. This means that the limit in (28) (and therefore f (0) itself) does not exist. Taking the limits of the previous inequality as h → 0. −(|aha−2 | + |cha−(c+2) |) ≤ aha−2 sin(h−c ) − cha−(c+2) cos(h−c ) ≤ |aha−2 | + |cha−(c+2) | When a > (2 + c) > 2. This means that the limit in (28) (and therefore f (0) itself) does not exist. 1] so we can establish bounds for the limited term. By the chain rule and product rule we know that the derivative of f when x = 0 is f (x) = (a2 − a)xa−2 − c2 xa−(2+2c) sin(x−c ) + (c2 + c − ca)xa−(2+c) − caxa−(1+c) cos(x−c ) (30) And I’m not going to screw around with limits and absolute values of something so annoying to type out.The range of the sin and cos functions is [−1. case 2: f (0) does not exist when a < 2 + c Deﬁne the sequences {hn } and {jn } such that hn = 1 2nπ 1/c jn = 1 (2n + 1)π a−(c+2) 1/c When a < 2 + c we have a − (c + 2) < 0 and therefore hn h→0 n→∞ → ∞.

The “if ” case: f is monotonically increasing if f is convex Let f be a function that is convex on the interval (a. b) and that f is not convex. b) then it is strictly concave on some interval (s. y ∈ (a. Exercise 5. d. We can now claim that equation (31) holds for all λ ∈ (α. b) and λ ∈ (0. 1] such that g(p) > 0 because we can choose p = λ which causes the previous equation to simplify to g(p) = f (λc + (1 − λ)d) − [λf (c) + (1 − λ)f (d)] which is > 0 from (31). By deﬁnition this means that for all x. means that the convexity conditions fails on the inteval (s. b) (theorem 4. And we know that there is at least one p ∈ [0.14 Lemma 1: If f is not convex then the convexity condition fails for all λ for some s. 1) such that f (λc + (1 − λ)d) > λf (c) + (1 − λ)f (d) (31) Having ﬁxed our choice of c. we deﬁne the function g(x) to be g(x) = f ((x)c + (1 − x)d) − (x)f (c) + (1 − x)f (d) We know that g is continuous on [0. y ∈ (a. Let Z1 be the set of all p ∈ [0. λ) for which g(p) = 0 and let Z2 be the set of all p ∈ (λ. d ∈ (a. and the sentiment is similar to that of lemma (1) of part (f).3) and therefore contain their supremums and inﬁmums. β).Exercise 5.23) and choose x.13g The proof is similar to that of part (d). d] ⊂ (a. b) for all λ ∈ (0. t = βc + (1 − β)d which. b) with y ≥ x. and λ so that the previous equation is true. 1) for s = αc + (1 − α)d. These sets are closed (exercise 4. b) This could also be stated as “if f is not convex on (a. And this is the same as saying that f (λs + (1 − λ)t) > λf (s) + (1 − λ)f (t) for all λ ∈ (0. By deﬁnition. b) and for all λ ∈ (0. t) ⊂ (a. t) ∈ (a. b) (see exercise 4. 1) it must be the case that f (λx + (1 − λ)y) ≤ λf (x) + (1 − λ)f (y) (32) We want to express the left-hand side of this inequality as f (x + h): we can do this by deﬁning h such that h = (λ − 1)(x − y) Rearranging this algebraically allows us to express λ and 1 − λ as λ=1− h . 1] because f is continuous on [c. 1] for which g(p) = 0 It’s immediately clear that these sets are nonempty since g(0) = g(1) = 0.9). b)”. t ∈ (a. 1). Let α = sup{Z1 } and let β = inf{Z2 }. y−x (1 − λ) = h x−y Substituting these values of λ and λ − 1 into (32) results in f (x + h) ≤ f (x) − which is algebraically equivalent to f (x + h) − f (x) f (y) − f (x) ≤ h y−x h f (x) h f (y) + y−x y−x 83 . by deﬁnition. Assume that f is continuous on the interval (a. this means that we can choose c.

this means that f is monotonically increasing if f is convex. By deﬁnition. Taking the limit of both sides of the previous equation as h → 0 gives us f (s) > f (t) − f (s) t−s 84 (36) h f (s) h f (t) + t−s t−s . y−x (1 − λ) = 1 − h x−y Substituting these values of λ and 1 − λ into (32) results in f (y − h) ≤ which is algebraically equivalent to f (y) − f (y − h) f (y) − f (x) ≥ h y−x Equation (32) had to be true for any value of λ ∈ (0. we concluded that f (y) ≥ f (x). 1). We can now follow the logic of the “if” case and deﬁne h = (λ − 1)(s − t) Rearranging this algebraically allows us to express λ and 1 − λ as λ=1− h . we can ﬁnd some subinterval (s. t−s (1 − λ) = h s−t (35) Substituting these values of λ and λ − 1 into (35) results in f (s + h) > f (s) − which is algebraically equivalent to f (s + h) − f (s) f (t) − f (t) > h t−s Equation (35) had to be true for any value of λ ∈ (0. t) ∈ (a. As λ → 0 we see that h → 0. 1).Equation (32) had to be true for any value of λ ∈ (0. The “only if ” case: f is monotonically increasing only if f is convex Assume that f is not convex. b) such that f (λs + (1 − λ)t) > λf (s) + (1 − λ)f (t) is true for all λ ∈ (0. we have have shown that f (y) ≥ f (x) We assumed only that f was convex and that y > x. As λ → 1 we see that h → 0. By lemma 1. we now want to express the left-hand side of (32) as f (y − h): we can do this by redeﬁning h such that h = −(λ)(x − y) Rearranging this algebraically allows us to express λ and 1 − λ as λ= h . Taking the limit of both sides of the previous equation as h → 0 gives us f (x) ≤ f (y) − f (x) y−x (33) Having established (33). 1). As λ → 1 we see that h → 0. 1). Taking the limit of both sides of the previous equation as h → 0 gives us f (y) ≥ f (y) − f (x) y−x (34) h f (x) h f (y) + f (y) − y−x y−x Combining equations (33) and (34).

As λ → 0 we see that h → 0. Otherwise the function f (x) = x would be a counterexample 2 to the claim that M1 ≤ 4M0 M2 . 2 ξ>a (38) f (x + 2h) f (x) f (x + 2h) f (x) − − hf (ξ)| ≤ | |+| | + |hf (ξ)| 2h 2h 2h 2h We’re given upper bounds for |f (x)| and |f (x)|. We assumed only that f was not convex.15) we can express f (x + 2h) as f (x + 2h) = f (x) + 2hf (x) + which can be algebraically arranged to give us f (x) = |f (x)| = | f (x + 2h) f (x) − − hf (ξ) 2h 2h 4h2 f (ξ) . and theorem 5. By contrapositive. we now redeﬁne h such that h = −(λ)(s − t) Rearranging this algebraically allows us to express λ and 1 − λ as λ= h . Exercise 5. 1). b). and M2 it appears that we must assume that these bounds are ﬁnite. this means that f is monotonically increasing only if f is convex. we have have shown that f (t) < f (s) for some t > s. t−s (1 − λ) = 1 − h s−t Substituting these values of λ and 1 − λ into (35) results in f (t − h) > which is algebraically equivalent to f (t) − f (t − h) f (t) − f (s) < h t−s Equation (35) had to be true for any value of λ ∈ (0. f is convex iﬀ f (x) ≥ 0 for all x ∈ (a. we concluded that f was not monotonically increasing.11 tells us that f is monotonically increasing iﬀ f (x) ≥ 0 for all x ∈ (a. b) We’ve shown that f is convex iﬀ f is monotonically inreasing. M1 .Having established (36). these bounds give us the inequality |f (x)| ≤ 2M0 + hM2 h 85 . Taking the limit of both sides of the previous equation as h → 0 gives us f (t) < f (t) − f (s) t−s (37) h f (s) h f (t) + f (t) − t−s t−s Combining equations (36) and (37). Using Taylor’s theorem (theorem 5.15 Note on the bounds of f and its derivatives When Rudin asks us to assume that |f | and its derivatives have upper bounds of M0 . 2 Proof that M1 ≤ 4M0 M2 for real-valued functions Choose any h > 0.

This inequality must hold for all x. f2 (x + 2h). . and M2 . . . . so we have M1 ≤ 2M0 + hM2 h We can multiply both sides by h and then algebraically rearrange this into a quadratic equation in h. . .15 we established the inequality |f (x)| ≤ | f (x) f (x + 2h) |+| | + |hf (ξ)| 2h 2h We are given an upper bound of M2 for |f (ξ)|. ∞) and is deﬁned by f (x) = ( f1 (x). lim |f (x)| ≤ lim | f (x + 2h) f (x) |+| | + hM2 2h 2h 86 x→∞ x→∞ . even when |f (x)| approaches its upper bound. But nothing in the proof following equation (38) requires us to assume that f is real-valued instead of vector-valued. fn (x) ) Assume that |f |. . we have f (x + 2h) = = [f1 (x) + 2hf1 (x) + 2h2 f1 (ξ1 )]. so we can take the limit of both sides of this equation as x → 0 (theorem 4. This occurs exactly when 2 M1 ≤ 4M0 M2 Does this apply to vector-valued functions? Yes. 2hfn (x) ) + 2h2 f1 (ξ).16 proof 1 In exercise 5. . . . f2 (x). . . If we evaluate f at x + 2h we have f (x + 2h) = ( f1 (x + 2h). and if (40) had two solutions then we would have h2 M2 − hM1 + M0 < 0 on some interval of h. . . fn (x) ) + ( 2hf1 (x). To make sure that there is at most a single real solution we need to make sure that the discriminant of (40) is either zero (one solution) or negative (zero solutions). Let f be a vector-valued function that is continuous on (a. . M1 . [fn (x) + 2hfn (x) + 2h2 fn (ξ1 )] ( f1 (x). so we know that both f and f are continuous (theorem 5. and |f | have ﬁnite upper bounds of (respectively) M0 . f2 (x). . 2h2 fn (ξ) = f (x) + 2hf (x) + 2h2 f (ξ) This tells us that equation (38) holds for vector-valued functions. [f2 (x) + 2hf2 (x) + 2h2 f2 (ξ1 )]. .2). So the proof following (38) still suﬃces to prove that 2 M1 ≤ 4M0 M2 Exercise 5. fn (x + 2h) ) Taking the Taylor expansion of each of these terms. h2 M2 − hM1 + M0 ≥ 0 The quadratic solution to this equation is h= M1 ± 2 M1 − 4M0 M2 2M2 (39) (40) We want to make sure that there are not two solutions to (40): we want (39) to hold for all values of h. . . . 2hf2 (x). . . . . 2h2 f2 (ξ).2). |f |. so this inequality can be expressed as |f (x)| ≤ | f (x + 2h) f (x) |+| | + hM2 2h 2h We’re told that f is twice-diﬀerentiable.

To show more explicitly that the value of these M terms depends on a. ∞). The last inequality therefore allows us to conclude that a→∞ x>a lim sup |f (x)|2 ≤ lim sup 4|0||M2 | = 0 a→∞ x>a It’s clear that |f (x)| must be less than or equal to the supremum of a set containing |f (x)|. we might express the previous inequality as sup |f (x)|2 ≤ sup 4|f (x)||f (x)| x>a x>a Each of these terms is continuous with respect to a (the proof of this claim is trivial but tedious) so we can take the limit of both sides as a → ∞. a→∞ x>a lim sup |f (x)|2 ≤ lim sup 4|f (x)||f (x)| a→∞ x>a We’re told that |f (x)| has an upper bound of M2 . the right-hand side becomes 0. lim |f (x)| = lim lim |f (x)| ≤ lim lim hM2 = 0 x→∞ h→0 x→∞ x→∞ h→0 x→∞ lim |f (x)| ≤ 0 This show that f (x) → 0 as x → ∞.15 we established the inequality 2 M1 ≤ 4M0 M2 where each of the M terms represented a supremum on the interval (a. h→0 x→∞ lim lim |f (x)| ≤ lim lim hM2 h→0 x→∞ The left-hand side of the previous inequality is independent of h and therefore doesn’t change. so we have x→∞ lim |f (x)| ≤ lim sup |f (x)|2 ≤ 0 a→∞ x>a x→∞ lim |f (x)| ≤ 0 This shows that f (x) → 0 as x → ∞. 2 6 ξ1 ∈ (−1. which is what we were asked to prove. 2 6 f (0)(1)2 f (ξ2 )(1)3 + . Exercise 5.We’re told that f (x) → 0 as x → 0 so this becomes x→∞ lim |f (x)| ≤ lim hM2 x→∞ This must be true for all h. 1) so we can take the Taylor expansions of f (−1) and f (1) around x = 0: f (−1) = f (0) + f (0)(−1) + f (ξ1 )(−1)3 f (0)(−1)2 + . ∞). We’re also told that f (x) → 0 as x → ∞.17 We’re told that f is three-times diﬀerentiable on the interval (−1. Exercise 5. so we can take the limit of both sides of this as h → 0 (we can do this because both |f (x)| and hM2 are continuous functions with respect to the variable h). 0) ξ2 ∈ (0. which is what we were asked to prove.16 proof 2 In exercise 5. 1) f (1) = f (0) + f (0)(1) + 87 . This inequality was proven to hold for all a such that f was twice-diﬀerentiable on the interval (a.

6 ξ1 ∈ (−1.18 The nth derivative of f (t) The exercise tells gives us the following formula for f (t): f (t) = f (β) − (β − t)Q(t) Since β is a ﬁxed constant. we have n−1 P (β) = f (α) + Q(α)(β − α) + k=2 Q(k−1) (α) (β − α)k − (k − 1)! 88 n−2 k=1 Q(k) (α) Q(n−1) (α) (β − α)k+1 − (β − α)n k! (n − 1)! . we have f (n) (t) = nQ(n−1) (t) − (β − t)Q(n) (t) which. And this is what we were asked to prove.When we evaluate f (1) − f (−1). 0)ξ2 ∈ (0.15) includes a P (β) term deﬁned as n−1 f (k) (α) P (β) = (β − α)k k! k=0 (41) If we isolate the k = 0 case this becomes n−1 P (β) = f (α) + k=1 f (k) (α) (β − α)k k! We can now use equation (41) to express the terms of the summation as functions of Q. 6 ξ1 ∈ (−1. the ﬁrst two derivatives of f with respect to t are f (t) = Q(t) − (β − t)Q (t) f (t) = 2Q (t) − (β − t)Q (t) And. f (−1). 1) We’re given the values of f (1). 1) If f (ξ1 ) ≤ 3 then f (ξ2 ) ≥ 3. and vice-versa: so f (x) ≥ 3 for either ξ1 or ξ2 ∈ (−1. becomes Q(n−1) (t) Q(n) (t) f (n) (t) (β − α)n = (β − α)n − (β − α)n+1 n! (n − 1)! n! Modifying the Taylor formula The formula for the Taylor expansion of f (β) around the point f (α) (theorem 5. ξ1 ∈ (−1. 1) f (ξ1 ) + f (ξ2 ) . and f (0) so this last equation becomes 1= which is algebraically equivalent to f (ξ2 ) = 6 − f (ξ1 ). 0)ξ2 ∈ (0. in general. many of these terms cancel out and we’re left with f (1) − f (−1) = 2f (0) + f (ξ1 ) + f (ξ2 ) . after multiplying by (β − α)n /n! and setting t = α. 0)ξ2 ∈ (0. Exercise 5. 1). n−1 P (β) = f (α) + k=1 Q(k−1) (α) (β − α)k − (k − 1)! n−1 k=1 Q(k) (α) (β − α)k+1 k! If we extract the k = 1 term from the leftmost summation and the k = n−1 term from the rightmost summation.

Therefore we’re justiﬁed in taking the limits in (42). giving us βn −αn lim Dn = f (0) · lim + f (0) · lim n→∞ n→∞ βn − αn n→∞ βn − αn = lim f (0) n→∞ βn − αn = f (0) βn − αn 89 . Therefore condition 1 is met.2) and therefore condition 2 is met (theorem 4.19a The given expression for Dn is algebraically equivalent to Dn = f (βn ) − f (0) βn f (0) − f (αn ) −αn + βn − 0 βn − α n 0 − αn βn − α n We really want to be able to evaluate this by taking the limit of each side of this previous equation in the following manner: lim Dn = lim f (βn ) − f (0) βn f (0) − f (αn ) −αn · lim + lim · lim n→∞ βn − αn n→∞ n→∞ βn − αn βn − 0 0 − αn (42) n→∞ n→∞ There two conditions that must be met in order for this last step to be justiﬁed: condition 1: theorem 3. Q(n−1) (α) (β − α)n (n − 1)! Q(n−1) (α) (β − α)n (n − 1)! f (α) − f (β) Q(n−1) (α) (β − α) − (β − α)n α−β (n − 1)! Exercise 5. The other two limits exist because they’re equal to f (0).2). and we’re told that f (0) exists. The fact that f (0) exists tells us that f is continuous at x = 0 (theorem 5. we see that this previous equation evaluates to P (β) = f (α) + This simpliﬁes to P (β) = f (α) + (f (β) − f (α)) − A simple algebraic rearrangement of these terms gives us f (β) = P (β) − which is the equation we were asked to derive. which tells us that at 2 of the 4 limits in (42) exist.We can then re-index the leftmost summation to obtain n−2 P (β) = f (α) + Q(α)(β − α) + k=1 Q(k) (α) (β − α)k+1 − (k)! n−2 k=1 Q(n−1) (α) Q(k) (α) (β − α)k+1 − (β − α)n k! (n − 1)! The two summations cancel one another. leaving us with P (β) = f (α) + Q(α)(β − α) − Q(n−1) (α) (β − α)n (n − 1)! If we replace Q(α) with the deﬁnition of Q given in the exercise.2 tells us that we must have lim f (αn ) = lim f (βn ) = f (0).3 tells us that each of the limits must actually exist (and must not be ±∞) condition 2: and theorem 4. The fact that αn < 0 < βn guarantees that 0 < βn /(βn − αn ) < 1 and 0 < −αn /(βn − αn ) < 1.

βn ). 1) and that lim αn = 0 so by theorem 4. We’re also told that f is continuous on (−1. giving us lim Dn = f (0) · lim = f (0) · lim = f (0) βn βn + f (0) · lim 1 − n→∞ βn − αn βn − αn βn βn +1− βn − αn βn − αn n→∞ n→∞ n→∞ Exercise 5. We’re told that f is continuous so by theorem 5. The other two limits exist because they’re equal to f (0). and we’re told that f (0) exists. βn ) such that f (βn ) − f (αn ) = f (γn ) (44) Dn = βn − αn Each γn is in the interval (αn . We’re told that lim αn = lim βn = 0.19c proof 2 The mean value theorem (theorem 5. 1). choose some γn ∈ (αn . Therefore we have lim Dn = = lim lim f (αn + hn ) − f (αn ) hn lim f (αn ) n→∞ n→∞ h→0 n→∞ We’re told that f is continuous on the interval (−1. Therefore condition 1 is met. Therefore we’re justiﬁed in taking the limits in (43). which tells us that at 2 of the 4 limits in (43) exist.2 we have n→∞ lim Dn = lim f (αn ) = f (0) n→∞ Exercise 5. lim Dn = lim f (γn ) = f (0) n→∞ n→∞ 90 .2).2 we can take the limit of equation (44). so we know that f is deﬁned on this interval.19b The given expression for Dn is algebraically equivalent to Dn = f (βn ) − f (0) βn f (αn ) − f (0) + βn − 0 βn − αn αn − 0 1− βn βn − αn We want to evaluate this by taking the limits of the individual terms as we did in part (a): lim Dn = lim f (βn ) − f (0) βn f (αn ) − f (0) · lim + lim n→∞ βn − αn n→∞ βn − 0 αn − 0 lim 1− βn βn − αn (43) n→∞ n→∞ n→∞ In order for this to be justiﬁed we must once again meet the two conditions mentioned in part (a). 0 = lim αn ≤ lim γn ≤ lim βn = 0 n→∞ n→∞ n→∞ which means that lim γn → 0. Therefore we can use the squeeze theorem to determine lim γn .2) and therefore condition 2 is met (theorem 4.Exercise 5. We’re told that 0 < βn /(βn − αn ) < M for some M . We can now express Dn as Dn = f (αn + hn ) − f (αn ) hn We know that αn → 0 and βn → 0 as n → ∞. so clearly hn → 0 as n → ∞.10) allows us to construct a sequence {γn } as follows: for each n. The fact that f (0) exists tells us that f is continuous at x = 0 (theorem 5.19c proof 1 Deﬁne the sequence {hn } where hn = βn − αn .

x ∈ (ai . bi ) ∧ bi = ∞ (x − ai )n+1 . poorly-deﬁned Calc II-ish way that f is inﬁnitely diﬀerentiable and that. To calculate the limit of f (x) in the ﬁrst case. Looking at the derivative with respect to x we have x∈E 0. University of South Florida 91 . f (n) (x) = 0 iﬀ x ∈ E. x ∈ (ai . Let {(an . (x − ai )n+1 (x − bi )n+1 . bi ) ∧ ai = −∞ (x−bi )2 . Similar results hold for the limits in the other two cases. Deﬁne f : R1 → R1 as x∈E 0.29 that any open set in R1 can be represented as a countable number of disjoint open segments. x ∈ (ai . bn )} represent such a countable collection of disjoint sets such that (ai . −2(bi −ai ) −1 x ∈ (ai . It’s fairly easy to conﬁrm that this function is continuous and that f (x) = 0 iﬀ x ∈ E. bi ) ∧ bi = ∞ (x−ai )3 exp (x−ai )2 . bi ) ∧ ai = −∞ f (x) = (x − ai )2 . Therefore p(n)exp(−1/n) will tend to zero as n → ∞. Note that the derivative also has the property that f (x) = 0 iﬀ x ∈ E. bi ) = E C . we can pretend that we’ve deﬁned the exponential function and deﬁne f to be x∈E 0. Every term of every derivative of f (x) will consist only of polynomial multiples of the exponential term. 6 this example was provided by Boris Shekhtman. (x − bi )2 . so L’Hopital’s rule is applicable. f (x) = −1 x ∈ (ai . I have no idea how to prove this.20 Exercise 5. bi ) ∧ −∞ < ai < bi < ∞ (x−ai )2 (x−bi )2 exp (x−ai )2 (x−bi )2 . the exponential limit limn→∞ exp(−1/n) will tend towards zero faster than limn→∞ p(n) tends towards ∞. bi ) ∧ ai = −∞ (x−bi )3 exp (x−bi )2 . −1 exp x ∈ (ai . we use L’Hopital’s rule. x ∈ (ai . so we see that lim f (x) = 0. −1 exp x ∈ (ai . bi ) ∧ −∞ < ai < bi < ∞ (x−ai )2 (x−bi )2 .21 We saw in exercise 2. so it will hold that for all k: lim f (k) (x) = lim f (k) (x) = 0 x→ai x→bi It seems clear in a vague. bi ) ∧ ai = −∞ f (x) = x ∈ (ai . bi ) ∧ bi = ∞ exp (x−ai )2 . exp x→a −1 (x−bi )2 lim f (x) = lim x→a (x − bi )3 The numerator and denominator of this last term both tend to ±∞. for all n. bi ) ∧ −∞ < ai < bi < ∞ It’s easy but tedious to verify that this a continuous function which is diﬀerentiable and that f (x) = 0 iﬀ x ∈ E. −2 −1 x ∈ (ai . Repeated applications of L’Hopital’s rule will eventually give us a constant term in the numerator and a term that tends to ±∞ in the denominator. bi ) ∧ −∞ < ai < bi < ∞ It’s similarly easy but tedious to verify that this is a continuous function that is n times diﬀerentiable and that f (x) = 0 iﬀ x ∈ E. Finally6 . It can also be seen that when k ≤ n we have f (k) (x) = 0 iﬀ x ∈ E.Exercise 5. for any polynomial term p(n). (x − bi )n+1 . x ∈ (ai . For the second part of the exercise we can deﬁne a function f to be x∈E 0. bi ) ∧ bi = ∞ (x − ai )2 (x − bi )2 . f (x) = −2 −1 x ∈ (ai . The general idea is that.

We’re told that |f (t)| ≤ A for all real t. Exercise 5. Then f (x2 ) − f (x1 ) x2 − x1 = =1 x2 − x1 x2 − x1 By theorem 5.11 we know that the sequence converges to some element x ∈ R1 .Exercise 5. So: |xn − xn−1 | ≤ A|xn−1 − xn−2 | ≤ A2 |xn−2 − xn−3 | ≤ A3 |xn−3 − xn−4 | ≤ · · · ≤ An−2 |x2 − x1 | We can use this general formula to determine the diﬀerence between xn+k and xn .22c Choose an arbitrary value for x0 and let {xn } be the sequence recursively deﬁned by xn+1 = f (xn ). ∃x ∈ R : lim xn = x n→∞ But the elements of {xn } are just the elements of {f (xn )}. This contradicts the presumption that f (x) = 1. so our initial assumption must be wrong: there cannot be two ﬁxed points.22b For f (t) to have a ﬁxed point it would be necessary that f (t) = t. since f (xn ) = xn+1 .22a Assume that f (x1 ) = x1 and f (x2 ) = x2 with x1 = x2 . And we’re converging in R1 so by theorem 3. −→ 1 =0 1 − et Exercise 5. so we can also conclude that ∃x ∈ R : lim f (xn ) = x n→∞ And we’re told that f is continuous. so taking the limit of eachs ide of this inequality as n → ∞ gives us n→∞ lim |xn+k − xn | ≤ lim An |x2 − x1 | = 0 n→∞ And this is just the Cauchy criterion for convergence for the sequence {xn }. so by theorem 4.6 we see that ∃x ∈ R : f (x) = x 92 . |xn+k − xn | = |(xn+k − xn+k−1 ) + (xn+k−1 − xn+k−2 ) + · · · + (xn+2 − xn+1 ) + (xn+1 − xn )| ≤ ≤ |xn+k − xn+k−1 | + |xn+k−1 − xn+k−2 | + · · · + |xn+2 − xn+1 | + |xn+1 − xn )| An+k−2 |x2 − x1 | + An+k−3 |x2 − x1 | + · · · + An |x2 − x1 | + An−1 |x2 − x1 )| −→ |f (xn−1 ) − f (xn−2 )| ≤ A |xn−1 − xn−2 | ≤ An |x2 − x1 | This shows us that |xn+k − xn | ≤ An |x2 − x1 | We’re told that A < 1.10 this means that f (x) = 1 for some x between x1 and x2 . So for any n we have f (xn−1 ) − f (xn−2 ) ≤A xn−1 − xn−2 which. gives us |xn − xn−1 | ≤ A|xn−1 − xn−2 | But this holds for all n. in which case we would have t = t + (1 − et )−1 This statement is not true for any t.

Now let xk be chosen from the interval (α. Therefore it converges to a some ﬁxed point in the interval (x1 . γ). 0) α is a ﬁxed point. Exercise 5. so αx > 1 and therefore αδx > δ. 1). 1). Case 1: xk ∈ (α. Theredeﬁnition of xk f is concave on (α.23a If x < α.11). We can express this x as xk = λα + (1 − λ)0 for some λ ∈ (0. so we know that f is monotonically increasing (theorem 5. γ) and is strictly concave on (α. γ) and so the function is convex on this interval (exercise 5. This is nonnegative. This gives us f (x) = f (α − δ) deﬁnition of x = = = (α−δ)3 −1 3 α3 −3α2 δ+3αδ 2 −δ 3 +1 3 −3α2 δ+3αδ 2 α3 +1 + 3 3 deﬁnition of f algebra algebra algebra deﬁnition of x α is a ﬁxed point = f (α) − αδ(α − δ) = f (α) − αδx − 1 δ 3 3 = α − αδx − 1 3 3δ 3 − δ3 − 1 δ3 3 We know that x < α < −1.23b The ﬁrst derivative of f is f (x) = x2 . The second derivative of f (x) = 2x is negative on the interval (α. from our choice of xk we have α < xk . Therefore we have xk+1 = f (xk ) deﬁnition of xk+1 = f (λα + (1 − λ)0) > λf (α) + (1 − λ)f (0) = λα + (1 − λ)(1/3) = xk + (1 − λ)(1/3) > xk We see that xk < xk+1 . 0) Let xk ∈ (α.Exercise 5. then we can express x as x = α − δ for some δ > 0. f (0) = 1/3 deﬁnition of xk 93 . We can express this as xk = λβ + (1 − λ)γ for some λ ∈ (0. 0) be chosen. The second derivative of f (x) = 2x is positive on the interval (β. γ) be chosen. We know that this sequence doesn’t converge because we’re told that f has no ﬁxed points less than α. beta]. and this point must be β. Combining these inequalities yields α < xk < xk+1 < β Therefore our initial choice of xk gives us an increasing sequence {xn } with an upper bound of β. Case 2: x ∈ (β. 0).14). The second derivative is f (x) = 2x. 0) and so the function is concave on this interval (corollary of exercise 5. γ) Let xk ∈ (β. Therefore if x1 < α then the sequence {xn } will be a nonincreasing sequence. from the monotonic nature of f we have xk < β → xk+1 = f (xk ) < f (β) = β.14). From this we have an inequality: 1 < α − δ − 3 δ3 <α−δ =x deﬁnition of x This establishes that x < α → f (x) < x. This is negative for x < 0 and positive for x > 0: therefore f is strictly convex on (0.

f (0) = 1/3 deﬁnition of xk > xk We see that xk < xk+1 . 1). Exercise 5. We can express this as xk = λ0 + (1 − λ)β for some λ ∈ (0. and this point must be β. Case 3: x ∈ (0. so γxδ > δ. β) Let xk ∈ (0. from our choice of xk we have xk < γ. Therefore it converges to a some ﬁxed point in the interval [β. γ) β and γ are ﬁxed points = xk deﬁnition of xk We see that xk+1 < xk . so this sequence clearly converges to β. If xk = 0 then xk+1 = f (0) = (1/3) and the remainder of the sequence {xn } converges to β by one of the previous cases. and this point must be β. from the monotonic nature of f we have xk < β → xk+1 = f (xk ) < f (β) = β. Case 4: xk = β or xk = 0 If xk = β then every term of {xn } is β. from the monotonic nature of f we have β < xk → β = f (β) < f (xk ) = xk+1 . Therefore it converges to a some ﬁxed point in the interval (x1 . From this we have an inequality : 1 > γ + δ + 3 δ3 >γ+δ =x deﬁnition of x 94 . from our choice of xk we have α < xk . Therefore we have xk+1 = f (xk ) deﬁnition of xk+1 = f (λ(0) + (1 − λ)β) < λf (0) + (1 − λ)f (β) = λ(1/3) + (1 − λ)β = λ(1/3) + x deﬁnition of xk deﬁnition of convexity β is a ﬁxed point. β) and so the function is convex on this interval (exercise 5. β) be chosen.14). Combining these inequalities yields β < xk+1 < xk < γ Therefore our initial choice of xk gives us an decreasing sequence {xn } with an lower bound of β. x1 ). This gives us f (x) = f (γ + δ) deﬁnition of x = = = = = = (γ+δ)3 −1 3 γ 3 +3γ 2 δ+3γδ 2 +δ 3 +1 3 2 2 3 γ 3 +1 + 3γ δ+3γδ + δ3 3 3 f (γ) + γδ(γ + δ) + 1 δ 3 3 1 f (γ) − γδx + 3 δ 3 γ + γδx + 1 δ 3 3 deﬁnition of f algebra algebra algebra deﬁnition of x γ is a ﬁxed point We know that γ > 1 and x > 1. Combining these inequalities yields α < xk < xk+1 < β Therefore our initial choice of xk gives us an increasing sequence {xn } with an upper bound of β.23c If x > γ then we can express x as x = γ + δ for some δ > 0.fore we have xk+1 = f (xk ) = f (λβ + (1 − λ)γ) < λf (β) + (1 − λ)f (γ) = λβ + (1 − λ)γ deﬁnition of xk+1 deﬁnition of xk f is convex on (β. The second derivative of f (x) = 2x is positive on the interval (0. beta].

24 The function f (x) has a derivative of zero at its ﬁxed point. We’re also told that f (x) > δ > 0 for all x ∈ (a.This establishes that x > γ → f (x) > x.2). The function g(x) does not have a derivative of zero at its ﬁxed point and therefore does not have this property (although. which means that f (x) is monotonically increasing. Therefore if x1 > γ then the sequence {xn } will be a nonincreasing sequence. √ a the mean Exercise 5. b).25b Lemma 1: xn+1 < xn if xn > ξ We’re told that f (b) > 0 and that f (x) = 0 only at x = ξ. Lemma 3: if {xn } → κ then f (κ) = 0 Suppose that limn→∞ xn+1 = κ.17. Exercise 5. as we saw in exercise 3.23) we know that f (x) > 0 when x > ξ (otherwise we’d have f (x) = 0 for some second x ∈ (ξ. Therefore we have xn − f (xn ) < xn . which means that c < xn implies f (c) ≤ f (xn ). so by the intermediate value theorem (theorem 4. Exercise 5. Lemma 2: xn+1 > ξ if xn > ξ We’re told that f (x) ≥ 0. We know that this sequence doesn’t converge because we’re told that f has no ﬁxed points greater than γ. this equation is equivalent to xn − which of course means that xn+1 ≥ ξ.25a Each xn+1 is chosen to be the point where the line tangent tangent to f (xn ) crosses the x-axis. it still converges albeit more slowly). so when xk and xk+1 are both close to value theorem guarantees us that f (xk ) and f (xk+1 ) will be very near one another: |f (xk ) − f (xk+1 )| = |(xk − xk+1 )f (ξ)| ≈ 0 The lefthand term of this equation converges very rapidly to 0 because both |xk − xk+1 | and f (x) are converging toward zero. Therefore the ratio f (xn )/f (xn ) is deﬁned at each x and is positive when x > ξ. Using the fact that f (ξ) = 0 and a bit of algebraic manipulation. f (xn ) xn > ξ This of course means that xn+1 < xn when xn > ξ. We know that f is continuous since it is diﬀerentiable (theorem 5. Then by the Cauchy criterion we have limn→∞ xn − xn+1 = 0 which is equivalent to f (xn ) lim xn − xn − =0 n→∞ f (xn ) or f (xn ) lim =0 n→∞ f (xn ) For this to hold it must either be the case that f (xn ) → 0 or f (xn ) → ±∞: but we know that f (xn ) is bounded from the fact that f (xn ) is bounded (mean value theorem : if f (xn ) were unbounded. b) (without the δ it might be the case that f (x) = 0 but lim f (x) = 0). then 95 f (xn ) ≥ξ f (xn ) . which means that f (xn ) − f (ξ) ≤ f (xn ) xn − ξ because the LHS of this inequality is equal to f (c) for some c < xn .

Therefore by theorem 4. By induction. so it must be the case that κ = ξ. Using the quotient rule (theorem 5. we have f (ξ) = f (xn ) + f (xn )(ξ − xn ) + f (tn )(ξ − xn )2 2 Subtracting f (xn ) from both sides then dividing by f (xn ) (we’re told f (x) ≥ δ > 0) gives us f (tn )(ξ − xn )2 f (ξ) − f (xn ) = (ξ − xn ) + f (xn ) 2f (xn ) Rearranging some terms and recognizing that f (ξ) = 0. by theorem 3. This means that {xn } converges to ξ. so the inequality in part (c) guarantees us that xn+1 − ξ ≤ M (xn − ξ)2 = A(xn − ξ)2 2δ This allows us to recursively construct a chain of inequalities. So we have limn→∞ xn = κ and limn→∞ f (xn ) = 0.3). This means that {xn } is a decreasing sequence with a lower bound of ξ. Therefore. we know that every element of {xn } will be > ξ. By lemma 3 we have f (κ) = 0.25e How does g behave near ξ? We’re told f and f are diﬀerentiable.25c Using Taylor’s theorem to expand f (ξ) around f (xn ). xn+1 < xn for all n. . But we’re told that f has only one zero. therefore g is diﬀerentiable (theorem 5.14. is equivalent to xn+1 − ξ = f (tn )(xn − ξ)2 2f (xn ) Exercise 5. we have xn − f (xn ) f (tn )(xn − ξ)2 −ξ = f (xn ) 2f (xn ) Which. Therefore it must be the case that f (xn ) → 0. we have xn+1 − ξ ≤ A2 n n−1 (x1 − ξ)2 n −1 (x1 − ξ)2 = n n 1 [A(x1 − ξ)]2 A Exercise 5. using lemma 2. A2 Collapsing the exponents of the rightmost term. and clearly x is diﬀerentiable. the sequence converges to some point κ.25d We’re told that f (x) ≥ δ and f (x) ≤ M . by the deﬁnition of xn+1 . by lemma 1. Exercise 5. xn+1 − ξ ≤ A1 (xn − ξ)2 ≤ A1 A2 (xn−1 − ξ)4 ≤ A1 A2 A4 (xn − 2 − ξ)8 ≤ · · · ≤ A1 A2 . the derivative of g is given by g (x) = 1 − f (x)2 − f (x)f (x) f (x)f (x) = 2 f (x) f (x) 96 .(f (xn ) − f (x1 ))/xn − x1 = f (c) would be unbounded).2 we have f (κ) = lim f (x) = 0 x→κ The actual proof Choose x1 > ξ. Therefore. .3 again).

lim |g (x)| ≤ lim A|f (x)| = 0 x→ξ x→ξ Which immediately implies x→ξ lim g (x) = 0 and this describes the behavior of g as x → ξ. xn+1 = m−1 m xn = m−1 m 2 xn−1 = m−1 m 3 xn−2 = · · · = m−1 m n x1 Taking the limit of each side as n → ∞. From the deﬁnition of g this would mean that g(κ) − κ = − f (κ) =0 f (κ) For this to hold it must either be the case that f (κ) = 0 or f (κ) = ±∞: but we know that f is bounded from the fact that f is bounded (mean value theorem : if f were unbounded. if m = 1 2 then equation (45) becomes lim xn+1 = lim (−1)n x1 n→∞ n→∞ This limit clearly doesn’t exist when x1 = 0. the RHS reduces to f (ξ) = 0 because f is continuous. Show that Newton’s method involves ﬁnding a ﬁxed point of g This is a slightly modiﬁed version of lemma 3 for part (b). Therefore it must be the case that f (κ) = 0 and therefore κ must be the unique point for which f (κ) = 0. if m is smaller than 2 then the limit in equation (45) is unbounded and so {xn } → ±∞. we have |g (x)| = f (x)f (x) f (x) Using the inequality we established in part (d) we have |g (x)| ≤ A|f (x)| When we take the limit of each side of this equation as x → ξ. In the speciﬁc case given in the exercise we have m = 97 1 3 and therefore the sequence {xn } fails to converge. Exercise 5. then f (x)−f (y) = f (c) would be x−y unbounded for some x. y). . Newton’s formula gives us the step function xn+1 = xn − f (xn ) xm xn n = xn − = = xn − f (xn ) m mxm−1 n m−1 m xn This gives us a recursive deﬁnition of xn . The latter case occurs when |m − 1| < |m| 1 1 This inequality holds iﬀ m > 2 . we have lim xn+1 = lim m−1 m n n→∞ n→∞ x1 (45) In order to have {xn } converge to 0 we must either have x1 = 0 or |m − 1/m| < 1.25f We’ll consider the more general case in which f (x) = xm . which is what we were asked to describe. Suppose that we have found a ﬁxed point κ such that g(κ) = κ.Taking the absolute value of each side. If m is larger than 2 the limit in equation (45) exists and is equal to zero and 1 so {xn } → 0. In this case the single real zero occurs at x = 0.

f (a) = 0. But y1 and y2 were arbitrary solutions for the intial-value problem. by exercises 5. This turns (48) into |f (x)| ≤ 0 . We need only perform this a ﬁnite number of times (speciﬁcally. and we’re told that there is a real number A such that |f (x)| = |y2 − y1 | = |φ(x.27 the initial-value problem has a unique solution. b]. so we have proven that there is only one unique solution. we see that f (x) is diﬀerentiable and that f (a) = a. Therefore we have M0 ≤ M1 (x0 − a) Suppose 0 < A(x0 − a) < δ < 1. x ∈ [a. Therefore f (x) = 0 on the entire interval [a. 98 . As in the previous problem.28 Let ya and yb be two solution vectors to the initial-value problem and deﬁne the vector-valued function f (x) = yb (x) − ya (x). by exericse 5.26 Following the hint given in the exercise. x0 ] (48) (47) From the fact that f (a) = 0 we know that M1 (x0 − a) is the maximum possible value that f could possibly obtain at f (x0 ) (otherwise. We know that |f (x)| ≤ M1 (x0 − a) (46) because otherwise we’d have f (x) − f (a) f (x) = = f (c) for some c ∈ (a. yb ) − φ(x. a maximum of of [(b − a)/(x0 − a)] + 1 times) before we have covered the entire interval. In comparison. And since we can force 0 < A(x0 − a) < 1 by choosing an appropriate x0 . Therefore. x ∈ [a. Then (48) would give us |f (x)| ≤ M1 (x0 − a) ≤ AM0 (x0 − a) ≤ δM0 which contradicts (49) unless M1 = M0 = 0. let M0 = sup |f (x)| and let M1 = sup |f (x)| for x ∈ [a. M0 is the maximum value that f actually does obtain at some x ∈ [a. b) > M1 = sup |f (x)| x0 − a x0 − a which is clearly contradictory. if f (x0 ) > M1 (x0 −a).26.27 The given hint is pretty much the entire solution. x0 ].26: it’s diﬀerentiable. Let y1 and y2 be two solutions to the initial-value problem and deﬁne the function f (x) = y2 (x) − y1 (x). therefore y1 = y2 . it must be the case that M1 = M0 = 0. we know that M1 = sup |f (x)| ≤ sup |Af (x)| = A sup |f (x)| = AM0 Combining (46) and (47) gives us |f (x)| ≤ M1 (x0 − a) ≤ AM0 (x0 − a) . b]. We can now repeat these steps using the interval [x0 . Additionally. y1 )| ≤ A|y2 − y1 | = A|f (x)| Therefore. we have y2 (x) − y1 (x) = f (x) = 0 for all x. b]. then the mean value theorem would give us some f (c) > M1 ). x0 ]. x0 ] which shows that f (x) = 0 on the interval [a. (49) Exercise 5. if there is a real number A such that |f (x)| = |yb − ya | = |φ(x.Exercise 5.26 and 5. y2 ) − φ(x. The function f meets all of the prerequisites spelled out in exercise 5. Exercise 5. ya )| ≤ A|y2 − y1 | = A|f (x)| then.

Using (50) and (51) we see that this inequality will hold if there exists some A such that |w1 (x) − v1 (x). wk−1 (x) − vk−1 (x). f ) = 0. We’re told that α is continuous at x0 .f ) = 0 Although it’s easy to construct a partition such that L(P. Let v and v be two solution vectors to the initial value problem and deﬁne the vector valued function f (x) = w(x) − v(x): f (x) = [w1 (x) − v1 (x). wk (x) − vk (x). . . f ) = 0. . so this is equivalent to |w2 (x) − v2 (x). . f ). wk−1 (x) − vk−1 (x). By choosing an appropriate partition we can make U (P. . x0 ]) then ∆αi = [α(x0 ) − α(x0 )] = 0 and therefore mi ∆αi = 0. . f ) < .29 Note: there’s probably a more subtle answer here. so by the deﬁnition of continuity we can ﬁnd some δ > 0 such that d(x. . f ) : P is a partition of [a. . let P be an arbitrary partition of [a. w1 (x) − v1 (x)| ≤ A|w2 (x) − v2 (x). . therefore sup L(P. . . g(x) · (w(x Exercise 6. x0 ) < δ → d(α(x). wk−1 (x) − vk−1 (x). α(x0 )) < 2 (52) 99 . . If mi contains only the point x0 (that is. Therefore b 0 = sup{L(P. b]} = a f dx inf U(P. . so mi ∆αi = 0. f ) = 0. . if the interval is [x0 . wk (x) − vk (x). w3 (x) − v3 (x). If mi contains any point other than x0 then inf f (x) = 0 on this interval. wk (x) − vk (x) which. therefore inf U (P. and the rightmost term is a dot product. . we have to show that 0 is the supremum of the set of all L(P. which means L(P. wk (x) − vk (x).28 we know that we want an inequality of the form |f (x)| ≤ A|f (x)|. Therefore mi ∆αi = 0 for all i. by the equivalences given in the exercise. . wk (x) − vk (x). f ) = 0 for every partition. becomes f (x) = w2 (x) − v2 (x). w2 (x) − v2 (x). . f ) = 0 for all P . .1 (Proof 1) Outline of the proof We’ll see that L(P. . . j=1 (50) k gj (wj − vj ) (51) From exercises 5. w3 (x) − v3 (x). f ) = 0. . w3 (x) − v3 (x). sup L(P. w2 (x) − v2 (x). w3 (x) − v3 (x). So let > 0 be given. . j=1 gj (wj − vj ) The order of the components don’t matter when we’re taking the norm. To do this. . so L(P. . probably involving Taylor’s theorem. f ) arbitrarily small.Exercise 5. b] and let mi be an arbitrary interval of this arbitrary partition. We conclude that f = 0. But P was an arbitrary partition.f )=0 We need to show that for any > 0 we can ﬁnd some partition P such that 0 ≤ U (P. . wk (x) − vk (x)| k ≤ A w2 (x) − v2 (x). .26-5. wk (x) − vk (x)] The derivative of this is given by f (x) = w1 (x) − v1 (x). w2 (x) − v2 (x). . .

x0 + µ). x0 + µ) = 2µ < δ. therefore b b f dα = a a f dα = 0 Exercise 6. Since |x0 − x| ≤ µ < δ for all x ∈ X. f ) is equivalent to Mi − mi on a single arbitrarily small interval around 0.1 (Proof 2) Thanks to Helen Barclay (hbarcla2@mail.Let 0 < µ < M1 ∆α1 M2 ∆α2 M3 ∆α3 δ 2 and let P be the partition {a. b] where κ is some arbitrary nonzero number.19). b}.10. therefore the inﬁmum of L(P. The proof Suppose. x0 − µ. Now let 0 < µ < δ and let P be 2 the four-element partition {a. Therefore there exists some δ > 0 such that |x0 − x| < δ → |f (x0 ) − f (x)| < κ . Every partition of [a.usf. The function f is discontinuous at only one point and α is continous at that point. therefore f is uniformly continuous (theorem 4. so b 0 = inf{U (P.edu) for this proof.3 Outline of the proof We’ll see that U (P. we have |f (x0 ) − f (x)| = |κ − f (x)| < κ for all x ∈ X. From this we conclude that sup L(P. b]. We’re told that f is continuous on a closed set. b] will contain a point other than x0 so mi = 0 for every interval. that f (x0 ) = κ for some x0 ∈ [a. f ) = mi ∆xi ≥ m2 ∆x2 = min f (X)[(x0 + µ) − (x0 − µ)] = min f (X)(2µ) > κ (2µ) > 0 2 f dx = 0. it must be the case that sup L(P. f ) : P is a partition of [a. For clarity. f dx = 0 Since L(P. This means that 2 L(P. f ) > 0 and therefore We assumed that f (x) = 0 for some x ∈ [a. f ) is zero. b}. b]. then we can construct a partition P for which L(P. 100 . For this 2 inequality to hold for all f (x) it must be the case that min f (X) > κ . f ) i=13 Mi ∆αi = 0 + + 0 = But this epsilon is arbitrarily small. M2 ∆α2 = α(x0 + µ) − α(x0 − µ) ≤ From this we have U (P. f ) > 0. To make this diﬀerence arbitrarily small we will need to make sup f (x) − inf f (x) arbitrarily small by restricting x to a suﬃciently small neighborhood of 0: this is possible iﬀ f is continuous at 0. We now calculate each Mi : = = = M1 [α(x0 − µ) − α(a)] M2 [α(x0 + µ) − α(x0 − µ)] M3 [α(b) − α(x0 + µ)] = = = 0 · [α(x0 − µ) − α(a)] = 0 1 · [α(x0 + µ) − α(x0 − µ)] 0 · [α(b) − α(x0 − µ)] = 0 To determine a bound for M2 ∆α2 we apply (52) which allows us to conclude that α(x0 + µ) − α(x0 − µ) < from the fact that d(x0 − µ. so f ∈ R by theorem 6. f ) − L(P. let X = (x0 − µ. b] and concluded that then f (x) = 0 for all x ∈ [a. x0 + µ. x0 − µ. b]} = a f dx Exercise 6. x0 + µ. if Exercise 6. for purposes of contradiction.2 Outline of the proof If we assume that f (x) = 0 for some x ∈ [a. f ) > 0 and therefore f = 0. f dx = 0: by contrapositive. f ) > 0 for this particular P .

f (0)) < and therefore. therefore ∆β = 0. . 0] and [0. ω]. by the triangle inequality.16 the points f −1 (Mω ) and f −1 (mω ) both exist in [0. 0] and [0. We’re told that f (0+) = f (0). f ) and L(P. f ) = Mω − mω = f (a) − mω ≥ f (a) − f (0) ≥ f (b) − f (0) ≥ 101 . f ) − L(P. So for any partition P let a be a point at which f (a) = Mω and let b be a point for which |f (b) − f (0)| ≥ . mω ) < and therefore U (P. Therefore the values of U (P. For all three of the β functions we have Mα ∆βα Mω ∆βω mα ∆βα mω ∆βω = Mα [β(0) − β(α)] = Mα β(0) = Mω [β(ω) − β(0)] = Mω [1 − β(0)] = mα [β(0) − β(α)] = mα β(0) = mω [β(ω) − β(0)] = mω [1 − β(0)] (53) (54) (55) (56) (57) Case 1: β(0) = 0 When β(0) = 0 we have Mi ∆β = Mα β(0) + Mω [1 − β(0)] = Mω mi ∆β = mα β(0) + mω [1 − β(0)] = mω Proof that continuity of f implies integrability: let > 0 be given. that 0 is an element of this partition.4 we can assume. Let Mα denote the supremum of f (x) on the interval [α. 0. the upper and lower Riemann sums depend entirely on the intervals containing 0 Let P be an arbitrary partition of [−1. therefore Mi ∆β = mi ∆β = 0. f (0)) < d(f −1 (mω ). f ) = Mω − mω < And since was arbitrary. so by deﬁnition of continuity at this point there exists some δ > 0 such that d(0. The interval [0. 0]. This shows that Mi ∆β = mi ∆β = 0 for every interval that doesn’t contain 0. For every xi−1 < xi < 0 we have β(xi−1 ) = β(xi ) = 0. ω] Let P be an arbitrary partition. f ) depend entirely on the intervals [α. . From the negation of the deﬁnition of continuity we therefore know that there exists some > 0 such that for all δ we can ﬁnd some 0 < x < δ such that |f (x) − f (0)| ≥ . This holds for arbitrary P . δ] is compact. etc. The general form of Mi ∆β and mi ∆β on [α. therefore ∆β = 0. for every 0 < xi−1 < xi we have β(xi−1 ) = β(xi ) = 1. This gives us f (b) ≤ f (a) = Mω and mω ≤ f (0). mω ) ≤ d(Mω . therefore U (P. therefore by theorem 4. Note that α and ω must both exist since P is a ﬁnite set. . 1}. 0) < δ → d(mω . 0) < δ → d(Mω . f (x)) < 2 Construct a partition P = {−1. f (0)) + d(f (0). < α < 0 < ω < . f ) − L(P. this is suﬃcient to show that f dx exists. 1]. We deﬁne α to be the partition element immediately preceding 0 and deﬁne ω to be the partition element immediately following 0 (so our partition has the form P = {x0 < x1 < . similarly deﬁne mα . δ] and therefore: d(f −1 (Mω ). mω . . Mω . although it’s possible that α = −1 or ω = 1.Lemma 1: For each of the β functions. 2 2 Proof by contrapositive that integrability of f implies right-hand continuity: Assume that f (0+) = f (0). < xn < xn+1 }). x) < δ → d(f (0). therefore Mi ∆β = mi ∆β = 0. By theorem 6. Similarly. d(Mω . without loss of generality. δ.

102 . f ) − L(P. δ] we know that Mω ≥ f (0). f ) ≥ f dβ (58) f dβ f . f ) − L(P.4 and/or the deﬁnition of f ). we saw in case (1) that [Mω − mω ] can be made arbitrarily small iﬀ f (0) = f (0+). f ) ≥ U (P. iﬀ f is continuous at 0. Case 3: β(0) = 1 2 f dβ − f (0) (59) > 0 we must have f dβ = f (0). This gives us f dβ mω = L(P. f ) ≤ f (by theorem 6.The partition P was arbitrary so this inequality must be true for all possible partitions. so we can never ﬁnd P such that inf U (P. f ) = 1 [Mα + Mω ] 2 1 [mα + mω ] 2 1 1 [Mα − mα ] + [Mω − mω ] 2 2 The function f is integrable iﬀ this term can be made arbitrarily small. Proof that this integral. that is. f ) − L(P. This gives us the inequalities Mω = U (P. f ) ≥ f (0) − But we also know that mω ≤ f (0) and that U (P. f ) in this case iﬀ f (0+) = f (0−) = f (0). f dβ = f (0) iﬀ f (0) = f (0−) is almost identical to the previous proof with α in the When β(0) = 1/2 we have Mi ∆β = Mα β(0) + Mω [1 − β(0)] = mi ∆β = mα β(0) + mω [1 − β(0)] = Therefore U (P. f ) − sup L(P. f ) ≤ f (0) from which we have U (P. f ) ≥ f (0) L(P. f ) ≤ and therefore f dx does not exist. f ) = If this is to be true for all Case 2: β(0) = 1 When β(0) = 1 we have Mi ∆β = Mα β(0) + Mω [1 − β(0)] = Mα mi ∆β = mα β(0) + mω [1 − β(0)] = mα From here the proof that place of ω. We also know that L(P. f ) − L(P. and we saw in case (2) that [Mα − mα ] can be made arbitrarily small iﬀ f (0) = f (0−). f ) ≤ from which we have U (P. f ) ≥ Combining (58) and (59) gives us f dβ − f (0) ≤ U (P. Therefore we can minimize U (P. if it exists. is equal to f (0): from the fact that 0 is an element of [0. f ) − L(P.

Proof that this integral, if it exists, is equal to f (0): from the fact that 0 is an element of [0, δ] we know that Mω ≥ f (0) and Mα ≥ f (0). We also know that L(P, f ) ≤ f (by theorem 6.4 and/or the deﬁnition of f ). This gives us the inequalities 1 1 [Mα + Mω ] = U (P, f ) ≥ [f (0) + f (0)] = f (0) 2 2 L(P, f ) ≤ from which we have U (P, f ) − L(P, f ) ≥ f (0) − f dβ f . This gives us (60) f dβ

But we also know that mω ≤ f (0) and that mα ≤ f (0) and that U (P, f ) ≥ U (P, f ) ≥ f dβ

1 1 [mα + mω ] = L(P, f ) ≤ [f (0) + f (0)] = f (0) 2 2 from which we have U (P, f ) − L(P, f ) ≥ Combining (58) and (59) gives us f dβ − f (0) ≤ U (P, f ) − L(P, f ) = If this is to be true for all Part (d) of the exercise If f is continuous at 0 then we have f (0) = f (0+) = f (0−), so by case (1) we have (2) we have f dβ2 = f (0) and by case (3) we have f dβ3 = f (0). f dβ1 = f (0) and by case > 0 we must have f dβ = f (0). f dβ − f (0) (61)

Exercise 6.4

Let P be an arbitrary partition of [a, b]. Let n = |P | − 1, so that n represents the number of intervals in the partition P . Every interval will contain at least one rational number (theorem 1.20b) and at least one irrational number (by pigeonhole principle, since |(xi , xi+1 )| = |R| > |Q|). Therefore we have Mi = 1, mi = 0 for all i. This gives us

n

Mi ∆xi =

1

1 · (xi+1 − xi ) = xn+1 − x0 = b − a

n

mi ∆xi =

1

0 · (xi+1 − xi ) = 0

But P was an arbitrary partition, so this holds for all partitions, and therefore sup L(P, f ) = 0, and therefore f ∈ R. inf U (P, f ) = 1

Exercise 6.5

Is f ∈ R if f 2 ∈ R? No. Consider the function f (x) = -1, 1, x∈Q x∈Q

We saw that f ∈ R in exercise 6.4, but clearly f 2 (x) = 1 ∈ R. 103

Is f ∈ R if f 2 ∈ R? Yes. The function φ(x) = 6.11 to claim that √ 3 x is a continuous function, so we let h(x) = φ(f 3 (x)) = f (x) and appeal to theorem f3 ∈ R → h ∈ R → f ∈ R Note √ that the claim that h(x) = φ(f 3 (x)) = f (x) relied on the fact that x → x3 is a one-to-one mapping, so 3 that x3 = x. In contrast, the mapping x → x2 is not one-to-one as can be seen by the fact that (−1)2 = −1.

Exercise 6.6

Outline of the proof We can deﬁne a partition for which f is discontinuous on 2n intervals each of which has length 3−n , so the Riemann sum across these intervals is proportional to (2/3)n . This sum approaches 0 as n → ∞. The proof Let Ei be deﬁned as in sec. 2.44. Note that E0 consists of one interval of length 1; E1 consists of two intervals of length 1 ; E2 consists of 4 intervals of length 1 ; and, in general, En will consist of 2n intervals of length 3−n . 3 9 With each En we can associate a set Fn that contains the endpoints of the intervals contained in En . That is, we deﬁne Fn = {a1 < b1 < a2 < b2 < . . . < a2n < b2n } where each [ai , bi ] is an interval in En . Since every point of the Cantor set – and therefore every point at which f might be discontinuous – is in an interval of the form [ai , bi ] we’ll choose a partition that lets us isolate these intervals. Let m and M represent the lower and upper bounds of f on [0, 1]. Choose an arbitrary large enough that 2n+1 (M − m) 3n−1 > and choose δ such that 0<δ< These choice will be justiﬁed later. Deﬁne Pn to be Pn = {0 = a1 < (b1 + δ) < (a2 − δ) < (b2 + δ) < (a3 − δ) < (b3 + δ) < . . . < (a2n − δ) < b2n = 1} This partition contains 2n+1 points and therefore contains 2n+1 − 1 segments. We must now show that we can make U (Pn , f ) − L(Pn , f ) arbitrarily small.

2n+1 −1

> 0. Choose n

1 3n+1

U (Pn , f ) − L(Pn , f ) =

i=1

(Mi − mi )∆xi

We’ll separate this sum into the intervals on which f is continuous (i = 2, 4, 6, . . .) and the intervals on which f might contain discontinuities (i = 1, 3, 5, . . .).

2n 2n

=

i=1

**(M2i − m2i )∆x2i +
**

i=0

(M2i+1 − m2i+1 )∆x2i+1

The function f is continuous on every interval of the form [bi + δ, ai+1 − δ], and these intervals are represented by the lefthand summation. We can therefore reﬁne Pn such that

2n

≤

2

+

i=1

(M2i+1 − m2i+1 )∆x2i+1

104

We know the exact value of the ∆x terms. For Pn , we have d(ai , bi ) = 3−n and therefore ∆x2i+1 = 3−n + 2δ. Similarly, we know that Mi ≤ M and mi ≥ m, so we have

2n

**≤ From our choice of δ <
**

1 3n+1

2

+

i=1

(M − m)

1 + 2δ 3n

<

1 3n

this becomes

2n

≤

2

+

i=1

(M − m)

1 3n−1

This summation is constant with respect to i so it becomes simply = From our choice of n we have 3n−1 > 2 + 2n (M − m) 1 3n−1

2n+1 (M −m)

and therefore 3−(n−1) < /2n+1 (M − m). + 2n (M − m) 2n+1 (M − m) 2 + 2 =

≤

2

=≤

Exercise 6.7a

If f ∈ R, then from theorem 6.12 we have

1 c 1

f dx =

0 0

f dx +

c

f dx

(62)

**Now consider the partition P of [0, c] deﬁned by P = {0, c}. This partition has only a single interval, so we have
**

c 0<x<c

inf f (x) · c = L(P, f ) ≤

0

f dx ≤ U (P, f ) = sup f (x) · c

0<x<c

From (62), adding

1 c

**f dx to each term of this inequality gives us
**

1 1 1

f dx + inf f (x) · c ≤

c 0<x<c 0

f dx ≤

c

f dx + sup f (x) · c

0<x<c

Taking the limit of each term of this inequality as c → 0, we see that f (x) · c → 0 while the center term remains constant. The resulting inequality is

1 c→0 1 1

lim

f dx ≤

c 0

f dx ≤ lim

c→0

f dx

c

**which of course implies that
**

1 c→0 1

lim

f dx =

c 0

f dx

which is what we wanted to prove.

Exercise 6.7b

Outline of the proof We construct a function f for which f = (see Rudin’s remark 3.46). We then see that 3.28).

n (−1)i i=1 i : this n |f | = i=1 1 : i

is the alternating harmonic series, which converges this is the harmonic series, which diverges (theorem

105

We want to choose δ such that each harmonic number in [c. 0. Let N be the smallest harmonic number greater than c. N − δ]. 1]. Let 1 Hc represent the partition containing the elements {c} ∪ n ± δ : n ∈ N. 1]. so we have mj ∆xj = inf f (x)∆xj = sup f (x)∆x = M = (−1)j therefore (Mj − mj )∆xj = 0 (66) 1 − 2δ(j + 1) j 106 .The proof Choose any c ∈ (0. c 1 • The partition Hc contains one interval of the form [c. M1 ∆x1 = sup f (x)∆x1 = (−1)N (N + 1)∆x1 = (−1)N (N + 1) = (−1)N (N + 1) 1 −δ N (N + 1) = (−1)N 1 1 −δ− N N +1 1 − (N + 1)δ N The function is constant over this interval. 1 − δ for which j 1 − 2δ j(j + 1) = (−1)j 1 − 2δ(j + 1) j = (−1)j (j+1) The function is constant over this interval. 1) and consider the following function deﬁned on [c. f (x) = (−1)n (n + 1). <x< otherwise 1 n+1 1 n for some n ∈ N Claim 1 : f is integrable on any interval [c. so we have m1 ∆x1 = inf f (x)∆x1 = sup f (x)∆x1 = M1 therefore |M1 − m1 |∆x1 = 0 • The partition Hc contains one interval of the form [1 − δ. 1] for which Mn ∆xn = sup f (x)∆xn = 0∆x = 0 mn ∆xn = inf f (x)∆xn = −2∆x = −2δ therefore |Mn − mn |∆xn = 2δ • The partition contains N − 1 intervals of the form 1 i (63) (64) − δ. 1] is contained in a distinct interval of radius δ: so choose δ such that 1 1 0 < δ < 2 [ N − c]. n < 1 ∩ [0. 1 + δ for which i Mi ∆xi = sup f (x)∆xi ≤ sup |f (x)|2δ ≤ |i + 1|2δ mi ∆xi = inf f (x)∆xi ≥ inf −|f (x)|2δ ≥ −|i + 1|2δ therefore |Mi − mi |∆xi ≤ |Mi | + |mi | ≤ 4δ|i + 1| • The partition contains N − 2 intervals of the form Mj ∆xj = sup f (x)∆xj = (−1)j (j+1) 1 1 −δ− −δ j j+1 1 j+1 (65) + δ. 1] 1 Choose an arbitrarily small > 0. We must also make sure that δ < 2(N +1)(N +2) for reasons that will become clear later.

1 c Claim 2 : The value of limc→0 f is deﬁned Adding the values for the M terms. our choice of δ came afterward. Earlier in the proof we chose δ such that δ < so we obtain the inequality |U (P. f ) = M1 + Mn + i=2 Mi + i=2 Mj N N −1 ≤ (−1)N 1 |i + 1|2δ + − (N + 1)δ + 0 + N i=2 (−1)j j=2 N −1 1 − 2δ(j + 1) j 1 − 2δ(j + 1) j = (−1)N 1 − (N + 1)δ δ[(N + 1)(N + 2) − 6] + N (−1)j j=2 Remember that we ﬁrst chose a value for c. we have N N −1 |U (P. Therefore lim U (Pc . which forced us to use a particular value for N . f )| = |(M1 − m1 )∆x1 + (Mn − mn )∆xn + i=2 (Mi − mi )∆xi + j=2 N (Mj − mj )∆xj | N −1 ≤ |(M1 − m1 )|∆x1 + |(Mn − mn )|∆xn + i=2 N N −1 |Mi − mi |∆xi + j=2 |Mj − Mj |∆xj ≤ 0 + 2δ + i=2 4δ|i + 1| + j=2 0 = = (N + 1)(N + 2) − 1 2 2δ(N + 1)(N + 2) 2δ + 4δ 2(N +1)(N +2) . we have N N −1 U (P.Summing equations (63)-(66). and adding them gives us the inequality 1 c→0 lim f dx ≤ 1 − ln(2) c Together. f )| ≤ And was arbitrary. so this proves that f ∈ R. so taking the limit of both sides as c → 0 gives us ∞ c→0 → 0 and N → ∞ as c → 0. f ) = j=2 (−1)j 1 + = 1 − ln(2) + j which proves that 1 c→0 lim f dx ≥ 1 − ln(2) c The values for the m terms are almost identical to the M terms. f ) − L(P. of course. f ) ≤ 1 N (−1)N + N N −1 (−1)j j=2 1 + j 1 N We deﬁned to be the smallest harmonic number greater than c. so we’re still free to select δ small enough to make this last inequality become: U (Pc . f ) − L(P. these two inequalities prove that 1 c→0 lim f dx = 1 − ln(2) c 107 .

so the lefthand limit doesn’t exist and therefore 1 limc→0 c f dx does not exist. f ) n−2 = i=1 n−2 mi ∆xi = i=1 f (i + 1) (Note: the index is going to n − 2 instead of n − 1 because of the awkward numbering of the partition: note that the partition only goes to xn−1 ). . Choose c. 3. . we ﬁnd that adding the values for the M terms gives us N N −1 U (P. Because f (x) > 0 this means that f (n) is unbounded (theorem 3. Exercise 6. c n f (x) dx 1 ≥ 1 f (x) dx ≥ L(Pn . assume that f (n) diverges. f ) n−1 = i=1 n−1 Mi ∆xi = i=1 f (i) 108 . n] to be Pn = {x0 = 1. so we conclude that the integral does not converge. 4. we select δ small enough to gives us 1 c→0 ∞ lim f dx = c j=2 1 + j But we know from chapter 3 that this series doesn’t converge. Next assume that f (n) converges to some κ ∈ R. The fact that f is monotonically decreasing allows us to derive the following chain of inequalities. Let n be the greatest integer such that n < c. .8 Choose an arbitrary real number c > 1.24). n. First. 2. The fact that f is monotonically decreasing allows us to derive the following chain of inequalities.Claim 3 : The value of limc→0 1 c |f | is not deﬁned If we follow the logic from claims 1 and 2. n = xn−1 }. Taking the limit as c → ∞. we have ∞ c n f (x) dx = lim 1 c→∞ f (x) dx ≥ lim 1 n→∞ f (i + 1) i=2 The integral on the left-hand side of this chain of inequalities is greater than the unbounded sum on the righthand side. f ) = M1 + Mn + i=2 Mi + i=2 Mj N N −1 ≤ 1 1 − (N + 1)δ + + |i + 1|2δ + N 2 i=2 j=2 1 − 2δ(j + 1) j N −1 = 1 1 − (N + 1)δ + δ[(N + 1)(N + 2) − 6] + N 2 j=2 1 − 2δ(j + 1) j Again following the logic of part 2.Deﬁne the partition Pn of [0. Pn+1 as deﬁned above. . c n+1 f (x) dx 1 ≤ 1 f (x) dx ≤ U (Pn+1 .

11). This leaves us with ∞ 0 cos x dx = lim c→∞ 1+x c 0 sin x dx (1 + x)2 By the deﬁnition in exercise 6.Taking the limit as c → ∞. We rewrite uv as the exponential of a natural log: 1 1 ln up + ln v q uv = (up )1/p (v q )1/q = exp(ln((up )1/p (v q )1/q )) = exp p q The exponential function f (x) = ex has a strictly positive second derivative. therefore its ﬁrst derivative is monotonically increasing (theorem 5. we have ∞ c n−1 f (x) dx = lim 1 c c→∞ f (x) dx ≤ lim → ∞ 1 n i=1 f (i) We’re told that f (x) > 0.usf. deﬁnition of convex in exercise 4. Exercise 6.14.9 By the deﬁnition given in exercise 6.23). therefore f (x) is a convex function (proof in exercise 5. so we can use integration by parts: ∞ c f (x)g (x) dx = lim f (c)g(c) − f (0)g(0) − 0 c→∞ 0 f (x)g(x) dx Applying this to the given function with f (x) = (1 + x)−1 and g(x) = sin x. so we know that 0 f (x) is a monotonically increasing function of c. The previous c c inequality tells us that 1 is bounded above. up vq + p q 109 . the non-integral terms both tend to zero as c → ∞. 1 − λ = 1/q we have exp 1 1 ln up + ln v q p q ≤ 1 1 up vq exp (ln up ) + exp (v q ) = + p q p q Combining these last two inequalities.14 we know that limc→∞ 1 f dx is deﬁned.8: ∞ c f (x)g (x) dx = lim 0 c→∞ f (x)g (x) dx 0 The integral from 0 to c is ﬁnite. Therefore by theorem 3.edu).10 a: method 1 This proof was due to Helen Barclay (hbarcla2@mail. Integration by parts therefore yields ∞ 0 sin c sin 0 cos x dx = lim − + c→∞ 1 + c 1+x 1 c 0 sin x dx (1 + x)2 Since −1 ≤ sin c ≤ 1.8 this is equivalent to ∞ 0 cos x dx = 1+x ∞ 0 sin x dx (1 + x)2 Exercise 6. we have uv ≤ which is what we were asked to prove. we have f (x) = −(1 + x)−2 and g (x) = cos x. By deﬁnition of convexity with λ = 1/p.

a = p . and f (u) ≥ 0 for all other u. t = v q . And this is what we were asked to prove. so we see that (67) holds when s ≥ t or when t ≥ s – that is. But both of these inequalities are just (67) with a variable change. We can show algebraically q 1 that p−1 = p . Making these substitutions gives us = v q − v pq/p = v q − v q 0 Therefore the unique minimum of f is f (v q/p ) = 0.Exercise 6.10 a: method 2 1 Using the variable substitutions s = up . Exercise 6. so the the point u = v q/p is the unique global minimum of f (u). If s ≥ t then s/t ≥ 1. (1 − a) = 1 q we can rewrite (67) uv ≤ as up vq + p q sa t1−a ≤ as + (1 − a)t Multiplying both sides by s−a ta−1 gives us 1 ≤ as1−a ta−1 + (1 − a)s−a ta We can rewrite this previous inequality in two equivalent ways: 1≤a 1≤a If s ≤ t then t/s ≥ 1 and therefore a t s a−1 t s s t a−1 + (1 − a) 1−a t s s t a (68) −a + (1 − a) (69) + (1 − a) t s a ≥ a(1)a−1 + (1 − a)(1)a = 1 so the inequality in (68) holds.10 a: method 3 This proof was due to Boris Shektman (don’t email him). which occurs at u = v 1/(p−1) . and therefore a s t 1−a + (1 − a) s t −a ≥ a(1)1−a + (1 − a)(1)−a = 1 so the inequality in (69) holds. For u < v q/p we have f (u) < 0 and for u > v q/p we have f (u) > 0. from this. it always holds. we also know that p + q = pq. Evaluating the function at this value of u gives us f (v q/p ) = 1 1 + p q vq vq + − v q/p v p q v q − (v (p+q)/p ) up vq + − uv p q = We’re given that 1/p + 1/q = 1 and. 110 . Deﬁne the function f (u) to be f (u) = This function has the derivative f (u) = up−1 − v Note that f (u) = 0 has only one positive real solution.

Exercise 6. we have b b f g dα ≤ a a fp gq 1 + dα = p q p b f p dα + a 1 q b g q dα = a 1 1 + =1 p q Exercise 6. by the deﬁnition of κ and λ. so in either of these cases we would have to be able to ﬁnd a neighborhood of ∞ or x0 for which this inequality still held.13 and part (b) of ˆ ˆ this exercise we know that |f | and |ˆ| are Riemann integrable and that g b b ˆˆ f g dα ≤ a a ˆ g |f ||ˆ| dα ≤ 1 ˆ By deﬁnition of f and g this becomes: ˆ b a f g b κ1/p λ1/q dα ≤ a |f | |g| dα ≤ 1 κ1/p λ1/q Multiplying both sides by the constant κ1/p λ1/q gives us b b f g dα ≤ a a |f g| dα ≤ κ1/p λ1/q which. is equivalent to b b b 1/p b 1/q f g dα ≤ a a |f g| dα ≤ a |f |p dα a |g|q dα which is what we wanted to prove. 111 . the fact that Holder’s inequality is valid for proper integrals shows that it’s true for improper integrals. λ= a |g|q dα ˆ Deﬁne f and g to be ˆ f (x) ˆ f (x) = 1/p . and therefore by theorem 6. we would be able to ﬁnd some function f such that c c→∞ c 1/p c 1/q lim f g dα > lim a b c→∞ |f |p dα a b 1/p a b |g|q dα 1/q or c→x0 lim f g dα > lim x0 c→x0 |f |p dα x0 x0 |g|q dα But this is a strict inequality.Exercise 6.10c Deﬁne κ and λ to be κ= a b b |f |p dα. And this would give us a proper integral for which Holder’s inequality doesn’t hold. By contrapositive.10d Proof by contrapositive.10b From part (a). κ These two functions are “normalized” in the sense that b b g (x) = ˆ b g(x) λ1/q κ =1 κ λ =1 λ ˆ |f |p dα = a b a b |f |p 1 dα = κ κ 1 |g|q dα = λ λ |f |p dα = a b |ˆ|q dα = g a a |g|q dα = a ˆ We know that f and g are Riemann integrable by theorem 6. If the inequality were false for the improper integrals.11.

b ||f + g||2 = a |f + g|2 dα deﬁnition of || · || ≤ = = ≤ b (|f | a b a b a b a + |g|)2 dα The triangle inequality is established for | · | |f g| dα + b a b |g|2 dα a 1/2 b a |f |2 + 2|f g| + |g|2 dα |f |2 dα + 2 |f |2 dα + 2 2 b a properties of integrals: theorem 6.12 1/2 |f |2 dα |g|2 dα + b a |g|2 dα Holder’s inequality (exercise 6. so it has upper and lower bounds m ≤ f (x) ≤ M . we have |f (t) − g(t)| ≤ |Mi − mi | = Mi − mi Similarly. f (xi )} ≤ Mi And of course we have similar bounds on f : mi ≤ f (t) ≤ Mi and therefore. (71) (70) 112 . Letting f = f − g and g = g − h this becomes ||f − h|| ≤ ||f − g|| + ||g − h|| which is what we were asked to prove. f (xi )} ≤ g(t).12 We’re told that f is continuous. mi min{f (xi−1 ). g(t) ≤ Mi and −f (t).10) deﬁnition of || · || = ||f ||2 + 2||f || ||g|| + ||g||2 = (||f || + ||g||) Taking the square root of the ﬁrst and last terms of this chain of inequalities. That is. Exercise 6. xi ] the function g(t) is bounded between f (xi−1 ) and f (xi ). g ∈ R. then we can prove the given inequality by letting f = f −g and g = g−h. α) − L(P.Exercise 6. t ∈ [xi−1 . since Mi ≤ M and −mi ≤ −m we have Mi − mi ≤ M − m We’ve now established all of the inequalities we need to complete our proof. we can ﬁnd some partition P such that 2 > 0. −g(t) ≤ −mi . Since U (P. This function is simply a series of straight lines connecting f (xi−1 ) to f (xi ). f. f. This becomes clearer if we rewrite g in an algebraically equivalent form: f (xi ) − f (xi−1 ) g(t) = f (xi−1 ) + (t − xi−1 ) xi − xi−1 We see that on any interval [xi−1 . since f (t). we have ||f + g|| ≤ ||f || + ||g|| This inequality must hold for any f. xi ] ≤ max{f (xi−1 ). α) < M −m Having now determined a particular partition P we can deﬁne g as suggested in the hint. Choose any f ∈ R(α).11 If we can prove that ||f +g|| ≤ ||f ||+||g||.

||f − g||2 = = ≤ = = ≤ = xi+1 n |f (x) − g(x)|2 dα i=0 xi xi+1 n |Mi − mi |2 dα i=0 xi n 2 xi+1 dα i=0 |Mi − mi | xi n 2 i=0 |Mi − mi | ∆α(xi ) n i=0 (M − m)(Mi − mi ) ∆α(xi ) n (M − m) i=0 (Mi − mi ) ∆α(xi ) 2 b a |f (x) − g(x)|2 dα deﬁnition of || · || integral property 6.12c from (70) integral property 6.22) with F = 1/2 u and g = sin u du gives us − cos u √ f (x) = 2 u which expands to cos x2 − cos(x + 1)2 + − 2(x + 1) 2x By the triangle inequality this gives us f (x) = |f (x)| = ≤ − cos(x + 1)2 cos x2 + − 2(x + 1) 2x (x+1)2 x2 (x+1)2 (x+1)2 − x2 x2 cos u du 4u3/2 cos u du 4u3/2 (x+1)2 x2 cos u du 4u3/2 cos u du 4u3/2 − cos(x + 1)2 cos x2 + + 2(x + 1) 2x 1 1 + + 2(x + 1) 2x (x+1) x2 2 (x+1)2 x2 < = = = = 1 du 4u3/2 1 1 1 1 1 + + − 2(x + 1) 2x 2 x+1 x 1 1 1 + + 2(x + 1) 2x 2x(x + 1) x + (x + 1) + 1 2x(x + 1) 2(x + 1) 1 = 2x(x + 1) x 113 .12a from (71) integral property 6. f. f. α) − L(P.19) we see that the given integral is equivalent to (x+1)2 f (x) = x2 sin u √ du 2 u (72) √ Integration by parts (theorem 6.13a Letting u = t2 we have du/dt = 2t and therefore dt = du du = √ 2t 2 u Using the change of variables theorem (6. Exercise 6. α)] ≤ (M − m) M −m 2 = we chose so that this would hold Taking the square root of the ﬁrst and last terms gives us ||f − g|| ≤ which is what we were asked to prove.12a Deﬁnition of U and L we chose P so that this would hold = (M − m)[U (P.

(x + 1)2 ] but cos is not a constant function so cos u < 1 for some u ∈ [x2 .Note that the strict inequality is justiﬁed by the fact that cos u ≤ 1 for all u ∈ [x2 . It’s also clear that this function is bounded above by 1 and bounded below by −1. by the triangle inequality. but it’s not immediately clear that these bounds are the lim sup and the lim inf.13c We’re asked to ﬁnd the lim sup and lim inf of the function xf (x) = 1 cos(x2 ) − cos([x + 1]2 ) + r(x) 2 We established in part (b) that r(x) → 0 as x → ∞. Exercise 6. |r(x)| ≤ ≤ = = = = < cos(x + 1)2 + 2x x+1 1 + 2x x+1 (x+1)2 x2 (x+1)2 x2 (x+1)2 x2 cos u 4u3/2 cos u du 4u3/2 cos(x + 1)2 − 2x x+1 (x+1)2 x2 (x+1)2 x2 cos u du 4u3/2 cos(x + 1)2 − 2x x+1 cos u du 4u3/2 cos u du 4u3/2 (x+1)2 x2 1 du 4u3/2 1 1 1 1 + 2x − x+1 2 x+1 x 1 1 +x x+1 x(x + 1) 1 1 + x+1 x+1 2 x+1 2 x So we see that 2xf (x) = cos x2 − cos[(x + 1)2 ] + r(x) where |r(x)| < 2/x.13b In the previous problem we used integration by parts to determine that − cos(x + 1)2 cos x2 f (x) = + − 2(x + 1) 2x Multiplying by 2x gives us −x cos(x + 1)2 2xf (x) = + cos x2 − 2x x+1 which is algebraically equivalent to 2xf (x) = cos x2 − cos[(x + 1)2 ] + Letting r(x) be deﬁned as r(x) = we have. (x + 1)2 ]. 114 . Exercise 6.

Proof: The lim sup of xf (x) is 1 Let 1 > > 0 and N ∈ N be given. we could rearrange it algebraically to give us π 2 [2(j − k) + 1]2 − 2π[2j + 3k + 1] + 1 = 0 This would allow us to use the quadratic formula to give an algebraic expression for π. we can see that these two equalities hold when π[2(j − k) + 1] − 1 √ 2 2π 2 =k (75) But this equation can’t hold. j. we have xf (x) = 1 1 cos(x2 ) − cos([x + 1]2 ) + r(x) = [cos(2kπ) − cos(2(j + 1)π + 2δπ) + r(x)] 2 2 115 j. M + δ]. Using this value of x in our original function.Remark: The supremum and inﬁmum of cos(x2 ) − cos([x + 1]2 ) are never obtained The supremum would be obtained if we could ﬁnd x such that cos(x2 ) − cos([x + 1]2 ) = 2 which would occur precisely when x2 = 2kπ. k ∈ N . The two important properties of κ are that p(κ) is an integer and that κ − M < δ (so that κ itself is “almost” an integer). Let δ be chosen so that 0<δ< Deﬁne the function p(m) to be p(m) = Its derivative with respect to m is p (m) = √ π[2m + 1] − 1 √ 2 2π 2 cos−1 (1 − ) 2π 2π(π[2m + 1] − 1) which is strictly increasing. Therefore we can make the derivative as large as we want by choosing a suﬃciently large m. Let κ represent one such x. If it did. We now have 2 π[2(M + δ) + 1] − 1 √ =k∈N p(κ) = p(M + δ) = 2 2π If we deﬁne j to be j = k + M this becomes π[2(j − k + δ) + 1] − 1 √ 2 2π 2 =k∈N Reversing the algebraic steps that led from (74) to (75) tells us that there exists some x ∈ R such that x2 = 2kπ. The inﬁmum of cos(x2 ) − cos([x + 1]2 ) is not obtained for the same reason. [x + 1]2 = (2(j + δ) + 1)π. M + δ) From the strictly increasing nature of p and our choice of M we have p(M + δ) − p(M ) = p (ξ)δ > p (M )δ > 1 Therefore p(x) must take an integer value for at least one x ∈ [M. By the mean value theorem we have p(M + δ) − p(M ) = p (ξ)δ. [x + 1]2 = (2j + 1)π. we are able to choose M ∈ N such that p (M )δ > 1 and M > N . Speciﬁcally. ξ ∈ (M. k ∈ N (74) (73) After some tedious algebra. but this is impossible as π is a transcendental number.

so the supremum is x→∞ lim sup xf (x) = 1 The proof that lim inf xf (x) = −1 is similar. this becomes xf (x) = 1 [1 − {cos(2(j + 1)π) cos(2 π) − sin(2(j + 1)π) sin(2 π)} + r(x)] 2 = 1 [1 − (−1) cos(2 π) − 0 + r(x)] 2 Finally. by the alternating series theorem (3.19) we see that the given integral is equivalent to ex+1 f (x) = ex sin(u) du u (79) 116 . Exercise 6.a that the sequence { n + 1 − n} is decreasing. Exercise 6. we know that r(x) → 0 as x → ∞: x→∞ lim xf (x) = 1 − 2 And was arbitrary.25) the series in (76) converges and therefore the integral in (76) converges.Using trig identities. Therefore.43) the series in (77) converges.14a Letting u = et we have du/dt = et = u and therefore dt = du du = et u Using the change of variables theorem (6. from our original choice of δ as an inverse cosine.6.13d ∞ ∞ √ √ (n+1)π sin(t ) dt 0 2 = 0 ∞ sin(t2 ) dt nπ (n+1)π (76) √ √ = 0 ∞ (−1)n | sin(t2 )| dt nπ √ √ (n+1)π = 0 ∞ (−1)n (−1)n 0 | sin(t2 )| dt nπ (n+1)π √ √ ≤ = √ π 1 dt nπ (and is similarly bounded below) √ n (77) ∞ (−1)n 0 √ n+1− (78) √ √ We saw in exercise 3. this becomes xf (x) = 1 [1 + (1 − ) + r(x)] 2 1 [r(x) − ] 2 =1+ From part (b). by the comparison test (3. Therefore.

We use integration by parts as we did in 6. x ex cos u du u2 117 .13(a). f (x) = which expands to f (x) = By the triangle inequality this gives us |f (x)| = ≤ cos ex cos ex+1 − − ex ex+1 ex+1 ex − cos u u ex+1 ex+1 − ex ex cos u du u2 cos ex cos ex+1 − − x e ex+1 ex+1 ex cos u du u2 cos u du u2 cos u du u2 cos ex cos ex+1 + + x e ex+1 1 1 + x+1 + ex e ex+1 ex ex+1 ex < < = = = 1 du u2 1 1 1 1 + x+1 + x − x+1 ex e e e 1 1 e−1 + x+1 + x+1 x e e e 2e ex+1 2 ex Therefore ex |f (x)| < 2. by the triangle inequality.14b In the previous exercise we used integration by parts to determine that f (x) = Multiplying by ex gives us ex+1 ex+1 ex cos ex cos ex+1 − − x e ex+1 cos u du u2 e f (x) = cos e − e Letting r(x) be deﬁned as x x −1 cos e x+1 −e x ex cos u du u2 ex+1 r(x) = −e we have. ex+1 ]. Exercise 6. Note that the strictness of the inequality is justiﬁed by the fact that cos u ≤ 1 for all u ∈ [ex . ex+1 ] but cos is not a constant function so cos u < 1 for some u ∈ [ex .

15a Using integration by parts with F = f 2 (x) and g = dx gives us b b 1= a f 2 (x) dx = xf 2 (x) b a − a b 2f (x)f (x)x dx which evaluates to 1 = bf 2 (b) − af 2 (a) − a 2f (x)f (x)x dx which.15b Applying Holder’s inequality to the last equation in part (a) gives us −1 = 2 b b 1/2 b 1/2 f (x)f (x)x dx ≤ a a |x f (x)| dx 2 2 · a |[f (x)] | dx 2 Since f (x)f (x) must be negative at some point (by mean value theorem. f (x)f (x)x dx a Exercise 6. becomes b 1=− a 2f (x)f (x)x dx b Dividing both sides by −2 gives us −1 = 2 which is the desired equality. this inequality must be strict: −1 = 2 b b 1/2 b 1/2 f (x)f (x)x dx < a a |x f (x)| dx 2 2 · a |[f (x)] | dx 2 Squaring the ﬁrst and last term in this chain of inequalities.ex+1 |r(x)| ≤ −e x ex ex+1 cos u du u2 1 du u2 1 du u2 = −ex ex ex+1 = = = < |e | ex x |ex | 1 1 − x+1 ex e e−1 e 2 e So we see that ex f (x) = cos ex − e−1 cos ex+1 + r(x) where |r(x)| < 2e−1 . since f (a) = f (b) = 0) while x2 f 2 and [f (x)]2 are strictly positive. Exercise 6. we have 1 < 4 b b |x2 f 2 (x)| dx · a a |[f (x)]2 | dx 118 . since f (a) = f (b) = 0.

this is equivalent to = 1 ns n=1 ∞ Exercise 6. so the norms are redundant: 1 < 4 b b x2 f 2 (x) dx · a a [f (x)]2 dx Exercise 6. as we saw in part (a). n + 1) so this becomes = n=1 ∞ s n n dx xs+1 n+1 x=n = n=1 ∞ sn n −1 sxs = n=1 1 1 − ns (n + 1)s ∞ We then split up the summation into three parts.16b We’re asked to evaluate the integral ∞ x − [x] s −s dx s−1 xs+1 1 Having determined that [x] was integrable in part (a) we can split up the integral as follows: = s −s s−1 ∞ 1 1 dx + s xs ∞ 1 ∞ 1 [x] dx xs+1 Elementary calculus allows us to calculate the left integral: = s s − +s s−1 s−1 ∞ [x] dx xs+1 =s 1 [x] dx xs+1 This.All of the normed values are squares. 1 = n=1 n 1 1 1 + n s− n s n n (n + 1)s n=2 n=1 ∞ ∞ Evaluating the ﬁrst summation and changing the index of the third gives us =1+ n=2 n 1 1 (n − 1) − ns n=2 (n)s n − (n − 1) ns n=2 1 ns n=2 ∞ ∞ ∞ =1+ =1+ And since 1 is clearly equal to 1/ns when n = 1.16a We can integrate the given function separately over inﬁnitely many intervals of length 1: ∞ s 1 [x] dx = s xs+1 n=1 ∞ n+1 ∞ n+1 n [x] dx xs+1 We have [x] = n on the interval [n. is equivalent to = ζ(s) 119 .

Proving this more formally Let > 0 be given. We’re asked to prove that b b α(x)f (x) dx = f (b)α(b) − f (a)α(a) − a a f (x) dα Most of the work is done for us by the theorem 6. xi ] such that α(xi−1 ) − α(xi ) = α (ti )[xi−1 − xi ] or. f. We’re told that f is continuous and that α is monotonically inreasing. at least) by letting f = G and f = g. α) − L(P. f.Exercise 6. and this completes the proof. to express the same equation with diﬀerent notation. We’re told that α is a monotonically increasing function on [a. ∆αi = α (ti )∆xi By theorem 6. On any interval [xi−1 . α) < 2 . but we can prove this more formally. we have b b f (x) dα − a a f (x)α (x) dx < which. b] such that U (P.17 I’m going to change the notation of this problem a bit to make it clearer (to me.22 (integration by parts) which tells us that b b α(x)f (x) dx = f (b)α(b) − f (a)α(a) − a a f (x)α (x) dx So our proof is reduced to simply proving that b b f (x) dα = a a f (x)α (x) dx (80) It’s tempting to appeal to elementary calculus to show that dα = α (x) dx. xi ] the mean value theorem tells us that there is some ti ∈ [xi−1 . b] and that f is continuous. Let P be an arbitrary partition of [a. requires that b b f (x) dα = a a f (x)α (x) dx This proves (80). in order to be true for any > 0. therefore f ∈ R(α). by (82) and (83) and the triangle inequality. 120 .7 (c) we have n b (81) f (ti )∆αi − i=1 a f (x) dα < 2 (82) and also n b f (ti )α (ti )∆xi − i=1 a f (x)α (x) dx < 2 (83) By the inequality (81) we have n n f (ti )∆αi = i=1 i=1 f (ti )α (ti )∆xi Therefore.

one-to-one. the lemma is proven. since Riemann integrals are deﬁned only for bounded functions and this function is unbounded: 1 = lim 2π |sin (2kπ) − 2kπ cos (2kπ)| lim |γ (t) dt| = lim γ t→0 k→∞ k→∞ 2kπ = lim 2π |±2kπ| = 2π sin k→∞ = lim 4π 2 k = ∞ k→∞ Therefore the integral in (84) doesn’t exist. So we have g(f (s)) = g(f (t)) but f (s) = f (t) and therefore s = t. By contrapositive. f (t) = y. 121 . therefore γ2 (t) is rectiﬁable and its length is given by 2π 2π |2ieit | = 0 0 2 = 4π The arc γ3 is not rectiﬁable Proof by contradiction. Lemma 2: If g ◦ f : A → C is one-to-one and f : A → B is one-to-one and onto.27) and its length is given by 2π 2π |ieit | = 0 0 1 = 2π The derivative γ2 (t) = 2ieit is continuous. Therefore g(f (x)) = g(f (y)) → f (x) = f (y) → x = y and so g ◦ f is one-to-one. d] is a continuous. then g ◦ f : A → C is one-to-one By deﬁnition of “one-to-one”.18 The derivative γ1 (t) = ieit is continuous. f (x ) = f (z ) g(s) = g(t) → s = t so that f is not one-to-one. Let f be a continuous real function. which means that γ3 is not rectiﬁable. y such that x ≤ x < y < z < z. Proof by contrapositive. the lemma is proven. y ∈ B such that g(x) = g(y) but x = y. So g ◦ f is not one-to-one. If f were not strictly increasing and not strictly decreasing then we could ﬁnd some x < y < z such that either f (y) ≤ f (x) and f (y) ≤ f (z) or such that f (y) ≥ f (x) and f (y) ≥ f (z). real function then f is either strictly decreasing or strictly increasing Proof by contrapositive. But f is one-to-one and onto. Lemma 3: If f : [a. so there exist unique s. t ∈ A such that f (s) = x. we have f (x) = f (y) → x = y. Suppose that g is not one-to-one.19 Lemma 1: If f : A → B and G : B → C are both one-to-one. Exercise 6. therefore γ1 (t) is rectiﬁable (theorem 6. Then we could ﬁnd x.Exercise 6. If γ3 were rectiﬁable then its length would be given by the integral 2π 2π |γ (t)| dt = 0 0 2πi sin 2π 1 t + −1 2πit cos t2 1 t e2πit sin(1/t) dt 1 1 1 dt (84) − cos t t t 0 But this integral isn’t deﬁned. By contrapositive. b] → [c. which means the arc length of γ3 is not deﬁned. From the intermediate value property of continuous functions we know that we must then be able to ﬁnd x . then g : B → C is one-to-one.

. We saw in lemma 5 that φ−1 (a) = c and φ (b) = d so we have −1 γ1 (a) = γ1 (φ(φ−1 (a))) ≡ γ2 (φ−1 (a)) = γ2 (c) = γ2 (d) = γ2 (φ−1 (b)) ≡ γ1 (φ(φ−1 (b))) = γ1 (b) Therefore γ1 (a) = γ1 (b). x1 . . so its inverse exists and is a continuous mapping from [a. φ−1 (xn+1 ) = d} which is a partition of [c. . which means that γ2 is a closed curve. And γ2 = γ1 ◦ φ so γ2 is one-to-one. . . We saw in lemma 4 that φ(c) = a and φ(d) = b. Proof 1: γ1 is an arc iﬀ γ2 is an arc If γ1 is an arc then γ1 is a one-to-one function. therefore by lemma 1 γ1 ◦ φ is one-to-one. By lemma 3. < φ(xn ) < φ(xn+1 ) = φ(d) = b and therefore {φ(xi )} = {a = φ(x0 ). . Proof 2: γ1 is a closed curve iﬀ γ2 is a closed curve Assume that γ1 is a closed curve so that γ1 (a) = γ1 (b). xn . φ(xn+1 ) = b} which is a partition of [a. b]. xn+1 = d} From the properties of φ we have a = φ(c) = φ(x0 ) < φ(x1 ) < φ(x2 ) < . Lemma 5: If P = {xi } is a partition of γ1 . then P = φ−1 (xi ) is a partition of γ2 We’re told that φ is a continuous one-to-one function from [a. Therefore γ1 is an arc. so φ−1 (b) = d. . which means that γ1 is a closed curve. b]: P = {xi } = {a = x0 . . φ(xn ). so we have γ2 (c) ≡ γ1 (φ(c)) = γ1 (a) = γ1 (b) = γ1 (φ(d)) ≡ γ2 (d) Therefore γ2 (c) = γ2 (d). . φ−1 (x1 ). d]: P = {xi } = {c = x0 . b] to [c. d] (theorem 4. . . this means that φ−1 is either strictly increasing or strictly decreasing. d]. φ−1 (xn ). . If γ2 is an arc then γ2 = γ1 ◦ φ is one-to-one. and since φ(c) = a it must be strictly increasing. There γ2 is an arc. 122 . xn . b] onto [c. Let P be an arbitrary partion of [a. We’re told that φ is a one-to-one and onto function. . . And φ−1 is onto. then P = {φ(xi )} is a partition of γ1 From lemma 3 we know that φ is strictly increasing or decreasing. d]. . . φ(x2 ). < φ−1 (xn ) < φ−1 (xn+1 ) = φ−1 (b) = d and therefore {φ−1 (xi )} = {c = φ−1 (x0 ).Lemma 4: If P = {xi } is a partition of γ2 .17). so φ(d) = b. . . therefore by lemma 2 we know that γ1 is one-to-one. and since φ−1 (a) = c it must be strictly increasing. x1 . . φ(x1 ). Now assume that γ2 is a closed curve so that γ2 (c) = γ2 (d). We’re told that φ is one-to-one. And φ is onto. . φ−1 (x2 ). Let P be an arbitrary partition of [c. xn+1 = b} From the properties of φ−1 we have c = φ−1 (a) = φ−1 (x0 ) < φ−1 (x1 ) < φ−1 (x2 ) < .

Let Mn = sup |fn | and let M = sup |f |. it’s clear that Λ(γ1 ) is ﬁnite iﬀ Λ(γ2 ) is ﬁnite and so γ1 is rectiﬁable iﬀ γ2 is rectiﬁable. There are only ﬁnitely many n ≤ N . . m > N → |fn (x) − fm (x)| < By choosing N ∗ > max{N. M1 . each of which is bounded by some Mn . γ2 ) This means that the set {Λ(P. this becomes n Λ(P. Using the notation from deﬁnition 6. d]: Λ(P. n. . Proof 4: γ1 is rectiﬁable iﬀ γ2 is rectiﬁable Having shown previously that Λ(γ1 ) = Λ(γ2 ).1 Let {fn } be a sequence of functions that converges uniformly to f . γ2 )} so clearly these sets have the same supremums.2a Let > 0 be given. . b]. Because of the uniform convergence of {fn } → f we can ﬁnd some N such that n > N → |fn (x)| = |fn (x) − f (x) + f (x)| ≤ |fn (x) − f (x)| + |f (x)| ≤ + M This tells us that for all n > N the function fn is bounded above by M + . if we let P be an arbitrary partition of [a. we have n Λ(P. M2 . which means that Λ(γ1 ) = Λ(γ2 ). MN } we have found a bound for all fn . 2 . γ1 ) = Λ(φ−1 (P ). γ1 ) Similarly. γ2 ) = i=1 |(γ2 (xi ) − γ2 (xi−1 )| < M From the deﬁnition of γ2 . We’re told that {fn } and {gn } converge uniformly on E so we can ﬁnd N. γ2 ) = Λ(φ(P ). .26. Exercise 7. γ2 ) = i=1 |(γ1 (φ(xi )) − γ1 (φ(xi−1 ))| < M From lemma 4 we see that {φ(xi )} describes a partitioning of [c. So by choosing max{M + . Let > 0 be given.Proof 3: γ1 and γ2 have the same length Let P be an arbitrary partitioning of [a. γ1 )} is identical to the set of {Λ(P. Exercise 7. b] we can use lemma 5 to show that Λ(P . m > N ∗ → |(fn (x) + gn (x)) − (fm (x) − gm (x))| ≤ |fn (x) − fm (x)| + |gn (x) − gm (x)| < which shows that fn + gn converges uniformly on E. M such that n. M } we have n. m > M → |gn (x) − gm (x)| < 2 123 .

Exercise 7.2b

If {fn } and {gn } are bounded functions then by exercise 7.1 they are uniformly bounded, say by F and G. Let > 0 be given and choose δ such that δ < min , , , 3 3F 3G

We’re told that {fn } and {gn } converge uniformly on E so we can ﬁnd N, M such that n, m > N → |fn (x) − fm (x)| < δ, n, m > M → |gn (x) − gm (x)| < δ

|fn gn − fm gm | = |(fn − fm )(gn − gm ) + fm (gn − gm ) + gm (fn − fm )| ≤ |(fn − fm )||(gn − gm )| + |fm ||(gn − gm )| + |gm ||(fn − fm )| ≤ δ 2 + |fm |δ + |gm |δ ≤ δ 2 F δ + Gδ

2

≤

3

+F

3F

+G

3G

=

Exercise 7.3

Let {fn } be a sequence such that fn (x) = x. This function obviously converges uniformly to the function 1 f (x) = x on the set R. Let {gn } be a sequence of constant functions such that gn (x) = n . This sequence obviously converges uniformly to the function g(x) = 0. Their product is the sequence {fn gn } where fn (x)gn (x) = x/n. It’s clear that this sequence converges pointwise to fn (x)gn (x) = 0. To show that fn gn is not uniformly convergent, let choose t ∈ R such that t > (n(n + 1)). We now have |fn (t)gn (t) − fn+1 (t)gn+1 (t)| = > 0 be given. Choose an arbitrarily large n ∈ Z and (n(n + 1)) = n(n + 1)

t t t − = > n n+1 n(n + 1)

This shows that the necessary requirements for uniform convergence given in theorem 7.8 do not hold.

Exercise 7.4

Holy shit this exercise is a mess. For what values of x does the series converge absolutely? (Incomplete) For values of x > 0 we have 1 1 = 2x 1+n x 1 1 ≤ 2 1/x + n x 1 n2

By the comparison test this shows that f (x) converges absolutely. If x = 0 then we have 1 = 1 + n2 x |1| = ∞

This series clearly doesn’t converge, absolutely or otherwise. If x < 0 things get more complicated: If x = −1/n2 for any n ∈ N then the nth term of the series is undeﬁned and therefore f (x) is undeﬁned. If x = −1/n2 for any n ∈ N then we can use the fact that For what intervals does the function converge uniformly? If E is any interval of the form [a, b] with a > 0 then we have sup |fn (x)| = fn (a) = 1 1 + n2 a

124

And therefore we have sup |fn (x)| =

1 1 ≤ 2a 1+n a

1 n2

This shows that sup |fn (x)| converges by the comparison test, and so by theorem 7.10 we see that the series fn (x) converges uniformly on E. If E is any interval of the form [a, b] with b < 0 that does not contain any elements of the form −1/n2 , n ∈ N For what intervals does the function fail to converge uniformly? Deﬁne the set X = {xn } with xn = −1/n2 . The function f will fail to converge uniformly on any interval that contains an element of X ∪ 0 or has an element of X ∪ 0 as a limit point of E. Proof: Let E be an arbitrary interval. If E contains any xn ∈ X then f (xn ) undeﬁned and so f fails to converge uniformly on E. If E contains 0 then f (0) = 1 = ∞, but we will never ﬁnd some ﬁnite N such that 1 − f (0) < so it’s clear that f fails to converge uniformly on E. Now suppose that some xn ∈ X is a limit point of E. The nth term of f is unbounded near xn , so f is unbounded near xn , and therefore limt→xn f (t) = ∞. From this we have

n→∞ t→xn N 1

**lim lim f (t) = lim ∞ = ∞
**

n→∞

On the other hand, if we ﬁrst ﬁx a value of t and take the limit of f as n → ∞ we have lim f (t) = lim 1 =0 1 + n2 t

n→∞

n→∞

and therefore

t→xn n→∞

**lim lim f (t) = lim 0 = 0
**

t→xn

Exercise 7.5

If x ≤ 0 or x > 1 then x = 0 for all n and therefore in these cases limn→∞ fn (x) = 0. For any other x, choose an integer N large enough so that N > 1/x. For all n > N we now have x > 1/n and therefore fn (x) = 0, so for this case we have limn→∞ fn (x) = 0. This exhausts all possible values of x, so {fn } converges pointwise to the continuous function f (x) = 0. To show that this function doesn’t converge uniformly to f (x) = 0 let 1 > arbitrarily large integer. We can easily verify that fn 1 n + 1/2 = sin2 (nπ + π/2) = 1 > 0 be given and let n be an

and therefore the deﬁnition of uniform convergence is not satisﬁed. The last part of the question asks us to use the series fn to show that absolute convergence for all x does not imply uniform convergence. The proof of this is simple: we’ve already shown that {fn } is not uniformly convergent but |fn (x)| converges to f (x) = | sin2 (π/x)| so |fn | is absolutely convergent.

Exercise 7.6

To show that the series doesn’t converge absolutely for any x: (−1)n x2 + n = n2 x2 1 + ≥ 2 n n 1 n

The rightmost sum is the harmonic series, which is known to diverge. By comparison test the leftmost series must also diverge.

125

To show that the series converges uniformly in every bounded interval [a, b], let be sup{|a|, |b|}. Deﬁne the partial sum fm to be

m

> 0 be given. Deﬁne X to

**fm = We can rearrange this algebraically to form
**

m

(−1)n

n=1

x2 + n n2

fm =

(−1)n

n=1

1 1 + x2 (−1)n 2 n n n=1 1/n2

m

We know that the alternating harmonic series converges (theorem 3.44, or example 3.40(d)) and that converges (theorem 3.28). Therefore by the Cauchy criterion for convergence we can ﬁnd N such that p>q>N →| and we can also ﬁnd M such that

p

1 (−1)n | < n 2 n=q

p

p>q>M →|

n=q

(−1)n

1 |< n2 2|X|2

**So, by choosing p > q > max{N, M } we have (for all x ∈ [a, b]):
**

p

|fp (x) − fq (x)| =

n=q p

(−1)n (−1)n

n=q

1 1 + x2 (−1)n 2 n n n=q 1 1 (−1)n 2 + x2 n n n=q

p

p

≤ ≤ ≤ + +

2 2

x2 2X 2 2 =

Exercise 7.7

We can establish a global maximum for |fn (x)| The derivative of fn (x) is fn (x) = 1 + nx2 − 2nx2 1 − nx2 = (1 + nx2 )2 (1 + nx2 )2

√ This derivative is zero only when nx2 = 1, which occurs only when x = ±1/ n, at which point we have fn ±1 √ n ±1 = √ 2 n

These extrema must be the global√ extrema for the function, since fn has no asymptotes and fn (x) → 0 as x → ±∞. Therefore |fn (x)| ≤ 1/(2 n). fn converges uniformly to f (x) = 0 Clearly fn converges pointwise to f (x) = 0, since for any x we have

n→∞

lim fn (x) = lim

n→∞

x =0 1 + nx2

126

so that p > q > N implies |cn | satisﬁes the Cauchy convergence criterion. 2. From the previously established bounds. we now have 1 |fn (x) − f (x)| = |fn (x) − 0| = |fn (x)| ≤ √ < 2 n By theorem 7. b)). so we can ﬁnd N such p |cn | < n=q For the same values of p > q > N we also have. The limit of fn is given by: lim fn (x) = lim n→∞ n2 x4 n→∞ 1 − nx2 = + 2nx2 + 1 0 1 x=0 x=0 This shows that it’s not necessarily true that [lim fn ] = lim[fn ] even if fn converges uniformly. Now suppose that t a limit point of {xn }. . By the Cauchy convergence of |cn | we can ﬁnd N such that |cn | < . From this we have ∞ n=N ∞ ∞ |s − t| < δ → f (s) − f (t) ≤ n=N +1 (cn I(s − xn ) − cn I(t − xn )) ≤ n=N +1 |cn [I(s − xn ) − I(t − xn )]| 127 . This t either is or isn’t a limit of some subsequence of {xn }. . by the triangle inequality and the fact that I(x − xn ) ≤ 1. If t is not a limit point of some subsequence then we can ﬁnd some neighborhood around t that contains no points of {xn }. p p p |fp − fq | = n=q cn I(x − xn ) ≤ n=q |cn I(x − xn )| ≤ n=q |cn | < so {fn } converges uniformly by theorem 7. so clearly f (x) = 0 for all x. And if f is constant on an interval around t then it is clearly continuous at t.9 this is suﬃcient to prove that fn converges uniformly to f (x) = 0. N .8 Proof of continuity Let t be an arbitrary point (not necessarily in the interval (a. Exercise 7. then I(s − xn ) = I(t − xn ) for n = 1. the function I(t − xn ) is constant on this interval for all n and therefore f (x) = cn I(x − xn ) is constant on this interval for all n. Let fm represent the partial sum m fm = n=1 cn I(x − xi ) We’re told that |cn | converges. . When does f (x) = lim fn (x)? We’ve established that f (x) = 0.8 Proof of uniform convergence Let > 0 be given. let δ represent the radius of this neighborhood.√ Now let > 0 be given and choose n suﬃciently large so that 1/(2 n) < : this can be done by choosing 1 n > 4 2 . If we choose s such that |s − t| < δ. Choose a neighborhood around t small enough that it does not contain the ﬁrst N terms of {xn }. (see below for a a more thorough justiﬁcation of this claim). .

so cn I(x − xn ) has the same value for every x ∈ Nδ (t). so we have ∞ ∞ |s − t| < δ → f (s) − f (t) ≤ n=N +1 |cn [I(s − xn ) − I(t − xn )]| ≤ n=N +1 |cn | < That is. which means that n > N → |fn (xn ) − f (xn )| < 2 (85) We’re also told that {fn } is a uniformly convergent sequence of continuous functions. We have therefore shown that f is continuous at t under all cases. so the converse does not hold. if t is not a limit point of some subsequence then we can ﬁnd some neighborhood around t that contains no points of {xn }. means that f is continuous at t. we have |fn (xn ) − f (x)| ≤ |fn (xn ) − f (xn )| + |f (xn ) − f (x)| < Converse statement 1 Suppose the converse is “Let {fn } be a sequence of continuous functions that converges uniformly to f . If |s − t| < δ then there are no elements of {xn } between s and t and so every element of A is greater than both t and s and every element of B is smaller than both t and s. and since {xn } → x we can ﬁnd M such that n > M → |f (xn ) − f (x)| < Using the triangle inequality and equations (85) and (86). Is it true that if lim fn (xn ) = f (x) then {xn } → x?”. so f is continuous. we’ve found δ such that |s − t| < δ → f (s) − f (t) < which. so it must also hold for the elements of {xn }. and let B be the set of elements of {xn } that are smaller than t. But t was an arbitrary point (not necessarily conﬁned to (a. so we have n→∞ 2 (86) 2 + 2 = lim fn (xn ) = lim fn (1) = 0 = f (0) n→∞ It’s clear that {xn } → 0. so we can ﬁnd N such that n > N → |fn (t) − f (t)| < 2 for all t ∈ E This holds for all t ∈ E. This means that I(x − xn ) has the same value for every x ∈ Nδ (t).9 We’re told that {fn } → f uniformly on E.Each I(x − xn ) is either 0 or 1 so |I(s − xn ) − I(t − xn )| ≤ 1. The answer to this question is “no”. From this we see that I(s − xn ) = I(t − xn ) = 0 if xn ∈ A and I(s − xn ) = I(t − xn ) = 1 if xn ∈ B. and therefore f (x) has the same value for every x ∈ Nδ (t). 1]. Consider the function fn (x) = x/n and the sequence {xn } = {1} (an inﬁnite sequence of 1s). Exercise 7. by deﬁnition. b)) and therefore f is continuous at all points. The sequence {fn } converges uniformly to f (x) = 0 on the set [0. Justiﬁcation of I(x − xn ) being constant on some interval As mentioned above. Let A be the set of elements of {xn } that are greater than t. 128 .

despite the fact that lim fn (xn ) = f (x) whenever {xn } → x ∈ E. so we can choose δ such that 0<δ< 1 min{(nx). Let fn (x) = 1/(nx). it’s clear that fn (x) does not converge uniformly to f (x) = 0 on E (to prove this. Although 0 is a limit point of E.10a f is continuous at every irrational number We know that 1/n2 converges. Let E be the harmonic set {1/n}. That is: |x − y| < δ → |f (x) − f (y)| < which means that f is continuous at x. More importantly. We now derive the following chain of inequalities: N ∞ (nx) (nx ± nδ) (nx) (nx ± nδ) |f (x) − f (x ± δ)| = − − + 2 2 n n n2 n2 1 N +1 N ≤ 1 (nx) (nx ± nδ) − n2 n2 N ∞ + N +1 (nx) + n2 + ∞ N +1 (nx ± nδ) n2 < 1 N (nx) (nx ± nδ) − n2 n2 4 + 4 < 1 (nx) (nx) (nδ) − 2 ± 2 n2 n n N + 2 = 1 ±(nδ) n2 + 2 We chose our value of δ after we ﬁxed a particular value of N . choose any and any n and then choose x < (n )−1 ). but for all y such that |x − y| < δ. so by the Cauchy criterion we can choose N such that b b>a>N → a 1 < n2 4 The fractional part of (nx) will never be zero because x is irrational. we have |f (x) − f (x ± δ)| < 2 + 2 = And this would hold not just for x ± δ. Then fn converges pointwise on E to f (x) = 0. So the only sequences of points xn ∈ E such that {xn } → x ∈ E are sequences where every term of the sequence is eventually just x itself (that is. 1 − (nx) : n < N } 2 This guarantees that 0 < (n[x − δ]) < (nx) for all n < N and (nx) < (n[x + δ]) < 1 for all n. Is it true that {fn } converges uniformly to f on E? The answer to this question is “no”. Exercise 7. 0 is not an element of E and this converse statement is only concerned with sequences {xn } that converge to some x ∈ E. this choice of δ guarantees that (n[x − δ]) = (nx) − (nδ) for n < N . So clearly for every such sequence we must have lim fn (xn ) = lim fn (x) = f (x). sequences for which there is some N such that n > N → xn = x). But. so we could have chosen δ small enough to make this sum less than /2.Converse statement 2 Suppose the converse is “Let {fn } be a sequence such that lim fn (xn ) = f (x) for some function f and for all sequences of points xn ∈ E such that {xn } → x ∈ E. 129 . In doing so.

then we would have |f (x+) − f (x−)| = 0. Instead.11): ∞ ∞ fn (x+) = f (x+). But when we actually calculate this diﬀerence we ﬁnd: ∞ ∞ |f (x+) − f (x−)| = n=1 fn (x+) − n=1 fn (x−) We determined that fn (x+) = fn (x−) unless nx is an integer. so we also have (by theorem 7. n=1 n=1 fn (x−) = f (x−) We know that 1/n2 converges. n2 n lim (nt) 1 = 2 n2 n t→x+ t→x− We know by lemma 1 that (nx)/n2 converges uniformly to f . Let >0 This is an immediate consequence of the fact that (nx) < 1 for all n. at which point nx = [mq]x = mq[p/q] = mp.10b We’ve shown that the discontinuities of f are the rational numbers. 130 . so by the Cauchy criterion we can choose N such that b b>a>N → a 1 < n2 4 If f were continuous at our rational point x. this occurs when n is a multiple of q. Most of the terms of the summations cancel out. we have lim (nt) 0 = 2. Exercise 7. But x was an arbitrary rational number. By the Cauchy criterion we can choose N such that b b>a>N → a 1 < n2 N and therefore we have f (x) − ∞ (nx) = n2 n=1 ∞ N (nx) (nx) − 2 n n2 n=1 n=1 (nx) < n2 ∞ ∞ = n=N +1 (nx) ≤ n2 n=N +1 n=N +1 1 < n2 f is discontinuous at every rational number When nx is not an integer the following limits are identical: lim (nt) (nx) (nt) = 2 = lim t→x− n2 n2 n t→x+ When nx is an integer then this equality no longer holds. leaving us with ∞ ∞ |f (x+)−f (x−)| = m=1 fmq (x+) − m=1 fmq (x−) = (mq) (mq) − lim = t→x+ [mq]2 t→x− [mq]2 m=1 lim ∞ 1 (0) − = n2 [mq]2 m=1 ∞ 1 [mq]2 m=1 ∞ This is clearly not equal to zero and it can’t be made arbitrarily small because we have no freedom to select particular values for m or q. and therefore f is not continuous at any rational number. and these are clearly countable and clearly dense in R. Therefore f (x+) = f (x−) and therefore f is not rational at x. x and that be given.Lemma 1: (nx)/n2 converges uniformly to f 1/n2 converges.

To do this. we have q q−1 q−1 |Aq − Ap−1 | = k=p fk gk = k=p An (gn − gn+1 ) + Aq gq − Ap−1 gp ≤ M k=p (gn − gn+1 ) + gq − gp 131 . we choose q > p > N . After a bunch of trivial calculus. Using this variable substitution in the previous integral: b−1 n−1 2 b j+1 (nx) (u) dx = du n n3 a j=0 j k=a Again. Following the logic of theorem 3. therefore therefore f ∈ R. so we can ﬁnally start integrating this thing. When x = j/n we have 1 u = j. We’re told that gn → 0 uniformly. k + 1) without changing the value of the integral.42. b−1 2 b k+1 (nx) (nx) dx dx = n n2 a k k=a For 0 ≤ δ < 1 we have (n(k + δ)) = (nk + nδ) = (nδ). b−1 2 b 1 (nx) (nx) dx dx = n n2 a 0 k=a As x ranges from 0 to 1. nx ranges from 0 to n. So we split up the interval (0. there is no diﬀerence between the value of (u) on the intervals (j. 1): b a (nx) dx = n 2 b−1 n−1 0 1 k=a j=0 (u) du n3 On the interval (0. f (x) exists and Exercise 7. Assume without loss of generality that a and b are integers. 1) into n intervals of length 1/n: b a (nx) dx = n 2 b−1 n−1 [j+1]/n j/n k=a j=0 (nx) dx n2 To make this integral a bit more manageable we make the variable substitution u = nx.11 Let > 0 be given. 1) we have (u) = u. We’re told that fn has uniformly bounded partial sums. Let [a. when x = [j + 1]/n we have u = j + 1. and we have dx = n du. we have 2 b u2 b−a−1 (nx) dx = [b − a − 1][n] 3 = n 2n n2 a And therefore a b b ∞ f (x) = a (nx) = n2 n=1 n=1 ∞ b a (nx) b−a−1 = n2 n2 n=1 b a ∞ We know this rightmost sum converges (speciﬁcally. so we can ﬁnd N such that n > N → gn (x) < n M for all x Let An represent the partial sum k=1 fk gk . Let n ∈ N be given. 1) instead of (k.10c To prove that f ∈ R we need only show that (nx)/n2 ∈ R for any ﬁxed value of n.16 to show that (nx)/n2 = (nx)/n2 = f .Exercise 7. j + 1) and the value of (u) on the interval (0. so let M represent the upper bound of the partial sums of | fn |. Our result is proven if we can prove that {An } satisﬁes the Cauchy converge criterion. so we can integrate over (0. b] be an arbitrary interval. We can then use the fact that (nx)/n2 converges uniformly to f (lemma 1) and theorem 7. it converges to π 2 [b − a − 1]/6).

leaving us with q−1 |Aq − Ap−1 | ≤ M gp − k=p (gn − gn+1 ) + gq + gp = 2M |gp | ≤ 2M |gN | = Exercise 7. so for all x and any speciﬁed value for c we can ﬁnd some M such that m > M → |fm (x) − f (x)| < from which we can conclude that c c 2c (88) m>M → 0 fm − f ≤ 0 |fm − f | ≤ c 2c = 2 (89) So. for the given value of > 0 we choose c large enough that (88) holds and then. we choose n to be large enough that (89) holds.30) and is strictly increasing (because g > 0) and has an upper bound of M . based on this choice of c. we know that lim G(t) = M t→∞ we can ﬁnd some N such that n > N → M − G(n) < which is equivalent to saying that ∞ c ∞ 2 n>N → 0 g− 0 g= c g< 2 (87) We’re given that fn → f uniformly. By the triangle inequality we then have ∞ c ∞ fn − f ≤ 0 c 0 fn − f + c ∞ fn − f ≤ 0 |fn − f | + c |fn − f | We can make this hold for arbitrary + = 2 2 by taking n suﬃciently large. Deﬁne G(t) to be G(t) = 0 g(x) dx Since G(t) is continuous (theorem 6. Let M = ∞ 0 g. We’re told that 0 ≤ fn .12 Let > 0 be given. which is simply saying that ∞ n→∞ ≤ lim fn − f = 0 0 ∞ ∞ from which we conclude n→∞ lim fn = 0 0 f 132 .The (gn − gn+1 ) terms telescope down. f ≤ g and that c g < ∞.

1] (theorem 3. b] with ﬁnitely many intervals of length δ/2 ([a. From each interval we choose a rational number qi . n→∞ if t is irrational and each Fn is continuous at t Exercise 7. The rational numbers are dense in R so we can construct a sequence {rn } of rational numbers such that {rn } → t. qi+1 such that qi < x < qi+1 and |qi+1 − qi | < δ. Let > 0 be given. so we can ﬁnd some subsequence of functions {fn } for which {fn (x1 )} and {fn (x2 )} both converge. We’re given that f is continuous on [a. so limk→∞ Fn (rk ) = Fn (t) for every n. Therefore we simply deﬁne the value of f (t) to be f (t) = lim Fn (t). Each Fn is continuous at t. If x = qi then we can ﬁnd qi . Therefore there exists some subsequence of functions {fn } for which the subsequence of real (1) (2) numbers {fn (x1 )} converges to some point in [0.13a For any choice of x1 ∈ [0.13). 1]. By the triangle inequality: |Fn (x) − f (x)| ≤ |Fn (x) − Fn (qi+1 )| + |Fn (qi+1 ) − f (qi+1 )| + |f (qi+1 ) − f (x)| Each Fn is monotonic. We have constructed {Fn } so that {Fn (t)} converges to some point.12.13: example An example where a sequence {fn } of monotonically increasing functions converges pointwise to f and f is not continuous: 0 x≤0 1 x≥1 1 − 1+n fn (x) = 1 1 − 1+nx 0 < x < 1 n→∞ lim fn (x) = 0 1 x≤0 x>0 Exercise 7. so we have |Fn (x) − Fn (qi+1 )| ≤ |Fn (qi ) − Fn (qi+1 )| so this last inequality becomes |Fn (x) − f (x)| ≤ |Fn (qi ) − Fn (qi+1 )| + |Fn (qi+1 ) − f (qi+1 )| + |f (qi+1 ) − f (x)| 133 . 1]. 1]. From the monotonicity of f we have x < a → |f (x) − α| < .19). If x = qi for some i then by (92) we have n > N → |Fn (qi ) − f (qi )| < . the sequence {fn (x1 )} is a bounded sequence of real numbers in the compact (1) domain [0.13b Let α = inf f (x) and let β = sup f (x) (these values may be ±∞). therefore we can ﬁnd some δ such that |x − y| < δ → |f (x) − f (y) < (91) We can cover [a. so we simply deﬁne f (t) for rational t to be f (t) = lim Fn (t).Exercise 7. x > b → |f (x) − β| < (90) Proving uniform convergence is now a matter of proving uniform convergence on [a. therefore it’s uniformly continuous on [a.30. b]. Now let t be a rational number or a point of discontinuity for some fn . so we don’t even need to rely on compactness for this). b] is ﬁnite. and 2. and the set of qi s is a ﬁnite set. Let a be a point at which |f (a) − α| < and let b be a point at which |f (b) − β| < . Call this subsequence of functions {fn } (2) and choose any x2 ∈ [0. b]. We know that f converges pointwise at each rational number. 1]. b] (theorem 4. This can be repeated a countable number of times to construct a sequence of functions {Fn } for which {Fn (x)} converges at all rational numbers and for all points of discontinuity for each fn (this set is countable by theorems 4. The function f must come arbitrarily close to its supremum and inﬁmum on R. The sequence {f (x2 ) is still a bounded sequence of real numbers in the compact (3) domain [0.2.6a). b]. so we can ﬁnd some integer N such that n > N → |Fn (qi ) − f (qi )| < for each qi (92) Now choose any x ∈ [a. n→∞ (1) if t is rational or a point of discontinuity Now t be a rational number and a point at which fn is continuous for all n.

But the function f (eiθ ) = −eiθ is a continuous function that is not in the closure of A . by theorem 7. Lemma 2: {Fn } is equicontinuous Let > 0 be given. which means that there exists some δ such that. 1) and all n ∈ N.The term |Fn (qi+1 ) − f (qi+1 )| is < by (92). Exercise 7. Each fn is equicontinous on [0. Exercise 7.19 This was solved during class Exercise 7.15 Let > 0 be given. Let δ = /M .25. we have x x b |Fn (x)| = a fn (t) dt ≤ a |fn (t)| dt ≤ a |fn (t)| dt ≤ (b − a)M and therefore {|Fn |} is uniformly bounded. |y| < δ → |f (0) − f (ny)| < By taking n suﬃciently large and choosing y appropriately we can cause ny to take on any value in (0. For this same function. 134 . we know that {Fn } contains a uniformly convergent subsequence.17 Exercise 7. For each n. ∞). 1].18 Lemma 1: {Fn } is uniformly bounded We’re told that {fn } is a uniformly bounded sequence: let M be the upper bound of {|fn |}. So we conclude that f is a constant function on [0. |f (eiθ )| = |eiθ | = 1 and so A vanishes at no point of K.16 See part (b) of exercise 7.14 Exercise 7.21 We know that A separates points because f (eiθ ) = eiθ ∈ A . The term |f (qi+1 ) − f (x)| is < |Fn (x) − f (x)| ≤ |Fn (qi ) − Fn (qi+1 )| + 2 by (91). y y |y − x| < δ → |Fn (y) − Fn (x)| = x fn (t) dt ≤ x |fn (t) dt| ≤ (y − x)M < δM < Therefore.13 Exercise 7. This leaves us with Additional applications of the triangle inequality gives us |Fn (x) − f (x)| ≤ |Fn (qi ) − f (qi )| + |f (qi ) − f (qi+1 )||f (qi ) − Fn (qi+1 )| + 2 The terms |Fn (qi ) − f (qi )| and |f (qi ) − Fn (qi+1 )| are both < by (92). for all n. |0 − y| < δ → |fn (0) − fn (y)| < By the deﬁnition of f this means that. ∞). for all y ∈ (0. The term |f (qi ) − f (qi+1 )| is < (91). This leave us with |Fn (x) − f (x)| ≤ 5 by Exercise 7.

Sign up to vote on this title

UsefulNot useful- Rudin Solution 1-8
- rudin solutions
- Baby Rudin
- Real and Complex Analysis Solutions Manual
- Rudin's Principles of Mathematical Analysis Solucionario
- Rudin - Real and Complex Analysis - Solutions
- Real Analysis Rudin Solution Manual(1-8)(10-13)
- Solutions to Real and Complex Analysis-By Walter Rudin-mathematic87.Blogfa
- rudin solutions 1.pdf
- rudin w ,solution manual of principles of mathematical analysis
- Rudin Solution to Real and Complex Analysis(Chapter 1-6)
- Solutions to Rudin Principles of Mathematical Analysis
- Rudin
- Principles of Mathematical Analysis
- rudin
- Dummit Solutions
- Royden Real Analysis Solutions
- Elementary Classical Analysis - Jerrold E. Marsden & Michael J. Hoffman
- Solutions Topology James Munkres Solutions
- rudin ch 11
- Solutions Rudin_2.pdf
- Advanced Calculus of Several Variables
- 34158129 Calculus on Manifolds Spivak M PDF
- Principles of Mathematical Analysis- Third Edition- Walter Rudin
- [Walter Rudin] Principios de Analisis Matematico
- m104_Rudin_exs
- Solucionario Rudin
- Apostol solution chapter 1
- Solutions Hatcher
- Problems and Solutions in Real and Complex Analysis (DeMeo)
- Rudin Solutions Ch 1 - 7