<< . .

. 7
( : 70)



. . >>

writing things down correctly. We split the proof into two parts.
Necessity Suppose f is continuous and U is an open set in Rp . If x ∈ f ’1 (U ),
then f (x) ∈ U . But U is open, so there exists an > 0 such that B(f (x), ) ⊆
U . Since f is continuous at x, we can ¬nd a δ > 0 such that

f (x) ’ f (y) < whenever x ’ y < δ.

We thus have B(x, δ) ⊆ f ’1 (U ). It follows that f ’1 (U ) is open.
Su¬ciency Suppose that f ’1 (U ) is open whenever U is an open subset of Rp .
Let x ∈ Rm and > 0. Since B(f (x), ) is open, it follows that f ’1 (B(f (x), ))
is open. But x ∈ f ’1 (B(f (x), )), so there exists a δ > 0 such that B(x, δ) ⊆
f ’1 (B(f (x), )). We thus have

f (x) ’ f (y) < whenever x ’ y < δ.

It follows that f is continuous.
Exercise 4.2.18. Show that sin((’5π, 5π)) = [’1, 1]. Give examples of
bounded open sets A in R such that (a) sin A is closed and not open, (b)
sin A is open and not closed, (c) sin A is neither open nor closed, (d) sin A
is open and closed. (Observe that … is automatically bounded.)
The reader may object that we have not yet derived the properties of sin.
In my view this does not matter if we are merely commenting on or illustrat-
ing our main argument. (I say a little more on this topic in Appendix C.)
However, if the reader is interested, she should be able to construct a poly-
nomial P such that (a), (b), (c) and (d) hold for suitable A when sin A is
replaced by P (A).
The next exercise gives a simple example of how Lemma 4.2.17 can be
used and asks you to contrast the new ˜open set™ method with the old ˜ , δ
method
Exercise 4.2.19. Prove the following result, ¬rst directly from De¬nition 4.2.14
and then by using Lemma 4.2.17 instead.
If f : Rm ’ Rp and g : Rp ’ Rq are continuous, then so is their
composition g —¦ f .
55
Please send corrections however trivial to twk@dpmms.cam.ac.uk

(Recall that we write g —¦ f (x) = g(f (x)).)
The reader who has been following carefully may have observed that we
have only de¬ned limits of sequences. Here is another notion of limit which
is probably familiar to the reader.
De¬nition 4.2.20. Let E ⊆ Rm , x ∈ E and a ∈ Rp . Consider a function
f : E \ {x} ’ Rp (or4 a function f : E ’ Rp ). We say that f (y) ’ a as
y ’ x through values of y ∈ E if, given > 0, we can ¬nd a δ( , x) > 0
such that, whenever y ∈ E and 0 < x ’ y < δ( , x), we have

f (x) ’ a < .

(We give a slightly more general de¬nition in Exercise K.23.)
Exercise 4.2.21. Let E ⊆ Rm , x ∈ E and a ∈ Rp . Consider a function
f : E \ {x} ’ Rp . De¬ne ˜ : E ’ Rp by ˜(y) = f (y) if y ∈ E \ {x} and
f f
˜(x) = a. Show that f (y) ’ a as y ’ x through values of y ∈ E if and only
f
if ˜ is continuous at x.
f
Exercise 4.2.22. After looking at your solution of Lemma 4.2.15, state and
prove the corresponding results for limits.
Exercise 4.2.23. [In this exercise you should start from De¬nition 1.7.2]
Let U be an open set in R. Show that a function f : U ’ R is di¬erentiable
at t ∈ U with derivative f (t) if and only if
f (t + h) ’ f (t)
’ f (t)
h
as h ’ 0 (through values of h with t + h ∈ U ).
Exercise 4.2.24. In Chapter 6 we approach the properties of di¬erentia-
tion in a more general manner. However the reader will probably already
have met results like the following which can be proved using Exercises 4.2.22
and 4.2.23.
(i) If f, g : (a, b) ’ R are di¬erentiable at x ∈ (a, b), then so is the sum
f + g and we have (f + g) (x) = f (x) + g (x).
(ii) If f, g : (a, b) ’ R are di¬erentiable at x ∈ (a, b), then so is the
product f — g and we have (f — g) (x) = f (x)g(x) + f (x)g (x). [Hint: f (x +
h)g(x + h) ’ f (x)g(x) = (f (x + h) ’ f (x))g(x + h) + f (x)(g(x + h) ’ g(x)).]
(iii) If f : (a, b) ’ R is nowhere zero and f is di¬erentiable at x ∈ (a, b),
then so is 1/f and we have (1/f ) (x) = ’f (x)/f (x)2 .
4
Thus it does not matter whether f is de¬ned at x or not (and, if it is de¬ned, it does
not matter what the value of f (x) is).
56 A COMPANION TO ANALYSIS

(iv) State accurately and prove a result along the lines of (ii) and (iii)
dealing with the derivative of f /g.
(v) If c ∈ R, c = 0 and f : R ’ R is di¬erentiable at x, show that the
function fc de¬ned by fc (t) = f (ct) [t ∈ R] is di¬erentiable at c’1 x and we
have fc (c’1 x) = cf (x). What happens if c = 0?
(vi) Use part (ii) and induction on n to show that if rn (x) = xn , then rn
is everywhere di¬erentiable with rn (x) = nrn’1 (x) [n ≥ 1]. Hence show that
every polynomial is everywhere di¬erentiable. If P and Q are polynomials
and Q(t) = 0 for all t ∈ (a, b) show that P/Q is everywhere di¬erentiable on
(a, b).

Exercise 4.2.25. (i) Use part (ii) of Exercise 4.2.24 to show that, if f, g :
(a, b) ’ R satisfy the equation f (t)g(t) = 1 for all t ∈ (a, b) and are di¬er-
entiable at x ∈ (a, b) then g (x) = ’f (x)/f (x)2 .
(ii) Explain why we can not deduce part (iii) of Exercise 4.2.24 directly
from part (i) of this exercise. Can we deduce the result of part (i) of this
exercise directly from part (iii) of Exercise 4.2.24?
(iii) Is the following statement true or false? If f, g : (a, b) ’ R are
di¬erentiable at x ∈ (a, b) and f (x)g(x) = 1 then g (x) = ’f (x)/f (x)2 .
Give a proof or counterexample.

Exercise 4.2.26. From time to time the eagle eyed reader will observe state-
ments like

˜f (x) ’ ∞ as x ’ ’∞™

which have not been formally de¬ned. If this really bothers her, she is probably
reading the wrong book (or the right book but too early). It can be considered
a standing exercise to ¬ll in the required details.
In Appendix D, I sketch a method used in Beardon™s elegant treatment [2]
which avoids the need for such repeated de¬nitions.


4.3 A central theorem of analysis
In this section we prove Theorem 4.3.4 which says that a real-valued con-
tinuous function on a closed bounded set in Rm is bounded and attains its
bounds. This result together with the intermediate value theorem (proved as
Theorem 1.6.1) and the mean value inequality (proved as Theorem 1.7.1 and
later in a more general context as Theorem 6.3.1) are generally considered
to be the central theorems of elementary analysis.
Our next result looks a little abstract at ¬rst.
57
Please send corrections however trivial to twk@dpmms.cam.ac.uk

Theorem 4.3.1. Let K be a closed bounded subset of Rm and f : K ’ Rp
a continuous function. Then f (K) is closed and bounded.
Thus the continuous image of a closed bounded set is closed and bounded.
Proof. By Theorem 4.2.2 (ii), we need only prove that any sequence in f (K)
has a subsequence converging to a limit in f (K).
To this end, suppose that yn is a sequence in f (K). By de¬nition, we can
¬nd xn ∈ K such that f (xn ) = yn . Since K is closed and bounded subset,
Theorem 4.2.2 (i) tells us that there exist n(j) ’ ∞ and x ∈ K such that
xn(j) ’ x as j ’ ∞. Since f is continuous,

yn(j) = f (xn(j) ) ’ f (x) ∈ f (K)

and we are done.
Exercise 4.3.2. Let N : Rm ’ R be given by N (x) = x . Show that
N is continuous. Deduce in particular that if xn ’ x as n ’ ∞, then
xn ’ x .
Exercise 4.3.3. (i) Let A be the open interval (0, 1). Show that the map
f : A ’ R given by f (x) = 1/x is continuous but that f (A) is unbounded.
Thus the continuous image of a bounded set need not be bounded.
(ii) Let A = [1, ∞) = {x ∈ R : x ≥ 1} and f : A ’ R be given by
f (x) = 1/x. Show that A is closed and f is continuous but f (A) is not
closed. Thus the continuous image of a closed set need not be closed.
(iii) Show that the function π : R2 ’ R given by π(x, y) = x is continu-
ous. (The function π is called a projection.) Show that the set

A = {(x, 1/x) : x > 0}

is closed in R2 but that π(A) is not.
We derive a much more concrete corollary.
Theorem 4.3.4. Let K be a closed bounded subset of Rm and f : K ’ R a
continuous function. Then we can ¬nd k1 and k2 in K such that

f (k2 ) ¤ f (k) ¤ f (k1 )

for all k ∈ K.
Proof. Since f (K) is a non-empty bounded set, it has a supremum M say.
Since f (K) is closed, M ∈ f (K), that is M = f (k1 ) for some k1 ∈ K. We
obtain k2 similarly.
58 A COMPANION TO ANALYSIS

In other words, a real-valued continuous function on a closed bounded
set is bounded and attains its bounds. Less usefully we may say that, in this
case, f actually has a maximum and a minimum. Notice that there is no
analogous result for vector-valued functions. Much popular economic writing
consists of attempts to disguise this fact (there is unlikely to be a state of
the economy in which everybody is best o¬).
Exercise 4.3.5. When I was an undergraduate, we used another proof of
Theorem 4.3.4 which used lion hunting to establish that f was bounded and
then a clever trick to establish that it attains its bounds.
(i) We begin with some lion hunting in the style of Exercise 4.1.14. As
in that exercise, we shall only consider the case m = 2, leaving the general
case to the reader. Suppose, if possible, that f (K) is not bounded above (that
is, given any κ > 0, we can ¬nd a x ∈ K such that f (x) > κ).
Since K is closed and bounded, we can ¬nd a rectangle S0 = [a0 , b0 ] —
[a0 , b0 ] ⊇ K. Show that we can ¬nd a sequence of pairs of intervals [an , bn ]
and [an , bn ] such that

f (K © [an , bn ] — [an , bn ]) is not bounded,
an’1 ¤ an ¤ bn ¤ bn’1 , an’1 ¤ an ¤ bn ¤ bn’1 ,
and bn ’ an = (bn’1 ’ an’1 )/2, bn ’ an = (bn’1 ’ an’1 )/2,

for all n ≥ 1.
Show that an ’ c as n ’ ∞ for some c ∈ [a0 , b0 ] and an ’ c as n ’ ∞
for some c ∈ [a0 , b0 ]. Show that c = (c, c ) ∈ K. Use the fact that f is
continuous at c to show that there exists an > 0 such that, if x ∈ K and
x ’ c < , then f (x) < f (c) + 1. Show that there exists an N such that

[an , bn ] — [an , bn ] ⊆ B(c, )

for all n ≥ N and derive a contradiction.
Hence deduce that f (K) is bounded above. Show also that f (K) is bounded
below.
(ii) Since any non-empty bounded subset of R has a supremum, we know
that M = sup f (K) and m = inf f (K) exist. We now produce our clever
trick. Suppose, if possible, that f (x) = M for all x ∈ K. Explain why,
if we set g(x) = 1/(M ’ f (x)), g : K ’ R will be a well de¬ned strictly
positive continuous function. Deduce that there exists a real number M > 0
such that g(x) ¤ M for all x ∈ K and show that f (x) ¤ M ’ 1/M for all
x ∈ K. Explain why this contradicts the de¬nition of M and conclude that
there must exist some k1 ∈ K such that f (k1 ) = M . We obtain k2 similarly.
(The author repeats the remark he made on page 38 that amusing as proofs
59
Please send corrections however trivial to twk@dpmms.cam.ac.uk

like these are, clever proofs should only be used when routine proofs do not
work.)

Our next theorem is just a particular but useful case of Theorem 4.3.4.

Theorem 4.3.6. Let f : [a, b] ’ R be a continuous function. Then we can
¬nd k1 , k2 ∈ [a, b] such that

f (k2 ) ¤ f (x) ¤ f (k1 )

for all x ∈ [a, b].

Later we will use this result to prove Rolle™s theorem (Theorem 4.4.4)
from which in turn we shall obtain the mean value theorem (Theorem 4.4.1).
Theorem 4.3.4 can also be used to prove the fundamental theorem of
algebra which states that every complex polynomial has a root. If the reader
cannot wait to see how this is done then she can look ahead to section 5.8.

Exercise 4.3.7. (i) Prove Theorem 4.3.6 directly from the one-dimensional
version of the Bolzano-Weierstrass theorem. (Essentially just repeat the ar-
guments of Theorem 4.3.1.)
(ii) Give an example of a continuous f : (a, b) ’ R which is unbounded.
(iii) Give an example of a continuous f : (a, b) ’ R which is bounded but
does not attain its bounds.
(iv) How does your argument in (i) fail in (ii) and (iii)?
(v) Suppose now we work over Q and write [a, b] = {x ∈ Q : a ¤ x ¤ b}.
Show that f (x) = (1+(x2 ’2)2 )’1 de¬nes a continuous function f : [0, 2] ’ Q
which is continuous and bounded but does not attain its upper bound. How
does your argument in (i) fail?
De¬ne a continuous function g : [0, 2] ’ Q which is continuous and
bounded but does not attain either its upper bound or its lower bound. De¬ne
a continuous function h : [0, 2] ’ Q which is continuous but unbounded.

We conclude this section with an exercise which emphasises once again the
power of the hypotheses ˜closed and bounded™ combined with the Bolzano-
Weierstrass method. The result is important but we shall not make much
use of it.

Exercise 4.3.8. (i) By picking xj ∈ Kj and applying the Bolzano-Weierstrass
theorem, prove the following result.
Suppose that K1 , K2 , . . . are non-empty bounded closed sets in Rm such
that K1 ⊇ K2 ⊇ . . . . Then ∞ Kj = …. (That is, the intersection of a
j=1
nested sequence of bounded, closed, non-empty sets is itself non-empty.)
60 A COMPANION TO ANALYSIS

(ii) By considering Kj = [j, ∞), show that boundedness cannot be dropped
from the hypothesis.
(iii) By considering Kj = (0, j ’1 ), show that closedness cannot be dropped
from the hypothesis.
Exercises K.29 to K.36 discuss a substantial generalisation of Exercise 4.3.8
called the Heine-Borel theorem.


4.4 The mean value theorem
Traditionally one of the ¬rst uses of the theorem that every continuous func-
tion on a closed interval is bounded and attains its bounds has been to prove
a slightly stronger version of the mean value inequality.
In common with Dieudonn´ ([13], page 142) and Boas ([8], page 118), I
e
think that the mean value inequality is su¬cient for almost all needs and
that the work required to understand the subtleties in the statement and
proof of Theorem 4.4.1 far outweigh any gain.
However, Theorem 4.4.1 is likely to remain part of the standard analysis
course for many years, so I include it here.
Theorem 4.4.1. (The mean value theorem.) If f : [a, b] ’ R is a
continuous function with f di¬erentiable on (a, b), then we can ¬nd a c ∈
(a, b) such that

f (b) ’ f (a) = (b ’ a)f (c).

Here are some immediate consequences.
Lemma 4.4.2. If f : [a, b] ’ R is a continuous function with f di¬eren-
tiable on (a, b), then the following results hold.
(i) If f (t) > 0 for all t ∈ (a, b) then f is strictly increasing on [a, b].
(That is, f (y) > f (x) whenever b ≥ y > x ≥ a.)
(ii) If f (t) ≥ 0 for all t ∈ (a, b) then f is increasing on [a, b]. (That is,
f (y) ≥ f (x) whenever b ≥ y > x ≥ a.)
(iii) If f (t) = 0 for all t ∈ (a, b) then f is constant on [a, b]. (That is,
f (y) = f (x) whenever b ≥ y > x ≥ a.)
Proof. We prove part (i), leaving the remaining parts to the reader. If b ≥
y > x ≥ a, then the mean value theorem (Theorem 4.4.1) tells us that

f (y) ’ f (x) = (y ’ x)f (c)

for some c with y > c > x. By hypothesis f (c) > 0, so f (y) ’ f (x) > 0.
61
Please send corrections however trivial to twk@dpmms.cam.ac.uk

Exercise 4.4.3. Prove Theorem 1.7.1 from Theorem 4.4.1.
The key step in proving Theorem 4.4.1 is the proof of the special case
when f (a) = f (b).
Theorem 4.4.4. (Rolle™s theorem.) If g : [a, b] ’ R is a continuous
function with g di¬erentiable on (a, b) and g(a) = g(b), then we can ¬nd a
c ∈ (a, b) such that g (c) = 0.
The next exercise asks you to show that the mean value theorem follows
from Rolle™s theorem.
Exercise 4.4.5. (i) If f is as in Theorem 4.4.1, show that we can ¬nd a real
number A such that, setting

g(t) = f (t) ’ At,

the function g satis¬es the conditions of Theorem 4.4.4.
(ii) By applying Rolle™s theorem (Theorem 4.4.4) to the function g in (i),
obtain the mean value theorem (Theorem 4.4.1). (Thus the mean value the-
orem is just a tilted version of Rolle™s theorem.)
Cauchy produced an interesting variant on the argument of Exercise 4.4.5.
which we give as Exercise K.51.
Exercise 4.4.6. The following very easy consequence of De¬nition 1.7.2 will
be used in the proof of Rolle™s theorem. Let U be an open set in R and let
f : U ’ R be di¬erentiable at t ∈ U with derivative f (t). Show that if
tn ∈ U , tn = t and tn ’ t as n ’ ∞, then
f (tn ) ’ f (t)
’ f (t)
tn ’ t
as n ’ ∞.
We now turn to the proof of Rolle™s theorem.
Proof of Theorem 4.4.4. Since the function g is continuous on the closed in-
terval [a, b], Theorem 4.3.6 tells us that it is bounded and attains its bounds.
More speci¬cally, we can ¬nd k1 , k2 ∈ [a, b] such that

g(k2 ) ¤ g(x) ¤ g(k1 )

for all x ∈ [a, b]. If both k1 and k2 are end points of [a, b] (that is k1 , k2 ∈
{a, b}) then

g(a) = g(b) = g(k1 ) = g(k2 )
62 A COMPANION TO ANALYSIS

and g(x) = g(a) for all x ∈ [a, b]. Taking c = (a + b)/2, we have g (c) = 0
(the derivative of a constant function is zero) and we are done.
If at least one of k1 and k2 is not an end point there is no loss in generality
in assuming that k1 is not an end point (otherwise, consider ’g). Write
c = k1 . Since c is not an end point, a < c < b and we can ¬nd a δ > 0 such
that a < c ’ δ < c + δ < b. Set xn = c ’ δ/n. Since c is a maximum for g,
we have g(c) ≥ g(xn ) and so

g(xn ) ’ g(c)
≥0
xn ’ c
for all n. Since
g(xn ) ’ g(c)
’ g (c),
xn ’ c
it follows that g (c) ≥ 0. However, if we set yn = c + δ/n, a similar argument
shows that
g(yn ) ’ g(c)
¤0
yn ’ c

for all n and so g (c) ¤ 0. Since 0 ¤ g (c) ¤ 0, it follows that g (c) = 0 and
we are done.
(We look more closely at the structure of the preceeding proof in Exer-
cise K.45.)
In his interesting text [11], R. P. Burn writes

Both Rolle™s theorem and the mean value theorem are geomet-
rically transparent. Each claims, with slightly more generality
in the case of the mean value theorem, that for a graph of a
di¬erentiable function, there is always a tangent parallel to the
chord.

My view is that the apparent geometrical transparency is due to our strong
intuitive feeling a function with positive derivative ought to increase ” which
is precisely what we are ultimately trying to prove5 . It is because of this strug-
gle between intuition and rigour that the argument of the second paragraph
of the proof always brings to my mind someone crossing a tightrope above
5
This should not be interpreted as a criticism of Burn™s excellent book. He is writing
a ¬rst course in analysis and is trying to persuade the unwilling reader that what looks
complicated is actually simple. I am writing a second course in analysis and trying to
persuade the unwilling reader that what looks simple is actually complicated.
63
Please send corrections however trivial to twk@dpmms.cam.ac.uk

a pool full of crocodiles. Let me repeat to any reader tempted to modify
that argument, we wish to use Theorem 4.4.1 to prove that a function with
positive derivative is increasing and so we cannot use that result to prove
Theorem 4.4.1. If you believe that you have a substantially simpler proof of
Rolle™s theorem than the one given above, ¬rst check it against Exercise K.46
and then check it with a professional analyst. Exercise K.43 gives another
use of the kind of argument used to prove Rolle™s theorem.
If the reader uses Theorem 4.4.1, it is important to note that we know
nothing about c apart from the fact that c ∈ (a, b).
Exercise 4.4.7. Suppose that k2 is as in the proof of Theorem 4.4.4. Show
explicitly that, if k2 is not an end point, g (k2 ) = 0.
Exercise 4.4.8. Suppose that g : R ’ R is di¬erentiable, that a < b and
that g(a) = g(b). Suppose k1 and k2 are as in the proof of Theorem 4.4.4.
Show that, if k1 = a, then g (a) ¤ 0 and show by example that we may have
g (a) < 0. State similar results for the cases b = k1 and a = k2 .
Exercise 4.4.9. (This exercise should be compared with Lemma 4.4.2.)
(i) Suppose that f : (a, b) ’ R is di¬erentiable and increasing on (a, b).
Show that f (t) ≥ 0 for all t ∈ (a, b).
(ii) If f : R ’ R is de¬ned by f (t) = t3 , show that f is di¬erentiable and
everywhere strictly increasing yet f (0) = 0.
Exercise 4.4.10. I said above that the mean value inequality is su¬cient
for most purposes. For the sake of fairness here is an example where the
extra information provided by Rolle™s theorem does seem to make a di¬erence.
Here, as elsewhere in the exercises, we assume that reader knows notations
like F (r) for the rth derivative of F and can do things like di¬erentiating a
polynomial which have not been explicitly treated in the main text.
Suppose that f : R ’ R is n times di¬erentiable and that

<< . .

. 7
( : 70)



. . >>