Continuous Functions
Continuous Functions
1
2 CHAPTER 3. THE SPACE OF CONTINUOUS FUNCTIONS
It is readily checked that Cb (X) is a normed space under the sup-norm. From now on,
Cb (X) is always regarded as a metric space under the metric induced by the sup-norm.
In order words,
d∞ (f, g) = kf − gk∞ , ∀f, g ∈ Cb (X).
Some basic properties of Cb (X) are listed below:
Property 1. Cb (X) is a complete metric space. Indeed, let {fn } be a Cauchy sequence
in Cb (X). For ε > 0, there exists some n0 such that kfn − fm k∞ < ε/4 for all n ≥
n0 . In particular, it means for each x, {fn (x)} is a Cauchy sequence in R. By the
completeness of R, the limit limn→∞ fn (x) exists and we define f (x) ≡ limn→∞ fn (x).
Assuming that f ∈ Cb (X), by taking m → ∞ in the inequality above, we immediately
obtain kfn − f k∞ ≤ ε/4 < ε, hence fn → f in Cb (X). To show that f ∈ Cb (X), we
let m → ∞ in |fn (x) − fm (x)| < ε/4 to get |fn (x) − f (x)| ≤ ε/4 for all x and n ≥ n0 .
Taking n = n0 we get |f (x)| ≤ |f (x) − fn0 (x)| + |fn0 (x)| ≤ ε/4 + kfn0 k∞ , hence f is
bounded. On the other hand, as fn0 is continuous, for each x we can find a δ such that
|fn0 (y) − fn0 (x)| < ε/4 whenever d(y, x) < δ. It follows that for all y, d(y, x) < δ,
3ε
|f (y) − f (x)| ≤ |f (y) − fn0 (y)| + |fn0 (y) − fn0 (x)| + |fn0 (x) − f (x)| ≤ < ε.
4
From this proof we see that the completeness of Cb (X) is inherited from the completeness
of R, so the underlying space X does not play any role in this aspect.
Property 2. Cb (X) = C(X) when X is a compact metric space. We need to show every
continuous function on a compact set is bounded. Assume on the contrary that for some
continuous f , there are points {xk } such that |f (xk )| → ∞. By compactness, there is
a subsequence {xkj } and z ∈ X such that limj→∞ xkj = z. But, by continuity we would
have limj→∞ |f (xkj )| = |f (z)| < ∞, contradiction holds.
Property 3. Cb (X) forms an algebra under pointwise product. Recall that an algebra
is a vector space V in which a product satisfying the association law is well-defined
between two points. The interaction between this product and the vector space structure
is reflected in the rules w(u + v) = (u + v)w = wu + wv and (au)(bv) = ab(uv) for all
u, v, w ∈ V and a, b ∈ R. It is clear the product of two bounded, continuous functions is
again a bounded, continuous function, so Cb (X) forms an algebra.
We will investigate various properties of the spaces of continuous functions. Recall that
a consequence of Weierstrass approximation theorem tells that every continuous function
on [a, b] can be approximated by polynomials with rational coefficients. In general a set
E in a metric space is called a dense set if its closure is equal to the space. Thus the
collection of all polynomials with rational coefficients forms a dense set in C[a, b]. Since
this set is countable, we know that every continuous function in [a, b] can be approximated
3.2. STONE-WEIERSTRASS THEOREM 3
from continuous functions chosen from a countable set. Our question is, in a general space
C(X), when does this property still hold? To obtain a result in the positive direction, we
need to establish a generalization of Weierstrass approximation theorem, namely, Stone-
Weierstrass theorem. That is what we are going to do in the next section.
Proof. Assume the interval to be [0, 1] first. Let f be a continuous function on [0, 1]. By
subtracting it from a linear function (a polynomial of degree 0 or 1) which passes (0, f (0))
and (1, f (1)) we may assume f (0) = f (1) = 0. Extend it to be a continuous function in
R which equals zero outside the unit interval and denote the extended function still by f .
Now we approximate f by introducing
ˆ 1
pn (x) := f (x + t)Qn (t)dt, x ∈ [0, 1],
−1
where Qn (x) := cn (1 − x2 )n . (In fact, as f vanishes outside [0, 1], the integration is in fact
´ 1 −x to 1 − x.) Qn is a polynomial and the normalizing constant cn is chosen so that
from
Q = 1, in other words, cn is given by
−1 n
ˆ 1
−1
cn = (1 − x2 )n dx.
−1
ˆ 1 ˆ 1
2 n
(1 − x ) dx = 2 (1 − x2 )n dx
−1 0
ˆ √
1/ n
≥2 (1 − x2 )n dx
0
ˆ √
1/ n
≥2 (1 − nx2 )dx
0
4 1
= √ >√ ,
3 n n
4 CHAPTER 3. THE SPACE OF CONTINUOUS FUNCTIONS
after using Bernoulli’s inequality in the form (1 − x2 )n ≥ 1 − nx2 for all x. It follows that
√
cn < n.
so, in particular, Qn (x) → 0 uniformly on [δ, 1] for any fixed δ ∈ (0, 1).
So far we have shown that trigonometric functions and polynomials are dense in the
space of periodic, continuous functions and the space of continuous functions respec-
tively. In this section we will establish a far-reaching generalization of these results in the
space of continuous functions defined in a compact metric space. In such a space both
3.2. STONE-WEIERSTRASS THEOREM 5
trigonometric functions and polynomials are not available, so we need to seek a reason-
able formulation. The answer relies on an extra algebraic structure we have exploited
explicitly.
We have pointed out that the space Cb (X) carries an extra property, namely, it is
an algebra under pointwise product. A subspace A is called a subalgebra of Cb (X) if it
is closed under this product. It is readily checked that all polynomials on [a, b] form a
subalgebra of C[a, b], so does the subalgebra consisting of all polynomials with rational
coefficients. Similarly, the vector space consisting of all trigonometric functions and its
subspace consisting of all trigonometric functions with rational coefficients are algebras
in in the space of 2π-periodic continuous functions in C2π . Note that C2π can be identified
with C(S 1 ) where S 1 = {(cos t, sin t) ∈ R2 : t ∈ [0, 2π]} is the unit circle.
Before proceeding further, recall that aside from the constant ones, there are many
continuous functions in a metric space. For instance, given any two distinct points x1
and x2 in X, it is possible to find some f ∈ C(X) such that f (x1 ) 6= f (x2 ). We simply
take f (x) = d(x, x1 ) and f (x1 ) = 0 < f (x2 ). It is even possible to find one in Cb (X),
e.g., g(x) = f (x)/(1 + f (x)) serves this purpose. We now consider what conditions a
subalgebra must possess in order that it becomes dense in Cb (X). A subalgebra is called
to satisfy the separating points property if for any two points x1 and x2 in X, there
exists some f ∈ A satisfying f (x1 ) 6= f (x2 ). From the discussion above, it is clear that
A must satisfy the separating points property if its closure is Cb (X). Thus the separat-
ing points property is a necessary condition
Pn for a subalgebra to be dense. On the other
2j
hand, the polynomials of the form j=0 aj x form an algebra which does not have the
separating point property, for it is clear that p(−x) = p(x) for such p. Another condition
is that, whenever x ∈ X, there must be some g ∈ A such that g(x) 6= 0. We will call this
the non-vanishing property. The non-vanishing
Pn property fails to hold for the algebra
j
consisting of all polynomials of the form j=1 aj x , for p(0) = 0 for all these p. The non-
vanishing property is also a necessary condition for an algebra to be dense in C(X). For,
if for some particular z ∈ X, f (z) = 0 holds for all f ∈ A, it is impossible to approximate
the constant function 1 in the supnorm by functions in A. Surprisingly, it turns out these
two conditions are also sufficient when the underlying space is compact, and this is the
content of the following theorem. This is for an optional reading.
Recall that Cb (X) = C(X) when X is compact. For the proof of this theorem two
lemmas are needed.
Lemma 3.4. Let A be a subalgebra in C(X) which separates points and non-vanishing
at all points. For x1 , x2 ∈ X and α, β ∈ R, there exists ϕ ∈ A such that ϕ(x1 ) = α and
ϕ(x2 ) = β.
Proof. Since A separates points, we can find some ψ ∈ A such that ψ(x1 ) 6= ψ(x2 ). We
claim that one can further choose ψ such that ψ(x1 ), ψ(x2 ) are both non-zero. For, if,
for instance, ψ(x1 ) = 0, fix some ξ ∈ A satisfying ξ(x1 ) 6= 0. This is possible due to the
non-vanishing property. Consider a function ψ1 ∈ A of the form ψ + tξ. We would like
to find t ∈ R such that (a) ψ1 (x1 ) 6= ψ1 (x2 ), (b) ψ1 (x1 ) 6= 0, and (c) ψ1 (x2 ) 6= 0. There
are two cases; when ξ(x2 ) 6= 0, it suffices to choose t such that t 6= 0, −ψ(x2 )/ξ(x2 ) (if
ξ(x2 ) 6= 0). When ξ(x2 ) = 0, we choose t such that t 6= ψ(x2 )/ξ(x1 ). Replacing ψ by ψ1 ,
we obtain our desired function which satisfies (a)–(c).
Now, we can find a and b such that the combination ϕ = aψ + bψ 2 ∈ A satisfies the
requirement in the lemma. Indeed, what we need are the conditions aψ(x1 ) + bψ 2 (x1 ) = α
and aψ(x2 ) + bψ 2 (x2 ) = β. As the determinant of this linear system (viewing a and b as
the unknowns) is equal to ψ(x1 )ψ(x2 )(ψ(x1 ) − ψ(x2 )) which is not equal to 0, a and b can
always be found.
Proof of Theorem 3.5. It remains to establish the necessary part of the theorem.
Let f ∈ C(X) be given. For each pair of x, y, there exists a function ϕx,y ∈ A satisfying
ϕx,y (x) = f (x) and ϕx,y (y) = f (y). This is due to the previous lemma when x and y are
distinct. When x is equal to y, such function still exists. Now, for each ε > 0, there exists
an open set Ux,y containing x and y such that
For fixed y, the sets {Ux,y : x ∈ X} form an open cover of X. By the compactness of X, it
admits a finite subcover {Uxj ,y }N
j=1 . The function ϕy = ϕx1 ,y ∨ · · · ∨ ϕxN ,y belongs to A ac-
cording to Lemma 3.3. Furthermore, ϕy > f − ε in X. For, let x ∈ X, there is some Uxj ,y
containing x. Therefore, ϕy (x) ≥ ϕxj ,y (x) > f (x) − ε. Next, Gy ≡ ∩N j=1 Uxj ,y is an open set
containing y and all these open sets together form an open cover of X when y runs over
3.3. SEPARABILTY 7
3.3 Separabilty
Recall that a set E in a metric space (X, d) is dense if for every x ∈ X and ε > 0, there
exists some y ∈ E such that d(y, x) < ε. Equivalently, E is dense if E = X. The space
X is called a separable space if it admits a countable dense subset. Equivalently, X
is separable if there is a countable subset E satisfying E = X. A set is separable if it
is separable as a metric subspace. When a metric space is separable, every element can
be approximated by elements from a countable set. Hence its structure is easier to study
than the non-separable ones. Here are two basic properties of separable spaces.
Proof. Let Y be a subset of the separable space (X, d) and D = {xj } a countable dense
subset of X. For each n, pick a point zjn from Y ∩ B1/n (xj ) if it is non-empty to form
the countable set E = {zjn }. We claim that E is dense in Y . For, let y ∈ Y and
each ε > 0, there is some xj ∈ D such that d(y, xj ) < ε/2. Therefore, for n > 2/ε,
B1/n (xj ) ∩ Y is nonempty and we can find some yjn ∈ B1/n (xj ) ∩ Y . It follows that
d(y, yjn ) ≤ d(y, xj ) + d(xj , yjn ) < ε/2 + 1/n < ε.
Proof. Every compact space is totally bounded. By Proposition 2.11, for each n, there
exist finitely many points x1 , · · · , xN such that the balls B1/n (xj ), j = 1, · · · N, form an
open cover of the space. It is clear that the countable set consisting of all centers of these
balls when n runs from 1 to infinity forms a dense set of the space.
Example 3.1. Consider the Euclidean space Rn . As it is well-known that the set of all
rational numbers Q forms a countable dense subset of R, R is a separable space. Simi-
larly, Rn is separable for all n ≥ 1 because it contains the dense subset Qn . According to
Proposition 3.5, all sets in the Euclidean space are separable.
Example 3.2. C[a, b] is a separable space. Although the proof was assigned as an exercise,
we repeat it here. Without loss of generality we take [a, b] = [0, 1]. Denote by P the
8 CHAPTER 3. THE SPACE OF CONTINUOUS FUNCTIONS
It is clear that S is a countable set. Given any polynomial p(x) = a0 +a1 x+· · ·+an xn , aj ∈
R, j = 1, · · · , n. For every ε > 0, weP
can choose some bj ∈ Q such that |aj −bj | < ε/(n+1)
for all j. It follows that for q(x) = j bj xj ∈ S, we have
X
|p(x) − q(x)| ≤ |a0 − b0 | + |a1 − b1 ||x| + · · · + |an − bn ||x|n
j
ε
< (n + 1)
2(n + 1)
ε
=
2
for all x. We conclude that kp − qk∞ ≤ ε/2 Now, for any f ∈ C[0, 1] and ε > 0, we apply
Weierstrass approximation theorem to obtain a polynomial p such that kf − pk∞ < ε/2
and then find some q ∈ S such that kp − qk∞ ≤ ε/2. It follows that
ε ε
kf − qk∞ ≤ kf − pk∞ + kp − qk∞ < + = ε,
2 2
that is, S is dense in C[0, 1].
Example 3.3. Let K be a closed, bounded set in Rn . Then C(K) is separable. This
is a generalization of the previous example. The proof is again aided by considering the
collection of all polynomials P in n-variables with rational coefficients. It is clear that
this countable collection forms an algebra in C(K). For the separating points property
we
Pn observe that for every two distinct points x0 and y 0 in Rn the polynomial p(x) =
0 2 0 0
k=1 (xk − xk ) satisfies p(x ) = 0 and p(y ) 6= 0. It is also nonvanishing because constant
functions are polynomials. Henceforth, it is dense in C(K) by Stone-Weierstrass theorem.
Proposition 3.7. The space C(X) is separable when X is a compact metric space.
To conclude this section, we note the existence of non-separable spaces. Here is one.
Example 3.4. Consider the space of all bounded functions on [a, b] under the supnorm.
It forms a metric space denoted by B[a, b]. We claim that it is not separable. For, let
fz ∈ B[a, b] be given by fz (x) = 0 for all x 6= z and fz (z) = 1. All fz ’s form an uncount-
able set. Obviously the metric balls B1/2 (fz ) are pairwise disjoint. If S is a dense subset
of B[a, b], S ∩ B1/2 (fz ) must be non-empty for each z. We pick wz ∈ S ∩ B1/2 (fz ) to form
an uncountable subset {wz } of S. We concldue that S must be uncountable, so there is
no countable dense set of B[a, b].
We pointed out that not every closed, bounded set in a metric space is compact. In Section
2.3 a bounded sequence without any convergent subsequence is explicitly displayed to
show that a closed, bounded set in C[a, b] needs not be compact. In view of numerous
theoretic and practical applications, it is strongly desirable to give a characterization of
compact sets in C[a, b]. The answer is given by the fundamental Arzela-Ascoli theorem.
This theorem gives a necessary and sufficient condition when a closed and bounded set in
C[a, b] is compact. In order to have wider applications, we will work on a more general
space C(K), where K is a closed, bounded subset of Rn , instead of C[a, b]. Recall that
C(K) is a complete, separable space under the sup-norm.
The crux for compactness for continuous functions lies on the notion of equicontinuity.
Let X be a subset of Rn . A subset F of C(X) is equicontinuous if for every ε > 0,
there exists some δ such that
|f (x) − f (y)| < ε, for all f ∈ F, and |x − y| < δ, x, y ∈ X.
Recall that a function is uniformly continuous in X if for each ε > 0, there exists some
δ such that |f (x) − f (y)| < ε whenever |x − y| < δ, x, y ∈ X. So, equicontinuity means
10 CHAPTER 3. THE SPACE OF CONTINUOUS FUNCTIONS
There are various ways to show that a family of functions is equicontinuous. Recall
that a function f defined in a subset X of Rn is called Hölder continuous if there exists
some α ∈ (0, 1) such that
for some constant L. The number α is called the Hölder exponent. The function is called
Lipschitz continuous if (3.1) holds for α equals to 1. A family of functions F in C(X)
is said to satisfy a uniform Hölder or Lipschitz condition if all members in F are Hölder
continuous with the same α and L or Lipschitz continuous and (3.1) holds for the same
constant L. Clearly, such F is equicontinuous. The following situation is commonly
encountered in the study of differential equations. The philosophy is that equicontinu-
ity can be obtained if there is a good, uniform control on the derivatives of functions in F.
Proposition 3.8. Let F be a subset of C(X) where X is a convex set in Rn . Suppose that
each function in F is differentiable and there is a uniform bound on the partial derivatives
of these functions in F. Then F is equicontinuous.
Proof. For, x and y in X, (1 − t)x + ty, t ∈ [0, 1], belongs to X by convexity. Let
ψ(t) ≡ f ((1 − t)x + ty). From the mean-value theorem
we have n
X ∂f
f (y) − f (x) = ψ(1) − ψ(0) = ((1 − t∗ )x + t∗ y)(yj − xj ).
j=1
∂x j
Therefore, √
|f (y) − f (x)| ≤ nM |y − x|,
where M = sup{|∂f /∂xj (x)| : x ∈ X, j = 1, . . . , n, f ∈ F} after using Cauchy-
Schwarz inequality. So F satisfies a uniform Lipschitz condition with Lipschitz constant
n1/2 M .
Theorem 3.9 (Arzela-Ascoli). Let F be a closed set in C(K) where K is a closed and
bounded set in Rn . Then F is compact if and only if it is bounded and equicontinuous.
3.4. COMPACTNESS AND ASCOLI-ARZELA THEOREM 11
A set E ⊂ C(K) is bounded means it is contained in a ball, or, more specifically, there
exists M > 0 such that
Proof. Since {fn (z1 )} is a bounded sequence, we can extract a subsequence {fn1 } such that
{fn1 (z1 )} is convergent. Next, as {fn1 (z2 )} is bounded, it has a subsequence {fn2 } such that
{fn2 (z2 )} is convergent. Keep doing in this way, we obtain sequences {fnj } satisfying (i)
{fnj+1 } is a subsequence of {fnj } and (ii) {fnj (z1 )}, {fnj (z2 )}, · · · , {fnj (zj )} are convergent.
Then the diagonal sequence {gn }, gn = fnn , for all n ≥ 1, is a subsequence of {fn } which
converges at every zj .
The subsequence selected in this way is sometimes called Cantor’s diagonal sequence.
|gn (x) − gm (x)| ≤ |gn (x) − gn (xjk0 )| + |gn (xjk0 ) − gm (xjk0 )| + |gm (xjk0 ) − gm (x)|
< ε + |gn (xjk0 ) − gm (xjk0 )| + ε.
Here n0 depends on xjk0 . As there are finitely many xjk0 ’s, we can choose some N0 such
that (3.2) holds for all xjk0 and n, m ≥ N0 . It follows that
i.e., {gn } is a Cauchy sequence in C(K). By the completeness of C(K) and the closedness
of F, {gn } converges to some function in F.
Conversely, let F be compact.
SN By Proposition 2.11, for each ε > 0, there exist
f1 , · · · , fN ∈ F such that F ⊂ j=1 Bε (fj ) where N depends on ε. So for any f ∈ F,
there exists fj such that
As each fj is continuous, there exists δj such that |fj (x)−fj (y)| < ε whenever |x−y| < δj .
Letting δ = min{δ1 , · · · , δN }, then
|f (x) − f (y)| ≤ |f (x) − fj (x)| + |fj (x) − fj (y)| + |fj (y) − f (y)| < 3ε,
The following special case of Arzela-Ascoli theorem, sometimes called Ascoli’s theorem,
is most useful in applications.
Proof. Let F be the closure of the sequence {fn }. We would like to show that F is
bounded and equicontinuous. First of all, by the uniform boundedness assumption, there
is some M such that
|fn (x)| ≤ M, ∀x ∈ K, n ≥ 1.
As every function in F is either one of these fn or the limit of its subsequence, it also
satisfies this estimate, so F is bounded in C(K). On the other hand, by equicontinuity,
for every ε > 0, there exists some δ such that
ε
|fn (x) − fn (y)| < , ∀x, y ∈ K, |x − y| < δ.
2
As every f ∈ F is the limit of a subsequence of {fn }, f satisfies
ε
|f (x) − f (y)| ≤ < ε, ∀x, y ∈ K, |x − y| < δ,
2
so F is also equicontinuous. Now the conclusion follows from the Arzela-Ascoli theorem.
3.4. COMPACTNESS AND ASCOLI-ARZELA THEOREM 13
From the proof we will see that a0 can be taken to be 0 < a0 < min{a, b/M } where
M = sup{|f (x, y)| : (x, y) ∈ R}. The theorem is also valid for systems.
Proof. Recalling in the proof of Picard-Lindelöf theorem we showed that under the Lips-
chitz condition the unique solution exists on the interval [x0 − a0 , x0 + a0 ] where 0 < a0 <
min{a, b/M, 1/L∗ } where L∗ is the Lipschitz constant. Let us first argue that the maxi-
mal solution in fact exists in the interval [x0 − a0 , x0 + a0 ] where 0 < a0 < min{a, b/M }.
In other words, the Lipschitz condition does not play any role in the range of existence.
Although this was done in the exercise, we include it here for the sake of completeness.
Take x0 = y0 = 0 to simplify notations. The functions w(x) = M x and z(x) = −M x
satisfy y 0 = ±M, y(0) = 0, respectively. By comparing them with y, our maximal
solution to (2.3), we have z(x) ≤ y(x) ≤ w(x) as long as y exists. In case y exists on
[0, α) for some α < min{a, b/M }, (x, y(x)) would be confined in the triangle bounded by
y = M x, y = −M x, and x = α. As this triangle is compactly contained in the interior
of R, the Lipschitz constant ensures that the solution could be extended beyond α. Thus
the solution exists up to min{a, b/M }. Similarly, one can show that the solution exists in
(− min{a, b/M }, 0].
With this improvement at our disposal, we prove the theorem. First, of all, by Weier-
strass approximation theorem, there exists a sequence of polynomials {fn } approach-
ing f in C([−a, a] × [−b, b]) uniformly. In particular, it means that Mn → M, where
Mn = max{|fn (x, y)| : (x, y) ∈ [−a, a] × [−b, b]. As each fn satisfies the Lipschitz condi-
tion (why?), there is a unique solution yn defined on In = (−an , an ), an = min{a, b/Mn }
for the initial value problem
dy
= fn (x, y), y(0) = 0.
dx
From |dyn /dx| ≤ Mn and limn→∞ Mn = M , we know from Proposition 3.8 that {yn }
forms an equicontinuous family. Clearly, it is also bounded. By Ascoli’s theorem, it
14 CHAPTER 3. THE SPACE OF CONTINUOUS FUNCTIONS
As {ynj } → y uniformly, {f (x, ynj (x))} also tends to f (x, y(x)) uniformly. By passing to
limit in the formula above, we conclude that
ˆ x
y(x) = f (t, y(t))dt, x ∈ I
0
In this section we discuss Baire category theorem, a basic property of complete metric
spaces. It is concerned with the decomposition of a metric space into a countable union
of subsets. The motivation is somehow a bit strangeSat first glance. For instance, we can
decompose the plane R2 as the union of strips R2 = k∈Z Sk where Sk = (k, k + 1] × R. In
this decomposition each Sk is not so sharply different from R2 . Aside from the boundary,
the interior of each Sk is just like the interior of R2 . On the other hand, one can make
the more extreme decomposition: R2 = α∈R lα where lα = {α} × R. Each lα is a
S
vertical straight line and is very different from R2 . It is simpler in the sense that it is
one-dimensional and has no area. The sacrifice is now we need an uncountable union.
The question is: Can we represent R2 as a countable union of these sets (or sets with
lower dimension)? It turns out that the answer is no. The obstruction comes from the
completeness of the ambient space.
We need one definition. Let (X, d) be a metric space. A subset E of X is called
nowhere dense if its closure does not contain any metric ball. Equivalently, E is nowhere
dense if X \ E is dense in X. Note that a set is nowhere dense if and only if its closure is
nowhere dense. Also every subset of a nowhere dense set is nowhere dense.
∞
Theorem 3.13 (Baire Category Theorem). Let {E j }1 be a sequence of nowhere
S∞
dense subsets of (X, d) where (X, d) is complete. Then j=1 Ej has empty interior.
A set with empty interior means that it does not contain any ball. It is so if and only
if its complement is a dense set.
3.5. COMPLETENESS AND BAIRE CATEGORY THEOREM 15
Proof. Replacing Ej by its closure if necessary, we may assume all Ej ’s are T closed
S sets. Let
B0 be any ball. The theorem will be established if we can show that B0 (X \ j Ej ) 6= φ.
As E1 is nowhere dense, there exists some point x ∈ B0 lying outside E1 . Since E1 is
closed , we can find a closed ball B 1 ⊂ B0 centering at x such that B 1 ∩ E1 = φ and its
diameter d1 ≤ d0 /2, where d0 is the diameter of B0 . Next, as E2 is nowhere dense and
closed, by the same reason there is a closed ball B 2 ⊂ B1 such that B 2 ∩ E2 = φ and
d2 ≤ d1 /2. Repeating this process, we obtain a sequence of closed balls B j satisfying (a)
B j+1 ⊂ Bj , (b) dj ≤ d0 /2j , and (c) B j is disjoint from E1 , · · · , Ej . Pick xj from B j to form
a sequence {xj }. As the diameters of the balls tend to zero, {xj } is a Cauchy sequence.
By the completeness of X, {xj } converges to some x∗ . Clearly x∗ belongs to all B j . If x∗
belongs to j Ej , x∗ belongs to some Ej1 , but then x∗ ∈ B j1 Ej1 T
S T
which S means that B j1
is not disjoint from Ej1 , contradiction holds. We conclude that B0 (X \ j Ej ) 6= φ.
Proof. Let E be of first category and let E = ∪∞ k=1 Ek where Ek are nowhere dense sets.
If it is also of second category, its complement is of first category. Thus, X \ E = ∪∞
k=1 Fk
where Fk are nowhere dense. It follows that X = E ∪ (X \ E) = ∪k (Ek ∪ Fk ) so the entire
space is a countable union on of nowhere dense sets, contradicting the completeness of
the space and the Baire category theorem.
Here we use an implicit argument to show there are far more such functions than contin-
uously differentiable functions.
We begin with a lemma.
We have
|f (y) − f (x)| ≤ L|y − x|, ∀y, |y − x| < δ0 ,
where L = |f 0 (x)| + 1. For y lying outside (x − δ0 , x + δ0 ), |y − x| ≥ δ0 . Hence
Proposition 3.16. The set of all continuous, nowhere differentiable functions forms a
set of second category in C[a, b] and hence dense in C[a, b].
3.5. COMPLETENESS AND BAIRE CATEGORY THEOREM 17
We claim that SL is a closed set. For, let {fn } be a sequence in SL which is Lipschitz
continuous at xn and converges uniformly to f . By passing to a subsequence if necessary,
we may assume {xn } to some x∗ in [a, b]. We have, by letting n → ∞,
Though elegant, a drawback of this proof is that one cannot assert which particular
function is nowhere differentiable. On the other hand, the example of Weierstrass is a
concrete one.
Separability is established by various approximation theorems. For the space C[a, b],
Weierstrass approximation theorem is applied. Weierstrass (1885) proved his approxima-
tion theorem based on the heat kernel, that is, replacing the kernel Qn in our proof in
Section 1 by the heat kernel. The argument is a bit more complicated but essentially the
same. It is taken from Rudin, Principles of Mathematical Analysis. A proof by Fourier
series is already presented in Chapter 1. Another standard proof is due to Bernstein,
which is constructive and had initiated a branch of analysis called approximation theory.
The Stone-Weierstrass theorem is due to M.H. Stone (1937, 1948). We use it to establish
the separability of the space C(X) where X is a compact metric space. You can find more
approximation theorem by web-surfing.
Arzela-Ascoli theorem plays the role in the space of continuous functions the same as
Bolzano-Weierstrass theorem does in the Euclidean space. A bounded sequence of real
numbers always admits a convergent subsequence. Although this is no longer true for
bounded sequences of continuous functions on [a, b], it does hold when the sequence is
also equicontinuous. Ascoli’s theorem (Proposition 3.11) is widely applied in the theory of
partial differential equations, the calculus of variations, complex analysis and differential
geometry. Here is a taste of how it works for a minimization problem. Consider
where ˆ 1
0
J[u] = u 2 (x) − cos u(x) dx.
0
First of all, we observe that J[u] ≥ −1. This is clear, for the cosine function is always
bounded by ±1. After knowing that this problem is bounded from −1, we see that inf J[u]
must be a finite number, say, γ. Next we pick a minimizing sequence {un }, that is, every
un is in C 1 [0, 1] and satisfies u(0) = 0, u(1) = 5, such that J[un ] → γ as n → ∞. By
Cauchy-Schwarz inequality, we have
ˆ y
0
un (x) − un (y) ≤ un (x)dx
sx ˆ sˆ
y y
≤ 12 dx u0n2 (x)dx
x x
sˆ s
ˆ
y 1
≤ 12 dx u0n2 (x)dx
x 0
p p
≤ J[un ] + 1 |y − x|
p
≤ γ + 2 |y − x|1/2
for all large n. From this estimate we immediately see that {un } is equicontinuous and
bounded (because un (0) = 0). By Ascoli’s theorem, it has a subsequence {unj } converging
to some u ∈ C[0, 1]. Apparently, u(0) = 0, u(1) = 5. Using knowledge from functional
analysis, one can further argue that u ∈ C 1 [0, 1] and is the minimum of this problem.
3.5. COMPLETENESS AND BAIRE CATEGORY THEOREM 19
Arzela showed the necessity of equicontinuity and boundedness for compactness while
Ascoli established the compactness under equicontinuity and boundedness. Google under
Arzela-Ascoli theorem for details.
There are some fundamental results that require completeness. The contraction map-
ping principle is one and Baire category theorem is another. The latter was first introduced
by Baire in his 1899 doctoral thesis. It has wide, and very often amazing applications in
all branches of analysis. Some nice applications are available on the web. Google under
applications of Baire category theorem for more.
Weierstrass’ example is discussed in some detailed in Hewitt-Stromberg, “Abstract
Analysis”. An simpler example can be found in Rudin’s Principles.
Being unable to locate a single reference containing these three topics, I decide not to
name any reference but let you search through the internet.