SPACES OF CONTINUOUS FUNCTIONS

If X is a set and F a normed vector space, the set BF (X) of bounded

functions from X to F , with the supremum norm, is also a normed vector
space. (And a Banach space, ifF is a Banach space.)
For X a metric space, denote by CFb (X) ⊂ BF (X) the set of continuous
bounded functions from X to F . With the sup norm that’s also a normed
vector space. (And a Banach space, if F is a Banach space.)
Proposition 1. CFb (X) is a closed subspace of BF (X); that is, the uniform
limit of continuous bounded functions is continuous.
Proof. Suppose fn → f uniformly in X. Let x0 ∈ X. Given  > 0,
choose N large enough so that supX |fN − f | ≤ , then a neighborhood V
of x0 so that |fN (x) − fN (x0 )| ≤  for all x ∈ V . Now note that, for each
x∈V:

|f (x) − f (x0 )| ≤ |fN (x) − f (x)| + |fN (x) − fN (x0 )| + |fN (x0 ) − f (x0 )| ≤ 3.

It is easy to give examples of sequences of continuous functions converg-

ing non-uniformly to a continuous function. However, this can’t happen for
monotone sequences on compact spaces:
Dini’s theorem. Let X be a compact metric space. If an increasing (or
decreasing) sequence (fn ) of continuous real-valued functions on X converges
to a continuous function f , then the convergence is uniform.
Proof. Given  > 0 and x ∈ X we may find an integer n(x) so that
f − fn(x) ≤ . By continuity, we may find a neighborhood V (x) of x in X so
that |f (x) − f (y)| ≤  and |f )n(x)(x) − fn(x) (y)| ≤ , for all y ∈ V (x). Then
for each y ∈ V (x) we have f (y) − fn(x) (y) ≤ 3. Take a finite sub cover
of {V (x)}x∈X and the maximum N of the n(xi ). Then for each n ≥ N we
have f (y) − fN (y) ≤ f (y) − fn(xi ) (y) ≤ 3, if y ∈ V (xi ). Since the V (xi )
cover X, this ends the proof.
Question. What goes wrong in this proof if the sequence is not mono-
tone?

Stone-Weierstrass approximation theorem.

Let A be a vector space over R. A is an algebra (or: commutative
algebra with unit) if there exists a ‘multiplication operation’ A × A → A,
(f, g) 7→ f ∗g which is bilinear (linear in f and g), commutative (f ∗g = g∗f )
and there is an element 1 ∈ A (the ‘unit’) satisfying f ∗ 1 = 1 ∗ f = f for

1
all f ∈ A. A subalgebra of A is a vector subspace of A which is closed under
the multiplication operation and contains the unit.
For any metric space X, the space CR b (X) of continuous, bounded real-

valued functions on X is an algebra over the field of real numbers (with the
operation of pointwise multiplication), satisfying, for the uniform norm:
||f g|| ≤ ||f ||||g||.
The unit is the constant function 1.
Exercise 1. Use this to show that the closure Ā of any subalgebra of
CRb (X) is also a subalgebra. Hint: recall Ā is the set of functions in C b (X)
R
which are uniform limits of functions in A. Given f = lim fn , g = lim gn ,
the main point is checking that f g = lim(fn gn ). Estimate |f g − fn gn |(x) in
the natural way.
Definition. We say a sub algebra of CR b (X) separates points if ∀x 6= y in

X we may find f ∈ A with f (x) 6= f (y).

Example. The polynomial functions in one variable (with real coeffi-
cients) form a subalgebra of CRb ([0, 1]). The polynomial functions in n real

variables form a subalgebra of CR b ([0, 1]n ). Both of these separate points.

The polynomials in one variable made up of even-degree monomials also

form a subalgebra of CRb ([−1, 1]), which doesn’t separate points (any such

polynomial takes the same value at 1 and −1).

Interpolation property. Assume A ⊂ CR b (X) is a subalgebra that sepa-

rates points. For all x 6= y in X and all real numbers a, b, there exists f ∈ A
so that f (x) = a, f (y) = b.
By assumption we know there exists g ∈ A so that g(x) 6= g(y). Set:
g − g(x)
f = a + (b − a) .
g(y) − g(x)
(Note that adding a constant to an element of A yields another element of
A, since the unit (the constant function 1) is in A.)
Stone-Weierstrass theorem. Let X be a compact metric space, A ⊂
CR (X) a subalgebra containing the constants and separating points. Then
A is dense in the Banach space CR (X).
Main Lemma. The pointwise max and the pointwise min of finitely many
functions in Ā is still in Ā.
We first give the proof of the theorem assuming the main lemma, then
prove the lemma. There are two steps:

2
 Step 1. Given f ∈ CR (X), x ∈ X and  > 0, we find gx ∈ Ā so that
g(x) = f (x) and gx (y) ≤ f (y) + , for all y ∈ X.
Step 2. Using compactness, argue there are finitely many points x1 , . . . , xN ∈
X so that ϕ(x) = max{gx1 (x), . . . , gxN (x)} (which is in Ā, by the main
lemma) satisfies:
f (y) −  ≤ ϕ(y) ≤ f (y) + ,
for all y ∈ X. Thus for any  > 0 we may find gϕ ∈ Ā so that ||f − ϕ|| ≤ 
(uniform norm). So f ∈ Ā.

Step 1. For each f ∈ CR (X), each x ∈ X and any  > 0, there exists a
function g ∈ Ā so that g(x) = f (x) and g(y) ≤ f (y) + ∀y ∈ X.
Proof. Given z ∈ X with z 6= x, let hz ∈ Ā satisfy hz (x) = f (x) and
hz (z) = f (z) + /2 (from the interpolation property.) By continuity, there
is an open neighborhood Vz of z in X so that, for each y ∈ Vz we have
hz (y) ≤ f (y) + . These define an open cover {Vz }z∈X of X. Taking a
finite subcover {Vzi }Ni=1 of X, we find (from the Main Lemma) the function
g = min{hzi |i = 1, . . . , N } is in Ā and satisfies the conditions required.
Proof of Step 2.
Let f ∈ CR (X) be arbitrary Given  > 0 and x ∈ X, let gx ∈ Ā be
the function from Step 1. By continuity there is a neighborhood U (x) of x
in X so that gx (y) ≥ f (y) −  for y ∈ U (x). Cover X by a finite number
of neighborhoods U (xi ), i = 1, . . . , N . Then (from the Main Lemma) the
function ϕ = max(gxi ) is in Ā and satisfies f (y) −  ≤ ϕ(y) ≤ f (y) + .

Proof of the Main Lemma.

√ Step 1. There exists a sequence (un ) of real polynomials approximating
t uniformly in [0, 1].
Define un by recurrence, letting u1 = 0 and setting:
1
un+1 (t) = un (t) + (1 − un (t)2 ).
2
√
We show by induction that un+1 ≥ un and un (t) ≤ t in [0, 1]. It follows
from the recursion relation that the first fact follows from the second. On
the other hand,
√ √ 1 √ 1 √
t − un (t) − (t − u2n (t)) = ( t − un (t))(1 − ( t + un (t))
t − un+1 (t) =
2 2
√
and from un (t) ≤ t it follows that √the second factor is positive. Thus we
have pointwise convergence of un to t (from the recurrence relation), and
then uniform convergence follows from Dini’s theorem.

3
 Step 2. If f ∈ Ā, then |f | ∈ Ā, the closure of A in CR (X).
Let a = ||f || (sup norm). The function f 2 /a2 is in Ā (since Ā is an
algebra) and takes values in [0, 1]. If un (t) are the functions from Step 1,
the compositions un ◦ (f 2 /a2 ) are in Ā (again since A is an algebra), and
converge uniformly to |f |/a.
Step 3. If f, g ∈ A, then max{f, g}, min{f, g} are also in Ā.

1 1
max{f, g} = (f + g + |f − g|), min{f, g} = (f + g − |f − g|).
2 2
Step 4 (final). The pointwise max and the pointwise min of finitely many
functions in Ā is still in Ā: follows from Step 3, since minimizing over a finite
set amounts to a finite number of pair comparisons.

Remark. The theorem is false for subalgebras of CC (X) (complex-valued

functions). This follows from the classical result in Complex Analysis:
Let fn : D → C be a sequence of complex analytic functions in a domain
D ⊂ C. Suppose fn → f uniformly on compact subsets of D. Then f : D →
C is analytic in D.

Corollary 1. (Polynomials in Rn .) Any real-valued continuous function

on a compact subset of Rn is the uniform limit of a sequence of polynomials.
Corollary 2. (Separability.) If X is a compact metric space, the space
CR (X) is separable.
Let (Un ) be a countable basis for the topology of X, and let gn (x) =
d(x; X \ Un ). The monomials g1m1 . . . grmr (with the mj integers) form a
countable set (hn ) of continuous functions on X, and the vector space they
span is the algebra A generated by the gn . So it suffices to use the Stone-
Weierstrass theorem to conclude A is dense in CR (X).
The family {gn } separates points: if x 6= y, we may find an Un so that
x ∈ Un , y 6∈ Un , and thus gn (x) 6= 0, gn (y) = 0.

Alternative approach: approximation by Bernstein polynomi-

als.
Cultural Remark 2: Neural Networks.

4
 Remark. A metric space X is precompact if for all  > 0 a finite number
of balls with diameter  suffices to cover X. The following three conditions
are equivalent for a metric space: (i) X is compact; (ii) X is sequentially
compact (any infinite sequence has an accumulation point.) (iii) X is pre-
compact and complete.
A precompact metric space X is separable. Indeed for each n there is a
finite set An so that, for each x ∈ X, d(x, An ) ≤ 1/n. Then A = ∪n An is
countable, and it is easy to see that Ā = X.

Equicontinuous sets. Let X be a metric space, E a normed vector

space. A family (subset) H ⊂ BE (X) of the set of bounded functions f :
X → E is equicontinuous at x0 ∈ X if (∀ > 0)(∃δ > 0)(∀f ∈ H)(d(x, x0 ) ≤
δ ⇒ |f (x) − f (x0 )| ≤ ).
H is equicontinuous if it is equicontinuous at every point of X. For
example, families of functions satisfying a uniform Lipschitz (or Hölder)
condition on X are equicontinuous.
1. If (fn ) ⊂ BE (X) converges pointwise in X to a function g ∈ BE (X),
and is equicontinuous at x0 ∈ X, then g is continuous at x0 .
2. If E is a Banach space, (fn ) ⊂ CEb (X) an equicontinuous sequence,
and fn (x) → g(x) pointwise for x in a dense subset D ⊂ X, then fn → g
pointwise on X. (Since E is complete, it suffices to show (fn ) is Cauchy on
X.)
3. If X is compact metric, (fn ) ⊂ CE (X) is equicontinuous and fn → g
pointwise in X, then fn → g uniformly in X.

Ascoli’s theorem. Let X be a compact metric space, E a Banach space.

For a family H ⊂ CE (X) to be relatively compact, the following conditions
are necessary and sufficient: (i) H is equicontinuous; (ii) for each x ∈ X, the
set H(x) = {f (x); f ∈ H} is relatively compact in E. (That is, its closure
is compact; if E is finite-dimensional, this just means H(x) is bounded, for
each x ∈ X. )
Remark. A relatively compact set is precompact; the converse is true in
a complete metric space.

Proof of Ascoli’s theorem.

(i) Necessity. If H is relatively compact, for each  > 0 there exist a finite
number of functions fi ∈ H so that for each f ∈ H we have ||f −fi || ≤ || (sup
norm), for some i. Thus for each x ∈ X we have |f (x)−fi (x)| ≤ , so the set
H(x) ⊂ E is precompact, and (since E is complete) also relatively compact.

5
Let x ∈ X and V (x) be a neighborhood of x so that |fi (x) − fi (y)| for each
y ∈ V (x), and all i. If follows that if y ∈ V (x) we have |f (x) − f (y)| < 3
for each f ∈ H, so H is equicontinuous.

(ii) Sufficiency. Since CE (X) is complete, it suffices to show H is pre-

compact.
Given  > 0, for each x ∈ X let V (x) be a neighborhood so that |f (x) −
f (y)| < , for each f ∈ H and y ∈ V (x). Take a finite sub covering V (xi ), i =
1, . . . , n. The set J = ∪ni=1 {f (xi ); f ∈ H} is (by hypothesis) relatively
compact, hence precompact; so there is a finite subset K = {c1 , . . . , cm } of
J so that each element of J is -close to some element of K.
We have to exhibit a covering of H by finitely many subsets of diameter
at most . For each function ϕ : {1, . . . , n} → {1, . . . , m} let Lϕ be the set
of all f ∈ H such that |f (xi ) − cϕ(i) | < . Some sets Lϕ may be empty, but
they certainly cover H, and there are only finitely many of them. We have
to check the Lϕ have small diameter.
Let f, g ∈ Lϕ , for some function ϕ as above. Then if y ∈ X, there is an
xi , i = 1, . . . n so that y ∈ V (xi ), so |f (y) − f (xi )| ≤  and |g(y) − g(xi )| < .
On the other hand, f (xi ) − cϕ(i) | ≤  and |g(xi ) − cϕ(i) | ≤ ; so in the end
we have |f (y) − g(y)| ≤ 4 for each i, or ||f − g|| ≤ 4 (sup norm.)