The Stone-Weierstrass Theorem

Matt Young
MATH 328 Notes
Queen’s University at Kingston
Winter Term, 2006

The Weierstrass Approximation Theorem shows that the continuous real-

valued fuctions on a compact interval can be uniformly approximated by
polynomials. In other words, the polynomials are uniformly dense in C([a, b], R)
with respect to the sup-norm. The original proof was given in [1] in 1885.
There are now several different proofs that use vastly different approaches.
One well-known proof was given by the Russian Sergei Bernstein in 1911.
His proof uses only elementary methods and gives an explicit algorithm for
approximating a function by the use of a class of polynomials now bearing
his name. It will be seen that the Weierstrass Approximation Theorem is in
fact a special case of the more general Stone-Weierstrass Theorem, proved by
Stone in 1937, who realized that very few of the properties of the polynomials
were essential to the theorem. Although this proof is not constructive and
relies on more machinery than that of Bernstein, it is much more efficient
and has the added power of generality.
First, Bernstien’s proof of the Weierstrass Approximation Theorem, which
is taken from [4], is examined. The Bernstein polynomials, which play a cen-
tral role in Bernstein’s proof, are introduced below.

Definition 14.1 (Bernstein Polynomials). For each n ∈ N, the nth

Bernstein Polynomial of a function f ∈ C([0, 1], R) is defined as
n   
X k n
Bn (x, f ) := f xk (1 − x)n−k .
n k
k=0

Theorem 14.1 (Weierstrass Approximation Theorem (1885)). Let

f ∈ C([a, b], R). Then there is a sequence of polynomials pn (x) that converges
uniformly to f (x) on [a, b].

Proof. Consider first f ∈ C([0, 1], R). Once the theorem is proved for this
case, the general theorem will follow by a change of variables. Since [0, 1]

1
is compact, the continuity of f implies uniform continuity. So, given  > 0,
there exists δ > 0 such that

|x − y| ≤ δ =⇒ |f (x) − f (y)| ≤ ∀x, y ∈ [0, 1].
2
Now, let M := kf k∞ . Note that M exists since f is a continuous function on
a compact set. Now, fix ξ ∈ [0, 1]. Then, if |x − ξ| ≤ δ, then |f (x) − f (ξ)| ≤ 2
by continuity. Alternatively, f |x − ξ| ≥ δ, then
 2
x−ξ 
|f (x) − f (ξ)| ≤ 2M ≤ 2M + .
δ 2

Combining the above two inequalities, we see that

 2
x−ξ 
|f (x) − f (ξ)| ≤ 2M + ∀x ∈ [0, 1].
δ 2

The Bernstein Polynomials can be used to approximate f on [0, 1]. First,

note that
n   
X k n
Bn (x, f −f (ξ)) = (f − f (ξ)) xk (1−x)n−k = Bn (x, f )−f (ξ)Bn (x, 1).
n k
k=0

And,
n  
X n
Bn (x, 1) = xk (1 − x)n−k = (x + (1 − x))n = 1
k
k=0

where the Binomial Theorem was used in the second equality. Thus,

|(Bn (x, f ) − f (ξ)| = |Bn(x, f − f (ξ))|

2  
≤ Bn x, 2M x−ξ δ
+2
2
= 2M Bn x, (x − ξ) + 2


δ2

where in the second step the fact that 0 ≤ Bn (x, f ) for 0 ≤ f and Bn (x, g) ≤
Bn (x, f ) if g ≤ f were used. Both can be proven directly from the definition
of Bn (x, f ). It can also be shown that
1
Bn (x, (x − ξ)2 ) = x2 + (x − x2 ) − 2ξx + ξ 2 .
n

2
So
 2M 2M 1
|(Bn (x, f ) − f (ξ)| ≤ + 2 (x − ξ)2 + 2 (x − x2 ).
2 δ δ n
In particular,
 2M 1
+ 2 (ξ − ξ 2 ).
|(Bn (ξ, f ) − f (ξ)| ≤
2 δ n
A simple calculation shows that on [0, 1], the maximum of z − z 2 is 14 . Thus,
 M
|(Bn (ξ, f ) − f (ξ)| ≤ + 2 .
2 2δ n
M
So, take N ≥ 2δ 2 
. Then, for n ≥ N ,

k(Bn (ξ, f ) − f (ξ)k∞ ≤ .

This proves the theorem for continuous functions on [0, 1]. Now, let
g ∈ C([a, b], R). Consider the function φ : [0, 1] → [a, b] defined by φ : x 7→
(b − a) x − a. φ is clearly a homeomorphism. Thus, the composite function
f := g ◦ φ is a continuous function on [0, 1]. By an application of the theo-
rem for functions on [0, 1], the case for an arbitrary interval [a, b] follows. 

Although tedious, the above proof relied on only elementary methods.

The introduction of some new concepts and results is required before the
proof of the Stone-Weierstrass Theorem can be approached.

Definition 14.2 (Unital Sub-algebra, Separating Points). Let K be

a compact metric space. Consider the Banach algebra

C (K, R) := {f : K → R | f is continuous}

equipped with the sup-norm,

kf k∞ := sup |f (t)|.
t∈K

Then,
1) A ⊂ C (K, R) is a unital sub-algebra if 1 ∈ A and if f, g ∈ A, α, β ∈ R
implies that αf + βg ∈ A and f g ∈ A.

2) A ⊂ C(K, R) separates points of K if for all s, t ∈ K with s 6= t, there

exists f ∈ A such that f (s) 6= f (t) .

3
 The above definition can also be generalized by replacing the requirement
that K be a compact metric space by requiring that K be a compact topo-
logical space.

Remarks. (a) It follows from the above definition that if A is a unital sub-
algebra, then all constant functions are elements of A.
(b) Let P([a, b] , R) be the space of polynomials from [a, b] to R. It is easily
checked that P([a, b] , R) is a unital sub-algebra and separates points.

Before stating and proving the Stone-Weierstrass Theorem a useful lemma

about closed sub-algebras will be proven. The following definition must first
be made.
Definition 14.3 (Lattice). A subset S ∈ C (K, R) is a lattice if, for all
f, g ∈ S, f ∨ g ∈ S and f ∧ g ∈ S, where
(f ∨ g)(x) := max{f (x), g(x)} and (f ∧ g)(x) := min{f (x), g(x)}.
p
Remark. In the following lemma, the Taylor series of the function 1 − g(t),
where
√ 0 ≤ g(t) ≤ 1, is investigated. It is useful to first study the Taylor series
of 1 − t. Formally,
∞
√ X
1−t=1− an t n
n=1
where, for n ∈ N,
n−1 
1 (−1)n−1 Y 1
  
(2n − 2)!
an = (−1) n−1 2 = − k = 21−2n .
n n! 2 n!(n − 1)!
k=0

Note that for all n ∈ N, an ≥ 0. Observe that

 
 = lim 2n − 1 = 1.
 an+1 
lim 
n→∞ an  n→∞ 2(n + 1)
Hence, by the ratio test, the series converges pointwise
√ for t ∈ (−1, 1). We
will Pshow that the series does indeed converge to 1 − t. Define ψ(t) :=
1− ∞ an tn . So, ψ(t) converges for t ∈ (−1, 1). Similarly, we obtain
n=1P
dψ(t)
dt
=− ∞ n=1 nan t
n−1
. Multiplying, we see that
dψ(t)
ψ(t) = −2(1 − t) .
dt
4
Then,
1 √
Z Z
dt dψ
− = =⇒ ψ(t) = c 1 − t
2 1−t ψ
for some c ∈ R. Evaluating both sides at t = 0 gives c = 1. Hence, ψ(t) =
√
1 − t pointwise for t ∈ [0, 1). Now, we must show that ψ(1) = 0. Stirling’s
inequality states
7 1 1
e 8 −n nn+ 2 < n! < e1−n nn+ 2 .
Hence, for n ≥ 2,
3 3
1−2n 1 (2n − 2)2n− 2 (n − 1)2n− 2
1 1 1
an < 2 7 1 1 <√ 7 2n
=√ 7 3 .
e 4 nn+ 2 (n − 1)n− 2 2e 4 (n − 1) 2e 4 (n − 1) 2
So,
∞ ∞
X 1 1 X 1
an < +√ 7 3 < ∞
n=1
2 2e 4 n=2 (n − 1) 2
where the comparison test was used in the first inequality and p-series test
P
k
in the second. Since n=1 an is monotonically increasing and bounded
k≥1
above, ψ(1) exists. Then, by Abel’s Theorem1 ,
∞ ∞
X X √
ψ(1) = 1 − an = 1 − lim− an tn = lim− 1 − t = 0.
t→1 t→1
n=1 n=1
√
Hence, again using Abel’s Theorem, ψ(t) = 1 − t uniformly for t ∈ [0, 1].

Lemma 14.1 Let A ⊂ C(K, R) be a closed unital sub-algebra. Then

√
i) if f ∈ A and f ≥ 0, then f ∈ A;

ii) if f ∈ A, then |f | ∈ A;

iii) A is a lattice.
1 n
P∞
Theorem (Abel). Let f (x) := n=0 cn (x − x0 ) , and assume that the series
converges at x = x0 + R, for some R ∈ (0, ∞). Then the series is uniformly convergent
on [x0 , x0 + R] and
∞
X
limx→(x0 +R)− = f (x) = f (x0 + R) = f (x) = cn R n .
n=0

5
Proof. To see i), consider without restriction

0 ≤ f ≤ 1.

Then, we can write

f =1−g
with
0 ≤ g ≤ 1.
Using a Taylor series expansion, we can write, formally,
∞
X
p p
f (t) = 1 − g(t) = 1 − an g n (t)
n=1

where the
√ coefficients are as in the above remark. The Taylor series approx-
imates f uniformly in k · k∞ . Indeed,
N N
√ X
n
p X
k f − (1 − an g )k∞ ≤ sup| f (t) − (1 − an g n (t))|
n=1 t∈K n=1
N
X
p
= sup | ζ − (1 − an ζ n )|.
ζ∈[0,1] n=1
√
So, given  > 0, by the uniform convergence of the Taylor series of 1−t
on [0, 1], there exists N̄ ∈ N such that
N
X
p
k f − (1 − an g n )k∞ <  ∀N ≥ N̄ .
n=1

That is,
N
X
p
lim k f − (1 − an g n )k∞ = 0.
N →∞
n=1

Since for all n ∈ N ∪ {0} g n ∈ A, and A√is a sub-algebra, 1 − N n

P
n=1 an g ∈ A.
And, since A is closed by hypothesis, f ∈ A. This completes the proof of
i). To prove ii), note that p
|f | = f 2 .

6
So, for f ∈ A, f · f = f 2 ∈ A, since A is an algebra. Applying i) to f 2 implies
|f | ∈ A. To prove iii), note that
1 1
f ∧ g = (f + g − |f − g|) and f ∨ g = (f + g + |f − g|).
2 2
Applying ii) to the above identities gives the desired result. 

We are now in a position to state and prove the Stone-Weierstrass the-

orem. The theorem was first proved by Stone in 1937 in [2]. However, he
greatly simplified his proof in 1948 into the one that is commonly used today.
See [3].

Theorem 14.2 (Stone-Weierstrass Theorem (1937)). Let K be a

compact metric space and A ⊂ C(K, R) a unital sub-algebra which separates
points of K. Then A is dense in C(K, R).

Remark. An equivalent statement is that if A is a closed unital sub-

algebra that separates points of a compact set K, and A ⊂ C(K, R), then
A = C(K, R). We will proceed using this formulation.

Proof. Let A ⊂ C(K, R) be a closed unital sub-algebra that separates points

of K. Let  > 0 be given. For any f ∈ C(K, R) we will show that there
exists g ∈ A such that
kf − gk∞ < .
Consider points s,t ∈ K. Since A separates points, there exists h ∈ A such
that h(s) 6= h(t). For some λ , µ ∈ R, define h̃ : K → R by

h(v) − h(t)
h̃(v) := µ + (λ − µ) ∀v ∈ K.
h(s) − h(t)

Note that h̃ ∈ A and h̃(s) = λ , h̃(t) = µ. Thus, for s 6= t, there exists

fs,t ∈ A such that
fs,t (s) = f (s)
and fs,t (t) = f (t).
So, fs,t approximates f in neighbourhoods around s and t. Now, fix s, and
let t vary. Put
Ut := {v ∈ K|fs,t (v) < f (v) + }.

7
Ut is open because it is the pre-image of an open set. Also, t ∈ Ut . So,
S
t∈K Ut is clearly an open cover of K. By the compactness of K, there exists
finitely many t1 , . . . , tn ∈ K such that
n
[
K⊂ Uti .
i=1

Put
hs := min fs,ti .
1≤i≤n

Then,
hs ∈ A
and hs (s) = f (s)
and hs < f + .
Now, define
Vs := {v ∈ K|hs (v) > f (v) − }.
S
Note that Vs is open and K ⊂ s∈K Vs . By compactness, there exists finitely
many s1 , . . . , sm ∈ K such that
m
[
K⊂ V sj .
j=1

Put g = max hsj . Then, g ∈ A and

1≤j≤m

f −  < g < f + .
That is,
kf − gk∞ < .
So, A is dense in C(K, R). And, since A is closed, A = C(K, R). 
Corollary 14.1 Let X be a compact subset of Rn for some n ∈ N. Then
the algebra of all polynomials P(X, R) in the coordinates x1 , . . . , xn is dense
in C(K, R).
Remarks. (a) The case in which n = 1 in the above corollary is the Weier-
strass Approximation Theorem.
(b) A result similar to the Weierstrass Approximation Theorem occurs in the
theory of Fourier series, and was also first proved by Weierstrass. It states
that a continuous 2π periodic function can be uniformly approiximated on R
by the trigonometric polynomials. The result is stated formally and proved
below.

8
Definition 14.4 (Trignometric Polynomials). The space of real-valued
trignometric polynomials T P(R, R) are functions f : R → R which are finite
sums of the form
N
X
f (x) = a0 + (an cos(nx) + bn sin(nx)).
n=1

Corollary 14.2 The set of all trigonometric polynomials are uniformly dense
in Cper ([0, 2π], R).
Proof. First note that T P(R, R) is an algebra. The only non-trivial re-
quirement to check is closure under multiplication. Indeed, for m, n ∈ R, we
have
1
sin mt cos nt = (cos (m − n)t − cos (m + n)t) ∈ T P(R, R).
2
Similar identities hold for other products. Also, note that 1 is the trigno-
metric polynomial with a0 = 1, an = bn = 0 for all n ∈ N. To use the
Stone-Weierstrass Theorem, we identify Cper ([0, 2π], R) with C(T, R), where
T = {(x1 , x2 ) ∈ R2 |x21 + x22 = 1}
is the unit circle. Let t 7→ (cos t, sin t). Then, the trigonometric polynomials
are polynomials in x1 , x2 on T, which are dense in C(T, R) by the previous
corollary with n = 2. 

Remark. Note that the above corollary can be easily generalized for Cper ([a, b], R).

We can also use the Stone-Weierstrass Theorem to prove the complex Stone-
Weierstrass Thereom. The complex version, however, requires additional
assumptions. The following proof is taken from [5].
Theorem 14.3 (Complex Stone-Weierstrass Theorem) Let A be a
(complex) unital sub-algebra of C(K, C), such that if f ∈ A, then f¯ ∈ A, and
A separates points of K. Then, A is dense in C(K, C).
Proof. Let f ∈ A. Then, Re(f ) = 12 (f + f¯) ∈ A and Im(f ) = 2i1 (f − f¯) ∈
A. Denote by AR the unital sub-algebra of A containing all real-valued
functions. Note that A separates points in K because AR does. By the
Stone-Weierstrass Theorem for real valued functions, AR is dense in C(K, R).
Since A = {f + ig|f, g ∈ AR }, we see that A is indeed dense in C(K, C) =
C(K, R) + iC(K, R). 

9
References
[1] K. Weierstrass (1885). Über die analytische Darstellbarkeit sogenannter
willkrlicher Functionen einer reellen Vernderlichen. Sitzungsberichte der
Kniglich Preuischen Akademie der Wissenschaften zu Berlin, 1885 (11).

[2] M. Stone (1937). Applications of the Theory of Boolean Rings to General

Topology. Translations of the Americain Mathematical Socienty 41 (3),
375-481.

[3] M. Stone (1948). The Generalized Weierstrass Approximation Theorem.

Mathematics Magazine 21 (21), 167-184 and 21 (5), 237-254.

[4] K. Davidson and A. Donsig. Real Analysis with Real Applications. Pren-
tice Hall, Upper Saddle River, N.J., 2002.

[5] H. Sohrab. Basic Real Analysis. Birkhäuser, New York, N.Y., 2003.

10