TMA4230 Functional analysis

2005

The Weierstrass density theorem

n
t k (1 t )nk = t + (1 t ) = 1.
(2)
k
k=0
We can show by induction on n that if b = 0 then all the coefficients n are zero. The
base case n = 0 is obvious. When n > 0, a little bit of binomial coefficient gymnastics shows that the derivative of a Bernshten polynomial can be written as another
Bernshten polynomial:

n1
X
n 1 k
b 0 (t ) = n
(k+1 k )
t (1 t )n1k .
k
k=0
In particular, if b = 0 it follows by the induction hypothesis that all k are equal,
and then they are all zero, by (2).
In other words, the polynomials t k (1 t )nk , where k = 0, . . . , n, are linearly independent, and hence they span the n + 1-dimensional space of polynomials of
degree 1. Thus all polynomials can be written as Bernshten polynomials, so there
is nothing special about these only about the way we write them.
To understand why Bernshten polynomials are so useful, consider the individual polynomials

n k
bk,n (t ) =
t (1 t )nk ,
k = 0, . . . , n.
(3)
k
If we fix n and t , we see that bk,n (t ) is the probability of k heads in n tosses of a
biased coin, where the probability of a head is t . The expected number of heads in
such an experiment is nt , and indeed when n is large, the outcome is very likely to
be near that value. In other words, most of the contributions to the sum in (1) come
from k near nt . Rather than using statistical reasoning, however, we shall proceed
by direct calculation but the probability argument is still a useful guide.
1 Named after Serge Natanovich Bernshten (18801968). The name is often spelled Bernstein.
2 When n = 3, we get a cubic spline. In this case, , , and are called the control points of the
0 1 2
3

spline. In applications, they are usually 2- or 3-dimensional vectors.

Version 20050417

The Weierstrass density theorem

1 Theorem. (Weierstrass) The polynomials are dense in C [0, 1].

k n
n
X

t k (1 t )nk ,
f (t ) f

n
k
k=0

(4)

We now use the fact that f is uniformly continuous: Let > 0 be given. There is then
a > 0 so that | f (t ) f (s)| < whenever |t s| < . We now split the above sum into
two parts, first noting that
k n
X

t k (1 t )nk
(5)

f (t ) f
k
n
|knt |<n
(where we used | f (t ) f (k/n)| < , and then expanded the sum to all indexes from
0 to n and used (2)). To estimate the remainder, let M = k f k , so that

k n
X
X
n k

t k (1 t )nk 2M
t (1 t )nk .
(6)
f (t ) f

n
k
k
|knt |n
|knt |n
To finish the proof, we need to borrow from the Chebyshev inequality in order to
show that the latter sum can be made small. First we find

n
n1
X
X n 1 k
n k
nk
k
t (1 t )
= nt
t (1 t )n1k = nt .
(7)
k
k
k=0
k=0
(Rewrite the binomial coefficient using factorials, perform the obvious cancellation using k/k! = 1/(k 1)!, put nt outside the sum, change the summation index,
and use (2).) Next, using similar methods,
n
X
k=0

k(k 1)

n2
X n 2 k
n k
t (1 t )nk = n(n 1)t 2
t (1 t )n2k = n(n 1)t 2 .
k
k
k=0

Version 20050417

The Weierstrass density theorem

Adding these two together, we get

k2

k=0

n k
t (1 t )nk = nt (n 1)t + 1 .
k

(8)

Finally, using (2), (7) and (8), we find

(nt k)2

k=0

n k
t (1 t )nk = (nt )2 2(nt )2 + nt (n 1)t + 1 = nt (1 t ).
k

The most important feature here is that the n 2 terms cancel out. We now have

n k
nt (1 t )
(nt k)
t (1 t )nk
k
|knt |n

X
n k
(n)2
t (1 t )nk ,
k
|knt |n
X

so that

1
n k
t (1 t )

.
t (1 t )nk
2
2
n
4n
k
|knt |n
X

(9)

Combining (4), (5), (6) and (9), we end up with

M
,
2n2

(10)

which can be made less than 2 by choosing n large enough. More importantly,
this estimate is independent of t [0, 1].
One final remark: There is of course nothing magical about the
interval [0,
1]. Any
closed and bounded interval will do. If f C [a, b] then t 7 f (1 t )a + t b belongs
to C [0, 1], and this operation maps polynomials to polynomials and preserves the
norm. So the Weierstrass theorem works equally well on C [a, b].
The StoneWeierstrass theorem is a bit more difficult: It replaces [0, 1] by any compact set
X and the polynomials by any algebra of functions which separates points in X and has no
common zero in X . (This theorem assumes real functions. If you work with complex functions, the algebra must also be closed under conjugation. But the complex version of the
theorem is not much more than an obvious translation of the the real version into the complex domain.) One proof of the general StoneWeierstrass theorem builds on the Weierstrass
theorem. More precisely, the proof needs an approximation of the absolute value |t | by polynomials in t , uniformly for t in a bounded interval.

Version 20050417

The Weierstrass density theorem

An amusing (?) diversion. Any old textbook on elementary statistics shows pictures of the
binomial distribution, i.e., bk,n (t ) for a given n and t ; see (3). But it can be interesting to look
at this from a different angle, and consider each term as a function of t . Here is a picture of
all these polynomials, for n = 20:

0.8

0.6

0.4

0.2

0.2

0.4

0.6

0.8

t
R
We may note that bk,n (t ) has its maximum at t = k/n, and 01 bk,n (t ) d t = 1/(n + 1). In fact,
(n + 1)bk,n is the probability density of a beta-distributed random variable with parameters
(k + 1, n k + 1). Such variables have standard deviation varying between approximately
p
1/(2 n) (near the center, i.e., for k n/2) and 1/n (near the edges). Compare this with the
distance 1/n between the sample points.
It is tempting to conclude that polynomials of degree n can only do a good job of approxp
imating a function which varies on a length scale of 1/ n.
We can see this, for example, if we wish to estimate a Lipschitz continuous function f ,
say with | f (t ) f (s)| L|t s|. Put = L in (10) and then determine the that gives the
best estimate in (10), to arrive at | f (t ) bn (t )| < 32 M 1/3 (L2 /n)2/3 . So the n required for a
given accuracy is proportional to L2 , in accordance with the analysis in the previous two
paragraphs.
Reference: S. N. Bernshten: A demonstration of the Weierstrass theorem based on the theory of probability. The Mathematical Scientist 29, 127128 (2004).
By an amazing coincidence, this translation of Bernshtens original paper from 1912 appeared recently. I discovered it after writing the current note.

Version 20050417