Math 857 Fall 2015

1 Orthonormal sets in Hilbert space

Let S ⊆ H. We denote by [S] the span of S, i.e., the set of all linear
combinations of elements from S. A set {uα : α ∈ A} is called orthonormal,
if huα , uβ i = 0 for all α 6= β and kuα k = 1 for all α. (Here A is some index
set.)
For every x ∈ H we define a transform x b : A → C by xb(α) = hx, uα i and
call these the Fourier coefficients of x with respect to {uα : α ∈ A}.
Let F ⊆ A be finite and set MF = [{uα : α ∈ F }]. We observe the
following facts.

1. If ϕ : A → C with ϕ(α) = 0 for α ∈

/ F , then y ∈ MF defined by
X
y= ϕ(α)uα
α∈F

satisfies yb(α) = ϕ(α). Also,

X
kyk2 = |ϕ(α)|2 .
α∈F

2. If x ∈ H and sF is defined by
X
sF,x = x
b(α)uα ,
α∈F

then
kx − sF,x k < kx − sk
for every s ∈ MF with s 6= sF . Moreover,
X
x(α)|2 ≤ kxk2 .
|b
α∈F

The first statement of these follows immediately from the orthogonality

conditions. For the second part note that sF,x and x have the same Fourier
coefficient for α ∈ F , i.e., x − sF,x ⊥ MF . Since sF,x ∈ MF , we obtain

x − sF,x ⊥ s − sF,x

1
for all s ∈ MF . Hence for s ∈ MF
kx − sk2 = kx − sF,x k2 + ksF,x − sk2
and the second term on the right is zero only if s = sF,x . The choice s = 0
gives the last inequality. (This means in particular that sF,x is the unique
best approximation to x in MF with respect to k.k)
Example. Rewrite these statements if H = L2 ([0, 1]) and the orthonormal
system is given by the exponentials un (t) = e2πitn where n ∈ Z.
We would like to remove the finiteness condition from the previous state-
ments. Let A be an arbitrary index set and 0 ≤ ϕ(α) ≤ ∞ for every α ∈ A.
Then X
ϕ(α)
α∈A
is short notation for the supremum of the set of all finite sums ϕ(α1 ) + ... +
ϕ(αn ) with αi ∈ A. (In Math 750 terms: the series is the Lebesgue integral
of ϕ with respect to counting measure on A.)
We write `2 (A) to indicate the class of functions ϕ with
X
|ϕ(α)|2 < ∞.
α∈A

We note that this is also a Hilbert space with scalar product

X
hϕ, ψi = ϕ(α)ψ(α).
α∈A

We note that the simple functions are dense in every Lp space. In par-
ticular, the set of functions ϕ that are zero on all but finitely many elements
of A is dense in `2 (A). For completeness we include a proof of the density
statement.
Theorem 1. Let µ be a Borel measure, and let S be the class of complex,
measurable, simple functions on X so that
µ({x : s(x) 6= 0}) < ∞.
If 1 ≤ p < ∞, then S is dense in Lp (µ).
Proof. Evidently S ⊆ Lp (µ). For the other direction, suppose first f ≥ 0
in Lp (µ). Let sn be a sequence of simple functions approximating f from
below. Then sn ∈ Lp and hence in S. Since |f − sn |p ≤ f p , dominated
convergence shows that the p-norm of the difference goes to zero, and the
complex case follows by taking real and imaginary parts, followed by taking
positive and negative parts for each.

2
Lemma 1. If ϕ ∈ `2 (A), then {α ∈ A : ϕ(α) = 0} is at most countable.

Proof. Let An = {α ∈ A : |ϕ(α)| ≥ 1/n}. Then

X X X
1≤ |nϕ(α)|2 ≤ n2 |ϕ(α)|2
α∈An α∈An α∈A

and the right side is finite. Hence An is a finite set, and the set of values
where ϕ is nonzero is a countable union of finite sets.

Definition 1. Let (X1 , d1 ) and (X2 , d2 ) be two metric spaces. A map

F : X1 → X2 is called an isometry, if

d2 (F (a), F (b)) = d1 (a, b)

for all a, b ∈ X1 .

The next goal is to show that the map F : H → `2 (A) defined by F(x) =
x
b is an isometry from the span of linear combinations of an orthonormal basis
{uα } onto `2 (A).

Theorem 2. Let X, Y be two metric spaces where X is complete. Assume

1. f : X → Y is continuous,

2. X has a dense subspace X0 on which f is an isometry,

3. f (X0 ) is dense in Y .

Then f is an isometry of X onto Y .

Proof. From continuity of f it is immediately clear that f is an isometry on

X. Let y ∈ Y . Let (xn ) ⊆ X0 be a sequence with f (xn ) → y. The sequence
f (xn ) is therefore Cauchy, and since f is an isometry on X0 , (xn ) is Cauchy
and by completeness of x has a limit. Continuity of f implies f (x) = y.

Theorem 3. Let U = {uα : α ∈ A} be an orthonormal set in H, and let P

be the space of finite linear combinations of U . Then for every x ∈ H,
X
x(α))|2 ≤ kxk2 ,
|b
α∈A

and F : H → `2 (A) defined by F(x) = x b is a continuous linear mapping

whose restriction to P is an isometry onto `2 (A).

3
Proof. We had seen that the inequality holds for every finite set F ⊆ A.
Theorem 1 implies that it holds for all x ∈ H. (This is also called Bessel’s
inequality.)
It follows from this inequality that F maps H into `2 (A). Evidently F
is linear, and an application of Bessel to x − y shows that F is continuous.
We had seen before that F is an isometry of P onto the subspace of
all elements in `2 (A) with finite support. This subspace is dense in `( A)
(Theorem 1 again). From Theorem 2 it follows that F is an isometry of P
onto `2 (A). (This is also called the Riesz-Fischer theorem.)

Theorem 4. Each of the following four conditions on an orthonormal set

uα implies the other three.

1. {uα } is a maximal orthonormal set in H,

2. The set P of all finite linear combinations of elements from {uα } is

dense in H,

3. The equality X
x(α)|2 = kxk2
|b
α∈A

holds for all x ∈ H,

4. The equality X
x y (α) = hx, yi
b(α)b
α∈A

holds for all x, y ∈ H.

By defining the scalar product

X
ha, bi`2 (A) = a(α)b(α),
α∈A

the last identity can be written as hb

x, ybi`2 (A) = hx, yiH . Maximal orthonor-
mal sets are also called orthonormal bases.

Proof. To say that {uα } is maximal means that no vector from H can be
added to this set in such a way that the resulting set is still orthonormal.
(See also the current homework.)
Assume that p is not dense in H. Then there exists x ∈ H\P . By the
⊥
theorem about closed subspaces, there exists y ∈ P of norm 1. This can
be added to {uα } to yield a larger orthonormal set.

4
 It follows that (1) impliies (2).
The previous theorem showed that the Fourier transform is an isometry
on P . If this is all of H, then (3) follows.
Polarization:

4hx, yi = kx + yk2 − kx − yk2 + ikx + iyk2 3 − ikx − iyk2 .

Hence every norm identity yields a corresponding scalar product identity.

In particular, (3) implies (4).
Finally, if (1) does not hold, then there exists u ∈ H with hu, uα i = 0
for all α and kuk = 1. The evidently every Fourier coefficient of u is zero,
hence so is the left side of (4), but the right side with x = y = u is 1.

2 Application to the triogonmetric system

From the previous theorem we know that in order to prove that the system
{un }n∈Z with
un (t) = e2πint
is an orthonormal basis of H = L2 ([0, 1]), we need to prove that it is dense
in H. Recall that X can be the real line or the unit circle. We denote by
Cc (x) the continuous functions with compact support in X. One more tool
from Math 750:

Theorem 5. For 1 ≤ p < ∞, Cc (X) is dense in Lp (µ) where µ is a Borel

measure on X.

Proof. Recall that S is the set of complex, measurable, simple functions on

X. Lusin’s Theorem (Folland, Section 2.4, Exercise 44):
For every s ∈ S and ε > 0 there exists g ∈ Cc (X) such that g = s except
on a set of measure < ε, and |g| ≤ ksk∞ .
Putting this together leads to

kg − skp ≤ 2ε1/p ksk∞ .

We had seen that S is dense in Lp (µ), hence any f ∈ Lp (µ) can be

approximated by functions in Cc (X).

For X = [0, 1] this means that the continuous functions are dense in
L2 ([0, 1]). Thus we need to prove that every continous function can be

5
approximated in arbitarily close in L2 by trigonometric polynomials. There
is a useful connection between k.k2 and k.k∞ on compact sets:
Z 1/2
|g|2 dµ ≤ kgk∞ ,

hence it suffices to approximate continuous functions by trigonometric poly-

nomials in L∞ norm.
Goal. Let f be continuous, and let ε > 0. Show that there exists a finite
sequence of coefficients an = an (ε) so that fε defined by
N
X
fε (t) = an e2πint
n=−N

satisfies
sup |f (x) − fε (x)| < ε.
x∈[0,1]

We constructR fε via convolution. Some preliminary ideas: Let ϕ ∈

L2 ([0, 1]) with ϕ = 1. Define
Z 1
fϕ (x) = f (x − u)ϕ(u)du.
0

Two crucial identities: First,

fbϕ (n) = fb(n)ϕ(n),

b

and secondly
Z 1
fϕ (x) − f (x) = ϕ(u)(f (x) − f (x − u))du.
0

This means that if ϕ is a trigonometric polynomial of degree N , then

automatically fϕ is as well, and if we want to estimate the difference, it is
sufficient to estimate the differences under the integral sign!
Assume that we can find a family {Qk } of trigonometric polynomials
with the following properties.

(i) Qk (x) ≥ 0 for all x ∈ R,

R 1/2
(ii) −1/2 Qk (x)dx = 1 for all k,

6
(iii) If 0 < δ < 1/2, then Qk (x) → 0 uniformly for all δ ≤ |x| ≤ 1/2 as
k → ∞.

We show first that the existence of such a family implies the desired
density statement. Let ε > 0. Let δ > 0 so that

|x − t| < δ implies |f (x) − f (t)| < ε (1)

for all x and t. (Note that our assumptions imply uniform continuity of f .)
We obtain from property (iii) that there exists k0 such that for all k ≥ k0
and δ ≤ |x| ≤ 1/2 we have
ε
Qk (x) ≤ . (2)
2kf k∞
Define
Z 1 Z 1/2
fk (x) = f (u)Qk (x − u)du = f (x − u)Qk (u)du.
0 −1/2

We note that fk is a trigonometric polynomial since Qk is a finite linear

combination of exponentials; plug the corresponding representation of Qk
into the first integral and change summation and integration. (The two rep-
resentations can be shown to be equal with a change of variable.) Property
(ii) implies that
Z 1/2
fk (x) − f (x) = (f (x − u) − f (x))Qk (u)du.
−1/2

Break the integral into two pieces, one over |u| ≤ δ and the other over
δ ≤ |u| ≤ 1/2. Note that (1) and property (ii) imply
Z
(f (x − u) − f (x))Qk (u)du ≤ ε
|u|≤δ

and that (2) implies

Z Z
(f (x − u) − f (x))Qk (u)du ≤ 2kf k∞ Qk (u)du ≤ ε.
δ≤|u|≤1/2 δ≤|u|≤1/2

It remains to show that a family {Qk } with the stated properties exists.
We define
1 + cos 2πx k
 
Qk (x) = ck
2

7
 R1
where ck is chosen so that 0 Qk (x)dx = 1. If you have seen Gamma func-
tions, √
k! π
ck = ,
2Γ[k + 21 ]
but this is not necessary to know.
Expanding Qk using the binomial theoremR implies that Qk is a trigono-
metric polynomial. Evidently Qk ≥ 0, and Qk = 1 by construction. It
remains to show that Qk goes to zero uniformly away from the origin. We
note first that Qk is decreasing on [0, 1/2]. Hence for δ > 0 and δ ≤ x ≤ 1/2
we have
Qk (x) ≤ Qk (δ),
and if we can show that this value goes to zero, the uniform convergence
follows. First, an inequality for ck : Since Qk is even, we have
Z 1/2  k Z 1/2  k
1 + cos(2πt) 1 + cos(2πt)
1 = 2ck dt > 2ck sin(2πt)dt,
0 2 0 2

note that sin(2πt) ≥ 0 on [0, 1/2]. The integral on the right hand side can
be evaluated to (π(k + 1))−1 . Hence

π(k + 1)
ck ≤ ,
2
and we obtain  k
π(k + 1) 1 + cos(2πδ)
Qk (δ) ≤ .
2 2
This goes to zero for fixed 0 < δ < 1/2 as k → ∞, since it is of the form
C(k + 1)η k with fixed C > 0 and 0 < η < 1.