4

Fourier Analysis in Hilbert Space

In the last section of Chapter 3 we introduced the Lebesgue Lp -spaces for general
measures and discussed their most basic properties. The most important Lp -space,
by far, is L2 . Its importance is its role in applications, especially in Fourier analysis.
The material of this chapter lies at the foundation of the branch of mathematics
called harmonic analysis.
In this chapter we will see that L2 is a Hilbert space (we already really have all the
bits of information we need to see this) and that in some sense the L2 -spaces (with
different µ’s) are the only Hilbert spaces. We will come to see how the problem
that Fourier examined, about decomposing functions as infinite sums of other —
somehow more basic — functions, is a problem best phrased and understood in the
language of abstract Hilbert spaces. One of the triumphs of functional analysis is
to take a very concrete problem — in this case Fourier decomposition — view it in
an abstract setting, and use theoretical tools to obtain powerful results that can be
translated back to the concrete setting. Fourier’s work certainly holds an important
spot at the roots of functional analysis, and it motivated much early work in the
development of the field.
Further Hilbert space theory appears in Section 5.4.

4.1 Orthonormal Sequences

During the second half of the eighteenth century and first decade of the nineteenth
century, infinite sums of sines and cosines appeared as solutions to physical prob-
 4.1 Orthonormal Sequences 75

lems then being studied. Daniel Bernoulli (1700–1782; Netherlands)1 suggested

that these sums were solutions to the problem of modeling the vibrating string,
and Joseph Fourier (1768–1830; France) proposed them as solutions to the prob-
lem of modeling heat flow. It is not really until the response to Fourier’s work
that we see other mathematicians coming to grips with the challenge that these
infinite sums truly posed: to understand the fundamental notions of convergence
and continuity. Over the decades following the appearance of Fourier’s works on
heat, the field of “real analysis” would be born in large part out of efforts to re-
spond to the challenges that Fourier’s work raised in pure mathematics. Many of
the great mathematicians of the period — perhaps most notably Cauchy, Riemann,
and Weierstrass — did their most important work in the development of this field.
For an excellent historical account of these mathematical developments, see [25].
Fourier begins with an arbitrary function f on the interval from −π to π and states
that if we can write

a0
∞
f (x)
+ ak cos(kx) + bk sin(kx),
2 k
1

then it must be the case that the coefficients ak and bk are given by the formulas

1 π
ak
f (x) cos(kx)dx, k
0, 1, 2, . . . ,
π −π

and
 π
1
bk
f (x) sin(kx)dx, k
1, 2, . . . .
π −π

The big question is this: When is this decomposition actually possible? Even if
the integrals involved make sense, does the series converge? If it does converge,
what type of convergence (pointwise, uniform, etc.) do we get? Even if the series
converges in some sense, does it converge to f ?
The immediate goal is to show you how these questions about Fourier series
can be treated in the abstract setting of an inner product space.
Let us now take stock of what we already know by gathering our information
about L2 . First, recall that L2
L2 (µ), for any abstract measure space (X, R, µ),
denotes the collection of all measurable functions f : X → C such that the integral

|f |2 dµ
X

1
Daniel Bernoulli is the nephew of James Bernoulli, who was mentioned at the beginning
of Section 3.1. The Bernoulli family produced several distinguished mathematicians and
physicists; at least twelve members of the family achieved distinction in at least one of these
fields.
76 4. Fourier Analysis in Hilbert Space

is finite. These functions are often called the “square integrable” functions on X.
With norm

f 2
|f |2 dµ,
X

this collection of functions becomes a Banach space. We can define an inner product
on L2 via

f, g
f gdµ.
X
It is easily seen that this is an inner product, and that the norm does indeed come
from this inner product. That is,


f 2
f, f 
|f |2 dµ.
X

Theorem 3.21 shows that L2 is a Hilbert space.

In the following definitions, the terminology should seem familiar from your
experiences with Rn .
Let (V , ·, ·) be an inner product space. We say that v √
and w in V are orthogonal
if v, w
0. We say that v is normalized if v
v, v
1. A sequence
{vk }∞
k
1 in V is an orthonormal sequence if vk , vj 
δkj , 1 ≤ k, j < ∞. The
function δkj is defined to be 1 if k
j and 0 if k
j .
In Exercise 4.1.1 you are asked to show that the trigonometric system (Figure
4.1)
1 cos(nx) sin(mx)
√ , √ , √ , n, m
1, 2, . . . ,
2π π π
is an orthonormal sequence in the inner product space L2 ([−π, π], m). From this,
you should find it plausible that the goal of Fourier analysis in its general setting
is this: Given an orthonormal sequence {fk }∞k
1 in an inner product space V and
an f ∈ V , find complex numbers ck such that


∞
f
c k fk .
k
1

The convergence of this infinite sum is in the norm induced by the inner product.
Further, it would be desirable to be able to do this for all f ∈ V . In general, this
cannot be done. Notice that Fourier was asserting that when {fk }∞ k
1 is the trigono-
metric system, the coefficients are of form f, fk  (an appropriate indexing of the
trigonometric system has not yet been established) whenever his decomposition
works.
Let {fk }∞
k
1 be an orthonormal sequence in V . If it is the case that for each f ∈ V
we can find constants ck (depending on f ) such that


∞
f
ck fk ,
k
1
 4.1 Orthonormal Sequences 77

cos (nx)
␲
0.4

0.2

–3 –2 –1 1 2 3

–0.2

–0.4

(a)

sin (mx)
␲
0.4

0.2

–3 –2 –1 1 2 3

–0.2

–0.4

(b)

FIGURE 4.1. (a) the functions cos(nx)

√
π
for n
1, 2, 3. (b) the functions sin(mx)
√
π
for m
1, 2, 3.

then we say that the sequence {fk }∞

k
1 is a complete orthonormal sequence in V . A
2

complete orthonormal sequence is sometimes called an orthonormal basis for V .

The latter terminology can cause confusion since a complete orthonormal system
is not a basis in the finite-dimensional sense discussed in Section 1.3.
The questions posed by Fourier’s work are, to some degree, answered by the fact
that the trigonometric system does indeed form a complete orthonormal sequence
in L2 . This important result appears as Theorem 4.6.
The trigonometric system is certainly an important complete orthonormal se-
quence (for the Hilbert space L2 ([−π, π])). But there are others, and we end this
section with a brief description of a few of them ([43] is a good general reference
for this topic). We can use the Gram–Schmidt process to construct an orthonormal
sequence in any inner product space.
For our first example, the Hilbert space is L2 ([−1, 1]). If one applies the Gram–
Schmidt process to the functions 1, x, x 2 , x 3 , . . . , one obtains the complete

2
Note that this is a new usage of the word “complete”; we now have at least two ways
we will use this adjective: a complete metric space, a complete orthonormal system.
78 4. Fourier Analysis in Hilbert Space

100

–1 –0.5 0.5 1

–100

–200

FIGURE 4.2. The Legendre polynomials, n
3, 4.

orthonormal sequence of Legendre polynomials (Figure 4.2),


2n + 1 1 d n 2
(x − 1)n , n
1, 2, . . . .
2 2n n! dx n
These polynomials are named for Adrien-Marie Legendre (1752–1833; France).
Next, consider the Hilbert space L2 ((0, ∞)). If one applies the Gram–Schmidt
process to the the functions x n e−x , n
0, 1, . . ., one obtains the complete or-
thonormal sequence of Laguerre functions. These appear in quantum mechanics
in the analysis of the hydrogen atom. This family is named for Edmond Laguerre
(1834–1886; France).
For our third example, the Hilbert space is L2 (R). If one applies the Gram–
−x 2
Schmidt process to the the functions x n e 2 , n
0, 1, · · ·, one obtains the
complete orthonormal sequence of Hermite functions. These also appear in
quantum mechanics. This family is named for Charles Hermite (1822–1901;
France).
The Legendre, Laguerre, and Hermite functions all show up as eigenfunctions
of certain linear operators (linear operators are the subject of the next chapter)
related to the Sturm–Liouville problem in differential equations.
The final family we discuss is the complete orthonormal sequence of Haar
functions. The Hilbert space is L2 ([0, 1]). This example is fundamentally different
from the previous examples in that the functions in this family are not continuous,
and they are not connected with differential equations. Haar functions appear in
the study of “wavelets.” Wavelet theory and its applications experienced explosive
development in the 1980s. There are several good books, at varying levels, on the
subject. A “brief” investigation of wavelets, their properties and uses, makes a good
student project ([28] gives an excellent overview and introduction to wavelets).
Wavelet series are used in signal and imaging processing and, in some contexts,
 4.1 Orthonormal Sequences 79

H2,1 H2,2
2 2

1 1

–2 –2

H2,3 H2,4
2 2

1 1

–2 –2

FIGURE 4.3. The Haar functions, H2,k (x).

are replacing the classical Fourier series. We define


 n
 −2 2 if k−1
k− 1
≤ x < 2n 2 ,
2n
H0,0 (x)
1, Hn,k (x)
2 n2 k− 1
if 2n 2 ≤ x < 2kn ,


0 otherwise ,
for n ≥ 1, 1 ≤ k ≤ 2n (Figure 4.3). This family is named for Alfréd Haar
(1885–1933; Hungary).

Jean Baptiste Joseph Fourier was born March 21, great talent in many areas by the age of
1768, in Auxerre, France (Figure 4.4). His fourteen. He wanted to join the military,
father had been a tailor, but both of his for some reason was rejected, and instead
parents were dead by the time Fourier was entered a Benedictine abbey to train for
ten. There seems to be some disagreement the priesthood. While there, he was able
among authors as to exactly how many to work on mathematics and submitted
siblings Fourier had, but by all accounts he his first paper in 1789. He never took
had many. According to [51], he was the his vows and returned to his school,
nineteenth (and not the last) child in the teaching math, history, philosophy, and
family. rhetoric. This was the time of the French
Fourier was distinguished in two fields: Revolution, and Fourier became quite
mathematics and Egyptology. He began involved in revolutionary politics. In 1794
both careers when he attended a military he was imprisoned and sentenced to be
school run by the Benedictines. He showed guillotined.
80 4. Fourier Analysis in Hilbert Space

after their arrival in Egypt, the Institut

d’Égypte opened in Cairo, and Fourier
´
was appointed secretaire perpetuel. He
had many duties in this post, including
investigating ancient monuments and
irrigation projects, but he managed to find
time to continue mathematical research.
Napoleon left for France in 1799. Fourier
followed in 1801 and was appointed by
Napoleon to a government position in
Grenoble. He held this post from 1802
until 1814. During this period, he devoted
much time to the writing of a massive
work entitled Description de l’Égypte.
This work was written by the team that
Napoleon brought with him to Egypt
FIGURE 4.4. Joseph Fourier. and is very important in the birth of the
modern field of Egyptology; it gave the
In 1795 the École Normale in Paris most comprehensive account, to date, of
opened to train teachers in an effort to ancient and contemporary Egypt. To put
rehabilitate the system of higher education this accomplishment in perspective, the
in France. The students were chosen and Rosetta Stone was discovered by this
financed by the republic. Fourier was team, and it was in 1822 that hieroglyphics
chosen to attend, and while there, he came were fully deciphered.
into contact with very good professors: It was also during his time in Grenoble
Lagrange, Laplace, and Gaspard Monge that Fourier did his work on heat diffusion.
(1746–18181; France). Unfortunately, the This work, done primarily during the period
school closed after a few months. At this 1804–1807, culminated in a monograph
time, Fourier went to teach at the École that was submitted to the Institut de
Polytechnique, which was designed as a France in Paris at the end of 1807. This
military academy to train the military elite. paper caused a great deal of controversy.
During this period he was, for a second One complaint was from Lagrange and had
time, arrested and subsequently freed. to do with the convergence of his “Fourier
Over the next few years, Fourier taught series.” Lagrange’s skepticism was on
(mathematics with military applications) target and, indeed, led to the rise of a new
and worked on mathematical research field of mathematics: “real analysis” (see
(mostly having to do with polynomials: [25]). The controversy caused Fourier to
extending Descartes’s rule of signs, ap- revise the paper and resubmit it in 1811.
proximating values of real roots, detecting Eventually, his Théorie analytique de
existence of complex roots). la chaleur was published in 1822. This
In 1798 Fourier was recommended, work is Fourier’s greatest contribution and
by Monge and the chemist Claude Louis certainly remains one of the masterpieces
Berthollet (1748–1822; France), to be of mathematical physics. It is important
Napoleon Bonaparte’s scientific advisor not only for the physical explanations that
on his expedition to Egypt. Very soon it gives, but also for the mathematical
 4.1 Orthonormal Sequences 81

techniques developed in the course of his later years focused on consequences

his attempting to explain the physics of of his earlier work. One of his other
heat flow. For example, he developed important mathematical projects during
techniques to find solutions for many this time was on problems that can now
differential equations that, up until that be viewed as precursors to the field of
point, had not been worked out. linear programming. He also did editorial
In the last fifteen years of his life, work and wrote several biographies of
Fourier continued to work on mathematics mathematicians during this period.
and on topics related to his work in Fourier died on May 16, 1830, after
Egypt. However, his most substantial being in a state of deteriorating health for
contributions had already been made, and several years.
much of his mathematical work during

4.2 Bessel’s Inequality, Parseval’s Theorem,

and the Riesz–Fischer Theorem
Let (V , ·, ·) be an inner product space, and {fk }∞
k
1 a specified orthonormal
sequence in V . Suppose we have an f ∈ V that we can decompose as


∞
f
c k fk .
k
1

What, then, are the ck ’s to be? We turn to the very simple case of R3 , with its
usual inner product, for inspiration. We take as our orthonormal family the three
Euclidean basis vectors e1
(1, 0, 0), e2
(0, 1, 0), and e3
(0, 0, 1). Then
every vector in R3 can be written in form


3
ck ek .
k
1

In this case we know that c1
v, e1 , c2
v, e2 , and c3
v, e3 . This example
illustrates the next theorem.

Theorem 4.1. Suppose that f
∞ ∞

k
1 ck fk for an orthonormal sequence {fk }k
1
in an inner product space V . Then ck
f, fk  for each k.
Before we prove this result, notice that this is, in fact, consistent with Fourier’s
assertion about the trigonometric system.

Proof. Let sn
nk
1 ck fk . Our hypothesis is thus that

lim sn − f 
0.
n→∞

Fix an m and let n ≥ m. Then

lim sn , fm 
f, fm . (4.1)
n→∞
82 4. Fourier Analysis in Hilbert Space

This is because
sn , fm  − f, fm 
sn − f, fm ,
and
|sn − f, fm | ≤ sn − f  · fm 
sn − f .
Therefore,


∞

∞

∞
sn , fm 
ck fk , fm 
ck fk , fm 
ck δkm
cm . (4.2)
k
1 k
1 k
1

Combining (4.1) and (4.2) gives the desired result. 2

Let {fk }∞
be an orthonormal sequence in V , and let f ∈ V . We call
k
1
∞ ∞
k
1 f, fk k the Fourier series of f with respect to {fk }k
1 , and f, fk  the
f
∞
Fourier coefficients of f with respect to {fk }k
1 . These objects are defined without
any assumptions or knowledge about convergence of the series.
The next theorem tells us something about the size of these coefficients.
Theorem 4.2 (Bessel’s Inequality3 ). Suppose that {fk }∞ k
1 is an orthonormal
inner product space V . For every f ∈ V , the series (of nonnegative
sequence in an

real numbers) ∞ k
1 |f, fk | converges and

2

∞
|f, fk |2 ≤ f 2 .
k
1

Proof. Consider the partial sum sn of the Fourier series for f . Then
f − sn , fk 
f, fk  − sn , fk 

n 

f, fk  − f, fj fj , fk
j
1
n 



f, fk  − f, fj fj , fk
j
1


n

f, fk  − f, fj (fj , fk 
j
1


n

f, fk  − f, fj δj k
j
1

f, fk  − f, fk 
0.
This shows that f − sn is orthogonal to each fk . Further,


n 
f − sn , sn 
f − sn , f, fk fk
k
1

3
Due to Friedrich Bessel (1784–1846; Westphalia, now Germany).
 4.2 Bessel’s Inequality, Parseval’s Theorem, and the Riesz–Fischer Theorem 83

n 



f − sn , f, fk fk
k
1


n

f, fk f − sn , fk ,
k
1

which equals zero by the previous argument. This shows that f − sn is orthogonal
to sn . Then, by Exercise 4.2.1,
f − sn 2 + sn 2
f 2 .
This shows that
sn 2 ≤ f 2 .
Since

n 2

sn 2
f, fk fk ,
k
1

which, by the same exercise and induction, is equal to

n
f, fk fk 2 ,
k
1

we have that


n

n
sn 2
|f, fk |2 · fk 2
|f, fk |2 .
k
0 k
1

Combining these last two sentences yields

n
|f, fk |2 ≤ f 2 .
k
1

Since this holds for each n,

∞
|f, fk |2 ≤ f 2 ,
k
1

as desired. 2
It is natural to want to determine conditions on {fk }∞
k
1 under which equality in
Bessel’s inequality holds.
Theorem 4.3 (Parseval’s Theorem4 ). As in the preceding theorem, suppose that
{fk }∞ ∞
k
1 is an orthonormal sequence in an inner product space V . Then {fk }k
1 is
a complete orthonormal sequence if and only if for every f ∈ V ,


∞
|f, fk |2
f 2 .
k
1

4
Due to Marc-Antoine Parseval des Chênes (1755–1836; France).
84 4. Fourier Analysis in Hilbert Space

Proof. This is left as Exercise 4.2.1(b). 2

We end this section with a sort of converse to Bessel’s inequality. Theorem
4.2 implies, as a special case, that if f ∈ L2 , then the sum of squares of the
Fourier coefficients of f , with respect to the trigonometric system {fk }∞ k
1 , is
always finite. The combination of Theorems 4.2, 4.3, 4.4, and 4.6 sets up a “linear
isometry” between L2 ([−π, π], m) and 2 . Specifically, for f ∈ L2 , define Tf

{f, fk }∞ ∞
k
1 , where {fk }k
1 denotes the trigonometric system. Then Tf ∈ 
2

(Theorem 4.2), T is linear, one-to-one (Theorem 4.6, together with Theorem 4.5
(c)), onto (Theorems 4.4 and 4.6), and f L2
Tf 2 (Theorem 4.3), for all
f ∈ L2 . This result, that L2 and 2 are isometrically isomorphic, is referred to as
the Riesz–Fischer Theorem (Theorem 4.6 sometimes goes by the same name).
Theorem 4.4. Assume that

∞ 2
(a) {dk }∞
k
1 is a sequence of real numbers such that k
1 dk converges, and
(b) V is a Hilbert space with complete orthonormal sequence {fk }∞k
1 .

Then there is an element f ∈ V whose Fourier coefficients with respect to {fk }∞

k
1
are the numbers dk , and


∞
f 2
dk2 .
k
1

Proof. Define


n
sn
d k fk .
k
1

For m > n,


m

m

m
sn − sm 2
dj dk fj , fk 
dk2 .
j
n+1 k
n+1 k
n+1

Therefore, {sn }∞
n
1 is Cauchy. Because V is a Hilbert space, there is an f ∈ V such
that
lim sn − f 
0.
n→∞

This is what we mean when we write

∞
f
dk fk ,
k
1

and Theorem 4.1 now says that dk
f, fk .

The remaining identity now follows from Parseval’s theorem. 2
In order for the Riesz–Fischer theorem to be true, we would need to know
that the trigonometric system is, in fact, a complete orthonormal sequence in
L2 ([−π, π], m). This is the goal of the next section.
 4.3 A Return to Classical Fourier Analysis 85

4.3 A Return to Classical Fourier Analysis

We now return to the classical setting, and the orthonormal family
1 cos(nx) sin(mx)
√ , √ , √ , n, m
1, 2, . . . ,
2π π π
in L2 ([−π, π], m).
Our first theorem of this section is proved mainly for its use in the proof of
Theorem 4.6 (that is why it appears here and not in the preceding section), but
it is interesting in its own right. Parseval’s theorem gives an alternative way to
think about “completeness” of an orthonormal family; this theorem gives a few
more ways. We state it only for orthonormal families in the specific Hilbert space
L2 ([−π, π], m); the result can be generalized to arbitrary Hilbert spaces.
Theorem 4.5. For an orthonormal sequence {fk }∞ 2
k
1 in L ([−π, π], m), the
following are equivalent:
(a) {fk }∞
k
1 is a complete orthonormal sequence.
(b) For every f ∈ L2 and  > 0 there is a finite linear combination


n
g
dk fk
k
1

such that f − g2 ≤ .

(c) If the Fourier coefficients with respect to {fk }∞ 2
k
1 of a function in L are all
0, then the function is equal to 0 almost everywhere.
Proof. It should be clear from the definition that (a) implies (b).
To prove that (b) implies (c), let f be a square integrable function such that
f, fk 
0 for all k. Let  > 0 be given and choose g as in (b). Then
 
n 
 
f 22
f 22 − f, dk fk 
|f, f − g|
k
1
≤ f 2 · f − g2 ≤ f 2 .
This implies that f 2 ≤ . Since  was arbitrary, f must be 0 almost everywhere.
To prove that (c) implies (a), let f ∈ L2 and put


n
sn
f, fk fk .
k
1

As in the proof of Theorem 4.4, we see that {sn }∞ 2

n
1 is Cauchy in L . And Theorem
3.21 then tells us that there is a function g ∈ L such that
2

lim sn − g2
0.

n→∞

That is,


∞
g
f, fk fk .
k
1
86 4. Fourier Analysis in Hilbert Space

Theorem 4.1 then tells us that the Fourier coefficients of g are the same as the
Fourier coefficients of f with respect to {fk }∞
k
1 , i.e., g, fk 
f, fk . By (b),
f − g must equal 0 almost everywhere. In other words,


∞
f
f, fk fk .
k
1

Since f was arbitrary, {fk }∞

k
1 is a complete orthonormal sequence. 2
Theorem 4.6. The trigonometric system
1 cos(nx) sin(mx)
√ , √ , √ , n, m
1, 2, . . . ,
2π π π
forms a complete orthonormal sequence in L2 ([−π, π], m). That is, if f is such
that |f |2 is Lebesgue integrable, then its (classical) Fourier series converges to f .
The convergence is convergence in the norm  · 2 , i.e.,
 π  a
n
n 2 
0
lim f (x) − + ak cos(kx) + bk sin(kx) dx
0.
n→∞ −π 2 k
1 k
1

This type of convergence is often called “in mean” convergence.

Proof. In Exercise 4.1.1 you are asked to prove that the trigonometric system is
orthonormal. We complete the proof of the theorem by verifying that condition (c)
of Theorem 4.5 holds. First consider the case that f is continuous and real-valued,
and that f, fk 
0 for each fk . If f
0, we then know that there exists an x0
at which |f | achieves a maximum, and we may assume that f (x0 ) > 0. Let δ be
small enough to ensure that f (x) > f (x2 0 ) for all x in the interval (x0 − δ, x0 + δ).
Consider the function
t(x)
1 + cos(x0 − x) − cos(δ).
This function is a finite linear combination of functions in the trigonometric system;
such functions are called “trigonometric polynomials.” It is straightforward to
verify
(i) 1 < t(x), for all x in (x0 − δ, x0 + δ), and
(ii) |t(x)| ≤ 1 for all x outside of (x0 − δ, x0 + δ).
Since f is orthogonal to every member of the trigonometric system, f is
orthogonal to every trigonometric polynomial and, in particular, is orthogonal
to t n for every positive integer n. This will lead us to a contradiction. Notice
that
 π
0
f, t 

n
f (x)t n (x)dx
−π
 x0 −δ  x0 +δ  π

f (x)t (x)dx +
n
f (x)t (x)dx +
n
f (x)t n (x)dx.
−π x0 −δ x0 +δ

By (ii) above, the first and third integrals are bounded in absolute value for each n by
b
2πf (x0 ). The middle integral, however, is greater than or equal to a f (x)t n (x)dx,
 4.3 A Return to Classical Fourier Analysis 87

where [a, b] is any closed interval in (x0 − δ, x0 + δ). Since t is continuous on

[a, b], we know that t achieves a minimum value, m, there. By (i) above, m > 1.
Then
 b
f (x0 )
f (x)t n (x)dx ≥ · mn · (b − a),
a 2
which grows without bound as n → ∞. This contradicts the assumption that
0
f, t n  for all n. Thus, any continuous real-valued function that is orthogonal
to every trigonometric polynomial must be identically 0.
If f is continuous but not real-valued, our hypothesis implies that
 π
f (x)e−ikx dx
0, k
0, ±1, ±2, . . . ,
−π

and thus also that

 π
f (x)e−ikx dx
0, k
0, ±1, ±2, . . . .
−π

If we add and subtract these two equations, we see that the real and imaginary
parts of f are orthogonal to each of the members of the trigonometric system. By
the first part of the proof, the real and imaginary parts of f are identically 0; hence
f is identically 0.
Finally, we no longer assume that f is continuous. Define the continuous
function
 x
F (x)
f (t)dt.
−π

For now let us assume that fk (x)
cos(kx)

√ .
Our hypothesis implies
π
 π
0
f (x) cos(kx)dx.
−π

Integration by parts yields

 π 
1 π
F (x) sin(kx)dx
f (x) cos(kx)dx
0.
−π k −π
Similarly, we can show that
 π
F (x) cos(kx)dx
0.
−π

We now have shown that F , and hence F − C for every constant C, is orthogonal
to each of the nonconstant members of the trigonometric system. We now take
care of the member √12π . Let
 π
1
C0
F (x)dx.
2π −π
Then F − C0 is easily seen to be orthogonal to every member of the trigonometric
system. Since F is continuous, F − C0 is also continuous, and the first part of the
88 4. Fourier Analysis in Hilbert Space

proof shows that F − C0 is identically 0. From this it follows that f
F  is 0

almost everywhere. 2
Is this theorem “good”? Note that  · 2 -convergence does not necessarily imply
either uniform or pointwise convergence (Exercise 4.1.4). With uniform conver-
gence, for example, we know that we cannot get the same result because the
partial sums of the Fourier series of f are always continuous functions, and if
the convergence of the series were uniform, then f would have to be continuous,
too. Since L2 ([−π, π], m) contains discontinuous functions, we see that uniform
convergence cannot always be achieved.
Theorem 4.6 has an important corollary, which we state as our next theorem.
See Exercise 3.6.8 for an alternative proof of this same result.
Theorem 4.7. C([−π, π]) is dense in L2 .
Proof. This is immediate, since the trigonometric polynomials are each con-
tinuous, and Theorem 4.6 shows that the smaller set is dense (see Theorem
4.5(b)). 2
This theorem gives us an alternative way to define L2 . First, it is not hard to see
that the interval [−π, π] can be replaced by any other closed and bounded interval
[a, b]. One can define L2 ([a, b], m) as the completion of C([a, b]) with respect to
the norm  · 2 . The advantage of this definition is that it gives a way of discussing
the very important Hilbert space L2 without ever mentioning general measure and
integration theory. Specifically, we define L2 to be the collection of functions f
defined on the interval [a, b] such that
lim fn − f 2
0
n→∞

for some sequence {fn }∞n
1 ∈ C([a, b]). Actually, L must be considered to be
2

the equivalence classes of such functions, where two functions are equivalent if
and only if they are equal almost everywhere (see the discussion preceding The-
orem 3.17). Therefore, it is not entirely true that this definition avoids discussing
measure. However, we can give this definition with only an understanding of
“measure zero,” and not general measure. (And measure zero can be defined in a
straightforward manner and is much simpler to understand than general measure.)

Exercises for Chapter 4

Section 4.1
4.1.1 Show that the trigonometric system
1 cos(nx) sin(mx)
√ , √ , √ , n, m
1, 2, . . . ,
2π π π
is an orthonormal sequence in L2 ([−π, π], m).
 4.3 A Return to Classical Fourier Analysis 89

4.1.2 In this exercise you will actually compute a classical (i.e., with respect
to the orthonormal sequence of Exercise 1) Fourier series, and investigate
its convergence properties. The function given is a basic one; in the next
exercise you are asked to do the same procedure with another very basic
function. You are being asked to do these by hand, and you can no doubt
appreciate that the computations get quite laborious once we depart from
even the most basic functions. There are tricks for doing these computations;
the interested reader can learn more about such techniques in a text devoted
to classical Fourier series.
(a) Let

1 if −π ≤ x < 0,
f (x)

0 if 0 ≤ x < π .
Show that its Fourier series is
1 1
∞
((−1)n − 1)
+ sin(nx).
2 π n
1 n
(b) Explain why this series converges in mean to f .
(c) What can you say about the pointwise and uniform convergence of this
series?
(d) Why are the coefficients of the cosine terms all zero?
4.1.3 In this exercise you will compute another Fourier series and investigate its
convergence properties.
(a) Let f (x)
x 2 . Show that its classical Fourier series is
π2
∞
(−1)n
+4 cos(nx).
3 n
1
n2
(b) Explain why this series converges in mean to f .
(c) What can you say about the pointwise and uniform convergence of this
series?
(d) Why are the coefficients of the sine terms all zero?
4.1.4 For a sequence {fn }∞ 2
n
1 in L ([−π, π], m), we have seen three ways for
∞
{fn }n
1 to converge:
(i) “pointwise,”
(ii) “uniformly,”
(iii) “in mean.”
The point of this exercise is to understand the relation between these three
types of convergence. For the counterexamples asked for below, use what-
ever finite interval [a, b] you find convenient. Please make an effort to
supply “easy” examples.
(a) Prove that uniform convergence implies pointwise convergence.
90 4. Fourier Analysis in Hilbert Space

(b) Give an example to show that pointwise convergence does not imply
uniform convergence.
(c) Prove that uniform convergence implies convergence in mean.
(d) Give an example to show that pointwise convergence does not imply
convergence in mean.
(e) Give an example to show that convergence in mean does not im-
ply pointwise convergence. (Note that the same example shows that
convergence in mean does not imply uniform convergence.)
4.1.5 Apply the Gram–Schmidt process to the functions 1, x, x 2 , x 3 , . . . to obtain
formulas for the first three Legendre polynomials. Then verify that they are
indeed given by the formula

2n + 1 1 d n 2
(x − 1)n , n
1, 2, 3.
2 2n n! dx n
4.1.6 Prove that the Haar family is an orthonormal family in the Hilbert space
L2 ([0, 1]).
4.1.7 (a) Show that the sequence
einx
√ , n
0, ±1 ± 2, . . . ,
2π
is a complete orthonormal sequence in L2 ([−π, π]).
(b) Show that the sequence

2
cos(nx), n
1, 2, 3, . . . ,
π
is a complete orthonormal sequence
  in L2 ([0, π ]). (Observe that
2
π
cos(nx) can be replaced by π2 sin(nx).)

Section 4.2
4.2.1 (a) Prove that in any inner product space (V , ·, ·), f and g orthogonal
implies
f 2 + g2
f + g2 .
√
Here, as usual,  · 
·, ·.
(b) Prove Parseval’s theorem.
4.2.2 Assume that {fn }∞ n
1 is a sequence in L and that fn → f in mean. Prove
2
∞
that {fn 2 }n
1 is a bounded sequence of real numbers.
4.2.3 Assume that f1 , f2 , . . . , fn is an orthonormal family in an inner product
space. Prove that f1 , f2 , . . . , fn are linearly independent.
4.2.4 For f and g in an inner product space, g
0, the projection of f on g is
the vector
f, g
g.
g2
 4.3 A Return to Classical Fourier Analysis 91

Show that the two vectors

f, g f, g
g and f− g
g2 g2
are orthogonal.
4.2.5 (a) Show that the classical Fourier series of f (x)
x is


∞
(−1)n+1
2 sin(nx).
n
1
n
(b) Use your work in (a), together with Parseval’s identity, to obtain Euler’s
remarkable identity

∞
1 π2

.
n
1
n2 6

(Note: The same procedure can be applied to the Fourier series of x 2

to obtain

∞
1 π4
4

,
n
1
n 90
and so on!)
http://www.springer.com/978-0-387-95224-6