Beruflich Dokumente
Kultur Dokumente
Editors
J.E. Marsden
L. Sirovich
M. Golubitsky
W.Jger
Advisor
G. Iooss
P. Holmes
Fourier Analysis
and Applications
Filtering, Numerical Computation, Wavelets
Translated by R. Ryan
With 99 Illustrations
Springer
Claude Gasquett Patrick Witomski
Universite Joseph Fourier (Grenoble I) Directeur du Laboratoire LMC-
IMAG
Transtator Tour IRMA, BP 53
38041 Grenoble, Cedex 09
Robert Ryan
France
12, Blvd. Edgar Quinet
75014 Paris
France
Series Editors
J.E. Marsden L. Sirovich
Control and Dynamical Systems, 107-81 Division of Applied Mathematics
California Institute of Technology Brown University
Pasadena, CA 91125 Providence, RI 02912
USA USA
M. Golubitsky W.Jger
Department of Mathematics Department of
University of Houston Applied Mathematics
Houston, TX 77204-34 76 Universitt Heidelberg
USA Im Neuenheimer Feld 294
69120 Heidelberg
tDeceased.
Germany
9 8 7 6 5 4 3 2 1
SPIN 10658148
Translator's Preface
This book combines material from two sources: Analyse de Fourier et ap-
plications: Filtrage, Calcul numerique, Ondelettes by Claude Gasquet and
Patrick Witomski (Masson, Paris, second printing, 1995) and Analyse de
Fourier et applications: Exercices corriges by Robert Delmasso and Patrick
Witomski (Masson, Paris, 1996). The translation of the first book forms
the core of this Springer edition; to this have been added ail of the exercises
from the second book. The exercises appear at the end of the lessons to
which they apply. The solutions to the exercises were not includcd bccause
of space constraints.
Whcn Springer offered me the opportunity to translate the book by Gas-
quet and Witomski, I readily acccpted bccause I liked both the book's
content and its style. I particularly liked the structure in 42 lessons and
12 chapters, and I agree with the authors that each lesson is a "chew-
able piece," which can be assimilated relatively easily. Believing that the
structure is important, I have maintained as much as possible the "look and
feel" of the original French book, including the page format and numbering
system. I believe that this page structure facilitates study, understanding,
and assimilation. With regard to content, again I agree with the authors:
Mathematics students who have worked through the material will be weil
preparcd to pursue work in many directions and to explore the proofs of
results that have been assumed, such as the development of measure theory
and the representation theorems for distributions. Physics and engineering
students, who perhaps have a different outlook and motivation, will be weil
equipped to manipulate Fourier transforms and distributions correctly and
to apply correctly results such as the Poisson summation formula.
Translating is perhaps the closest scrutiny a book receives. The process
of working through the mathematics and checking in-text referenccs always
uncovers typos, and a number of these have been corrected. On the other
hand, I have surely introduced a few. I have also added material: I have
occasionaily added details to a proof wherc I felt a few more words of
explanation wcre appropriate. In the case of Proposition 31.1.3 (which is
vi Translator's Preface
a key result), Exercise 31.12 was added to complctc thc proof. I havc also
completed the proofs in Lesson 42 and added some comments. Sevcral new
referenccs on wavelets have been included in thc bibliography, a few of
thcm with annotations. All of these modifications have bcen madc with
the knowledge and concurrence of Patrick Witomski.
Although the book was written as a textbook, it is also a useful refcrence
book for theorctical and practical results on Fourier transforms and distri-
butions. Therc arc several places where the Fourier transforms of specific
functions and distributions are summarized, and therc are also summaries
of general results. These summaries havc been indexed for easy refcrence.
The French edition was typeset in Plain TEX and printed by Louis-Jean
in Gap, France. Monsieur Albert at Louis-Jean kindly sent me a copy of
the TEX source for the French cdition, thus allowing many of the equations
and arrays to bc copied. This simplified the typcsetting and helped to avoid
introducing errors. My sincere thanks to M. Albert. Similarly, thanks go
to Anastis Antoniadis (IMAG, Grenoble) for providing the lb-TEX sourcc
for the cxercises, which was clegantly prcpared by his wife. I had the good
fortune to have had the work edited by David Kramcr, a mathcmatician
and freclance editor. He not only did a masterful job of straightening out
the punctuation and othcr language-based lapses, but he also added many
typesetting suggestions, which, I believe, manifestly improved the appear-
ance of the book. I also thank David for catching a few of the typos that I
introduced; those that remain are my responsibility and embarrassmcnt.
Robcrt Ryan
Paris, July 14, 1998
Preface to the French Edition
This is a book of applied mathematics whose main topics are Fourier anal-
ysis, filtering, and signal processing.
The development proceeds from the mathematics to its applications,
whilc trying to make a connection betwecn the two perspectives. On one
hand, specialists in signal processing constantly use mathematical concepts,
often formally and with considerable intuition based on experience. On the
other hand, mathematicians place more priority on the rigorous develop-
ment of the mathcmatical conccpts and tools.
Our objective is to give mathematics students somc understanding of
the uses of the fundamental notions of analysis they are learning and to
providc thc physicists and engineers with a theoretical framework in which
the "wcll known" formulas are justified.
With this in mind, the book presents a development of the fundamentals
of analysis, numerical computation, and modeling at levels that extend
from the junior year through the first year of graduate school. One aim is
to stimulate students' interest in the coherence among the following three
domains:
Fourier analysis;
signal processing;
numerical computation.
On completion, students will have a general background that allows them
to pursuc more spccialized work in many directions.
A progressive approach
The program we have adopted is progressive; it is written on levels that
range from the third year of college through the first year of graduate
school.
JUNIOR LEVEL
Lessons 1 through 7 are accessible to third-year students. They intro-
duce, at a practical level, Fourier series and the basic ideas of filtering.
Here one finds some simple examples that will be re-examined and studied
in more depth later in the book. The Lebesgue integral is introduced for
convenience, but in superficial way. On the other hand, emphasis is placed
on the geometric aspects of mean quadratic approximation, in contrast to
the point of view of pointwise representation. The notion of frequency is
illustrated in Lesson 7 using musical scales.
SENIOR LEVEL
The reader will find a presentation and overview of the Lebesgue integral
in Chapter IV, where the objective is to master the practical use of the
integral. The lesson on measure theory has been simplified. This chapter,
howcver, serves as a good guide for a morc thorough study of measure and
integration. Chapter VI contains concentrated applications of integration
techniques that lead to the Fourier transform and convolution of functions.
One can also include at this level the algorithmic aspects of the discrete
Fourier transform via the fast Fourier transform (Chapter III), thc concepts
of filtcring and linear differential equations (Chapter VII), an easy version
of Shannon's theorem, and an introduction to distributions (Chapter VIII).
MASTER LEVEL
According to our experience, the rest of the book, which is a good half
of it, demands more maturity. Herc one finds precisc results about thc
fundamental relation -r;g = j g, the Young inequalities (Chaptcr VI),
and various aspects of Poisson's formula related to sampling (Chapter XI).
Finally, time-frequency analysis based on Gabor's transform and wavelet
analysis (Chapter XII) call upon all of the tools developed in the first cleven
chapters and lead to recent applications in signal processing.
out ofthin air, where one ignores all of the fundamental background for a
very short-term advantage.
Juniors
Chapters I, II, and III.
Seniors and Masters in Mathematics
Chapters IV, V, VI, VIII, and IX.
Seniors and Masters in Physics
Chapters VII, X, XI, and XII.
was established by Raoul Robert. His initiative in this subject, which was
not his area of research, has played a decisive role, and the current cxplo-
sion of numerical work based on wavelets shows that hisvisionwas correct.
Our thanks go equally to Pierrc Baras for thc numerous animated discus-
sions we have had. Their ideas and comments have been a valuable aid and
irreplaceable inspiration for us.
The sccond printing of this book is an opportunity to make several rc-
marks. We have chosen not to include any new developments. We havc
listed at the end of the book several references on wavelets, which show
that this area has exploded during these last years. But for the student or
the teacher to whom we address the book, the path to follow remains the
same, and the basics must be even more solidly established to understand
these new areas of applications. It seems to us that our original objective
continues to be appropriate today.
We have made the necessary corrections to the original text, and a book
of exercises with solutions will soon be available to complete the project.
Claude Gasquet
Patrick Witomski
Grenoble, June 30, 1994
Contents
Translator's Preface V
References 433
Index 437
Chapter I
Signalsand Systems
Lesson 1
x(t)
-2 -1 0 2 3
1\
x(t) y(t)
Transmission system
input output
signal signal
When there are several input or output signals, the functions x(t) and
y(t) are vectors. We willlimit our discussion to the scalar case, where there
is a single input signal and a single output signal.
In signal theory, one is not necessarily interested in the system's compo-
nents, but rather in the way it transforms the input signal into the output
signal. It is a "black box." It will be modeled by an operator acting on
functions, and we write
y=Ax,
where x E X, the sct of input signals, and y E Y, the set of output signals.
Examples of systems:
An electric circuit
An amplifier
The telephone
The Internet
1.1 General considerations 5
One distinguishes:
Analogsystems that transform an analog signal into another analog sig-
nal (Figure 1.3)
Discrete systems that transform a discrete signal into another discrete
signal (Figure 1.4)
T T
-3 -2 -1 0 1 2 3 n -3 -2 -1 0 1 2 3 n
FIGURE 1.4. Discrete system.
-1 0 2 3 -1 0 1 2 3
FIGURE 1.5. The clamper.
6 Lesson 1. Signals and Systems
{
1 if ltl < a,
r(t) = 0 if ltl > a.
r(t)
-a 0 a
Signal values are, in principle, real numbers, and the frequency is a pos-
itive number. However, for reasons of convenience (Fresnel representation,
derivation, multiplication, ... ) a complex-valued function
z(t) = a:ei(wt+<p)
L
00
z(t) = Cne2i1rAnt.
n=-cx::>
1.3.3 Differentiator
FIGURE 1.8.
1.3.5 An RC circuit
\
7)
x(t)
L.,__ _ _ _ _ _ _ T....____c__
FIGURE 1.9. An RC circuit.
The input to the circuit shown in Figure 1.9 is the voltage x(t); the
output is the voltage v(t) across the capacitor. Thus,
A: X f---> V.
Y(t)
x(t)
k m
z)zzzz;zzzzJzzztzz~zzzzzzzizzzz~zzzft;:;,zzz
FIGURE 1.10.
R R R R R
R' R'
FIGURE 1.11.
Consider the electrical circuit in Figure 1.11. The input is the constant
voltage E, and the currents i 0, i 1, ... , iN constitute the output:
A: E t-+ (io,il, ... ,iN)
10 Lesson 1. Signals and Systems
A B
D c
FIGURE 1.12.
On the other hand, by summing the voltage drops around thc loop
DABC we have (see Figure 1.12)
E = (R + R')io- R'ii!
0 = R'ik-1- (2R' + R)ik + R'ik+l,
0 = iN+1
This is a second-order linear recursion system with boundary conditions.
Lesson 2
2.1.1 Linearity
Consider the system
A:X-+Y.
A is said to be linear if
and
A(..\x) = ..\A(x)
for all x, u EX and all ,\ER (or C if Xis complex). This is also called the
principle of superposition. The systems in Section 1.3 are alllinear, which
is easily verified by examining the governing equations (1.1) through (1.4).
2.1.2 Causality
A is said to be realizable (or causal) if the equality of any two input signals
up to time t = to implies the equality of the two output signals at least to
time to:
12 Lesson 2. Filters and Transfer Functions
2.1.3 lnvariance
A is said to be invariant, or stationary, if a translation in time of the input
leads to the same translation of the output; that is,
which says that A commutes with all translations. For discrete systems,
one considers only a that are multiples of the sampling interval.
(iii) The norm for convergence "in energy" (mean quadratic convergence):
is also used. By this we mean that the sequence Xn tends to x if the limit
exists for each of the components:
y(t) = [ e-(t-s)x(s) ds
00
It is thus sufficient to know the outputs for each of the inputs (e~)nEZ to
know thc image of an arbitrary periodic signal. Furthermore, it is easy to
2.4 A standard analog filter: the RC cell 15
determine the image f>. of the signal e.x, assuming that the latter belongs
to the space of input signals for the filter. Indeed, for all values oft and u,
which we write as
2.4.1 Systemresponse
t
Writing the unknown function v as v(t) = w(t)e-RC reduces the initial
equation
RCv'(t) + v(t) = x(t)
to
1 t
w'(t) = RCeRCx(t).
16 Lesson 2. Filtersand Transfer Functions
Assuming that the input signal x(t) is such that the second member is
integrable on every interval ( -oo, t), we have
1
w(t) = RC
1t s -oo eRCx(s)ds+K
and
v(t)
1
= RC
1t t-s
-oo e- RCx(s)ds+Ke- RC.
t
The constant K is determined by an auxiliary condition. For example, if
we assume that the response to the zero input is zero, we see that K = 0.
One can define the response of the system A to the input x to be
v(t) = Ax(t) = RC
1 1t t-s
-oo e- RC x(s) ds. (2.1)
It is clear from this expression that Ais linear, realizable, and invariant. It
is also continuous, for example, in the uniform norm, since
and thus
IIAxlloo :S llxlloo
This shows that thc RC cell is a filter.
Ax(t) 1 +oo
= -oo h(t- s)x(s) ds = (h * x)(t). (2.2)
This operation is, by definition, the convolution of the two signals h and
x. It is denoted by h * x, and we havc
Ax = h*x.
In this situation, one is said to have a convolution system. The function h,
called the impulse response of the system, characterizes the filter because
knowing h implies that the output of the filter is known for any input x.
Throughout the book, we will use h to denote the impulse response of a
system. A companion notion, the response of a system to the unit step
function u(t), will be defined in Lesson 24.
2.4 A standard analog filter: the RC cell 17
and we have
H(.>..) = 1 + 2i~.ARC
We see that signals for which I.AI is small, the low-frequency signals, are
transmitted by the filter almost as if it were the identity mapping (see
Figure 2.1). On the other hand, the high-frequency signals, for which I.AI
is large, are almost completely attenuated. This explains why this filter is
called a low-pass filter. The action of the filter on different frequencies is
clearly apparent from the graph of the function
which is called the energy spectrum of the filter. The function IH(.A)I is
called the spectral amplitude.
0 ,\ _ _1_
c- 21rRC
(2.3)
The analysis of this system follows that of the last example. Wc try to
express the output explicitly as a function of the input. Thus we change
the unknown by letting Yk = akvk, which transforms (2.3) into
(2.4)
and
k
Yk = bak + L ak-nXn
n=-oo
if n ~ 0,
if n < 0,
which is called the discrete convolution of the two signals x = (xn) and
h = (hn) This system is linear and invariant. One can easily verify that
the condition hn = 0 if n < 0 implies that it is realizable. The system is
continuous in the uniform norm whenever Iai < 1, and we have
The signal h is called the impulse response of the system. It is the response
to a unit impulse at time t = 0 defined by
where ek = { ~ if k
if k
=
=/:
0,
0.
Periodic Signals
Lesson 3
Trigonometrie Signals
We have seen that the pure sinusoidal signals are eigenfunctions for all
filters. They are also the simplest periodic signals. These two facts explain
their importance. We will see in the next two lessons that they enter into
the structure of all periodic signals.
f(t + a) = f(t).
2" t
(Note that a is not necessarily the smallest period.) Since en(t) = e mna
has period a for each integer n, the same is true for functions p of the form
where I is any fixed, finite set of integers and the Cn are arbitrary complex
numbers. By adding zero terms if necessary, we may assume that
+N
p(t) = I: (3.1)
n=-N
p(t)
-a 0 a 2a
FIGURE 3.1.
n=l
where, for n ~ 0,
an= Cn + C-n,
(3.3)
bn = i(cn- C-n)
3.3 Orthogonality
A simple computation shows that the following important relation holds
for the functions en(t):
(3.5)
3.3 Orthogonality 25
thc relation (3.5) expresses the fact that the functions en, E Z, are or- n
thogonal:
(en,em) = 0 if n "Im, and llenll2 = ya.
It follows that the vectors en are independent and that the dimension of
TN is exactly 2N + 1. If p is of the form (3.1), we have
(p,en) = Cnllenll~ = acn,
11a
and
an=
2r
~ Jo
t
p(t) cos ( 27rn~) dt,
bn
2r
= ~ Jo
t
p(t) sin ( 2nn~) dt.
(3.7)
and this, combined with (3.5), yields Parseval's equality for trigonometric
polynomials:
L N 1
lcnl 2 = - Jo
r jp(t)j 2 dt. (3.8)
m=-N a 0
26 Lesson 3. Trigonometrie Signals
3.4 Exercises
Exercise 3 .1 If f : R --> R is a periodic function with period a, integrable
J:+a
on bounded intervals, show that the integral f(t) dt does not depend on x.
lesson 4
(4.1)
The immediate answer is "no" if one considers only finite sums. The sum
on the right is infinitely differentiable, while there is no reason for f to be.
For example, f could be the periodic window function in Figure 4.1.
f(t)
-a -h 0 h a t
exists and is finite. We will not dwell on issues of integration at this point;
they will be discussed in Lessons 11 through 15. We introduce the notation
llfii2=JU,f)=( 1 a
o lf(tWdt)
1/2
.
It is important to note that the norm of f can be zero even though the
function f E L~(O, a) is not zero at every point. For example, f could be
zero at all but a finite nurober of points. Thus, to have a true norm, it is
necessary to identify such a function with the function that is identically
zero, which we sometimes call the null function. Generally, we must identify
any two functions fand g for which f 0a lf(t)-g(t)l dt = 0 (see Section 13.3).
In this case, we say that the two functions are equal almost everywhere,
and we write f = g a.e. At this point, it is suflicient to remernher that
Oe
L
FIGURE 4.2. Orthogonal projection on a subspace.
From (4.2) it is perfectly clear that the minimum is attained when Xn = Cn,
and only for this value.
In summary, the bcst approximation fN exists and is unique, and it is
given by
N
fN(t) = L Cnen(t).
n=-N
!N(t) = L (4.3)
n=-N
whcrc
Cn = -
a o
11a
f(t)e
-2i1Tn!
a dt. (4.4)
L
N
lcnl
1
2 :::; - Jo
r lf(tW dt, NEN,
n=-N a o
n=-oo
1.2268
0.7268
0.2268
-2732 X
-0.7732
-1 .2732 -1---.-~--..--.-->-4--.-~--..-~.--.,--"-,-"--,
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 11.0
10.0 12.0
FIGURE 4.3. fl(t) = ~sint.
0.7996
0.2996
-Q.2004 0 X
-Q.7004
-1.2004+-....--.---.---lo.f--~'-.--.---..-~-+~--.-.>.<;
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 11.0
10.0 12.0
0.8126
0.3126
-o.6874
-1.1874 +--.---.---.--"-r-..,~c..-...,....---.--.-.....J./.-.--l,/
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 11.0
10.0 12.0
FIGURE 4.5. j5(t) = ~(sint + i sin3t + i sin5t).
The proof of this theorem requires more background than is available
in these early lessons. We will prove it in Lesson 16 as an illustration of
results about Lebesgue integration.
It is this theorem that gives meaning to the formula
+oo t
!( t ) = """'
L....J Cne
2i7rn-
a, (4.6)
n=-oo
is equal to the sum of the series on the right. In the last example, the sum
ofthe series is zerofort = 1r, whereas j(1r) = -1. Similarly, if f is modified
at a point t 0 , the Fourier series is unchanged for t = t 0 . This touches on
the problern of pointwise representation that will be studied in the next
lesson.
(b) A more scholarly way to express Theorem 4.3.1 is to say that the
family of functions (en)nEZ is a topological basis for the space L~(O, a) and
that the series
+oo
L
1
lcnl 2 = -
r lf(t)i dt.
Jo 2 (4.8)
n=-oo a 0
(4.9)
4.5 Formulary
+oo t
!( t ) = ""'
L....t Cne 2i1fn-a
n=-oo
Cn =-11a f(t)e
a o
-2i1fnl
a dt
4.6 Exercises
Exercise 4.1 Calculate the Fourier series expansions of the following func-
tions and verify the symmetric properties of the coefficients:
(a) f has period 2 and f(t) = ltl if ltl < 1.
(b) f has period a and f(t) = ! if 0 :S t < a.
a
(c) f(t) =I sintl.
(d) f(t) = sin3 t.
Exercise 4.2 If the Fourier coefficients of f(t) are cn, what are the Fourier
coefficients of the translated function f(t - to)? Deduce from Exercise 4.1 the
Fourier series expansion of f (t) = Icos t I
Exercise 4.3 Prove relation (4.9). (Hint: Use (4.8) and (3.4).)
Exercise 4.4 Write Parseval's equality for each function in Exercise 4.1.
Exercise 4.5 Assurne that f E L~(O, a) and Cn are its Fourier coefficients.
Then f is also in L~(O, 2a) with Fourier coefficients c~. How are the coefficients
Cn and c~ related? Verify that the two Fourier series are identical.
36 Lesson 4. Periodic Signals and Fourier Series
Exercise 4.6 Find the Fourier series expansion ofthe function f with period
a = 2 defined on [-1,+1) forzE C\Z by
(H1) Cn(f) = 1 1
j(t)e- 2 irrnt dt = 0, n E Z,
(H2) 1: 1
2
j(t)p(t) dt = 0
Define
L L
N-1 k
C!N(t) =~ e2irrnt
k=O n=-k
1: 1
2
C!N(t) dt = 1.
1
(b) Show that
lim CJN(t)dt=O.
N~oo a<itl<!
- -2
4.6 Exercises 37
I:
(c) Deduce from this that
[~ 2 UN(t)j(t) dt > 0
2
for sufficiently large N and deduce the result.
Lesson 5
Pointwise Representation
This is false for the Riemann integral; take, for example, the function equal
to 1 on the rationals and -1 on the irrationals.
It follows immediately that the Fourier coefficients of a periodic function
Cn(/) = ~
a lo
r j(t)e - 21rn~ dt
exist if and only if f is integrable on (0, a). We introduce the notation
1
IInl :::; 21rlnl (lf(a)l + lf(b)l + lb
a lf'(x)l dx).
The right hand-side, and thus In, tends to 0 as when lnl--> +oo.
We now use a density argument that is based on the following provisional
assumption: The functions that are continuously differentiable on [a, b] are
densein L~(a, b). This means that given c: > 0, there exists a 9c: E C 1 ([a, b])
l
suchthat
b c
a lf(x)- 9c:(x)l dx:::; 2'
which implies that
From the previous argument, there exists N > 0 such that the last integral
is dominated by t:/2 if lnl ~ N. Thus, lnl ~ N implies IInl :::; c:, and this
proves the theorem. o
convergence of /N at a given point. For this, one needs more refined assump-
tions about the function. In practice, these hypotheses will generally hold.
However, we emphasize that theseadditional hypotheses are essential, since
several "natural" results that one might expect to hold for f E L~(O, a) are
indeed false. We cite three of these:
(i) !N -t f in the L~(O, a) norm.
(ii) !N(t) - t f(t) for almost all t.
(iii) If f is also continuous on IR, then !N(t) f(t) for all t E IR.
-t
It has even been shown (see [KF74]) that there exists an f in L~(O, a) such
that !N(t) diverges for all t as N-t +oo! These examples represent difficult
problems that have played a central role during the last century in research
on the theory of functions.
At the end points a and b we require that only the one-sided limits exist. We
denote this function space by Cpw[a, b], where "pw" stands for "piecewise."
Then
f is bounded on [a, b],
f E Cpw[a, b] ===? {
f is integrable on [a, b].
The integral of f involves only the integration of continuous functions, since
{b n {ak+I
Ja f(t) dt ={;Jak J(t) dt,
where ao = a, an+ 1 = b, and where a1, ... , an are the points of discontinu-
ity of f in (a, b). Each integral in the sum is understood tobe the integral
of the continuous extension of f to the interval [ak, ak+t] Note that as far
as the integral is concerned, f does not need to be defined at the points ak.
EXAMPLE: Define f on [-1, 1) by
- 1::; t < 0,
f( t) = { t 2 + 1 if
-t +2 if 0 ::; t ::; 1.
42 Lesson 5. Pointwise Representation
- 1::; t < 0,
f '(t) = { 2t if
-1 if 0 < t::; 1,
The implication from right to left is immediate. For the other direction,
see, for example, [KF74]. We deduce from this last cquivalence that
fN(to) = ~
1~~a (
LN
e
221 to-x) f(x) dx.
rn-a-
-2 n=-N
t
n=-N
e2i1rnt = sin 1r(~N + 1)t.
sm 1rt
(5.3)
5.2 Pointwise convergence? 43
Thus we obtain
From (4.3) and (4.4), we see that f = 1 implies fN = 1 for all N. Thus, for
all N::::: 0,
11~ sin7r(2N + 1)~ 1
- adx--
a o sin1r~ - 2
a
!N(to)- Yo =
1 ~~ sin7r(2N + 1)~
- Jo [f(to+x)-f(to+)+f(to-x)-f(to-)] . x a dx.
a 0 Sill 1ra (5.4)
tend to finite limits as X ......... o-. Hence, the same is true for
<p(x) = i(to + x)- i(to+ ~ + {(to- x)- i(to-).
Slll1f-
a
Thus, there exist a > 0 and M > 0 suchthat l<p(x) I :S: M for all x E (0, a].
Since i E L~(O, a), <p is integrable on [a, a/2], and
This implies that <p is integrable on (0, a/2), and by the Riemann-Lebesguc
theorem (Theorem 5.1.1), the integral
X)
iN(to)- Yo = -;;;1 Jo{~ <p(x) sm (2N + 1)-;;: dx
. (
REMARK: The fact that iN(t 0) has a limit for each fixed t 0 as N ---> +oo
does not imply that either of the series
+oo t 0
LCne 2i1rn-a
'"'
n=O
or 2:
n=-oo
is convergent. On the other hand, this does imply the convergence of the
series (4. 7) of sines and cosines because it is obtained by symmetrically
regrouping the terms of iN
5.3 Uniform convergence of Fourier series 45
n=-oo
1
Cn(f) = -.-
2z1rn
1a f
0
1
(t)e -2i11"n!a dt,
since f(O+) = f(a- ), and hence the Fourier coeflicients cn(/1 ) of / 1 are
given by
2i1l"n
Cn(/) = -cn(f),
1
a
which proves (i). We deduce (ii) directly from the inequality
To prove (iii), note that E!:-oo len(f)l < +oo implies that (/N) is a
Cauchy sequence in the uniform norm on [0, a] (and hence on IR), so !N
converges uniformly on IR to some continuous function g. Since uniform
convergence implies convergence in the L~(O, a) norm, it follows from 4.3.1
and the uniqueness of the limit that f = g almost everywhere. But both f
and g are continuous, so f = g everywhere. o
The last part of this proof established the following corollary.
n=-oo
5.3.3 Conclusions
We use the terms "regular" and "smooth" informally to mean that a func-
tion has a number (undetermined) of derivatives. Thus the more regular,
or the smoother, a function f is, the faster the coefficients cn(J) tend to 0.
This can be seen by repeatedly integrating by parts. This is summarized
in the following display, where the regularity of f is increasing and where
c;(o, a] denotes the space of functions f E Ck(R) that are a-periodic.
The property
lnkcn(J)i- 0 for all k E N as lnl- +oo
will be abbreviated by the expression the sequence cn(J) is rapidly decreas-
ing or variationssuch as cn(J) decreases rapidly. Note that these expres-
sions, although widely used, are a slight abuse of the language: They in no
way imply that the sequence len(J)I is monotonic.
We now show that implication (e) is in fact an equivalence.
5.3.4 Proposition Assurne that f E L~(O, a). Then the following two
properties are equivalent.
(i) The Fourier coefflcients of f are rapidly decreasing.
(ii) The function f is infinitely differentiable.
Proof. Wehaveseen that (ii) =? (i). Conversely, ifthe cn(J) are rapidly de-
creasing, then in particular, n2 lcn(J)I - 0. Hence :L~:'-oo lcn(J)I < +oo,
and the sequence of functions
N
!N(t) = L
n=-N
5.4 Exercises
Exercise 5.1 Prove that a piecewise continuous function on [a, b] is bounded.
Note that a and b must be finite! Find a counterexamples if a = -oo or b = +oo.
f(x) ={ .
x sm
0
x1 l'f
x r_J. 0 ,
if X= 0.
Show that f is continuous on [0, 1] and differentiahte (0, 1] but that it is not of
bounded variation on [0, 1].
Exercise 5.5 Develop the Fourier series ofthe function J, with period a = 2,
defined on [-1, 1) by
f(t) = COS11'Zt, z E C\Z.
From this deduce the equalities
1 00
1
71'COt71'Z =-
z
+ 2z""' - - -,
L....., z 2 -n2
n=l
L
00
1 2i7rn:!?.
f(x) = -e a, 0 < x < a.
n
n=-oo
n,t:O
48 Lesson 5. Pointwise Representation
f(x) = ;o + Lancosnx.
n=l
Cn(f) {~n =
if n < 0,
if n 2: 0.
n!
1
(b) Deduce that
L
2rr 00 2n
e2x cos t dt = 211" (~!)2 .
0 n=O
(c) Define
Show that
Bo(x) = 1,
B~(x) = kBk-1(x) and 1 1
Bk(x)dx = 0, k 2:1.
(a) Compute B1, B2, B3 and draw their graphs in the same coordinate system.
(b) Show that each of these graphs is mapped into itself by reflection about
one or two axes. Is this generally true for the polynomials Bk? Express this
property algebraically for Bk.
(c) Let fk be the function with period 1 that coincides with Bk on [0, 1). Show
that for k 2: 3,
fk E C (JR) and fk = kfk-1
1 I
(d) Justify the term-by-term differentiation of the series for fand show that
Exercise 5.11
(a) Show that if f E c;[o, a], then lcn(f)l :S ~.
n
(b) Show that f E C;'[O, a] implies limlnl~oo lnkcn(f)l = 0 for all k E N.
Exercise 5.12 Take f E L~(O,a) and Iet h be a sequence in L~(O,a) such
n=-N
gN(t) = L N t
Cn(g)e2irrn;;:
n=-N
Show that
N
Cn(fNgN) = L Cn-k(f)ck(g).
k=-N
(c) Prove that f NgN tends to fg in L~(O, a) and use Exercise 5.12 to show
that
L
00
Cn(fg) = Cn-k(f)ck(g), n E Z,
k=-oo
f(t)
a 0 b
FIGURE 6.1.
f(t)
/"-----
f(t)
a 3a 4a
f(t)
FIGURE 6.4.
with
and
II!- /NII2 = min II/- Pll2o
pEVN
If, for all f E L 2 (a, b), fN tends to f in the norm of L 2 (a, b) as N tends
to infinity, the family {cl>n }nEN is said to be a topological basis for L 2 ( a, b) o
One also says that this family is a complete system or a total family (see
Section l6o3)o In this case, we have the Parsevalrelation
One can also use functions defined on an unbounded interval, for example
(a, b) = IR or (a, b) = IR+, where IR+ = {x E IR I x ?: O}o A problern in
these cases is that the polynomials are no Ionger in L 2 (a, b)o This can be
solved, however, by multiplying the polynomials by an appropriate weight-
ing function that tends to 0 sufficiently fast at infinityo One then obtains a
family of functions in L 2 (a, b) that can be used to expand f E L 2 (a, b)o
54 Lesson 6. Expanding a Function in an Orthogonal Basis
6.2.2 Examples
(a) LEGENDRE POLYNOMIALS
Take (a, b) = (-1, 1) and consider only real-valued functions. An orthog-
onal basis is obtained by "orthogonalizing" the set {1, t, t 2 , ... } of linearly
independent polynomials with respect to the inner product
1
(f,g) = /_ 1 f(t)g(t)dt.
In general,
Pn(t) = -,1-ddn (t2- 1)n.
n.2n tn
These polynomials are orthogonal, and
F(x) = f(cosx),
which is even and 271"-periodic. The scalar product for L 2 (0, 7r) is
(F G) =
'
11 v'f=t2
-1
f(t)g(t) dt.
(! ) =
,g w
11 v'f=t2
-1
f(t)g(t) dt
6.2 Expansion of a function in an orthogonal basis 55
(f,g)H = l +oo
-oo j(t)g(t)e-t 2 dt,
</>n(t) = Hn(t)e - t /2 ,
2
6. 3 Exercises
Exercise 6.1
(a) Expand the functions in Figures 6.2-6.4 in Fourier series.
(b) Determine the rates at which their Fourier coefficients Cn converge to 0.
(c) Write Parseval's equality for these three cases.
(d) Express the pointwise convergence in the three cases fort= a/2 and t = a.
In which cases do (c) and (d) produce interesting identities?
Exercise 6. 2
(a) Use the Legendre polynomials P0 , H, P 2 , ?3 to compute the best approx-
imations J;, i = 0, 1, 2, 3, to J(t) = ltl on [-1, + 1] in the sense of the usual
L 2 ( -1, 1) norm.
(b) Represent J, JI, h on the same graph.
(d) Compute 1 1
J(x) dx and deduce that
= 1 ~2
~ (-1t = ~-
-
L.,_; (2n + 1) 2 -- 8
"'""' and
L.,_; 2n + 1 4
n=O n=O
(g) Expand f in a series of cosines and address the same questions as above.
lesson 7
+=
J(t) = I: (7.1)
n=-(X)
then the spectrum of f is defined tobe the set of pairs (nja, cn), n E Z.
amplitude
phase
o,
o_,
FIGURE 7.2. Phase spectrum (of a real signal).
n=-oo
The positions of the spectrallines are not changed; only their relative values
change as they are multiplied by H(nja). Thus, this process is properly
called frequency filtering.
Electromagnetic waves:
very long: (telegraph) 1.5 104 to 6 10 4 Hz
long: (radio) 6 104 to 3 10 5 Hz
medium: (radio) 3 105 to 3 106 Hz
short: (radio) 3 106 to 3 107 Hz
meters: (tclevision) 3 107 to 3 108 Hz
centimeters: (radar) 3 108 to 10 11 Hz
visible light: 3.7 10 14 to 7.5 10 14 Hz
The human ear can, in the best cases, detect sounds whose frequencies
range from 20 to 20,000 Hz.
27.5 55 880 .. .
A-1 Ao A4 .. .
f 2f 3f 4f 5f 6f
c c G c E G
60 Lesson 7. Frequencies, Spectra, and Scales
Thus the common cord (C, G, E) is found within the C-scale. Bringing
these notes back to the same octave, we have the cord (C, E, G, C):
f ~! ~! 2/
c E G c
If we similarly analyze a G, we find the cord (G, D, B):
~! 3f V 6f 125 f
G G D G B
f
C
~!
D
~!
E
uG ~5 f
B
2f
C
Starting with E leads to no new note, at least in the first three harmonics:
~! ~! 1f
1 5f
E E B E
The frequencies of all of these notes are simple fractions of the fundamental
frequency f. Furthermore, all of the denominators are powers of 2. The first
simple fraction with denominator 3 leads to the discovery of the F with
frequency ~f, from which the cord (F, C, A) is constructed:
~! ~! 4f 136 f 2~ f
F F c F A
f
c
uG ~!
D
different lengths: the major interval, the minor interval, and the semitone
or half-step. The following table illustrates this idea:
It is easier to see the nuances between the harmonic scale (H.S.) and the
tempered scale (T.S.) by looking at the decimal values:
c D E F G A B c
H.S. f 1.125! 1.250! 1.333! 1.5! 1.667! 1.875! 2f
T.S. f 1.122! 1.260! 1.335! 1.498! 1.682! 1.888f 2f
The tempered scale has no need for the comma. It is this scale that is,
in principle, used for the piano.
62 Lesson 7. Frequencies, Spectra, and Scales
7. 3 Exercises
Exercise 7.1 The Pythagorean scale is built on the "fifth," which is defined
by two vibrating strings whose lengths are in the ratio 3 to 2:
F c G
~~ f ~~
(a) Bring these frequencies back to the same octave and compare this scale
with those described in Beetion 7.2.2 and 7.2.4.
(b) Note that this scale has only one major tone T and a very "tight" half-tone
m. Describe the succession of tones and half-tones.
Exercise 7.2 Let f be a real periodic signal. Show that its amplitude spec-
trum is even. Investigate the properties of its phase spectrum.
Chapter 111
Cn
11a
= -
a o
f(t)e -2i7rn!a dto (801)
FIRST METHOD: Integrating (801) by the trapezoid formula gives the ap-
proximate value
N-1 k
1 1 """' -2i7rn-
Cn = N L..... Yke N '
k=O
or
N-1
1 1 """' -nk with (8o2)
cn = N L..... YkWN
k=O
66 Lesson 8. The Discrete Fourier Transform
SECOND METHOD: One can also compute the Fourier coeffi.cients, denoted
by c!;[, of the trigonometric polynomial
N
2-1
p(t) = ""'
~ cnN e2i?rn!a (8.3)
N
n=-2
k = 0, 1, 2, ... , N- 1. (8.4)
For convenience, we bring all of the indices n into the interval [0, N -1) by
translating thc negative indices to the right by N. This is possible because
the functions involved are N-periodic. Thus,
By defining
N
if O<n<--1
- - 2 '
N
if -<n<N-1
2 - - '
the system (8.4) can be written as
N-1
L Ynw'Nk = Yk, k = 0, 1, 2, ... , N- 1.
n=O
This system has the advantage that it can be solved explicitly. Let p be
an integer between 0 and N - 1 and compute the sum
N-1 N-1N-1 N-1 N-1
""' -kp _ ""' ""' v k(n-p) _ ""' v ""' k(n-p)
~ YkWN - ~ ~ .InWN -~.In~ WN
k=O k=O n=O n=O k=O
~
~WN
k(n-p) _
-
{0 N
if p
l"f
=f: n,
k=O P = n.
8.1 Computing the Fourier coefficients 67
We see that this is the same as formula (8.2)! After a change of indices, we
discover that
8.1.1 Conclusions
Integrating (8.1) using the trapezoid method yields N approximate Fourier
coefficientsc;;;
that are equal to the Fourier coefficients of the trigonometric
polynomial (8.3) that interpolates f at the points tk = k(a/N). We have
the equivalent formulas
N-1
Yk = LYnwRrk, k = 0, 1, 2, ... , N - 1,
n=O
(8.5)
N-1
Y. 1 ~ -nk n = 0, 1, 2, ... , N- 1,
n = N W YkWN '
k=O
N
if :s;n< 2 ,
0
N
if - 2 :::; n < 0.
1 1 1
N-1
1 WN
2(N-1)
1 WN
2(N-1)
WN
and
8.1.3 Remarks
(a) Be careful to note that the computation of the Yn yields the Fourier
coefficients c!;; in the following order (N = 8):
Yo Y1 Y2 Y3 Y4 Y5 Y6 Y1
c~ c~ c~
(b) It is convenient, particularly for manipulating the formulas, to extend
the vector y E CN to a periodic sequence with period N, which just comes
back to its original definition, since Yk = f(kajN), where f has period a.
This allows us to write Yk with k E Z. We will use this convention in the
rest of the book. Formula (8.5) shows that the Yn are also periodic with
period N. We will use the same convention and define Yn for n E Z.
These conventions mean that the approximate Fourier coefficients c!;;
also form a periodic sequence. Here it is important to remernher that c!;; is
an approximation (see Figure 8.1) of Cn only for
N -1 0 1 2 N N-1 N N+ 1
2 2
FIGURE 8.1. Fourier coefficients exact () and computed (x).
Yn = N1 "~ Y-kwN
I -nk 1
= N
"~ D
k=O k=-N+l
8.2.3 Theorem Let (xk) and (yk) be two complex sequences with pe-
riod N and Jet (Xk) and (Yk) denote their discrete Fourier transforms.
70 Lesson 8. The Discrete Fourier Transform
Zk = L XqYk-q, kE z,
q=O
(ii) The transform of the pointwise product of the sequences (xk) and
(Yk) is
wherc
N-1
Pn = L XqYn-q
q=O
Proof. By definition,
N-1 N-1
Zn = N1 "
L "
L -nk '
XqYk-qWN
k=O q=O
N-1 N-1
L iYki 2 =N L IYnl 2
k=O n=O
(xk) ~ (Xn),
(yk) ~ (Yn)
By linearity,
Zn= Xn +iYn
But notc that Xn and Yn arc not necessarily real! With this notation,
1 -
Xn = 2(Zn + ZN-n),
1 -
Yn = 2i(Zn- ZN-n)
It is only necessary to compute these values for n = 0, 1, ... , N /2, since the
values for n bctween N /2 and N - 1 will have already appeared as their
conjugates. Thus it is sufficient to compute the transform of (zk) to obtain
the transforms of (xk) and (yk)
It is also possible to compute the Fourier transform of a single real vector
with N components by using a transform of length N/2 (see [CLW70]).
+oo t
!( t ) = '""'
~ Cne
2i7rn-
a (8.6)
n=-oo
72 Lesson 8. The Discrete Fourier Transform
n=-oo
nk
Cn+qN ) WN,
m=-oo q=-oo
From this we see that for a fixed N, the faster Cn tends to zero as n tends
to infinity, the better will be the approximation
N N
for - -2 -< n -< -2 - 1.
L
+6 t
f(t) = cnii1rn;;;.
n=-6
8.5 Exercises 73
For N = 8: cg =Co,
c~ = c1.
This simple example illustrates the following general result, which follows
directly from (8.7):
Fora trigonometric polynomial of degree P, the values of the approxi-
mate coefflcients computed with the discrete Fourier transform are exact
whenever N ~ 2P + 1.
8.5 Exercises
Exercise 8.1 Consider two consecutive discrete Fourier transforms:
Compute Zq as a function of Yk
Exercise 8.2 Let (xk) and (yk) be two complex periodic sequences (with
period N) such that
XN-k = Xk and YN-k = Yk
for all k E Z. Show that the discrete Fourier transforms (Xn) and (Yn) arereal
and that they can be computed with a single transform of order N.
A Famous, Lightning-Fa st
Algorithm
Computing the vector (Yo, Y1 , ... , YN-d using formula (8.5) requires
(N- 1) 2 complex multiplications,
N(N- 1) complex additions,
assuming that the values of w~, the sines and cosines of the given angles,
have already been computed and stored.
A typical value of N is of the order of 1000, which implies about a million
operations of each kind. Considering the frequency of this computation, it
was natural to seek to lower the cost. In 1965, two American scientists,
J. W. Cooley and J. W. Thkey, developed a much more efficient algorithm
that has since been known as the fast Fourier transform (FFT). This al-
gorithm takes into consideration the special form of the transformation
matrix, which is constructed from the roots of unity. From the beginning,
the FFT, including its many extensions, has enjoyed enormous success. In
fact, it is safe to say that it has been the backhone of signal and image pro-
cessing in the last half of the twentieth century. Furthermore, it has been
the inspiration for numerous investigations in algebra independently of its
intensive use in signal processing. It was indeed a marvelous discovery. The
fast Fourier transform marked an important step in the theory of the com-
plexity of algorithm. This field of research is concerned with determining
and minimizing the cost of a given computation or dass of computations,
where the cost is measured by the number of numerical Operations. For
example, we will see that the cost of the FFT is of the order N log N.
Pk = -1 (
Yo+Y2WN -2k -(N-2)k) ,
+Ooo+YN-2WN
m
Ik=- 1 (
Yl+Y3WN -2k -(N-2)k)
+ooo+YN-lWN
m
Note that we have the relations
j j
FIGURE 901.
\!x!, !!x!
}b .Yl Y4 .Ys Y1 .Y3 .Ys Y1
Yo Y4 Y2 Ye
II II II II
Po lo Po lo
l Xl ! X!
FIGURE 9.2. Rearrangement of the data.
yk = 21 ( pk + WN-k Ik ) '
Yk+m = 21 ( Pk- WN-k lk ) ,
for k = 0, 1, ... , m - 1. Figure 9.3 illustratcs the complete algorithm for
N = 23 . The wiggly lines separate the independent computations. In an
actual program, a single vector is used. This is ultimately the output vector
(Yo, Y1. ... , YN-1); it is the result of successively transforming the vector
obtained by appropriately rearranging the original data.
~
Yo
Xl y1
~X~
Yo y1
~X~
Yo y1
~X!
Yo y1
II II II II II II II II
Po p1 lo /1 Po p1 lo /1
j j j):Kj
Yo y1 y2 Ya Yo Y1, y2 Ya
II II II II II II II II
p1 /2 Ia
Multiplications by w-;/
(k ~ 1): [2P- 1 - 1; 0];
Additions: [0; 2Pj.
From these relations we have
M1 =0, A1 = 2,
Mv = 2Mp-l + 2p-l - 1, Ap = 2Ap-1 + 2P.
A computation, which is left as an exercise, provides an explicit expressions
for Mp and Ap, namely,
Mv = (p- 2)2P- 1 + 1,
Ap = p2P.
We see from this that the global cost, as a function of N, is
Multiplications Additions
N Old FFT Ratio Old FFT Ratio
Method Method
2 0 0 2 2 1
4 0 0 12 8 1.5
8 49 5 10 56 24 2.3
16 225 17 13 240 64 3.8
32 961 49 20 992 160 6.2
64 3,969 129 31 4,032 384 10
128 16,129 321 50 16,256 896 18
256 65,025 769 85 65,280 2,048 32
512 261,121 1,793 145 261,632 4,608 57
1,024 1,046,529 4,097 255 1,047,552 10,240 102
TABLE 9.1.
We note that compilers deal with these recursions more or less wcll, par-
ticularly on microcomputers. While the program itsclf is concisely writtcn,
which is very attractive to the programmcr, its execution, by contrast, re-
quires a great deal of processing and a large amount of memory: At each
call to FFT, the procedure is completely recopied with new parameters.
Finally, it is not obvious that this procedure does indced compute the
desired FFT. For example, the second call to FFT is cxecuted only after
many other such calls.
9.5 Exercises 81
9. 5 Exercises
Exercise 9.1 Consider the discrete Fourier transform of order N defined by
the formulas (8.5)
N-1
v
<n = N1 ~
LYkWN
-nk
, n = 0, 1, ... , N- 1.
k=O
Write the discrete Fourier transform in its matrix form Y = SNY What is the
matrix associated with the inverse transform?
**Exercise 9.2 Let u and v be two complex periodic sequences with period
N. Consider the periodic convolution w = u * v defined by (10.1):
N-1
Wn = L Un-qVq, n = 0, 1, ... , N- 1.
q=O
with
1 0 0 0
-2i7r-\-
0 e 2 0 0
L2 = -2i7T~
0 0 e 2 0
-2i7r~
0 0 0 e 2
where
h]
-h '
!2 being the 4 x 4 identity matrix.
Write v 0 = y and cut v 0 , a vector of length 8, into two vectors,
v8 and v?, of length 4:
Second step:
This step is to compute Y* = R 1v1 .
Show that R 1 can be written as the product of 3 matrices:
Rl = R21~1,
l'
where
0 0 0 0
1] [~ ~
R2 = T1 0 l= L1 0
IT 0
0
T1
0
0 h
l
0 0 L1
[I,
and
h 0
~1 =! h -h 0 0
2 0 0 h I0 .
0 0 h _}1
h is the (2 x 2) identity matrix, T1 = HS1 is a 2 X 2 matrix,
l
and
0.
-2t1f~
1
e 2
R2 = R32~2,
1 1 0 0 0 0 0 0
1 -1 0 0 0 0 0 0
0 0 1 1 0 0 0 0
0 0 1 -1 0 0 0 0
~2 =!
2 0 0 0 0 1 1 0 0
0 0 0 0 1 -1 0 0
0 0 0 0 0 0 1 1
0 0 0 0 0 0 1 -1
(3) How does one proceed to obtain Y0 , Y1, ... , Y7 in this order?
(4) How must one modify the algorithm to compute the inverse discrete Fourier
transform?
Theorem 8.2.3 points to another way to proceed: Let (Xk), (Hk), and (Yk)
be the discrete Fourier transforms of the sequences (xn), (hn), and (Yn)
Equation (10.1) becomes
(10.3)
If we assume that the length N of complex vectors is a power of 2, N = 2P,
then this computation proceeds as follows:
Computation Cost
Step 1: Compute the transforms 9N [N(p- 2); 2Np]
(xn) ~ (Xk)
(hn) ~ (Hk)
Step 2: Compute the products (10.3) [N;O]
Step 3: Compute the transform .!J?iv - 1 [(N/2)(p- 2); Np]
!T. -1
(Yk) ~ (Yn)
The total cost is
For N = 64, the cost is [832; 1, 504] versus [16, 384; 16, 256] given by (10.2).
10.2 Nonperiodic convolution 87
Xn = 0 if n < 0 or n 2:: M,
hn = 0 if n < 0 or n 2:: Q. (Q $ M).
N
2 (3log2 N - 4) complex multiplications,
3N log 2 N complex additions.
MQ = 105 multiplications,
MQ- (M + Q- 1) ~ 104 additions.
The cost using the FFT is
We see that the FFT method is still advantageaus in this case. On the
other hand, this advantage is lost when the lengths of the two signals are
disproportionate. This happens frequently in "real-time" signal processing,
where the sequence (xn) is practically "infinite" and where the support of
the filter (hq) is relatively small. This is the case, for example, when one
smoothes data with a "sliding window."
EXAMPLE: Suppose
4
Yn = LhqXn-q (10.6)
q=O
88 Lesson 10. Using the FFT for Numerical Computations
is represcnted by
(10. 7)
This represcntation is convenient for computing the values of P for differ-
ent values of x (Horner's algorithm). It is less convenient if one wishes to
computc the product of P and another polynomial
Q(x) = bo + b1x + + bqxq.
The coefficients of the product
PQ(x) = c0 + c1x + + cp+qxp+q
are given by thc convolution
k
Ck=Lanbk-n, k=0,1, ... ,p+q. (10.8)
n=O
P(xj)Q(xj), j = 0, 1, ... , N- 1.
(We have assumed that p+q:::; N -1.) On the other hand, the computation
of P(x) for an arbitrary value x is more complicated in this representation.
We are going to examine these two representations and the problern of
going from one to the other.
~ (3log2 N- 4) multiplications,
3Nlog2 N additions.
EXAMPLE: Take p = 13, q = 15, and N = 32. The direct method costs
435 multiplications,
406 additions,
176 multiplications,
320 additions.
However, as indicated in Section 10.2, the FFT method loses its advan-
tage when p and q are not of the same order of magnitude.
90 Lesson 10. Using the FFT for Numerical Computations
(10.10)
The two representations (10.7) and (10.9) are then related by the equations
N-1
Yk = "'"' nk
~ anWN, k = 0, 1, ... , N- 1. (10.11)
n=O
N1 "'"' n = 0, 1, ... , N - 1.
-nk
an= ~ YkWN ,
k=O
We know from Section 9.2 that the cost of going from one representation
to the other is
N
2(log 2 N- 2) multiplications,
Nlog 2 N additions.
To(x) = 1,
T1(x) = x,
T2(x) = 2x 2 - 1,
T3(x) = 4x 3 - 3x,
T 4 (x) = 8x 4 - 8x 2 + 1, etc.
10.4 Polynomial interpolation and the Chebyshev basis 91
Since we wish to remain in the real domain, we cannot use the Xk given by
(10.10) directly, but our choices are derived from (10.10). In particular, we
take the abscissas Xk (which are called the Chebyshev abscissas) to be
From (10.12),
(10.15)
and
N
Yk = P(xk) = L
ancos ( nk ~ ). (10.16)
n=O
This formula is not exactly a DFT, but it is not far from one. By writing
the cosines in terms of exponents, we see that
N -N N
Yk = "'"' nk
2 LJ anw2N
1 1 "'"' nk
+ 2 LJ a_nW2N = "'"' nk
LJ CnW2N
n=O n=O n=-N (10.17)
r
where
2 an if 0 < n::; N,
Cn = ~0 if n=O, (10.18)
2 a_n if -N:Sn<O.
92 Lesson 10. Using the FFT for Numerical Computations
Yk = L:
n=-N
k = 0, 1, ... , 2N- 1. (10.19)
Applying the technique used in Section 8.1, we try to invert this system by
computing
2N-l
L
Ykw-;J:/, p = 0, 1, ... ,N.
'Yp =
k=O
By substituting (10.19) for Yk, this becomes
L L L L
2N-l N N 2N-1
'Yp = Cnw~'J.r-p)k = Cn w~'J.r-p)k.
k=O n=-N n=-N k=O
with
Eo = EN = 2; c1 = = EN-1 = 1.
Formulas (10.16) and (10.20) are the reciprocals of each other; they resolve
theoretically the problern of going from onc representation to the other.
It is possible to reduce this cost by taking into consideration the fact that
the sequences arereal and even: In (a) Y2N-k = Yk, andin (b) Y2N-n = Yn.
Thus the computation can be clone with a discrete Fourier transform of
order N /2 rather than order 2N. The cost is
f(()) = L an cosnO.
n=O
94 Lesson 10. Using the FFT for Numerical Computations
that satisfies
!(k;) =yk, k=O,l, ... ,N.
The coefflcients an are given by (10.20).
10. 5 Exercises
Exercise 10.1 Take two vectors of lengths M = 4 and Q = 3. Extend
them as described in Section 10.2 and verify that the computations with the new
vectors do indeed give the original convolution.
Exercise 10.3
(a) Show that the function f(O) in Theorem 10.4.6 can be written
L f(fh)gk(O),
N
f(O) =
k=O
where
N
k 1
9k(O) =La~ cosnO, and an= -N cosnOk,
E:n
n=O
We are going to introduce the Lebesgue integral here andin the next three
lessons. Experience has shown that this notion of integration is particularly
well suited for operations such as
taking limits under the integral sign,
taking derivatives under the integral sign,
interchanging the order of integration.
We do not intend to give a complete and rigorous development of the the-
ory of Lebesgue integration. Readers wishing a deeper understanding of
the fundamentals can consult any of numerous references such as [KF74],
[Hal64], and [Roy63]. We wish to go as directly as possible to the applica-
tions while at the same time presenting the essential ideas.
1b f(x)dx
is the area of the region bounded by the curve y = f(x), the x-axis, and
the two lines x = a and x = b. It was in 1853 that Bernhard Riemann gave
a rigorous definition of the integral that bears his name.
For a fixed n, consider a parti~ion ~n of the interval [a, b] of the form
1 a
c-a f(x)dx + Jb
c+
f(x)dx
as o:, --+ 0, o:, > 0. Riemann also proved that the Fourier coefficients of a
periodic, integrable function tend to zero. This result was later generalized
by Henri Lebesgue.
Riemann's work stimulated numerous studies aimed at generalizing the
definition of the integral to include the widest possible class of functions
(T. Stieltjes, E. Borel, H. Lebesgue, F. Riesz). It was around 1900 that
Lebesgue proposed his theory of integration. The Riemann sums are well
behaved for only a particular class of discontinuous functions; they require
that f(x) not vary too much in the intervals (xi-I.Xi) Lebesgue inverted
the situation: He considered the range of values [m, MJ taken by f(x) and
partitioned this interval into segments (Yi-1, Yi) He then considered the
set of x such that Yi-1 :S: f(x) < Yi He gave this set a measure mi and
formed the sums
L m(rli where Yi-l''li < Yi.
20 F r--
10 F r- --
5F r-- f--
-
H
2F
1F a1 ~ ra;- a4
1 2 3 4 5 6 7 8 9 10 11 12 13
~
first second'
bill bill SR =sum of the a;
II II
7F 11 F
FIGURE 11.1. Riemann's point of view.
20 F -1
10 F -1 2 3
5F r-
1
.....-.---
2 3
-4
2F -1
1F Ir #
SL =(1 F x 2) + (2 F x 2) + (5 F x 4) + (10 F x 3) + (20 F x 1)
The goal of measure theory is to extend the notions of length (of an interval
in JR.), of area (of a reetangle in JR. 2 ), of volume, and so on, to more complex
sets. In general, given a set X (which in applications will be a part of JR.n),
one considers a restricted family of subsets of X called a a-algebra, and it is
on the elements of this a-algebra that one defines a measure. A measure is
a function with values in JR.+ U { +oo} = i:+, where JR.+ = { x E lR. I x ;:::: 0},
that has certain desirable additivity properties.
12.1.1 Definition Given a set X, let .9 (X) be the dass of all subsets
of X . ." c .9 (X) is a a-algebra if the following hold:
(i) 0 and X are in ." ;
(ii) S E ." implies X \ S E ." ;
(iii) 81, S2, ... E ..9'" =? U:'=l Sn E ..9'" .
." is said to be closed under the formation of complements and count-
able unions. Note that .9 (X) is itself a a-algebra, but in general it contains
too many sets to be of interest.
f-l( usn) =
00 00
LJ-L(Sn)
n=l n=l
Then /-La and /-ld are measures on .!Y . However, these measures do not
generalize the notion of length when X = IR.!
v( U Sn)= Lv(Sn)
n=l n=l
Given a measure on an algebra d , it can be extended to a a-algebra
containing d . This is the content of the following theorem, which we will
assume (sec, for example, [Hal64]).
This procedure is used on lR. to define Lebesgue measure. One can take
the elements of d to be the sets that are finite Unions of intervals. The
measure v is defined for intervals by
v( U In)= Lv(In)
n=l n=l
It is possible to show that v is a measure on d . (Note that the delicate
part of the argument is to prove condition (ii), the countable additivity.)
From Theorem 12.1.7 we know that v can be extended to a measure J.L
on a a-algebra 5f containing d and hence containing the Borel sets. This
measure J.L is constructed by defining
L J.L(Sn) = 0.
CXJ
J.L(S) =
n=l
j(x) = {1 0
if X E_Q,
otherw1se
XA( x) = {ol if X E A,
if X.;. A.
In practice, it is easy to verify that the functions being used are measur-
able. The following properties are particularly useful in this regard.
12.4 Exercises
Exercise 12.1 Let X be a set and l!f c .9' (X). Show that the intersection
of all the u-algebras containing l!f is a u-algebra.
108 Lesson 12. Measuring Sets
Exercise 12.2
(a) Show that every closed set F of Rn is a Bore! set of Rn. For example,
a E Rn is a Bore! set.
(b) Show that Q and R\Q are Bore! sets in R.
(a) (-oo,a); (b) (-oo,a]; (c) [a,b); (d) the closed sets.
A= n An,
00
n=l
and assume that p,(Al) < oo. Show that p,(A) = !im p,(An)
n--+oo
Hint for (a): Write An as a countable union of pairwise disjoint measurable sets:
Exercise 12.6 Verify that /-La and /-Ld defined in Section 12.1.5 are measures.
Exercise 12.7 Let (X, Y , p,) be a measure space. Consider the space E of
measurable functions with values in i:. Show that the relation f "' g if f - g = 0
almost everywhere (a.e.) is an equivalence relation on E.
Exercise 12.8 What is the Lebesgue measure ofQn(O, 1]? Ofthe irrationals
in (0, 1]?
We will assume that the value of L.:~=l O!iJ.L(Si) does not depend on the
particular representation of e (see Definition 12.3.6), hence that the integral
of e is well-defined.
112 Lesson 13. Integrating Measurable Functions
Jx f df..L = 1, and one can show that f is not integrable in the sense of
Riemann.
(b) X= [0, 1] and
f(x)={1/x ~fx>O,
0 1f X= 0.
for all measurable sets E. The two integrals an the right make sense; how-
ever, in case they are both +oo, JE
f dJ.L is not defined. This leads to the
following definition.
The technique for proving this theorem can be found, for example, in
[Ber70]. This result is used to prove linearity for nonnegative functions,
which in turn is use to prove the next result.
we have
t+ + g+ + h- = h+ + r + g-.
Using the linearity of the integral for nonnegative functions and rearranging
terms shows that
lfd~-t:::; lgdJ1-.
13.3 The integral and sets of measure zero 115
f(x) = {+1 -1
~f x st Q,
If xEQ.
The proof of this theorem is based on the definitions of the two integrals
and on the monotone convergence theorem. The following sufficient condi-
tion for a function to be Lebesgue integrable is much easier to establish.
Proof. Take C:n > 0, C:n ~ 0, and define fn = f X[a+cn,b] Clearly, fn(x)
converges to f(x) as n ~ oo for all x E [a, b]. Furthermore, the sequence
f n is increasing. By the monotone convergence theorem,
{
J[a,b]
fdt-L = lim {
n-+oo J[a,b]
fndt-L = lim
n-+oo
1b
a+cn
f(x)dx =I. o
When fisnot positive, we use this result for I/I to conclude integrability
but not, for the moment, to compute the value of fra,b] f dt-L. We will see
further results relating the Riemann and Lebesgue integrals in Lesson 14.
13.4.5 Examples
(a) f(x) = .)x on (0, 1] is not Riemann integrable on [0, 1]. It is inte-
grable on every interval [c:, 1] with 0 < c: < 1, and
[ 1
1"' .rx
dx
= [2vxJ! = 2- 2vre.
Thus f is Lebesgue integrable on [0, 1] and
r
J(o,l]
fdt-L = 2.
1
lf(x)l:::; -;;x
118 Lesson 13. Integrating Measurable Functions
lim
cn -->O+
1b a+cn
lf(x)l dx = +oo,
then f is not Lebesgue integrable on [a, b]. In the case where f is nonneg-
ative, it is clear that the (generalized) Riemann integral
1ba
f(x) dx =
n
lim
-->O+
1b a+cn
f(x) dx
does not exist either. If, however, f takes both positive and negative val-
ues, the generalized Riemann integral can exist without f being Lebesgue-
integrable. Take as an example f(x) = k sink Set
In = 1-1 1. 1
cn X
s1n- dx
X
1
= ~ /"nsinu
1
[-cosu]1/cn ~ 1 /"ncosu
--du = - - -
U
- - 2- du.
U 1 1 U
The term
[ - cosu]1/cn
-- = -E:nCOS-
1
+cos1
!
U 1 n
1 /cn COSU
converges as E:n --+ 0. The integral - 2- du converges absolutely
1 u
!
because
11"n 1 cos ul
---du
[
< - -
1/cn
= 1-
1] n
1 2 u - u 1
11
Thus In converges as E:n --+ 0. On the other hand, J" = / " Isi: u Idu
tends to +oo as c --+ 0 because
1I 1
and
mr . n-1 1 (k+1)1r 2 n-1 1
7r
81
: I
u du ?. L
k=1
k1r
k?r
Isin ul du = ; L
k=1
k.
Proposition 13.4.4 can be proved for a generalized Riemann integral on
an interval [a, +oo], and Remark 13.4.6 is also true in this case.
13.4. 7 A convention
For the Riemann integral, the symbol I: f(x) dx makes sense when b < a
by the relation I:
f(x) dx = - Iba
f(x) dx, which comes directly from the
Riemann sums. On the other hand, the Lebesgue integral is taken over a
nonoriented set (a, b). When the integrals in a given context are all Lebesgue
integrals, we will adopt this sign convention. For example, I 1 f(x) dx will
denote - I[o, 1] f dJ.L.
13.5 Exercises 119
13.5 Exercises
*Exercise 13.1 Prove Proposition 13.2.1 for simple functions. Extend this
result to nonnegative measurable functions.
Exercise 13.2 Use Theorem 13.2.2 to prove the linearity of the integral for
nonnegative functions.
Hints:
B = {x E A llf(x)l = +oo}
and
An= {XE AI n ~ lf(x)l < n + 1 },
and decompose Aas A = B u BN u (A\(B u BN)) with BN = u:=O An.
lfndj.t~M
for all n E N. Show that the sequence fn converges almost everywhere to an
integrable function f on E and that
Hint: Use Theorem 13.2.2 on the functions 9n = inf fk Recall that for a
k;o:n
sequence ofreal numbers a,., n E N, the limit inferior, denoted by lim inf an,
n-+oo
is the quantity supkEN{infteN ak+t}.
(b) Investigate the sequence fn = -~X[o,n], which does not satisfy the non-
negativity hypothesis, and verify that Fatou's inequality does not hold.
Exercise 13.10 Consider the u-algebra .9" (N) of all subsets of N and endow
the elements E of .9" (N) with the counting measure defined by
J.t(E) = the number of integers in E.
Show that f : N ----> lR is integrable with respect to the measure J.t if and only if
n=O
u(E) =l f dj.t,
Integral Calculus
This lesson contains the essential tools for putting into practice integral
computations: It is the Lebesgue vcrsion of integral calculus. We present
rules for manipulating integrals that depend on a parameter. In particular,
we discuss continuity and derivation with a view toward applications to
the Fourier transform. The lesson also contains the formulas for changing
variables and the rules for interchanging the order of integration in double
integrals, the celebrated Fubini's theorem.
f(x)={+l ifxE~n(0,1],
0 otherw1se.
r
shows that
lim
n-oo}x r
fn(x) dJ.l = lim fn(x) dJ.l,
Jxn-oo
or
Proof. The proof is essentially the same as the one above. Here we write
is well-defined for all t E ~ because le- 2i1rtx f(x)l :::; lf(x)l. Formally, the
J
derivative of is
f'(t) = -2Q(x)(t).
is finite.
8cpn(x) cpn(x)
axl xn
Jcp(x) denotes the determinant of the Jacobian, and !Jw(x)! denotes its
absolute value.
J(x) = 1x f(t) dt
is differentiable for all x E [a, b] and J'(x) = f(x) for all x E [a, b].
(ii) If f is continuously differentiable on [a, b], then for all x E (a, b)
Proof. u(x) = u(a) +I: u'(t) dt. Multiplying both sides by v' and inte-
grating with respect to x shows that
Write
Apply Fubini's theorem and interchange the order of integration; thc last
integral becomes
and this proves the result. (We leave it to the reader to show that thc
hypotheses of Fubini's theorem are fulfilled.) o
128 Lesson 14. Integral Calculus
14.6 Exercises
Exercise 14.1 Consider the sequence of functions fn: R-+ R+ defined by
lim
n---+oo
lbn
-an
f(x) dx = r
}IR.
f(x) dx.
1
(b) Show that
L },.,
00 r(n+l)a
f(x)dx= f(x)dx
R n=-CXl na
f:
n=l
11/n(x)l dx
IR
< +oo,
lim /r
n---+oo
r (1 +X+n y)n dxdy.
J
14.6 Exercises 129
Exercise 14.6 Let f : IR ----> IR be such that the function X ----> xn f(x) is
integrable on IR for all n ;::: 0. Show that the Fourier transform
is infinitely differentiable.
1
*Exercise 14.7 Show that f(x, t) = e-tx is integrable on IR+ for all t > 0.
Use the following two methods to verify that I(t) = 00
e-tx dx is infinitely
differentiable:
(a) Compute I(t) explicitely.
(b) Apply Proposition 14.2.2.
Use this result to deduce that
1:
on the square = [-1, 1] X [-1, 1]. Show that
1 1 1
[
1
dx [ 1
f(x, y) dy = dy [ 1
f(x, y) dx.
defined on = [0, 1] x [0, 1] (except at (0, 0)). Show that the iterated integrals
exist and are different. Conclusion?
f(x, y) = e-x-y (x + Yt
on =IR+ x IR+ by making the change of variables u = x and v = x + y.
130 Lesson 14. Integral Calculus
with p > 0, -~ < 01, ... , (}n-2 < ~ and 0 < On-1 < 271". Let
Spaces
lesson 15
Function Spaces
We have collected in this lesson the definitions and essential results for the
commonly encountered spaces of functions (function spaces or functional
spaces). The lesson is somewhat like a catalog, and it can be used as a refer-
ence to find one's way around function spaces that are perhaps unfamiliar.
Several proofs are technical and can safely be skipped on first reading.
L
p 1/2
(ii) N2(f) = (
11/(k) II~) ;
k=O
(iii) Noo(f) = k=O,l,
max 11/(k)lloo,
... ,p
wherc 11/(k)lloo = maxlf(k)(x)l.
xEI
134 Lesson 15. Function Spaces
15.1. 7 Definition 9J (IR) (or 9J (I)) will denote the space offunctions
in c=(IR) (or c=(J)) that have bounded SUpport.
if lxl ~ 1,
otherwise
is in 9J (IR).
In the last definition it is important tobe precise about the interval I. For
example, if I= [a, b], one can have f E 9J (I) without f being zero at a and
b. If I = (a, b), the support of f must be a closed set in (a, b). In this case,
f can be extended by continuity to all of IR (f(x) = 0, x E IR\supp (!)),
and we have 9J ((a, b)) c 9J (IR).
15.2 Spaces of integrable functions 135
15.2.3 Proposition The spaccs LP(I), 1 :::; p :::; +oo, are complete
normed linear vector spaces when endowed with the following norms:
(i) 11/IIP = (llf(t)IP dt) l/p if 1:::; p < +oo;
(ii) 11/lloo = inf{c I measure {x llf(x)l ~ c} = 0}.
When p = oo it is easy to show that lf(x)l :::; 11/lloo except on a set of
measure zero.
We assume from now on that 1 :::; p, q :::; +oo. The proof of Proposition
15.2.3 is not immediate; we first establish the following result.
1 I
1 1
lf(t)g(t)l dt:::; -II/II~ + -IIYII~
p q
136 Lesson 15. Function Spaces
f lf(t)g(t)l
I
p-1
dt::; ~ 11!11~ + -llgll~;
p qa
1
1
Note that lf + giP- 1 E Lq when -
p
+-q1 = 1; thus by Hlder's inequality
1
II! + gll~::; (jlf(t) + g(t)lq(p- 1 ) dt) '1 (llfllp + llgllp)
= II! + gll~- 1 (llfllp + llgllp),
from which we have II! + giiP :S IIJIIP + llgllp
Showing that thc spaces LP (I) are complete is more technical; we refer
to [Bre83], for example, for a proof. D
The vector space L~(O, a) was introduced in Lesson 5. We now know that
15.3.2 Proposition Assurne that J.L(I) < +oo. Then the following
relations hold:
(i) L 00 (I) c LP(I) for all p ~ 1.
(ii) Lq(I) c LP(I) for all q ~ p ~ 1.
(iii) There exists a constant c = c(p, q) such that llhiiP :::; cllhllq for all
h E Lq(I), 1 :'S p :'S q :'S +oo.
Proof.
(i) For f E L=(I), JI lf(t)IP dt :'S J.L(I)IIfll~ < +oo.
(ii) Take p < q and f E Lq(I). Write S = {t EI llf(t)l ~ 1}. FortES,
lf(t)IP :'S lf(t)lq on S; thus
(iii) The inequality llhiiP ::; cllhllq is another way of saying that the injec-
tion of Lq(I) in LP(J) is continuous. Take h in Lq(I) and apply Hlder's
inequality with
. 1
Smce -
r
+-1s = 1, we see that
q-p
It follows that llhiiP ::; cllhllq with c = (J-t(J)) pq. D
When J-t(I) = -t-oo these results are false. The spaces L 1 (JR) and L 2 (JR)
are not comparable.
Thrning to the integrability properties of the regular functions, we see
immediately that if I is not bounded, then there is no inclusion of cm(I)
in LP(I) for any m ~ 0 and p ~ 1. It is sufficient to take f = 1. If I is
closed and bounded, then cm(J) C L 00 (I) c LP(I) for all m ~ 0 and all
p~l.
The following theorem about approximation is morc precise and will be
used often in the rest of the book.
d(x, I\ rl)
g(x) = d(x, I\ rl) + d(x, F)'
15.4 Exercises 139
fI
iXE - gl djJ. = r
10\F
IXE - gl d~-t :$ 2JJ.(n \ F) < TJ
This proves the result for the case when I is bounded. If I is unbounded,
we first approximate f in the L 1 norm with XJ f, where J is a bounded
open interval. o
15.3.4 Remark Theorem 15.3.3 is also true for 1 < p < +oo. It is false
for p = +oo, as can be seen by taking I = (0, 1) and f 1 (see [Bre83]). =
We will sec later (Lesson 21) that !JJ (R) is densein L 1 (JR). Forthis we will
use the convolution.
The display below summarizcs the inclusion relations among the function
spaces that we have introduced so far.
The space L 2 (JR) plays a central role in Fourier analysis. Its norm is
derived from a scalar product, and it is a Hilbert space. This will be the
subject of the next lcsson.
15.4 Exercises
Exercise 15.1 Let n be an Open set in Rn and let f : n -> R be measurable.
Assurne that f satisfies the following property (B):
(B) There exists c > 0 suchthat lf(x)l ::::; c for almost all XE n.
Define
llflloo = inf { a I lf(x)l ::::; a a.e. on n }.
(a) Show that lf(x)l ::::; llflloo a.e. on n.
(b) Let L oo (n) be the set of functions defined on n with values in R that satisfy
(B). Show that II lloo is a norm on L 00 (!l).
140 Lesson 15. Function Spaces
Exercise 15.3 Let f and g be in LP(I), 1 :::; p :::; oo. Show that II!- giiP =0
if and only if f = g a.e.
Exercise 15.4 Show that if f and g are in L 2 (J), the product fg is inte-
grable. Give an example where J,g E L 1(/) but fg is not integrable on I.
(b) Write
n
9n(x) = lfa(l)(x)l + 2.:: lfa(k+l)(x)- fa(k)(x)l.
k=l
Hilbert Spaces
In this lesson we present the ba.sic elements of the theory of Hilbert spaces.
These spaces generalize several aspects of lR"'. Hilbert spaces are endowed
with a "Euclidean" geometry in the sense that there is a distance function
and the notion of angle between two vectors. Hilbert spaces are complete,
and this allows one to develop the notion of an infinite-dimensional basis.
The prototypic Hilbert space is L 2 (I). 8ome of the points introduced in
Lesson 4, including orthogonal projections, will be formalized.
In (82), (y,x) is the complex conjugate of (y,x). (82) and (83) imply
that (x, ay) = a( x, y). When K = JR, the conjugation bars are clearly
superfluous. 8ince we will be working most of the time in C, we present the
results for this field.
Define llxll = J(x,x). We will see that this is a norm once the following
lemma is established .
i(x,y)i::::; llxiiiiYII
(ii) Parallelogram identity:
0::::; llx- (x, y)yll 2 = llxll 2 + i(x, y)i 2 - (x, (x, y)y)- ((x, y)y, x)
= llxll 2 -i(x,yW,
which proves (i).
To prove (ii), just expand the left-hand side:
16.1.6 Examples
Hilbert spaces:
IR.n with the scalar product (x, y) = :L:~=l XiYi, where x = (xb ... , Xn)
and y = (yl, ,yn)
cn with the scalar product (x, y) = E~l Xi'fk
L~(O, a) with the scalar product (f, g) = Ioa f(t)g(t) dt (sec Section 4.1).
Lp(O, a) is complete by Proposition 15.2.3.
Pre-Hilbert spaces:
c0 ([a, bJ; IR) with the scalar product u, 9) =I: J(t)g(t) dt .
c0 ([a, bJ; q with the scalar product u, 9) =I: J(t)g(t) dt.
16.2 Best approxirnation in a vector subspace 143
More gencrally, if (h, (h, ... , cPn arc pairwisc orthogonal, tlwn
The uniquencss comes from thc scalar product via thc parallclogram
identity. Suppose h and h are two solutions; then d = II!- h II = II!- hll
Since (h + h)/2 is in V,
1 1 1
d::; II!- 2 (h+h)ll::; 2 11!-hll+ 2 11!-hll =d.
16.2.2 Remark Note that the proof of Theorem 15.2.1 remains valid
whcn V is any set of vcctors that is convex and closed in the norm.
Consequently,
for all 0 and all w E V. This implies that (w, f- f*) = 0, which can be
seen by taking (} = 0, n/2, n, and 3n/2. D
then condition (16.1) translates into a system of linear equations in the Ak:
n
LAk(rfJk,r/Jj) = (f,rpj), j = 1, ... ,n.
k=l
The matrix G of this system, with G;j = (rp;, r/Jj), i, j = 1, ... , n, is called
the Gram matrix associated with the basis rjJ 1 , ... , r/Jn. This clearly shows
one reason why we want to have an orthogonal basis: In this case the matrix
is diagonal. (Recall the orthogonal polynomials in Beetion 6.2.)
16.2.4 Proposition Assume that the rp; form a basis for V. The Gram
matrix with general term (rpi, r/Jj ), i, j = 1, ... , n, is Hermitian and positive
definite.
146 Lesson 16. Hilbert Spaces
Proof. Gis clearly Hermitian: Gij = (c/Yi,c/Yj) = (c/Yj,c/Yi) = Gji Now let
X= (xl, ... 'Xn) be an element of cn and compute (X, GX). Since
n n
(GX)i = L_(c/Yi,c/Y1 )xi = ("f:.xjc/Yj,c/Yi),
j=l j=l
it follows that
n n n n
(X,GX) = "f:.xi(GX)i = "f:.xi(Lxic/Yi,cPi) = II "f:.xic/Yill 2 2 0.
i=l i=l j=l i=l
n
If (X, GX) = 0, then L XicPi = 0, and hence X = 0 because the cPi are
i=l
linearly independent. 0
This proof provides another way to show that the best approximation
exists and is unique when V has finite dimension. When the basis is or-
thogonal, f* is given by
(16.2)
The last condition means that (!, <Pn) = 0 for all n E N implies f = 0.
When !1J is an orthogonal system, the numbers cn(f) = (f,</Jn)/(</Jn,<Pn)
are called the Fourier coefficients of f relative to !1J .
L
00
J= Cn (f)c/Jn-
n=l
Proof. By Bessel's inequality, 2::~ 1 lcn(f)I 2 II<Pnll 2 ::; 11/11 2 < +oo; the
Fourier series converges by Proposition 16.3.4. o
Wo now move to question Q2.
(ii)* (iii) TakefEH = [~] and c > 0. There exists a finite number of
scalars a 1 , a 2 , ... , am such that
m
By Theorem 16.2.1,
m m
k=l k=l
If we let ak = 0 for k = m + 1, ... ,p, this relation remains true for all
p?: m, again by Theorem 16.2.1. Thus
p
for all p ?: m, which proves that the Fourier series S(f) converges to f.
Since f was arbitrary, this means that ~ is an orthogonal basis.
*
(iii) (iv) We saw in the proof of Proposition 16.3.2 that
p p
(iv)* (i) Take f E ~ j_. Then for all n E N, (f, 4>n) = 0, and Parseval's
relation implies IIJII = 0. Hence f = 0, which proves that ~ is a total
system. o
In summary,
f = f
n=l
(f, l/>n) lf>n
(l/>n, l/Jn)
if and only if 4>n is an orthogonal basis. In general, the difficult step in this
theory is to show that a given family of functions is a basis, for example,
to show that the 4>n(x) = e2 i1rnx form a basis in 1I = L~(O, 1).
150 Lesson 16. 1-Iilbert Spaces
a f:
n=-N
lenUW + r lf(t) - !N(tW dt = Jor lf(tW dt.
Jo (16.3)
L:
00
16.4 Exercises
Exercise 16.1 Verify that C 0 [-1, +1] endowed with the scalar product
(f,g) =[ 1
1 f(t)g(t)dt
(f,g) = Lf(n)g(n).
n=1
Exercise 16.4 Show that if r/J1, r/J2, ... , rPn arc n nonzero elements of apre-
Hilbert space that are pairwise orthogonal, then thcy are linearly independent.
Exercise 16.6 Show that a Hilbert spacc H that has a countable orthonor-
mal basis is isomorphic to l2 (N). (Consider thc mapping <I>: H-> l 2 (N) defined
by <I>(f) = Cn(f).)
Exercise 16.7 Let rPn be an orthorrormal basis for the Hilbert space H.
Show that for all f and g in H,
L Cn(f) Cn(g).
00
(!, g) =
n=l
Chapter VI
~
f(~) =
{b- 'Tf~e- a)~
sin
a
e-i7r(a+b)E
if
if
e= 0,
e"/= 0.
J
In this case, is not is L 1 (IR) because Isin 1ri~-a)E I is not integrable (Section
13.4.6). We will refer to this example several times.
Proof.
f
(i) The continuity of follows directly from the continuity of the integral
(17.1) with respect to the parameter ~ The function ~ t--+ e- 2 i1r~x f(x) is
continuous on IRandis dominated by lf(x)l, which is in L 1(1R). Proposition
14.2.1 applies.
(ii) For all ~ E IR we have Ii(~) I~ J lf(x)l dx = llfll1 Thus fis bounded,
and 5 is continuous from L 1(1R) to L 00 (IR).
(iii) For f = X[a,bJ we have Ii(~) I~ 1/nl~l for ~ =f. 0 (Section 17.1.2). Thus
liml~l->oo i(~) = 0; clearly this is true for all simple functions. Now take
f in L 1(1R). Since the simple functions are dense in L 1(1R), there exists a
sequence gn of simple functions suchthat limn....., 00 II/- gnlh = 0 and, for
each fixed n, liml~l->oo l9n(~)l = 0. From (17.3), li(O- 9n(01 ~ II/- gnll1
uniformly in ~ E IR for each fixed n. It follows that liml~l->oo f(~) = 0. D
Proof.
(i) The function h : ~ ~--+ e- 2 i-rrt;x f(x) is infinitely differentiable; further-
more, h(k)(~) = ( -2i1rx)ke- 2 i-rrt;x f(x) and lk<k)(~)l :::; 27rlxk f(x)l. Proposi-
tion 14.2.2 applies for k = 1, 2, ... , n, and
Assurne for the moment that f(a) has a limit as a ---+ +oo. Since f is
integrable, this limit must be zero. As a ---+ +oo, (17. 7) becomes
J
Since f' E L 1 (1R.), lima---++oo 0a f'(t) dt exists, and hence lima---++oo f(a) ex-
ists. A similar argument shows that lima---++oo f( -a) exists.
(iii) If f E L 1 (JR.) has bounded support (f(x) = 0 a.e. for lxl greater than
some K > 0), it is clear that xk f(x) is integrable for all k E N; thus by (i),
f E C 00 (1R.). 0
We are now going to examine how the Fourier transform behaves with
respect to translation and parity. We will use the following notation.
17.2.2 Notation
(i) If f has values in C, then f(x) = f(x), the complex conjugate of f(x).
(ii) fu denotes the reflection of f defined by fu(x) = f( -x).
(iii) The translate Ta! of f is defined by Taf(x) = f(x- a).
Taf(O = J
e-2ixt;x f(x- a) dx = J e-2ixt;(a+t) f(t) dt = e-2ixat; [(~).
J(kl(O = (2i7r~)kf(~)
(iii) a i- 0. f(ax)
.r
I-+ m
1 ~(~)
f ~
Xk ax ( ) 5 -1
kfe u -x ~-----+ ( -a + 2i7r0k+l
e-alxl ~ 2a
a2 + 47r2e
. (X) e- a Ix I I!T -4i7r~
- + -::----=---=-
Slgll
a2 + 47r2~2
.r sin 2a7r~
X[-a,+a] (x) I-+ 7r~
17.4 Exercises 161
17.4 Exercises
Exercise 17.1 Assurne that f E L 1 (JR).
(a) ForA E lR\0, define g(x) = j(Ax). Show that
g(~) = ~~~e-2"~<1(D.
Exercise 17.2 Let f(t) = (1- e)X[-l,lJ(t). Show that
~ (sin27r~
J(~) = 1r2
1
e ~ - cos 21r~ ) .
The Fourier transform allows us to pass from the time domain to the fre-
quency domain. It is remarkable that the inverse operation is obtained
very simply from Y itself. In fact, it is just Y . However, one must be
cautious, for as we have seen in the last lesson, f being integrable does
not imply that 1
is integrable (Section 17.1.2). We will need additional
hypotheses on f to invert f ~--+ f
2
Proof. Foreach n > 0, we introduce the function 9n(x) = e- :1xl, whose
Fourier transform is
~ ( ) 1 n
9n ~ = ; 1 + n2~2.
The functions 9n and Yn are in U(R). We can apply formula (17.4) to the
two functions fand e2i?Ttx9n(x), which in view of Proposition 17.2.4(ii) is
For all x ER, IJ(x)gn(x)e 2 i?Ttxl ~ IJ(x)l, and limn-+oo 9n(x) = 1. Since 1
is in L 1 (R), we can apply Lebesgue's theorem and pass to the limit under
the integral sign. Thus
Given E: > 0, there exists 'f} > 0 suchthat ly-tl ~ 'f} implies lf(y)- f(t)l ~ E:.
We decompose (18.2) as follows:
IJ(t) 11~1>'7
9n(0 d~~ = lf(t)l ( 1- ~ arctan n'f}),
7r
(18.3)
111~1>'7
f(t + ~)gn(~) d~l ~ 9n(TJ)IIJII1 (18.4)
As n tends to +oo, the right-hand sides of (18.3) and (8.4) tend to 0, and
this proves the theorem. D
5 J(x) = fa(x). D
We will now use the results of Section 17.3.4 to compute several inverse
transforrns. We take a E C with Re( a) > 0.
k
(a) h(x) = ~! e-axu(x) is integrable for all k E N, but it is continuous
~ 1
only for k;::: 1. Wehave 91(~) = h(~) = (a + 2 i1r~)k+l E L 1(1R) for k;::: 1.
Thus
( -x)k
g1(x) = ~ eaxu(-x)
for k > 1.
- xk
(b) In the same way, h(x) = kfeaxu(-x) for k;::: 1, and we have
~ -1 ( -x)k
92(~) = h(~) = ( -a + 2 i7r~)k+l and g2(x) = ~e-axu(x).
166 Lesson 18. The Inverse Fourier Transform
Y3(x) = e-alxl.
(d) f 4 (x) = sign(x)e-alxl is not continuous at 0. We note that
is in L 1 (1R). Hence
18.2.2 Summary
l ~ 1!:e-2.,..al~l
a2 + x2 a
is not defined. This does not exclude the possible existence of the limit
ja e2i1rtef(f.)df.
Then
lim = _21 (f(t+) +f(t-)).
a--+oo -a
Proof. Note first that (i) and (ii) imply that the limits f(t+) and f(t-)
exist for all t. Let g(~) = e 2 i7rteX[-a,aJ(~). Since fandgare in L 1(R), it
follows from Theorem 17.1.4 that
~ ~ sin 21ra(x- t)
g(x) = TtX[-a,aj(X) = 1r(x _ t)
Thus,
() 1
v a = f( t+u ) sin27rau du
=1
R 1l"U
00 sin21rau
(f(t + u) + f(t- u)) du. (18.5)
0 1l"U
.
(1') 1Im
R--+oo
1 R . X
O
smx dX = -1r.
--
X 2
(18.6)
and
lim s(y)
y--++oo
= 0. (18.8)
168 Lesson 18. The Inverse Fourier Transform
fbHt
- Jb h~(u)s(21rau) du].
3
1
a-+oo
bj+t
lim h~(u)s(21rau) du
a-+oo bi
lh~(u)s(21rau)!::; M!h~(u)!
by (18.9). Since f' is integrable, h~(u) is in L 1 (JR); we can apply the domi-
nated convergence theorem and conclude that
lim
a-+oo
1 bj+l
bj
h~(u)s(21rau) du= 0.
lim
a-jo+<X>
1 e i1r~tsm~ d~
a
-a
2
.
~
=
{7f
7r/2
0
if
if
if
ltl < 1/(27r),
ltl = 1/(27r),
ltl > 1/(27r).
18.4 Exercises
Exercise 18.1 Consider the following two statements:
(a) f is equal almost everywhere to a continuous function.
(b) f is continuous almost everywhere.
Show that (b) implies (a) butthat the converse is false.
1
2 -.. 7r2 c2
of f(x) = e-ax , a > 0, showed that f(e) = Ke-a"' . Determine the constant
K using Proposition 18.2.1, and use this result to evaluate e-ax 2 dx directly.
(c) Use the expression for the Fourier coefficients of a product (Exercise 5.13)
to deduce Shannon's formula: For all t ER,
We have seen in the last few lessons how it is necessary to restriet the
choice of functions in L 1 (~) if we wish to use the differentiation formulas
and define the inverse Fourier transform. In this lesson, we are going to
introduce a subspace of L 1 (~) that is invariant under the Fourier transform,
differentiation, and multiplication by polynomials.
Proof. Since f decays rapidly, there is an M > 0 suchthat for alllxl > M
we have lxP+ 2 f(x)i ::=:; 1. Thus
Jffi.
::=:; r ~~xP+2 j(x)idx
r lxPf(x)idx+ lfxf>M
r lxPf(x)idx lfxf<5_M X
::=:; MP r
lfxf<5_M
if(x)i dx + r
lfxf>M
~ dx < +oo.
X
D
172 Lesson 19. The Space .:? (JR)
The space .9 (JR) is called the Schwartz class of functions, and it is named
for the l<rench mathcmatician Laureut Schwartz.
The proof is left as an exercise. The main result of this lesson is contained
in the next theorem.
19.2 The space sP (IR) 173
19.2.3 Theorem The space .5I' is invariant under the Fourier trans-
form; that is, f E .5I' => 1E .5I' .
Proof. Assume f E .5I'. Then f is in L 1 (IR) and decays rapidly. Thus
1E 0 00 by Proposition 19.1.3. Since J(k) is rapidly decreasing for all k E N,
-.1(
)P !JT (((-2i1fX)qj(x))(p))(~) = e!JT ((-2i7rX)qj(x))(~) = e1(q)(~).
2z7f (19.1)
We will need the notion of convergence in .5I' in later parts of the book.
Proof.
(i) This is part of the definition.
(ii) It is sufficient to prove this for P(x) = xk; thus we must show that
xP(xk fn(x))(q) converges uniformly on R Butthis follows directly from the
definition and from Leibniz's formula for the derivatives of a product.
(iii) Take c: > 0. Since fn -+ 0 in .5I' , there is an N > 0 such that for
all n ~ N and all x E IR, 1(1 + x 2 )fn(x)l :::; c:. Thus for all n ~ N,
J lfn(x)l dx:::; c: J(1 + x 2 )- 1 dx = c7f, which proves that fn-+ 0 in P(IR).
174 Lesson 19. The Space .'f' (JR)
(19.2)
1
If f is in .9 , then is in .9 and hence is integrable. Since f is continuous
everywhere, (19.2) is true for all x E :IR.. In other notation,
J=Y(5!) (19.3)
for all f E .9 . In the same way, f = 5 (5 !) . This means that 5 is
a 1-to-1 mapping on .9 , and its inverse is
5 - 1 =5. (19.4)
19.4 Exercises
Exercise 19 .1 Find an exarnple of a function in coo (R.) that decays rapidly
but whose derivative does not decay rapidly.
*Exercise 19.2
(a) Show that if f and gare in.'? (R.), then the product fg belongs to .'? (R.).
(b) Show that the rnapping f ~--+ fg is continuous frorn .'? (R.) to .'? (R.).
Suppose that f E C 00 (Ji) and all of its derivatives are slowly increasing, and
suppose that g E .'? (R.). Show that fg E .'? (R.).
Exercise 19.4 Assurne that g E .'? (R.) and that f = P/Q, where P and Q
are polynornials. Show that if Q has no real zeros, then fg E .'? (R.).
The convolution, like the Fourier transform, is one of the essential tools
of signal processing. The results that we establish in this lesson will be
restricted to functions. Our development will often rely on the theorems on
integration established in Lesson 14. The limits of the notion of "function"
and practical applications lead naturally to generalize the Fourier trans-
form, convolution, and associated concepts to distributions. The study of
distributions will begin with Lesson 26; the convolution for distributions
will be developed in Lesson 32.
20.1.2 Examples
(a) Let f = g = X[o,I] Then
L f(x- t)g(t) dt = 1 1
X[o, 1J(x- t) dt = measure ([0, 1] n [x- 1, xl),
178 Lesson 20. The Convolution of Functions
if X :S 0,
Jg(x)~f 2-x
if 0 :S X :S 1,
if 1 :S X :S 2,
0 if X 2': 2.
0 X 0 2 X
f= g=X[0,1J
Proof.
(i) Since fandgare in L 1 (R), Fubini's theorem implies that the function
(y, z) ~ f(y)g(z) is in L 1 (R2 ). By making the change of variables y = x- t
and z = t, we have
iJrr
IRxiR
f(y)g(z)dydz=jrf
J Rx!R
f(x-t)g(t)dxdt.
Thus
and thus
f * g(x) = l +a
-a f(x- t)g(t) dt = (X[a-a,+aJI) * g(x).
f * g coincides on [a, ] with the convolution of two functions in L 1(JR), so
by Proposition 20.2.1(i) it is defined a.e. and is integrable. Thus f * g is
defined a.e. and is integrable on all compact sets.
(ii) If f E L00 (1R), then
Second case: p = 2, q = 2.
From Schwarz's inequality we have
Since IJI E L 1(IR) and 191 2 E L 1(IR), the function u 1---4 lf(u)ll9(x- u)l 2 is
integrable for almost all x (Proposition 20.2.1(i)). The right-hand term of
(20.4), being the product of two square integrable functions, is integrable.
Thus f * 9(x) is defined for almost all x.
(ii) Using the Schwarz inequality and (20.4) we see that
20.4.1 Definition
c~ = {! E C 0 (1R) I supp(f) c [a,+oo] for some a E IR}.
Cpw+ = {! is piecewise continuous I supp(f) C [a, +oo] for some a E IR}.
The function spaces C9.. and Cpw- are defined similarly for functions
whose support is limited on the right.
Recall that C2(1R) denotes the continuous functions with bounded sup-
port and that Cpw denotes the functions that arc piecewise continuous,
that is, f is continuous except for a finite number of points a 1 , ... , ak
where f(aj) and f(aj) exist (see Section 5.2.1).
20.4.3 Remark In the statement of this last result onc can assume
that f and g are in Cpw and have bounded support. The convolution f * g
is again in C2 (IR).
20.5 Summary
Ll * Ll c Ll
Ll * L'XJ c L 00 n C 0
L2 * L2 c L 00 n C 0
L2 * Ll c L2
20.6 Exercises
**Exercise 20.1 Let f: lR-> IC be a measurable function. With the notation
of Definition 20.1.3, supp(f) = lR\0. Show that f = 0 a.e. on 0.
Hint: The proof of this theoretical result is delicate. Write 0 = Kn, where U::"=l
Kn = {x E 0 I distance(x, lR\0) ;::: .!.n and lxl ::; n}.
Note that Kn is compact and hence is in the union of a finite number of the open
sets 0;.
Exercise 20.2 Let f = X[o, 11 . Show that h = f * (f *!) makes sense and
compute h. What is the regularity of h?
Exercise 20.5 Suppose f E L 1 (R) ancl g E LP(R), 1 $ p < +oo. Show that
f * g E LP(R) ancl that II! * gjjp $ llflh llgllp
Hint: The case p = 1 is clone in Proposition 20.2.1, ancl the case p = 2 is clone
in Proposition 20.3.2. For p 1= 1, 2, imitate the proof in Proposition 20.3.2 by
writing
!_ 1-!_
lf(u)g(x- u)! = (!f(u)l!g(x- u)IP)P lf(u)l P;
then use Hlcler's inequality (Lemma 15.2.4).
*Exercise 20.7 Assurne that f ancl gare in L~(O, a) ancl that their perioclic
convolution is
(a) Show that f * g exists ancl belongs to L~(O, a) n C 0 ([0, a]) ancl that
L
(i) Pn(x) 2': 0 for all x ER
(ii) Pn(x) dx = 1.
-1 e -~
{ ~
1-x if lxl :::; 1,
p(x) =
if lxl > 1,
1
with
1 1
c= e dx,
-~
1-x
-1
and let Pn(x) = np(nx). In practice, regularizing sequences are used with-
out defining them explicitly. As we will see, the details are not important;
one uses only properties (i), (ii), and (iii).
Thus
IJE:(x)- 9n(x)l ::; sup IJE:(x)- JE:(x- t)l,
lti:S:En
and
(21.3)
Proof. Take c: > 0. By Theorem 21.3.3 there is a gE; in ! (JR) such that
II!- 9Eih::; c:/4. From (20.1) we see that
II/- f * Pnlll ~ II/- g.,III + llg.,- g., * Pnll1 + IIYe * Pn- f * Pnlh
~ 211/- g.,III + llg.,- g., * Pnlll
Thus there is an N such that for all n > N, llg., - g" * Pn ll1 ~ c: /2, in which
case II/- f * Pnlh ~ c:. This proves that f * Pn tends to f in L 1(IR). 0
21.3. 7 Remark A similar argument can be used to show that the reg-
ularizations f * Pn of a function f E LP (IR), 1 < p < oo, tend to f in LP (IR).
(See Exercise 21.4.)
Proof.
(i) f*Y E C 00 (IR) (Proposition 21.2.1). We look at the behavior at infinity.
First,
lim f * g(x) = lim f f(x- t)g(t) dt =0
lxl->oo lxl->oo }IR
by dominated convergence: f E Y and lf(x- t)g(t)l ~ 11/llooiY(t)l, which
is integrable. To study limlxl->oo xP(f * g)(q)(x) we use the formula
p
xP(f * g)(q)(x) = Li(xP-i!) * (xig(q)),
j=O
21.4.2 Remark Note that we did not use the regularity of f in the last
proof. By modifying this proof we can show that g f-+ f * g is continuous
from !7 (JR) to !7 (JR) when g E !7 (JR) and f E Lfoc(JR) decreases rapidly.
21.5 Exercises
Exercise 21.1 Assurne that f is in L 1 (R) and g(x) = e2 i7l"X. Compute f*Y
*Exercise 21.3 Show that !JJ (R) is densein LP(R), 1 :::; p < +oo. Deduce
from this that .9' (IR) is dense in LP(R), 1 :::; p < +oo.
Exercise 21.4 Take f E LP(R), 1 :::; p < +oo. Use Exercise 20.5 to show
that lim llf- f * PniiP = 0.
n-++cx:>
*Exercise 21.5 Let I be an open interval in IR. Show that !JJ (I) is densein
LP(I), 1 :::; p < +oo.
1:
and from this deduce that
In signal processing, L 2 (IR) models the space of signals that are functions
of a continuous variable (usually time) and that have finite energy. Until
now, the Fourier transform has been defined only for integrable functions,
and L 2 (IR) is not included in L 1 (IR). The purpose ofthis lesson is to extend
the Fourier transform to L 2 (IR); we will do this using results that have been
established for .9 (IR).
Proof. The first equation follows directly from (17.4): Let h(~) = g(~).
kJ(~)h(~) d~ k
From (17.4),
= f(x)h(x) dx.
Thus k(J- 5 !)'1/J = 0 for all '1/J E .5I' (IR). Since j- 5 f is in Lfoc(IR),
for f 6 (x) = e-caxu(cx) with E: = 1 and Re(a) > 0, fc(~) = 2. ( E:a + Z7r
which is not in L 1 (JR). It is, however, in L 2(IR), and in this case we can
compute Y (fc).
22.2.1 Proposition
(i) If f E L 2 (1R), then Y Y f = fa a.e.
(ii) If f E L 1 (JR) n L 2 (1R), then Y f = fa a.e.
Proof. To prove (i), we first show that Y f = Y fa Thus take se- a
quence fn in!/' (IR) suchthat limn---+oo llf- fnll2 = 0. Wehave Y fn =
ff(fn)a by Proposition 18.2.1, and in the limit, Y f = Y fa lf f is
f
also in Ll(JR), then = Y f, and this implies (ii). o
sinx sr
(ii) -X- ~ KX[-...L ...LJ(~)
21r '271"
With these results and those of Lessons 17 and 18, we can compute the
Fourier transform of any rational function P(x)jQ(x) by decomposing it
into partial fractions.
22.3 The uncertainty principle 197
a~=
f
f e1ReWde
JJR
(energy dispersion in frequency).
Proof. We assume the following two results (sec Exercises 22.6 and 22. 7):
(i) lim xlf(x)l 2 = 0.
lxi-HlO
(ii) f'(e) = 2i7fef(e).
The second formula will be proved in the more general context of tempered
distributions (Proposition 31.2.4). Also, note that f being differentiable
almost everywhere does not imply (ii) (take f(x) = X[-l,lJ(x)).
Using (ii) and Theorem 22.1.4(iii) we see that
a~=
f
1 f lf'(eW de
47f 2 }JR = 4 1 2 f l!'(x)l 2 dx.
7f jR
198 Lesson 22. The Fourier Transform on L 2 (R)
I Lx(f(x)7(x))' dxl ~ L L
lxf'(x)7(x)l dx + lxf(x)]'(x)l dx
~(L f (L f
lx7(xWdx 12 lf'(xWdx 12
L+( Llxf(xWdx)
112
( lf'(x)l 2 dx)
112
=2( L lf'(xWdx)
112
( L lxf(x)l 2 dx) 112
=47ro-j CJj.
But
Thc next proposition shows that a Gaussian signal has the minimum
effective duration for a given cffective bandwidth.
-
c 1m e _211"2>-.2
c d.A
4rr2 IR c
4rr 2
D
22.4 Exercises 199
22.4 Exercises
*Exercise 22.1
(a) Let a and b be two real numbers with a < b. Compute the Fourier transform
") sin 1r(b- a)~ -irr(a+b)~ C e;r f h b 1
o f !( ." = 7r~ e . ompute c.T w en a = - = - 27!".
Exercise 22.2 Let f(x) = ~ with a E JR, a :f= 0. Compute the Fourier
a +x
transform of f two different ways: by direct computation (see Section 18.2.2) and
by decomposing f into partial fractions and using the Fourier transform on L 2(JR)
(see Section 22.2.2).
Exercise 22.3 Let f(x) =~ with a E JR, a :f= 0. Show that f ~ L 1 (JR).
a +x
Compute the Fourier transform of f.
Exercise 22.4 In Exercise 18.5 on Shannon's formula, show that the as-
f imply that f E L 2(1R) and that 2:::~:-oo lf(na)l 2 < +oo.
sumptions on
22 . 5
E xerc1se Evaluate 1 IR
-
X
2-x dx and
sin2
- 1 IR
( dx 2) 2 dx.
1+x
Hint: Use Theorem 22.1.4(iii). The results are 1r and i
Exercise 22.6 Take f E C 1 (JR)nL 2 (JR) and suppose in addition that f' and
xf are in L (1R).
2
**Exercise 22.7 Assurne that f E C 1 (JR) n L 2(JR) with f' E L2(JR). We wish
to establish the formula
(a) Let hn(~) =I: e 2 irrt;x J'(x) dx. Show that hn converges to .'T J' in L 2(JR).
From this deduce the existence of a strictly increasing sequence (nk)kEN
suchthat hnk converges a.e. to .'T f' (use [Bre83] p. 58).
(b) By integrating by parts and using [Bre83] p. 130, show that there is a
strictly increasing sequence (kj) suchthat hnk. converges a.e. to 2i7r~.'T f.
J
Lesson 23
-*
convolution and multiplication. Formally, we have these relations:
J ~
g(e) = f(e) 9(e),
h(e) = 1* 9(e).
We will establish conditions under which these formulas are valid.
Proof. Since !/ (IR) is densein L1(1R), there exists a sequence fn in!/ (IR)
suchthat limn--cx> II!- fnll1 = 0. As we have noted, .!T h(t) = fn(t) for all
n E N and all t E IR. We are going to show that JR(f(t) -.!T f(t))cp(t) dt =
0 for all <p E !/ (IR), and this implies (by Exercise 21.6 or by (27.6)) that
.!T 1= f a.e.
Since fn is in!/ (IR) C L 1(1R), we have by (17.4)
1
This result implies that if fand are in L 1 (JR.), then f is continuous, or
more precisely, the equivalence dass to which f belongs contains a contin-
uaus representative, namely, 5 f
23.1.2 Proposition Given fand g in L 1 (JR.), wc havc
(i) j;g(~) = f(~) g(~) for all ~ E lR..
(ii) If in addition 1 and gare in L (JR.), then
1
Proof.
(i) f * g is in L 1 (JR.) by 20.2.1. The computation of j;g(~) is a direct
application of Fubini's theorem:
Le- i1r~x
2 f * g(x) dx = Le- i1r~x L
2 ( f(x- t)g(t) dt) dx
= L Le- i1r~x
g(t) ( 2 f(x- t) dx) dt
(ii) Note that (i) is true for 5 by changing i to -i. Since 1 and gare
in U(JR.), we can apply (i) and Proposition 23.1.1:
23.2.2 Remark With reference to the last proposition, note that the
formula r;g i = g does not make sense a priori, since f * g is only in
L 00 (!R). This formula is true whenever f * g is in L 1 (JR).
When f E L 2 (IR) and g E L 1 (IR), the convolution and the Fourier trans-
form are well defined, and we have the next result.
Properties:
Fourier transforms:
(i) a E C, Re(a) > 0, c: = 1, k = 0, 1, 2, ... ;
xk r c:
-e-wxu(c:x) ~ --:-------:-;---:--:-
k! (ca + 2i71'0k+ 1 '
-alxl .r 2a
e 1--------t a2 + 471'2~2 '
23.3.4 Convolution
f g f*9 Continuity
LI LI LI II! * 9III :::; IIJIIIII9III
LI Loo L n C0
00
II/ * 9lloo :S IIJih ll9lloo
Lz Lz L n C0
00
llf * 9lloo :S llfllzll9llz
Lz LI Lz II! * 9llz :S llfllzii9III
.9' .9' .9'
f E LI(~)}
for all ~ E ~'
gE LI(~)
f
j, E LI(~)}
for all ~ E ~'
g, g E LI(~)
/EL 2 (~)}
===} f * g(t) = (j. g)(t) for a.e. t ER
g E LI(~)
23.4 Exercises
Exercise 23.1 Compute N for f = X[o,IJ
23 . 2
E xerCISe Compute f *f when f(t) = sin27rAt .
Wlth).. > 0.
?rt
1
and deduce that
sin 2 x d _
-2- X -?r.
IR X
23.4 Exercises 207
Analog Filters
lesson 24
The tools we have just developed (convolution and the Fourier transform
for functions) are going to be used to study analog filters that are governed
by a linear differential equation with constant coefficients,
q p
is a filter.
Proof. We have proved the first part of the result and thus need only
to show that A is a filter on !7 . The linearity and invariance present
no difficulty. To prove continuity in the topology of !7 , suppose that a
sequence f n tends to 0 in !7 . Then fn
tends to 0 in !7 , as does Yn given
by (24.3). Thus Yn tends to 0 by Theorem 19.3.1. o
The differential equation (24.1) has a unique solution without initial
conditions being specified. This is because we require the solution g to be
in !7 , which means that g and all of its derivatives vanish at infinity.
We assume in what follows that P/Q has no poles on the imaginary axis.
Also, note that P '/=. 0, since we assume that ap =F 0.
(24.5)
24.2 Generalized solutions of the differential equation 213
This is the same kind of formula that we obtained in Section 2.4 for
the RC filter. The response is the convolution of the input with a fixed
function h that is called the impulse response. Note that if dcg P ~ deg Q,
thc computations we havc just made no Ionger make sense.
(24.10)
where z1, ... , Zq are the poles. From Section 22.2.2, read for !JT -l, we
conclude that
(24.11)
where
g" +w2g = J,
where P(x)fQ(x) = 1/(x2 + w2 ) has two poles are on the imaginary axis.
In this case h is a sinusoid and the Fourier transform of H (when H is
considered to be a function) is no Ionger defined. This problern will be
resolved in Section 35.2.3 in the context of distributions.
24.4 Stability 215
g = h1 * !' or g = h2 * !".
In the general case, we change the unknown function to g0 = g + >..j and
find that q q
I: bkgak) = I:(ak - >..bk)f(k).
k=O k=O
Taking >.. = aqjbq reduces the degree of the right-hand side and brings us
back to the case p < q. We can then write
g = >..f + ho * f. (24.14)
24.3.5 Summary
When P / Q has no poles on the imaginary axis and deg P ~ deg Q, a unique
generalized solution of (24.1) can be defined under the sole condition that
f E L 1 (JR) U L 2 (JR) U L 00 (lR). A(f) = g is a filter that we will call the
generalizedfilter A associated with (24.1). The output g is given by g = h*f
or possibly by a formula like (24.14).
24.4 Stability
24.4.1 Definition An analog system A: X-+ Y is said tobe stable if
there exists an M > 0 suchthat IIA/IIoo ~ Mllflloo for all f E L 00 (1R) nX.
216 Lesson 24. Analog Filters Governed by a Differential Equation
g(t) = Jo
r+oo h(s)f(t- s) ds
is 0 for t < t 0 when f(t) = 0 for t < t 0 . We prove the other direction by
contradiction. Thus suppose that there is a h < 0 such that h( h) > 0.
Since h is continuous at t 1 , there is an interval (a, b) suchthat b < 0 and
a < h < b implies that h(t) > 0. For the causal input signal
f(t) = X[O,b-aJ(t),
24.6 Gain and response time 217
g(t) = it
t-b+a
h(s) ds
with g(b) > 0. This contradicts the fact that Ais causal. This is the proof
when deg P < deg Q. In case deg P = deg Q, one uses the trick introduced
in Section 24.3.4.
From formulas (24.11) and (24.13) we see that supp(h) c [0, +oo) if and
only if K+ is empty. Thus if deg P:::; deg Q, we have the following result:
which is the ratio between the asymptotic value of the step response and
the height of the input step function. The response time is defined to be the
time it takes the step response to reach and maintain a certain percentage
of its limit, in general 95%:
with
r13 : 2 ~l.
and thus the real parts of the roots are not all negative.
(b) Q(x) = x 3 + (2k + 1)x2 + (k + 1) 2 x + k2 + 1 = 0.
The Routh matrix is
1 (k + 1) 2 0
2k + 1 k2 +1 0
2k(k 2 + 2k + 2)
0
2k + 1
k2 + 1
24.8 Exercises 219
For the elements in the first column all to have the same sign, we must
have 2k + 1 > 0 and 2k > 0. Thus the real parts of the roots of Q are
strictly negative if and only if k > 0.
24.8 Exercises
Exercise 24.1 Compute explicitly the output g of the generalized filter de-
fined by
g' - ag = j, a > 0,
and show that it is stable. Is it realizable? Compute the step response.
g" + 2ag' + bg = f
given a, b E R.
(a) Determine the regions of the (a, b)-plane where the poles of Q are not on
the imaginary axis.
(b) Determine the regions corresponding to a realizable filter.
(c) Show that the filter is unstable if b = 0.
Exercise 24.5 Compute the transfer function and the impulse response of
the generalized filter
g/11 + g = !" + f.
Is the filter stable? Realizable?
lesson 25
RCg' + g = J,
and
P(x) 1 1
Zl=--
Q(x) 1 +RCx' RC.
The filter is stable and realizable (fortunately!). Formula (24.11) shows that
1 __t_
h(t) = - e RCu(t)
RC
and
1
g(t) = RC
lt t-s
-oo e- RC f(s) ds.
h1 (RC)= 1- e- 1 ~ 0.63,
h1 (3RC) = 1 - e- 3 ~ 0.95.
h(t)
1
RC
0 RC
FIGURE 25.1. Impulseresponse of the RC filter.
u(t) h,(t)
95%
0 0 RC 2RC 3RC
FIGURE 25.2. Step response of the RC filter.
Thus
P(x) 1
Q(x) LCx 2 + RCx + 1'
25.2 The RLC circuit 223
and there are three cases to consider that depend on the sign of
The two poles are complex conjugates and have negative real parts:
z = -a + i and z = -a - i.
H(>.) = _1_ [ 1 _ 1 ]
iwC 2i1r>.-z 2irr>.-z '
v(t) = - 20
w
Jt
-(X)
e-a(t-x) sin(t- x)f(x) dx
h(t)
2
wC
...--
/
...... ......
/
2 /
-wC
FIGURE 25.4. Impulseresponse of the RLC circuit (R small).
224 Lesson 25. Examples of Analog Filters
R
(1 + 8- id) - - - - - - - - - ~---r--
0 !!. 21T
FIGURE 25.5. Step response of the RLC circuit (R small).
h (t ) = -1t e - !i.t ()
2L u t , (25.2)
LC
and
v(t) = L~ j_too (t- s)e- 2~ (t-s) f(s) ds.
The step response is
25.3 Another second-order filter: - ~ g" + g = f 225
0 s.!:.
R
FIGURE 25.6. Step response of the RLC circuit in the critical case.
Zt = --u,
R+w
Z2 = --u,
R-w
where w = J R2 - 4~.
H is decomposed as
H(>.) = __1 [ 1 _ 1 ]
wC 2i7r). - Zt 2i7r). - Z2 '
and
-1
h(t) = wC [eztt- ezt] u(t). (25.3)
()
htt
2L z t 2L z
= [ 1+Cw(R+w)el -Cw(R-w)e
t] ut.
()
The response is slower than in the critical case = 0. The gain is 1 in
all three cases. The RLC filter is stable and realizable.
so that
w2
H(>.) = 411'2)...2 +w2
From Section 18.2.2, the impulse response is (Figure 25.8)
h(t) = ~we-wltl.
2
Thus the output is
g(t) = ~w
2
r e-wlt-sl f(s) ds,
}IR
and the step response is (Figure 25.9)
if t :::;; 0,
if t ~ 0.
u(t) h,(t)
0 0
0 0
There is one pole at the origin, and the results of Lesson 24 do not apply.
If f is in .9 , one cannot in general find a g in .9 . It is easy to study this
directly: g is a primitive of f, and if we limit the search to causal signals,
then g is determined by having to be causal. In this case,
9= U* j.
IH(A)I h(t)
h(t)
-
IH(A)I
Ac> 0. (25.5)
and they occur in conjugate pairs. We want H(A) tobe a rational function
H(.X) = - 2i7r I ~1 2 [ 1 1 ]
p+ p 2i7r .X+ a - 2i7r .X+ a '
where a = a(1 - i). Referring to Beetion 22.2.2,
h(t) = -ia(e-at- e-at)u(t) = 2ae-at sinat u(t).
This impulse response has the same form as that of the RLC circuit,
which is equation (25.1).
IH(A)I
25.8 Exercises
Exercise 25.1 Show that it is possible to choose the constants R, L, and C
such that the RLC circuit is a Butterworth filter of order 2.
Hint: Take R = ..j2L7C and compute IH(.XW as in Section 25.2, First case. One
finds that
1
with .Xe=-----==
27r..;YC
R c
T
X (t) ( i (t)
r } V{t)
where x is the input and where the voltage v across the resistance r is the output.
(a) Show that x and v are related by RrCv' + (R + r)v = rx + RrCx'.
(b) Compute the transfer function and the step response.
(c) Assurne that rissmall with respect to R. What is the role of this filter?
Chapter VIII
Distributions
lesson 26
We are going to take a turn here that will lead to a new environment in
which signals are no Ionger modeled solely by functions. The two themes
for this heuristic introduction are impulse and derivation.
-h 0 h X
-h 0 h
We imagine that the duration of the force becomes shorter and shorter
(h ----> 0) while always imparting the same energy EJ to S. These applied
26.2 Uncontrolled skid on impact 237
forces become more and more intense, and in the limit we have an instan-
taueaus shock at time t = 0. If vh and 'Yh = v~ are respectively the velocity
and the acceleration of S, its kinetic energy at time t is
1:
Newton's second law, fh = m"(h, implies that
!h(t) dt = c.
By taking C = 1, we see that the forces fh satisfy the following three
conditions:
(i) fh(t) ~ 0 for all t ER
0 if ltl ~ h.
L
(ii) !h(t) =
(iii) fh(t) dt = 1.
At the limit, we will have a shock f(t) that has the following properties:
(i) f(t) ~ 0 for all t ER
(ii) f(x) = 0 if t i= 0.
(iii) Lf(t)) dt = 1.
the contradiction. They had a useful tool, even if it was not conceptually
satisfying. They used the imp(t) "function"-it was, in fact, called (t),
but we change the name temporarily for clarity-which, desirable or not,
was thought of as satisfying the conditions
if t # 0,
imp(t) = {O
+oo if t = 0,
kimp(t) dt = 1,
imp(t)
k
of the integral
I= imp(t)f(t) dt.
l x
-oo
.
1mp ( )
t dt = u ( x ) =
{0 if
.
1 1f
X<
X> 0,
0, (26.1)
which is quite natural in view of (26.1). (u(O) is not defined, but this is not
important.) Hereis the evaluation of I:
I=f(+oo)- Jo
r+oo !'(x)dx=f(+oo)-f(+oo)+/(0);
I= /(0).
This relation makes sense even if f is only continuous at the origin, and
it is thus possible to make practical use of integrals containing the im pulse
26.3 A new-look derivation 239
L
function by letting
imp(t)f(t) dt = f(O)
Du= imp.
u(t)
.../
1~8~~~/
'\. new
'\. derivation
0
0
-h 0 h usual
derivation
-h 0 h
From the modeling point of view, the unit step represents, for example,
the instantaneous establishment of a constant electric current. We consider
this phenomenon from a microscopic point of view, without going to the
level of electrons, where the model would necessarily be discrete. Physically,
there is no discontinuity at t = 0, but rather the continuous and very rapid
(of order w- 7 seconds) establishment of the current. A more precise model
would thus be a function uh(t) like the one shown in Figure 26.5.
For convenience, we put the time origin at the center of the transition
phase. The usual derivative u~, (t) must satisfy the following conditions:
(i) u~(t) ~ 0 for all t ER
(ii) u~(t) = 0 if itl > h.
(iii) l u~(t)dt = 1.
for all regular functions cp that have bounded support. This point of view
is taken in distribution theory, which was officially born in 1947 with the
publication of Schwartz's first article in the Annals of the Fourier Insti-
tute at Grenoble. Little by little the idea that all continuous functions
were differentiable spread throughout the mathematics community, to the
general amazement of all! While the theory at first seemed rather esoteric
and complicated (probably because of its heavy use of topology and the
Lebesgue integral), mathematicians quickly realized that its actual use was
much simpler than the theory: One could work formally and quickly with-
out worrying about whether functions satisfied certain conditions, such as
differentiability. A distribution is always differentiable, and in fact infinitely
differentiable. A series divergent in the usual sense will often be convergent
in the sense of distributions. One important property that we have already
verified for the Heaviside function is the continuity of derivation:
where the limits are taken in the sense of distributions. Finally, the Fourier
transform, an indispensable tool in so many areas, was until the advent
of distributions defined only for integrable and square-integrable functions.
One was not able to spcak, for example, of the Fourier transform of the
242 Lesson 26. Where Functions Prove to Be Inadequate
What Is a Distribution?
Tt('P) = l
-oo
+oo
f(x)cp(x) dx. (27.1)
This idea is analogous to that of identifying a real number a with the linear
function
x f--7 ax.
It is this concept that allows one to go from the notion of derivative to that
of differential.
Clearly, the integral in (27.1) does not always exist. If we want it to exist
for rather general functions f, it is necessary to impose severe restrictions
on the function <p, which is called a test function. We first require that <p
vanish outside a bounded interval so that there will be no problern with
convergence at infinity. Thus all test functions <p have bounded support.
To generalize derivation (or the derivative), we examine what happens
when f is continuously differentiable. The functional associated with f' is
T!'('P) = l
-oo
+oo
f'(x)cp(x) dx,
T!'('P) =- l +oo
-oo f(x)cp'(x) dx. (27.2)
This formula has the advantage that the derivative of f no Ionger ap-
pears, and thus the derived functional can be defined even though the
244 Lesson 27. What Is a Distribution?
In other words,
O(x) = {
exp ( - -1- )
1- x2
if lxl < 1, (27.3)
0
if lxl 2 1.
One can verify, with a little patience, that this function is infinitely differ-
entiable and that the derivatives all vanish at x = 1. (Show that the nth
derivative of 0 is of the form Fn(x)O(x), where Fn isarational function.)
It follows that ! =j:. {0}, but this is still a modest result.
27.3 The definition of a distribution 245
This second notation brings to mind a close relative of T(r.p), namely, the
scalar product in L 2 (1R) expressed by (27.1).
The continuity ofT means the following:
We still must specify the meaning of "r.pn -+ r.p in 9J"; thus we need to
define a topology (or at least the concept of convergence) for 9J. Here, and
elsewhere in the book, we settle for a direct definition of convergence---
the notion of a limit of a sequence---and avoid discussing the underlying
topology. We will define what 'Pn -+ 0 means; by linearity, 'Pn -+ r.p will
mean that 'Pn - r.p -+ 0.
There does not exist a distance function, much less a norm, on .! that
gives this notion of convergence. There is, however, a well-defined topology
on .! (sec [Sch65b)). It is suffi.cient for our purposes to have the notion of
convergence.
The following useful property follows directly from the definition.
It is clear that the set of distributions has the structure of a complex vec-
tor space with the obvious addition of two linear forms and multiplication
by scalars. This space is called the topological dual of .! or the space of
continuous linear functionals on .!. We denote this dual space by .! '(JR)
or simply .! '. This conforms with historical and customary notation.
27.3.4 Examples
(a) Point distributions: Let a be a real number and let 8a be the mapping
defined on .! by
8a('P) = cp(a).
+oo
T = L AnOna
n=-oo
+oo
T(cp) = L Ancp(na).
n=-oo
This sum is, in fact, finite for each cp; hence there is no problern about
convergence. Furthermore, T is continuous: If 'PP ~ 0 and if [A, B] is a
bounded interval containing the supports of all the 'Pv then
T(cpp) = L Ancpp(na)
A:o:=;na:o:=;B
is a distribution.
Proof. Tt is clearly linear. The continuity on !iJ follows directly from
(27.4): Let (a, b) be an interval containing the supports of the elements '{)n
of a sequence in !iJ that tends to zero. o
It is clear from (27.5) that if I and g are locally integrable and equal
almost everywhere, they define the same distribution Tt on !iJ; that is,
I= g a.e. ===} Tt = T9
We are going to investigate the converse of this implication.
Tt = 0 ===} I = 0 a.e.
(See Figure 27.1.) The proof of this property was given as an exercise
(Exercise 21.6). Thus for all I in Lfoc(R)
L/oc
0
FIGURE 27.2. Identification of Lloc as a subspace of .! '.
Tu(cp) = Jo
r+oo cp(x) dx.
The constant function f = K, K E C, is identified with the constant
distribution of the form
1 fcp = cp(O)
for all cp E .!JJ . We argue indirectly. Thus, assume that such an f exists and
let p E ..!JJ be the function defined by (27.3). Then we have
1 f(x)p(nx) dx = p(O)
27.5 Exercises
Exercise 27.1 Are the following functionals distributions?
(a) T('P) = I'P(O)I.
(b) T('P) = a, a E IC.
+oo
(c) T('P) = E 'P(n)(O).
n=O
Exercise 27.2 Suppose 'PE 9J (IR) and n E N*. Show that 'P can be written
in the form
L %, 'P(k)(O) + xn'I/Jn(x)
n-1 k
'P(x) =
k=O
with 1/Jn E C""(IR).
Hint: Use Taylor's formula with integral remainder.
250 Lesson 27. What Is a Distribution?
Exercise 27.3
(a) Take yj E C, 0 ~ j ~ n, and a ER We wish to find r.p E !iJ (JR) suchthat
(I)
for some A, B > 0. Fora E JR, show that there exists r.p E !iJ (JR) suchthat
r.pUl(a) = yj for all jE N.
Elementary Operations on
Distributions
la(x) = 1(-x).
The function I is said tobe even if Ia =I; it is said to be odd if Ia = -I
We will try to define the reflection Ta of a distribution T. For a regular
distribution Tf, the identification with the function I imposes the relation
The last formula is our guide for extending "reflection" to all distributions.
for all <p E !lf. The distribution T is said to be cven if Ta = T and odd if
Ta= -T.
f(x- a) = f(x)
Taf = J.
Thus it is sufficient to definc the translation operator for distributions
to establish the desired definition. For a regular distribution Tt, its idcnti-
fication with f forces the rclation
for all <p E !lf. The last formula can be extended to all distributions.
for all <p E !lf. A distribution T is said to be periodic with period a =f. 0 if
EXAMPLES:
We will assume (28.3). This is slightly more general than the condition
discussed in Section 27.4.2, which was that
Wehave
which we deduce directly from (27.2). We need to show that T' is a dis-
tribution. It is clearly a linear mapping defined on .!, since c.p' is in .!
for all c.p E .!. It is also continuous: If 'Pn --+ 0 in .!, then c.p~ --+ 0 in .!
(Proposition 27.3.3), and hence T'(c.pn)--+ 0.
Since T' is a distribution, it has a derivative T" : .! --+ C given by
and so on.
28.4.2 Examples
(a) The derivative of the point distribution E!a is given by
(Tu,c.p) = Jo
roo c.p(x)dx,
and hence
(T~,c.p) = -(Tu,c.p') = -
r+oo c.p'(x)dx = c.p(O) =
Jo (fJ,c.p).
-l l
shows that
(Tj,cp) = fcp' = f'cp = (Tt,cp'),
and hence that
Tj = Tf'.
The identification of a locally integrable function with its associated dis-
tribution leads to the following result: For a function that is absolutely
continuous an all compact intervals [a, b], the derivative of f in the sense
of distributions agrees a. e. with the usual, or ordinary, derivative. For ex-
ample,
f(x) = lxl
and
J'(x) = sign(x) for a.e. x.
In the general case of a locally integrable function J, this derivative-
whether or not it is a function-will be denoted by f' when there is no
ambiguity. The derivative in the sense of distributions is expressed by the
relation
For example,
u' = 8. (28.9)
j +oo
(Tj,cp)=- -oo fcp'=-
1a
00
/(x)cp'(x)dx-
lb
a f(x)cp'(x)dx
r+oo
- Jb f(x)cp'(x) dx.
28.4 The derivative of a distribution 257
f(x)
b X
f'(x)
(Tj, cp) =- f(a- )cp(a) + /_~ f'(x)cp(x) dx- f(b- )cp(b) + f(a+ )cp(a)
+ la
r f'(x)cp(x) dx + f(b+ )cp(b)
r+oo f'(x)cp(x) dx.
+ Jb
This can be written
(28.10)
258 Lesson 28. Elementary Operations on Distributions
-a 0 a 2a 3a X
FIGURE 28.2.
We are now going to see what can be clone, from the point of view of
distributions, with certain functions that are unbounded and not integrable
in the neighborhood of a point.
the same problern with any rational function having a real pole. This is
something of a wrench in our distribution machinery, for one of our main
goals with distributions is to extend the notion of function. Furthermore,
the functions in question often arise in practical problems. We will see here
how this problern can be resolved. Although f is not locally integrable, its
primitive F(x) = ln(lxl) is locally integrable. The distribution that comes
to our aid is simply the derivative of F in the sense of distributions. This
distribution is well-defined (Section 28.4.3) and is denoted by pv(.!), where
X
pv stands for "principal value." But what is the value of
(pv(~),<p)?
Wehave
(Tj..,<p) = -(Tp,<p1) = -faln(ixl)<p'(x)dx.
with
JE=-/_: ln(lxl)<p'(x) dx -l+oo ln(ixl)<p'(x) dx.
Integrating by parts shows that
shows that the integrated term tends to 0 as c --+ 0. Thus we see that
(pv (.!.), <p) = lim { <p(x) dx = lim 1+oo <p(x) - <p( -x) dx.
X E-+O lixi?_E X E-+O E X (28.11)
does not exist in general, but taking symmetric limits (-c- and c) around
the origin guarantees the existence of the limit (28.11) as c--+ 0. This is a
particular case of what is called the "principal value" of an integral.
One easily deduces from (28.11) that
xpv(~) = 1. (28.12)
260 Lesson 280 Elementary Operations on Distributions
(fp ( 21 ) , cp) = hm
X E-+0
0 1 lxl?_c:
-cp(x)
X
2- dx,
since this time the limit generally does not existo Again, the plan is to
consider a primitive of g, but of a higher ordero The function F(x) = ln(lxl)
is, up to a sign, a second primitive of g, so we define
fp(:2 ) = -F",
where the derivative is taken in the sense of distributionso The new distri-
bution is thus the negative of the derivative of pv (.!):
X
(28014)
(fp(~), cp) =
X
lim
c:-+0
l+oo cp'(x)- cp'( -x) dxo
E X
+ 2 X o
E X
28.6 Exercises 261
(fp ( _!.._)
2 , <p
) = 1"1m 1+oo <p(x) + <p( -x)- 2<p(O) 2
d
X.
X E->0 E X
(28.15)
x2 .fp ( :2 ) = 1. (28.16)
In summary:
(1)
(pv - , <p) =-
1 1n lxl<p'(x) dx = lim 1+oo <p(x)- <p(-x) dx,
-11 I I "( )
X IR E->0 E X
(fp ( _!.._)
X
)-
2 , <p -
IR
n X <p X d X -- 1"Im
E->0
1+oo <p(x) + <p( -x)- 2<p(O) d
E X
2 X.
28.6 Exercises
Exercise 28.1
(a) Compute the successive derivatives in the sense of distributions of the fol-
lowing functions:
(o:T)(n) = ~ (~)o:(k)T(n-k).
Exercise 28.4 Campare (T')a and (Ta)'. Compute (8')a and (8a)'.
Exercise 28.5 Suppose f E Lfoc(JR). Show that F(x) = 1x f(t) dt, a E JR,
has f as its derivative in the sense of distributions.
Exercise 28.7 For f : ~ ---+ C measurable and >. E ~*, the operator h>-. is
defined by
h>-.f(x) = f(>.x), x E ~.
Extend this definition to distributions.
Exercise 28.8
(a) Define the conjugate T of a distribution TE~~(~).
(b) Define Re(T) and Im(T). Characterize a distribution as being real or imag-
inary.
x
(e -1)pv ;
( 1) = -x-.1
ex-
pv(~) = S + f,
where S is a distribution with bounded support and f E L 2 (~).
(2) Show that one can find T, U E !. 1 (JR) such that 0 E supp(U) and xT = U.
Exercise 28.15 Take b E Lfoc(JR) and define a by a(x) = xb(x).
(a) Solve the equation xT = a, TE!. 1 (1R).
(b) Apply this to a(x) = sinx.
Hint: For (a), use Exercise 28.11(c) to show that T = b + K8, K E IC.
Exercise 28.17
(a) Assurne <p E !. (JR) with supp(<p) C R\{0}.
Show that there exists an a > 0 suchthat <p(x) = 0 for all x E [-a, a].
Deduce from this that for each k E N there exists 'ljJ E !. (JR) such
that <p(x) = xk'lj;(x) for all x.
(b) Show that
Hint: If <p E D(JR) is suchthat supp(<p) c lR\{0}, by (a) there exists 'ljJ E !. (JR)
suchthat <p(x) = xn'lj;(x) for all x. Then (T, <p) = (T, xn'lj;) = (xnT, 'lj;) = 0, and
hence supp(T) C {0} ).
264 Lesson 28. Elementary Operations on Distributions
**Exercise 28.18 Let g E C""(JR) and Iet its derivative be odd and bounded.
(a) For x < 0, show that the integral
1 A
x g'(t) dt
t , A < 0,
has a finite Iimit as A ---> -oo. Let f be the even function defined for x < 0
1"'
by
f(x) = g'(t) dt = lim {"' g'(t) dt.
_ 00 t A-+-oo}A t
(b) Show that
lf(x)l:::; C(l -log lxl) for 0< lxl:::; 1.
Deduce that f E Ltoc(JR).
(c) Show that the derivative of f in the sense of distributions is
j'(x) = g'(x)pv(;).
lesson 29
Convergence of a Sequence of
Distributions
We saw in Beetion 26.2 that from the point of view of physies, the impulse
is a limit. For these and other reasons it is important to investigate the
notion of limit in ! '.
for all cp E ! .
29.1.2 Examples
(a) Ifthe real sequenee (an)nEN eonverges to a, then Dan eonverges to Oa,
sinee cp( an) ---. cp( a) for all cp in !.
(b) Consider the sequenee of functions Un)nEN defined by
fn(x) = sin27rnx.
For fixed x (not an integer or half-integer), the sequenee fn(x) does not
eonverge as n ---. +oo. Thus the sequenee of functions Un) does not tend
266 Lesson 29. Convergence of a Sequence of Distributions
for all r.p E .!. We conclude that the sequence Un) tends to 0 in !lJ 1 , and
we have the following negative result:
fn __.. f in.! 1 ~ J J fnr.p __.. fr.p for all r.p E !lJ. (29.1)
We wish to see whether this sequence has a limit in the sense of distribu-
tions. For r.p E .!, the mean value theorem implies that
for some cn, -1/n < Cn < 1/n. Asn-.. +oo, r.p(cn)-.. r.p(O); hence,
fn-.. 8 in .! 1
Although we have taken a particular form for the functions fn, this argu-
ment is easily generalized to more general sequences, and it shows that the
29.3 Relations with the convergence of functions 267
1 X
4
FIGURE 29.1. The impulse as the limit of functions fn
point distribution at the origin is the model of the impulse that we sought
in Lesson 26.
0
268 Lesson 29. Convergence of a Sequence of Distributions
(29.2)
n=-oo
where the an do not necessarily tend to 0. This series is not, a priori, the
Fourier series of a function, and furthermore, it is not, in general, pointwise
convergent. Nevertheless, this series converges in !JJ' under rather general
conditions and thus defines a distribution. Recall that the series (29.2)
converges if the sums
N
!N(x) = L
n=-N
L
N (
:7r
2 . )k+2O:ne2i11'nax ---+ p(k+2)
n=-N
n#O
as N ---+ +oo. This proves that the series {29.2) converges in the sense of
distributions to the distribution
T = ( 2;7r) k+2 p(k+2) + a:o. o
n=O
Thus
+oo
(Lla, <p) = L
n=-oo
<p(na)
for all <p E !lJ , and it is clear that Lla is periodic with period a.
Y= f(t)
_.-"
- ----1"'----i', /
/
/
/
"'
' 1',
''
'
t'
-2a -a 0 a 2a 3a
f(t)
Writtcn as distributions,
+oo
flla = L f(na)Ona
n=-oo
+oo +oo
f ~ L
n=-oo
f(na) Tna X[-~~)~
2' 2
L
n=-oo
f(na)Ona, or f ~ aflla.
272 Lesson 29. Convergence of a Sequence of Distributions
f(t)
-a 0 a 2a X
FIGURE 29.5.
! '( x ) -
_ - -1 "'"'~e
2i7rn:!C
a.
a
nof-0
n=-oo
The series on the right diverges in the sense of functions, but it converges
in the sense of distributions to Dirac's comb. This series is the Fourier series
expansion of Dirac's comb, and it illustrates the general situation described
in Section 29.4.
29.6 Exercises
Exercise 29.1 Let T be a distribution and let (hn) be a sequence of real
numbers that tends to 0 (hn =/: 0). Show that h1n (T- ThnT)--+ T' in !'(IR) as
n--+ +oo.
Exercise 29.2 Show that every sequence of functions (in) that satisfies the
following three conditions converges to 6 as n --+ +oo:
fn E L 1 (1R) and JIRfn(x)dx = 1.
fn :::=: 0 a.e.
fn = 0 a.e. on ( -oo, -1/n) U (1/n, +oo).
.
l1m
x-+oo
1"' 0
sinx
- -
X
d x=-.
1r
2
L
00
Tn = Ak(n)Cka
k=-oo
Show that Tn tends to 0 in ! '(IR) if and only if for all k E Z,
lim Ak(n)
n-oo
= 0.
Exercise 29.6 Assurne that f: IR--+ Cis piecewise continuous on all closed,
bounded intervals.
(a) Prove that the sequence of distributions
n=-oo
sN = ~ L
n=-N2
J(~)c*.
274 Lesson 29. Convergence of a Sequence of Distributions
Hint: Use the Fourier series expansion of the function with period 2 defined on
( -1, 1) by f(x) =x (see Section 29.5.3).
Exercise 29.8 Show that the sequence of functions fn(x) = ne-1rn 2 "' 2 tends
to 6 as n -> +oo.
Hint: First compute 1fn(x) dx knowing that 1 e-rrx
2
dx = 1; then show that
Exercise 29.9 Let (an)nez be a real sequence and let (.An)nez be a complex
sequence. Give a sufficient condition on (an) that implies that
n=-oo
is a distribution.
Primitives of a Distribution
For cp E .!, define '1/J"' = cp- I(cp)O, where I(cp) = JR cp. Then
and so
f(x) = f(O) +lax g(t) dt.
Clearly,
Fcp' cp and (Fcp)' = cp- I(cp)O = 1/J.
=
Since the derivative of Fcp is in! 0 , we know that Fcp is in!.
Now define U : ! --4 C by
(U, cp) = -(T, Fcp)
(At this point we do not know that U is a distribution, so writing (U, cp) is
a slight abuse of notation.) If U is a distribution, then it is a primitive of
T, since
(U',cp) = -(U,cp') = (T,Fcp') = (T,cp).
To prove that U is a distribution, it is suflicient to show that the mapping
cp f-t F cp is linear and continuous from ! to ! .
The mapping is clearly linear, so we focus on continuity. Let (cpn) be a
sequence in! that tends to 0. The sequence of integrals (I(cpn)) tends to
0; hence
(30.4)
tends to 0 in !. If [A, BJ is an interval containing the supports of all the
1/Jn, then the supports of the functions
30.3 Exercises
*Exercise 30.1
(a) Solve the differential equation
T' + aT = 0 (a # 0)
*Exercise 30.2 Use Exercise 28.1 to derive the general solutions of the fol-
lowing differential equations:
*Exercise 30.3 For -oo < a < b < +oo we consider the space
where the derivative is taken in the sense of distributions. Show that W 1 1 (a, b)
is the space of absolutely continuous (AC) functions on [a, b].
Remark: If f E W 1 1 (a, b), one shows that f has a unique representative in the
space of absolutely continuous functions by using g(x) =I:
f'(t) dt, x E [a, b].
**Exercise 30.4 (the Sobolev space H 1 (a, b)) Let (a,b) an inter-
val, bounded or not. The vector space H 1 (a,b) is defined by
*Exercise 30.6 Show that if f E H 1 (1R), then lim 1"'J-oo f(x) = 0 (apply the
argument used in Proposition 17.2.1(ii) to the function f ). Verify that f E H 1 (1R)
does not imply f E L 1 (IR).
Chapter IX
This suggests that the Fourier transform for distributions should be defined
by transposition:
(T, r.p) = (T, (/5). (31.1)
We know that the expression (T, (/5) make sense when (/5 E !fJ (lll), and we
have seen that (/5 E coo (Proposition 17.2.1). But does (/5 have compact
support? We will see in Proposition 31.5.4 that this is never the case. We
have, however, shown that the space of rapidly decreasing functions .5I' (lll)
is invariant under the Fourier transform (Theorem 19.2.3). This leads to
the introduction of the subspace of tempered distributions.
31.1.1 Definition .5I' '(lll) denotes the vector space of continuous lin-
ear functionals T defined on .5I' (lll). Thus
'Pn ~ 0 in .5I' ===} (T, 'Pn) ~ 0 in C.
284 Lesson 31. The Fourier Transform of Distributions
If we take cp in !lJ (ffi.), then T(cp) is well-defined, since !lJ (ffi.) C SC (ffi.)
(Figure 31.1). Furthermore, convergence in !lJ (ffi.) implies convergence in
SC (ffi.) (see Definitions 19.2.4 and 27.3.2). This means that the elements
of SC '(ffi.) restricted to !lJ (ffi.) are distributions. Since !lJ (ffi.) is dense in
SC (ffi.) (Exercise 19.7), we can identify SC '(ffi.) with a subspace of !lJ '(ffi.)
(see Exercise 31.1).
FIGURE 31.1.
Proof.
(i) The mapping cp ---+ xkcp is continuous in .Y (~) (Proposition 19.2.5).
We have (xkT, cp) = (T, xkcp) for all cp E ~ (~). If a sequence 'Pn E ~ (~)
tends to 0 in .Y, then Xk'Pn---+ 0 in .Y. Thus (xkT, 'Pn) ---+ 0, which shows
that xkT E .Y 1 (~) (Proposition 31.1.3).
(ii) Take 'Pn E ~ (~) suchthat 'Pn ---+ 0 in .Y. The mapping cp---+ cp(k) is
continuous in .Y (~)(Proposition 19.2.5). Since (T(k), 'Pn) = ( -1)k(T, cp~k)),
we can pass to the limit and conclude, as in (i), that T(k) E .Y 1 (~).
(iii) Suppose that Tn is in .Y 1 (~) and that Tn---+ Tin .Y 1 (~). Then for all
cp in .Y (~), (xkTn, cp) = (Tn, xkcp) -+ (T, xkcp) = (xkT, cp). This proves the
continuity ofT f---t xkT. The same technique is used to show that T f---t T(k)
is continuous in .Y 1 (~). o
2 2N 'Pn(x)l
::;Csupl(1+x) ( dx )N"
2
xEIR IR 1 +X
is a tcmpered distribution.
Proof. Wehave already seen that T is a distribution (Section 27.3.4(b)).
Let 'Pv be a sequence in !lJ (IR) that tends to 0 in Y (IR). We must show
that limv_,oo I(T, 'Pv) I = 0.
Since (Yn)nEZ is slowly increasing (see Definition 29.4.1), there is an
N E N suchthat IYnl ::; C(1 + n 2)N for all n E Z (take any N ;::: k/2 for
the k in (29.4)). With this we have
+oo
where c2 = Cl L 1 + tan)2. By hypothesis, p~~ Mp = 0, and this
n=-ex>
proves that TE .9' '(IR). o
+oo
As an example, Dirac's comb a = L 8na is a tempered distribution.
n=-oo
To finish this section, we state without proof the structure theorem for
tempered distributions [Sch65b, Kho72].
T = Lfknk)
k=l
Proof.
(i) Since T E Y '(IR), xkT E Y '(IR) (Proposition 31.1.5(i)). Thus ;;kT
exists, and we have (n, cp) = (T, xkrp) for all cp E Y (IR). From Proposi-
tion 17.2.1(ii) we have xkrp = -.1(
)kcp(k); hence
2m
Proof.
(i) Take cp E Y (IR). By Proposition 17.2.3(ii),
(5 (Tl)"), cp) = (Trr, 5 (cp)) = (T, (5 cp)l)") = (T, 5 cp) = (5T, cp).
and 5 (5 T) =Tu D
290 Lesson 31. The Fourier Transform of Distributions
hence
{31.4)
In the same way,
so
{31.5)
We note that the Fourier transform of a is a C 00 function. For the deriva-
tives of the Dirac impulse we have {Proposition 31.2.4{i))
{31.6)
and by taking the Fourier transform of both sides of {31.6) we see that
;k = 1 {k) {31. 7)
( -2i7r)k .
The Fourier transform of a polynomial (which cannot be computed by
integration because xk fj L 1 {IR)) is a linear combination of the derivatives
of the Dirac distribution at the origin.
is continuous,
n=-oo n=-oo
thus
-
a 1
= -1. (31.10)
a - a
The Fourier transform of the Dirac comb with grid a is a- 1 times the
Dirac comb with grid a- 1 . For a = 1, it is a distribution equal to its
Fourier transform, which is a property shared by the Gaussian g(t) = e-1rt 2
(Proposition 1(.3.2).
for all r.p E C 00 (IR). iif '(IR) isalinear subspace of .9 '(IR) [Kho72].
292 Lesson 31. The Fourier Transform of Distributions
31.4.2 Examples
(a) A function f E Lfoc(IR) with compact support is in g '(JR.).
(b) Let Oa be the Dirac distribution at a. For all p E N, oip) is in lf I (JR.).
Warking with the elements of g '(JR.) is facilitated by knowing their struc-
ture; thus we are going to assume the following theorem [Sch65b, Kho72].
T= L Jti>.
j=l
31.5.1 Theorem IfT is in lf '(JR.), then T and all ofits derivatives are
slowly increasing functions in C 00 (1R.). Furthermore, f<k>(e) = (T, 'Pkk)) for
k = 0, 1, 2, ... , where 'Pe(x) = e- 2i71"Xe; that is,
Proof. We use Theorem 31.4.3 and write T = L:;~=l J}ni). Since J}ni) is
in g '(JR.) for j = 1, 2, ... ,p, we have
p
(Tx, e-2i71"Xe) = L ut;)' e-2inxe)
j=l
= I)-1)n; 1(-2i7re)n;e-2inxe/j(x)dx
j=l IR
p
= L(2i1re)ni jj(e).
j=l
= t
J=ljR
r Jtj)(x) (-2i7rx)k ( jRr e-2i1rxer.p(0 d~) dx
= 1r.p(~) t 1
R
(
J=l R
Jt 1 )(x)( -2i1rx)k e- 2 i7rXe dx) d~,
dj .
IT(k)(~)l ~ C sup sup 1-. (( -2i7rx)ke- 2'7rXe)l,
j'.S_m xEK dxJ
Proof. We know that f(~) = T(~) is the C 00 function f(O = (Tx, e- 2 i1rxe).
Define [(z) = (Tx, e- 2i1rxz) for z E IC. One shows by direct computation
1
(as above) that is holomorphic on IC, i.e., it is infinitely differentiable on
IC. The inequality follows from the continuity ofT. o
31.6 Formulary
(i)
(ii)
~: !T -2i1rae
(iii) ua 1--+ e
. !T
e2t11"Xa I--+ 6a
6(k) ~ (2i7r{)k
xk ~ ( -2i11")-k6(k)
(iv) u(x)
!T
I--+
1 + 2i11"1 pv (1)
26 e
sign(x) ~ i~ pvG)
pv(~) ~ -i1rsign({)
(v) .!T (Y T) =Tu
(Y T)u =Y Tu =Y T
Y Y T=Y .!T T=T
31.7 Exercises
Exercise 31.1 (.9' '(JR) as a subspace of ! '(JR))
ForT E .9' 'R) we write j(T) = 11 y (IR.) to denote the restriction ofT to !lJ (R).
(a) Show that j(T) E !IJ'(R).
(b) Show that j: .9' '(R) 1-+ !ll'(R) is injective (j(T) = 0 implies T = 0).
(c) Show that j is continuous.
31.7 Exercises 295
*Exercise 31.2
(1) Suppose that f E SC (R) and that TE SC 1 (R). For <p E SC (R) we define
(JT, <p) = (T, f'P).
(a) Show that JT E SC 1 (R).
(b) Show that the mapping (!, T) >--> fT from SC (R) x SC 1 (R) to SC 1 (R)
is continuous with respect to each variable.
Exercise 31.4 Show that log lxl is a tempered distribution. Deduce from
this that pv ( _!) and fp (-.;.) are also tempered distributions.
X X
Exercise 31.7 Compute the Fourier transform of fp (-.;.). Use this to com-
x
pute the Fourier transform of lxl.
~ 1 (1) 1- e2rrl<l
!(~) = 2ipv "l - 2i~
296 Lesson 31. The Fourier Transform of Distributions
Exercise 31.10 1
(a) Compute the Fourier transform of arctan- from Exercises 31.6 and 31.9
X
and the formula
(b) Prove this result starting with the derivative of arctan .!.
X
and proceeding
as in Exercise 31.9.
Convolution of Distributions
= k(h.g(x-u)cp(x)dx)f(u)du. (32.2)
From this it appears that one should study the quantities fu * cp and Yu * cp
when f and g are distributions.
Proof.
(i) When cp E !i! (~), the expression (rxT, cp) makes sense for all x E R
We wish to show that the function 'lj;(x) = (rxT, cp) is differentiable. Thus
Iet hn be a sequence of nonzero reals that tends to 0 as n ~ oo. Define
1
O'.n(Y) = hn [cp(y +X+ hn)- cp(y + x)];
then
1
hn ['1/J(x + hn)- 'l/;(x)] = (T, O'.n)
(iii) Again, (rxT,<p) makes sense because <p E C 00 (1R.) and TE ~'(IR.). To
prove that '1/J is differentiable it is sufficient to show that ahq) converges
to (r_xt.p')(q) uniformly on all compact subsets of IR.. This is done using
inequalities similar to those used in the proofs of (i) and (ii). o
coo *~'.
32.1.3 Proposition (convolution Y *
Y ') Assume<p E Y (IR.)
and T E Y '(IR.). The convolution <p * T and all of its derivatives are slowly
increasing coo functions.
Proof. By definition t.p*T(x) = (Ty, <p(x-y)). We use Theorem 31.1.10 and
write T = E~=l fknk), where the continuous slowly increasing functions fk
satisfy lfk(x)l:::; Ck(l +x 2 )Nk. Then
= t
k=l Ja
rfk(x- y)<p(nk)(y) dy,
and
is
The results of this section show that the convolution of a distribution and
a C 00 function is a smoothing operation. We will see in the next section,
where we introduce the convolution of two distributions, that !(IR) is even
dense in !'(IR).
32.2 The convolution if 1 * .9J 1 301
32.2.1 Theorem *
(~ 1 !lf ') Assume SE&' '(IR) and TE !0'(IR).
(i) There exists a distribution called the convolution of S and T and
denoted by S * T such that for all <p E !0 (IR),
(S * T, <p) = (St, (Tx, <p(x + t))) =(Tu, (Sx, <p(x + u)) ). (32.7)
(ii) The mapping (S, T) f-> S * T from ?f '(IR) x !0'(IR) to !0'(IR) is con-
Proof.
(i) Since A is compact, A + B is closed. Let n = lR \ (A + B) and take
<p E ~ (JR) with supp(<p) c 0. We will show that (S * T, <p) = 0. Wehave
(S * T, <p) =(Tu, (Sx, <p(x + u))) and 7f(u) = (Sx, <p(x + u)) = S * <p 17 ( -u).
Thus we wish to show that supp( 7f) n B = 0.
If u E supp( 7f) n B, then -u E supp( 'Pa) + supp( S) (recall that by
Proposition 32.1.6 supp(S * 'Pa) C supp<p 17 + suppS). This means that
-u = y + x with -y E supp(<p) and x E supp(S). But then -y = u + x
with u E B, x E A, and hence
supp(<p) n (A + B) -j. 0,
for some constant C. This shows that 'Pn ~ 0 in Y' (IR) implies that an
converges to 0 in Y' (IR). o
We use this result to prove the next one.
32.3.3 Proposition
(i) Let Sn be a sequence in iif 1 (lR) that converges to 0 in iif 1 (lR); that is,
(Sn, r.p) ---+ 0 for all r.p E C 00 (JR). Then Sn * T ---+ 0 in Y 1 (lR), and
hence in~ 1 (lR), for all TE Y 1 (JR).
(ii) Let Tn be a sequence in Y 1(JR) that converges to 0 in Y 1 Then for
all SE iif 1 (lR), S * Tn---+ 0 in Y 1 (JR), and hence in~ 1 (JR).
Proof.
(i) By Proposition 32.1.1, the function 1/J(t) = (Tx,r.p(x+t)) is in C 00 (lR)
for all r.p in Y (JR). Thus
(ii) Similarly, a(u) = (Sx, r.p(x+u)) is in Y (JR) for all <p is Y (JR) (Propo-
sition 32.3.1). Hence limn-+oo(S*Tn,r.p) = limn-+oo(Tn,r.p) = 0. 0
32.3.4 Proposition
(i) Let Sn be a sequence in iif 1 (lR) that converges to 0 in~ 1 (JR). Assurne
that there exists a compact set K such that Sllpp(Sn) C K for all n.
Then Sn* T---+ 0 in~ 1 (JR) for all TE Y 1 (JR).
(ii) Let Tn a sequence in Y 1 (JR) that converges to 0 in ..2? 1 (JR). Then for
all SE iif' 1 (lR), S*Tn---+ 0 in ~ 1 (lR).
Proof.
(i) Take <p E ..2? (JR). The function 1/J(t) = (Tx, r.p(x + t)) is in C 00 (lR).
Since supp(Sn) C K, (Sn, 1/J) = (SnJN), where () is a function in ~ (JR)
suchthat O(x) = 1 for x E K. Then 01/J is in~ (JR) and (Sn, 01/J) ---+ 0.
(ii) If <p E ~ (JR), then the function a(u) = (Sx, r.p(x + u)) is in ~ (JR),
and hence (Tn, a) ---+ 0. 0
32.4.1 Proposition Suppose that TE.!'~ and r.p E C 00 (!R) and that
supp(T) C [a,+oo) and supp(r.p) C (-oo,b]. Then (T,r.p), defined by
(T, r.p) = (T, Or.p), (32.11)
where () is a function in.!' (IR) equal to 1 on an interval [-M, MJ containing
a and b in its interior, is well-defined.
Proof. ()r.p E .!'(IR), so (T, Or.p) makes sense. We must show that the defini-
tion of (T, r.p) does not depend on the choice of 0. Let 01 be another function
in .!'(IR) equal to 1 on [-Mll M1] containing a and b. Then (()- OI)r.p van-
ishes on [-m, +oo), where m = min{M, M 1}. Since supp(T) c [a, +oo),
we have supp(T) n supp((O- OI)r.p) = 0 and (T, (()- 01)r.p) = 0. o
To define the convolution, it is necessary to give meaning to the expres-
sions (St, (Tx, r.p(x + t))) and (Tu, (Sx, r.p(x + u))) for S and Tin .!' ~ and
r.p in .!'(IR).
32.4.2 Proposition Suppose that TE .!' ~ and r.p E C 00 (!R) and that
supp(T) C [a, +oo) and supp(r.p) C ( -oo, b]. Then 'lj;(t) = (Tx, r.p(x + t)) is
defined of all t E IR, supp('lj.;) C ( -oo, b- a] , and 'lj; E C 00 (IR).
Proof. The function Lt'P is in C 00 (!R) with supp(Ltr.p) c (-oo, b- t] for
all t E IR. Thus by Proposition 32.4.1, (Tx, r.p(x + t)) is well-defined.
Now, 'lj;(t) = 0 if supp(r-tr.p) n supp(T) = 0, which is the case when
b - t < a. Hence, supp('lj;) c (-oo, b - a]. That 'lj; E coo is a consequence
of (32.11) and Proposition 32.1.1(i). o
These preliminary results lead to the next theorem, which we state with-
out proof (see, for example, [Sch65b]).
Proof. Take <p E .21 (IR.) with supp(<p) C ( -oo, a 1 + a2). The support of
= (Tx, <p(x + t)) is in ( -oo, ai) by Proposition 32.4.2. Thus
"P(t)
supp("P) n supp(S) = 0
and (S * T, <p) = 0, which proves that supp(S * T) c [a1 + a2, oo). D
P(T) = U. (32.13)
L aml5(m) * T = u. (32.14)
m=O
A distribution E is said to be an elementary solution of (32.14) if
S * T1 = S * T2 = U.
308 Lesson 32. Convolution of Distributions
32.6 Exercises
Exercise 32.1 Let Sand T be two even (or two odd) distributions. Show
that S *T is even.
Exercise 32.4 Suppose f(x) = (1- x)u(x) and g(x) = exu(x). Show that
f * g makes sense and compute this convolution.
Exercise 32.6 Keeping in mind the last exercise, show that it is possible to
find an elementary solution for the operator
dm dm-1 d
P = - - +am-1--
dxm dxm- 1
+ +a1-
dx
+ao
Exercise 32.7
(a) If At and A2 have inverses in! t (I~), show that At* A2 is invertible. What
is the inverse of At * A2?
(b) Use this result to find the elementary solution of the operator
d2 d 2
p = dx 2 - 2 ).. dx + )..
Exercise 32.8 What are the inverses in !Pt (I~) of u, 15', and 15'- a/5?
Lesson 33
---
As is the case for functions (Lesson 23), the Fourier transform interchanges
---- --
convolution and multiplication of distributions. We wish to determine under
what conditions the relations T * U = T U and T U = T...... * U are true.
The first thing to notice is that one must be careful manipulating these
-
relations, since the product of two distributions is not generally defined.
We faced a similar problern with the convolution in the last lesson. There
we were able to establish several conditions under which the convolution is
well-defined and consistent with the convolution for functions.
Proof.
(i) 1/J * T is in Y '(IR) by Proposition 32.1.3. For all r.p E Y (IR),
-
Applying Proposition 23.1.2(ii), we see that
~
;j 'P = :(}j * fi5 = 1/Ju * cp = (1/Jx, cp(x + u));
hence -:;;;T = ;j T.
(ii) ;j E !? (IR) and T E !? '(IR) imply that ;j * T is defined and is in
!? '(IR). Applying the operator .!T to (i), we have
<p
~
= S * cp = Su * cp.
Thus
-
(T, S <p) =(Tu, (Su(x), cp(u- x)))
=(Tu, (Sx, cp(u + x)))
= (T * S, cp)
= (T;S,<p),
0
33.3 The Fourier transform and convolution L 2 * L2 313
.;._ pv(!) = S
~7r X
+g
with SE lt'(IR) and g E L 2 (1R) (see Exercise 28.10). Equation (33.4) is
1
valid because Sigii * makes sense, and
A=~pvG)*B, (33o5)
Proof.
(i) Wehave just seen that His well-defined on L 2(!R)o H(f) is in L 2(1R),
since
and f E L2(!R)o His clearly linear; from (3307) and Theorem 22ol.4(iii) we
have
IIHJII2 = IIHJII2 = llfll2 = llfll2,
which proves that H is an isometryo To find the inverse, apply (33o7) to
H(f):
5 [H(H(f))(~)] = -i sign(~)ii(J)(~) = - f(~)o
Thus H(H(f)) =- f for all f E L 2(1R), and H- 1 =-Ho
(ii) This is a restaterneut of (33o5)o 0
G= j. u.
33.5.1 Definition The signal whose Fourier transform is 2G, which is
the signal g = 25 G, is called the analytic signal associated with the real
signal f.
From the definition we have
hence
= f +iHf.
g
33.6 Exercises
Exercise 33.1 Compute the Fourier transform of f(x) = cos ~x X[-l,lJ(x)
Exercise 33.2 Show that the convolution pv(~) * pv(~) makes sense and
compute it.
Exercise 33.3 Compute the Fourier transform of pv(~) from the relation
xpvG) = 1.
Tn = Tn-1 * T1, n 2 2.
(c) Define fn(e) = Tn( 2 1l"~;rJ Show that fn converges in .'7 '(R) and find its
limit.
(d) Study the convergence of T!n in .'7 '(R).
2
Remark: Exercise 33.4 shows that any function of the form e-ax , a > 0, is the
limit in the sense of ~ '(R) of a sequence of finite linear combinations of Dirac
distributions. In fact, this is an immediate result of Exercise 28.7 and the fact
that forT E .'7 '(R) and A E R\0, 5 (h>..T) = l~l ht.!T T.
Filters for functions have been studied in Lessons 1, 2, 24, and 25. We
are going to recast and complete this analysis in the light of what we now
know about distributions. We will see that the basic tools developed so far,
namely, convolution and the Fourier transform, play an essential role in the
study of generalized filters, in the same way they did in the study of filters
for functions.
Proof. In case (i), A is well-defined for any subspace X C !lf ', and it is
linear. The invariance comes from Propositions 32.5.1 and 32.2.2: For all
a E JR,
and in terms of A,
Ta(Af) = A(Taf)
for all f E X. A is continuous by Theorem 32.2.1. For case (ii), Af is
in !lf ~- A is linear and invariant by Proposition 32.5.2; for continuity, we
invoke Theorem 32.4.3. o
A(e.x) = H(.X)e,x,
where e,x(t) = e2 i1r-Xt. These two definitions are equivalent for convolution
systems. One assumes, of course, that the sinusoidal signals e,x are in the
set of input signals. We Iook at several cases.
Case 1: h E 'if '.
From Proposition 33.2.1, Section 31.6, and (28.5),
A(e,x) = h(.X)e,x,
so
H=h. (34.2)
Case 2: h E !7 .
The equalities in (34.1) arevalid in this case by Proposition 33.1.1; hence
(34.2) is also true.
34.2 Realizable, or causal, filters 321
for all t 0 ER
Proof. We prove the result in the simple case where {j E X. If Ais realiz-
able, the relations h = A8 and supp( 8) C [0, +oo) imply that the support of
h is in [0, +oo). Conversely, if supp(h) c [0, +oo) and supp(f) c [t0 , +oo),
then we have supp(h *!) C [t0 , +oo) by Proposition 32.4.4. D
where
where we define O.j = 0 if j < 0 {the polynomial part is zero if p < q) and
where z 1 , ... , Zq are the simple poles of P(x)/Q(x) in C. Then
p-q
h(t) = L O.jO(j) + L kezktu(t)- L kezktu( -t), {34.6)
j=O kEK- kEK+
where K_ are K+ are defined in Section 24.3.1.
Case 2: H has multiple poles.
The polynomial part contributes a sum E~;;;;g O.jO(j) to h as in Case 1.
Thus we can limit ourselves to the case p < q. Using the same notation we
used in Section 24.3, we obtain the same result:
Pk(t) = fl
mk tm-1
k,m (m _ 1)!
We would like to take the inverse Fourier transform and apply the results
of Sections 33.1 and 33.2, but first we need to represent h as en element of
lf 1 + !7 . For this we write
O(t) = {10 if
if
ltl ~ 1,
ltl ~ 2.
h * f = hl * f h2 = (hl h2)!,
and
g = h* J,
34.3.3 Remarks
(a) The generalized solution in terms of functions that was given for
this equation in Section 24.3.5 is the tempered solution that we have just
obtained. This does not lessen our interest in Lesson 24 where we found
that the solution is a function when the input f is a function (p ::; q).
However, in Lesson 24 we assumed that f E L 1 n L 2 n L 00 ; here we have a
much wider range of inputs f, even if we restriet f to be a function.
(b) If P(x)jQ(x) has a pole on the imaginary axis, then the solution g
in !/ 1 is no Ionger unique. For example,
g" + w2 g = 8, w > 0,
has as solutions in !/ 1 all of the functions
Acoswt + Bsinwt, t < 0,
g(t)= { (
Acoswt + B +
1)
w
sinwt, t > 0.
This equation will be studied in Lesson 35. Here we merely note that the
moderated growth at infinity imposed by the space !/ 1 is not enough to
guarantee uniqueness. It is causality that will be determinant.
34.3.4 Causality
Proposition 34.2.2 and equations (34.6) and (43. 7) give us a necessary and
sufficient condition for causality.
(34.3) is realizable (or causal) if and only if the real parts of the poles of
P(x)jQ(x) are strictly negative.
34.4 Exercises
Exercise 34.1 Suppose the filter A: Y 1 (R) ~ Y '(R) is governed by the
equation
g" - w 2 g = !", w > 0.
(a) Compute its impulse response.
(b) Compute and represent graphically the step response.
This lesson is a direct continuation of the last one. We are going to look
for the causal solutions of a linear differential equation with constant coef-
ficients; thus by assumption, the filterwill be realizable (Section 34.2). For
convenience we write the equation with bq = 1:
p
L bkg<k) + g<q) = L
q-1
ajJ(j). (35.1)
k=O j=O
(35.2)
I
9q-2 = 9q-1,
9~-1 = -(bq-19q-1 ++bog)+ <p.
326 Lesson 35. Realizable Filters and Differential Equations
If we define
0 1 0 0
0 0 1 0
gl
M=
0 0 0 1
G=
rgq~l. J'
g h flJ,
-bo -bl -b2 -bq-1
then (35.2) can be written as the matrix cquation
and we havc
where X 0 is an arbitrary fixed vector. Assurne supp( <p) C [t0 , +oo ). Then
<I>(t) = 0 and G(t) = etM X 0 for all t < t 0 . If G(t) is causal, then necessarily
X 0 = 0. Equation (35.3) thus has the unique causal solution
(35.6)
and (35.1) has a unique causal solution g. Since g(t) = 0 fort E ( -oo, t 0 ),
the system is realizable.
35.2 Examples 327
k=O j=O
(35.7)
35.1.3 Remark Notice once again, the fact that the differential equa-
tion has a unique solution is a consequence of a constraint on g; in this
case, it is that g have support limited on the left. This restriction takes the
place of initial conditions. -
We will look at some examples of how to find actual solutions.
35.2 Examples
35.2.1 The RC filter RCg' +g= f
We are going to find the same result as we did in Section 25.1 but by a
different method. The impulse response h is the solution in .! ~ of
RCh' +h=o. (35.8)
328 Lesson 35. Realizable Filters and Differential Equations
RCh' +h=O
kRC6 = 6,
so
1
k= RC'
The impulse response (Figure 25.1) is
1 ~t t-s
g(t) = RC -oo e- RC f(s) ds,
which is indeed what we found in Section 25.1. Taking f = u gives the step
response (Figure 25.2)
1
35.2.2 The filter - 2 g'' + g = f (w > 0)
w
This second-order filter was studied in Section 25.3 where the conditions
imposed by the use of the Fourier transform led to noncausal, slowly in-
creasing solutions. Here, in contrast, the causality assumption willlead to
a solution that grows exponentially.
35.2 Examples 329
To find h, we use the same method that we used above. This time, however,
we willlook directly for a solution h of the form h = yu where y is a function
in C 00 R Thus
h' = y'u + y(O)o, h" = y"u + y'(O)o + y(O)o',
and (35.10) becomes
(y"- w2 y)u + y'(O)o + y(O)o' = -w 2 o.
This prompts us to look for a function y that satisfies the equation
y"- w2 y =0
with the boundary conditions
y(O) = 0,
y'(O)=-w 2 ,
which is a problern completely in terms of functions. The general solution is
y(t) = .Xcoshwt + J.LSinhwt;
with the initial conditions this becomes
y(t) = -w sinhwt.
The impulse response is
h(t) = -wsinhwt u(t),
and the solution g is given by
This filter is realizable but unstable: The input is bounded, but the out-
put is not. We note in this regard that the implications of the positions of
the poles given in Sections 24.4 and 24.5 do not apply here.
330 Lesson 35. Realizable Filters and Differential Equations
u(t) h,(t)
1 11
FIGURE 35.1. Causal step response of - 2 g + g = f.
w
1
35.2.3 The resonator 29" + g = f
w
We encountered this filter in Lesson 24 (Section 24.3.3), but we were not
able to analyze it using the Fourier transform because the poles were on
the imaginary axis. This equation describes the mechanical example Section
1.3.6 where the friction is negligible (zero in the equation). The equation
also represents a weight suspended on a spring where there is no "air"
friction, which means that the coefficient of the first derivative is zero.
We wish to find the impulse response h using the method we used in the
last section. Thus
h" + w 2 h = w2 6,
and by assuming h = yu, we have the system
y" +w2 y = 0,
y(O) = 0,
y'(O) = w2
The solution is
y(t) = .Xcoswt + J.tsinwt
with A = 0 and J.t = w. Thus
for a locally integrable, causal input f. The step response (Figure 35.2) is
h1{t) = (h * u)(t) = {1- coswt)u(t).
35.3 Exercises 331
u(t)
0 0 7T
w w
35.3 Exercises
Exercise 35.1 We wish to solve the equation
1 11
- - g +g=f
w2
by the matrix method described in Section 35.1.
332 Lesson 35. Realizable Filters and Differential Equations
e
tA _
-
[e-wt 0
ewt
J
0
(d) Deduce the integral expression for g(t) found in Section 35.2.2.
(1) Solve (1) by the matrix niethod of Section 35.1 (see Exercise 35.1).
(2) Show that changing the unknown to go = g- f leads to the differential
equation
(2)
Periodic Distributions
We are going to return to the topics of Lessons 4 and 5 armed with what we
now know about tempered distributions and their Fourier transforms. Our
objective is to show the connection between Fourierseriesand the Fourier
transform.
and that this series converges to f in the norm of L~(O, a). The Fourier
coeffi.cients are given by
Cn = ~
a lo
r f(t)e - 2i1rn~ dt. (36.2)
n=-oo
In this general case, we no Ionger know how to interpret the sum in (36.1).
And even if we knew that the series converged in some sense, we would still
336 Lesson 36. Periodic Distributions
have to show that its limit is f. We recall a negative result from Section 5.2:
N
L
t
f E L~(O, a) ~ cnii.".na ~ f in L~(O, a) as N ~ +oo.
n=-N
It is this formula that establishes the connection between the Fourier trans-
form of f and its Fourier series.
Several questions arising from this scenario need to be addressed. Let f
be periodic and locally integrable.
Ql: Does the Fourier seriesoff converge in some sense?
Q2: lf yes, does it converge to f?
Q3: Is f tempered?
Q4: Does the Fourier series of f converge to f in .57 '?
We need to answer "yes" to Q4 (thus also to Ql, Q2, Q3) if we are to
write (36.3) and expect it to make sense. In fact, this formula requires that
f be tempered and that the Fourier transform and the infinite summation
can be interchanged.
The answers, all positive, are given in the next result.
and
+oo
1 I:
n=-oo
a
(36.5)
Cn =- 11a
a o
f(t)e -2i7rn.!a dt.
36.2 The Fourier series of a periodic distribution 337
This proves (i). Next take the Fourier transform of (36.6); applying Propo-
sition 33.2.1 and (31.10), we have
- -
f- = fo .6.a = -1-fo.6. 1 ,
a -a
which by (28.5) is
~ 1 +oo ~ n
f =-
a
L !o(-)8!}..
a a
(36.7)
n=-oo
(36.8)
with equality in Y 1
338 Lesson 36. Periodic Distributions
( ?!'.,cn)
a nEZ
(36.10)
n=-oo
an= ~s(~)-
Proof. The idea is the sameasthat used in Proposition 36.1.3. We would
like to define S as the "restriction" ofT to (0, a), that is, to write
S= X[o,a] T.
for all t E ~. Assurne for the moment that we have such a function. The
distribution S = (}T will do the job: S has bounded support and
+oo +oo
S*a = L S*Ona = L TnaS.
n=-oo n=-oo
For all cp E ! ,
+oo +oo
(S * a, cp) = L (S, T-na'P) = L (T, OT-na'P)
n=-oo n=-oo
Since T is periodic,
(T, 1/J) = (T, Tna1/J)
for all 1/J E ! and all n E Z; using (36.11) shows that
+oo
(S * a, cp) = L (T, TnaOcp) = (T, cp),
n=-oo
which proves (ii). Statements (i) and (iii) follow from Proposition 36.2.1.
The Fourier coefficients of T are
1 ~(n) (36.12)
on = ~S -;;: .
The only way for this to hold is to have s(~) = 0 for all n.
To finish the proof, we need to show the existence of a function (} E !
satisfying (36.11). Let
cp(t) = P(~),
340 Lesson 36. Periodic Distributions
exists because for each t there are at most two nonzero terms. For the same
reason, ip, like cp, is infinitely differentiable. <i5 is a-periodic and strictly
positive; thus a suitable choice for () is
cp(t)
O(t) = ip(t)" D
(36.13)
n=-oo
T is a periodic tempered distribution with period .!. Thus its Fourier co-
a
efficients Y-n are slowly increasing. D
Proof. Write
N
en(t) = e
2i7rn!
a and /N(t) = L ckek(t).
k=-N
Then
N +oo
/NT= L L CkO.nen+k =
k=-Nn=-oo n=-oo
where
N
n(N) = L
CkO.n-k
k=-N
We will study what happens to n(N) as N-+ +oo.
(1) The series whose general term is CkO.n-k is absolutely convergent for
each fixed n. lndeed, the sequence (an-k) is slowly increasing in k (Theorem
36.2.2), so
lan-kl ~ Alklm
lkl. By Section 5.3.3(e), we also have
for sufliciently large
lckl ~ Blkl-m- 2 .
Thus ickan-kl ~ Clkl- 2 for sufliciently large lkl, and the sequence n(N)
converges as N -+ +oo.
(2) We wish to show that !NT-+ fT in .9 '. If cp E .9, then
36.4 Exercises
*Exercise 36.1 Let x = E!:-oo
xn8nh be a discrete periodic signal with
period a = Nh and grid h > 0, and let
N-1
Xk = N1 "'""
L....J -nk
XnW N , kEZ,
n=O
X = L Xj'T"jhll.a
j=O
Xke
-2i"->.kh A
'-" 1
-
k=O a
We are now going to tackle the problern of sampling analog signals. This
operation is a prerequisite of digital signal processing. For example, an
analog speech signal must be sampled before it can enter a digital tele-
phone system. A sampler records the level of the signal every a seconds
and transforms it into a sequence of impulses (Figure 37.1). An analog-
ta-digital converter (ADC) codes these impulses as numbers that can be
processed digitally.
+oo +oo t
"" ( ) 1 "" ~(n) 2i11'n-a,
L_-ft-na=~L_-f~e (37.1)
n=-oo n=-cx::>
+= +=
L
n=-oo
Tnaf = L J * Ona = J * ~a, (37.3)
n=-oo
where specific values of the function do not appear. Thus f can be any
distribution for which the series converges, say, in ! 1
Finally, we note once again that the conve;:pence of these series must be
interpreted as the (symmetric) limit of 2:::~=-N as N ----t +oo. We know,
for example, that the series on the right in (37.1) converges in! 1 if is a 1
slowly increasing function (Theorem 29.4.2). As in the case of the left-hand
side, the variable t plays only a symbolic role in the expression on the right.
There are, however, cases where the Poisson formula holds for all t E IR.
This happens, for example, when f E !/ (see Exercise 37.1).
(i) Theseries (37.4) converges in L 1 (0, a), and FE L~(O, a). The Fourier
coeflicients of F are
k E 71..
(ii) If in addition, f' E L 1 (~) (where the derivative f' is taken in the
sense of distributions), then the series (37.4) converges uniformly on
~' and thus F is continuous on R
Proof.
(i) We show that the restriction of the series (37.4) to (0, a) converges in
the complete space L 1 (0, a) by showing that the sequence
N
FN(t) = L f(t- na)
n=-N
IIFN+P- FNIIu(o,a) ~ L
N<lni<5.N+PJo
r lf(t- na)l dt ~ 1lxi~Na lf(x)l dx.
Since f is in L 1 (~), the last integral tends to 0 as N and P tend to +oo.
Thus the sequence (FN) converges to F on (0, a), and F E L~(O, a) by
periodicity. The Fourier coefficients of F are the limits of those of FN:
But
Ck (FN ) = -1 ""
L...J 1a
f ( t - na ) e -2inkla dt = -1 ~(N+l)a J( x )e -2ink'Ea dx
a lni<5.N o a -Na
tends to
~ ~+oo f(x)e-2ink~ dx = ~f(~),
a _ 00 a a
for all t E [0, 1] and all n E Z, and suchthat 2:::~:-oo Un < +oo.
37.2 Poisson's formula in L 1 (JR) 347
The assumption in (ii) means that T/ = Tf', and an argument like that
given in Section 30.1.2 shows that f is absolutely continuous on all bounded
intervals. Thus fort E [0, 1],
f (t - n) -
r+l f (t -
Jn x) dx =
r+l [f (t - n) -
Jn f (t - x) Jdx,
and
lf(t-n)l ~ 1 n
n+l
[1/(t-x)l+l/'(t-x)l] dx ~
~-(n-1)
-(n+l)
[lf(y)l+lf'(y)l] dy.
This proves that the series (37.4) converges uniformly on [0, 1] (or [0, a]
in general). Since F is periodic, the convergence in uniform on JR, which
proves the result. o
37.2.2 Theorem
(i) If f E L 1 (JR), then Poisson's formula (37.1) expresses the equality in
.9 1 between the function FE L~(JR) and its Fourier series.
(ii) If, in addition, f' E L 1 (JR) (where the derivative f' is taken in the
sense of distributions), equality (37.1) holds for all t E JR. More pre-
cisely, the series on the left converges uniformly on lR to a continuous
periodic function F, and the Fourier series of F, which is the series
on the right, converges uniformly on lR to F.
348 Lesson 37. Sampling Signalsand Poisson's Formula
Finally,
In this situation, f is said to be band limited, which is to say that E iif '. 1
We saw in Section 29.5 that sampling f every a seconds can be expressed by
+oo
af a = a L f(na)8na
n=-oo
1
This makes sense because E iif I implies f E C 00 (lR).
We wish to determine how sampling modifies the spectrum of f. For this
we need to compute the Fourier transform of af a. By Proposition 33.2.1,
Y(f * ~a) = J a.
Taking the Fourier transform and using (31.10) shows that the spectrum
of the sampled function is
af7):;. = 1* !.
a
37.3 Application to the study of the spectrum of a sampled signal 349
f(t) f(A)
0 0
?
-a Oa 2a
(37.7)
n=-oo n=-oo
f(t) f(>.)
8> _1_
2Ac
-3a -2a -a 0 a 2a 3a
a< _1_
- 2Ac
0 a 2a 1 >.
a a
FIGURE 37.4. A high sampling rate separates the components of the spectrum.
then we see that the right-hand side of (37.8) is exactly the right-hand side
of (37.1) with a = 1. We will compute F using f. From Section 18.2.2 we
know that
!( t) = ie-2''-bltl,
F(t) = i +oo
L e-27rblt-nl (37.9)
n=-oo
for all t ER This series converges much faster than (37.8), andin this case
we can compute the sum explicitly. Fort E [0, 1],
1 2i7rnt 7r 1
n~oo
+oo
n2 + b2 e = b sinh 7rb cosh [27rb ( t- 2)] .
3 7. 5 Exercises
Exercise 37.1 (Poisson's formula in Y (JR.))
(a) Use equation (31.10) to show that
f
n=-oo
~(~) =a f
n=-<Xl
~(na), a~O,
for all ~ E Y (R)
(b) Deduce from this that
t > 0.
(a) Show that there exists a frequency f E [-T/2,T/2) suchthat the signal
h(t) = cos(27rft + <p)
Exercise 37.2 Use the method illustrated in Section 37.4 to transform the
Fourier series
+oo
F(t) = L e-bn2 e2i"nt, b > 0.
n=-oo
Does the transformed series converge faster than the original series?
Lesson 38
(38.1)
with equality in Y '. When the sampling rate is high enough, which is
when
1
a < -
- 2-Xc'
the translates of the spectrum 1 in the left-hand side of (38.2) do not
overlap; they are separated by a- 1 - 2-Xc ~ 0.
The idea behind Shannon's formula is to isolate the central copy of 1
and use it to reconstruct f, which will then be expressed in terms of its
values f(na) (see Figure 38.1). The next assumptionisthat [, and hence
f, is square integrable:
(38.3)
The left-hand side of (38.2) is then a periodic function F(.\) with period
354 Lesson 38. The Sampling Theorem and Shannon's Formula
F(.\)
FIGURE 38.1.
1/a that is square integrable over one period. Thus F can be expanded in
a Fourier series
+oo
F()..) = L Cne2i11'-\na (38.4)
n=-oo
The sequence (c-n) is square integrable, and the equality (38.4) holds in
L 2 (0, 1/a). Since Fourierexpansions of periodic tempered distributions are
unique (Theorem 36.2.2),
C-n = af(na).
First conclusion: Under the assumptions (38.1) and (38.3) the equality
(38.2) holds in L~(O, 1/a) and
+oo
L if(naW < +oo.
n=-oo
we see that
+oo
!(>.) = a L f(na)r(>.)e-2i11'-\na, (38.5)
n=-oo
which holds in L 2 (R). From the continuity of 5 on L 2 (R) and from the
relation
_ . _ sin~(t-na)
5 [r(>.)e- 2m-\na] = (5 r)(t- na) = a ,
rr(t- na)
38.1 Shannon's theorem 355
1
and for all a ~ 2>.c ,
then the series converges uniformly on IR and equality holds for all t ER
REMARK: Shannon's formula can also be written as
+oo ( 1)n
f(t) = ~ sin '!!..t "" f(na)---. (38.10)
1r a ~ t-na
n=-oo
This causes poles, which do not belong to f, to appear in the series at the
points tn = na.
356 Lesson 38. The Sampling Theorem and Shannon's Formula
N
38.2 The case of a function f(t) = L cne 2i1rAnt
n=-N
This equality holds in L 2 (0, 1/a), but in view of Theorem 5.2.4, it is also
true for all t in the open interval (-1/(2a), 1/(2a)). Hence, for all A ER,
LN
l(na)
1sin~(t-na)
7r a cp(t) dt --t
1 l(t)cp(t) dt
n=-N lR 0;(t- na) lR
for all cp E Y. This, however, is not the case: One can find such signals I
and functions cp for which the sum does not converge to the integral on the
right (see Exercise 38.5).
I
Va = { v E L 2 (IR) supp(v) c [- 2~, 2~] }
It is easy to show that Va is a closed subspace of L 2 (JR).
38.4.1 Proposition
(i) The family offunctions (san)nEZ is an orthogonal basis for the Hilbert
space Va.
(ii) If (aj)jEN is any sequence such that _lim aj = 0, then ujEN Vaj is
J-++oo
densein L2 (R).
1 -- 212~
IR
SanSap - a
_ _l_
2a
e-2i11".\(n-p)a d'/\ -- {a
0
if n = p,
if n -=1- p,
(38.15)
2
This proves density and completes the proof of (i).
To prove (ii), take f E L (JR.), c > 0, and definc 9n by
f(t)
N
Cn = ' '' Cn-2N + Cn-N + Cn + Cn+N + Cn+2N + ' '' ,
which we saw in connection with the discrete Fourier transform. Here the
approximate spectrum c;:[ is "contaminated" with extra copies of the real
spectrum Cn that appear as the terms Cn+pN, p =f. 0. Prefiltering eliminates
these terms, which can be too large for practical computations, even though
they eventually tend to zero.
38.6 Exercises 361
1
Then by (38.5), for a < 2..\c,
+oo
[(>.) = a L f(na)e-2i1rna>.
n=-oo
L
N-1
SN(>.)= a Xne-2i1r>.na,
n=-N
38.6 Exercises
Exercise 38.1 Let f be an element of .Y '(R) suchthat
2 +oo 1
7TCOt2-rrt = (16t - 1) ~ (4 2 )( )
L......t n - 1 2t- n
n=-<X>
1 +oo 1 1
cotx=-+~(-+-)
X L......t X - n7T X + n7T
n=l
e2i7rt + e-2i7rt
f(t) = cos21rt = 2
and notice that in this case one can apply Theorem 5.2.4 for all t ER
Exercise 38.4 Write equality (38.12) at the points t = 1/(2a) using The-
orem 5.2.4.
SN('P) = L f(n)In('P)
n=-N
with
In('P) = 1r(t)<p(t+n)dt
and r = x[_! !J
2'2
(a) Show that
In('P) = 1: 1
2
( 5<.p)(x)e- 2 i"nx dx.
38.6 Exercises 363
L
N
SN(cp) = ncn(?/J),
n=-N
where 1/J is the function with period 1 that agrees with Y cp on the interval
( -1/2, 1/2).
(c) Take cp = g, where g is an element of !lJ (R) suchthat
g(x) = x if lxl <~
(Exercise 27.4). Compute SN(cp) and conclude that Shannon's formula is
not generally true in .7 '.
lesson 39
We are going to study several specific questions about discrete signals and
filters in this and the following lesson. The current lesson concentrates on
the convolution of discrete signals and its application to discrete filters.
will be called a discrete signal; we denote the set of discrete signals by Xa:
This is a vector space that is usually endowed with the topology induced
by that of ! 1 , which is the topology of pointwise convergence:
lim XN =X in! 1 <====> lim XNn = Xn for all n E Z.
N-+oo N-+oo
Dx=h*x, xEX,
for all cp E !fJ, and these sums have only a finite nurober of terms. Hence,
+oo
(h*x,cp) = (hk,(Xm,cp((m+k) a))) = L hk'I/Jk,
k=ko
where
+oo
'1/Jk = L Xm'P((m + k)a).
m=mo
The result follows by interchanging the (finite) sums and by the change of
variable n = m + k:
are such that (hk) is rapidly decreasing and (xn) is slowly increasing.
Then the convolution h * x is well defined. Furthermore,indexdiscrete sig-
nals!convolution of
(i) h * x is a tempered distribution.
+oo +oo
(ii) h *X = L Yn8na with Yn = L hkXn-k, and the series for Yn
n=-oo k=-oo
converges absolutely.
Proof. The proof is based on Fubini's theorem for the discrete measure
space (Z x Z, .!T , f..l) where .!T is the a-algebra generated by the finite sub-
sets ofZxZ and J.l is the measure defined by J.t(S) = the number of point inS.
A function u : Z x Z --+ C is integrable (or summable) if and only if Iu I is
integrable. One part of Fubini's theorem states that the condition
L
+oo
n=-oo
(
L
+oo
k=-oo
lun,kl
)
<+oo
1
z
2
U= L
(n,k)EZ 2
Un,k = L
+oo
n=-oo
(
L
+oo
k=-oo
Un,k
)
= L
+oo
k=-oo
(
L
+oo
n=-oo
)
Un,k
with
+oo
'1/Jk = L Xmcp((m + k)a).
m=-oa
(39.2)
(39.3)
If (39.3) is finite, Fubini's theorem tells us that all of the series involved in
the formal computations are absolutely convergent and Summation in any
order gives the same answer.
The sequence lhkl is rapidly decreasing and lxnl is slowly increasing. If
we define
+oo +oo
L
t t
f(t) = lhkle2i1rk;;: and T = "'"'
~
IXn Ie2i7rn-a,
k=-oo n=-<X:l
and they are slowly increasing (Theorem 36.2.2(iv)). Thus there exist A > 0
and o: > 0 such that
n :::; A(l + lnl"')
for all n E Z. On the other hand, since cp E SC , there is a B > 0 such that
370 Lesson 39. Discrete Filters and Convolution
This shows that the sum (39.3) is finite and hence that (39.2) is summable.
We thus can sum (39.2) in any order, andin particular,
(h * x, (f?) = L
+oo ( +oo
L )
hkXn-k (f?(na),
n=-oo k=-oo
which proves that h * x makes sense and is given by (ii). The estimate
shows that (Yn) is slowly increasing, so (ii) follows from Corollary 36.2.3.in-
dexdiscrete signals!convolution of o
and
we see (by a computation similar to the one above) that h * x exists for all
h,x E l~ and that h * x E l~. (The convolution l! * l~ does not need tobe
studied as a special case since l~ c l':'.)
39.4 Summary
The convolution h * x is defined for the distributions
+oo
and X= L
XnOna
n=-oo n=-oo
In cases (a) and (b), the series for Yn isafinite sum; in the other cases, the
series is absolutely convergent. In case (c),
,;;; = h. x.
These results show that the mapping
D: X___, Xa,
x t--+ D(x) = h *x
is a discrete filter in the following cases:
Case 1: h is finite, and X = Xa.
Case 2: h is causal, and X= Xa n !iJ ~-
Case 3: h is rapidly decreasing, and X = Xa n 5I' I (slowly increasing).
Case 4: h E l~, and X= l':.
Case 5: h E l~, and X= l~.
Case 6: h E l':, and X= l~.
Case 7: h E Xa, and X = Xa n ~ 1 = Ya (finite inputs).
In Cases 1, 2, and 3, the topology on X isthat induced by !iJ 1 In Cases
4, 5, and 6, one can take the topologies of the l~ spaces. In Case 7, one has
many choices.
{: : }
following:
[l~:~:~:]
(or causal).
[xn = 0 for all n < 0 =? Yn = 0 for all n < 0.]
The next result characterizes these two properties in terms of the im pulse
response.
(For c =lclei 0 , sign(c) = e-iiJ.) The signals xP are finite, so they are in
X n l': for Cases 2, 5, 6, and 7, and llxPIIoo :::; 1. Then
oo n
Y~ = L hkx~-k = L hksign(hp-n+k),
k=-oo k=-2p+n
and
p
y~ = I: lhkl
k=-p
for all p 2:: 0. From the definition of stability we conclude that
p
IY~I = L lhkl:::; A
k=-p
39.6 Exercises
Exercise 39.1 Let x = :L:=-oo XnDna, a > 0, be a discrete signal. Compute
the impulse responses of the following filters y = Dx.
(a) Yn = Xn-1
1
(b) Yn = 2(Xn + Xn-1).
1
(c) Yn = 3(Xn+1 + Xn + Xn-1)
Which of these filters are realizable?
Exercise 39.2 Show that-;;;;; = hx when h and x are in l~ (use the result
in Section 39.3.4).
Exercise 39.3 Consider the discrete filter whose impulse response h = (hn)
is given by
if n::::; 0,
if n > 0,
and that belongs to Case 7 in Section 39.4. Show that the response of every finite
signal (which is necessarily bounded) is bounded but that the filter is not stable.
(40.3)
The z-transform of a discrete signal does not always exist. For example,
there is no z-transform for Dirac's comb.
376 Lesson 40. The z- Transform and Discrete Filters
EXAMPLES:
(a) Fora > 0 and > 0, define
if n < 0,
if n ~ 0.
Then
-1 +CXl
X(z) = " nz-n + "anz-n = _z_ + _z_
.L...J .L...J -z z-a
n=-CXJ n=O
for values of z satisfying iz/l < 1 and la/zl < 1. Thus the z-transform
exists if a < . It is defined and holomorphic in the annulus a < izl < .
(b) The discrete version of the unit step function (Heaviside function) is
defined by
0 if n < 0,
x -u - {
n - n - 1 if n ~ 0,
and
+CXJ 1
U(z) =L z-n = -1---z---=-1
n=O
if lzl > 1. Here the annulus of convergence is the exterior of the unit disk:
r = 1, R= +oo.
It follows that
Y(z) = H(z) X(z)
for all z E A. One should not be surprised that the z-transform of a con-
volution of two signals is the product of their z-transforms!
so
+oo
X= Lrn8na
n=O
By the second method,
1
Xn = -.- -Zn 1
- - dz.
2m r z- r
378 Lesson 40. The z- Transform and Discrete Filters
)Y
FIGURE 40.2.
Ifn ~ 0, the residue of f(z) = zn(z-r)- 1 atz= r is rn. Ifn < 0, another
pole appears atz= 0. The residue atz= 0 is obtained by expanding f(z)
around z = 0:
1 +oo zn+p
=- - - =-""""" -.
Zn
f(z)
r 1- ~ L... rP+l
r p=O
n=-oo
40.2 Applications to discrete filters 379
!Y
____ poles of H(z)
n=-oo
When the filter is realizable, hn = 0 for all n < 0. In this case, the hn
are obtained from thc recurrence
ho = ao,
n
hn =an- :Lbkhn-k, n = 1,2, ... ,
k=1
for all n E Z.
40.2.3 Example
The discrete form of the realizable RC filter, RCv' + v = f, is
Yn- Yn-1 + Yn = Xn,
Re ;::_____::.___ nEZ.
a
In this case, the annulus of convergence is the exterior of the unit disk:
r = 1, R = +oo. The discrete filter has the form
40.3 Exercises
Exercise 40.1 Invert the z-transform defined by X(z) = z(z- r)- 1 in the
< lzl < r
annulus 0 and compare the result with the example in 40.1.2.
z2 + 1
H(z)=~1, z -
knowing that the associated filter is realizable.
Current Trends:
Time-Freq uency Analysis
lesson 41
Thus truncating the signal results in convolving its spectrum with the car-
dinal sine (Figure 41.1).
2A
f
The approximation of by g becomes better as A increases, that is, as s A
better approximates the Dirac impulse. Unfortunately, the computations
for this process quickly become very voluminous. The cardinal sine decays
slowly and has important lobes near the origin. To avoid these problems,
one replaces X(-A,A] with a more regular function. These functions are all
called windows, and they are concentrated around the origin.
EXAMPLES:
(a) Triangular window (Figure 41.2)
-A 0 A
w(t)
a= 0.54
A 0 A
2 2
FIGURE 41.3. Hamming and Hanning windows.
Wt(.X, b) = l
-oo
+oo
f(t)w(t- b)e- 2i1r>.t dt. (41.1)
Wt(A, b) replaces f(.X). The mapping f ~---+ Wt is called the sliding window
Fourier transform or simply the windowed Fourier transform.
388 Lesson 41. The Windowed Fourier Transform
w(t) = Ae-" 12
K
A Y7i'la
0 0
The parameter A plays the role of a frequency, localized around the ab-
scissa b of the temporal signal. W 1 (.\, b) thus provides an indication of how
the signal behaves at time t = b for the frequency A. We use the function w
rather than w in (41.1) for reasons of convenience and because we wish to
allow complex-valued windows. Thus, W1 becomes a scalar product in L 2 :
Wt(.\,b) = l
-oo
+oo
f(t)w>.b(t)dt.
Since
~ (t:) ') ,
W>.b ." = e-2i1T(~->.)b w
~((:
." - ,~~, (41.5)
this becomes
so
= ll!ll~llwll~ = 11!11~
This establishes (a).
390 Lesson 41. The Windowed Fourier Transform
To prove (b), we first show that YA is well-defined for all A > 0 by showing
that (A,b) ~ WJ(A,b)w>.b(x) is integrable on the strip [-A,A] x R Let
I:
By Schwarz's inequality and Parseval's relation, we have (Theorem 22.1.4)
with
g(A) = 1+oo
-oo 5 df(~)(~- A)](b)w(x- b)e 2 i1r>.(x-b) db,
so
41.3 Dennis Gabor's formulas 391
The next step is to verify that the function of (.>., ~) under the double
integral (41. 7) is integrable on [- A, A] x ffi.. Since Iw I is even,
Denote the second integral by ct~A(~). Then 0::::; tt~A(~) : : ; 1, since llwll2 = 1.
Since ct~A is bounded, 1ct~A is in L 2 and 9A = !T (1 ct~A) The last step
is to show that 9A tends to f in L 2 as A-+ +oo. Forthis we evaluate the
norm of the difference:
(41.8)
so
A
[ A
llt;l?.-z
[1- ct'A(~)fl1(~)1 2 d~::::; 1 A
lt;l?.-z
li(~W d~,
which also tends to 0 as A -+ +oo. These two estimates show that c:(A)
(41.8) tends to 0 as A tends to infinity, and this proves (b). 0
392 Lesson 41. The Windowed Fourier Transform
contains information in both time and frequency around the point (b, .\).
For numerical computations, the coefficients Wt(A, b) are evaluated on
a grid (m.\o, nbo) with m, n E Z and Ao, bo > 0. One thus obtains a double
sequence Wm,nU) = Wt(m.\o, nbo), which is a discretized version of the
function of the two real variables ,\ and b.
f(x) = f-oo
+oo
J(~)e2ine d~,
f(x) = { Wt(.\,b)w>-.b(x)d.\db,
JJR2
can be interpreted as decomposing the signal f in terms of functions that
play the role of basis functions, except that sums are replaced by integrals.
In the Fourier transform, these functions are sinusoids; in the Gabor
transform, they are strongly attenuated sinusoids, or looked at the other
way, modulated Gaussians (Figure 41.5). In the frequency space, we have
the representations illustrated in Figure 41.6.
With Fourier's method, the "basis functions" are completely concen-
trated in frequency (Dirac impulses) and totally distributed in time ( unat-
tenuated sinusoids extending from -oo to +oo ). This is another way to
explain that taking the Fourier transform gives the maximum amount of
information about the distribution of the frequencies but completely loses
information relative to time.
With Gabor's method, the figures show that time-frequency information
remains coupled, although there is always a compromise: The uncertainty
principle limits the simultaneaus localization in time and frequency. In spite
41.4 Comparing the methods of Fourier and Gabor 393
Re[e2hrxf] Re[w,~.bCxl] ,'~,
+1 / \ y= w(x-b)
\
-1
b X
Fourier Gabor
FIGURE 41.5. Basis functions for Fourier and Gabor decompositions.
Re[w_;;(~)]
Fourier
,\
Gabor
FIGURE 41.6. Basis functions for Fourier and Gabor in frequency space.
b-1
f(t)'W>-.b(t) dt = 0.
to analyze signals that contain features on scales that range over several
orders of magnitude. This is the case, for example, with speech. Consider
the word "school." It begins with a short high-frequency attack followed by
a Ionger relatively lower-frequency component. Fluid mechanics provides
another important example. In fully developed turbulence, one observes
events on scales that range from the macroscopic to the microscopic.
The geophysicist Jean Morlet encountered these kinds of problems in
connection with seismic exploration for oil. Here it is necessary to analyze
signals that result from a pulse being reflected (and delayed and com-
pressed) from various layers in the earth. This led Morlet to introduce a
new method where the window is not only translated but is also dilated
and contracted. This was the beginning of the use of wavelets for numerical
signal processing.
41.5 Exercises
Exercise 41.1 With the notation and hypotheses of Theorem 41.3.1, show
that for fand g E L 2 (1R),
//JR w
2
1 (>.,b)W 9 (>.,b)d>.db= 1f(t)g(t)dt.
Exercise 41.2 Consider the signal f(t) = e2 i"at, a E JR, and the Gaussian
window w(t) = e-?Tt 2
1
(a) Verify that
W 1 (>., b) = f(t)w(t- b)e- 2 i"M dt
1
to compute
Wj(A, b) = j(t)w(t- b)e- 2 i1rM dt.
Lesson 42
Wavelet Analysis
Gabor's method dates from the 1940s. With wavelets we enter a dynamic
contemporary research environment; what is now known as the modern
theory of wavelets emerged in the 1980s, notably with the article [GM84]
by Alex Crossmann and Jean Morlet. We say "modern" wavelet theory
because looking back over the mathematicallandscape from a late twentieth
century perspective we can identify many earlier ideas and techniques that
are now logically included in this theory. Work by Haar in 1909; work in the
late 1920s by Strmberg; results from the 1930s by Littlewood and Paley,
Lusin, and Franklin; and later work in the 1960s, particularly the result of
Calder6n on operators with singular kernels-all these efforts and others
are now interpreted in the language of wavelets.
What happened in the 1980s was qualitatively different; there occurred
a conjunction of requirement and solution. Jean Morlet, a geophysicist,
wished to analyze a particular dass of signals associated with seismic ex-
ploration, and he had an idea about how this should be clone. He sought
the collaboration of Alex Grossmann, who, being a theoretical physicist,
had command of certain mathematical tools, particularly those associated
with coherent states and group representations from quantum theory. The
immediate result was their celebrated 1984 paper; it was also the begin-
ning of a productive collaboration between mathematics and other sectors
of science and technology. We will say more about contemporary research
at the end of the lesson, once some basic results have been established.
b E IR, a > 0.
396 Lesson 42. Wavelet Analysis
The properties of '1/J are quite different from those of a window, which
has more or less the aspect of a characteristic function, while '1/J, on the
other hand, oscillates and its integral is zero. We also want '1/J and :(jJ to
be well localized, which means that they both converge to zero at infinity
fairly rapidly. In this way one obtains a function that looks like a wave: It
oscillates and quickly decays. This is the source of its name. Morlet used
the function
t2
7/J(t) = e-2 cos5t,
which is now known as Morlet's wavelet; derivatives of the Gaussian are
widely used in practice. Figures 42.1-42.4 illustrate differences in the be-
havior of the Gabor functions W>.b(t), which have a ridged envelope, and
wavelets, which are dilated and contracted. With wavelets one sees the
action of an accordion. (The factor a- 112 has not been used in the figures.)
Unlike Gabor functions, wavelets do not have a rigid envelope.
l{l(t)
Re[wA.o(t)] Re[wA.o(t)]
I
\
''
I
''
I
I
'' ' I
0.0 0.0
1 2
FIGURE 42.2. Gabor functions W>.b(t) = e -2(t-b) e 2inAt: The envelope is rigid,
and the number of oscillations varies with frequency.
42.2 The wavelet transform 397
rfr(x)
0.8227 a=1
0.5727
0.3227
0.0727
-0.1773
-0.4273
-0.6773 L--.....__._____.___..'--'---.1.....-'----- '----'-"'--+
-5.0 -3.0 -1.0 0.0 1.0 2.0 3.0 4.0 5.0
-4.0 -2.0
FIGURE 42.3. A mother wavelet (8th derivative of a Gaussian).
'1/Jab(t) = 1
VfaT'I/J (t--a-b) , a, b E IR, a =1- 0,
CJ(a, b) = f-oo
+oo
f(t)""ijjab(t) dt.
K
1 !J F
ICJ(a, bW-
da db =
2
a
f+oo lf(tW dt.
-oo
(b) Rcconstruction formula:
f(x) = K
1 Jr1r
IR 2 CJ(a,b)'I/Jab(x)~
dadb
fc(x) = K
1 Jr1rlal~c CJ(a, b)'I/Jab(x)~,
dadb
bEIR
398 Lesson 42. Wavelet Analysis
1/Js.o(x)
0.82261
a=3
0.57261
0.32261
0.07261
-0.17739
-0.42739
-0.67739 .__....._....-......_...._---'_..l---L-...L-~~...~....-l.
-5.0 -4.0 -3.0 -2.0 -1.0 0.0 1.0 2.0 3.0 4.0 5.0
1/1! o(x)
2'
0.8227
a=0.5
0.5727
0.3227
),07271-----
-0.1773
-0.4273
-0.67731..-....J---L-..L---L...II-L....IL-1---L-....I--1..---l
-5.0 -4.0 -3.0 -2.0 -1.0 0.0 1.0 2.0 3.0 4.0 5.0
FIGURE 42.4. Wavelets at low and high frequency: They have the same form
and the same number of oscillations; they are dilated for large a and contracted
for small a.
Proof. First two observations: The '1/Jab are normalized so that II'I/Jabll2 = 1,
and the proof is similar tothat of Theorem 41.3.1. Thus, as before, we find
another expression for CJ(a, b):
and since
(42.1)
we have
(42.2)
42.2 The wavelet transform 399
By the change of variable ~ = a>., we see that the last integral is constant
and equal to K, which proves the result.
To prove (b), we first compute
and since
it follows that
(42.3)
Define
g"(x) = {
lial?.c
J(a) d~ = {
a llal?.c
(l+oo
-oo Jc>.)l;j}(a>.We2i7rAx d>.) dlal. (42.4)
a
The next step is to show that the function of (a, >.) under the integral signs
is integrable on (Iai ~ e) x R
By the change of variable ~ = a>., we see that
A= l+oo r l;j}(a>.W
-oo li(>.)l ( Jial?.c Iai
da) d>.
A1 = 1 -1
1
lf(A)I ( f
llel~el>.l
1~(~)1 2 d~) dA
1~1
~ K j_11 lf(A)IdA ~ K\1'211/112
For lAI ~ 1,
so
A2
C 1>.1~1
11\AI)I
1A C 1>.1~1
~;)
A
2
< +oo.
This means that we can interchange the order of integration in (42.4); thus
(42.5)
42.2.2 Remarks
(a) Hypothesis (i) implies that ~(0) = JIR '!jJ(t) dt = 0, since ~ is continu-
ous. In all practical cases this condition is also suffi.cient. For example, if 'ljJ
and x'ljJ are integrable, then ~ E C 1 (JR) and ~---> 1~1- 1 1~(~)1 2 is continuous
at ~ = 0. There is no problern with the integral at infinity, since 'ljJ E L 2 (JR).
f
(b) For signals f belanging to L 1 (JR) n L 2 (JR) such that is also in L 1 (JR),
the proof of the theorem is simplified because all of the integrals exist in
the usual sense when c = 0. From (42.3) we deduce that
42.2.3 Examples
(a) The wavelet first used by Morlet (Figures 42.5 and 42.6),
t2
'!fJ(t) = e -2 cos 5t, (42.6)
is not normalized, but this is not a problem. On the other hand, the hy-
pothesis (i) is not satisfied, since
Thus K = +oo! However, the value of ~(0) is on the order of 10- 5 . For
numerical computations this is essentially zero, and in practice things work
well. Nevertheless, the theorem does not apply to Morlet's wavelet.
(b) The simplest example of a wavelet is the piecewise constant function
'ljJ defined by
1 if 0 <X<~'
{
'1/J(x) = -~ 1"f 1 1
2 <x < '
elsewhere.
This is the Haar wavelet (Figure 42.7), and
402 Lesson 42. Wavelet Analysis
I/I( X)
0.92411
0.6741
0.4241
0.1741
--o.0759 0 X
--o.3259
--Q.5759
--Q.8259 L---L.-...l---~--L.--IL...l---L....---L.-...l--
-4.0 -3.0 -2.0 -1.0 0.0 1.0 2.0 3.0 4.0
1.25.
1.00
0.75
0.50
0.25
1/l(X)
0.50
0.25
0 1.: 1:
2: :
X
L--.
-10.0 -5.0 0.0 5.0 10.0 ~
FIGURE 42.7. The Haar wavelet and the modulus of its spectrum.
( ) =- 2 ( 2 ) _.!_x2
'1/Jx -1 1-x e 2 (42.7)
v'371"4
42.2 The wavelet transform 403
Both 'lj; and ;j} belang to Y and are well localized. Figure 42.9 illustrates
the 8th derivative of the Gaussian and its spectrum.
tf!(g) ~(g)
0.86295
1.50
0.61295 1.25
0.36295 1.00
0.11295 0.75
0.50
-0.13705 X 0.25
-Q.38705 0.00 L..._----L-LL-.L----...J-"---L-__.__.. ,
-6.0 -4.0 -2.0 0.0 2.0 4.0 6.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 g
-5.0 -3.0 -1.0 1.0 3.0 5.0
FIGURE 42.8. The Mexican hat and its spectrum.
~(g)
y
1.50
0.8227
1.25
0.5727
0.3227 1.00
0.75
0.07271---..::::-:r-+--++.:+--+---'<:::::=----
-o.1773 X
0.50
-0.4273 0.25
-o .8773 L.-..L-..J---'--'"----'-----"--'---'-.......__---' 0.00
-4.0 -2.0 0.0 2.0 4.0 -1.5 -1.0 0.5 0.0 0.5 1.0 1.5 g
-5.0 -3.0 -1.0 1.0 3.0 5.0
FIGURE 42.9. The 8th derivative of the Gaussian and its spectrum.
In this case, ICJ(a, b)l depends only on a, and when ;j} is real, the argument
of C1 (a, b) is proportional to b modulo 21r. Here we sec that the wavelet
coefficients tell us something about the behavior of the function f, and
404 Lesson 42. Wavelet Analysis
which means that the wavelet transform "ignores" constants. For f(t) = t,
Ct(a, b) = alal 1 / 2 l
-oo
+oo
x""i(x) dx
alall/2 :::::.'
f(t) = t ===} Ct(a,b) = - -2-.-7/J (0).
Z1r
This last property plays an important role when the wavelet transform is
used to analyze the singularities of a function.
One would thus havc an orthogonal basis, in the usual sense, for the Hilbert
space L 2 (1R.), where the coefficients Cjk = Cf(j, k) are independent of one
another. What is special about this basis is that the functions are all de-
termined from one wavelet 'ljJ by dilations and translations.
1/Jik(x)
-2-~ - - - - - -
Un
q
= ~ J(ll(o) -oo lT:jj}(nx) dx + l+oo
l+oo xl
-oo R(x):jj}(nx) dx
with
R(x) =
lo
r (x- t)q j(q+ll(t) dt.
q!
42.3 Orthogonal wavelets 407
q f(l)(O)Mt
Un = L
t=O
ll l+l
.n
+rn.
An easy computation shows that lrnl :::; Cn-(q+2) for some constant C.
Thus
_ f'(O) M1
+
f"(O) M2
+
...
+
j(q)(O) Mq
+
o(-1-)
Un - n 2 11. n3 21. n q+l q.I n q+2 '
(42.10)
for t/J E L 2(1R) with llt/JII2 = 1, and t/J is a wavelet of some order r ~ 0.
Later in the lesson we will see how to construct orthonormal bases of the
form (42.10) where all we know about t/J is that it is in L 2(JR). This is not
particularly interesting for applications, as we have seen in the case of the
Haar wavelet (which is of order 0). For efficient numerical computations
it is necessary to use higher-order wavelets, which means that the wavelet
and its Fourier transform have reasonably good localization and regularity.
c
4 1 1 2 4 ~
3 3 3 3 3
FIGURE 42.11.
The curve BC is required to have the sameform as AB, but reversed and
stretched:
w(2~) = ~- w(2(1- ~)), ~ E [~, ~].
Then ;j} is defined by
It is easy to see that ;j} is in !l! . From this it follows that 'ljJ E .9 and is
given by
(42.12)
Observe that'lj; is real and its graph is symmetric with respect to t = 1/2.
It turns out that there is not much leeway in the choice of w, and the
wavelets constructed all have about the same appearance as the one shown
in Figure 42.12. We note that although this function decreases rapidly,
42.3 Orthogonal wavelets 409
1.25 1/J(X)
1.00
0.75
0.50
0.25
0.00
-0.25 X
-0.50
-0.75
-1.00
-8 -5 -2 0 4 7
-1 1
FIGURE 42.12. Meyer's C 00 wavelet.
it has a rather large "numerical" support. Meyer proved (and this is not
simple [LM86]) that the 1/Jjk form an orthonormal basis for L 2 (IR). Thus
for all f E L 2 (IR),
L: L: u, 1/Jjk)'l/Jjk,
+oo +oo
!= (42.13)
j=-oo k=-oo
and
j
+oo +oo +oo
L: L: lU, 1/Jjkw =
-oo
!f(tW dt.
j=-oo k=-oo
L: u, 1/Jjk)'l/Jjk
+oo +oo
1 = L: 1j with IJ =
j=-oo k=-oo
v(t)
-2 -1 0 2 3 4
orthogonal. Wc do not assume that the vectors Tk9 in Definition 42.4.2 are
orthogonal, and in important cases they are not.
N
II!- L ckekll-+ 0 as N-+ +oo. (42.14)
k=-N
+oo
Allfll ~2 L lckl 2 ~ Bllfll 2 , (42.15)
k=-oo
which means that f f-+ ( LkEZ Iek 12 ) 112 defines a norm on H that is
equivalent to the original norm on H.
j +oo
cp(t- p)rp(t- q) dt = {0 if p =/: q,
=q
-oo 1 if p
j +oo
-00
cp(.X)~(.X)e-2i7r(p-q).A d.X = {0 if P =/: q,
1 if p = q,
j +oo
-oo
lcp(.X)I2e-2i7rn.A d.X = {0
1
~f
If
n =/: 0,
n = 0.
(42.17)
42.5 Multiresolution analysis and wavelet bases 415
+oo +oo
L ltP'(.X + kW = L 5 ltP"I 2 (n)e 2i7rn\
k=-oo k=-oo
in the sense of Y 1 lf (42 .16) holds (in the sense of L 1 ) , then the Fourier
coeffi.cients of ltP"I 2 satisfy condition (42.17). On the other hand, if we have
(42.17), then
+oo
L ltP'(.X + kW = 1
k=-oo
in the sense of Y '. But by Exercise 21.6, this implies that thc relation
holds for almost every .X E R o
(42.18)
Proof. Since cp E V0, thcre exists a sequence (mk) in l2 (Z) such that
+oo
cp(t) = L mkg(t- k)
k=-oo
where
L
+oo
M(.X) = mke- 2i1rk>-.
k=-oo
One can show that the Poincare process leads essentially to the relation
1
{1 1+oo {1
0<C Jo lm(.XW d.X :S -oo lm(.XWI9(.XW d.X :S D Jo lm(.X)I 2 d.X.
1 +oo
The middle integral is equal to jlm( .X) 12 L 19( .X+ kWd.X, so we have
0 k=-00
1 1 +oo {1
0 < C fo1m(.X)I 2d.X :S fo1m(.XW k~oo 19(.-\ + k)l 2d.X :S D Jo lm(.XW d.X
42.5 Multiresolutionanalys is and wavelet bases 417
for all (ak) in l 2 (Z), which is to say, for all m E L~(O, 1). Butthis can be
true if and only if
+oo
O<C:::; L 19(-\+kW:::;D
k=-00
for almost every ,\ E JH?., which is what we wished to prove. For convenience
we write
Then
1 1
VC:::; N(,\):::; Vf5 and ..fJ5:::; M(,\):::; v7J'
and we conclude that both M and N are in Lgc'(O, 1) and hence in L~(O, 1).
This implies that 4?(-\) = M(,\)g(,\) is well-defined as an element of L 2 (JH?.).
It is clear from this definition that 4? satisfies (42.16); hence {rkcp} is an
orthonormal family.
It remains to show that {rkcp} spans V0, which is by now close to obvious:
Since g(,\) = N(-\)4?(-\) and NE L~(O, 1), there is a sequence (nk) in l 2 (Z)
suchthat
+oo
g(t) = L nkcp(t- k).
k=-oo
Since the functions g(t- k) span V0, this shows that the functions cp(t- k)
span V0 and completes the proof. o
J=-oo J=-oo
The plan is to look for a function 'ljJ E Wo suchthat the functions Tk'l/J form
an orthonormal basis for Wo.
~(2..\) = A(.A)~(.A)
and
L
+oo
~(2..\) = ake-2i1rk-\~(.A) = A(.A)~(.A),
k=-oo
On the other hand, all of the functions cp(t- k) are orthogonal to '1/J, and
therefore
We write this as
B(.>..) = e- 2 in>.A(.>.. + 1/2)0(.>..)
with
0(.>..) = -e2 in>.[A(.>..)B().. + 1/2)- A(.>.. + 1/2)B(.>..)]
and make two observations. Note that 0().. + 1/2) = 0(.>..) and, since A and
B must satisfy (42.20) and (42.22), that
IO(.>..)i = 1. (42.27)
Conversely, it is easy to show that any function 0 with period 1/2 satis-
fying (42.27) will work. A simple family of functions 0 is
42.5.8 Theorem If 'ljJ is defined by (42.21) and (42.28), then the set
of functions {Tk'l/J hEz is an orthonormal basis for Wo and the functions
Proof. The first task is to sort out what has been proved and what remains
to be proved. We assume that we have a multiresolution analysis of L 2 (JR.)
and that we have in hand a function r.p suchthat the 'Pk(t) = r.p(t-k) form an
orthonormal basis for V0 . By definition, W 0 is the orthogonal complement
of Vo in VI, so vl = Vo E9 Wo, and this implies by a change of scale that
\'JH = \rj ffi Wj for all j E Z. The fact that
+=
L2 = E9 Wj
j=-=
This is how we proceed: Define B by (42.28) and 'ljJ by (42.21); then work
the arguments of Proposition 42.5.6 backwards to show that B satisfies
(42.22) and (42.23) and that 'ljJ satisfies relations (42.24) and (42.25). These
are Straightforward computations, and we leave this part as an exercise.
This proves that the functions '1/Jk are in W0 and that they form an orthonor-
mal family in W 0. What remains tobe shown isthat the '1/Jk span W 0. To
do this, we will show the existence ofsequences (ck), (dk), (ek), (fk) E l 2 (Z)
such that the functions cp(t), cp(2t), and 'lj;(t) are related by the following
equations:
L L
+oo +oo
cp(2t) = Ckcp(t- k) + dk'lj;(t- k),
k=-oo k=-oo
+oo +oo
(42.29)
cp(2t- 1) = L ekcp(t- k) + L !k'I/J(t- k).
k=-oo k=-oo
Once we have (42.29) we have the result: Theserelationsshow that for each
n E Z, the function cp(2t- n) can bc expressed as linear a combination of
the 'Pk and the '1/Jk If the '1/Jk do not span W 0, there is a nonzero element
h E Wo such that (h, '1/Jk) = 0 for all k. Since W0..L Vo, (h, 'Pk) = 0 for all
k. But h E V1 = Vo ffi Wo and the cp(2t- k)) form a basis for V1 ; this and
(49.29) imply that h = 0. Hence the '1/Jk must span W0.
The last step is to show that wc do indeed havc (42.29). In the Fourier
domain, (42.29) is equivalent to thc existence of four functions C, D, E, F
belanging to L~(O, 1) such that
0( ~) = C(.X)0(.X) + D(.X)~(.X),
(42.30)
e-i7r>-0( ~) = E(.X)0(.X) + F(.X)~(.X).
While this completes the proof, much can be said about this result and the
questions it raises. A few comments are given below. o
422 Lesson 42. Wavelet Analysis
42.5.9 Remarks
(a) Formula (42.28) provides a relation between the Fourier coefficients
ak and bk of A and B:
bk = (-1) 1-k1-k
This allows one to obtain 'lj; in terms of cp without a Fourier transform,
since by (42.26),
with
L u, 'lj;jk)'lj;jk,
+oo
,j =
k=-00
Fn = L /j.
j=-oo
(42.31)
42.5 Multiresolutionanalysis and wavelet bases 423
[ A(>.) B(>.) ]
A(>. + 1/2) B(>. + 1/2)
is unitary for almost all >., are called conjugate quadraturc filtcrs.
(c) As indicated several times, regularity and localization of the scaling
function <p and the wavelets '1/ljk are necessary for efficient numerical com-
putation. Thus the "minimal" assumptions made about <p in (b) do not
lcad to practical wavelets. If we assumc, however, that
for all m E Z, in addition to assuming that <p satisfies (42.16) and (42.31)
and that I<P(O)I = 1, the whole situation becomes much "smoother." In
this casc, the coefficicnts ak decrcase rapidly at infinity and A E coo.
Furthermore, not only do assumptions about the regularity of <p and the
localization of its derivatives lead to the regularity of '1/J, but thcy also imply
that '1/J has vanishing moments. This analysis can be found in [CR95].
This relation implies that the V0 c V1 and thus that \;j c \;j +1 by a changc
of scale. It was argucd following Definition 42.4.2 that {\;j} is a multireso-
lution analysis of L 2 (JR).
For later use, we take thc Fourier transform of both sides of (42.32) and
write
9(2-\) = G(-\)9(-\), (42.33)
wherc G(,\) = (1 + cos 271'-\)/2. Note that Gis real and even.
Sincc the translates Tkg, k f= 0, are not orthogonal to g, it is necessary
to transform the rkg into an orthonormal family. For this, wc usc Theorem
42.5.3 and definc rp by (42.19):
1k=-oo
+oo
L 19(,\ +
1
kWe2i1rn,.\ d,\ = 119(-\)l2e2i7rn,.\ d,\
IR
l
0
= 9(-\)g(,\)e2ioornA d,\
= l g(t)g(t- n) dt.
i
A simple computation shows that
~
if n = 0,
J. g(t)g(t- n) dt { if n = 1,
othcrwisc.
42.5 Multiresolution analysis and wavelet bases 425
2 1 1
L
+~
k=-~
19(>. + kW = 3 + 3 cos 21r>. = 3[1 + 2cos2 1r>.],
L
+~
.!T M =m = O:nOn,
n=-oo
where (an) is a slowly increasing sequence. In fact, since ME L~(O, 1), O:n
tends to zero as lnl --+ +oo. An application of Proposition 33.2.1 shows
that
--
g * .!T M = g M,
and since the Fourier transform in 1-to-1 on .'? ', we must have
1/J(X)
e 2 i7r>.'I/J(2>.) = S(>.)g(>.),
implies that
'I/JC~1) =(s*g)(t).
Both s and g are even, and it follows that '1/J( (t + 1) /2) is even. The function
'ljJ(t + 1/2) obtained by replacing t with 2t is also even; thus its translate
'ljJ(t) is symmetric araund 1/2.
To summarize, starting with a multiresolution of L 2 (JR.) generated by
the function g = r * r, we have used the constructions describcd in this
lesson to generate the scaling function <p and the spline wavelet '1/J. We
have shown that <p and 'ljJ are real spline functions of degree 1, that <p is
even, and that 'ljJ is symmetric about t = 1/2. We also argued that the
support of <p is JR.; similarly, since S cannot be a polynomial, the support
of 'ljJ is R However, both functions decay exponentially (for a proof sec
[Dau92]). These results generalize to the multiresolution generated by gn.
For a systematic discussion of spline wavelets, we suggest the article by
Charles Chui in [RBC+92]. The spline wavelet 'ljJ and the modulus of its
spectrum are shown, respectivcly, in Figures 42.15 and 42.16. The spline
wavelet of degree 3 is shown in Figure 42.17 and the modulus of its spectrum
is illustrated in Figure 42.18.
42.5 Multiresolutionanalysis and wavelet bases 427
1~(~)1
0 2 3 4 5 6 7 8 9 10 11 12 ~
l{l(x)
1~(~)1
1 2 3 4 5 6 7 8 9 10 1112 ~
42.6 Afternot es
This lesson has been but a brief introduction to the theory and applications
of wavelets. We have presented only a few topics from what has become
a dynamic and productive area of research with a rich theory and a wide
range of applications. In this last section we indicate some other aspects
of the field and provide a few pointers to the literature, which is now
substantial.
A first point concerns history and the sociology of science. Since the be-
ginning in the 1980s of what we call "modern wavelet theory," the field
has been characterized by a healthy interplay between theory and applica-
tions. Simply put, mathematicians have worked in close collaboration with
researchers from other areas of science and engineering, and wavelet theory
has been strongly influenced by applied problems. These revolve naturally
araund signaland image processing, but the signals and images arrive from
diverse fields: astronomy, biology, medicine, hydrodynamics, geophysics,
and, .of course, telecommunications-to mention but a few. The challenge
is to find a field of science or engineering where wavelet techniques have not
been applied, or at least tried. This was not always the case. As mentioned
at the beginning of the lesson, we now sec many older results in mathe-
matics and in signal processing that are now interpreted in the language
of wavelet theory. These results were for the most part unknown outside
their respective communities. Since the initial collaboration between Marlet
and Grossmann, the tradition of interaction and cross-fertilization among
disciplines continues, and there is a resonance in this when we recall that
Fourier was motivated by problems in heat conduction.
We have introduce two kinds of wavelet analysis: continuous wavelet
analysis associated with a family of the form
'lj;ab(t) 1 (t- b)
= ya'lj; -a- , b E IR., a > 0,
(42.38)
R(x) = ~ sin(1rn 2 x)
L.... n2
n=l
42.6 Afternotes 429
suchthat the support of '1/Jr is in [0, 2r + 1], the moments J tn'I/Jr dt = 0 for
0 ::=; n ::=; r, and '1/Jr has about r /5 continuous derivatives. A complete ac-
count can be found in Daubechies's book [Dau92]. Another significant step
was the discovery by Daubechies, Cohen, and Feauveau of a general way to
generate biorthogonal wavelet bases. (A particular example had previously
been constructed by Philippe Tchamitchian.) This means there are two
families {'1/ljk} and {.;j;jk}, each of the form (42.32), that are unconditional
bases for L 2 (JR.) and suchthat
except when j = j' and k = k', in which case it equals 1. A complete discus-
sion of this construction and of why biorthogonal wavelets are interesting
for applications is given in [CR95].
The original French version of this lesson appeared in 1990 at a time when
the only book on wavelets was Yves Meyer's Ondelettes et operateurs I:
Ondelettes [Mey90]. Professor Meyer and his students have played a central
role in the development of wavelet theory, and Meyer's books, both the
technical work cited above and his more widely accessible account [Mey93],
have had an influence on both sides of the Atlantic.
Ten Lectures on Wavelets [Dau92] by lngrid Daubechies was the first
book in English, and it has deservedly become a "best seller." Full accounts
of most of the material in this lesson can be found there.
There are now many books on wavelets in English. Furthermore, all are
accessible to anyone who has understood the material of these 42 lessons.
We have included several books in the References, usually anPotated, that
have not been cited in the text.
430 Lesson 42. Wavelet Analysis
42.7 Exercises
Exercise 42.1 With the notation and hypotheses of Theorem 42.2.1, show
that
K
1 JJ - dadb
R2CJ(a, b) C 9 (a, b) ~=}IR[ _
f(t) g(t) dt
where '1/J(x) = 1 on (0, 1/2); '1/J(x) = -1 on [1/2, 1); and '1/J(x) = 0 otherwise.
(1) Show that {'1/Jikh.kez is an orthonormalsystem in L 2 (Ji).
(2) We know that {'1/Jjk} is an orthonormal basis for L 2 (Ji). Consider the scaling
function t.p = X[o,l) associated with the wavelet basis {'1/Jik}. If n E N* and
dool
and write
.
B = [coo
cw
.
Cll
[CLW67] J.W. Cooley, P.A.W. Lewis, and P.D. Welch. The Fast Fourier
Transform algorithm and its applications. Technical report,
I.B.M. Research, 1967.
[CLW70] J.W. Cooley, P.A.W. Lewis, and P.D. Welch. The Fast Fourier
Transform algorithm. Programming considerations in the cal-
culation of sine, cosine and Laplace transforms. J. Sound Vi-
brations, 12(3):315-337, 1970.
[CR95] A. Cohen and R.D. Ryan. Wavelets and Multiscale Signal Pro-
cessing. Chapman & Hall, London, 1995.