Professional Documents
Culture Documents
Signal Processing
Springer-Verlag Berlin Heidelberg GmbH
Pierre Bremaud
Mathematical Principles of
Signal Processing
Fourier and Wavelet Analysis
, Springer
Pierre Bremaud
Ecole Polytechnique Federale de Lausanne
Switzerland
and
INRIAJEcole Normale Superieure
France
bremaud@ens.fr
987 6 5 4 3 2 1
SPIN 10845428
ToMarion
Contents
Preface xi
A Fourier Analysis in LI 1
Introduction 3
Al Fourier Transforms of Stable Signals 7
A 11 Fourier Transform in L I 7
Al2 Inversion Formula . . . . . . . . 16
B Signal Processing 49
Introduction 51
BI Filtering 55
B 11 Impulse Response and Frequency Response 55
Bl2 Band-Pass Signals . . . . . . . . . . . . . . 68
viii Contents
B2 Sampling 75
B21 Reconstruction and Aliasing . 75
B22 Another Approach to Sampling 82
B23 Intersymbol Interference . 84
B2-4 The Dirac Formalism . . . . . 88
B3 Digital Signal Processing 95
B31 The DFf and the FFf Algorithm 95
B32 The Z-Transform . . . . . . . . . 100
B33 All-Pass and Spectral Factorization 109
B4 Subband Coding 115
B41 Band Splitting with Perfect Reconstruction . 115
B42 FIR Subband Filters 120
References. . . . . . . . . . . 126
Appendix 239
The Lebesgue Integral 241
References. . . . 261
Glossary of Symbols 263
Index 267
Preface
Fourier theory is one of the most useful tools in many applied sciences, part-
icularly, in physics, economics, and electrical engineering. Fourier analysis is a
well-established discipline with a long history of successful applications, and the
recent advent of wavelets is the proof that it is still very alive. This book is an
introduction to Fourier and wavelet theory illustrated by applications in commun-
ications. It gives the mathematical principles of signal processing in such a way
that physicists and electrical engineers can recognize the familiar concepts of their
trade.
The material given in this textbook establishes on firm mathematical ground the
field of signal analysis. It is usually scattered in books with different goals, levels,
and styles, and one of the purposes of this textbook is to make these prerequisites
available in a single volume and presented in a unified manner.
Because Fourier analysis covers a large part of analysis and finds applications
in many different domains, the choice of topics is very important if one wants
to devise a text that is both of reasonable size and of meaningful content. The
coloration of this book is given by its potential domain of applications-signal
processing. In particular, I have included topics that are usually absent from the
table of contents of mathematics texts, for instance, the z-transform and the discrete
Fourier transform among others.
The interplay between Fourier series and Fourier transforms is at the heart of
signal processing, for instance in the sampling theory at large (including multireso-
lution analysis). In the classical Fourier theory, the formula at the intersection of the
Fourier transform and the Fourier series is the Poisson formula. In mathematically
oriented texts, it appears as a corollary or as an exercise and in most cases receives
little attention, whereas in engineering texts, it appears under its avatar, the formula
xii Preface
giving the Fourier transfonn of the Dirac combo For obscure reasons, it is believed
that the Poisson sum fonnula, which belongs to classic analysis, is too difficult,
and students are gratified with a result of distributions theory that requires from
them a higher degree of mathematical sophistication. Surprisingly, in the applied
literature, whereas distribution theory is implicitly assumed to be innate, the basic
properties of the Lebesgue integral, such as the dorninated convergence and the
Fubini theorem, are never stated precisely and seldom used, although these tools
are easy to understand and would certainly answer many of the questions that alert
students are bound to ask. In order to correct this unfortunate tradition, which has
a demoralizing effect on good students, I have insisted on the fact that the c1assical
Poisson fonnula is all that is needed in signal processing to justify the Dirac
symbolism, and I have devoted some time and space to introduce the Lebesgue
integral in a concise appendix, giving the precise statements of the indispensable
tools.
The contents are organized in four chapters. Part A contains the Fourier theory
in LI up to the c1assical results on pointwise convergence and the Poisson sum
fonnula. Part B is devoted to the mathematical foundations of signal processing.
Part C gives the Fourier theory in L 2 . Finally, Part D is concemed with the time-
frequency issue, inc1uding the Gabor transfonn, wavelets, and multiresolution
analysis. The mathematical prerequisites consist of a working knowledge of the
Lebesgue integral, and they are reviewed in the appendix.
Although the book is oriented toward the applications of Fourier analysis, the
mathematical treatment is rigorous, and I have aimed at maintaining a balance
between practical relevance and mathematical content.
Acknowledgments
Michael Cole translated and typed this book from a French manuscript, and Clau-
dio Favi did the figures. Jean-Christophe Pesquet and Martin Vetterli encouraged
me with stimulating discussions and provided the illustrations of wavelet analy-
sis. They also checked and corrected parts of the manuscript, together with Guy
Demoment and Emre Telatar. Sebastien Allam and Jean-Fran~ois Giovanelli were
always there when TEX tried to take advantage of my incompetence. To all of them,
I wish to express my gratitude, as well as to Tom von Foerster, who showed infinite
patience with my prornises to deliver the manuscript on time.
Fourier Analysis in L1
Introduction
ae a2e
-=K-,
at a2x
where e(x, t) is the temperature at time t and at loeation x of an infinite rod, and
Kis the heat eonduetanee. The initial temperature distribution at time 0 is given:
e(x,O) = f(x).
Fourier claimed that his solution was general beeause he was eonvineed that alI21T-
periodie funetions ean be expressed as a trigonometrie series with the eoefficients
Cn = cn(f) = -
21T
1
1
0
2lT
f(t)e -int dt.
lThe definitive form of his work was published in Theorie Analytique de la Chaleur,
Finnin Didot ed., Paris, 1822.
4 Part A Fourier Analysis in L I
true for - l ( < x < +l(. But the mathematicians of that time were skeptical about
Fourier's general conjecture. Nevertheless, when the propagation of heat in solids
was set as the topic for the 1811 annual prize of the French Academy of Sciences,
they surmounted their doubts and attributed the prize to Fourier's memoir, with the
explicit mention, however, that it lacked rigor. Fourier's results that were in any case
true for an initial temperature distribution that is a finite trigonometric sum, and be
it only for this, Fourier fully deserved the prize, because his proof uses the general
tricks (for instance, the differentiation rule and the convolution-multiplication
rule) that constitute the powerful toolkit of Fourier analysis.
Nevertheless, the mathematical problem that Fourier raised was still pending,
and it took a few years before Peter Gustav Dirichlet2 could prove rigorously,
in 1829, the validity of Fourier's development for a large class of periodic func-
tions. Since then, perhaps the main guideline of research in analysis has been the
consolidation of Fourier's ingenious intuition.
The classical era of Fourier series and Fourier transforms is the time when the
mathematicians addressed the basic question, namely, what are the functions adrnit-
ting a representation as a Fourier series? In 1873 Paul Dubois-Reymond exhibited
a continuous periodic function whose Fourier series diverges at O. For almost one
century the threat of painful negative results had been looming above the theory.
Of course, there were important positive results: Ulisse Dini3 showed in 1880 that
if the function is locally Lipschitz, for instance differentiable, the Fourier series
represents the function. In 1881, Carnille Jordan4 proved that this is also true for
functions of locally bounded variation. Finally, in 1904 Leopold Fejeii showed
that one could reconstruct any continuous periodic function from its Fourier coef-
ficients. These results are reassuring, and for the purpose of applications to signal
processing, they are sufficient.
However, for a pure mathematician, the itch persisted. There were more and
more examples of periodic continuous functions with a Fourier series that diverges
at at least one point. On the other hand, Fejer had proven that if convergence is
taken in the Cesaro sense, the Fourier series of such continuous periodic function
converges to the function at all points.
2Sur la convergence des series trigonometriques qui servent arepresenter une fonetion
arbitraire entre des limites donnees, J. reine und angewan. Math., 4,157-169.
3 Serie di Fourier e altre rappresentazioni analitiche delle funzioni di une variabile reale,
Pisa, Nistri, vi + 329 p.
4Sur la serie de Fourier, CRAS Paris, 92, 228-230; See also Cours d'Analyse de l'lfcole
Polytechnique, I, 2nd ed., 1893, p. 99.
5Untersuchungen ber Fouriersehe Reihen, Math. Ann., 51-69.
Introduction 5
j(v) = L
J(t)e-2irrvt dt
is its Fourier transform, where ~ is the set of real numbers. This striking formula
found very nice applications in the theory of series and is one of the theoretical
results founding signal analysis. The Poisson sum formula is the culrninating result
of Part A, which is devoted to the classical Fourier theory.
A 11 Fourier Transform in LI
This first chapter gives the definition and elementary properties of the Fourier
transform of integrable functions, which constitute the specific calculus mentioned
in the introduction. Besides linearity, the toolbox of this calculus contains the
differentiation rule and the convolution-multiplication rule. The general problem
of recovering a function from its Fourier transform then receives a partial answer
that will be completed by the results on pointwise convergence of Chapter A3.
We first introduce the notation: N, Z, Q, ~, C are the sets of, respectively,
integers, relative integers, rationals, real numbers, complex numbers; N+ and ~+
are the sets of positive integers and nonnegative real numbers.
In signal theory, functions from ~ to C are called (complex) signals. We shall
use both terminologies (function, or signal), depending on whether the context is
theoretical or applied.
We denote by L~(~) (and sometimes, for short, LI) the set offunctions f(t) 10
from ~ into C such that
In analysis, such functions are called integrable. In systems theory, they are called
stable signals.
IOWe shall often use this kind of loose notation, where a phrase such as "the function
f(t)" means "the function f : lR ~ c." We shall also use the notation "I" or "fO" with
a mute argument. For instance, "f( - a)" is the function t --+ f(t - a).
Let A be a subset of IR. The indicator function of Ais the function lA : IR 1-+
{O, 1} defined by
lA(t) = 1o
1 if tE A,
ift ~ A.
The function I(t) is called locally integrable if for any closed bounded interval
[a, b] C IR, the function I(t)l[a,bj(t) is integrable. We shall then write
I E L~ loe(lR)
L I/(t)1 2 < 00
E = L I/(t)1 2 dt.
The function I(t) is called locally square-integrable if for any closed bounded
interval [a, b] C IR, the function I(t)l[a,bj(t) is square-integrable. We shall then
write
I E L~,loe(lR)
EXERCISE AI.I. Give an example 01 a function that is integrable but not 01finite
energy. Give an example 01 a function that is 01finite energy but not integrable 01
finite energy. Show that
Fourier Transform
We can now give the basic definition.
DEFINITION AI.I. Let s(t) be a stable complex signal. The Fourier transform (FT)
ofs(t) is thefunctionfrom:IR into C:
(Note that the argument of the exponential in the integrand is -2i7rvt.) The
mapping from the function to its Fourier transform will be denoted by
Doppler s(at) Fr
--+ -IsA{V}
-
lai a
2(s(v - Vo ) + sA( v
Fr 1 A
s(t) cos(27rvot) --+ + Vo )). (2)
o v -vo o +vo v
T
1
-T/2 o +T/2 1 2 3
T T T
We will show that the Gaussian pulse is its own Fr, that is,
(4)
In order to compute the corresponding Fourier integral, we use contour integration
in the complex plane. First, we observe that it is enough to compute the Fr s( v)
for v 2: 0, since this Fr is even (see Exercise A1.4). Take a 2: v (eventually, a
will tend to 00).
Consider the rectangular contour Y in the complex plane (see Fig. A1.3),
Y = Yl + Y2 + Y3 + Y4,
where the Yi 's are the oriented line segments
')'4
-a +a
Figure A1.3. The integration path in the proof of (4)
A 11 Fourier Transform in L I 11
We denote by -Yi the oriented segment whose orientation is opposite that of Yi.
i
Wehave
e- rrz2 dz = 11 + /z + h + 14,
where li is the integral of e- rr Z2 along Yi. Since the latter integrand is a holomorphic
function, by Cauchy's theorem (see, for instance, Theorem 2.5.2, p. 83, of [Al],
or Theorem 2.2, p. 101, of [A6]),
i e- rrz2 dz = 0,
and therefore,
h + /z + h + 14 = O.
We now show that
lim /z !im 14 = o.
= a-+oo
a~oo
l l
then
/z = v
e-rr(a+itf i dt = v
e-rr(a2-t2)e-2irrat i dt.
l/zl .:::: l a
e-rr(a-t)(a+t) dt .:::: l a
e-rra(a-t) dt
= e- rra 21o a
errat 1
dt = -(1 _
Jra
e- rra 2 ),
where the last quantity tends to 0 as a tends to +00. A sirnilar conclusion holds
for 14, with sirnilar computations. Therefore,
!im (h
a-HXl
+ h) = 0,
that is,
(5)
wehave
1-)'3
= l+ a e-n(iv+tf dt
-a
Therefore,
The Fr of the Gaussian pulse can be obtained by other means (see Exercise
A 1.16). However, in other cases, contour integration is often necessary.
Using contour integration in the complex plane, we show that, for a > 0,
1
s(t) = e-at IlR+(t) Fr
~ s(v)
A
= . (6)
a+2mv
1
First observe that
s(v) = 1
o
00
e- 2znvt
.
e- at dt = . 1
2mv + a 0
00
.
e-2znvt-a\2inv + a) dt
= 1
2inv + a
1 Y
e- Z dz.
(The reader is refered to Fig. A1.4 for the definition ofthe lines y, YJ. Y2, and Y3.)
Therefore, it suffices to show that
i e- z dz = 1.
By Cauchy's theorem,
1 Yl
e-Z dz + 1Y2
e- Zdz + 1 )'3
e- Zdz = 0.
A 11 Fourier Transform in LI 13
2i7f1/
Convolution-Multiplication Rule
THEOREM AI.I. Let h(t) and x(t) be two stable signals. Then the right-hand side
oJ
is defined Jor almost all t and defines almost everywhere a stable signal whose FT
is y(v) = h(v)x(v).
Proof" By ToneIli's theorem and the integrability assumptions,
The integral IIR h(t - s )x(s) ds is therefore weIl defined for almost aIl t. Also,
=L Lh(t - s)e-2irrv(t-s)x(s)e-2irrvs ds dt
= h(v)x(v).
The funetion y(t), the convolution of h(t) with x(t), is denoted by
y(t) = (h * x)(t).
We therefore have the convolution-multiplication rule,
(h * x)(t) --+
Fr A
h(v)x(v). (8)
EXAMPLE AI.L The convolution of the rectangular pulse reeT (t) with itself is the
triangular pulse of base [- T, + T] and height T,
TriT(t) = (T - Itl)1[-T,+T](t).
T2
&
T
A
I
I
I
i
C'>~
I
I
. 1
~C'>
j I
-T +T -~ -~ -~ 0 ~ ~ ~
TriT(t)
Proof' The Fr of a rectangular pulse s(t) satisfies Is(v)1 ::; K/lvl [see Eq. (3)].
Hence every signal s(t) that is a finite linear combination of indicator functions
of intervals satisfies the same property. Such finite combinations are dense in
Lb(l~) (Theorem 28 of the appendix), and therefore there exists a sequence sn(t)
of integrable functions such that
we deduce that
::; -K n +
lvi
iIR
Is(t) - sn(t)1 dt,
11 Riemann, B., (1896), Sur la possibilite de representer une fonetion par une serie
trigonometrique, Oeuvre Math., p. 258.
16 Al. Fourier Transfonns of Stable Signals
THEOREM Al.3. Let f(t) be a 2:rr-periodic locally integrable function, and let
g: [a,b] CbeinC I , where [a,b] ~ [-:rr, +:rr]. Then
f-*
b
lim l fex - u)g(u) sin(Au) du = 0
....... 00 a
uniformly in x.
Proof For arbitrary E: > 0, choose a 2:rr-periodic function h(t) in Cl such that
/(A) =l b
hex - u)g(u) sin(Au) du
COS(AU) Ib b
=- hex - u)g(u) - - +l [hex - u)g(u)]
, COS(AU)
du.
A a a A
Since h E Cl and is periodic, h and h' are uniformly bounded. The same is true of
g, g' (g is in Cl). Therefore,
lim /(A)
....... 00
=0 uniformly in x .
Now,
11 b
fex - U)g(U)Sin(AU)dUI
:S I/(A)I +l b
Ih(x - u) - fex - u)llg(u)1 sin(Au) du
THEOREM At.4. Let set) be an integrable complex signal with the Fourier
transform s(v). Under the additional condition
holdsfor almost all t.lf set) is, in addition to the above assumptions, continuous,
equality in (12) holds for all t.
(Note that the exponent ofthe exponential ofthe integrand is +2irrvt.)
EXERCISE At.ll. Check that the above result is true for the signal
(a E lR, a > 0, a E C).
Proof' We now proceed to the proof of the inversion formula. (lt is rather tech-
nical and can be skipped in a first reading.) Let set) be a stable signal and consider
the Gaussian density function
with the Fr
We first show that the inversion formula is true for the convolution (s * h u )(t).
Indeed,
= JR
{ s(u)hu(u)e Zc;2' -'<-(t)du
I
(12
= (s * h u )(t).
18 Al. Fourier Transforms of Stahle Signals
Thus, we have
L
to
s(v)e2i1rvt dv
for all t E IR. If we can prove that when u ..I- 0 the function on the left-hand side of
(14) converges in L~(IR) to the function s(t), then, for almost all t E IR, we have
the announced equality (Theorem 25 of the appendix).
To prove convergence in L~(IR), we observe that
(using the fact frr~. hu(u)du = 1), and therefore, defining f(u) = iJR Is(t - u)-
s(t)1 dt,
Now, If(u)1 is bounded (by 2 iJR Is(t)1 dt). Therefore, iflimu-l-o f(u) = 0, then, by
dominated convergence,
f(u) ::::: d(s( - u), sn(' - u)) + L ISn(t - u) - sn(t)1 dt + d(s(), snO),
where
Suppose that, in addition, set) is continuous. The right-hand side of (12) defines
a continuous function because s(v) is integrable. The everywhere equality in (12)
follows from the fact that two continuous functions that are almost everywhere
equal are necessarily everywhere equal (Theorem 8).
f ~dU,
l(t)= t > 0.
J.R.t+u
EXERCISE Al.13. Deduce from the Fourier inversion formula that
L(Si~(t) Y dt = Jr.
Exercise 1.14 is very important. It shows that for signals that cannot be called
pathological, the version of the Fourier inversion theorem that we have in this
chapter is not applicable, and therefore we shall need finer resuIts, which are given
in Chapter A3.
EXERCISE AI.14. Let set) be a stable right-continuous signal, with a limit from
the left at all times. Show that if s(t) is discontinuous at some time to, its FT cannot
be integrable.
Regularization Lemma
In the course of the proof of Theorem A1.4, we have used a special case of the
regularization lemma below, which is very useful in many circumstances.
DEFINITION AI.2. A regularizing function is a nonnegative function h a : IR ---+ IR
depending on a parameter a > and such that
lim
a'\-O
l-a
+a
ha(u) du = 1, forall a > 0,
lim
a'\-O
f
JIR
I(s * h a )(t) - s(t)1 dt = 0.
20 Al. Fourier Transforms of Stable Signals
Proof" We ean use the proof of Theorem A1.4, starting from (15). The only
plaee where the speeifie form of h" (a Gaussian density) is used is (16). We must
therefore prove that
lim ( J(u)h,,(u) =0
"-1-0 JIR
independently. Fix e > O. Sinee limuto J(u) = 0, there exists a = aCe) such that
We shall see how differentiation in the time domain is expressed in the frequeney
domain.
THEOREM A1.S. (a) Ifthe integrable signal set) is such that tks(t) E LU~)Jor
alt 1 :s k :s n, then its FT is in Cn , and
(b) IJ the signal set) E cn and if it is, together with its n first derivatives,
integrable, then
s(v) = L e-2invIs(t)dt,
A12 Inversion Forrnula 21
we can differentiate k times under the integral sign (see Theorem 15 and the
hypothesis t ks(t) E Lb(IR)) and obtain
(b) It suffices to prove this for n = 1, and iterate the result. We first observe that
limlaltoo s(a) = O. Indeed, with a > 0, for instance,
s(a) = s(O) + l a
s'(t) dt,
and therefore, since s'(t) E Lb(IR), the limit exists and is finite. This limit must be
obecause s(t) is integrable. Now, the Fr of s'(t) is
[ e-2irrvts'(t)dt = lim [+a e-2irrvts'(t)dt.
JIR atoo La
Integration by parts yields
a
i : e-2irrvts'(t)dt = (e-2irrvts(t)):: + i:a(2i:n:V)e-2irrvtS(t)dt.
It then suffices to let a tend to 00 to obtain the announced result.
EXERCISE AI.IS. Let s(t) be a stable signal with a Fourier transform with compact
support. Show that s(t) E Coo, that all its derivatives are integrable, and that the
kth derivative has the FT (2i:n:v)k s (v).
EXERCISE AI.16. Give a differential equation satisjied by the Gaussian pulse, and
use it to deduce its Fourier transform. Could you do the same to prove (6)?
where B(x, t) is the temperature at time t and at location x of the rod with heat
conductance K, and with the given initial temperature distribution
B(x, 0) = f(x). (18)
or
As we mentioned earlier, Fourier considered the finite rod heat equation, which
receives a similar solution, in terms of Fourier series rather than Fourier integrals
(see Chapter A2). The efficiency of the Fourier method in solving differential or
partial differential equations of mathematical physics has been, after the pioneering
work of Fourier, amply demonstrated 12 .
12See, for instance, the classic text of 1. N. Sneddon, Fourier Transfonns, McGraw-Hill,
1951; Dover edition, 1995.
A2
Fourier Series of Locally Stable
Periodic Signals
Fourier Coefficients
A periodic signal is neither stable nor of finite energy unless it is almost everywhere
null, and therefore, the theory of the preceding Chapter is not applicable. The
relevant notion is that of Fourier series. (Note that Fourier series were introduced
before Fourier transforms, in contrast with the order of appearance chosen in this
text.) The elementary theory of Fourier series of this section is parallel to the
elementary theory of Fourier transforrns of the previous section. The connection
between Fourier transforrns and Fourier series is made by the Poisson sum formula,
of which we present a weak (yet useful) version in this chapter.
A complex signal s(t) is called periodic with period T >
for all t E ~,
(or T -periodic) if,
s(t + T) = s(t).
AT -periodic signal s(t) is locally stable, or locally integrable, if s(t) E Lb([O, Tl),
that is,
l T
Is(t)1 dt < 00.
A T -periodic signal s(t) is locally square-integrable if s(t) E L~([O, Tl), that is,
l T
Is(t)1 2 dt < 00.
One also says in this case that s(t) hasfinite power, since
Sn = -I
T
l0
T
s(t)e- 2'lJr TI
n dt,
(19)
(20)
EXERCISE A2.3. Compute the Fourier coefficients of the T -periodic signal s(t)
such that on [-T /2, +T /2), s(t) = 1[-a~,+a~l(t), where a E (0, 1).
EXERCISE A2.4. Let s(t) be a T -periodic locally stable signal with nth Fourier
coefficient Sn. Show that limlnltoo Sn = O.
One often represents the sequence {Sn}nEZ of the Fourier coefficients of a T-
periodic signal by "spectrallines" separated by 1/ T from each other along the
frequency axis. The spectralline at frequency n / T has the complex amplitude Sn
(see Fig. A2.1). This is sometimes interpreted by saying that the FT of s(t) is
s(v) = I)n(V -
nEZ
f)'
where (t) is the Dirac generalized function (see Section B2-4).
EXERCISE A2.5. Let s(t) be a T -periodic locally stable signal with nth Fourier
coefficient Sn. What is the nth Fourier coefficient of s(t - a), where a E llV What
can you say about the period and the Fourier coefficients of the signal s(t / a),
where a > O?
A21 Fourier Series in Lloc 25
Convolution-MuItiplication Rule
THEOREM A2.1. Let x(t) be a T -periodie locaily stable signal, and let h(t) be a
stable signal. The signal
is almost everywhere weil defined, T -periodie, and locaily stable. Its nth Fourier
coefficient is
Yn
A
= hA(n)
T x n, A
(22)
Now
x(v) fI(v)
-~
I -q,I I I I I I
-~ 0
123
T T T v / "
" ,
and hence by the usual argument (see the proof of Theorem A 1.1),
for almost aIl t E lR. Thus, y(t) is almost everywhere weIl defined by (21). Also,
= L x(t - s)h(s)ds,
which shows that y(t) is periodic with period T. The same argument as in the proof
of Theorem AU shows that y(t) is locally stable. FinaIly,
l
Yn = -1
T 0
T
y(t)e- Z'l1!'j't
n dt
=- l 1 1 T T
hT(t
- n dt ds
- s)x(s)e- Z'l1!'j't
T 0 0
A22 Inversion Formula
and therefore,
Pr(t) :::: O. (24)
Also,
-1 I+ / Pr(t) dt
T 2
= 1. (25)
T -T/2
In view of the above expression of the Poisson kernel, we have the bound
1 [ (1 - r 2 )
- Pr(t)dt <
1 1 _ e 2i1ry 1
T [-t,+t]\[-e,+s] - 2 '
!im
rtl
-1
1
T
+t
-t
cp(t)Pr(t) dt = cp(O),
for all bounded, continuous cp : ffi. -+ <C (same proof as in Lemma ALl).
The following result is similar to the Fourier inversion formula for stable signals
(Theorem A1.4).
THEOREM A2.2. Let set) be aT -periodie localty stable complex signal with Fourier
coefficients {sn}, n E Z. lf
(27)
lfwe add to the above hypotheses the assumption that set) is a continuousfunction,
then the inversion formula (28) holds for all t.
Proof The proof is similar to that of Theorem A1.4. We have
LSnrlnle2i1ryl
11+[
= _ 2 s(u)Pr(t - u)du, (29)
nEZ T -t
and
10r I10r = 0,
T T
!im s(u)Pr(t - u) du - S(t)1 dt
rtl T
that is: The right-hand side of (29) tends to set) in Lb([O, T]) when r t 1. Since
LnEZ ISn I < 00, the function of t in the left-hand side of (29) tends toward the
28 A2. Fourier Series of Locally Stable Periodie Signals
function LnEZ sne+2irr(n/T)t, pointwise and in L~([O, T]). The result then follows
from Theorem 25.
The statement in the case where set) is continuous is proved exactly as the
corresponding statement in Theorem A1.4.
As in the case of stable signals, we deduce from the inversion formula the
uniqueness theorem.
COROLLARY A2.1. Two locally stable periodic signals with the same period T
that have the same Fourier coefficients are equal almost everywhere.
EXERCISE A2.6. Compute
using the expression ofthe Fourier coefficients ofthe 2-periodic signal set) such
that
fort E [-1,+1].
EXERCISE A2.7. Let x(t) be a T -periodic locally stable signal with nth Fourier
coefficient x n such that
(30)
This aesthetic formula has a number of applications in signal processing (see Part
B).
The next result establishes the connection between the Fourier transform and
Fourier series, and is central to sampling theory. It is a weak form of the Poisson
sum formula (see the discussion after the statement of the theorem).
'THEOREM A2.3. Let set) be a stable complex signal, and let 0 < T < 00 be
fixed. The series LnEZ set + nT) converges absolutely almost everywhere to a
T -periodic locally integrable function <I>(t), the nth Fourier coefficient of which is
(l/T)s(n/T).
We paraphrase this result as follows: Under the above conditions, the function
<I>(t) := L set + nT) (31)
nEZ
A22 Inversion Forrnula 29
In particular,
Therefore, the series LnEZ set +nT) converges absolutely for almost all t. In part-
icular, <I>(t) is weH defined (define it arbitrarily when the series does not converge).
This function is c1early T -periodie. We have
{T 1<I>(t)ldt = {T ILS(t+nT)ldt
10 10 nEZ
cn(<I = -1
T
l0
T
<I>(t)e- 2"''in l dt
=~
T
{T
Ja
!LS(t + kT)e-
kEZ
2i1C ',f(t+kTl! dt
1 ~ (n )
= T1 JJR
{
s(t)e- l1C'it dt = T S T .
2 n
We have a function as weH as its formal Fourier series. When both are equal
everywhere, we obtain the strong Poisson sum formula. The next exercise gives
conditions for this.1t will be improved by Theorem A3.12.
EXERCISE A2.S. Let set) be a stable signal with the FT s(v), and suppose that
(a) LnEZ set + nT) is a continuous function, and
(b) LnEZ Is(n/T)I < 00.
This result challenges one to obtain conditions that a locally integrable 2rr-
periodic function f must satisfy in order for its Fourier series to converge to
f. Recall that the Fourier series associated with a 2rr-periodic locally integrable
function f is the formal Fourier series
(33)
cn(f) = _1
2rr
j+Jr f(u)e-inu du.
-Jr
(34)
The series (33) is calledformal as long as one does not say something about its
convergence in some sense (pointwise, almost everywhere, in LI, etc). If one has
no more than the condition that f is 27T -periodic and locally integrable, the worst
can happen, as Kolmogorov's theorem shows.
The purpose of this section is to find reasonable conditions guaranteeing
convergence as n -+ 00 of the truncated Fourier series
+n
sI (x) = L ck(f)e ikx . (35)
k=-n
We have to specify (1) in what sense this convergence takes place and (2) what
the limit iso Ideally, the convergence should be pointwise and to fitself. The next
exercise gives a simple instance where this is true.
EXERCISE A3.1. Assume that the trigonometric series
+n
Sn(t) = L Ck eikt
k=-n
converges uniformly to some function f(t). Show that in this case, for all k E Z,
Ck = ck(f).
Dirichlet's Integral
We will first express the truncated series sI in a form suitable for analysis. For
this we write
IL I
-Jr
(the function in the right-hand side is called the Dirichlet kerne!) and therefore,
f 1 j+Jr sin((n + i)(x - s))
Sn (x) = -2
7T -Jr
. (( _ )/2)
sm x s
fes) ds.
Performing the change of variable x - s = u and taking into account the fact that
fand the Dirichlet kernel are 27T-periodic, we obtain
I
S!(x)-A=-
j+Jr sin((n
.
+ ! )u)
2 (f(x+u)-A)du (39)
27f -Jr sm(u/2)
iJr
or, equivalently,
1 sin((n + ! )u)
S!(x) - A = - . 2 {fex + u) + fex - u) - 2A}du. (40)
27f 0 sm(u/2)
Therefore, in order to show that, for a given x E IR, S! (x) tends to A as n -+ 00,
it is neeessary and suffieient to show that the Diriehlet integral in the right-hand
side of (39) eonverges to zero as n -+ 00.
The localization principle states that the eonvergenee of the Fourier series is a
loeal property. More preeisely:
THEOREM A3.2. lf fand gare two locally integrable 27f -periodic complex-valued
functions such that, for a given x E IR and some 8 > 0, it holds that f (t) = g(t)
whenever t E [x - 8, x + 8], then
lim{S!(x) - S!(x)} = O.
ntoo
= -
1 j+Jr sin((n + !)u) w(u) du,
27f -Jr
where
fex + u) - g(x + u)
w(u) = l lul ->8 . ( u /2)
sm
is integrable over [0, 27f]. The last integral therefore tends to zero as n -+ 00 by
the Riemann-Lebesgue lemma.
lim S!(x) = A
ntoo
34 A3. Pointwise Convergence of Fourier Series
:s Ti,
.1
if,for some < 8
11m
8
.
smn ~(u)
+ !)u) - - du = 0, (41)
ntoo 0 uj2
where
~(u) = f(x + u) + f(x - u) - 2A. (42)
1 8
sinn + !)u) v(u)du, (44)
where
Dini's Theorem
THEOREM A3.4. Let f be a 2Ti-periodic locally integrable complex-valued
function and let x E IR. Iffor some < 8 :s Ti and some A E IR, the function
f(x + t) + f(x - t) - 2A
t --+
lim
ntoo
S! (x) = A.
Proof" The hypothesis says that the function ~(u)ju, where ~ is defined in (42),
is integrable, and therefore condition (41) ofTheorem A3.3 is satisfied (Riemann-
Lebesguelemma).
COROLLARY
A3.1. If a 2Ti -periodie locally integrable complex-valued function
f(t) is Lipschitz continuous of order IX > about x E IR, that is,
If(x + h) - f(x)1 = O(lhIO') as h --+ 0,
for some constant K and for all t in a neighborhood of zero, and 1/ltI 1-0: is
integrable in this neighborhood, because I-ex< 1. Dini's theoremA3.4 concIudes
the proof.
COROLLARY A3.2. Let f(t) be a 21T-periodic locally integrable complex-valued
function, and let x E lR be such that
f(x + 0) = lim f(x + h) and f(x - 0) = lim f(x - h)
hW hW
exist and are finite, and further assume that the derivatives to the left and to the
right at x exist. Then
. SI() _
I1m n X -
f(x + 0) + f( x - 0)
ntoo 2
Prao!" By definition, one says that the derivative to the right exists if
lim f(x + t) - f(x + 0)
ttO t
exists and is finite, with a similar definition for the derivative to the left. The
differentiability assumptions imply that
lim f(x + t) - f(x + 0) + f(x - t) - f(x - 0)
ttO t
as announced in the last corollary. For t = n /2, we obtain the remarkable identity
nIl 1 1
4=1-3+:5-7+
Jordan's Theorem
Jordan 's convergence theorem features funetions of bounded variation.
DEFINITION A3.1. A real-valued function q; : lR f-+ lR is said to have bounded
variation on the interval [a, b] C lR if
n-l
sup L !q;(Xi+l) - q;(Xi)! < 00, (45)
'D i=O
where the supremum is over all subdivisions D = {a = Xo < Xl < ... < Xn = b}.
We quote without proof the fundamental result on the strueture of bounded
variation funetions.
THEOREM A3.5. A real-valued function q; has bounded variation over [a, b] if and
only ifthere exist two nondecreasing real-valuedfunctions q;l, q;2 such that,for all
tE [a, b],
In partieular, for all X E [a, b), q; has a limit to the right q;(x + 0); for all
X E (a, b], it has a limit to the left q;(x - 0); and the diseontinuity points of q;(t)
in [a, b] form a denumerable set, and therefore a set ofLebesgue measure zero.
The proof is omitted.
EXERCISE A3.2. Let f E L~(lR). Show that, for any B > 0,
I-B B
+ j(v)e2irrvt dv = 2B { f(t
JR
+ s)sine (2Bs) ds,
and use this to study the pointwise convergence of the left-hand side as B tends to
infinity, along the lines of the current chapter.
The funetion
2B sine (2Bt)
is also ealled Dirichlet's kernei.
A31 Dini's and Jordan's Theorems 37
EXERCISE A3.3. Let!t and h be the 2rr -periodic functions defined on (-rr, +rr]
by
!t(x) = x,
Compute their Fourier coefficients, and use this to compute
L (_l)n,
n~l n
an = -1
rr
10
2Jr
f(t)cos(nt)dt, bn = -1
rr
1
0
2Jr
f(t)sin(nt)dt.
Of course, the series in (48) is purely formal when no additional constraints are
put on f(t) in order to guarantee its convergence. Now, the function F(t) defined
for tE [0, 2rr) by
is 2rr-periodic, is continuous (observe that F(O) = F(I) = 0), and has bounded
variation on finite intervals.
Therefore, by Jordan's theorem its Fourier series converges everywhere, and for
all x E lR,
00
An = -1
rr
10
2Jr
F(t) cos(nt) dt
1 [
rr
sin(nx)
=- F(x)--
n
]2Jr
0
-1
nrr
1 0
2Jr
(f(t) - !ao) sin(nt) dt
= - -1
nrr 0
1 2Jr
b
f(t)sin(nt)dt = _..!:,
n
38 A3. Pointwise Convergence of Fourier Series
1
F(x) = zA ~
o+ ~ {an-;; sm(nx) bncos(nx) }.
. - -;; (50)
1 ~bn (51)
zA o = L...- - .
n=l n
Since A o is finite we have shown, in particular, that L~l bn/n converges for any
sequence {b n }n2:1 of the form
I~ ;
Consider the 2n -periodic function defined in the interval ( - n, + n] by
ifx > 0,
J(x) = if x < 0,
n x
---- if x < O.
2 2
The partial sum of its Fourier series is
f _ ~ sin(nx)
Sn (x) - L...- - - .
k=l n
phenomenon, which can be observed whenever the function has a point of discon-
tinuity. The proof of (52) for this special case keeps most of the features of the
general proof, which is left for the reader. In this special case,
f
Sn (x) = l
o
x sin((n + !)t) dt -
2 sin(!t)
x
-.
2
Now,
l o
x sin((n
2 sin('it)
+ !)t)
1 dt
= l
o
x sin(nt)
- - dt
t
+ l 0
x ,
sm(nt)
( cos(!t)
1
2 sin('it)
-
1)
-t dt
+ - ll
2 0
X
cos(nt)dt.
The last two integrals converge uniformly to zero (by the uniform version of the
Riemann-Lebesgue lemma). Also,
1 o
~ sin(nt)
- - dt
tot
= l n sin(t)
- - dt ::::: 1.18 -:rr:rr
> -
2 2
.
A32 Fejer's Theorem
sI
The Fourier (t) series of a 2:rr-periodic locally integrable function I converges
to I(t) for a given t only under certain conditions (see the previous section).
However, Cesaro convergence of the series requires much milder conditions. For
a 2:rr -periodic locally stable function I, Fejer's sum
(53)
behaves more nicely than the Fourier series itself. In particular, for continuous
functions, it converges pointwise to the function itself. Therefore, Fejer's theorem
is a kind of inversion formula, in that it shows that for a large dass of periodic
functions (see the precise statement in Theorem A3.11 ahead), the function can be
recovered from its Fourier coefficients.
40 A3. Pointwise Convergence of Fourier Series
Fejer's Kernel
Take the imaginary part of the identity
n-l 1_ inu
' " ei(k+l/2)u = e iu / 2 e .
~ 1- e lU
k=O
to obtain
sI
Starting from Dirichlet's integral expression for (t) [cf, Eq. (37)], we obtain, in
view of the identity just proven, Fejer's integral representation of a! (x),
{+Jr (+Jr
a!(x) = LJr Kn(u)f(x-u)du= LJr Kn(x-u)f(x)du, (54)
where
I sin2(~nt)
Kn(t) = 2 1 (55)
2n:rr sin ('it)
is, by definition, Fejer's kernel. It has the following properties:
(56)
i:
and [letting f(t) = 1 in (54)],
Jr Kn(u) du = 1. (57)
lim
ntoo
j -6
+C
Kn(u) du = 1. (59)
The last four properties make ofFejer's kernel a regularization kerneion [-:rr, +:rr]
(by definition of a regularization kernel).
Cesaro Convergence for Fourier Series of Continuous Functions
We first treat the case of continuous functions, because the result can be ob-
tained from the basic principles of analysis, in particular, without recourse to
the Riemann-Lebesgue lemma.
THEOREM A3.7. Let f(t) be a 2:rr-periodic continuousfunction. Then
lim sup la! (x) - f(x)1 = O. (60)
ntoo XE[-Jr,+Jr]
A32 Fejer's Theorem 41
i:
Proof' From (54) and (56), we have
n
la! (x) - l(x)1 ::s KnCu) I/(x - u) - l(x)1 du
= 1+ +8
-8
[
[-n,+n]\[-8,+8]
=A+B. (61)
For a given 8 > 0, ehoose 8 sueh that I/(x - u) - l(x)1 ::s 8/2 when lul ::s 8.
Note that I is uniformly eontinuous and uniformly bounded (being aperiodie and
eontinuous funetion), and therefore 8 ean be ehosen independently of x. We have
A::s
81+8 Kn(u)du ::s 2'8
2 -8
By (57) and (59), B ::s 8/2 for n sufficiently large. Therefore, for n suffieiently
large, A + B ::s 8.
Fejer's theorem for eontinuous periodie funetions is the key to important ap-
proximation theorems. The first one is for free. We eall a trigonometrie polynomial
any finite trigonometrie sum of the form
L
+n
p(x) = Ck eikx .
-n
Proof' Use Theorem A3.7 and observe that a! (x) is a trigonometrie polynom-
iill.
From this, we obtain the Weierstrass approximation theorem.
THEOREM A3.9. Let I : [a, b] 1-+ C be a continuousfunction. Select an arbitrary
8 > O. There exists a polynomial P(x) such that
Proof' First, suppose that a = 0, b = 1. One ean then extend I : [0, 1]] 1-+ C
to a funetion still denoted by I, I : [-n, +n]] 1-+ C, that is eontinuous and
sueh that IHn) = I( -n) = O. By Theorem A3.8, there exists a trigonometrie
42 A3. Pointwise Convergence of Fourier Series
We shall first obtain for the Fejer's sum the result analogous to Theorem A3.3.
First, from (54), we obtain
a!(x)-A= -
1 irr sin. 222( !nu) {f(x+u)+f(x-u)-2A}du. (63)
1
2nrr D sm (:zu)
THEOREM A3.10. For any x E IR. and any constant A,
lima!(x) = A (64)
ntoo
. 1
11m
ntoo
-
n
1 D
8 . 4J(u)
sm 2(!nu) - 2 -
u
du = 0, (65)
where
!
In 18
r sin 2(!nu) 4J(U) du l < ! r 14J(u)1 du
sin 2 (!u) - n 18 sin 2 (!u)
A33 The Poisson Formula 43
t
1
tends to 0 as n 00. We must therefore show that
1 8 sin 2( !nu)
_ 2 fjJ(u) du
n 0 sin 2 (!u)
tends to 0 as n t 00. However, (65) guarantees this because
:'S -
n
11 (10
8
. 2( 1
SIll ZU
) - 1I
12 IfjJ(u)1 du
ZU
1
11
-
n 0
8 sin2(!nu)
u
; fjJ(u) du
1
< -
c 1ry sin 2(!nu)
du + -1 1 -IfjJ(u)1-
8
du.
n 0 u2 n ~ u2
The last integral is bounded; and therefore, the last term goes to 0 as n t 00. As
for the penultimate term, it is bounded by Ac, where
A = 1o
00 sin2(!v)
v
2 dv < 00.
A33 The Poisson Formula
The following corollary of F6jer's theorem will play the key role for the proof of
the Poisson sum formula (Theorem A3.l2).
44 A3. Pointwise Convergence of Fourier Series
COROLLARY A3.1. Let f be a 2rc -periodie locally integrable function and suppose
that, for some x E lR.,
lim
ntoo
0'1 (x) = A.
From F6jer's theorem and (a),
lim
ntoo
0'1 (x) = fex).
We have already given a weak version of the Poisson sum formula in Section
A22. A most interesting situation is when the function cI>(t) defined by (31) is
equal to its Fourier series for all t E lR., that is,
THEOREM A3.12. Let set) be a stable complex signal, and let 0 < T < 00 be
fixed. Assume in addition that
Here are two important cases for which the strong Poisson sum formula holds.
COROLLARY A3.1. Let set) be a stable complex signal, and let 0 < T < 00 be
fixed. If, in addition, L set + nT) converges everywhere to a continuousfunction
that has bounded variation, then the Poissonformula (68) holds.
Praof' We must verify conditions (1) and (2) ofTheorem A3.12. Condition (1)
is part of the hypothesis. Condition (2) is a consequence of Iordan's theorem
A3.6.
EXAMPLE A3.1. If set) is continuous, has bounded support, and has bounded
variation, the Poisson sumformula (68) holds.
A33 The Poisson Formula 45
for some a > 1, then the Poisson formula (68) holds for alt < T < 00.
The Poisson formula can be used to replace aseries with slow convergence by one
with rapid convergence, or to obtain some remarkable formulas. Here is a typical
example. For a > 0,
Since
Ls(t +n) = Le-2nalt+nl
nEZ nEZ
is a continuous function with bounded variation, we have the Poisson formula, that
is,
Therefore,
1 n 1 + e- 2na 1
L a 2 + n 2 = 2a 1 - e- 2na 2a 2 .
nO": 1
Letting a ~ 0, we have
The general feature of the above example is the following. We have aseries
that is obtained by sampling a very regular function (in fact, C OO ) but also slowly
46 A3. Pointwise Convergence of Fourier Series
h(t-2T)
~ 1
[jU[jU[j[j[j~
-3T -2T -T 0 T 2T 3T
-3T -2T -T O T 2T 3T
decreasing. However, because of its strong regularity, its Fr has a fast decay. The
series obtained by sampling the Fr is therefore quickly converging.
Radar Return Signal
Let s(t) be a signal ofthe form
s(t) = (I>(t -
nE'L
nT) f(t). (70)
(We may interpret h(t - nT) as a return signal of the nth pulse of a radar after
reftection on the target, and f (t) as a modulation due to the rotation of the antenna. )
The Fr ofthis signal is (see Fig. A3.1)
References
[Al] Ablowitz, M.J. and Jokas, A.S. (1997). Complex Variables, Cambridge University
Press.
[A2] Bracewell, R.N. (1991). The Fourier Transform and Its Applieations, 2nd rev. ed.,
McGraw-Hil1; New York.
[A3] Gasquet, C. and Witomski, P. (1991). Analyse de Fourier et Applieations, Masson:
Paris.
[A4] Helson, H. (1983). Harmonie Analysis, Addison-Wesley: Reading, MA.
[A5] Katznelson, Y. (1976). An Introduetion to Harmonie Analysis, Dover: New York.
[A6] Kodaira, K. (1984). Introduetion to Complex Analysis, Cambridge University Press.
[A7] Krner, T.W. (1988). Fourier Analysis, Cambridge University Press.
References 47
[A8] Rudin, W. (1966). Real and Complex Analysis, McGraw-Hill: New York.
[A9] Titchmarsh, E.C. (1986). The Theory of Funetions, Oxford University Press.
[AlO] Tolstov, G. (1962). Fourier Series, Prentice-Hall (Dover edition, 1976).
[All] Zygmund, A. (1959). Trigonometrie Series, (2nd ed., Cambridge University Press.
Part B
Signal Processing
Introduction
The Fourier transform derives its importance in physics and in electrical engineer-
ing from the fact that many devices mapping an input signal x(t) into an output
signal y(t) have the following property: If the input is a complex sinusoid e2invt,
the output is T(v)e2invt, where T(v) is a complex function characterizing the de-
vice. For example, when x(t) and y(t) are, respectively, the voltage observed at the
input and the steady-state voltage observed at the output of an Re circuit (see Fig.
BO.I), the input-output mapping takes the form of a linear differential equation:
1'~1'
x(t)
C y(t)
l""" ,1,,)
Figure BO.I. The Re circuit
52 Part B Signal Processing
(a) If Yl (t) and Y2(t) are the outputs corresponding to the inputs Xl (t) and X2(t),
then AlYl (t) + A2Y2(t) is the output corresponding to the input AlXl (t) + A2X2(t);
(b) If y(t) is the output corresponding to x(t), then y(t - -r) is the output
corresponding to x(t - -r).
Such physical devices are caIled (homogeneous linear) filters.
A basic example is the convolutional filter, for which the input-output mapping
takes, in the time domain, the form
where h(t) is caIled the impulse response, because it is the response of the filter
when the Dirac pulse 8(t) is applied at the input. Indeed,
If the impulse response is integrable, the output is weIl defined and integrable, as
long as the input is integrable. Then, by the convolution-multiplication rule, the
expression of the input-output mapping in the frequency domain is
y(v) = T(v)x(v),
where T (v) is the frequency response, that is, the Fr of the impulse response:
Observe that if the input is x(t) = e-2i1Cvt, the output is weIl defined and equal to
ISee J.R. Higgins, Five short stories about the cardinal series, Bult. Amer. Math. Soc.,
12, 1985,45-89.
54 Part B Signal Processing
cmumute for base-band signals. This is not a difficult result, but it is of course a
fundamental one because in signal processing, one first sampies and then performs
the filtering operation in the sampled domain, since one of the advantages of digital
processing comes precisely from the difficulty of making analog filters.
One advantage of analog processing is that it is instantaneous. To maintain
competitivity, the signal processing algorithms have to be fast. For instance, the
discrete Fourier transform is implemented by the so-called fast Fourier transform,
an algorithm whose principle we briefty explain in this Part. Subband coding also
has a fast algorithm associated with it. It is a data compression technique. The
signal is not directly quantized, but instead, it is first analyzed by a filter bank, and
the output of each filter bank is quantized separately. This allows one to dispatch
the compression resources unequally, with fewer bits allocated to the subbands
that are less informative (see the discussion in Chapter B4). Subband coding is
the last topic of Part Band introduces the sections on multiresolution analysis in
Part D.
BI
Filtering
DEFINITION Bl.l. The transformation fram the stable signal x(t) to the stable
signal y(t) defined by the convolution
where h(t) is stable, is ca lied a convolutional filter. This filter is called a causal
filter if h(t) = 0 for t < O.
The signal y(t) is the output, whereas the signalx(t) is the inputofthe linear filter
with impulse response h(t). Informally, if x(t) is the Dirac generalized function
8(t) (an impulse at time 0), the output is (see Fig. Bl.1)
8(t) I f\h(t~
> >
0 0 V
impulse impulse response
y(t) = [t
oo
h(t - s)x(s) ds. (2)
L
DEFINITION
(Note that the output is weH defined by the convolution formula, even though in
this particular case the input is not integrable.)
EXERCISE Bl.1. Let y(t) be the output of a stable and causal convolutional filter
with impulse response h(t) [see (2)]. Let
z(t) = 1 t
h(t - s)x(s)ds, t ~ 0,
be the output of the same filter, when the input x(t) is applied only from time t = 0
on. Show that
lim Iz(t) - y(t)1
tt+oo
= O.
The meaning ofproperties (a) and () is the following: (a) XI (t), X2(t) E D('c),
AI, A2 E C ===} AIXI (t) + A2X2(t) E D('c); and () x(t) E D('c), T E lR ===}
x(t - T) E D('c).
The meaning of properties (i) and (ii) is the following: (i) XI (t), X2(t) E D('c),
c c C
AI, A2 E C,XI(t)~ YI(t),X2(t) ~ Y2(t) ===} AIXI(tHA2X2(t) ~ AI(t)YI(tH
A2Y2(t); (ii) x(t) E D('c), T E lR, x(t) ~ y(t) ===} x(t - T) ~ y(t - T).
where G(v) = IT(v)1 is the amplitude gain and (v) = Arg T(v) is the phase.
EXAMPLE Bl.l. Let
or
EXAMPLE B1.2. Let h(t) E L~(lR). We shalt see in Part C that the FT h(v) = T(v)
of h(t) can be defined and that it is in L~(lR). We take D('c) = LUlR) and define
,C by the input-output relationship
where xCv) is the FT ofthe input x(t) E L~(lR). The right-hand side of(7) has a
meaning since T(v) and xCv) being in L~(lR) implies that T(v)x(v) is in Lt(lR)
(see Theorem 20 of the appendix).
EXAMPLE B 1.3. If T ( v) is an arbitrary function, not necessarily in L~ (lR), one can
always define a filter ,C by the input-output relation (7), provided one chooses for
domain D('c) the set of signals x(t) such that the right-hand side has a meaning.
58 B 1. Filtering
-B o +B
Low-pass (B)
-VQ o +VQ
Band-pass (vQ, B)
One possible domain for Hilbert's filter is the set of stable (resp., finite-energy)
signals whose FT has compact support. The amplitude gain of Hilbert's filter is 1
I
(except for v = 0, where the gain is zero), and its phase is
Jr /2 if v > 0,
(v) = 0 if v = 0, (11)
-Jr /2 if v < O.
+i r - ,- - - - - -
,0
- - - - - - - ' , -i
Hilbert filter
There is no function admitting the frequency response (10). There is, in fact,
a generalized function (in the sense of the theory of distributions) with Fr equal
to T(v). However, in signal theory, the Hilbert filter is used only in the theory of
band-pass signals (see Section BI2). For such signals the Hilbert filter coincides
with a bona fide convolutional filter:
EXERCISE Bl.3. Show that the output y(t) ofthe Hilbertfilter, corresponding to
a stable signal x(t) having an FT x(v) that is null outside the frequency band
[- B, + B], can be expressed as
y(t) =- 1R.
x(t - s)
2 sin 2 (n Bs)
ns
ds.
where
y(t) = [
x.(v) e2iJrvt dv
2mv JIR
is in the domain of the preceding filter (the differentiator), and therefore,
x(t)
~----,-----:~ y (t)
The filter 'cl is the forward loop filter, whereas 'cl is the feedback loop filter.
The forward loop processes the total input, which consists ofthe input x(t) plus
the feedback input, that is, the output y(t) processed by the feedback loop filter.
Give the impulse response of the convolutional filter with the above jrequency
response T (v). Interpret the filter as a feedback filter.
DEFINITION Bl.4. The signal set) is called decomposable ifit can be put into the
form
where /LI and /L2 are signed measures of finite total variation. A signed measure
of finite total variation is a mapping /L : B(IR.) ~ IR. of the form
for some A E B(IR.) and all C E B(IR.), where /LI and /L2 are measures on (IR., B(IR.))
of finite total mass.
EXAMPLE Bl.4. Let set) be a periodic signal with period T that is stable over its
period and whose Fourier coefficients satisfy the condition
p,(dv) = LSn8q,(dv).
nEZ
62 BI. Filtering
This measure is signed, has total variation LnEZ ISn I < 00, and since
L
/Lx:
Proof
and hence
y(t) = L
(h(t - s) L e2invs flxCdV )) ds
= L
e 2inv(t-s) (L e- 2inv(t-s)h(t - s) dS) flxCdv),
andhence
(19)
EXAMPLE BI.5. In light of(18), one can interpret Eq. (22) ofTheorem A2.I:
Yn = h
A A(n)A
T x n,
where h(t) is a stable impulse response and x(t) is a locally stable periodic signal
withperiod T: IfLnEZ l.inl < 00, then LnEZ IYnl < 00, since h(v) is bounded.
B 11 Impulse Response and Frequency Response 63
The two signals x(t) and y(t) are therefore decomposable, and (19) can be written
= h(v)ILAdv) = T(v)iLAdv).
We define a linear time-invariant filter CC, D(I: as follows. First, we define the
domain
D(I:) = {x(t);x(t), xCv), and vqx(v) E L~(IR)}. (23)
We first observe that any function x(t) in the domain is differentiable up to order
q and that its jth derivative is
One has to verify that the integral in (25) is weIl defined. Indeed, I/I P (2i n v) I is
bounded because P(z) has no imaginary root. In particular,
IT(v)llx(v)1 .:::: KIQ(2inv)llx(v)1 for some K < 00.
Therefore, IT(v)x(v)1 is integrable for all x(t) E D(I:), and the integral in (25) is
weIl defined for all such x(t).
In fact, T(v)x(v)v k is integrable for all k, 0 .:::: k .:::: p. To check this, observe
that Ivl k /IP(2inv)1 is bounded for all k .:::: p because P(2inv) is bounded away
64 BI. Filtering
from zero (P(z) has no imaginary root) and Ivl k /IP(2iJTv)1 behaves as Ivl k - p at
00. Therefore, the output y(t) is differentiable up to order p, and for j :::s p,
From (24), (26), and (22), it follows that the input x(t) and the output y(t) are
linked by the differential equation
+L = box(t) + L
p q
aoy(t) aly(I)(t) b1x(l)(t),
1=1 1=1
( Q(2iJTv) x(v)e2invt dv ?
Ai P(2iJTv)
The answer is "no, in general" and "yes, asymptotically" if we impose the following
condition:
P(z) is strict1y stable, (28)
that is, the real parts of all the roots of P(z) are strict1y negative. Then, y(t), t ~ to,
the solution of (27) with arbitrary initial conditions y(to) = Yo, Y6j)(to) = Y6 j )
(1 :::s j :::s p - 1), satisfies
ttoo
( l
lim y(t) -
IR
Q(2iJTv)
.
P(2m v)
x(v)e 2lJrvt dv
A )
= O. (29)
Proof" The general solution of (27) is the sum of a particular solution of (27) and
of the general solution of the differential equation without a right-hand side,
Therefore, since
( Q(2iJTv) x(v)e2invt
JIR P(2iJTv)
is a particular solution of (27), we have to show that limttoo z(t) = 0 for the gen-
eral solution of (30). This follows from the theory of linear differential equations
B 11 Impulse Response and Frequency Response 65
because the characteristic polynomial P(z) of (30) has all its roots in the open left
half complex plane (see [B5]).
If P(z) is not strictly stable, there are initial conditions such that
Q(2inv) ~ .
y(t) - ImIR .
P(2mv)
x(v)e217fvt dv
EXAMPLE Bl.6. Consider the LRC circuit (see Fig. BI.5). Its input and output
are related through the differential equation
LCy(T) + RCj(t) + y(t) = x(t),
where jet) and y(t) are thefirst and second derivatives ofy(t). The roots ofthe
characteristic polynomial
-R JR2 - 4L/C
z = ----'--,----'--
2L
and their real parts are always strictly negative. Therefore, the system is strictly
stable, and the permanent regime when the input is x(t) E D(C) is
( 1 2'
y(t) = JlR 1 + RC(2inv) + LC(2inv)2 x(v)e !7fvt dv.
We note that Q(z) == 1 in this example, and therefore,
D(C) = {x(t) : x(t), xCv) E L~(lR)}.
Rational Filters as Convolutional Filters
Going back to the general case described by Eqs. (20)-(25), we pose the problem:
Is the filter of convolutional type? The answer is yes if and only if q < p. Indeed,
consider the factorization of P(z):
P(z) = ap n
k=l
r
(z - Zk)m k ,
R L
1'~1'
x(t) c 1- y(t)
l111111111111I11111
and therefore,
1 -00
0 tj- 1 .
- - - eZkte-217rvt
(j - I)!
dt =-
1
(2inv - Zk)1
..
1 o
00 tj-l
_ _ _ ezkte-217rvt
(j - I)!
.
dt =+
(2inv - Zk)1
..
(Remember that the case Re (Zk) = 0 has been exc1uded.) Defining
'""' ~
k tj - e Zkt
1 j
for t < 0,
~ ~('-1)'
k;Re(zkl>O j=l } .
h(t) = (31)
we have
Fr
r mk
h(t) -+
L L (2inv
k=l j=l
k~ Zk)j .
The input-output mapping is therefore
q-p (
y(t) = LCXkX(k)(t) + J.IR h(t - s)x(s)ds. (32)
k=O lR
In particular,
y(t) = 1
h(t - s)x(s) ds (33)
if and only if q < p. If, moreover, P(z) is strict1y stable (no roots in the c10sed
right half-plane), then, as the expression (31) of the impulse response shows, the
impulse response is causal, and the filter is then called realizable.
EXERCISE BI.5. Give the impulse response of the LRC filter in the case R 2 <
4LjC.
We observe that the input-output relationship (32) is meaningful for all x(t) E
D(,C'), where D(,C') consists of the stable complex signals that are differentiable
B 11 Impulse Response and Frequency Response 67
up to order max(O, q - p). Therefore, one can consider that the filter (C', D(C',
where L' is described by (32), is an extension ofthe original filter (C, D(C. We
may consider that (32) is an extension of the differential equation (27). For some
functions of the extended domain, the input-output relationship is not a differential
equation.
y" - ~y
2
= x" - ~x
3 .
Butterworth Filters
(ir'
(34)
1+
(As n -+ 00, the filter looks more and more like an ideallow-pass filter.) One
seeks T (v) of the form
K
T(v) = P(2inv)'
where P(z) has all its roots strictly to the left of the imaginary axis in order to
guarantee stability and causality.
The roots of the polynomiall + (v / B)2n are
O::;l::;2n+1.
We reorder these roots in such a way that VI, vi, ... , vn , v~, are the 2n roots, where
VI, .. , Vn have strictly positive imaginary parts. We shall allocate VI, . , Vn to
T(v), thus proposing
T(v) = (35)
This is the frequency response of a real filter (i.e., T*(v) = T( -v because any
root among VI, ... , Vn is purely imaginary or it can be associated with another root
symmetrie with respect to the imaginary axis (see Fig. B 1.6).
In the case n = 2, we find
68 BI. Filtering
1/1 1/0
,, /
,, /
/
\ /
, / ........... \ I .........
,,
/ ... /
B .... I \ ....
,,
/
/ ...
/ / \
,
/
/
/
/
/
1/* 1/*
1 o
I/i
n=2 n=3
Figure B1.6. Roots of 1 + (vi B)2n for n = 2 and n = 3
T(v) = -Vi V2 V3
(v - vd(v - V2)(V - V3)
EXERCISE B1.7. Show that the Butterworth filter 0/ order n = 2 can be
implemented by an LRC circuit.
Complex Envelope
DEFINITION B1.1. A band-pass (vo, B) signal, where B < VO, is a stable signal
whose FT is null if Iv I f/. [- B + vo, Vo + B]. A base-band (B) signal is a stable
signal s(t) whose FT is null outside the interval [- B, +B].
It will be assumed, moreover, that set) is real and hence that its Fr is Hermitian
even:
sC-v) = s(v)*.
We are going to show that a real band-pass signal set) has the representation
wherem(t) andn(t) aretwo signals thatare real, andbase-band (B). The base-band
signals met) and n(t) are the quadrature components ofthe band-pass signal set).
Bl2 Band-Pass Signals 69
/\/\ I
0
/\1'\ 8(1/)
0
/\1'\ ~S:(I/)
d\ 0
~u(1/ )
sa(t) = 2 10 00
s(v)e2iJrvt dv, (37)
EXERCISE BI.S. Show that the FT ofthe signal Re {u(t)e2iJrvot} is s(v), and thus
s(t) = Re {u(t)e2iJrvot}. (39)
Let m(t) and n(t) be the real and imaginary parts of u(t):
u(t) = m(t) + in(t). (40)
The quadrature decomposition (36) follows from (39) and (40).
EXERCISE BI.9. Show that
~() u(v) + u(-v)* (41a)
mv=
2 '
A u(v) - u(-v)*
n(v) = 2i (41b)
and that
Band-Pass Filtering
When the band-pass signal (36) is passed through a filter with frequency response
T(v), we may, without loss of generality, consider that T(v) = 0 if lvi fj [vo -
B, Vo + B], since filtering is expressed in the frequency domain by multiplication
of the Frs. Hence it will be assumed that the impulse response h(t) of the filter is
also a band-pass (vo, B) function.
The output signal y(t) has as Fr
y(v) = T(v)s(v), (43)
cos(27rl/ot) 2 cos(27rl/ot)
sin(27rl/ot) 2 sin(27rl/ot)
We shall describe two effects that are specific of frequency transposition. The
first one is the phenomenon of cross-talk in quadrature multiplexed channels.
Cross-Talk
Suppose we use quadrature multiplexing; we thus send two band-pass messages
m(t) and n(t) in the form
Let us assume that the distortion s(t) -+ s'(t) is a linear filtering with frequency
response T(v). Let us note that T(v) is Hermitian symmetric, as it is the Fr of a
real impulse response h(t). We have s'(v) = s(v)T(v), and therefore,
m'(v) = {T(v + va)s(v + va) + T(v - va)s(v - va)}l[-B,+Bl(v),
then
m'(v) = G(v)m(v), n'(v) = G(v)n(v), (47)
72 B 1. Filtering
T(v)
~+13_
~:::: !-~=~
v
-va-B -Va -va+B va-B Va va+B
where
G(v) = T(v + vo). (48)
where K is a complex constant that will be taken equal to unity. This channel trans-
forms the complex sinusoid e2irrvt into the delayed complex sinusoid e i (2rrvt+ (v)),
where (v) is the phase ofthe filter at the frequency v.
Let s(t) be a real signal, band-pass (vo, B), of the form s(t) = m(t) cos 2:rrvot.
Let y(t) be the signal obtained by passing s(t) through the dispersive channel. The
corresponding base-band equivalent filter has the frequency representation
v(v) = T(v + vo)m(v),
where v(v) is the Fr of the complex envelope v(t) of y(t) [see (43)].
Suppose that in the band [vo - B, Vo + B], the dispersion has a first-order
expansion
then (approximately)
where
(vo)
T ---- (50)
P - 2:rrvo
and
(51)
Therefore, we have
Now,
y(t) = Re {v(t)e2invot}.
Hence we have
y(t) = m(t - Tg ) cos 2:rrvo(t - Tp ). (52)
The constants Tp and Tg are the phase delay and group delay, respectively.
B2
Sampling
The first question that arises is: To what extent does the sampie sequence re-
present the original signal? This cannot be true without further assumptions since
obviously an infinite number of signals fit a given sequence of sampies.
The second question is: How do we efficiendy reconstruct the signal from its
sampies?
We begin with a general result that will then be applied to the study of
undersampling and both oversampling.
THEOREM B2.1. Let s(t) be a stable and continuous complex signal with Fourier
transform s( v) E LJ::OR.), and assume in addition that, jJr some 0 < B < 00,
Then
Proof" By Theorem A2.3, the 2B-periodic function <P(v) = LjEZ s(v + j2B)
is locally integrable, and its nth Fourier coefficient is
1 [ 2 n
2B J~ s(v)e- "'2li V dv,
that is, since the Fourier inversion formula for set) holds (s(v) is integrable) and
it holds everywhere (s(t) is continuous), the nth Fourier coefficient of <p(v) is in
fact equal to
2~ ~s(2~)e-2iJrfsv.
nE",
In view of condition (53), the Fourierinversion formula holds a.e. (Theorem A2.2),
that is, <p(v) is almost everywhere equal to its Fourier series. This proves (54).
Since the frequency response T(v) E Lb(~), the impulse response h(t) given
by (55) is bounded and uniformly continuous, and therefore set) is bounded and
continuous (the right-hand side of (56) is a normally convergent series-by (53)-
of bounded and continuous functions). Also, upon substituting (55) in (56), we
obtain
set) = _1 L S ( ~) [ T(v)e2iJrv(t-fsl dv
2B nEZ 2B J~
=[ ! L _1 s( ~)
e-2iJrvfs! T(v)e2iJrvt dv.
J~ nEZ 2B 2B
B21 Reconstruction and Aliasing 77
L
krna Is( ~
2B
)IIT(V)I dv = (L Is( ~ )1) (rk
na 2B
IT(v)1 dV) < 00.)
Therefore,
!
where
THEOREM B2.2. Let s(t) be a stable and continuous signal whose FT s( v) vanishes
outside [- B, + B], and assume condition (53) is satisfied. We can then recover
s(t)from its sampies s(nj2B), n E Z, by theformula
s.(t) = _1
1 2B
"s(~)
L;,2B
8(t - ~).
2B
(59)
nEa-
78 B2. Sampling
!21 s (t)
_~~~f~f~f~f'f'f'f",
Figure B2.l. The Dirae eomb of (59)
1/2B
1,T(t ) --
S -
2B 2B gT (t - -2B'
lL:(n) S n)
-
nEZ
ST(t) = - I1
r 0
T
s(t - u) du.
(Observe that we eannot use Theorem B2.1 as such; why?). Condition (53) implies
that Si,T(t) is integrable and has an Fr given by
The signal ST(t) is obtained by low-pass (B) filtering of the stable signal Si,T(t).
Sinee the impulse response of a low-pass is not integrable, we eannot use the
eurrent version of the eonvolution-multiplieation rule as it iso However, we shall
proeeed formally beeause the result is justified by a more appropriate version of
B21 Reconstruction and Aliasing 79
~ ~
s,(v) = g,(v) (1"
2B L;s (n)
2B e- 2in 2Bn l[-B,+Bl(v) ) = g,(v)s(v).
nE",
~ A
The result then follows by the inversion formula and the convolution-multiplication
formula (the current version, this time).
Aliasing
What happens in the Shannon-Nyquist sampling theorem if one supposes that the
signal is base-band (B), although it is not the case in reality?
Suppose that a stable signal s(t) is sampled at frequency 2B and that the
resulting impulse train is applied to the low-pass (B) with impulse response
h(t) = 2Bsinc (2Bt), to obtain, after division by 2B, the signal
where
If s(t) is integrable, then s(v) is its Fr, by the Fourier inversion theorem. This
Fr is obtained by superposing, in the frequency band [- B, +B], the translates
by multiples of 2B of the initial spectrum s( v). This superposition constitutes the
phenomenon of spectrum folding, and the distortion that it creates is called aliasing
(see Fig. B2.3).
EXERCISE B2.1. Show that if the signal
-w
~
-B +B +W s(v)
o B 2B 3B 4B
-W -B
<--~,
+B +W
~(v)
Figure B2.3. Aliasing
and let s(t) be a stable and continuous base-band (B) signal such that
LkEZ Is(kjvo)1 < 00. Consider the train ofimpulses
Si(t) = ~ L s (~) 8 (t - ~) .
Vo nEZ Vo Vo
Passing this train through a low-pass (vo + B), one obtains a signal a(t). Passing
this train through a low-pass (B), one obtains a signal b(t).
Show that
The following exercise gives aversion of the sampling theorem for band-pass
signals.
EXERCISE B2.3. Let Vo = 2K B for some integer K 2: 1, and let m(t) be a stable
base-band (B) signal. Consider the jrequency-transposed version of this signal,
that is, s(t) = m(t) cos(2Jl'vot). Suppose that
_1
2B
L s (~)
nEZ 2B
8 (t - ~)
2B
Oversampling
We have seen the effeets of inadapted sampling, that is, sampling at a too slow
rate (undersampling) that results in aliasing, or speetrum folding. We now show
that oversampling ean be exploited to obtain faster rates of eonvergenee in the
reeonstruetion formula.
Assurne that the situation of the Shannon-Nyquist theorem prevails; in partie-
ular, we have the reeonstruetion formula (58). The quantity sine (2Bt - n) therein
is of the order of 1/ n in absolute value and of altemating sign. Therefore, the speed
of eonvergenee of the series on the right-hand side is, roughly, eomparable to that
of
L (_1)n S (..!!:.-).
nEZ n 2B
In order to aeeelerate eonvergenee, one ean use oversampling in the following way.
Assurne that supp(s(v)) is eontained in the frequeney interval [- W, + W] for
some 0< W < 00. In formula (57) ofTheorem B2.1, ehoose
B = (1 +a)W (62)
for some a > 0 and take any integrable function T(v) such that
T(v) =1 ifv E [-W, +W]. (63)
The resulting signal is then a perfect replica of s(t) since
which is a replica of the original signal s(t) provided the frequency response of
the filter verifies condition (63).
EXERCISE B2.4. Suppose that
v+B -v+B
T(v) = B _ W 1[-B,-w] + 1[-w,+w] + B _ W 1[+B,+w].
Give the corresponding impulse response, and study the rate 0/ convergence 0/ the
series on the right-hand side 0/(60).
The series in the reconstruetion formula can decay faster by choosing a smoother
frequency response T(v), since increasing the smoothness of a function increases
the decay of its Fourier transform.
82 B2. Sampling
set) = e2iTrAt ,
where A E R This signal is neither stable nor of finite energy, and therefore it
does not fit into the framework ofthe L'- and L 2 -versions of Shannon's sampling
theorem. However, the Shannon-Nyquist formula remains essentially true.
. (rr )
e2iTrAt = L e2iTrnTt sm rr-T (A - nT) , (65)
nEZ - (A - nT)
T
where the series converges uniformly for all t E [-B, +B] for any B < 1/2T.
The result then follows by exchanging the roles of t and A.
Let g(t) be the 1/2T-periodic function equal to e2irrt on (-1/2T, +1/2T].
The series in (65) is the Fourier series of g(t). We must therefore show uniform
pointwise convergence of this Fourier series to the original function.
Without loss of generality, we do this for the Fourier series of the 2rr-periodic
function equal to e iat on (-rr, +rr], where the convergence is uniform on any
interval [-c, +c] C (-rr, +rr). By (39) ofSectionA31, it suffices to show that
lim
ntoo
I-Tr
+Tr
leia(t-s) - eiat 1
sin((n + .! )s)
sin(s /2)
2 ds =0
lim
ntoo
I +Tr 1 sin(as /2)1
-Tr
.
sm(s /2)
sin((n + !)s) ds = O.
set) =L s(nT)
sin (!f
7T:
(t - nT)
. (70)
nEZ T (t - nT)
Proof" Sinee ,Lis finite and the eonvergenee in (64) is uniform in A E [-B, +B],
set) =[ {L e2ilrvnT
sm - (t - nT
. (7T:
7T: T
)} J.L(dv)
[-B,+B] nEZ - (t - nT)
T
- (t - nT))
. (7T:
sm
=L {[ e2ilrvnT J.L(dv) }
nEZ [-B,+B]
7T: T
- (t - nT)
T
.
This c10ses for the moment our study of the Shannon-Nyquist sampling theory.
It will be eompleted in Seetion B32 by the theorem of equivalenee of analog and
digital filtering, and in Seetion C22 by the L 2 -version of the sampling theorem.
84 H2. Sampling
where g(t) is areal or complex function (the "pulse"). Such a "coding" of the
information sequence is referred to as pulse amplitude modulation. Here, T > 0
determines the rate of transmission of information and also the rate at which the
information is extracted at the receiver.
EXERCISE B2.2. Assume that g(t) is a stable signal with FT g(v) and that an is
stable, with transferfunction A(z). Show that s(t) is stable, andgive its FTin terms
of g(v) and A(z).
that is,
akg(O) + L ak-j g(jT).
jEZ
NO
If one only wants to obtain ak from the sampie s(kT), the term
L ak_jg(jT)
jEZ
NO
is parasitic. This term disappears for every sequence ak if and only if
g(jT) =0 for all j i= O. (72)
It turns out that this is equivalent to
(73)
The weak version of the Poisson sum formula of Section A22 is actually all
that is needed to prove the result. Indeed,
THEOREM B2.6. Let g(t) be a continuous and integrable function, and assume
that its FT g( v) is in LUIR). The following two conditions are equivalent:
B23 Intersymbol Interference 85
2W=2B~ ~
T
In this case, there is no other choice for the corresponding pulse than
A I
g(v) = 2B 1[-B,+Bj(v),
that is,
sin(27r Bt)
g(t) = . (75)
27r Bt
One dis advantage of such a pulse is linked to questions of numerical stability.
Indeed, let us assurne that the sampling of s(t) is not carried out at the time kT but
at the time kT + ~, where ~ > O. We obtain
sin(27r B~)" sin(27r B~)
s(kT +~) = ak 27rB~ + f;;;oa k- j 27rB(~ _ jT)'
We see that the error
Is(kT +~) - akl (76)
does not stay bounded for all bounded sequences {ad, because
L I . I
NO I~ - JT
=00. (77)
whose Fr is
In fact,
sine (4BD.)
sekT + D.) = ak 1 _ 16B 2D. 2
'" sin(47fBD.)
+ f#oa n- j 47f B(D. - jT)(1 - 16B2(D. - jT)2)'
and the error (76) is seen to remain bounded whatever the bounded sequenee {ad.
Partial Response Signaling
Another disadvantage of the pulse (75) is that one eannot realize signals with an
Fr that has an "infinite slope" (at - B and + B).
We shall see that, with clever encoding, we ean attain the Nyquist limit (74)
(which says that in order to transmit a "symbol" an every T seeonds without
intersymbol interferenee, a bandwidth of at least 2W = 2B ~ 1fT is needed),
without resorting to an unrealizable pulse (with a very large slope).
For example, in the duobinary encoding teehnique, instead of transmitting (7),
one transmits
S'(t) = L(an + an+l)g(t - nT), (80)
nEZ
that is,
S'(t) = Lang'(t - nT), (81)
nEZ
where
g'(t) = g(t) + g(t + T). (82)
With the pulse (75) ofminimal bandwidth 2B, starting from (80) we obtain
s'(kT) = ak + ak-l = Cb
and from the sequence {Ck} and the initial datum ao we recover the sequenee {ak}'
The interest of this teehnique is that we do not seek to implement Si (t) in the form
(80) using the unrealizable pulse g(t), but rather in the form (81) with a realizable
pulse g'(t). Indeed,
8'(V) = (1 + e-2iJrvT)g(v)
= 2T eos(7fvT)e-2iJrvT 1[-B,+Bj(v).
B23 Intersymbol Interference 87
This pulse has minimal bandwidth 2B, and, furthermore, it is easier to realize, not
having an infinite slope.
The above is a particular case of the technique of partial response signaling. 3
The general principle is the following: We pretend to use the unrealizable pulse
g(t) given by (75), but in (71) we replace the symbol an by an encoding Cn , say, a
linear encoding
(83)
which gives
S'(t) = I>ng(t - nT).
nEZ
where
g'(t) = g(t) + y,g(t + T) + ... + Ykg(t + kT) (84)
is a base-band (B) pulse, in general easily realizable, with FT
8'(v) = T(v)g(v), (85)
where
k
T(v) = 1 + LYje-2inVjT = P(e-2invT) (86)
j=!
and
k
P(z) = 1 + LYjz j . (87)
j=!
(we assume that 1/ P(z) is stable and therefore that the corresponding filter is
causal; these notions are discussed in detail in Section B32).
L qJ(t)o(t) dt = qJ(O),
for all funetions qJ(t). They are aware that there exists no such funetion in the
usual sense with such property, and they take the above formula as a symbolie
way of dealing with a limit situation. In the "prelimit," o(t) is replaeed by a proper
funetion, depending on a parameter, say, n. There are many ehoiees for this proper
funetion on(t), the simplest one being
on(t) = nree~(t).
Then for sufficiently regular funetion qJ(t) (say, eontinuous),
-1
T
l
0
T
8(t)e- 2''" TI
n dt = 1.
Now if we write the eorresponding formal Fourier series
-1
T
Le 2irr!!.v
T
'
neZ
~T(V) = ~ L 8 (v - !!..) .
T neZ T
This overdose of heuristics may weIl be fatal for the more critical mind. However,
in most basic courses in signal analysis, it is administered with the best intentions,
with the exeuse that it saves the student from a painful exposition to distributions
theory. This apology of mathematical euthanasy is founded on wrong premiees.
The first question that one should ask is: Do we need the Dirae comb in signal
analysis? Looking back at the previous chapters, we ean immediately answer NO.1t
is not needed to derive the Shannon-Nyquist theorem, because the Poisson formula
is all that is needed there. Is the Poisson formula harder than distributions theory?
Again, the answer is NO, without surprise, because the distributions theory version
90 B2. Sampling
of the Poisson sum formula is only a small ehapter of distributions theory. (I shall
add that the heuristie derivation of the Poisson sum formula-see the eomment
following the statement of Theorem A2.3 of Chapter l-is mueh more eonvineing
than the usual heuristie derivation of the Fr of the Dirae eomb.)
In fact, the reader may skip this ehapter and proeeed to Chapters 3 and 4 without
damage. On the other hand, the Fourier transform of the Dirae eomb is part of a
well-established tradition in signal analysis that is bound to be etemal due to its
aesthetie appeal. I have therefore devoted the next seetion to the expression of the
classical results of Fourier analysis in the Dirae formalism. It is, however, a purely
symbolie analysis.
The Dirac Generalized Function
The principal formal objeet of the Dirae formalism is the Dirae generalized funetion
8(t), and the first formal rule is the symbolie formula
L
EXAMPLE
x(t) =L
nEZ
r
J[{
e2iJrvtxn8(V - -f) dv
= LXne2iJrft,
nEZ
EXAMPLE B2.3. Let x(t) be as in the previous example. /fit is the input of afilter
with (stable) impulse response h(t) and withfrequency response T(v) = h(v),
symbolic calculations give for the output
y(t) = 1 h(v)x(v)e2iJrvt dv
that is,
T xn,
a result that we already know.
The seeond symbolie fomula, that we now introduee, gives the FT of this
generalized funetion:
(D2)
1 cp(t)I:lT(t) dt = 1 qJ'(v)Lr;:.(v) dv
I:lT(v) = ~L 8 (v - -f)'
nEZ
Multiplication Rule
The third symbolie formula of the Dirae formalism eoneerns the multiplieation of
a Dirae generalized funetion by a funetion in the usual sense:
s(t)8(t - a) == s(a)8(t - a). (D3)
92 B2. Sampling
EXAMPLE B2.5. Sampling and Spectrum Folding. The train oJ sampled pulses
Si(t) = Ls(nT)8(t - nT)
netz:
may, in view oJ(D3), beJormally written
Si(t) = s(t)Llr(t).
Its symbolic FT is thereJore
Si(V) = s(v) * K;(v)
that is,
~
Sie v) "~(
= -1 ~ s v- n) .
-
T netz: T
Ifwe input TSi(t) into a low-pass [-I/T, + l/T]filter, we thereJore obtain at the
output the signal set) with FT
s(v) = LS(V -
netz:
-f) l[-t,+tl(v).
=~
T
L h(v)8(v -
netz:
!!.-)
T
B24 The Dirac Formalism 93
Thuswe have
L
N-i
Sk = sn e - i (21rkn/N).
n=O
The DFr is an approximation of the Fr, the quality of which depends on the
parameters N and .1.. The first question to ask is: How to choose these parameters
to attain a given precision? As we shall see, the answer is given by the Poisson
sum formula. For the time being, we shall give the basic properties of the DFr
without reference to a sampled signal.
Let a = (ao, ... , aN-i) be a finite sequence of complex numbers. For the Nth
root of unity, we adopt the following notation:
N-I N-I
~ ~ m(k-n)
= L...t ak L...t w N .
k=O m=O
But if k =1= n,
N-I WN(k-n) _ 1
~ m(k-n) N =0
L...t w N = 1
m=O w Nk-n -
In the sequel, we use the above periodie extensions. The relation between the
sequences {an} and {Am} will be symbolized by
(93)
We observe that
(94)
B31 The DFf and the FFf Algorithm 97
(95)
(96)
Making n = 0 in (96) and taking (94) into account, we obtain the Plancherel-
Parseval equation for the OFf
N-I 1 N-I
LakbZ = N L AmB~. (97)
k=O m=O
With an == bn we obtain the energy conservation formula
N-I 1 N-I
L
k=O
lall = N L IA
m=O
I
m 2 (98)
where one unit corresponds to one multiplication. The fast Fourier algorithm, 5 , also
called the fast Fourier transform (FFT), considerably reduces the computational
complexity. It is based on the following remark.
Let an ~ Am be a DFT pair (note that we are considering a DFT of order 2N
2N
with 2N terms an and 2N terms Am). Define
(0 ::S n ::s N - 1),
and
(the latter DFTs are of order N). A direct calculation shows that
(0 ::s m ::s 2N - 1). (99)
. that Bm+N
Ob servmg = B m, Cm+N = Cm, and m+N
W 2N = 2N , we can sp lt
-wm 1
Eq. (99) in two parts:
(O::s m ::s N - 1) (100)
and
(0 ::s m ::s N - 1). (101)
In order to calculate B m and C m for 0 < m ::s N - 1, we need 2(N - 1)2
computational units. When (100) is used we need N -1 additional multiplications.
The multiplications in (101) are for free since they were done in (100). In total,
the method requires
2(N - 1)2 +N - 1 = (N - 1)(2N - 3)
units instead of (2N - 1)2 for the direct method. If we have to calculate a DFT of
order N such that
N = 2s , (102)
!
the FFT will take F(N) ::S N 10g2 N computational units. The result is obtained
by induction. Indeed, F(2) = 1, and the considerations above show that
F(2N) = 2F(N) + N - 1 ::S 2F(N) + N.
But if F(N) ::S !N 10g2 N, then 2F(N) + N ::S N(log2 N + 1) = !2N log2 2N.
The gain in computational complexity with respect to the direct method is thus
of the order of
1 10g2 N
---
2 N
5Cooley, J.w., Lewis, P.A.w., and Welch, P.D., The Fast Fourier Transform Algor-
ithm, eonsiderations in the ealeulation of sine, eosine, and Laplaee transforms, 1. Sound
Vibrations, 1970, 12(3),315-337.
B31 The DFT and the FFT Algorithm 99
The above discussion just gives the basic idea of the FFT. For a detailed account
of the algorithmic aspects of the discrete-time Fourier transform, see, for instance,
[B8]. We now turn to the numerical issues behind the DFf.
Numerical Analysis of the DFT
The Poisson sum formula is useful in numerical analysis when approximating a
Fourier integral by a Darboux sum, and this is of course related to the finite Fourier
transform.
Let us recall the Poisson sum formula, assuming that the conditions of validity
are satisfied:
(103)
The expression (103) elucidates the relation between the Fr s(v) of the signal
s(t) and the DFr ofits sampled and truncated version (s( - M ll), ... , s( +M ll,
+M
L s(nll)e-2irrn2.J+,.
n=-M
1 ~( k )
II s (2M + l)ll
would remain. The DFf of (s( -M ll), ... , s( +M ll would then be a sampled
version of the PT, that is,
( II1~s(-Mvd,, II1~)
s(+MvI) ,
where VI = Ij[(2M + l)ll]. But one cannot have a signal s(t) with bounded
support which has FT s(v) also with bounded support. There will thus always be
an error, equal to
100 B3. Digital Signal Processing
This error is the aliasing error. It can be controlled by choosing t:.. small enough for
s(v) to be negligible outside the interval [-B, +B] = [-1/2t:.., 1/2t:..]. But then
M must be adjusted so that s(t) remains zero outside [-M t:.., +M t:..]. Increasing
M increases the computational complexity.
We shall retain the approximate relation linking the effective bandwidth 2B =
1/ t:.., the effective temporal extension T = 2M t:.., and the complexity N = 2M + 1:
2BT :::::: N. (106)
Band T are chosen such that s(t) is negligible outside [-T /2, +T /2] and s(v) is
negligible outside [- B, + B]. Precision requires large T and large B, in order to
capture a large amount of the time-frequency content of the signal. This results in
large complexity (measured by N) of the DFT. This in turn requires sophisticated
algorithms such as the FFT in order to reduce the computationalload.
EXERCISE B3.2. Give the impulse response of the filter with frequency response
exp(cos(w))ei sin(w).
and, in particu1ar,
L IXkllhn-kl < 00 for al1 n E Z.
kE71
The filter is called causal because if the input Xn is zero for n .::: no the output
Yn is zero for n .::: no. The input-output relation (111) takes, for a causal filter, the
form
n
Yn = L Xkhn-k. (114)
k=-oo
is the frequency response of the convolutional filter with stahle impulse response
hn.
102 B3. Digital Signal Processing
Ifwe write i(w) and y(w), respectively, for the Fourier sums ofthe input X n and
the output Yn, the input-output relation (6) reads
y(w) = h(w)i(w). (116)
Indeed,
y(w) = LYne-inW
nEZ
_" 1
- ~
kEZ
Xk e -ikw " h - -i(n-k)w
~ n-k e
nEZ
I
EXAMPLE B3.1 (The pure delay). The input-output relation X n -+ Yn defined by
Yn = Xn-k
is a homogeneous filtering with impulse response
hn = 1o 1 ifn
ifn
= k,
=1= k,
and jrequency response
h(w) = e- ikw .
EXERCISE B3.3 (The smoothing filter). This is the filter defined by the input-
output relation
1 +N
Yn = 2N + 1 L
k=-N
Xn-k
Let s(t) be a stable continuous signal, base-band (B), sampled at the Nyquist
frequency 2B. We obtain the sampled signal
S
. n s(v)dv
( - n) = j+B e 2lJrV2Jj
2B -B
= _1 j+Jt eiJtW2BS(!!.. w) dw
2Jr -Jt Jr
(118)
Yn = 2~ t; h (2~ )x(n 2~ k )
is stable, and its Fourier sum is
Hence we have
Yn = -1 j+Jt 2Bh~ (B
- W) - w ) e lnW
x (B . dw
2Jr -Jt Jr Jr
y(t) = j +B
e2iJtvtx(v)h(v)dv,
-B
and therefore Yn = y(nj2B).
104 B3. Digital Signal Processing
Transfer Functions
To every discrete-time signal Xn is associated itsformal z-transform, which is the
formal series
(120)
The input-output relation (111) reads as a function of the z-transforms of Xn, Yn,
andh n
Y(z) = H(z)X(z). (122)
Note, however, that the z-transform of a signal only takes a meaning as a function
of z E C if one gives the domain of convergence of the series defining it.
We use the unit delay operation z defined symbolically by
iXn = Xn-k
With this notation the relation (6) is written
Yn = Lhk(ixn)
nEZ
In some cases (see the examples below) a function H(z) holomorphic in a ring
{rl < Izl < r2} containing the unit circle {Izl = I} is given. This function defines
a convolutional filter whose impulse response h n is given by the Laurent expansion
(see [B6], Theorem 1.22, p. 53)
H(z) = L hnz n (rl < Izl < r2). (124)
nEZ
i
Recall that the Laurent expansion is explicitly given by the Cauchy formula
where C is a c10sed path without multiple points that lies within the interior of the
ring of convergence, for example the unit circ1e, taken in the anti-c1ockwise sense.
The method of residues can be used to compute the right-hand side of formula
(126). This equality also takes the form
(127)
The integral in (126) can also, have been computed by the method of residues:
If C is a simple c10sed contour on which f is analytic, except for a finite number
of isolated singular points Z I, ... , ZN, then
1. f(z)dz = 2irr
~
tab
k=1
where ak is the residue of f at Z = Zk (see [B 1], Chapter 4, pp. 207 and following).
This is the Cauchy residue theorem. In the case where f has a pole of order m at
Z = Zk, the residue at this point is given by formula
1 dm- I
ak = (m _ I)! dz m- I [f(z)(z - Zk)m-I]lz=Zk
The series filter ,C = 'c2 * 'cl is, by definition, the convolutional filter with
impulse response h n = (h I * h 2)n and transfer function H(z) = H I(Z)H2(Z). It
operates as follows: The input X n is first filtered by ,cl. and the output of'ci is then
filtered by 'c2, to produce the final output Yn.
The parallel filter ,C = 'cl + 'c2 is, by definition, the convolutional filter with
impulse response hi + h~ and frequency response H(z) = HI(z) + H2(Z). It
operates as follows: The input X n is filtered by 'cl, and "in parallel," it is filtered
by 'c2, and the two outputs are added to produce the final output Yn.
The feedback filter ,C = ,CJ/(l - 'cl * 'c2) is, by definition, the convolutional
filter with impulse response frequency response
H(z) = HI(Z)
1 - H I (Z)H2(Z)
This filter will be a convolutional filter if and only if this frequency response is the
FT of a stable impulse response.
106 B3. Digital Signal Processing
Xn Yn
Xn
hn = (~r l{n:::O).
is holomorphic in the ring Cr j,r2 = {rl < Izl < r2} (in the open disk {Izl < r2}
if rl = 0) which contains the unit circle since r2 > 1. We thus have a Laurent
expansion in Crj ,r2
H(z) = I>nzn, (130)
nEZ
which defines a filter with stable impulse response h n and frequency response
Q(e- iW )
H(e iW ) =
P(e-' W )
(see [BI], Section 3.3, or [B6], Theorem 1.22, p. 153).
1 zn-r+l
L n(n -
00
=- - 1) ... (n - r + 2) - - ,
y n=r-l yn
(- IY ~ (j + r - I)! j
1 ~
(z - yY ( 1)
yr r- . j=O '1'
}.yJ Z,
Changing Z into l/ s,
(-s1- r )-r = (-1)r-rrsr ( s - -r1)-r Isl< - .
Irl
1
If Yn is the output of the filter with transfer function (129) corresponding to the
stable input signal X n, we have y(w) = H(e-iw)x(w), that is,
P(e-iw)ji(w) = Q(e-iw)x(w).
Now P(e-iw)y(w) is the Fourier surn of the signal Yn + L:~=l ajYn-j, and
Q(e-iw)x(w) is the Fourier surn of X n + L:i=l bexn-e. Therefore,
p q
P(Z)Yn = Q(z)xn.
The general solution of the recurrence equation (131) is the surn of an arbitrary
solution and of the general solution of the equation without right-hand side
p
Yn + LajYn-j = O.
j=l
This latter equation has for a general solution a weighted surn of terms of the form
r(n)p-n,
where p is aroot of P(z) and r(n) is a polynornial of degree equal to the multiplicity
of this root minus one. If we are given X n , n E Z, and the initial conditions
Yo, Y-l, ... , Y-p+l, the solution of (131) is cornpletely deterrnined.
B33 All-Pass and Spectral Factorization 109
In order that the general solution never blows up (it is said to blow up if
limlnltoo /Yn/ = (0) whatever the stable input X n, n E Z, and for any initial con-
ditions Y-p+l, ... , Y-l, Yo, it is necessary and sufficient that all the roots of P(z)
have modulus strict1y greater than unity.
A particular solution of (131) is
Yn =L hkxn-k .
k::::O
The output Yn is stable when the input X n is stable since the impulse response h n
is itself stable, and therefore Yn does not blow up.
Therefore, we see that in order for the general solution of (131) with stable input
Xn to be stable, it is necessary and sufficient that the polynomial P(z) has all its
roots with modulus strict1y greater than 1.
DEFINITION B3.1. The rational filter Q(z)/ P(z) is said to be stable and causal if
P(z) has all its roots outside the closed unit disk {/z/ ::: I}.
Causality arises from the property that if P(z) has roots with modulus strictly
greater than unity Q(z)/ P(z) = H(z) is analytic inside {/z/ < rz} where rz > 1.
The LaUfent expansion of H(z) is then an expansion as an entire series H(z) =
Lk::::O hkz k, and this means that the filter is causal (hk = 0 when k < 0).
DEFINITION B3.2. The stable rational filter Q(z)/ P(z) is said to be causally
invertible if Q(z) has all its roots outside the closed unit disk {/z/ :::
I}.
In fact, writing the analytic expansion of P(z)/ Q(z) in the neighborhood of zero
as Lk::::O WkZ k, we have
that is,
X n= L WkYn-k . (132)
k::::O
satisfies
<I if Izl < 1,
{
IH(z)1 = 1 iflzl = 1, (134)
Hi() ZZi* - 1
Z=---
Z - Zi
be an arbitrary factor of H(z). If Izl 1, we observe that IHi(z)1 1, using
Fejer's identity
IHiCI* ) 1= IHiI(z) I
we see that the resultjust obtained implies that IHi(z)1 > 1 if Izl > 1.
A filter with frequency response H(e- iw ) is a pure phase filter, or all-pass filter,
by definition.1t is called all-pass because its gain is unity: IH(e-iW)1 = 1.
A(z) = aN n(z -
N
j=l
Zj).
The effect offiltering X n with an all-pass filter (zrz - I)/(z - Zl) is to replace the
factor z - Zl in A(z) by zrz - 1, but it does not change the energy of the signal.
B33 All-Pass and Spectral Factorization 111
B(z)
* -
= A(z) zlz 1
,
Z - Zl
is such that
and therefore,
Thus,
N N
L la n l2 = L Ib l n 2 (136)
n=O n=O
At a time 0 ::::: k ::::: N the two signals (ao, ... , aN) and (b o, ... , b N) have already
dissipated the energies
k k
Ea(k) = L lajl2 and Eb(k) =L Ibj l2 .
j=O j=O
There is an interesting relation between these partial energies. Writing
where
and therefore,
(137)
This shows that if Izii < 1, then (ao, ... , an) is always late with respect to
(b o, ... , bN ) in dissipating its energy.
112 B3. Digital Signal Processing
Fejer's Lemma
EXERCISE B3.2. Let X n be a stable signal with z-transJorm X(z). Define its
autocorrelationfunction Cn by
cn = L Xn+k X;
kEZ
where
R(z) = X(z)X(z)*.
The 2:rr-periodic function R(e- iUJ ) in the above exercise has the following
properties:
i: n
(139)
R(z) = bz ro T1 ( z - --;Zk1 )m
kEK
k
B33 All-Pass and Spectral Factorization 113
Therefore, if Izi = 1,
Two rational fractions that coincide when Izi = 1 coincide for all Z E <C. In
particular, a = b, and whenever we have in R(z) the factor (z - zd with IZk I i= 1,
then we also have the factor (z - 1.). We therefore have
Zk
Using Fejer's identity (135), we therefore find that R(z) can be put under the form
R(z) = ciIG(z)1 2 ,
where
G(z) = n(z - Zj)Sj n(z - ze)".
jE] tEL
The function R(e- iw ) can remain real and nonnegative if and only if c ~ 0 and
d = O. Finally, we can always suppose that IZj I < 1 for all j E J (a root Zj is
paired with another root l/zj).
EXERCISE B3.3. Find a constant c and polynomials P(z) and Q(z) as in Theorem
B3.1, such that
2
5 - 2cos(w) 1 Q(e- i "') 1
----)=c .
3 - cos(w) P(e-'''')
The proof ofTheorem B3.1 can be specialized to obtain that for any polynomial
p(z) such that p(e- i"') ~ 0 for all w E lR, there exists a polynomial A(z) with
A(O) = 1 and no root inside the closed unit disk, and a constant c ~ 0, such that
B(z) = H(z)A(z)
Let x(t) be a stable base-band (B) real signal that we seek to analyze in the
following sense. For fixed N = 2k we wish to obtain for all I ::::: i ::::: 2k the signals
Xi(t) with Fourier transforms
i- I i ]
Bi = [T B '2 kB .
From a theoretical point of view the problem is stated with its solution: For each
i, do no more than filter x(t) with a pass-band filter offrequency response IB/v)!
From the practical point of view of digital processing, in the sampIe domain, an
ideal band-pass filter has an infinite impulse response-actually one with rather
slow decay-and this makes the above pure band-pass filters of poor value from a
numerical point of view.
that is,
2k
x(t) = LXi(t),
i=l
where Xi (t) is obtained from x(t) by approximate band-pass filtering on the band Bi.
This means that leakage between contiguous bands must be mutually compensated.
The above is a summary of the numerical problem associated with subband
decomposition of a signal by a filter bank. The second problem is algorithmic:
How to perform efficiently analysis and synthesis? The standard example of an
efficient algorithm is the FFf, which involves successive splitting, and subband
decomposition is another avatar of this idea: The basic block of the algorithm con-
sists of splitting a given band in two, that is, of solving the subband decomposition
problem for N = 2.
Subband coding is one way of performing data compression. Instead of sam-
pling the original signal and then quantifying the resulting sampies with a view
of digitizing them, one performs the sampling and quantifying operations on each
of the outputs Xi(t). If a subband Bi is deemed unimportant it will be allocated
fewer compression resources, that is, only coarsely quantified. The appraisal of
the importance of each subband is generally based on psychological experiments.
The subjective difference between subbands is very marked in two-dimensional
signal processing, where it has been observed that low-frequency components are
the most important from a subjective point of view.
The Basic Algorithm
Since all signals and filters considered in the present chapter are real, we need only
consider positive frequencies, those in the frequency band [0, B]. Ideal splitting of
the frequency band [0, B] uses two ideal band-pass filters, one for the band [0, B 12]
and the other for the band [BI2, B]. We call To(v) and Tl(V) their frequency
responses. Then, as the Shannon-Nyquist theorem suggests, we sampie each output
at rate B, and reconstruction is perfo,Pled by t~o ideal band-pass filters, [0, B 12]
and [B 12, B], respectively. We call To(V and Tl (v) their fr~quency responses (of
course, ifweuse ideal pass-band filters, To(v) = To(v), and Tl (v) = Tl (v); wekeep
different notations because in the nonideal case, the analysis and reconstruction
filters need not be the same).
Consider Fig. B4.1. In the ideal case (ideal pass-band filters), the signals in the
upper branch at levels (){ (Xl (t and Y (Yl (t are identical and equal to the original
signal x(t) filtered by the band-pass [0, B 12]. This follows from the theory of
sampling of Chapter B2, and the details of the operations in the lower branch are
shown in Figure B4.2. Similarly, in the lower branch of Fig. B4.1, the signals at
levels ()( (X2(t and Y (Y2(t are identical and equal to the original signal x(t)
filtered by the band-pass [B 12, B].
As we explained ~efore, the ideal band-pass filters will be replaced by approx-
imations To( v) and To( v) that have most of their energy inside the band [0, B 12],
B41 Band Splitting with Perfect Reconstruction 117
13 //\f'......//\f'......//\f'.....
~~ ~~ ~
,
~ after sampling at rate B
- 0 +B +2B
and Tl (v) and Tl (v) that have most of theirs inside [B /2, B]. We insist once more
on the fact that we do not require that To(v) = To(v) nor that Tl(v) = Tl(V),
because we need some freedom in the choice of To(v) and Tl(v) to guarantee
perfect reconstruction. Analysis of the original signal yields the decomposition
(Xl (t), xz(t, whereas synthesis reconstructs y(t) = Xl (t) + xz(t). Synthesis is
called perfect when y(t) = x(t).
Xl(t) = _1 L>(~)
. ~
2B JEa.
2B
ho(t - ~),
2B
where ho(t), h l (t), ho(t), h l (t) are the r~spective imp~lse responses correspond-
ing to the frequency responses To(v), To(v), Tl(V), Tl(V). Sirnilarly the signal
Yl (t) at level y in Fig. B4.1 is
YI(n2B) = Lk _1
2B j
LX(~)
2B
ho(~B - ~)
2B
iio(!!.-
2B
- ~).
B
with a similar expression for the output Y2(t) ofthe lower branch ofFig. B4.1.
Down- and Up-sampling
We shall now express the resuIts in terms of the operations of down-sampling and
up-sampling, and then go back to (142).
Let {xnlnez be a sequence of complex numbers and let m E N. Consider the
sequences {Ynlnez and {znlnez defined by
Yo YI Y2
and
Xo 0 Xl 0 X2 0 X3 0 X4 0 Xs
Zo Zl Z2 Z3 Z4 Zs Z6 Z7 Zg Z9 ZIO
The sequence {YnlneZ is said to be obtained from the original sequence {xnlnez
by down-sampling by a factor m. The corresponding operation is denoted as m,/...
Up-sampling by a factor m, denoted as mt, is the operation that transforms {xnlnez
into {znlnez.
In this chapter, we are concemed with the case m = 2. For future use, we shall
express the operation of down-sampling by 2 followed by up-sampling by 2 in
terms of z-transforms (see Figure B4.3).
Denote X(z) and R(z) thez-transforms ofthe sequences {x(n)lnez and {r(n)}neZ,
respectively. The sequence {r(n)}neZ is therefore obtained from {x(n)}neZ by
B41 Band Splitting with Perfect Reconstruction 119
x( n)) @1--_Y--'l(~_)_--1CWI-_~)_r_(n_)
XW RW
Figure B4.3. Down-sarnpling and up-sarnpling
1-------------- r--------------I
1
1
" - - - - - - - - ______ 1
ANALYSIS SYNTHESIS
that is,
Going back to (142) and the similar expression for the lower branch of Fig. B4.1,
we see that the whole system is equivalent in the z-domain to Fig. B4.4.
= L ( - l tho(n) (~)n
nEZ Z
Therefore, if Ho is symmetrie,
6Esteban, D., and Galand, C. (1977), Applications of quadrature mirror filters to split-
band voice-coding schemes, Proc. IEEE Inf. Conf ASSP, Hartford, Connecticut, 191-195.
B42 FIR Subband Filters 121
for some K ::: 1, which means that we accept a delay of K time units to recover
the input, and in this case FIR filters do exist.
EXAMPLE B4.1. Taking the no-aliasing condition (147) into account, the relaxed
condition (149) with K = 1 gives
Ho(z)2 - Ho( - d = 2z. (150)
I ____________________ J
~---------------------~
ANALYSIS SYNTHESIS
stage 1
stage 2
stage J
stage 1
stage 2
stage J
Give the impulse response of each of the fOUf filters in Fig. B4.7.
Another Solution
Another class of solutions7 for the no-aliasing condition (145) is
~l (z)= Z-l Ho( - Z-l),
Ho(z) = Ho(C 1),
{
(152)
H 1(z) = zHo( - z).
The perfect reconstruction condition (146) then becomes
Ho(z)Ho(Z-l) + H o(- z)Ho(- Z-l) = 2. (153)
Since Ho is a real filter,
HO(Z-l) = H(z)* for z = e-ia>,
and (153) takes the form
Ho(e-ia 12
1 + IHo(- e-i a1 2 = 2. (154)
1 ja>
ml (w) = ,J2Hl(e- ).
7Smith, M.J.T., and Barnwell, m T.P. (1986), Exact reconstruction techniques for tree-
structured subband coders, IEEE Transactions ASSP, 34, 434-441.
124 B4. Subband Coding
1 + ei"')N
mo(w) = ( - 2 - L(w),
+
we have
1+ 2N N
= (cos2 (~))
i", 1
Mo(w) = 1 IL(w)1 2 IL(w)1 2.
Therefore, (157) is
(1 - y)N a(y) + yN a(1 - y) = 1.
B42 FIR Subband Filters 125
Therefore, P(y) = a(y) is a solution of (156). We have thuse proven that (156)
admits at least one solution, and by the uniqueness in Bezout's theorem, this
solution is the only one of degree :s N - 1. We have
L
N-l
a(y) = (1 - y)-N[l - yN a(1 - y)] = (f+k-l) l + G(yN).
k=O
= L (f+k-l) l
N-l
a(y)
k=O
This solution is the unique one with degree :s N - 1. Observe that it is nonnegative
for aH y E [0, 1], and therefore a solution to the initial problem. Call it PN and let
P be the general solution. We have
P(y) L (f+k-l) l + yN R
= N-l (1 )
-- , (158)
k=O 2
where R(y) is any odd polynomial such that P(y) so defined remains nonnegative
for all y E [0, 1].
Having obtained Mo(w), it remains to extract its square root mo(w). But this can
be done by spectral factorization, using Fejer's lemma.
We shall elose this chapter on the basic principles of subband coding. Note, how-
ever, that other solutions were proposed, most notably "biorthogonal solutions,"8
which are more versatile and yield finite impulse response subband filters with
better properties (of symmetry, for instance). We refer to the monograph [B12],
where the reader will find a full and detailed treatment of this topic, as weH as
additional references.
References
[BI] Ablowitz, M.J. andJokas, A.S. (1997). Complex Variables, Cambridge University
Press.
[B2] Daubechies, I. (1992). Ten Lectures on Wavelets, CBSM-NSF Regional Conf.
Series in Applied Mathematics, SIAM: Philadelphia, PA.
[B3] Gasquet, C. and Witomski, P. (1991). Analyse de FourieretApplications, Masson:
Paris.
[B4] Haykin, S. (1989). An Introduction to Analog and Digital Communications, Wiley:
New York.
[B5] Hirsch, M.W. and Smale, S. (1974). Differential Equations, Dynamical Systems,
and Linear Algebra, Academic Press: San Diego.
[B6] Kodaira, K. (1984). Introduction to Complex Analysis, Cambridge University
Press.
[B7] Lighthill, MJ. (1980). An Introduction to Fourier Analysis and Generalized
Functions, Cambridge University Press.
[B8] Nussbaumer, H.J. (1981). Fast Fourier Trans/orm and Convolution Algorithms,
Springer-Verlag: New York.
[B9] Orfanidis, S. (1985). Optimal Signal Processing, McMillan: New York.
[BIO] Papoulis, A. (1984). Signal Analysis, McGraw-Hill: New York.
[B 11] Rudin, W. (1966) Real and Complex Analysis, McGraw-Hill: New York.
[BI2] Vetterli, M. and Kovacevic, J. (1995). Wavelets and Sub-Band Coding, Prentice-
Hall: Englewood Cliffs, NJ.
Part C
Fourier Analysis in L2
Introduction
The modem era of Fourier theory started when the tools of functional analysis-
in particular, Lebesgue's integral and Hilbert spaces-became available. Fourier
theory then seemed to have reached the promised land, which is called L 2, the
space of square-integrable complex functions, indeed a Hilbert space.
F. Riesz and E. Fischer were the first to study Fourier series in the L 2 framework. 1
Many ideas of the modem theory of Hilbert spaces were already contained in the
work of these two mathematicians, and they had a c1ear view of the geometrie
aspect ofthe L 2-spaces. They were inspired by aseries of articles by David Hilbert
written after 1904 on the theme of integral equations and in which he gives the
properties of 4(Z). Note, however, that the notion of abstract Hilbert spaces made
its appearance much later than one usually believes, in the years 1927-1930, with
the work of John von Neumann, who was motivated by quantum mechanics. 2
In short, a Hilbert space is a vector space H on the field <C (or lR.), with a
Hermitian (or scalar) product, denoted (., .) or (., .) H, and a special topological
property that we shall now briefly introduce. The Hermitian product induces a
norm, the norm IIxll, or IIxIlH, ofthe vector x E H being
IIxll = (x,x}'1.
I
IF. Riesz, Sur les systemes orthonormaux de fonctions, CRAS Paris, 144, 1907,615-
619; and E. Fischer, Sur la convergence en moyenne, CRAS Paris, 144, 1907, 1022-1024;
Applications d'un theoreme sur la convergence en moyenne, CRAS Paris, 144, 1907, 1148-
1151.
2His theory was published in the reference text Mathematische Grundlagen der Quantum
Mechanik in 1932.
130 Part C Fourier Analysis in L 2
This allows us to define a limit in H: We say that lim n-4oo x n = x iflimn-4oo IIxn -
xII = O. Having this notion of a limit, we have the notion of a Cauchy sequence:
A sequence {xnlnEN in H is called a Cauchy sequence if
Note that for any positive integer k, Ck , considered as a vector space on C with the
usual Hermitian product, is indeed a Hilbert space. B ut there are more sophisticated
Hilbert spaces. For instance, L~(lR), the space of functions f : lR --+ C that are
square-integrable:
In L~(lR), one does not distinguish two functions that are almost everywhere equal.
The Hermitian product is
j +Jr
In L~([-rr, +rr]), one also does not distinguish two functions that are almost
everywhere equal.
A third example is .e~(Z), the set of complex sequences a = {xnlnEz such that
L Ix l n 2 < 00,
nEZ
Introduction 131
The Hilbert space LUlR.) is a paradise ofFourier transforms, since every function
thereof admits a Fourier transform, and moreover the mapping that associates to a
function its Fourier transform is a bijection from L~(lR.) to itself, and the inversion
formula for Fourier transforms, which gives the latter in terms of the former is
This is not apreeise statement. In particular, the integrals appearing in the definition
of the transform and in the inversion formula are in some extended sense, and the
equality in the inversion formula is "almost everywhere." To be exact,
where the limit is in the sense of L~(lR.). A similar interpretation is needed for the
integral defining the Fourier transform.
The beautiful formula of the L 2- theory is the Plancherel-Parseval's formula
in other words,
The above results are stated for the Fourier integral transform, but similar results
hold for the Fourier series of periodic functions: Let f be a 2n -periodic function
square-integrable on [0, 2n]; then it admits the representation
f(t) =L cn(f)e int .
neZ
This is the inversion formula for Fourier series. Similarly to the Fourier transform
in L~ (lR.), this equality is only almost everywhere, and the sum has to be interpreted
in an extended sense:
where the limit is in the sense of L~([ -n, +n]). This result is in fact a particular
case of the Hilbert basis theorem, which gives the orthonormal expansion
x = L(x, en)en
neZ
132 Part C Fourier Analysis in L 2
The Fourier series development is a particular case of the above very general resuIt,
where H == L~([-x, +x]), and
1 .
en(t) == ~elnt.
",2x
The Plancherel-Parseval formula for Fourier series reads
-
1 j+Jr f(t)g(t)* dt = L cn(f)cn(g)*,
2x -Jr nEZ
where f and gare 2x-periodic functions in L~([ -x, +x]. In terms ofHermitian
products,
Cl
Hilbert Spaces
Pre-Hilbert Spaces
In the above definition and in the sequel, 0 represents the zero of IR or C, or the
neutral element of addition in E. The context will remove ambiguity.
From now on, we shall consider complex pre-Hilbert (and later Hilbert) spaces.
The other choice for the scalar field, IR, leads to formally analogous results.
Consequently, two Hermitian products (., h and (., h on E such that 11 . 111 =
11 . 112 are identical.
THEOREM CI.I. For all x, y E E, we have the Schwarz inequality
I(x, y)1 :s IIxll x lIyll, (5)
with equality holding if and only if there exist a, E C such that ax + y = O.
Proof" We do the proof for the real case and leave the complex case to the reader.
We may assume that (x, y) =1= 0; otherwise, the result is trivial. For all E lR,
IIxll 2 + 2(x, y)2 + 2(x, y)211Y1l2 = IIx + (x, Y)YIl2 ~ O.
This second-degree polynomial in E lR therefore cannot have two distinct real
roots, and this implies a nonpositive discriminant:
I(x, y)1 4 - IIxIl 2 1(x, y)1 211y1l2 :s 0,
and thus the inequality (5) holds. Equality in (5) corresponds to a null discriminant,
and this in turn implies a double root of the polynomial. For such a root, IIx +
(x, Y)YIl2 = 0, that is, by Property (d) in Definition CU,
x + (x, y)y = O.
DEFINITION CI.2. Two elements x, y E E are said to be orthogonal if (x, y) = O.
Let Xl, ... , X n E E be pairwise orthogonal. We have Pythagoras' theorem:
(6)
EXERCISE C1.2. Show that if h(t) and x(t) are both in L~(lR), then Y = h * x is
weil defined. Find h E L~(lR) such that IIh 11 = 1 and maximizing y(T)for a given
time T. What is the corresponding maximum?
Cl 2 Continuity Properties
Closed Subspaces
EXERCISE Cl.3. Let (X, X, J-t) be a measure space. For somefixed constant K, let
G = L~(J-t) n {f; sup If(t)1 :::: K, J-t - a.e.}. Is G a Hilbert subspace of L~(J-t)?
Answer the same question, with G = L~(J-t) n {f; sup If(t)1 < 00, J-t - a.e.}.
WhataboutLUJ-t)n{f;suplf(t)l:::: K(f) < 00, J-t-a.e.}?
Paraphrasing the above result, we see that X E span {XI> t E T} if and only if X is
the limit in H of a sequenee of finite linear eombinations of elements of {XI, t E T}.
Continuity of the Hermitian product
THEOREM CI.4. Let H be a Hilbert space over C with the Hermitian product
( " .). The mappingfrom H x H into C defined by (x, y) t-+ (x, y) is bicontinuous.
Proof: We have
I(x + h l , Y + h 2 ) - (x, y)1 = I(x, h 2 ) + (h l , y) + (h l , h 2 )1
By Sehwarz's inequality, l(x,h 2 )1 ::: IIxllllh 2 11, l(hl,y)1 ::: lIyllllhlll, and
l(h l ,h 2 )1::: Ilh 1 1l1lh 2 11. Therefore,
lim
II h l II.IIh2 11-1-0
I(x + h l , Y + h 2 ) - (X, y)1 = O.
In partieular, the norm X t-+ IIx 11 is a eontinuous funetion from H to lR+
EXERCISE C1.5. Let (X, X, fL) be a measure space, where fL is a finite measure.
Let {fn}n ::: 1 be a sequence of LUfL) converging to f. Apply Theorem Cl.4 to
prove that limntoo fL(fn) = fL(f). Give a counterexample ofthis property when the
hypothesis that fL is finite is dropped. (Hint: f = 1[0, I] , fn = (1-1/ n) 1[0, 1] +- .. .)
Show that when fL is finite,
G = L~(fL) n {f; fL(f) = O}
is a Hilbert subspace of L~(fL).
Note that when fL is not finite, G need not be a Hilbert subspaee of L~(fL).
Wavelet multiresolution analysis will provide a speetacular eounterexample.
Isometry Extension Theorem
DEFINITION CI.5. Let Hand K be two Hilbert spaces with Hermitian products
denoted ( " .) Hand ( " .) K, respectively, and let q; : H t-+ K be a linear mapping
such that,for all x, y E H,
(q;(x), q;(Y)K = (x, y)H. (8)
Then q; is called a linear isometry from H into K. If, moreover, q; isfrom H onto
K, then Hand Kare said to be isomorphie.
Note that a linear isometry is neeessarily injeetive, sinee q;(x) = q;(y) implies
q;(x - y) = 0, and therefore,
0= 1Iq;(x - y)IIK = IIx - yllH,
and this implies x = y. In particular, if the linear isometry is onto, it is neeessarily
bijeetive.
138 Cl. Hilbert Spaces
THEOREM CI.5. Let Hand K be two Hilbert spaces with Hermitian products
( ., .) Hand ( ., .) K, respectively. Let V be a vector subspace of H that is dense
in H, and let cp : V f-+ K be a linear isometry from V to K (cp is linear and (8)
holds for all x, y E V). Then there exists a unique linear isometry (l : H f-+ K
such that the restriction of (l to V is cp.
Proof We sha11 first define (l(x) for x EH. Since V is dense in H, there exists
a sequence {xn}ne:! in V converging to x. Since cp is isometric,
However,
Cl 3 Projection Theorem
Let G be a Hilbert subspace of the Hilbert space H. The orthogonal complement
of G in H, denoted G1., is defined by
G1. = {z EH: (Z, x) = 0 for all x E G}. (9)
Clearly, G1. is a vector space over Co Moreover, it is c10sed in H since if {Zn}n~l
is a sequence of elements of G 1. converging to zEH, then, by continuity of the
Hermitian product,
0= lim(Zn,x) = (Z,x) forallx E H.
ntoo
Proof" Let d(x, G) = infzEG d(x, z) and let {Yn}n~l be a sequence in G such
that
d(x, Gi :s d(x, Yni
1
:s d(x, G)2 + -. (*)
n
The parallelogram identity gives, for all m, n ::: I,
IIYn - Ym 11 2 = 2(lIx - Yn 11 2 + IIx - Ym 11 2 ) - 411x - ~(Ym + Yn)1I 2
Since ~(Yn + Ym) E G,
therefore,
Therefore,
lIy - y'1I 2 ~ 0,
which implies that lIy - Y'1I 2 = 0 and therefore, y = y'.
It now remains to show that x - y is orthogonal to G, that is,
(x-y,z)=O forallZEG.
This is trivially true if z = 0, and we shall therefore assume z =j::. O. Because
y + AZ E G for all A E IR,
IIx - (y + Az)1I 2 ::: d(x, Gi,
that is,
Since
wehave
- 2A Re {(x - y, z)} + A211z 11 2 ::: 0 .for allA E IR,
which implies Re {(x - y, z)} = O. The same type of calculation with A E ilR
(pure imaginary) leads to
Im{(x - y, z)} = o.
Therefore,
(x - y, z) = O.
That y is the unique element of G such that y - x E G.L follows from the
observation made just before the statement of Theorem C 1.6.
The element y in Theorem C 1.6 is called the orthogonal projection of x on G
(see Fig. CU) and is denoted PG(x).
Projection Principle
The projection theorem states, in particular, that for any x E G there is a unique
decomposition
x = Y +z, Y E G, Z E G.L, (11)
C13 Projection Theorem 141
xeH
Xj = Pa(Xj) + Wj (i = 1,2),
where Wj e Gl. (i = 1,2). Therefore,
Xl + X2 = Pa(XI) + Pa(X2) + WI + W2
= Pa(xt} + Pa(X2) + W,
where W e Gl.. Now, Xl + X2 admits a unique decomposition of the type
= y+w,
Xl +X2
where W e Gl., y e G: namely, y = Pa(XI + X2). Therefore,
Pa(XI + X2) = Pa(XI) + Pa(X2).
One similarly proves that
Pa(ax) = aPa(x) for all a e G, X e H.
142 Cl. Hi1bert Spaces
Thus PG is linear.
From Pythagoras' theorem applied to x = PG(x) + w,
Hence, PG is continuous.
() The unique decompositions of x on G and G.L and of PG(x) on F and F.L
are
x = PG(X) + w,
The next result says that the projection operator PG is "continuous" with respect
to G.
THEOREM CI.S. (i) Let {Gn }n:::l be a nondecreasing sequence ofHilbert subspaces
ofH. ThentheclosureGofUn:::l Gn isaHilbertsubspaceofH and,for all X E H,
lim PGn(x) = PG(x).
ntoo
Proof: (i) The set Un>l G n is evidently a vector subspace of H (in general,
however, it is not closedflts closure, G, is a Hilbert subspace (Theorem C1.3). To
any Y E Gone can associate a sequence {Yn}n:::h where Yn E G n, and
lim IIY-Ynll=O.
n->oo
and therefore,
(ii) Devise a direct proof in the spirit of (i) or use the fact
NOTATION. IIG I and G 2 are orthogonal Hilbert subspaces olthe Hilbert space
H,
GI ffi G2 := {z = XI + X2 : XI E G, X2 E G2}
y = f(z)* IIzzll 2
x = (x - f(x)
f(z)
z) + f(x)
f(z)
z= u + w,
where u E N and w is colinear to z.
C2
Complete Orthonormal Systems
t
implies that
.,,-11-------;;-11
j=!
en +! =
I
p - "t(fp, ej}ej
Hilbert Basis
The following theorem gives the preliminary results that we shall need for the
proof of the Hilbert basis theorem.
THEOREM C2.1. Let {en}n2:0 be an orthonormal system 01 Hand let G be the
Hilbert subspace 01 H generated by {en}n~!. Then:
(a) For an arbitrary sequence {an }n~O 01complex numbers, the se ries Ln>o anen
is convergent in H if and only if {an}n~! E e~, in which case -
(14)
verges if and only if Ln>O lan 12 < 00. In this case equality (14) follows from the
continuity of the norm, by letting n tend to 00 in the last display.
(b) Accordingto (a) ofTheorem Cl.7, IIxll :::: 11 PGn (x)lI, where G n is theHilbert
subspace spanned by {el, ... , en }. But
n
PGn(x) = ~)x, ei}ei,
i=O
and by Pythagoras' theorem,
n
IlpGn (x)11 2 = L I(x, ei}1 2.
i=O
Therefore,
n
IIxll 2 :::: L I(x, ei}1 2,
i=O
from which Bessel's inequality follows on letting n -+ 00.
(c) From (15) and result (a), it follows that the series Ln>o (x, en}en converges.
For any m :::: 0 and for all N :::: m, -
Leuing N -+ 00, we obtain (17) (using (16) and the continuity of the Hermitian
product).
DEFINITION C2.2. The sequence {Wn}n~O oJvectors oJ His said to be total in H
ifit gene rates H.
In other words, the finite linear combination of the elements of {w n }n~O forms
a dense subset of H.
EXERCISE C2.2. Prove that a sequence {wn}n~O oJthe Hilbert space H is total in
H if and only if there is no element oJ H orthogonal to alt the Wn, n ::: 0, except
0, that is, if and only if
IIxII 2 = L I(x, e }1
n 2; (19)
n~O
(c)Jor alt x E H,
therefore,
IIx - PG (x)1I 2 = 0,
which implies
x = PG(x).
Since this is true for an x E H, we must have G == H, that is, {en}n;::O is total in
H.
A sequence {en }n;::O satisfying one (and then an) of the conditions of Theorem
C2.2 is caned a (denumerable) Hilbert basis of H.
EXERCISE C2.3. Let 1/f be a function in L~ (IR) with the FT VI
= 2~ 1/, where
I = [-2rr, -rr] U [+rr, +2rr]. Show that {1/fj,n}jEZ,nEZ is a Hilbert basis of L~(IR),
where 1/fj,n(x) = 2 j / 2 1/f(2 j / 2 x - n).
EXERCISE C2.4. Let {gj}j;::o be a Hilbert basis of L 2 O, 1]). Show that {gA -
n)ICn,n+1l0}j;::O,nEZ is a Hilbert basis of L 2 (1R). (Here, L 2(I) Denotes the Hilbert
space (equivalence classes) of measurable complex-valued functions defined on
I, with the Hermitian product (f, g) = 1/
f(t)g(t)*dt.)
Biorthonormal Expansions
DEFINITION C2.3. Two sequences {en}n;::O and {dn}n;::O of aHilbert space H form
a biorthonormal system if
(0:) (e n , d k ) = Oforall n =f. k,
() (e n , dn ) = Ifor all n ~ O.
This system is ca lIed complete if, in addition, each ofthe sequences {en}n;::O and
{dn}n;::oforms a total subset of H.
Then we have the biorthonormal expansions
x = L(x, en}dn, x = L(x, dn}en
n;::O n;::O
whenever these series converge. Indeed, with the first series, for example, calling
its sum y, we have for any integer m ~ 0,
An interesting theoretical question is: For what type of Hilbert spaces is there a
denumerable Hilbert basis? Here is a first (theoretical) answer.
Proof Let {fn }n:::O be a sequence defined in Definition C2.4. Construct from
it the orthonormal sequence {en}n:::O by the Gram-Schmidt orthonormalization
procedure. It is a Hilbert basis because (a) of Theorem C2.2 is satisfied. Indeed,
forany zEH,
The following theorem is the fundamental result of the theory of Fourier series of
finite-power periodie signals.
=
{en ( )} def { 1 2i Jr
../Te !l. . }
T , n E
'7J
fLj,
For this, let f(t) E L~([O, T]) and let fN(t) be its projection on the Hilbert sub-
space generated by {enO, -N ::: n ::: N}. The coefficient of en in this projection
is cn(f) = (f, en)L~([O.Tl)' we have
+N {T (T
L Icn(f)1 + Jo
n=-N
2
0
If(t) - fN(t)1 2 dt = Jo
0
If(t)1 2 dt. (21)
C22 Two Important Hilbert Bases 151
qJ(X) = l T
Jex + t)Jet)* dt,
where
defines a T -periodic and continuous function qJ. Its nth Fourier coefficient is
Cn(qJ)
- + t)f(t)*
= T1 10[T( fex - dt ) e- 2.IITTXn dx,
1 {T _ {{T _
= T 10 10 2. n }
f(t)* fex + t)e- IITTX dx dt
=
- 2
T1 10{T f(t)* {ft+T f(s)e-
t
n
IITTS ds e
} 2 n
IITTt dt
Since LnEZ Icn (f)1 2 < 00 and qJ(X) is continuous, it follows from the Fourier
inversion theorem for locally integrable periodic functions that, for all x E IR,
In particular, for x = 0,
11 f - cP 11 2 , and therefore,
(22)
where
bn = I-B
+B
s(v) e2irrvfB dv,
Proof' Let L~(lR; B) be the Hilbert subspaee of LUlR) consisting of the finite-
energy eomplex signals with a Fourier transform having its support eontained in
[ - B, +B]. The sequenee
(24)
where h(t) == 2B sine (2Bt), is an orthonormal basis of L~(lR; B). Indeed, the
functions of this system are in L~(lR; B), and they form an orthonormal system
sinee, by the Planeherel-Parseval formula,
= I-B
+B
e2irrvkz-; dv = 2B x ln=k.
C22 Two Important Hilbert Bases 153
It remains to prove the totality of the orthonormal system (24) (see Theorem C2.2).
We must show that if g(t) E L~(IR.; B) and
+N 1 n
s(t) = lim LCn Mnh(t - - ) , (27)
Ntoo -N v2B 2B
where the limit and the equality in (27) are taken in the L 2-sense (as in (23, and
Cn = I -B
+B 1
./fii
. n
s(v) - - e 2l1tv2ij dv.
An Apparent Paradox
Note that since s( v) is in L 2 and of compact support, it is also in LI, and therefore
the Fourier inversion formula is true and the reconstruction formula takes the
farniliar form
The flaw in the above "proof' is that the Fourier inversion formula holds only
almost everywhere, and maybe not at the sampling times. Therefore, formula (28)
is true only if the Fourier inversion formula can be applied at all the times of the
form nj2B. This is the case if s(t) is continuous, because the inversion formula
then holds everywhere.
We see that the continuity hypothesis always pops up. We cannot expect a much
better version of the sampling theorem in the LI or L 2 framework. Indeed, since
s(v) is integrable, the right-hand side of
s(t) = Ls(v)e2invtdv
THEOREM C3.1. Let set) E LUlR). The mapping Irom lR into L~(lR) defined by
t -+ set + .)
is uniformly continuous.
and called the autocorrelation function of the finite-energy signal s(t). Note that
it is the convolution s(t) * s(t), where s(t) = s( -t)*.
THEOREM C3.2. lfthe complex signal s(t) lies in L~(~) n L~(~), then its FT s(v)
belongs to L~(~) and
Praof: The signal s(t) admits s(v)* as FT, and thus by the convolution-
multiplication rule,
(31)
Since ha(v) = e- 2:rr 2a2x 2 t I when a .,j.. 0, the left-hand side of (32) tends to
IR Is(v)1 2 dv, by dominated convergence.
On the other hand, since the autocorrelation function (s * s)(t) is continuous and
bounded, the quantity
L (s * s)(x)ha(x)dx = L (s * s)(ay)hj(y)dy
tends when a .,j.. 0 toward
by dominated convergence.
C31 Fourier Transform in L 2 157
From the last theorem, we have that the mapping cp: s(t) -+ s(v) from LboR.) n
L~(lR) into L~(lR) thus defined is isometrie and linear. Sinee LI n L 2 is dense
in L 2, this linear isometry ean be uniquely extended into a linear isometry from
LUlR) into itself (Theorem C1.5). We will eontinue both to denote by s(v) the
image of s(t) under this isometry and to eall it the FT of s(t).
lim ( 1s(v)
Ttoo J~
-1 -T
+T
2
s(t)e-2irrVldt 1 dv. (33)
THEOREM C3.3. If SI (t) and S2(t) are finite-energy, complex signals, then
Im ( Sin(JTv))2
~ JTV
dv = 1.
sinee Ih(t)1 and Ix(t)1 2 are in Lb(lR). Therefore, for almost all t,
and y(t) is almost everywhere well defined. We now show that y(t) E LUlR).
158 C3. Fourier Transforrns of Finite-Energy Signals
we have YA(V) = h(V)XA(V), Also, lim YA(t) = y(t) in L~(IR) [use (37)], and thus
limYA(v) = y(v) in L~(IR). Now, since limxA(v) = x(v) in L~(IR) and h(v) is
bounded, lim h(V)XA(V) = h(v)x(v) in L~(IR). Therefore, we have (36).
EXERCISE C3.3. Use the Plancherel-Parseval identity to prove that
{dt 7t:
that is,
s(t) = lim
Atoo
j-A
+A
s(v)e 2i Jl'vt dv, (40)
We shall prepare the way for the proof with the following result.
LEMMA C3.1. Let u(t) and v(t) be two jinite-energy signals. Then
Proof' If (41) is true for u(t), v(t) E L~(~) n L~(~), then it also holds for
u(t), v(t) E L~(~). Indeed, denoting XA(t) = x(t)I[-A,+Aj(t), we have
that is, (UA, VA) = (UA, VA)' Now UA, VA, UA, and VA tend in LU~) to u, V, u,
and V, respeetively, as A t 00, and therefore, by the eontinuity of the Hermitian
produet, (u, v) = (u, v).
=L v(y)u(y)dy.
160 C3. Fourier Transforms ofFinite-Energy Signals
Proof of(39); Let g(t) be a real signal in L~(lR), and define f(t) = (g~)(t), where
g-(t) = g(-t). We have j(v) = g(v)*. Therefore, by (41):
L g(x)j(x)dx = L g(x)f(x)dx
= f (x)(x)* dx.
Therefore,
IIg - f112 = IIgll 2 - 2Re (g, j) + IIfII 2
= IIgll 2 - 211g11 2 + IIfII 2 .
But IIgll 2 = IIgll 2 and IIfll 2 = IIg1l 2 . Therefore, IIg - f112 = 0, thatis,
g(t) = j(t). (42)
In other words, every real (and, therefore, every complex) signal g(t) E L~(lR)
is the Fourier transform of some function of LUIR). Hence, the mapping q; is
onto.
EXERCISE C3.5. Show that if a stable signal is base-band (that is, if its FT has
compact support), then it also has afinite energy.
We dose this seetion by showing how the LI Fourier inversion theorem was
lirnited in scope, since it does not take much for a stable signal not to have an
integrable FT.
EXERCISE C3.6. Show that if a stable signal is discontinuous at a point t = a, its
FT is not integrable.
C4
Fourier Series of Finite Power Periodic
Signals
f:
and the Hilbert space L~([O, T], dt/T) of complex signals x = {x(t)}, t E lR.,
such that Ix(t)1 2 dt < 00, with the Hermitian product
Jorx(t)y(t)* T'
T dt
(x, Y}L~([O,T],~) = (44)
Sn = -1
T
l
0
T
s(t)e- 2"17r'i/
n dt
(45)
defines a linear isometry sO --+ {Sn} lram L~([O, T], dt / T) onto e~, the inverse
01 which is given by
s(t) = L::Sne2i1l"f/ , (46)
nEZ
where the se ries on the right-hand side converges in L~([O, T], dt / T), and the
equality is almost everywhere. This isometry is summarized by the Plancherel-
Parseval identity:
Prao!" The result follows from general results on orthonormal bases of Hilbert
spaces, since the sequence
+N L~([O, Tl)
L Sn e2i1t (n/T)t ------+ set).
-N Ntoo
THEOREM C4.2. The Fourier se ries 01 a T -periodie signal s(t) with finite power
converges almost everywhere to set).
This result shows that the situation for finite-power periodic signals is pleasant,
in contrast with the situation prevailing locally stable periodic signals (remember
Kolmogorov's result, Theorem A3.1). The proof ofCarleson's result is omitted; it
is rather technical. It also shows that the L 2 framework is very adapted to Fourier
series, since everything works "as expected."
Let lb be the space of sequences In, n E Z, such that LnEZ Ilnl < 00 (stable
discrete-time signals).
THEOREM C4.3. lb c l~, that is, a discrete-time stahle signal has finite energy.
Prao!" Let A = {n: IX n I ~ I}. Since LnEZ IX n I < 00, then necessarily
card(A) < 00. On the other hand, if Ixnl ::::: 1, Ix n l2 ::::: Ixnl, whence
The situation for discrete-time signals is therefore in contrast with that of
continuous-time signals, for which there exist stable signals with infinite energy
that and finite-energy signals that are not stable.
C42 Orthonormal Systems of Shifted Functions 163
Let L~(2n) be the Hilbert space of functions j: [-n, +n] ~ C such that
f~: Ij(w)1 2 dw < 00 provided with the Hennitian product
(], g)L~(2n) = _1
2n
l+
-n
n
j(w)g(w)* dw.
THEOREM C4.4. There exists a linear isomorphism between LU2n) and .e~
dejined by
fn = l+ -n
n
j(w)e inw dw ,
2n
j(w) =L
nEZ
fn e- inw . (48)
L fng: = -2n1 l+
nEZ -n
n
-
f(w)g(w)* dw. (49)
L
nEZ
Ig(v + -f )1 2
= T almosteverywhere. (51)
L L
follows from the Plancherel-Parseval formula:
Riesz's Basis
The following notion will play an important role in multiresolution analysis.
DEFINITION C4.1. A system offunctions of LU~)
{w( - nT)}nEz (52)
AL 1ck1 2 ::s
k
(IL
Jffi. kEZ
CkW(t - kT)1
2
dt ::s B L
kEZ
Icd, (53)
(IL
in (53) is equal to
Cke-2itrkTVW(V)12 dv
Jffi. kEZ
Also,
Now, any function c(v) E L~([O, I/T]) has the form LkEZcke-2itrkTv, where
LkEZ ICkl 2 < 00, and (53)is therefore equivalent to
AT 1 1fT
Ic(v)1 2 dv::s 1 IC(V)121~
1fT
Iw(v + f )1 1
2
dv
::s BT Jot fT
Ic(v)1 2dv
C42 Orthononnal Systems of Shifted Functions 165
THEOREM C4.6. Let (w( - n T) }neZ be a Riesz basis of some Hilbert subspace
Vo C L~(R). Define the function g E L~(R) by its Fourier transform
(55)
Therefore, the generators of Vo are in Vo. The converse is true by the same argument
since
g(v) = a(v)-Iw(v),
where a(vr l is also, in view of condition (54), a (ljT)-periodic of finite
power.
EXERCISEC4.1. Let h(t) be the function of FT (ljJ2B)I[-B.+Bj(v) for some
B > O. Show that there exists no orthonormal basis of L~(R; B) of the form
{k g (t - 2~) }nd where
g(t) = LCkh (t - ~),
keZ 2B
166 C4. Fourier Series of Finite-Power Periodic Signals
where {cn} nEZ is in CU7-), unless only one of the Cn is nonzero. Show that if the
Fourier sum c(w) ofthe sequence {cn}nEZ is such that A < !c(w)!2 < B for some
o< A :s B < 00 and for all w, then {k g (t - 2~)}
nEZ
is a Riesz basis of
L~(lR; B).
References
[Cl] Daubechies, I. (1992). Ten Lectures on Wavelets, CBSM-NSF Regional Conf.
Series in Applied Mathematics, SIAM: Philadelphia, PA.
[C2] Gasquet, C. and Witomski, P. (1991). Analyse de Fourier et Applications, Masson:
Paris.
[C3] Halmos, P.R. (1951). Introduction to Hilbert space, Che1sea, New York.
[C4] Rudin, W. (1966). Real and Complex Analysis, McGraw-Hill, New York.
[C5] Young, N.Y. (1988). An Introduction to Hilbert Spaces, Cambridge University
Press.
Part D
Wavelet Analysis
Introduction
f(t)w*(t - b),
where w(t) is the time window function, a function negligible outside a relatively
small interval around zero.
Given a window w(t), the local information at time b is obtained by computing
the Fourier transform of the last display:
where
Wv,b(t) = w(t - b)e2invt.
We see that the time information is collected in a time interval around time b of
width of the order of that of the time window. Now, from the Plancherel-Parseval
identity,
(Note that as a varies, the energy of lai! !(at) remains the same.) The so-called
Heisenberg uncertainty principle makes the above limitation more explicit and
states that
2 Sampling theory and wave propagation, in NATO ASI series, Vol. 1, Survey in Acoustic
Signal! Image Processing and Recognition, C.H. ehen, ed., Springer-Verlag, Berlin, 233-
261,1983.
Introduction 171
In the wavelet transfonn, the role the family of functions Wv,b(t) plays in the
windowed Fourier transfonn is played by a family
(t - b)
1fta,b(t) = lai -1/2 1ft -a- , a, b E IR, a =I- 0,
where 1ft(t) is called the mother wavelet. The wavelet transfonn (WT) of the
function f E L~(IR) is the function
where
:(fra,b(V) = laI 1/ 2 e- 2i 1l'Vb:(fr(av).
From the above expression, it appears that the wavelet transfonn C j (a, b) analyzes
the function f(t) in the time-frequency box
Q= 28-
is independent of the frequency variable a. The area of the time-frequency box
is constant, but its shape varies with the frequency v = m/
a analyzed. For high
frequencies it has a large time dimension, and for small frequencies it has a small
time dimension, which is the desired effect.
The interest of the wavelet transfonn for signal analysts is that they can "read"
it to extract infonnation about the time-frequency structure that is otherwise
blurred in the brute signal by concurrent phenomena and subsidiary effects. For in-
stance, they can detect the appearance time of a phenomenon linked to a particular
frequency (e.g., the time at which a particular atom starts to be excited).
A fonnula, called the identity resolution, allows us to reconstruct, under mild
conditions that we shall make precise in due time, the function from its wavelet
transfonn:
f = L LU, 1J!j,k)1J!j,k.
JEZ kEZ
where
j .
1J!j,n(t) = 2 2 1J!(2J t - n),
and where 1J!(t) is the mother wavelet. The multiresolution analysis of Stephane
Mallat3 is one particular way of obtaining such orthonormal bases and is the main
topic of Part D. Similarly to the continuous wavelet transform, the coefficients of
the multiresolution decomposition can be used to analyze a signal. But multires-
olution analysis is also a tool for data compression. Indeed, with a good design of
the mother wavelet, many wavelet coefficients are small and can be neglected. The
coefficients that are not neglected are quantized more or less coarsely, depending,
for instance, on the frequency index. This is of course reminiscent of subband
coding, and this resemblance is not at all a coincidence. Mallat's algorithm of
analysis-synthesis is of the same form as the subband analysis-synthesis algor-
ithm. Multiresolution analysis can be considered as a systematic way of doing
subband analysis. One of its advantages is to place the latter in a framework where
the mathematical issues ensuring the efficiency of the algorithms are more easily
dealt with.
Perhaps one of the most striking advantages of multiresolution analysis over
classical Fourier analysis is in the way it handles discontinuities. Consider, for
instance, the signal on top of Figs. DO.la and DO.lb,4 which has a spike. In
both figures the middle signal is a Fourier series approximation of the top sig-
nal, whereas the bottom signal is a wavelet approximation of the top signal. In Fig.
DO.I a, only the first 60 coefficients of the expansions (Fourier or wavelet) are used
to produce the approximated signals, and it appears that there is not a dramatic
difference between the two approximations. This is not the case, however, when,
as in Fig. DO.I b, one uses the 60 largest coefficients. The advantage of the wavelet
approximation is then obvious.
3Multiresolution approximation and wavelets, Trans. Amer. Math. Soc., 315, 1989,69-
88.
4Reproduced with the kind permission of Martin Vetterli.
Introduction 173
0.5
of---------'
o 100 200 300 400 500 600 700 800 900 1000
0.5
o
o 100 200 300 400 500 600 700 800 900 1000
0.5
01------"--"--"'\
o 100 200 300 400 500 600 700 800 900 1000
(a)
0.5
01------'
o 100 200 300 400 500 600 700 800 900 1000
0.5
o
o 300 400 500 600 700 800 900 1000
0.5
01------'
o 100 200 300 400 500 600 700 800 900 1000
(b)
mw = -1
Ew IR
1 t Iw(t)1 2 dt,
mfjj = _1_
Ew
r
JIR
v Iw(v)1 2 dv,
Then define
and
The numbers o"w and 0"1ii are the root mean-square (RMS) widths of w and w,
respectively. Note that m w and mlii are not always defined. When they are weIl
defined, o"w and 0"1ii are always defined but may be infinite. Therefore, we shall
always assurne that
and
to guarantee at least the existence of the centers of w(t) and of its Fr.
EXERCISE D1.I. Check if the centers m w and mlii are well defined, and then
compute o"w and 0"1ii and the product O"wO"Iii in thefollowing cases:
w(t) = l[-T,+Tl
w(t) = e-a1tl , a > 0,
w(t) = e- at2 , a > O.
EXERCISEDI.2. Suppose that the centerm w ofthefunction w E L 2 iswell defined.
Show that the quantity
L It - toI 2 Iw(t)1 2 dt
is minimized by to = m w .
Heisenberg's Inequality
THEOREM DI.I. Under the conditions stated above, we have Heisenberg's
inequality
(1)
Proof: We assurne that the window and its Fr are centered at 0, without loss of
generality (see Exercise D1.3). Denoting the L 2 -norm of a function f by IIfII, we
have to show that
~ 1 2
IItwll x IIvwll :::: 4n IIwll ' (2)
By Schwarz's inequality,
IItwll x IIw'lI ::: I(tw, w')1 ::: IRe{(tw, w')}I.
Now,
wherea > O.
Given a window w(t), the local information at time b is obtained by computing
the Fourier transform of (5):
lim ( I/(t)
Atoo JR.
-1 JR.(
Ivl:",A
Wf(V, b)Wv,b(t) dv dbl2 dt = O. (10)
IA(t) = 1 JR.(
Ivl:",A
Wf(v, b)Wv,b(t) dv db.
and, therefore,
L Iw(v)1 2 dv = L Iw(t)1 2 dt = 1,
We show that the function JA is weH defined, that is, (v, b) -+ Wj(v, b)Wv,b(t) is
integrable over [- A, + A] x R In view of (12),
:s ( L l.r-I{j(.)w(. - v)*}1 db
2 )1/2 (
L
Iw(t - b)1 2 db
)1/2
This function h(v) is in LI, being the convolution product oftwo LI-functions. In
particular,
IA(t):s
j +A
-A Ih(v)1 dv < 00,
JA(t) = j+A
-A g(v) dv,
where
Therefore,
fACt) = i: (L A
!(f.1.)lw(f.1. - v)1 2 e2in /Lt df.1. ) dv.
In order to change the order of integration in the above integral, we first verify that
(v, f.1.) -+ 1!(f.1.)llw(f.1. - v)1 2 is integrable over [- A, + Al x lR. But I!I E L 2 and
Iwl 2 E L 1 ; therefore, I!I * Iwl 2 E L 2 , and the integral of an L 2-function over a
i: (L i: (L
finite interval is finite. But this integral is just
A A
1!(f.1.)llw(v - f.1.)1 2d f.1.) dv = 1!(f.1.)llw(f.1. - V)1 2d f.1.) dv
fACt) =L(i: A
!(f.1.)lw(f.1. - v)1 2e2in /Lt dV) df.1.
where 0:::: C{JA(f.1.) :::: 1, in view of assumption (b). In particular, !C{JA E L 2, and
1
fA = F- ~
(jC{JA).
We now show that limA too fA = f in L 2. For this, we write, using the Plancherel-
Parseval identity,
IIf - fAlli2 = IIF- 1! - F- 1!C{JAlli2
= IIF- 1{!(1 - C{JAmi2
Wehave
= j /L-A
-00
Iw(Y)1 2 dy +
1+00
/L+A
Iw(Y)1 2 dy,
182 D 1. The Windowed Fourier Transform
::; j -A 12
Iw(Y)1 2 dy +
1+00 Iw(Y)1 2 dy
-00 AP
= y(A) -+ 0 (A -+ (0).
Also,
-+ 0 (A -+ (0)
lim
Atoo
111 - IAlli2 = O.
From (7) and the Planeherel-Parseval identity, we obtain the two expressions
for the windowed Fourier transform:
(13)
with equality if and only if w(t) == Ae-ct2 , where c = 4n2a~. The last result
shows that.
THEOREM D1.3. The Gabor window is optimal, in the sense that it minimizes the
uncertainty aJijaw.
Dl2 The WFf and Gabor's Inversion Formu1a 183
Although such "atomic" windowed Fr bases do exist, they turn out to be very
bad from the view point of time-frequency resolution, as the following result,
called the Balian-Low theorem,5, shows.
THEOREM 01.4. If{ lj!m,n}mE71,nE71, where lj!m,n is defined by (15)(with gEL 2 (ffi.),
is an orthonormal basis of L 2 (ffi.), then at least one of the following equalities is
true:
where 1/I(t) is called the mother wavelet. The function 1/Ia,b is obtained from the
mother wavelet 1/1 by successively applying a change of time scale (accompanied
by a change of amplitude scale in order to keep the energy constant) and a time
shift (see Fig. D2.1).
DEFINITION D2.1. The wavelet transform ofthe function f E LUIR) is the function
C j : (IR - {On x IR 1-* C defined by
1lV2
11
,-------, 1
11 1
11 1y'2
1
1 .
11
1 )
-1 0 11 1+1
1
3.51 1 17.5
1
t
1 1 1
1 1
-y'2
-1 L - J 1 1
-V2 L.....
'ljJ(t) 'ljJ! ,3.5 (t) 'ljJ2,7.5(t)
where
(19)
Let m", and a", be, respectively, the center and RMS width of the mother wavelet
1/1 , respectively defined by
m", = -1
E",
1lR I 11/I(t)1 2 dl,
aJ = _1_ r(I -
E", llR
m",)211/1(1)1 2 dt,
and similarly define m;;; and a;;;, where VI is the Fourier transform of 1/1. The center
and width of 1/1a,b are, respectively,
b+am"" aa""
whereas the center and width of Vla,b are
1 1
-m;;;, -a;;;.
a a
We shall simplify notations by writing
We see that C j(a, b) is the result of the analysis of the function f in the time-
frequency box (see Fig. D2.2)
[b+am-aa,b+am+aa]x [ma - ~a m
a
+ ~J.
a
D22 The Wavelet Inversion Forrnula 187
i
o b
Figure D2.2. Time-frequency tiling in wavelet analysis
Let us assurne that fii > O. The frequency window is then centered at v = fii / a
and has width 2a / a; therefore,
center frequency fii
Q- --
- bandwidth - 2a
is independent of the frequency variable a. This is called constant-Q jiltering.
Calling v = fii / a the center frequency, we see that the area of the box is constant
a,
and equal to 4a but that its shape changes with the frequency v = 1/a analyzed.
For high frequencies it has a large time dimension, and for small frequencies it has
a small time dimension (see Fig. D2.2). The interest of such features is discussed
in the introduction to this chapter.
We shall see in the next subsection that in order to guarantee perfect reconstruc-
tion of the signal from its wavelet transform, the center of :(f must be zero. Also,
the center of the wavelet itself can be taken equal to zero without loss of generality.
The Fourier transform of a wavelet has bumps at positive and negative frequencies
(see Example D2.3, the Mexican hat). The centers of the bumps then play the role
of the center of the wavelet in the first part of the above discussion (where fii was
assumed to be nonzero ).
THEOREM D2.1. Let 1jJ : ~ f-+ ~ be a mother wavelet such that 1jJ E LI n L 2,
L 11jJ(t)1 2 dt = 1, (20)
188 D2. The Wavelet Transfonn
and
and
Je(t) = -1 ~~ dadb
C j(a, b)1/Ia,b(t) - 2- .
K R.x/lal:::eJ a
Proof' First observe that 1/Ia,b has the same energy as 1/1, equal to unity.
From (18) and (19),
The function inside the curly brackets is in L I because it is the product of two
L 2-functions, and it is in L 2 as it is the product of an L 2-function with a bounded
function (~ is bounded because 1/1 E LI). By the Plancherel-Parseval identity,
= laILI!(v)121~(av)12 dv,
and therefore,
11R. R.
dbda
IC j(a, b)1 2 - 2- =
a
11
R. R.
A ~
IJ(v)1 2 11/I(av)1 2 da
dv-
la I
= lal l / 2 L F;I{!(v)V;*(av)}(b)1/!a,b(t)db
= lal l / 2 L !(v)V;*(av)F;I{1/!a,b(t)}(v)dv,
J(t) = ~ ( l(a) d~
K J11al?} a
1 = 11
R lal?
~ ~
IJ(v)II1/!(av)1 -
lai
1
da dv
= { 1!(v)1 (
JR
(
J1xl?lvl
1V;(x)1 2
Ix I
dx) dv
= 1+ + {
-I
1
J1VI?1
= 11 +h
But
11 = 1+ -I
1
1!(v)1 ( (
J1xl?IVI
1V;(x)1 2
lxi
dX) dv
:s K 1 -I
+1
1!(v)1 dv < 00
190 D2. The Wavelet Transfonn
lz = 1 Ivl:::l
l!ev)1 (1 ~ 2
11/Iex)1 dX) dv
Ixl:::slvl Ix I
But
l11/let)1 2 dt = 1,
and, therefore, using Schwarz' inequality,
lz < ! 1
- e Ivl:::l
l!ev)1 dv
lvi
= { IK - gsev)1 2 1!ev)1 2 dv =
JlR Ivl:o:e-
+
Ivl>e-
1 1
= A + B.
1/ 2 1/ 2
On {lvi :s c- I / 2 },
gsev) = { l1freav)1 2 da = { l1frex)1 2 dx
Jlal:::s lai Jlxl:::SIVI lxi
D22 The Wave1et Inversion Formu1a 191
where
Ke :s K and lim K e
e-+O+
= K.
Therefore,
Also,
11
Proof" We start from (24):
1 Ja
[Ca) -da
a2
=
Ja Ja
~ ~
f(v)Il{!(av)1 2 2 dvda
e "rvt - -
a2
= K L g(v)e2irrvt dv.
This quantity is almost everywhere equal to K f (t) by the Fourier inversion formula
in LI.
Recall that if f is continuous the equality in the Fourier inversion formula holds
for all t and, therefore, (23) is then true for all t E R
Oscillation Condition
Since ~ is continuous (l{! E LI), the assumption (21) implies that ~(o) = 0, that
is, to say,
L l{!(t) dt = o. (26)
192 D2. The Wavelet Transfonn
In most situations it suffices to verify (26), and then (21) follows. For example, if
1/I(t) and t1/l(t) are integrable, then Vi is Cl; therefore, if Vi(o) = 0, the quantity
IVi(v)1 2 /lvl is integrable in a neighborhood of zero and therefore on ~, since at
infinity there is no problem, due to the hypothesis 1/1 E L 2 (which implies that
Vi E L2 ).
EXAMPLE 02.1 (Modet's pseudo-wavelet). Morlet used the mother wavelet
1/I(t) = ye- t2 / 2 cos(5t),
where y is a normalization factor that makes the energy equal to unity. The
theoretical problem here is that
However, the numerical results obtained with this wavelet were satisfactory
because the value of Vi(O) is in fact very smalI.
EXAMPLE 02.2 (Haar wavelet). The Haar wavelet
ifO ~ t < ~,
if ~ ~ t < 1,
otherwise,
satisfies the conditionsfor the reconstructionformula (23) to be valid. Here
~ 1 - cos(7l' v)
1/I(v) = le. -;rrv .
7l'V
~ilr-----~----II
(a)
-10 -8 -6 -4 -2 0 2 4 6 8 10
JJ(/:::
(b)
1
-10 -8 -6 -4 -2 0 2 4 6 8 10
(c)
j:~
-10 -8 -6 -4 -2 0 2 4 6 8 10
0.8
0.6
0.4
0,2
01 - - - - - - - ' L _ _ _ _ _- - - '
~.2 ~ __ ~ __ ~ __ ___
~ __ ~ __ __
~ __ ~ ~ ~~
L 1<7J(w + 2kn)1
kEZ
2 = 1, a.e.
One reason for abandoning the definition in terms of the frequency v is that
the topic of MRA involves a mixture of analog signals and of digital filtering,
and digital signal processing is traditionally-as in the present text---dealt with in
terms of the pulsation w.
(b)for all j E Z, Vj ~ Vj+1 (the Vj 's are said to be nested; see Fig. 4.7),
The function cp is called the scaling function of the MRA. The index j represents
a resolution level: The projection Pj f of a function f E L 2 on Vj is interpreted
as the observation of this function at the resolution level j.
Usually, the projection on Vo is the function itself, in which case the projections
at all levels j 2: 0 are identical. The projection at level 0 is, in applications, the
one offered by the recording device.
Observe that, since the mapping f -+ ,J2 f(2 ) is an isometry from Vo onto VI
and since (cp(. - n) }neZ is an orthonormal basis of Vo, the set (,J2 cp(2 . - n) }neZ
is an orthonormal basis of VI. More generally, {CPj,n}nez is an orthonormal basis
of Vj , where
(27)
](1/)
j(.) E Va
-7f o +7f 1/
}(I/)
~.
o +27f 1/
j(.) E Vi
](1/)
j(.) E V- 1
1/
Shannon multiresolution
Figure D3.2. Nesting in the Shannon MRA
and
$(w) = ll[-rr,+rrj(w)
(see Fig. D3.2). Verify that {Vj }jE71 is a multiresolution analysis 0/ L 2 associated
with the scaling function rp.
rp(t) = ll(o,l](t)
(see Fig. D3.3). Verify that {Vj }j E71 is a multiresolution analysis 0/ L 2 associated
with the scaling function rp.
We shall see later that some regularity of the scaling function is desirable.
198 D3. Wavelet Orthonormal Expansions
,...........J--I
L...-
I ~
j(.) E Vo
16 j(.) E V-I
2 I
Haar multiresolution
DEFINITION D3.2. The function ({J is said to belong to Sr for some rEN if ({J is r
times continuously differentiable with rapidly decreasing derivatives, in the sense
that
The Haar and Shannon scaling functions are not in Sr (for any rEN).
Conditions (a), (b), (e), and (d) of Definition 4.3 are not independent. In fact,
the first part of (d) is always true under conditions (a), (b), (e), whereas the latter
conditions are almost sufficient for the second part of (d). The result below makes
this statement precise.
THEOREM D3.1. Suppose that eonditions (a), (b), and (e) of the definition of an
MRA are satisfied. Then njEZ Vj = 0. Moreover, ifep is eontinuous at the origin,
then
Proo!" The first statement will be proven in the more general Theorem D4.1.
We now prove the second statement. (The proof is technical and can be skipped
in a first reading.) Denote by Ta the translation operator defined by Taf(x) =
f (x - a). We shall first show that the Hilbert space W = clos ( Uj EZ Vj ) is
invariant under translations.
D31 Mother Wave1et 199
h = L c{q;(2 j . -k),
kE'lL
and Tm 2-( h E Vj This means that Tm 2-( f is c:-c1ose to Vj for all j ::: e. From this
and the arbitrariness of c:, we deduce that Tm 2-( f E W.
Let now a E lR be arbitrary. Given c: > 0, there exists 8 such that, for all
c E (a - 8, a + 8),
(use Theorem C3.1 stating that the map a 1--+ Taf is uniformly continuous). In
particular, we can find a dyadic number c for which the above inequality is satisfied.
Since Tc! E W and c: is arbitrary, we deduce that Ta f E W.
We now proceed to the proof of (31). We assurne that ~is continuous at 0 and
that I~(O)I -=1= O. Therefore, ~(w) -=1= 0 on (-c, +c), for some c > O. Consider any
function g orthogonal to W, that is, orthogonal to all f E W. Since W is invariant
under translations, for all x E lR and for all f E W,
for all x ER The function jg* E L~(lR), and therefore, by the Fourier inversion
theorem in L I , j g* = 0 almost everywhere.
In particular, with f(t) = 2 j q;(2 j t) (indeed, such f E Vj C W), we obtain
~(Tjw)g(w)* = 0, a.e.
Assurne now that W = L~(IR). Let 1 be the function with the Fr j = 1[-1.+1].
In particular,
By (29),
Itim 11
J +00
LU, qJj.k)qJj,kiI~ = ~.rr
ke'Z
For large enough j, [-2- j, +2- j] c [-rr, +rr], and therefore the last displayed
expression is 2 j times the sum of the squared absolute values of the Fourier
coefficients of 1[-2-i,+2-ijr. Therefore, by the appropriate Plancherel-Parseval
identity,
Wavelet Expansion
We shall suppose in the sequel that the scaling functions cp satisfy 1qJ'(0) I > 0, and
then take (without further loss of generality)
qJ'(0) = 1. (32)
DEF1NITION D3.3. A wavelet orthonormal basis 01 L 2 = L~(lR) is an orthonormal
basis olthelorm {1/Ij,n}j,nez, where
(33)
The function 1/1 is then called the mother wavelet of the wavelet basis. The
expansion
1 = LL(f, 1/Ij,k)1/Ij,k (34)
jeZ keZ
is called the wavelet expansion of I.
A wavelet orthonormal basis can be obtained from an MRA in the following
way. Let Wj be the orthogonal complement of Vj in Vj+l:
Vj+l = Vj EB Wj. (35)
From property (d) of the definition of MRA,
L2 = EBWj . (36)
jeZ
ift
E
E
(0,
(!' 1],
does it. To see this, it suffices to verify that any f E VI with support (0, 1] is a
linear combination of cp and 1/1 and that cp and 1/1 are orthogonal. Orthogonality
is obvious. Any fE VI such that supp(f) E (0, 1] is oftheform
!],
f(t) = l
a
ift
ift
E
E
(0,
(!' 1].
and we therefore have the decomposition (see Fig. D3.4)
a+ a-
f = - 2 - cp + -2-1/1
The function 1/1 is called the Haar mother wavelet.
'Cl
/
~x
5 1
4" ----, f(t)
l--- L --
1
'ljJ(t)
~
1 +1
4"
I
01 1 1
~
lx
2
21
1
1
1
1 1
-1 --~
The scaling function ({J E Vo and therefore, ({J E VI. Requirements (a) and (e) in
the definition of an MRA imply that {({JI,n}nEZ is a Hilbert basis of VI. and we
therefore have the expansion ({J = LnEZ hn({JI,n, that is,
({J = v'2 L h n({J(2 . -n), (39)
nEZ
where
(40)
In the Fourier domain (39) reads
~
((J(w) = 1M ~hne-
'""' inW~(W)
'i({J "2 '
-v2 nEZ
that is,
(41)
mo(w) = 1M ~hne
'""' -inw . (42)
-v2 nEZ
It is called the low-pass filter MRA, because mo(O) = 1 (recall the running
assumption that $(0) = 1; see (32)). Substituting identity (41) in (38) gives
204 D3. Wavelet Orthonormal Expansions
300~
200
100
O~ __ ~ ____- L____ ~ ____L -_ _ ~ _ _ _ _- L____~____L-~~____~
300~
200
100
O~ __- J_ _ _ _ ~ _ _ _ _- L_ _ _ _~_ _ _ _- L_ _ _ _~_ _ _ _~_ _ _ _~_ _ _ _~_ _~~
300~
200
100
O~ __ ~ ____- L____ ~ ____L -__ ~ ____- L____ ~ ____L -_ _ ~ ____ ~
300~
200
100
O~ __- J_ _ _ _ ~ _ _ _ _- L_ _ _ _- L_ _ _ _- L_ _ _ _~_ _ _ _~_ _ _ _L -_ _ _ _~_ _~~
(a)
,oo~
-10:
50 100 150 200 250 300 350 400 450 500
,oo~
-10:
50 100 150 200 250 300 350 400 450 500
100~
-10:
50 100 150 200 250 300 350 400 450 500
(b)
Therefore,
or, equivalently,
(43)
The filter with frequeney response eiwmo(w + n)* is ealled the high-pass filter
of the MRA. Eqn. (43) shows that the high-pass and the low-pass filters altogether
extraet the whole energy eontained in the band [-n, +n].
We now eharaeterize the spaees V-I and Vo. This will be a preliminary to the
eharaeterization of W_I, the orthogonal eomplement of V-I in Vo. Onee this is
done, we shall obtain the eharaeterization of Wo and then the mother wavelet
itself.
LEMMA D3.1. f E V-I if and only if it has an FT of the form
f(w) = m(2w)mo(w)qJ'(w), (44)
Proof" Indeed, any f E V-I ean be deeomposed along the orthonormal basis
{CP-I,n}nEZ, that is,
1 1
f(t) = M I>kCP( -t - k), (45)
",2 kEZ 2
where {Cn}nEZ E l~. Taking the FT, we obtain
f(w) = hI>ke-i2kWqJ'(2w).
kEZ
This is (44) (using (41) and defining m(w) = v'2 LkEZ Cke-ikw).
f
Conversely, eonsider a funetion defined by (44), where m is a 2n-periodie
funetion in L~([-n, +n]). We show that fis in L~(lR.). First, observe that it is
of the form h(w)qJ'(w), where h is a 2n-periodic funetion in L~([ -n, +n]) (sinee
m E L~([ -n, +n]) and sinee mo is bounded in view ofEq. (43. Now
= j +Jr 2
-Jr Ih(w)1 dw < +00.
206 D3. Wavelet Orthononnal Expansions
f
This proves that E L~(lR). Since f
E L~(lR), it is the Fr of a function I E
LUlR). Tracing back the computations in the first part of the proof, we obtain that
(45) holds true, with {cn}nEZ E e~, which implies that I E V-j.
LEMMA D3.2. I E Vo if and only if it has an FT 01 the lonn
f(w) = d(w)$(w), (46)
lor some 2rr -periodie function d E L~([ -rr, +rr]).
Proof Indeed, let I E Vo. It can be decomposed along the orthonormal basis
{fPo,n }nEZ, that is,
I(t) = L dkfP(t - k), (47)
kEZ
where d E L~([-rr, +rr]). Arguing as in the proof ofLemma D3.l, we can show
f
that any function of the form (46) is the Fr of a function I E Vo.
We are now ready to state and prove the Fourier characterization of Wo, the
Hilbert space of details at level O.
THEOREM D3.2. The function I E Wo if and only if
f(w) = ei~mo(~ + rr)* V(w)$(~) , (50)
for all2Jr-periodie funetion m E L~([ -Jr, +Jr]). This ean also be written
and therefore,
j(w) = d(w)ifI..w),
where
d(w) = eiwmriw + n)*.
Since Imo(w)1 ::: 1, this implies that d(w) E L~([-n, +n]). Therefore, I E Vo
(Lemma D3.2). Also, from the expression of d(w), do(w) = eiwv(w)m'o(w), and
therefore
do..lmo(w),
that is, d(w)mo(w)* + d(w + n)mo(w + n)* = O. By Lemma D3.1 and Eq. (48),
this implies that 1.1 V-I. But also I E Vo. Therefore, I E Wo.
We are now ready for the main result of this subsection, the Fourier characteri-
zation of the mother wavelet in terms of the scaling function and of the high-pass
filter.
r
THEOREM D3.3. The junction 'tjJ is a mother wavelet if and only if
r
the form
s(w)v(w)* = LCke-inw,
nEZ
D32 Mother Wavelet in the Fourier Domain 209
L 11fr(w + 2br)1
kEZ
2 = 1, a.e.,
Conversely, let 1fr be an orthonormal wavelet. Being in Wo, it is of the form (50).
By the usual calculations, we find that
L 11fr(w + 2krr)1
kEZ
2 = Iv(w)1 2 ,
(54)
(a) (b)
- r-
0.5 0.5
0 o
-0.5 -0.5
-1 -1 ~
-3 -2 -1 0 2 3 -3 -2 -1 o 2 3
(c) (d)
1.5 1.5 ,-----------~--~--~--____,
0.5 0.5
OL-------~--~--~--~
o 0.1 0.2 0.3 0.4 0.5 0.1 0.2 0.3 0.4 0.5
Figure D3.6. Haar scaling function and the corresponding wavelet (left: scaling
function; right: wavelet; top: time domain; bottom: frequency domain)
and, therefore,
forn = 0, I,
otherwise,
and, using (57),
fi(w) = I[-n,+nj(w),
and, therefore,
sin(nt)
q;(t) =
nt
D33 Mallat's A1gorithm 211
Therelore, necessarily,
mo(w) = $(2w) on [-n, +n],
that is,
By periodicity,
mo(w) =L $(2w + 2kn).
kEZ
= - e- iw (L kEZ
$(2w + 2kn + 1) $(w)
=- e- iW ($(2w + n) + $(2w - n
Pjl = LCj,n'Pj,n,
nEZ
where
(58)
(a) (b)
0.5 0.5
0 0
-0.5 -0.5
-1 -1
-15 -10 -5 0 5 10 15 -15 -10 -5 0 5 10 15
(c) (d)
1.5 1.5
0.5 0.5
0.1 0.2 0.3 0.4 0.5 0.1 0.2 0.3 0.4 0.5
Figure D3. 7 Shannon sealing funetion and the eorresponding wavelet (left: sealing
funetion; right: wavelet; top: time domain; bottom: frequeney domain)
Dj/ = Ldj,nVrj,n,
nEZ
where
(59)
and wehave
(60)
Denote by Cj and d j the sequenees {Cj,n}nEZ and {dj,n}nEZ, respeetively. The pur-
pose of Mallat's algorithms is to decompose the funetion f, that is, to pass from
CM to dM-I, dM-I, ... , do, Co, and to reconstruct that is to pass, from co, do,d l ,
" " d M to CM.
The sequenee d M - lo d M - lo ... , do, Co is the wavelet encoding ofthe wavelet data
CM' We shall explain the interest of this eneoding onee we have derived Mallat's
algorithm.
D33 Mallat's Algorithm 213
Since the function cp(t/2) is in V-I, and V-I C Vo, and since {cp(. - n)}nEZ is a
Hilbert basis of Vo, we have the decomposition
1 1
2CP(2 t ) = L ancp(t + n),
nEZ
where
a n =- 11m cp(-t)cp(t+n)dt.
1
2 lR 2
Therefore,
L::!. 1 1 .
=2 2 -cp( -(2' t - 2n))
2 2
where
n = ~ [ 1/1'(~t)cp(t+n)dt.
21lR 2
Therefore, it follows by computations sirnilar to those above that
In Theorem D3.3, we now make the particular choice of the mother wavelet
corresponding to v(w) = 1:
that is,
L n einw = L ( _l)n+la;_neinw.
nEZ nEZ
Therefore,
(63)
d j - 1n = hLZCj,Zn-k. (65)
kEZ
These are the basic recursions of the decomposition algorithm (see Fig. D3.8).
The recursions for the reconstruction algorithm (see Fig. D3.9) are obtained from
(60), (61), and (62). This gives
(~ (6~ (6~
dM-I dM-2
/ /
~CM-I~CM
and
Cj,2k-1 - Cj.2k
dj - I k = "fi
We shall now evaluate the algorithmic complexity of the decomposition
algorithm. (Similar results hold for the reconstruction algorithm.)
For this we suppose that the low-pass and high-pass filters, mo and ml, respect-
ively, have finite impulse responses, that is, the sequences {an }nEZ and {n }nEZ have
a finite-number (say, K) of nonzero terms. Suppose that the infinite-dimensional
vector CM has in practice a finite number N of nonzero terms (say, after truncation).
Then there are approximately N /2 terms in CM-I, and therefore, in view of (64),
the passage from CM to CM-I costs approximately K N /2 multiplications; so does
the passage from CM to d M -I. For the decomposition algorithm, we therefore have
approximately
N
(K2 N N) N
+ K 4' + ... + K 2 M + K 2M = K N
multiplications. The complexity of Mallat's algorithm is therefore linear in data
size.
Note that Mallat's algorithm encodes N numbers into N numbers. Thus the
compression gain seems to be null. However, only a few terms in the sequence of
details dj,e, e E Z, j = M - 1, ... ,0, are nonnegligible, provided the MRA is
sufficiently smooth. The smoothness issue is discussed in Chapter D5.
D4
Construction of an MRA
An inspired choice of q; will make the Vj 's nested as required, and this has to
be verified because there is no reason why it should be so when one starts from an
arbitrary orthonormal system {q;(. - n)}nEZ, A necessary and sufficient condition
for this is that
for some sequence {cn}nEZ E e~(Z) or, equivalently, that the dilation equation
holds for some 2JT -periodic function mo in L~( -JT, +JT). We must also verify that
conditions (d) in the definition of, an MRA are satisfied. By Theorem D3.1, it
suffices that fibe continuous at the origin and that Ifi(O) I = 1.
Meyer's Wavelet
Define cp by
. 2JT
lflwl S 3'
fi(w) = . 2JT 4JT (69)
l f - < Iwl <
3 - - 3 '
where v is a smooth function (C k or C OO ) such that
otherwise,
ifx SO,
v(x) = {~ ifx:::l
(70)
and
L lfi(w + 2kJT)1 2 = 1,
kEZ
and, therefore, {cp(. - n)}nEZ is an orthonormal system. We must now verify that
the Vj are nested, and for this it suffices to verify that Vo c Vj or, equivalently,
that cp E Vj. But this is true if and only if there exists a 2JT -periodic function mo
of finite power such that
mo(w) = Lfi(2w+4kJT)
kEZ
=~ ~(w)
cp(w)cp "2 '
D41 MRA from an Orthononnal System 219
since the supports of $(w + 2klr) and of $(w /2) do not overlap if k =1= O. But since
$(~) = 1 if w E supp(~,
we have
$(w)$(~) = $(w),
as desired. We obtain a mother wavelet by formula (53) of with v(w) = 1. This
gives
which gives
0 otherwise.
EXERCISE D4.1. Let P be a probability measure on lR with support in [-8, +8] C
[-1' +1]' and define q;(t) by its Fourier trans/orm
Check that q;(t) is indeed in LUlR) and that the system {q;(. -n ]}nEZ is orthonormal.
I
Check that the dilation equation (55) holds with
~ 4JT
mo(~) = q;(w) iflwl::S 3'
o otherwise.
Show that q;(t) so defined is the scaling function 0/ some multiresolution analysis
and that a mother wavelet is given by its Fourier trans/orm
kEZ
Here we shall also have to verify that
nVj
JEZ
= 0. (79)
Proof The inequalities (77) are equivalent to the existence of A > 0, B < 00
such that
0< AllfII 2 ~ L l(f, wo,k)1 2 ~ BllfII 2 , (80)
kEZ
for all f E Vo, and therefore equivalent to
0< AllfII 2 ~ L l(f, wj,k)1 2 ~ BllfII 2 < 00,
kEZ
D42 MRA from a Riesz Basis 221
By (80),
we have
where
A(c,j) = I)k-Tjc, k+Tjc].
kEZ
222 D4. Construction of an MRA
w(w) = ( sinz'-~(' I ) )2
Wehave
L Iw(w + 2krr)1 2 = ~ + ~ cos(w)
kE'L
= ~ ( 1 + 2 cos2 (~) )
(One way to prove this is to compute the Fourier coefficients of the left-hand side
~ ~ .f3
cp(w) = w(w) ( (W))1/2 . (83)
1 + 2cos 2 "2
If we can compute, at least numerically, the Fourier coefficient Cn in
1 + 2cos (
W))1/2
(
(jJ(2w) w 2 -
= -:::::-- =
mo(w)
q;(w)
cos 2
(2) 1+2cos (w)2 ,
_
2
and this leads to an expression for the mother wavelet's Fourier transform. Again
the (numerical) evaluation of the Fourier coefficients of the function factoring (jJ( w )
yields an evaluation of 'ifr(t) in terms of the translates of q;(2x).
For n = 3, we have
ifO~t~1,
if 1 ~ t ~ 2,
if t < 0,
the rest of the function being obtained by symmetry around 2.
In the general case, Bn(t) is (for n :=:: 1) in cn- I , its support is the interval
[0, n + 1], and
L Bn(x)dx = 1.
Wehave
Bo(w) = e-i~sinc (2:)'
and, therefore, in the Fourier domain, the recurrence defining the B-splines
becomes by the convolution-multiplication rule
Bn(w)
A
= (.W
e-'Tsinc (w))n+!
2Jr (84)
We shall now seek a scaling function for the B-spline of order n. From the
observation
e-iIsinc (2:) = e-i~ cos (4:) x e-i~sinc (4:)'
it follows that
where
.fi n+!
hk = 2n+! (k)' 0~k~ n + 1,
and the scaling equation is
1 n+!
Bn(t) = --;; L(~+!)Bn(2t - k).
2 k=O
and {Bn (. - k)} kEZ constitutes a Riesz basis of the Hilbert subspace that it generates.
In order to compute the scaling function of the MRA, we need the following lemma.
D43 Spline Wavelets 225
Moreover, the coefficients of this polynomial are rational and can be computed
recursively.
Proof Denote the left-hand side of (87) by Fn(w). Inserting (86) in(84) gives
and, therefore,
We introduce the new variable y = cos(w), and define the function Pn by Fn(w) =
Pn(y). Since Fo(w) = 1, we have Po(y) = 1. The recursion in the last display
becomes
2 n+ld (Pn-1(y)
Pn(y) = n(2n + 1) (1 - y) dw (1 _ y)n .
+ (1 - y)2(1 + y)P~'--l)'
Therefore, if Pn - 1 is a polynornial of degree n - 1, then Pn is a polynornial of
degree n. The conclusion follows since Po is indeed a constant.
226 D4. Construction of an MRA
The general method of the previous section gives for scaling function (fi = (fin
Therefore,
(Pn(z+r' )
1
r = LCkl.
!
kEZ
Observe that Ck = Ck. Also, since the function in the left-hand side is analytic,
ICkl ~ plk l ,
for some Ipl < 1. In particular, the scaling function cp(t) has exponential decay.
We now proceed to compute the mother wavelet. We have to compute the impulse
response of the low-pass filter mo(w). We have
A 1
_ cp(2w) _ Bn(2w) ( Pn(cos(w ) 2:
mo(w) - - - - - A - - ,
cp(w) Bn(w) Pn(cos(2w
that is,
. n+! !
l+e- WW ) (Pn(COS(W)2
mo(w) = (
2 Pn(cos(2w
We compute the Fourier expansion
1
( Pn(cos(w ) 2:
= '~qke
" -ikw
,
Pn(cos(2w kEZ
where
i
1
rr ( Pn(cos(w )
qk = q-k = - 2:
cos(kw)dw.
n 0 Pn(cos(2w
Therefore,
where
br = hL(-ll-ILk-1qr-k.
kEZ
D5
Smooth Multiresolution Analysis
<C 2 L 1 1 1
- O,k+2 nEIU'" (1 + Ix - nD 2 (l + It - nl)2 (1 + Ix - tDk '
(90)
In particular, for each t E IR, the function qt : IR --+ C defined by qt(x) = q(x, t)
is in L 2 , and the development of any function f E Vo along the orthonormal basis
{qJn}nEZ = {qJ(' - n)}nEZ
DEFINITION D5.3. Let E be some set, and let H be a Hilbert space offunctions f :
E --+ CwiththeHermitianproduct(, }.IfthereexistsafunctionK: ExE --+ C
such thatfor each x E E, thefunction K(x, .) E H, and fex) = (K(x, .), j}, H
is called an autoreproducing Hilbert space with reproducing kernel K.
EXERCISE D5.1. Let E be some set, and let H be a Hilbert space of functions
f : E --+ C with the Hermitian product (., .). Suppose that for each x E E,
the mapping f --+ fex) from H to C is continuous. Show that H is then an
autoreproducing Hilbert space.
Equation (91) therefore teUs us that Vo is an autoreproducing Hilbert space with
reproducing kernel q(x, t). Similarly, for all mEZ, Vm is an autoreproducing
Hilbert space with reproducing kernel qm (x, t), where
(92)
We know (Theorem D3.1) that 1$(0)1 = 1, and we can assume without loss of
generality that $(0) = 1. Therefore, in view of property (38),
$(2klr) = l[k=o}. (93)
D52 Pointwise Convergence Theorem 231
It follows from this and the weak Poisson formula (Theorem A2.3) that
L qJ(x - n) = 1. (94)
where qm is the autoreproducing kernel of Vm, defined by (92). This kernel re-
presentation allows us to obtain pointwise convergence results, in the manner
of Dirichlet's pointwise convergence analysis of Fourier series. We need some
preliminary results on the kernel.
DEFINITION D5.4. Let {m}mEZ be a sequence of functions m : lR x lR ~ C.
It is called a quasi-positive delta sequence if it satisfies the three following three
conditions:
(a) There exists afinite nonnegative constant K such that
lim
mtoo
ft-c
t+c
m(x, t)dx = 1, (97)
=L Iq(x, 2m t)1 dx
<C2
-
r 1
J[{ (1 + lxi?
dx=K<oo
'
where we have used inequality (90).
We now prove property (b) ofDefinition D504: Let c > 0 be finite. We have
f t+c
qm(X, t)dx =
12m (t+c)
q(x,2mt)dx
t-c ~Q-~
Then
tim fm{t)
m--+oo
= f(t)
uniformlyon any compact subinterval [a, ] E (a, b).
Proof" For y > 0, write
= f(t) f HY
t-y
8m(x, t) dx
= A+B +(C).
Let [a, ] E (a, b), t E [a, ]. Let c be as in (b) of D5.4. Choose y such that
o< y < c, + y < b, a - y > a. For any 0 < 8 < 1, further restrict y so
that If(x) - f(t)1 < 8 whenever t E [a, ] and Ix - tl < y (in which case both t
and x are in a compact subinterval contained in (a, b), and we can then invoke the
uniform continuity of f in this closed interval). We then have
uniformly with t E ~, and the same for the limit of J/~:). Therefore, If(t) - AI ~
8.
234 D5. Smooth Multiresolution Analysis
it is highly desirable from a numerical point ofview that the coefficients dj,n decay
rapidlyas IJ I, Iml -+ 00, thus ensuring fast convergence ofthe wavelet expansion.
This is not the case, however, even for smooth functions (say, I E C oo n L 2 ) if
no further conditions are imposed on the mother wavelet 1{1. To understand this
and see what type of conditions 1{1 should satisfy, let us examine the asymptotic
behavior of
d(N) = JN l l(x)1{I(Nx)* dx
= JNa(N).
A Taylor expansion of I (assumed to be C OO ) with Lagrange residue gives
where
RK(X) = l
o
x (x
-
K!
t)K
lK+l)(t) dt.
We assurne that the scaling function has a Fourier transform at 0 equal to 1, which
implies that the mother wavelet has a null Fourier trans form at 0 or, equivalently,
that it integrates to O. Therefore,
J,Lk = L x k1/l(x)*dx,
Then
(99)
Let N be a dyadic integer (that is, N = 2- j oko) such that 1/I(N) #- 0 (the
existence of N follows by the density of dyadic integers in IR and by the fact that
1/1 is continuous and not identically zero).
Let j > 1 be sufficiently large for 2j N to be an integer. By orthogonality
o= 2j L 1/1 (x )1/t(2 j x - 2 j N) dx
1/I(N) L 1/I(y)dy = O.
Therefore, (99) is proved for k = O.
Suppose that (99) is true for k = 1, ... , n - 1, where n .::: r. We have the Taylor
expansion
=L
n (x N)k (x N)n
1/t(x) 1/I(k)(N) - + rn(x) - ,
k=O k! n!
where rn(x) is uniformly bounded. Choose N such that 1/I(n)(N) #- O. Substituting
LI~
in (*), we obtain
1jI(n)(N) l yn1jl(y)dy = 0 .
l1jl(X)dx = 0,
which is always satisfied and implies that the projection Im = Pvml satisfies
l Im(x)dx = 0, (100)
a surprising fact at first glance, since the function I that is analyzed is in general
not such that
l Im(x)dx = O. (101)
In Mallat's algorithm one first computes the projection Po I that is the approxim-
ation of I at the resolution level 0, and then the coarser resolution approximations
Pj I, j :s -1. As we have just seen, the moment conditions on 1jI are useful for
the first part of the algorithm. For the second part fast decay of the coefficients
hn = h l <p(x)<p(2x - n)* dx
is needed for rapid numerical convergence. An ideal situation is when only a finite
number of h n are nonzero, which is guaranteed ifthe scaling function has compact
support. Note that if this is the case, then the compactness of the scaling function
carries over to the mother wavelet, and this is why one usually talks of compact
wavelets rather than compact scaling functions.
Let us mention at this point that if we start from a Riesz basis of Vo, as in the
method explained in Section D42, the compactness of w (there defined) does not
imply compactness of the scaling function. In the face of this negative statement
one needs to be reassured about the transmission of exponential decay from w to
<p. As a matter of fact the situation is not too bad, and a result in this direction is,
for example, Proposition 5.4.1 in [D3]. We end this section by showing how the
decay ofthe scaling function is transmitted to the coefficients h n Localization of
References 237
the scaling function can be taken in many related senses. We mentioned previously
one of them, namely ({J E Sr. Another definition of localization could be
1 Ixl~A
1({J(x)12 dx::: : : ' (103)
Ihni::: ../211
Ixl~A
({J(x)({J(2x -n)dxl
: : . /2 (1
Ixl~A
1({J(2x - n)1 2dx)1/2 +../2 (1 Ixl::::A
1({J(x)1 2 dx)1/2 ,
and therefore with a proper choice of A, saya = n, in view of the tail majorization
(103), we obtain
Dm
Ihn I ::: for all m E N, (104)
(1 + n)m
where the Dm are finite. Thus, the Fourier coefficients of mo are rapidly decaying
and this implies that mo E C oo .
The topic of compact wavelets is an important one, but it is rather technical. The
interested reader is refered to [D3] for the detailed theory.
References
[01] Blatter, C. (1998). Wavelets, a Primer, A. K. Peters: Natick, MA.
[02] Chui, C.K. (1992). An lntroduction to Wavelets, Academic Press: New York.
[03] Oaubechies, I. (1992). Ten Lectures on Wavelets, CBSM-NSF Regional Conf
Series in Applied Mathematics, SIAM: Philadelphia, PA.
[04] Hemandez, E. and Weiss, G. (1996). A First Course on Wavelets, CRC Press: Boca
Raton,FL.
[05] Kahane, J.-P. and Lemarie-Rieusset, P.G. (1998). Stfries de Fourier et Ondelettes,
Cassini: Paris.
[06] Mallat, S. (1998). A Wavelet Tour 0/ Signal Processing, Wiley: New York.
[07] Meyer, Y. (1993). Wavelets Algorithms and Applications, SIAM: Philadelphia, PA.
[08] Vetterli, M. and Kovacevic, J. (1995). Wavelets and Sub-Band Coding, Prentice-
Hall: Englewood Cliffs, NJ.
[09] Walter, G. (1994). Wavelets and Other Orthogonal Systems with Applications,
CRC Press: Boca Raton, Fl.
Appendix
The Lebesgue Integral
Introduction
Integration is almost as old as mathematics. It is at least as old as Greek
mathematics,8 since Eudoxus and Archimedes used the exhaustion method to
compute the volume ofvarious solids, in particular, the pyrarnid and the cone. 9
The Riemann integral has a few weak points, the two main ones being that
8Sir Thomas Heath, A History of Greek Mathematics; Vol. I: From Thales to Euclid,
Clarendon Press, Oxford, 1921; Dover edition, 1981.
9Exhaustion is the procedure by which we compute, for instance, the volume of the cone
of height h and circular base of radius R, as the limit of a heap of circular tiIes:
ntoo k=! n
h
-R -.
n
242 Appendix
(1) The dass of nonnegative functions which are Riemann-integrable is not large
enough. Indeed, some functions have an "obvious" integral, and Riemann's integ-
ration theory denies it, while Lebesgue's theory recognizes it (see Example 9), and
its stability properties under the limit operation are too weak.
(2) The Riemann integral is defined with respect to the Lebesgue measure (the
"volume" in ffi.n), whereas the Lebesgue integral can be defined with respect to a
general abstract measure, a probability for instance.
The last advantage is an excellent argument to convince a student to invest a
little time in the study of the Lebesgue integral, because the return is considerable.
Indeed, the Lebesgue integral ofthe function f with respect to the measure p, (see
the meaning in the first chapter), modestly denoted by
Ix fex) p,(dx),
contains a variety of mathematical objects, for instance, the usual Lebesgue integral
on the line,
L f(x)dx,
can also be viewed (with profit) as a Lebesgue integral with respect to the counting
measure on Z. The Stieltjes-Lebesgue integral
L f(x)dF(x)
e (the "length") on JEt It says: There exists a unique measure e on JR: that gives to
the intervals [a, b] the measure b - a. Of course, in order to understand what all the
fuss is about, and what kind of mathematical subtleties hide behind such a harmless
statement, we shall have to be more precise about the meaning of "measure". But
when this is done, one is very much ready to approve the statement although the
proof is not immediate. Of course, in this appendix, the proofs of such "obvious"
results are not given. In fact, the goals of this appendix are to provide a tool and to
give a few tips as to how to use it safely. The reader who has no previous knowledge
ofintegration theory will therefore be very much in the situation of the new recipient
of a driving license who takes the road in spite of her inexperience. Experience is
best acquired on the road, and the main text contains many opportunities for the
student to check her reflexes and to apply the roles that are briefly explained in the
appendix. The student wishing to purehase good insurance is directed to the main
companies, a few of which are listed in the bibliography of this appendix.
Farewell and bon voyage!
THEOREM 2. Let (X, X) and (E, E) be two measurable spaces, where E = a(C)
for some collection C of subsets of E. Then f : (X, X) f-+ (E, E) if and only if
f-I(C) E X for all C E C.
One immediate application of this result is
EXAMPLE 1. Let (X, X) be a measurable space and let n ~ I be an integer. Then
f = (fl, ... , fn) : (X, X) f-+ (IRn, Bn) if and only iffor all a = (al, ... , an) E
IQt, {f :s a} E X. (Here
{f:S a}:= {x E IR n : !;(x):s a;joralll:s i:s n}.)
The proof follows immediately from 2 and the definition of Bn.
EXAMPLE 2. Let X and E be two topological spaces with respective Borel sigma-
fieids B(X) and B(E). Any continuous function f : X f-+ E is measurable with
respect to B(X) and B(E).
The above result is a direct consequence of Theorem 2 and of the abstract
definition of continuity: f : X f-+ Eis said to be continuous if f-I( 0) is an open
set of X whenever 0 is an open set of E.
Measurability is stable by composition:
THEOREM 3. Let (X, X), (Y, y), and (E, E) be three measurable spaces, and let
q; : (X, X) f-+ (Y, y), g : (Y, y) f-+ (E, E). Then f := g 0 q; : (X, X) f-+ (E, E).
Measures
DEFINITION 5. Let (X, X) be a measurable space and let /-L : X f-+ [0,00] be
a setfunction such thatforany denumerablefamily {An}n:::1 ofmutually disjoint
sets in X,
(105)
246 Appendix
The setfunction JL is called a measure on (X, X), and (X, X, JL) is ca lied a measure
space.
Property (105) is the sigma-additivity property.
The following three properties are easy to check:
JL(0) = 0;
(A ~ B and A, B E X) ===} (JL(A):::: JL(B));
(An E X for all nE N) ===} (JL(U~oAn):::: L~o JL(An)).
EXAMPLE 3. Let a E X. The measure Ba defined by Ba(C) = lc(a) is the Dirac
measure ata EX. The setfunction JL : X 1-+ [0,00] defined by
L aj lai(C),
00
JL(C) =
j=O
I
EXAMPLE
f-LO, tD ift 2: 0,
FJL(t) = (107)
-f-Lt,OD ift < 0.
This is a c.d.f (use next lemma), and, moreover,
From the last formula, we deduce that any point set {a}, a E lR has null Lebesgue
measure, and therefore, any countable subset of lR (Ql, for instance), has null
Lebesgue measure.
The following lemma features the sequential continuity properties of measures.
LEMMA 1. Let (X, X, f-L) be a measure space. Let {An}n~1 be a non-decreasing
(that is, An ~ A n+ 1for all n 2: 1) sequence of X. Then
f-L (0 n=1
An) = lim t
ntoo
f-L(A n). (l08)
Let {Rn }n~ 1 be a nonincreasing (that is, Rn+1 ~ Rn for all n 2: 1) sequence of X
such that f-L(R no ) < 00 for some no E N+. Then
f-L (n n=1
Rn) = lim +f-L(Rn).
n,!,oo
(l09)
Proof: We shall prove (108). This equality follows direct1y from sigma-additivity
since
n-I
f-L(A n) = f-L(Ad +L f-L(Ai+1 - Ai)
i=1
and
v (0 Rn) = v(0) = 0.
The following result is the key to the construction of the Lebesgue integral:
THEOREM 9. Let I : (X, X) 1-+ (iR, B) be a nonnegative Borelfunction. There ex-
ists a nondecreasing sequence {/n}n2:l olnonnegative elementary Borelfunctions
that converges pointwise to I.
Proof" Take
n2- n -l
In(x) = L kr n 1Ak ,n(x),
k=O
where
Ak,n = {x EX: kr n < I(x) ~ (k + l)r n }.
For any nonnegative elementary Borel function I : (X, X) 1-+ (lR., B) of the
form
k
I(x) = Lai lA/X),
i=l
where ai E lR.+, Ai E X for all i E {l, ... , k}, one defines the integral of I with
respect to J1" denoted
by
(110)
of the approximating sequence. Note that the quantity (111) is nonnegative and
can be infinite. It can be shown that if I ::: g, where I, g : (X, X) 1--+ (~, 13) are
nonnegative, then
In particular, if
wehave
and therefore,
(112)
Integrable Functions
(114)
is meaningful and defines the left-hand side. Moreover, the integral of I with
respect to JL defined in this way is finite.
(115)
This leads to one of the forms "finite minus finite," "finite minus infinite," and
"infinite minus finite." The case JL(f+) = JL(f-) = + 00 is rigorously excluded
from the definition, because it leads to the indeterminate form "infinite minus
infinite."
EXAMPLE 7. Any funetion I : Z t-+ lR. is measurable with respeet to P(Z) and .
With the measure IJ. defined in Example 4, and with I ~ Olor instanee, we have
=L
00
1J.(f) anl(n).
n=!
EXAMPLE 8. Let Ca be the Dirae measure at point a E X. Then any I : (X, X) t-+
(lR., ) is ca-integrable, and
ca(f) = I(a).
Elementary Properties of the Integral
First, recall that for all A EX,
(117)
The extension to complex Borel functions of the properties (a), (b), (d), and (f) in
Theorem lO is immediate.
Riemann and Lebesgue
The following result tells us that all the time spent learning about the Riemann
integral has not been in vain.
THEOREM 11. Let I : (lR., ) t-+ (R ) be Riemann-integrable. Then it is
Lebesgue-integrable with respeet to l, and the Lebesgue integral is equal to the
Riemann integral.
EXAMPLE 9. The eonverse is not true: The funetion I defined by I (x) = 1 if x E Ql
and I (x) = 0 if x f/. Ql is a Borel funetion, and it is Lebesgue-integrable with its
integral equal to zero beeause {I =I- O} = Ql, has l-measure zero. However, I is
not Riemann-integrable.
252 Appendix
lim
Atoo
j-A
+A
--2
l+x
x
dx = O.
d/L = lim t [
Jx[ f ntoo Jx fn d/L.
The next result is a useful technical tool called Fatou 's lemma.
THEOREM 13. Let fn : (X, X) ~ (i, B), n ~ 1, be such that fn ~ 0 /L-a.e. for
all n ~ 1. Then
The domina ted convergence theorem is also called the Lebesgue theorem:
THEOREM 14. Let fn : (X, .1:') ~ (i, B), n ~ 1, be such that, for some function
f: (X,.1:') ~ (i, B) and some /L-integrablefunction g : (X,.1:') ~ (i, B),
(i) lim fn
ntoo
= f, /L-a.e.,
The Lebesgue Integral 253
The results in Theorems 12 and 14 ensure that under certain circumstances limit
and integration may be interchanged (that is, JL(lim In) = lim JL(fn. The classical
counterexample that shows this is not always true is the following:
EXAMPLE 11. For (X, X, JL) = (IR., {3, l), define
1
In(x) =0 if Ixl>-
n
1
n
:s x :s 0,
1
n
One has
lim In(x) =0 if x =j:. 0,
ntoo
that is, limntoo In = 0 JL-a.e. Therelore, JL(limntoo In) = O. However, JL(fn) =1
lor alt n ~ 1.
THEOREM 15. Assume thatlor JL-almost alt x thefunction t "rl I(t, x) is continu-
ousatto E (a, b)andthatthereexistsaJL-integrablefunctiong : (X, X) f-+ (i:, )
such that I/(t, x)1 :s Ig(x)1 JL-a.e. lor alt t in a neighborhood V 01 to. Then
I : V f-+ IR. is welt defined and is continuous at to. Furthermore, assume that
(a) t -+ I(t, x) is continuously differentiable on V lor JL-almost alt x,
() For some JL-integrable function h : (X, X) f-+ (i:, )
Proof" Let (tn}n~l be a sequence in V \ {to} such that limntoo tn = to, and define
fn(x) = f(t n, x), f(x) = f(to, x). Then, by dominated convergence,
lim I(tn) = I(to).
ntoo
1
Also,
l(tn) - I(to) f(tn, x) - f(to, x) dx
---'--'-~ = JL( ),
tn - to x tn - to
and for some 9 E (0, 1), possibly depending upon n,
{ af
= Jx at
(to, x) JL(dx).
The Fubini Theorem
Product Measures
Let (Xl, Xl, JLl) and (X2, X2, JL2) be two measure spaces where JLl and JL2 are
sigma-finite measures.
Define the product set X = Xl X X2 and the product sigma-field X = Xl X X 2,
where by definition the latter is the smallest sigma-field on X containing all sets
ofthe form Al x A 2, where Al E Xl. A 2 E X2.
THEoREM 16. There exists unique measure JL on (Xl X X2, Xl x X2) such that
JL(A l x A2) = JLl(A l )JL2(A2) (121)
(B) Fubini. If I is Ji,-integrable, then, lor Ji,1-almost all XI, the function X2 --+
I(XI, X2) is Ji,2-integrable and XI --+ JX2 I(XI, X2) Ji,2(dx2) is Ji,2-integrable, and
(123) is true.
In this text we shall refer to the global result as the Fubini-Tonelli theorem.
Part (A) says that one can integrate a nonnegative Borel function in any order
of its variables. Part (B) says that the same is true of an arbitrary Borel function if
that function is Ji,-integrable. In general, in order to apply Part (B), one must use
Part (A) with I = 1I1 to ascertain whether or not J 1I1 dJi, < 00.
Wehave
= h(X2) ~ 0,
However;
1 1
h(X2) dx2 1= 1
00
(- h(xI)) dxl,
since h ~ f-a.e. on (0, 00). We therelore see that successive integrations yield
different results according to the order in which they are perjormed. As a matter
ollact, I(XI, X2) is not integrable on (0,1) x (1,00).
256 Appendix
Integration by Parts
THEOREM 18. Let 111 and 112 be two sigma-finite measures on (R 8). For any
interval (a, b) c lR,
111 a, b ])112a, b]) = { 111 a, t]) 112(dt) + ( 112a, t)) 111 (dt). (124)
~a.bl ~a.bl
Observe that in the first integral we have (a, t] (c1osed on the right), whereas in
the second integral we have (a, t) (open on the right).
Proof' The proof consists of computing the l1-measure of the square (a, b] x
(a, b] in two ways. The first one is obvious and gives the left-hand side of (124).
The second one consists of observing that 11a, b] x (a, b]) = I1(Dd + I1(D2),
where D I = {(x, y);a < y ::::: b, a < X::::: y} and D 2 = (a, b] x (a, b] \ D I . Then
I1(D 1) and I1(D2) are computed using Tonelli's theorem. For instance,
and
L g(x) F/i-(dx)
stands for IIR g(x) l1(dx). When this integral is used, it is usually called the
Lebesgue-Stieltjes integral of g with respect to F w With this notation, (124)
becomes
The Spaces LP
For a given P :::: 1, L~(I1) is, roughly speaking (see the details below), the collection
of complex-valued Borel functions J defined on X such that Ix
IfIP dl1 < 00.
We shall see that it is a complete normed vector space over C, that is, a Banach
space. Of special interest to Fourier analysis is the case P = 2, since L~(I1) has
additional structure that makes of it a Hilbert space.
Let (X, X, 11) be a measure space and let J, g be two complex-valued Borel
functions defined on X. The relation R defined by
(fRg) ~ (f = g l1-a.e.)
The Lebesgue Integral 257
(126)
THEOREM 20. Let p and q be positive real numbers different Iram 1 such that
1 1
-p + -q = 1
(p and q are then said to be conjugate), and let I, g : (X, X) t-+ (i,8) be
nonnegative. Then, we have Hlder's inequality
(127)
Proof' Let
We may assurne that 0 < A < 00, 0 < B < 00, because otherwise Hlder's
inequality is trivially satisfied.
Define F = J/ A, G = g/ A, so that
Ix FP dJl Ix = Gq dJl = 1.
The inequality
1
F(x)G(x)::::: - F(x)P + -1 G(x)q (*)
p q
is trivially satisfied if x is such that F(x) = 0 or G(x) = O. If F(x) > 0 and
G(x) > 0, define
1x
(FG)dJl::::: -
1
p
1
+ - = 1,
q
and this is just (127).
THEOREM 21. Let P ~ 1 and let J, g : (X, X) 1-+ (i:, B) be nonnegative and such
that
and
One may assume that the right-hand side of (128) is finite and that the left-
hand side is positive (otherwise the inequality is trivial). Therefore, !x(f +
g)P dlL E (0,00). We may therefore divide both sides of the last display by
t
[Jx (f + g)P dlL q. Observing that 1 - 1/q = 1/P yields the desired inequality
(128).
For the last assertion of the theorem, take p = q = 2.
THEOREM 22. Let p ::: 1. The mapping vp : L~(IL) 1-+ [0,00) defined by
Riesz-Fischer Theorem
We shall denote vp(f) by IIfll p. Thus L~(IL) is a normed vector space over C,
with the norm 11 . 11 p and the induced distance
dp(f, g) = 111 - gllp.
THEOREM 23. Let p ::: 1. The distance d p makes 01 L~ a complete normed space.
In other words, L~(IL) is a Banach space for the norm 11 . I p
ProoJ- To show completeness one must prove that for any Cauchy sequence
(fn}n~1 of L~(IL) there exists I E L~(IL) such that limntoo dp(fn, f) = O.
Since {fn}n~1 is a Cauchy sequence (that is, limm,ntoo dp(fn, Im) = 0), one can
select a subsequence (fn,}i~1 such that
dp(fni+l - In) S Ti. (*)
Let
k
gk =L I/ni+l - Inil,
i=1
260 Appendix
00
g= L
;=1
I/ni+' - Ini I
L (jni+' (X) -
00
converges absolutely for /L-almost all x. Call this limit I(x) (set I(x) = 0 when
this limit does not exist). Since
k-I
we see that
One must show that I is the limit in Lt(/L) of Unkk~:I' Let e > O. There exists an
integer n = N(e) such that II/n - Im I p :s e whenever m, n 2: N. For all m > N,
by Fatou's lemma we have
lim
m--+oo
111 - Imllp = O.
Terminology. For p 2: 1, Lt(/L) is a Banach space (a complete normed vector
space) over Co This phrase will implicitly assume that the norm is defined as in
(129). When /L is the Lebesgue measure on jRn, we write Lt(jRn) instead of Lt(/L)
(with a slight symbolic inconsistency).
In the proof of Theorem 23 we obtained the following result.
THEOREM 24. Let Unk:1 be a convergent sequence in Lt(/L), where p 2: 1, and
let I be the limit.
A subsequence {/ni k::1 can then be chosen such that
lim f,n
;too '
=I /L- a.e. (130)
Note that the statement in (130) is about functions and not about equivalence
c1asses. The functions thereof are any members of the corresponding equivalence
c1ass. In particular, since when a given sequence of functions converges /L-a.e. to
two functions, these two functions are necessarily equal/L-a.e.
References 261
Approximation Theorems
We now quote the approximation results used in the main text.
THEOREM 27. Let f E L~(lR.), P ~ 1. There exists a sequence {fnln:o:! of con-
tinuous functions fn : IR t-+ C with compact support that converges to f in
L~(IR).
(To have compact support means, for a continuous function, to be null outside
some c10sed bounded interval.)
THEOREM 28. Let f E L~(IR), P ~ 1. There exists a sequence {fnln:o:! offunctions
fn : IR t-+ C which are finite linear combinations ofindicatorfunctions ofintervals,
that converges to f in L~(IR).
THEOREM 29. Let fE Lt([-n, +nD be a 2n-periodicfunction (that is, f(t) =
f(t + 2n)forall t E IR, and J~: If(t)1 d t < (0). There exists a sequence {fnln:o:!
of functions fn : IR t-+ C with continuous derivatives that converges to f in
Lt([ -n, +n D.
References
[Dl] de Barra, G. (1981). Measure Theory and Integration, EIlis Horwood: Chichester.
[D2] Halmos, P.R. (1950). Measure Theory, Van Nostrand: New York.
[D3] Royden, H.L. (1988). Real Analysis, 3rd ed., MacMillan: London.
[D4] Rudin, W. (1966). Real and ComplexAnalysis, McGraw-Hill: New York.
[D5] Taylor, A.E. (1965). General Theory of Functions and Integration, Blaisdell,
Waltham, MA, Dover edition, 1985.
Glossary of Symbols
1 : Ca, b]
f:
L~([a, b]), the set (equivalence c1asses) ofmeasurable functions ~ C
such that I/(t)IP dt < 00.
l~(Z), the set of complex sequences {Xn}nEZ such that LnEZ IXn 12 < 00.
cn(f) = (I/2n) J~rr f(t)e- int dt, the nth Fourier coefficient of f.
Sn(f) = L~: ck(f)e+ ikt , the Fourier series.