Correction of the exercises from the book

A Wavelet Tour of Signal Processing

Gabriel Peyr
Ceremade
Universit Paris-Dauphine
Stphane Mallat
CMAP
cole Polytechnique
November 16, 2009
Abstract
These corrections refer to the 3
rd
edition of the book A Wavelet Tour of Signal Processing
The Sparse Way by Stphane Mallat, published in December 2008 by Elsevier. If you
nd mistakes or imprecisions in these corrections, please send an email to Gabriel Peyr
(gabriel.peyre@ceremade.dauphine.fr). More information about the book, including how
to order it, numerical simulations, and much more, can be nd online at wavelet-tour.com.
1 Chapter 2
Exercise 2.2. If
_
[h[ = +, for all A > 0 there exists B > 0 such that
_
B
B
[h[ > A. Taking
f(x) = 1
[A,A]
sign(h(x)) which is integrable and bounded by 1 shows that
f h(0) =
_
B
B
sign(h(t))h(t)dt > A.
This shows that the operator f f h is not bounded on L

, and thus the lter h is unstable.

Exercise 2.4. One has
f
r
(t) = Re[f(t)] = [f(t) +f

(t)]/2 and f
i
(t) = Ima[f(t)] = [f(t) f

(t)]/2
so that

f
r
() =
_
f(t) +f

(t)
2
e
it
dt =

f()/2 + Conj
__
f(t)e
it
dt
_
/2
= [

f() +

f

()]/2,
where Conj(a) = a

is the complex conjugate. The same computation leads to

f
i
() = [

f()

f

()]/2.
Exercise 2.6. If f = 1
[,]
then one can verify that

f() =
2 sin()

.
1
It result that
_
sin()

=
1
2
_

f()d = f(0) = 1.
If g = 1
[1,1]
then g()/2 = sin()/. The inverse Fourier transform of g()
3
is g g g(t) so
_
sin
3
()

3
d =
1
8
_
g()
3
d =
2
8
g g g(0) =
3
4
,
where we used the fact that
g g g(0) =
_
1
1
h(t)dt = 3
where h is a piecewise linear hat function with h(0) = 2.
Exercise 2.8. If f is C
1
with a compact support, with an integration by parts we get

f() =
1
i
_
f

(t)e
it
dt
so that
[

f()[
C

with C =
_
[f

(t)[dt < +,
which proves that f() 0 when [[ +.
Let f L
1
(R) and > 0. Since C
1
functions are dense in L
1
(R), one can nd g such that
_
[f g[ /2. Since g() 0 when [[ +, there exists A such that [ g()[ /2 when
[[ > A. Moreover, the Fourier integral denition implies that
[

f() g()[
_
[f(t) g(t)[ dt
so for all [[ > A we have [

f()[ which proves that f() 0 when [[ +.

Exercise 2.10. For a > 0 and u > 0 and g a Gaussian function, dene
f
a,u
(t) = e
iat
g(t u) +e
iat
g(t +u).
We verify that

(f
a,u
) increases proportionally to u. Its Fourier transform is

f
a,u
() = e
iu
g( a) +e
iu
g( +a)
so

(f
a,u
) increases proportionally to a. For a and u suciently large we get the the result.
Exercise 2.12. a) Denoting u(t) = [ sin(t)[, one has g(t) = a(t)u(
0
t) so that
g() =
1
2
a() u(/
0
)
where u() is a distribution
u() =

n
c
n
( n)
and c
n
is the Fourier coecient
c
n
=
_

[ sin(t)[e
int
dt =
_
0

sin(t)e
int
dt +
_

0
sin(t)e
int
dt.
2
The change of variable t t + in the rst integral shows that c
2k+1
= 0 and for n = 2k,
c
2k
= 2
_

0
sin(t)e
i2kt
dt =
4
1 4k
2
.
One thus has
u() =
1
2

n
c
n
a( n
0
) =
2

k
a( 2k
0
)
1 4k
2
.
b) If a() = 0 for [[ >
0
, then h dened by

h() =

2
1
[
0
,
0
]
guarantees that g

h = a and
hence a = g h.
Exercise 2.14. The function (t) = sin(t)/(t) is monotone on [3/2, 0] and [0, 3/2] on which
is variation is 1 +
2
3
. For each k N

, it is also monotone on each interval [k +1/2, k +3/2] on

which the variation is
1

[(k + 1/2)
1
+ (k + 3/2)
1
]. One thus has
[[[[
V
= 2(1 +
2
3
) +
2

k1
[(k + 1/2)
1
+ (k + 3/2)
1
] = +.
For = 1
[a,b]
, [

[ =
a
+
b
and hence [[[[
V
= 2.
Exercise 2.16. Let
f(x) = 1
[0,1]
2(x
1
, x
2
) = f
0
(x
1
)f
0
(x
2
) where f
0
(x
1
) = 1
[0,1]
(x
1
).
One has

f(
1
,
2
) =

f
0
(
1
)

f
0
(
2
) =
(e
i
1
1)(e
i
2
1)

2
.
Let
f(x) = e
x
2
1
x
2
2
= f
0
(x
1
)f
0
(x
2
) where f
0
(x
1
) = e
x
2
1
.
One has

f(
1
,
2
) =

f
0
(
1
)

f
0
(
2
) = e
(
2
1
+
2
2
)/4
.
Exercise 2.18. We prove that the Gibbs oscillation amplitude is independent of the angle and
is equal to a one-dimensional Gibbs oscillation. Let us decompose f(x) into a continuous part
f
0
(x) and a discontinuity of constant amplitude A:
f(x) = f
0
(x) +Au(cos()x
1
+ sin()x
2
)
where u(t) = 1
[0,+)
(t) is the one-dimensional Heaviside function. The lter satises h

(x
1
, x
2
) =
g

(x
1
) g

(x
2
) with g

(t) = sin(t)/(t). The Gibbs phenomena is produced by the discontinuity

corresponding to the Heaviside function so we can consider that f
0
= 0. Let us suppose that
[[ /4, with no loss of generality. We rst prove that
f h

(x) = f g

(x) (1)
where g

(
1
,
2
) = 1
[,]
(
2
). Indeed f(x) is constant along any line of angle , one can thus
verify that its Fourier transform has a support located on the line in the Fourier plane, of angle
+/2 which goes through 0. It results that

f()

() =

f() g

() because the ltering limits

the support of

f to [
2
[ . But g

(x
1
, x
2
) = (x
1
) sin(x
2
)/(x
2
). The convolution (1) is thus
a one-dimensional convolution along the x
2
variable, which is computed in the Gibbs Theorem
2.8. The resulting one-dimensional Gibbs oscillations are of the order of A0.045.
3
2 Chapter 3
Exercise 3.2. If Supp(

f) [/s, /s], then

f() =

f()1
[/s,/s]
() =
1
s

f()

s
()
and hence using the Fourier convolution theorem and the fact that
s
is symmetric,
f(u) =
1
s
f
s
(u) =
1
s
f(t),
s
(t u)).
Exercise 3.4. Let A(t) =

n
f(t n), then in the sense of distribution

A() =

f()

n
e
in
=

f()

n
( 2n) =

f(2n)( 2n).
Taking the inverse Fourier transform leads to

n
f(t n) =

f(2n)e
2in
.
Exercise 3.6. The sucient condition comes from
[[Lf[[

[[f[[

[[h[[
1
.
If [[h[[

= +, then for any A > 0 it exists B such that

|k|<B
[h[k][ > A. Taking f[k] =
sign(h[k])1
[B,B]
[k] that satises [[f[[

1 shows that
Lf[0] =

|k|B
[h[k][ > A.
The operator f f h is not bounded on

, so that the lter is unbounded.

Exercise 3.8. Let e
k
[n] = e
2i
N
kn
=
kn
where = e
2i
N
. If k ,= k

, one has a geometrical sum

e
k
, e
k
) =

kn

n
=

n
(
kk

)
n
=
1
N(kk

)
1
kk

= 0
because
N
= 1. Since [[e
k
[[ =

N, the family e
k
/

N
k
is an orthonormal basis of C
N
.
Exercise 3.10. a) One has

f(ns) = f
s
where
s
= 1
[s/2,s/2]
.
b) Since

f() =

f()

s
(), supp(

f) [/s, /s], and hence

f is recovered using Shannon
interpolation formula.
c) One has

f() =

f()
sin(s/2)
s/2
and thus
f =

f
s
with

s
() =
s/2
sin(s/2)
.
d) For [/s, /s], one has

s
() > 2/
4
and hence [[f[[ /2[[

f[[ which shows that the reconstruction is stable.

Exercise 3.12. One has
[

f[k][ = [

n
f[n]e

2i
N
nk
[

n
[f[n][[e

2i
N
nk
[ =

n
[f[n][.
Exercise 3.14. Since (1)
n
= e
in
, one has
g() =

n
h[n]e
in
e
in
=

h( +).
If h is a low-pass lter, then g is a high-pass lter.
Exercise 3.16. a) Since

h()

h
1
() =

() = 1, one has

h
1
= 1/

h.
b) Up to translation we suppose that h is causal. The lter h
1
and h have nite support if and
only if

h() = P(e
i
) where P(z) and z
k
/P(z) are polynomial for some k N. This can only
happens if P(z) is a monomial P(z) = az
p
, so that h[n] = a[n p].
Exercise 3.18. a) For h
0
[n] = a
n
1
[0,+)
[n], one has the following geometrical sum

h
0
() =

n
(ae
i
)
n
=
1
1 ae
i
.
Let h
p
() = (1 ae
i
)
p
. Observe that
h

p
() =
paie
i
(1 ae
i
)
p+1
=

n
h
p
[n](in)e
in
.
The inverse Fourier transform of h
p+1
() = (1 ae
i
)
p1
thus satises
h
p+1
[n] =
(n + 1) h
p
[n + 1]
a p
.
Iterating on this relation with h
1
[n] = a
n
1
[0,+)
[n] gives
h
p
[n] =
h
1
[n +p 1] ,

p1
k=1
(n +k)
a
p1
(p 1)!
.
b) Taking the Fourier transform of the recursion formula shows that
g() =

h()

f() where

h() =

K
k=0
a
k
e
ik

M
k=0
b
k
e
ik
.
c) Let a
k
be the roots of the polynomial

M
k=0
b
k
z
k
with multiplicity p
k
. Then

h() = P(e
i
)

k
(1 a
k
e
i
)
p
k
,
where P is a polynomial. If one has [a
k
[ < 1 for all k, then using question a), each (1a
k
e
i
)
p
k
is the Fourier transform of a stable causal lter, and so is h.
If there exists l such that [a
l
[ = 1, then it can be written a
l
= e
i
so [

h()[ = + and the

lter can therefore not be stable.
5
If, there exists l with [a
l
[ > 1 then let h
l
() = (1 a
l
e
i
)
p
l
. Observe that
h
l
() = (1 a
l
e
i
)
p
l
= a
p
l
l
e
ip
l
(a
1
l
e
i
1)
p
1
.
Its inverse Fourier transform is h
l
[n] and with question (a) we verify that it is not a causal
lter. To prove that h is then not a causal lter, one can decompose

h() =
L

k=0
Q
k
(e
i
)
(1 a
k
e
i
)
n
k
where each Q
k
is a polynomial of degree strictly smaller than n
k
and observe that the component
for k = l is not causal so h can not be causal.
Exercise 3.20. a) One has
x() =

n
y[Mn]e
in
=

n
(y h)[n]e
in/M
=
1
2
y

h(/M)
where we have use the convolution result of Exercise 3.7 and where
h[n] =

k
[n kM]
which satises, in the sense of distributions

h() =
2
M

k
( 2k/M).
This shows that
x() =
1
M
_

y()

k
(/M 2k/M)d =
1
M
M1

k=0
y(( 2k)/M).
b) If for [, ], y() = 0 for / [/M, /M], then
y() = x(M)

M
() where

M
() = M1
[/M,/M]
().
Let (x M)[Mn] = x[n] and (x M)[k] = 0 if k mod M ,= 0 be the up-sampled signal. The
Fourier transform of x M is x(M), so that
y = (x M)
m
where
M
[n] =
sin(n/M)
n/M
.
Exercise 3.22. a) One has

f[] =
N
2
([ k] +[ +k]) .
So
f
a
[n] = e
2i
N
kn
.
b) One has g[n] = (f[n] + f

[n])/2, and using a change of variable n n in the summation

gives
g[k] =

f[k]/2 +

n
f

[n]e

2i
N
kn
=

f[k]/2 +

f[k]

/2.
6
c) The denition of

f
a
shows that g = Re(f
a
) satises
g[k] = (

f
a
[k] +

f
a
[k]

)/2 =

f[k],
and hence g = f.
Exercise 3.24. We dene the sub-images
0 k
1
, k
2
< L, 0 n
1
, n
2
< M, f
k
[n] = f[n +kM].
One has
f[n] =

k
f
k
[n kM],
so that
f h[n] =

k
f
k
[n kM] h[n] =

k
(

f
k
h)[n kM]
where we have denoted by

f
k
the image of size (2M 1) (2M 1) obtained by zero-padding
from f
k
. This allows one to compute f h using L
2
circular FFT of size (2M 1)
2
, followed by
4N additions to reconstruct the full image. The overall complexity of this overlap add method
is thus approximately 8KN log(M).
The complexity of a direct evaluation of the convolution is K

NM
2
, where K

2 (1 addition
and 1 multiplication).
For K = 6 and K

= 2, using the overlap-add algorithm is better in the range of M such that

48N log
2
(M) 2NM
2
M
2
/ log
2
(M) 24
which we found numerically to be M 9.
3 Chapter 4
Exercise 4.2. The formula (2.32) for the Fourier transform of the Gaussian implies
Ag(, ) =
1

2
_
exp
_
1
2
2
_
(v +/2)
2
+ (v /2)
2

_
e
iv
dv
=
1

2
e

2
4
2
_
e
v
2
e
iv
dv = exp
_

2
4
2


2

2
4
_
Exercise 4.4. One can write the discrete Fourier transform in (4.27)
Sf[m, l] = e
i2lm/N
f g
l
[m] with g
l
[n] = g[n] e
i2ln/N
.
It is thus computed with N convolutions. Since g
l
has a support of size L, the fast overlapp-add
convolution algorithm of Section 3.3.4 computes each convolution with O(N log L) operations.
The windowed Fourier transform is thus computed with O(N
2
log L) operations. Exercise
4.6. One has
b(t) =
1
C

_
s
0
0
f

s

s
(t)
ds
s
2
+
1
C

s
0
f

s
0

s
0
.
Taking the Fourier transform leads to

b() =

f()
C

__
s
0
0
[

s
()[
2
ds
s
2
+
1
s
0
[

s
0
[
2
_
.
7
Using the fact that [

s
()[
2
= s[

(s)[
2
and
[

s
0
[
2
=
_
+
1
[

(s
0
s)[
2
ds
s
=
_
+
s
0
[

(s)[
2
ds
s
leads to the result.
Exercise 4.8. Let
b(t) =
1
C
_
+
0
f
s
(t)
ds
s
3/2
.
Its Fourier transform reads, after a change of variable = s,

b() =

f()
C
_
+
0

(s)
ds
s
3/2
=

f()
C
_
+
0

()
ds
s
=

f().
Exercise 4.10. For f(t) = cos((t)) with (t) = a cos(bt), the width s of the window is small
enough if
s
2
[

(t)[ = s
2
ab
2
[ cos(bt)[ 1
and thus s
2
ab
2
1.
For (t) = cos(
1
(t)) + cos(
2
(t)) with
1
(t) = a cos(bt) and
2
(t) = a cos(bt) + ct, there is
enough frequency resolution if
[

1
(t)

2
(t)[ = [c[

s
and enough spacial resolution if
s
2
[

1
(t)[ = s
2
[

2
(t)[ = s
2
ab
2
1.
This shows that one needs
_
c
ab
2

1.
Exercise 4.16. Example 4.20 shows that P

f is a spetrogram with a Gaussian window g

if
(u, ) = g

(u)g

() = P
V
g

(u, ) = g
/

2
(u)g

2
(),
where we have used (4.125) for the last equality. If = 1/2, then one can take =

2.
Otherwise, if > 1/2, one decomposes =
0

1
where

0
(u, ) = g
/

2
(u)g

2
() and
1
(u, ) = g
/

2
(u)g

2
()
where is chosen so that /

2 > 0 and

2 > 0. This shows that P

is a smoothing
with
1
of P

0
which is a spectrogram and hence positive.
Exercise 4.18. One has
A(t) =
_
(

(t))
2
P
V
f(t)d =
_
h()
__
(

(t))
2
e
i
d
_
d
where
h() = a(t +/2)a(t /2)e
i[(t+/2)(t/2)]
.
8
In the sense of distribution, one has the following equality
_
(

(t))
2
e
i
d = 2[

() + 2i

(t)

() +

(t)
2
()]
where ,

are the Dirac distribution and its rst and second derivatives. This shows that
A(t) = 2[h

(0) + 2i

(t)h

(0) +

(t)
2
h(0)].
After computing the derivatives of h and simplication, one nds
A(t) = (a(t)a

(t) a

(t)
2
),
which is the result.
4 Chapter 5
Exercise 5.2. A change of variable t Kt gives
f,
p
) =
_
1
0
f(t)e
2ipt/K
dt = K
_
1/K
0
f(Kt)e
2ipt
=
_
1
0
f
K
(t)e
p
(t)dt,
where e
p
(t) = e
2it

p
is an orthonormal basis of L
2
[0, 1]. Let us dene f
K
(t) = f(Kt) for
t [0, 1/K] and f
K
(t) = 0 otherwise. For K 1, we get

p
[f,
p
)[
2
=

p
[f
K
, e
p
)[
2
= [[f
K
[[
2
= K[[f[[
2
So
p

p
is a tight frame of L
2
[0, 1] of frame bound K.
Exercise 5.4. One has
tr(U
1
U
2
) =

i,j
U
1
[i, j]U
2
[j, i] = tr(U
2
U
1
).
Exercise 5.6. If one does not constrain [[
p
[[, one can choose
0
/N, N
1
,
2
. . . ,
N1
. If one
constrains [[
p
[[ = 1, one can choose
p
[n] = [n + p mod N] so that f = f and thus one
should impose

[][
2
= N and [

[][ > 0.
This can be achieved by setting

[] = 1/N for ,= 0 and

[0] =
_
N (N 1)/N
2
. This
implies
A = min

[][
2
= 1/N
2
0
and
B = max

[][
2
= N (N 1)/N
2
+.
Exercise 5.8. One has

m
f[p] = f,
m+p
) = f
m
[p] =

m
9
where
m
is a convolution operator, whose eigenvalues are

m
[], and eigenvectors are the
discrete Fourier vectors. The eigenvalues of

are thus

m
[

m
[][
2
.
Exercise 5.10. a) Since 1/

Ke
2i
K
nk

0k<K
is an orthogonal basis of the signal supported in
I = [K/2, K/2 1] +mM, and g[n mM]f[n] is supported in I, one has

n
[g[n mM][
2
[f[n][
2
=
K1

k=0
[f,
1

K
g
mnk
)[
2
.
b) Summing over m leads to
K

n
[f[n][
2
N/M1

m=0
[g[n mM][
2
=

k,m
[f, g
m,k
)[
2
,
and hence the result of Theorem 5.18.
Exercise 5.12. For m = 3, one has

h()/

2 = cos
4
(/2)
4
= P(z) = (z/2 +z
1
/2)
4
= (z
2
+ 4z + 6 + 3z
1
+z
2
)/2
4
and one can choose
g()/

2 = sin
2
(/2) = Q(z) = (z 2 +z
1
)/4.
If

h = h, then

g() =
2 [

h()[
2
g

()
= 2
1 cos
4
(/2)

2 sin
2
(/2)
=

2[1 + cos
2
(/2)] =

2

Q(z)
where

Q(z) = 1 + (z + 2 +z
1
)/4, so g/

2 = [1/4, 3/2, 1/4].

Exercise 5.14. As 2, the frame becomes less orthogonal (A/B +), so the dual
vectors become more and more dierent from the primal ones, and hence the windows g and g
dier more and more.
Exercise 5.16. a) Each
s
(t ns ks/K)
n
is an orthogonal basis of U
s
, the set of signals
with Fourier support included in [/s, /s]. This proves that
sn/Ks

n
is a tight frame with
frame bound K.
b) One has
Pa[n] =

a[n

]H[n, n

] where H[n, n

] =
s
(t n/Ks),
1
K

s
(t n

/Ks)),
and, using Plancherel formula,
H[n, n

] =
1
2K
s
1/2
1
[/s,/s]
()e
in/Ks
, s
1/2
1
[/s,/s]
()e
in

/Ks
)
=
s
2K
_
/s
/s
e
i(nn

)/Ks
d = h[n n

] where h[n] =
sin(n/K)
n
.
10
c) Since

h = 1
[/K,/K]
, one has
Im(P) = a h a = a = a supp( a) [/K, /K] .
d) Theorem 3.5 proves that s
1/2
f
s
(ns) = f(ns).
e) Let a[n] = f
s
(ns
0
). For [, ], one has
a() =
1
s
0
T(f
s
)(/s
0
) = supp( a) [/K, /K]
where T is the Fourier transform. This shows that a Im() and hence Pa = a.
One has
PY [Kn] = a[Kn] +P(W)[Kn] = [PY [Kn] s
1/2
f(ns)[
2
= [W h[Kn][
2
.
One then has
E([W h[Kn][
2
) =
2

k
[h[k][
2
=

2
2
_

h()[
2
=

2
K
0 when K +.
Exercise 5.18. We prove the right hand side of the inequality

j=
[

(2
j
)[
2
=
1
2

j=
[ g(2
j1
)[
2
[

(2
j1
)[
2
B

j=
(1 [

h(2
j1
)[
2
)[

(2
j1
)[
2
B

j=
([

(2
j1
)[
2
[

(2
j
)[
2
) = B([

(2
1
)[
2
[

(2

)[
2
).
Since

(0) = 1 and () tends to zero as increases, letting and go to and + proves
that

j
[

(2
j
)[
2
B. A similar proof applies to the left hand side.
Exercise 5.21. We use the change of variables
x = cos() usin() and y = sin() +ucos().
For k + < p, one can expand the monomial P(x, y) = x
k
y

as a polynomial in u of degree less

than p
P(x, y) =

t<p
A()u
t
.
This shows that
P, ) =

t<p
_
A()
_ _
u
t
( cos() usin(), sin() +ucos())du
_
d = 0.
11