Wavelet Tour Corrections

Correction of the exercises from the book
A Wavelet Tour of Signal Processing

Gabriel Peyré
Ceremade
Université Paris-Dauphine
gabriel.peyre@ceremade.dauphine.fr
Nov. 2009
Abstract
rd
These corrections refer to the 3 edition of the book A Wavelet Tour of Signal Processing
– The Sparse Way by Stéphane Mallat, published in December 2008 by Elsevier. If you
find mistakes or imprecisions in these corrections, please send an email to Gabriel Peyré
(gabriel.peyre@ceremade.dauphine.fr). More information about the book, including how
to order it, numerical simulations, and much more, can be find online at wavelet-tour.com.
1 Chapter 2
Exercise 2.1. For all t, the function ω 7→ e−iωt f (t) is continuous. If f ∈ L1 (R), then for all ω,
|e−iωt f (t)| 6 R|f (t)| which is integrable. One can thus apply the theorem of continuity under the
integral sign which proves that fˆ is continuous.
If fˆ ∈ L1 (R), using the inverse Fourier formula (2.8) and a similar argument, one proves that
f is continuous.
R RB
Exercise 2.2. If |h| = +∞, for all A > 0 there exists B > 0 such that −B |h| > A. Taking
f (x) = 1[−A,A] sign(h(−x)) which is integrable and bounded by 1 shows that
Z B
f ? h(0) = sign(h(t))h(t)dt > A.
−B
This shows that the operator f 7→ f ? h is not bounded on L∞ , and thus the filter h is unstable.
Exercise 2.3. Let fu (t) = f (t − u), by change of variable t − u → t, one gets
Z Z
fû (ω) = f (t − u)e−iωt dt = f (t)e−iω(t+u) dt = e−iωu fˆ(ω).
Let fs (t) = f (t/s), with s > 0, by change of variable t/s 7→ t, one get
Z Z
fˆs (ω) = f (t/s)e−iωt dt = f (t)e−iωst |s|dt = |s|fˆ(sω).
1
Let f by C 1 and g = f 0 , the by integration by parts, since f (t) → 0 where |t| → +∞,
Z Z
ĝ(ω) = f (t)e0 −iωt
dt = − f (t)(−iω)e−iωt dt = (iω)fˆ(ω).
Exercise 2.4. One has
fr (t) = Re[f (t)] = [f (t) + f ∗ (t)]/2 and fi (t) = Ima[f (t)] = [f (t) − f ∗ (t)]/2
so that
f (t) + f ∗ (t) −iωt
Z Z
fˆr (ω) = e dt = fˆ(ω)/2 + Conj f (t)eiωt dt /2
2
= [fˆ(ω) + fˆ∗ (−ω)]/2,
where Conj(a) = a∗ is the complex conjugate. The same computation leads to
fî (ω) = [fˆ(ω) − fˆ∗ (−ω)]/2.
Exercise 2.5. One has Z

fˆ(0) = f (t)dt = 0.
If f ∈ L1 (R), one can apply the theorem of derivation under the integral sign and get
R
Z Z
d ˆ −iωt ˆ0
f (ω) = −itf (t)e dt =⇒ f (0) = −i tf (t)dt = 0.
dω
Exercise 2.6. If f = 1[−π,π] then one can verify that
2 sin(πω)
fˆ(ω) = .
ω
It result that Z Z
sin(πω) 1
= fˆ(ω)dω = f (0) = 1.
πω 2π
If g = 1[−1,1] then ĝ(ω)/2 = sin(ω)/ω. The inverse Fourier transform of ĝ(ω)3 is g ? g ? g(t) so
sin3 (ω)
Z Z
1 2π 3π
dω = ĝ(ω)3 dω = g ? g ? g(0) = ,
ω3 8 8 4
where we used the fact that Z 1
g ? g ? g(0) = h(t)dt = 3
−1
where h is a piecewise linear hat function with h(0) = 2.

R
Exercise 2.7. Writing u = a − ib, and differentiating under the integral sign , one has
Z
2
f 0 (ω) = −ite−ut e−iωt dt.
By integration by parts, one gets an ordinary differential equation

−ω ˆ
f 0 (ω) = f (ω)
2u
2
whose solution is
ω2
f (ω) = Ke− 4u
for some constant K = fˆ(0). Using a switch from Euclidean coordinates to polar coordinates
(x, y) → (r, θ) which satisfies dxdy = rdrdθ, one gets
Z Z ZZ
2 2 2 2
K 2 = e−ux dx e−uy dy = e−u(x +y ) dxdy
Z 2π Z +∞ Z +∞
2 2 π
= e−ur rdrdθ = 2π re−ur dr = ,
0 0 0 u
which gives the result.
Exercise 2.8. If f is C1 with a compact support, with an integration by parts we get
Z
ˆ 1
f (ω) = f 0 (t)e−iωt dt
iω
so that Z
C
|fˆ(ω)| 6 with C= |f 0 (t)|dt < +∞,
ω
which proves that f (ω) → 0 when |ω| → +∞.
1 1 1
R Let f ∈ L (R) and ε > 0. Since C functions are dense in L (R), one can find g such that
|f − g| 6 ε/2. Since ĝ(ω) → 0 when |ω| → +∞, there exists A such that |ĝ(ω)| 6 ε/2 when
|ω| > A. Moreover, the Fourier integral definition implies that
Z
|fˆ(ω) − ĝ(ω)| 6 |f (t) − g(t)| dt
so for all |ω| > A we have |fˆ(ω)| 6 ε which proves that f (ω) → 0 when |ω| → +∞.
Exercise 2.9. a) For f0 (t) = 1[0,+∞) (t)ept , we get
Z +∞
1
fˆ0 (ω) = e(p−iω)t dt = .
0 iω − p
n pt
For fn (t) = t 1[0,+∞) (t)e , an integration by parts gives
Z +∞
ˆ n ˆ
fn (ω) = tn e(p−iω)t dt = fn−1 (ω),
0 iω −p
so that
n!
fˆn (ω) = .
(iω − p)n
b) Computing the Fourier transform on both sides of the differential equation gives
PK
ak (iω)k
g = f ? h where ĥ(ω) = Pk=0 M
.
k
k=0 bk (iω)
PM
We denote by {pk }L
k=0 the poles of the polynomial
k
k=0 bk z , with multiplicity nk . If K < M ,
one can decompose the rational fraction into
L
X Qk (iω)
ĥ(ω) =
(iω − pk )nk
k=0
3
where each Qk is a polynomial of degree strictly smaller than nk . It results that h(t) is a sum of
derivatives up to a degree strictly smaller than nk of the inverse Fourier transform of
1
fˆpk ,nk (ω) =
(iω − pk )nk
which is
1 nk
fpk ,nk (t) = t 1[0,+∞) (t)epk t .
nk !
Each filter fpk ,nkis causal, stable and nk times differentiable. It results that that h is causal and
stable.
If, there exists l with Re(pl ) = 0 then for the frequency ω = −ipl we have |ĥ(ω)| = +∞ so h
can not be stable.
If, there exists l with Re(pl ) > 0 then by observing that fˆpl ,nl (−ω) = (−1)nl (iω + pl )−nl and
by applying the result in a) we get
1 nl
fpl ,nkl (t) = t 1(−∞,0] (t)e−pl t
nl !
which is anticausal. We thus derive that h is not causal.
c) Denoting α = eiπ/3 , one can write
1
|ĥ(ω)|2 =
1 − (iω/ω0 )6
with
1/ĥ(ω) = (iω/ω0 + 1)(iω/ω0 + α)(iω/ω0 + α∗ ) = P (iω).
Since the zeros of P (z) have all a strictly negative real part, h is stable and causal. To compute
h(t) we decompose
a1 a2 a3
ĥ(ω) = + + ,
iω/ω0 + 1 iω/ω0 + α iω/ω0 + α∗
we compute a1 , a2 and a3 and by applying the result in (a) we derive that
∗
ĥ(t) = ω0 (a1 1[0,+∞) (t) e−tω0 + a2 1[0,+∞) (t) e−tαω0 + a3 1[0,+∞) (t) e−tα ω0
).
Exercise 2.10. For a > 0 and u > 0 and g a Gaussian function, define
fa,u (t) = eiat g(t − u) + e−iat g(t + u).
We verify that σω (fa,u ) increases proportionally to u. Its Fourier transform is
fâ,u (ω) = e−iuω ĝ(ω − a) + eiuω ĝ(ω + a)
so σω (fa,u ) increases proportionally to a. For a and u sufficiently large we get the the result.
Exercise 2.11. Since f (t) > 0
Z Z
|fˆ(ω)| = | −iωt
f (t) e dt| 6 f (t) dt = fˆ(0) .
Exercise 2.12. a) Denoting u(t) = | sin(t)|, one has g(t) = a(t)u(ω0 t) so that
1
ĝ(ω) = â(ω) ? û(ω/ω0 )
2π
4
where û(ω) is a distribution X
û(ω) = cn δ(ω − n)
n
and cn is the Fourier coefficient

Z π Z 0 Z π
cn = | sin(t)|e−int dt = − sin(t)e−int dt + sin(t)e−int dt.
−π −π 0
The change of variable t → t + π in the first integral shows that c2k+1 = 0 and for n = 2k,
Z π
4
c2k = 2 sin(t)e−i2kt dt = .
0 1 − 4k 2
One thus has

1 X 2 X â(ω − 2kω0 )
û(ω) = cn â(ω − nω0 ) = .
2π n π 1 − 4k 2
k
π
b) If â(ω) = 0 for |ω| > ω0 , then h defined by ĥ(ω) = 2 1[−ω0 ,ω0 ] guarantees that ĝ ĥ = â and
hence a = g ? h.
1X ˆ 1X ˆ
ĝ(ω) = fn (ω) ? [δ(ω − 2nω0 ) + δ(ω + 2nω0 )] = [fn (ω − 2nω0 ) + fˆn (ω + 2nω0 )].
2 n 2 n
Each fˆn (ω ± 2nω0 ) is supported in [(−1 ± 2n)ω0 , (1 ± 2n)ω0 ], and thus ĝ is supported in
[−2N ω0 , 2N ω0 ].
Since the intervals [(−1 ± 2n)ω0 , (1 ± 2n)ω0 ] are disjoint, one has
fˆn (ω ± 2nω0 ) = 2ĝ(ω)1[(−1±2n)ω0 ,(1±2n)ω0 ] (ω).
The change of variable ω ± 2nω0 → ω and summing for n and −n gives
fˆn (ω) = [ĝ(ω − 2nω0 ) + ĝ(ω + 2nω0 )]ĥ(ω),
where ĥ(ω) = 1[−ω0 ,ω0 ] (ω). Denoting gn (t) = 2g(t) cos(2nω0 t), one sees that fn is recovered as
fn = gn ? h.
Exercise 2.14. The function φ(t) = sin(πt)/(πt) is monotone on [−3/2, 0] and [0, 3/2] on which
2
is variation is 1 + 3π . For each k ∈ N∗ , it is also monotone on each interval [k + 1/2, k + 3/2] on
which the variation is π1 [(k + 1/2)−1 + (k + 3/2)−1 ]. One thus has
2 2X
||φ||V = 2(1 + )+ [(k + 1/2)−1 + (k + 3/2)−1 ] = +∞.
3π π
k>1
For φ = λ1[a,b] , |φ0 | = λδa + λδb and hence ||φ||V = 2λ.

Exercise 2.16. Let
f (x) = 1[0,1]2 (x1 , x2 ) = f0 (x1 )f0 (x2 ) where f0 (x1 ) = 1[0,1] (x1 ).
5
One has
(eiω1 − 1)(eiω2 − 1)
fˆ(ω1 , ω2 ) = fˆ0 (ω1 )fˆ0 (ω2 ) = .
ω1 ω2
Let 2 2 2
f (x) = e−x1 −x2 = f0 (x1 )f0 (x2 ) where f0 (x1 ) = e−x1 .
One has 2 2
fˆ(ω1 , ω2 ) = fˆ0 (ω1 )fˆ0 (ω2 ) = πe−(ω1 +ω2 )/4 .
Exercise 2.17. If |t| > 1, the ray ∆t,θ does not intersect the unit disc, and thus pθ (t) = 0. For
|t| < 1, the Radon transform is computed as the length of a cross section of a disc
p
pθ (t) = 2 1 − t2 .
Exercise 2.18. We prove that the Gibbs oscillation amplitude is independent of the angle θ and
is equal to a one-dimensional Gibbs oscillation. Let us decompose f (x) into a continuous part
f0 (x) and a discontinuity of constant amplitude A:
f (x) = f0 (x) + A u(cos(θ)x1 + sin(θ)x2 )
where u(t) = 1[0,+∞) (t) is the one-dimensional Heaviside function. The filter satisfies hξ (x1 , x2 ) =
gξ (x1 ) gξ (x2 ) with gξ (t) = sin(ξt)/(πt). The Gibbs phenomena is produced by the discontinuity
corresponding to the Heaviside function so we can consider that f0 = 0. Let us suppose that
|θ| 6 π/4, with no loss of generality. We first prove that
f ? hξ (x) = f ? gξ (x) (1)
where ĝξ (ω1 , ω2 ) = 1[−ξ,ξ] (ω2 ). Indeed f (x) is constant along any line of angle θ, one can thus
verify that its Fourier transform has a support located on the line in the Fourier plane, of angle
θ + π/2 which goes through 0. It results that fˆ(ω)ĥξ (ω) = fˆ(ω)ĝξ (ω) because the filtering limits
the support of fˆ to |ω2 | 6 ξ. But gξ (x1 , x2 ) = δ(x1 ) sin(ξx2 )/(πx2 ). The convolution (1) is thus
a one-dimensional convolution along the x2 variable, which is computed in the Gibbs Theorem
2.8. The resulting one-dimensional Gibbs oscillations are of the order of A × 0.045.
2 Chapter 3
Exercise 3.1. One has φs,n (t) = s−1/2 1[ns,(n+1)s) , which satisfies ||φs,n || = 1 and hφs,n , φs,n0 i = 0
for n 6= n0 because [ns, (n + 1)s) and [n0 s, (n0 + 1)s) are disjoint. If f (x) = an on each interval
[ns, (n + 1)s), then X X
f (x) = an 1[ns,(n+1)s) = hf, φs,n iφs,n
n n
So {φs,n }n is an orthonormal basis of functions that are piecewise constant on each interval
[ns, (n + 1)s).
Exercise 3.2. If Supp(fˆ) ⊂ [−π/s, π/s], then
1
fˆ(ω) = fˆ(ω)1[−π/s,π/s] (ω) = fˆ(ω)φ̂s (ω)
s
6
and hence using the Fourier convolution theorem and the fact that φs is symmetric,
1 1
f (u) = f ? φs (u) = hf (t), φs (t − u)i.
s s
Exercise 3.3. a) The function f is an interpolation function if and only if

X
f (n)δ(t − n) = δ(t)
n
in the distribution sense. Using the sampling Theorem 3.1 with s = 1, one gets equivalently the
equality of the Fourier transform X
fˆ(ω + 2kπ) = 1.
k
b) One has X
fˆ(ω) = h[n]θ(ω)e−inω = h(ω)θ(ω).
n
and thus X X
fˆ(ω + 2kπ) = h(ω)A(ω) where A(ω) = θ̂(ω + 2kπ)
k k
If for all ω, A(ω) 6= 0, one can set h(ω) = 1/A(ω) and f (ω) = θ(ω)/A(ω). Thus f ∈ L2 (R) if it
exists B > 0 such that |A(ω)| > B, in which case ||f || 6 ||θ||/B.
P
Exercise 3.4. Let A(t) = n f (t − n), then in the sense of distribution
X X X
Â(ω) = fˆ(ω) e−inω = fˆ(ω) δ(ω − 2nπ) = fˆ(2nπ)δ(ω − 2nπ).
n n n
Taking the inverse Fourier transform leads to

X X
f (t − n) = fˆ(2nπ)e2inπω .
n n
Exercise 3.5. The orthogonal projection of f on Us is defined by
∀ g ∈ Us , hf˜ − f, f˜ − gi = 0.
It thus satisfies
∀ g ∈ Us , ||f − g||2 = ||f − f˜||2 + ||f˜ − g||2 > ||f − f˜||2
and hence f˜ minimizes ||f − f˜|| subject to f˜ ∈ Us .
Exercise 3.6. The sufficient condition comes from
||Lf ||∞ 6 ||f ||∞ ||h||1 .

P
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 satisfies ||f ||∞ 6 1 shows that
X
Lf [0] = |h[k]| > A.
|k|6B
7
The operator f 7→ f ? h is not bounded on `∞ , so that the filter is unbounded.
! !
Z π X X
−inξ −ip(ω−ξ)
ĝ ? ĥ(ω) = h[n]e g[p]e dξ
−π n p
Z π X
= h[n]g[p]e−iξ(n−p) e−iωp dξ.
−π n,p
R P
Exchanging signs and , and using the fact that
Z π
e−iξ(n−p) dξ = δ[n − p]
−π
shows that X
ĝ ? ĥ(ω) = h[n]g[n]e−inω = fˆ(ω).
n
2iπ 2iπ
Exercise 3.8. Let ek [n] = e N kn = ω kn where ω = e N . If k 6= k 0 , one has a geometrical sum
0
X 0 X 0 1 − ω N (k−k )
hek , ek0 i = ω kn ω −k n = (ω k−k )n = =0
n n
1 − ω k−k0
√ √
because ω N = 1. Since ||ek || = N , the family {ek / N }k is an orthonormal basis of CN .
Exercise 3.9. Denoting X
fd (t) = f (ks)δ(t − ks),
k
one has using Theorem 3.1

1X ˆ
fˆd (ω) = f (ω − 2kπ/s).
s
k
Since fˆ is supported in In = [−(n + 1)π/s, −nπ/s] ∪ [nπ/s, (n + 1)π/s], and the intervals In + 2kπ
are disjoint, one has
fˆ(ω) = fˆd (ω)φ̂s (ω) where φs (ω) = s1In (ω),
which corresponds to the reconstruction formula
X
f (t) = f (ns)φs (t − ns)
n
with the kernel obtained by inverse Fourier transform formula

Z −nπ/s Z (n+1)π/s
iωt
φs (t) = s e dω + s eiωt dω
−(n+1)π/s nπ/s
1
= [sin((n + 1)π/s) − sin(nπ/s)]
πω/s
Exercise 3.10. a) One has f˜(ns) = f ? φs where φs = 1[−s/2,s/2] .

b) Since fˆ˜(ω) = fˆ(ω)φ̂s (ω), supp(fˆ
˜) ⊂ [−π/s, π/s], and hence f˜ is recovered using Shannon
interpolation formula.
8
c) One has
˜(ω) = fˆ(ω) sin(ωs/2)
fˆ
ωs/2
and thus
ωs/2
f = f˜ ? ψs with ψ̂s (ω) = .
sin(ωs/2)
d) For ω ∈ [−π/s, π/s], one has
φ̂s (ω) > 2/π
and hence ||f || 6 π/2||f˜|| which shows that the reconstruction is stable.
Exercise 3.11. a) φ is supported in [−1, 1], on t ∈ [−1, 0], φ(t) = t + 1, on t ∈ [0, 1], φ(t) = 1 − t.
b) If f (t) is linear on each interval [n, n + 1], then
X
f (t) = f (n)φ(t − n).
n
One has, using (7.21),

X X sin4 (ω/2) 1 1
|φ̂(ω − 2kπ)|2 = 2
= (1 + 2 cos2 (ω/2)) > ,
(ω/2 + kπ) 3 3
k k
so using Theorem 3.4, {φ(t − n)}n is a Riesz basis of the space of piecewise linear functions on
each interval [n, n + 1].
c) The dual basis satisfies
3 sin2 (ω/2) ĥ(ω)

φ̃ˆ(ω) = =
(ω/2)2 (1 + 2 cos2 (ω/2)) −ω 2
where ĥ(om) = −12 sin2 (ω/2)/(1 + 2 cos2 (ω/2)) is the Fourier series of a discrete filter. It results
P
that φ(t) is obtained by integrating twice h(t) = n h[n]δ(t − n). The Fourier series ĥ(ω) is a
rational fraction of e−iω which is not reductible to a polynomial so h[n] has an infinite support,
which proves that φ(t) also has an infinite support.
2iπ 2iπ
X X X
|fˆ[k]| = | f [n]e− N nk |6 |f [n]||e− N nk |= |f [n]|.
n n n
Exercise 3.13. a) Let h[0] = 1, h[−1] = −1 and h[n] = 0 for n ∈

/ {−1, 0}. Then
X
||f ||V = |f
? h[n]|
n
and
2iπ
ĥ[k] = 1 − e N k =⇒ |h[k]| = 2| sin(kπ/N )|.
b) By applying the result of the exercise 3.12 we have that the Fourier transform fˆ[k] ĥ[k] of
f? h[n satisfies
|fˆ[k] ĥ[k]| 6 ||f ||V
9
so for |k| 6 N/2 we verify that
||f ||V ||f ||V N ||f ||V

|fˆ[k]| 6 6 6
2| sin(kπ/N )| 2| sin(kπ/N )| 2k
since sin(x) > 2x/π for |x| 6 π/2.

Exercise 3.14. Since (−1)n = e−iπn , one has
X
ĝ(ω) = h[n]eiπn einω = ĥ(ω + π).
n
If h is a low-pass filter, then g is a high-pass filter.

Exercise 3.15. If g = f ? h then ||g||1 6 ||f ||1 ||h||1 and we can exchange the summations order as
follow
X X X X
ĝ(ω) = e−iωn f [p]h[n − p] = f [p]e−iωp e−iω(n−p) h[n − p]
n p p n
X X
= f [p]e −iωp
e −iωn
h[n] = fˆ(ω)ĥ(ω)
p n
where the second line is obtained by change of variable n − p → p in the summation.

Exercise 3.16. a) Since ĥ(ω)ĥ−1 (ω) = δ̂(ω) = 1, one has ĥ−1 = 1/ĥ.
b) Up to translation we suppose that h is causal. The filter h−1 and h have finite support if and
only if ĥ(ω) = P (e−iω ) 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] = aδ[n − p].
Exercise 3.17. a) One has
K
Y |a∗k − e−iω |2
|h(ω)|2 =
|1 + e−iω |2
k=1
and one verifies that |a∗k

− e | = 1 + |ak | + 2Re(ak e−iω ) = |1 + e−iω |2 .
−iω 2 2
b) {h[n − m]}m is an orthogonal basis if and only if for all m, m0
hh[n − m], h[n − m0 ]i = h ? h[m − m0 ] = δ[n − m0 ].
Taking the Fourier transform of this relation leads to ||h(ω)||2 = 1 for all ω.
Exercise 3.18. a) For h0 [n] = an 1[0,+∞) [n], one has the following geometrical sum
X 1
ĥ0 (ω) = (ae−iω )n = .
n
1 − ae−iω
Let hp (ω) = (1 − ae−iω )−p . Observe that
−paie−iω X
h0p (ω) = −iω p+1
= hp [n](−in)e−inω .
(1 − ae ) n
The inverse Fourier transform of hp+1 (ω) = (1 − ae−iω )−p−1 thus satisfies
(n + 1) hp [n + 1]
hp+1 [n] = .
ap
10
Iterating on this relation with h1 [n] = an 1[0,+∞) [n] gives
Qp−1
h1 [n + p − 1] , k=1 (n + k)
hp [n] = .
ap−1 (p − 1)!
b) Taking the Fourier transform of the recursion formula shows that
PK
ˆ ak e−ikω
ĝ(ω) = ĥ(ω)f (ω) where ĥ(ω) = Pk=0 M
.
−ikω
k=0 bk e
PM
c) Let ak be the roots of the polynomial k=0 bk z −k with multiplicity pk . Then
Y
ĥ(ω) = P (e−iω ) (1 − ak e−iω )−pk ,
k
where P is a polynomial. If one has |ak | < 1 for all k, then using question a), each (1−ak e−iω )−pk
is the Fourier transform of a stable causal filter, and so is h.
If there exists l such that |al | = 1, then it can be written al = eiα so |ĥ(α)| = +∞ and the
filter can therefore not be stable.
If, there exists l with |al | > 1 then let hl (ω) = (1 − al e−iω )−pl . Observe that
hl (−ω) = (1 − al eiω )−pl = a−p
l
l −iωpl
e (a−1
l e
−iω
− 1)−p1 .
Its inverse Fourier transform is hl [−n] and with question (a) we verify that it is not a causal
filter. To prove that h is then not a causal filter, one can decompose
L
X Qk (e−iω )
ĥ(ω) =
(1 − ak e−iω )nk
k=0
where each Qk is a polynomial of degree strictly smaller than nk and observe that the component
for k = l is not causal so h can not be causal.
Exercise 3.19. One has, using the inverse Fourier transform,
2N −1
1 X ˆ 2iπ
f˜[2n] = f˜[k]e 2N k2n
2N
k=0
N/2−1 2N −1
1 X ˆ 2iπ 1 X 2iπ 1 2iπ
= f [k]e N kn + fˆ[k]e N kn + fˆ[N/2]e N N/2n ,
N N N
k=0 k=3N/2+1
and one thus has

N −1
1 X ˆ 2iπ
f˜[2n] = f [k]e N kn = f [n].
N
k=0

X X 1
x̂(ω) = y[M n]e−iωn = (y · h)[n]e−iωn/M = ŷ ? ĥ(ω/M )
n n
2π
where we have use the convolution result of Exercise 3.7 and where
X
h[n] = δ[n − kM ]
k
11
which satisfies, in the sense of distributions
2π X
ĥ(ω) = δ(ω − 2πk/M ).
M
k
This shows that

Z π M −1
1 X 1 X
x̂(ω) = y(ξ) δ(ω/M − ξ − 2πk/M )dξ = ŷ((ω − 2kπ)/M ).
M −π M
k k=0
b) If for ω ∈ [−π, π], ŷ(ω) = 0 for ω ∈

/ [−π/M, π/M ], then
ŷ(ω) = x̂(M ω)φ̂M (ω) where φ̂M (ω) = M 1[−π/M,π/M ] (ω).
Let (x ↑ M )[M n] = x[n] and (x ↑ M )[k] = 0 if k mod M 6= 0 be the up-sampled signal. The
Fourier transform of x ↑ M is x̂(M ω), so that
sin(nπ/M )
y = (x ↑ M ) ? φm where φM [n] = .
nπ/M
Exercise 3.21. a) One has, using the change of variable r = n − pN ,

−1
X NX
− 2iπ
X
− 2ikπ 2kπ
fˆp [k] = fd [n − pN ]e N kn = fd [r]e N r = fˆd .
N
p∈Z n=0 r∈Z
Using the sampling Theorem 3.1, one gets

ˆ ˆ 2kπ 1 X ˆ 2kπ 2`π
fp [k] = fd = f − .
N s Ns s
`
b) In order to have
2kπ
sfˆp [k] ≈ fˆ ,
Ns
ˆ
that ω0 π/s. To be able to interpolate without too much aliasing f from the values
one needs
ˆ 2kπ
f N s , one needs that t0 N s. Since a function cannot be compactly supported in both time
and space, no exact interpolation formula is possible.
c) The Fourier transform fˆ(ω) is proportional to the convolution of the indicator of [−π, π]
convolved with itself 4 times. Its support is thus [−4π, 4π].
N
fˆ[`] = (δ[` − k] + δ[` + k]) .
2
So
2iπ
fa [n] = e N kn .
∗
b) One has g[n] = (f [n] + f [n])/2, and using a change of variable n → −n in the summation
gives
2iπ
X
ĝ[k] = fˆ[k]/2 + f ∗ [n]e− N kn = fˆ[k]/2 + fˆ[−k]∗ /2.
n
12
c) The definition of fâ shows that g = Re(fa ) satisfies
ĝ[k] = (fâ [k] + fâ [−k]∗ )/2 = fˆ[k],
and hence g = f .
X
hek1 [n1 ]ek2 [n2 ], ek10 [n1 ]ek20 [n2 ]i = ek1 [n1 ]ek2 [n2 ]e∗k10 [n1 ]e∗k20 [n2 ]
n1 ,n2
! !
X X
= ek1 [n1 ]e∗k10 [n1 ] ek2 [n2 ]e∗k20 [n2 ]
n1 n2
= hek1 , e ihek2 , e i = δ[k1 − k10 ]δ[k2 − k20 ].

k10 k20
Exercise 3.24. We define the sub-images
∀ 0 6 k1 , k2 < L, ∀ 0 6 n1 , n2 < M, fk [n] = f [n + kM ].
One has X
f [n] = fk [n − kM ],
k
so that X X
f ? h[n] = fk [n − kM ] ? h[n] = (f˜k
? h)[n − kM ]
k k
where we have denoted by f˜k the image of size (2M − 1) × (2M − 1) obtained by zero-padding
from fk . This allows one to compute f ? h using L2 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 0 N M 2 , where K 0 ∼ 2 (1 addition
and 1 multiplication).
For K = 6 and K 0 = 2, using the overlap-add algorithm is better in the range of M such that
48N log2 (M ) 6 2N M 2 ⇔ M 2 / log2 (M ) > 24
which we found numerically to be M > 9.

Exercise 3.25. The computation of fˆ is performed by applying a 1D FFT to each direction of
the 3D array. This requires
KN (log(N1 ) + log(N2 ) + log(N3 )) = KN log(N )
operations.
3 Chapter 4
Z
Sf (u, ξ) = eiφ(t) g(t − u)e−iξt dt = ĥ(ξ) where h(t) = eiφ(t) g(t − u).
13
Using Parseval conservation of energy and the fact that ||g|| = 1,
Z Z Z
|Sf (u, ξ)|2 dξ = 2π |h(t)|2 dt = 2π |g(t)|2 = 2π.
b) One has, using the fact that ĥ0 (ξ) = iξ ĥ(ξ) and Parseval conservation of inner product,
Z Z Z
ξ|Sf (u, ξ)|2 dξ = ξ ĥ(ξ)ĥ∗ (ξ)dξ = 2iπ h0 (t)h∗ (t)dt,
and thus, expanding h0 (t)h∗ (t),

Z Z Z
ξ|Sf (u, ξ)| dξ = 2π φ (t)|g(t − u)| dt + 2π g 0 (t − u)g ∗ (t − u)dt,
2 0 2
and the second integral vanishes because g 0 is an odd function.

If one interpret Pu (ξ) = |Sf (u, ξ)|2 /(2π) as a probability density then this result proves that
the average frequency of this density for a fixed t is equal to the averaged instantaneous frequency
φ0 (u) over a neighborhood of t defined by the density |g(t − u)|2 .
Exercise 4.2. The formula (2.32) for the Fourier transform of the Gaussian implies
Z
1 1 2 2
−iγv
Ag(τ, γ) = √ exp (v + τ /2) + (v − τ /2) e dv
πσ 2 2σ 2
τ2 γ 2 σ2
Z
1 τ2 2
=√ e− 4σ2 e−v e−iγv dv = exp − 2 −
πσ 2 4σ 4
Exercise 4.3. The reconstruction (4.66) is exact if and only if for all signal f ,
J
log(a) X 1 1
f= j
? ψj∗
f ? ψj + J ? φ∗J
f ? φJ .
Cψ j=1 a a Cψ
Taking the finite discrete Fourier transform of this relation leads to

 
J
log(a) X 1 1
∀ k, fˆ[k] = fˆ[k] ·  |ψj [k]|2 + J |φJ [k]|2 , 
Cψ j=1 aj a Cψ
and hence the result.

Exercise 4.4. One can write the discrete Fourier transform in (4.27)
Sf [m, l] = e−i2πlm/N f ? gl [m] with gl [n] = g[−n] e−i2πln/N .
It is thus computed with N convolutions. Since gl 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 4.5. a) Applying (4.19) to f = gu0 ,ξ0 proves
ZZ
1
gu0 ,ξ0 = hgu0 ,ξ0 , gu,ξ igu,ξ dξdu.
2π
14
Taking the inner product with gv,ν proves
ZZ
1
K(u0 , v, ξ0 , ν) = K(u0 , u, ξ0 , ξ)K(u, v, ξ, ν)dudξ,
2π
which implies P P Φ = P Φ so that P is a projector on Im(P ), which is exactly, by Theorem 4.2

the functions Φ(u, ξ) that are windowed Fourier transform of some f ∈ L2 .
The projector P is symmetric because
K(u0 , u, ξ0 , ξ) = K(u, u0 , ξ, ξ0 )∗ .
A projector P is orthogonal (which means that Ker(P )⊥Im(P )) if and only it is a symmetric
operator P = P ∗ where P ∗ is the dual (conjugated transposed) operator of P for the inner
product ZZ
1
hΦ1 , Φ2 i = Φ1 (u, ξ)Φ2 (u, ξ)∗ dudξ.
2π
Indeed for all Φ1 ∈ Ker(P ) and Φ2 = P Φ2 ∈ Im(P ),
hΦ1 , Φ2 i = hΦ1 , P Φ2 i = hP Φ1 , Φ2 i = 0.
b) Since an orthogonal projector P is 1-Lipschitz, computing P S̃f reduces the quantization error
because
||Sf − P S̃f || = ||P (Sf − S̃f )|| 6 ||Sf − S f˜||.

Z s0
1 ds 1
b(t) = f ? ψs∗ ? ψs (t) + f ? φ∗s0 ? φs0 .
Cψ 0 s2 Cψ s0
Taking the Fourier transform leads to
fˆ(ω)
Z s0
2 ds 1 2
b̂(ω) = |ψ̂s (ω)| 2 + |φ̂s0 | .
Cψ 0 s s0
Using the fact that |ψ̂s (ω)|2 = s|ψ̂(sω)|2 and

Z +∞ Z +∞
ds ds
|φ̂s0 |2 = |ψ̂(s0 sω)|2 = |ψ̂(sω)|2
1 s s0 s
leads to the result.

Exercise 4.7. Using Plancherel formula and then Fubini,
Z Z Z +∞
2 1 2 1 ds
||φ|| = |φ̂(ω)| dω = |ψ̂(sω)|2 dω
2π 2π 1 s
Z +∞ Z
1 ds
= |ψ̂(sω)|2 dω .
2π 1 s
The change of variable sω → s in the inner integral, and the fact that ||ψ|| = 1 leads to
Z +∞ Z Z +∞
2 1 2 ds ds
||φ|| = |ψ̂(ω)| dω 2
= = 1.
1 2π s 1 s2
15
Exercise 4.8. Let Z +∞
1 ds
b(t) = f ? ψs (t) .
C 0 s3/2
Its Fourier transform reads, after a change of variable ξ = sω,
fˆ(ω) +∞ √ fˆ(ω) +∞
Z Z
ds ds
b̂(ω) = sψ̂(sω) 3/2 = ψ̂(ξ) = fˆ(ω).
C 0 s C 0 s
Exercise 4.9. a) If f is Cp , then fˆ(p) (ω) = (iω)p fˆ(ω) and thus the inverse Fourier transform
formula gives the result.
b) On the set Ωρ = {z \ Im(z) > ρ}, one has
|(iω)p fˆ(ω)eizω | 6 |ω|p |fˆ(ω)|e−ρω ,
which is integrable. Using classical result of holomorphy under the sign , one sees that f (p) (z)
R
is holomorphic on Ωρ for all ρ > 0.

c) One has
Z +∞
1
f (p) (x + iy) = (iω)p fˆ(ω)eixω e−yω dω.
2π 0
A wavelet transform of f is written over the Fourier domain
√
Z
1
W f (u, s) = fˆ(ω) sψ̂(sω)eiωu dω
2π
so one sees that f (p) is indeed a Fourier transform if one sets
ψ̂(ω) = 1[0,+∞[ (ω)(iω)p e−ω .
Using the inverse Fourier transform, and the fact that ĥ(p) (ω) = (iω)p ĥ(ω), one has
Z +∞
1 1 1
ψ(t) = h(p) (t) where h(t) = e−ω eiωt = ,
2π 0 2π 1 − it
so
ip!
ψ(t) = .
2π(t + i)p
Exercise 4.10. For f (t) = cos(θ(t)) with θ(t) = a cos(bt), the width s of the window is small
enough if
s2 |θ00 (t)| = s2 ab2 | cos(bt)| 1
and thus s2 ab2 1.
For θ(t) = cos(θ1 (t)) + cos(θ2 (t)) with θ1 (t) = a cos(bt) and θ2 (t) = a cos(bt) + ct, there is
enough frequency resolution if
∆ω
|θ10 (t) − θ20 (t)| = |c| >
s
and enough spacial resolution if
s2 |θ10 (t)| = s2 |θ200 (t)| = s2 ab2 1.
This shows that one needs r

c
∆ω ' 1.
ab2
16
Exercise 4.15. One has,
Z Z
1
P f (u, ξ)du = |f (u)|2 du|fˆ(ω)|2 = |fˆ(ω)|2
||f ||2
Using Plancherel conservation of energy formula, one has

Z Z
1 1
P f (u, ξ)dξ = |f (u)|2 |fˆ(ω)|2 dω = |f (u)|2 .
2π ||f ||2
Since P f is not quadratic in f (it is of degree 4), one cannot apply Theorem 4.11.
Exercise 4.16. Example 4.20 shows that Pθ f is a spetrogram with a Gaussian window gµ if
θ(u, ξ) = gσ (u)gβ (ξ) = PV gµ (u, ξ) = gµ/√2 (u)g√2µ (ξ),

√
where we have used (4.125) for the last equality. If σβ = 1/2, then one can take µ = 2σ.
Otherwise, if σβ > 1/2, one decomposes θ = θ0 ? θ1 where
θ0 (u, ξ) = gµ/√2 (u)g√2µ (ξ) and θ1 (u, ξ) = gσ−µ/√2 (u)gβ−√2µ (ξ)

√ √
where µ is chosen so that σ − µ/ 2 > 0 and β − 2µ > 0. This shows that Pθ is a smoothing
with θ1 of Pθ0 which is a spectrogram and hence positive.
Exercise 4.17. Since {gn (t)}n∈N is an orthonormal basis,
+∞
X
gn∗ (t) gn (t0 ) = δ(t − t0 ) .
n=0
Indeed, for all f ∈ L2

Z +∞
X +∞
X
f (t) gn∗ (t) gn (t0 ) dt = hf, gn ign (t0 )
n=0 n=0
Z
= f (t0 ) = f (t) δ(t − t0 )dt .
It results that
+∞
X
gn∗ (u − τ /2) gn (u + τ /2) = δ(τ ) .
n=0
Since P V gn (u, ξ) is the Fourier transform of gn∗ (u − τ /2) gn (u + τ /2) with respect to τ , it results
that
+∞
X
PV gn (u, ξ) = 1 .
n=0

Z Z Z
A(t) = (ξ − φ0 (t))2 PV f (tξ)dξ = h(τ ) (ξ − φ0 (t))2 e−iξτ dξ dτ
where
h(τ ) = a(t + τ /2)a(t − τ /2)ei[φ(t+τ /2)−φ(t−τ /2)] .
17
In the sense of distribution, one has the following equality
Z
(ξ − φ0 (t))2 e−iξτ dξ = 2π[−δ 00 (τ ) + 2iφ0 (t)δ 0 (τ ) + φ0 (t)2 δ(τ )]
where δ, δ 0 , δ 00 are the Dirac distribution and its first and second derivatives. This shows that
A(t) = 2π[−h00 (0) + 2iφ0 (t)h0 (0) + φ0 (t)2 h(0)].
After computing the derivatives of h and simplification, one finds
A(t) = −π(a(t)a00 (t) − a0 (t)2 ),
which is the result.
4 Chapter 5
Exercise 5.1. One has a union of K orthogonal bases
K−1
[
D = {φp }06p<KN = {φKp+i }06p<N ,
i=0
so that D is tight frame of frame bound KN since

−1 −1
KN K−1 N
! K−1
X
2
X X √ X
|hf, φp i| = N |hf, φpK+i / N i|2 = N ||f ||2 = KN ||f ||2 .
p=0 i=0 p=0 i=0
Exercise 5.2. A change of variable t → Kt gives

Z 1 Z 1/K Z 1
hf, φp i = f (t)e2iπpt/K dt = K f (Kt)e2iπpt = fK (t)ep (t)dt,
0 0 0
where {ep (t) = e−2iπt }p is an orthonormal basis of L2 [0, 1]. Let us define fK (t) = f (Kt) for
t ∈ [0, 1/K] and fK (t) = 0 otherwise. For K > 1, we get
X X
|hf, φp i|2 = |hfK , ep i|2 = ||fK ||2 = K||f ||2
p p
So {φp }p is a tight frame of L2 [0, 1] of frame bound K.

Exercise 5.3. The operator ΦΦ∗ is bounded on H = Im(Φ) which is of finite dimension, so
B < +∞. If A = 0, then it exists f 6= 0 in H such that n |hf, φn i|2 = 0 so that hf, φn i = 0
P
for all n. Since {Φn }n is generator of H, one has f = 0 which is a contradiction.
X
tr(U1 U2 ) = U1 [i, j]U2 [j, i] = tr(U2 U1 ).
i,j
Exercise 5.5. One has Φf [p] = f ? h where h = δ[·] − δ[· − 1]. As the discrete Fourier basis
diagonalizes Φ, the frame bounds are
π
A = min |ĥ[ω]|2 = min 4 sin2 ω = 4 sin2 (π/N )
ω6=0 ω6=0 N
18
and B = 4 (if N is even). One has A/B → 0 when N → +∞, so the frame is unstable when N
is large.
Exercise 5.6. If one does not constrain ||φp ||, one can choose {δ0 /N, N δ1 , δ2 . . . , δN −1 }. If one
constrains ||φp || = 1, one can choose φp [n] = φ[n + p mod N ] so that Φf = f ∗ φ and thus one
should impose X
|φ̂[ω]|2 = N and |φ̂[ω]| > 0.
ω
p
This can be achieved by setting φ̂[ω] = 1/N for ω 6= 0 and φ̂[0] = N − (N − 1)/N 2 . This
implies
A = min |φ̂[ω]|2 = 1/N 2 → 0
ω
and
B = max |φ̂[ω]|2 = N − (N − 1)/N 2 → +∞.
ω
Exercise 5.7. If x ∈ Null(U ∗ ) and y = U z ∈ Im(U ), then
hx, yi = hx, U zi = hU ∗ x, zi = 0
so that Null(U ∗ ) = Im(U )⊥ .

X
Φm f [p] = hf, φm+p i = f
? φm [p] =⇒ Φ∗ Φ = Φ∗m Φm
m
where Φm is a convolution operator, whose eigenvalues are φ̂m [ω], and eigenvectors are the
discrete Fourier vectors. The eigenvalues of Φ∗ Φ are thus m |φ̂m [ω]|2 .
P
Exercise 5.9. Let
φk,p (t) = g(t − 2pπ/ω0 )eikω0 t so φ̂k,p (ω) = ĝ(ω − kω0 )eiω/ω0 2pπ .
√
Since {eiω/ω0 2pπ }p is an orthonormal basis of each interval [−ω0 /2, ω0 /2] + kω0 , {1/ 2π φ̂k,p }k,p
√
is an orthogonal basis of L2 (R). Since f 7→ 1/ 2π fˆ is an isometry, this prove that {φk,p }k,p is
also an orthogonal basis.
√ 2iπ
Exercise 5.10. a) Since {1/ Ke K nk }06k<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
K−1
X X 1
|g[n − mM ]|2 |f [n]|2 = |hf, √ gmnk i|2 .
n k=0
K
b) Summing over m leads to

N/M −1
X X X
2
K |f [n]| |g[n − mM ]|2 = |hf, gm,k i|2 ,
n m=0 k,m
and hence the result of Theorem 5.18.

Exercise 5.11. For m = 2, the filters are written, with the convention z = e−iω
√
ĥ(ω)/ 2 = cos3 (ω/2)e−iω/2 = P (z) = (z −1 + 3 + 3z + z 2 )/8
19
√
ĝ(ω)/ 2 = −i sin(ω/2)e−iω/2 = Q(z) = z/2 − 1/2.
Applying Bezout algorithm to the polynomials (1 + 3z + 3z 2 + z 3 )/8 and (z 2 − z)/2 leads to
17 7
(8z − 7z 2 ) × P (z) + ( z + 5z 2 + z 3 )Q(z) = 1
4 4
and the reconstruction filters are given by
ˆ (ω)/√2 = 8z −1 − 7z −2
h̃ and g̃ˆ(ω) =
17 −1 7
z + 5z −2 + z −3 .
4 4
Exercise 5.12. For m = 3, one has

√
ĥ(ω)/ 2 = cos4 (ω/2)4 = P (z) = (z/2 + z −1 /2)4
= (z 2 + 4z + 6 + 3z −1 + z −2 )/24
and one can choose

√
ĝ(ω)/ 2 = sin2 (ω/2) = Q(z) = (z − 2 + z −1 )/4.
If h̃ = h, then
2 − |ĥ(ω)|2 1 − cos4 (ω/2) √ √

g̃ˆ(ω) = ∗
=2√ 2
= 2[1 + cos2 (ω/2)] = 2Q̃(z)
ĝ (ω) 2 sin (ω/2)
√
where Q̃(z) = 1 + (z + 2 + z −1 )/4, so g̃/ 2 = [1/4, 3/2, 1/4].
Exercise 5.13. Using the same proof as Theorem 5.18 and exercise 5.10, with M = 1, K = N ,
one sees that
2iπ
gm,` [n] = g[n − m] exp `n
N m,`
is a tight frame of frame bound

X
N |g[n − m]|2 = N ||g||2 = N.
n
Theorem 5.5 implies that the dual frame is {gm,` /N }m,` .

The reproducing kernel reads

0 0
X
0 2iπ 0
K(m, m , `, ` ) = hgm,` , gm0 ,`0 i = g[n − m]g[n − m ] exp n(` − ` ) .
n
N
Exercise 5.14. As βη → 2π, the frame becomes less orthogonal (A/B → +∞), so the dual
vectors become more and more different from the primal ones, and hence the windows g and g̃
differ more and more.
1ˆ
∀ π ∈ [−π, π], x̂(ω) = f (ω/s)
s
for s = T /K, so supp(x̂) ⊂ [−π/K, π/K].
20
b) Since x − x̃ = W
J(h) = E(||x̃ ? h − x||2 ) = E(||x ? (h − δ) + W ? h||2 ).
Since W is a white noise of variance σ 2 independant of x
J(h) = E(||x ? (h − δ)||2 ) + E(||W ? h||2 ).
Since W is a white noise of variance σ 2 its power spectrum is σ 2 . Let R̂x (ω) be the power
spectrum of the stationary random vector x[n],
Z π
1
J(h) = (R̂x (ω)|1 − ĥ(ω)|2 + σ 2 |ĥ(ω)|2 ) dω .
2π −π
To minimize J(h) for each ω we minimize the value under the integral, which is obtained with
π
σ 2 R̂x (ω) σ2
Z
R̂x (ω) 1
ĥ(ω) = and J(h) = dω 6 .
σ 2 + R̂x (ω) 2π −π σ 2 + R̂x (ω) K
c) In this case x̃ = x ? hp + W so
J(h) = E(||x ? (h ? hp − δ)||2 ) + E(||W ? h||2 )

Z π
1
= (R̂x (ω)|1 − ĥ(ω) ĥp (ω)|2 + σ 2 |ĥ(ω)|2 ) dω .
2π −π
The minimum is obtained with

π
σ 2 R̂x (ω)
Z
R̂x (ω) ĥp (ω) 1
ĥ(ω) = and J(h) = dω.
σ 2 + R̂x (ω) |ĥp (ω)|2 2π −π σ 2 + R̂x (ω)|hp (ω)|2
Since ĥp (ω) = (1 − e−iω )−p ,

π/K
σ2 σ2 π 2p
Z
J(h) 6 |2 sin(ω/2)|2p dω 6 .
2π −π/K K K 2p (2p + 1)
For a fixed K > π, increasing p decreases the error.

Exercise 5.16. a) Each {φs (t − ns − ks/K)}n is an orthogonal basis of Us , the set of signals
with Fourier support included in [−π/s, π/s]. This proves that {φs−n/Ks }n is a tight frame with
frame bound K.
b) One has
X 1
P a[n] = a[n0 ]H[n, n0 ] where H[n, n0 ] = hφs (t − n/Ks), φs (t − n0 /Ks)i,
K
n0
and, using Plancherel formula,

1 0
H[n, n0 ] = hs1/2 1[−π/s,π/s] (ω)e−in/Ksω , s1/2 1[−π/s,π/s] (ω)e−in /Ksω i
2Kπ
Z π/s
s 0 sin(πn/K)
= ei(n−n )/Ksω dω = h[n − n0 ] where h[n] = .
2Kπ −π/s πn
21
c) Since ĥ = 1[−π/K,π/K] , one has
Im(P ) = {a \ h ? a = a} = {a \ supp(â) ⊂ [−π/K, π/K]} .
d) Theorem 3.5 proves that s−1/2 f ? φs (ns) = f (ns).

e) Let a[n] = f ? φs (ns0 ). For ω ∈ [−π, π], one has
1
â(ω) = F(f ? φs )(ω/s0 ) =⇒ supp(â) ⊂ [−π/K, π/K]
s0
where F is the Fourier transform. This shows that a ∈ Im(Φ) and hence P a = a.
One has
P Y [Kn] = a[Kn] + P (W )[Kn] =⇒ |P Y [Kn] − s−1/2 f (ns)|2 = |W ? h[Kn]|2 .
One then has

π
σ2 σ2
X Z
E(|W ? h[Kn]|2 ) = σ 2 |h[k]|2 = |ĥ(ω)|2 = →0 when K → +∞.
2π −π K
k
1
Exercise 5.17. a) Theorem 5.11 applied to the generator 2j/2 ψ(u/2j )
= φn (u) proves the result.
b) ψ has p vanishing moments if and only if ψ̂ (k) (0) = 0 for all k < p. Thus, since
X
|ψ̂(2j ω)|2 > A > 0,
j
one has that ψ̃ˆ(k) (0) = 0 and thus ψ̃ has also p vanishing moments.
c) One has for g = {gj (u)}j
XZ XZ
(PV g)j (u) = gj 0 (u0 )hψ2j ,u , ψ̃2j0 ,u0 idu0 = gj 0 (u0 ) · (ψj ? ψ̃j 0 )(u − u0 )du.
j0 j0
This shows that X

(PV g)j = ψj ? gj 0 ? ψ̃j 0 .
j0
Exercise 5.18. We prove the right hand side of the inequality

β β
X 1X
|ψ̂(2j ω)|2 = |ĝ(2j−1 ω)|2 |φ̂(2j−1 ω)|2
j=α
2 j=α
β
X
6B (1 − |ĥ(2j−1 ω)|2 )|φ̂(2j−1 ω)|2
j=α
β
X
6B (|φ̂(2j−1 ω)|2 − |φ̂(2j ω)|2 ) = B(|φ̂(2α−1 ω)|2 − |φ̂(2β ω)|2 ).
j=α
Since φ̂(0) = 1 and φ(ω) tends to zero as ω increases, letting α and β go to −∞ and +∞ proves
that j |ψ̂(2j ω)|2 6 B. A similar proof applies to the left hand side.
P
22
X
Zgn,k (u, ξ) = e2iπξ` 1[0,1] (u − ` − n)e2iπk(u−`) .
`
Since 1[0,1] (u − ` − n) is zero unless ` = −n, one obtains
Zgn,k (u, ξ) = e−2iπξn e2iπuk .
Since {e−2iπξn e2iπuk }n,k is an orthogonal basis of L2 [0, 1]2 , the Zak transform transforms an
orthogonal basis into an orthogonal basis, and is thus an unitary transforms.
b) One has
Z 1 XZ 1
Zf (u, ξ)dξ = e2iπξ` dξf (u − `) = f (u).
0 ` 0
Note that this formula is valid to recover f (u) for u ∈ [0, 1], but it can be extended to arbitrary
u ∈ R using the fact that
Zf (u + 1, ξ) = e2iπξ Z(u, ξ).
c) Let gn,k (t) = e2iπk g(t − n). Similarly to question a), one proves that
Zgn,k (u, ξ) = e−2iπξn e2iπuk (Zg)(u, ξ).
One has, using the fact that the Zak transform is unitary, and then Plancherel formula for Fourier
series of functions defined on [0, 1]2 ,
X X X Z Z 2
−2iπξn 2iπuk ∗
|hf, gn,k i|2 = |hZf, Zgn,k i|2 =

e e (Zf )(u, ξ)(Zg) (u, ξ)dudξ

n,k n,k n,k
ZZ
= |(Zf )(u, ξ)|2 |(Zg)(u, ξ)|2 dudξ.
This shows that if

∀ (u, ξ), A 6 |(Zg)(u, ξ)|2 6 B,
then, using the fact that Z is an isometry,
X
A||f ||2 = A||Zf ||2 6 |hf, gn,k i|2 6 B||Zf ||2 = B||f ||2 ,
n,k
which proves that {gn,k }n,k is a frame with bounds A and B.

d) Denoting Φf [n, k] = hf, gn,k i the frame operator, we have proved that
ZZ
2
||Φf || = |(Zf )(u, ξ)|2 |(Zg)(u, ξ)|2 dudξ,
so that
Z(Φ∗ Φf )(u, ξ) = |(Zg)(u, ξ)|2 (Zf )(u, ξ)
and hence
(Zf )(u, ξ)
Z((Φ∗ Φ)−1 f )(u, ξ) = .
|(Zg)(u, ξ)|2
The dual frame {g̃n,k }n,k is defined as
g̃n,k = (Φ∗ Φ)−1 gn,k
23
and hence
(Zgn,k )(u, ξ) (Zg)(u, ξ)e−2iπξn e2iπuk
Z g̃n,k (u, ξ) = =
|(Zg)(u, ξ)|2 |(Zg)(u, ξ)|2
−2iπξn 2iπuk
e e
= ∗
= (Zhn,k )(u, ξ)
(Zg) (u, ξ)
where we have defined
hn,k (t) = e2iπk h(t − n) where (Zh)(u, ξ) = (Zg)∗ (u, ξ)−1 .
Exercise 5.21. We use the change of variables
x = ρ cos(α) − u sin(α) and y = ρ sin(α) + u cos(α).
For k + ` < p, one can expand the monomial P (x, y) = xk y ` as a polynomial in u of degree less
than p X
P (x, y) = A(ρ)ut .
t<p
This shows that

XZ Z
t
hP, ψi = A(ρ) u ψ(ρ cos(α) − u sin(α), ρ sin(α) + u cos(α))du dρ = 0.
t<p

X A C
nk n∗k = ,
C B
k
where, denoting αk = 2kπ/K,

X X X
A= cos2 (αk ), B= sin2 (αk ) and C = cos(αk ) sin(αk ).
k k k
One has C = 0 and

K X K
A=B= + cos(2αk ) = ,
2 2
k
∗ K
P
so k nk nk
= 2 Id2 , which proves that {nk }k is a tight frame with frame bound K/2.
b) One has
∂θ ∂θ
ψ k = h[ , ], nk i =⇒ ψ̂ k (ω) = θ̂(ω)hi[ωx , ωy ], nk i.
∂x ∂y
This shows that
K X
X X X
|ψ̂ k (2j ω)|2 = |θ̂(ω)|2 22j |h[ωx , ωy ], nk i|2
k=0 j j k
and hence the result since

X K
|h[ωx , ωy ], nk i|2 = ||ω||2 .
2
k
24
5 Chapter 6
Exercise 6.1. a) It is an immediate consequence of Definition 6.1 of Lipschitz and uniform
Lipschitz regularity.
b) For t 6= 0 f is C 1 so it is 1-Lipschitz, and |f (t)| 6 |t| so f is also 1-Lipschitz at 0.
If f is uniformly Lipschitz α then there exists K > 0 such that for all (u, v) ∈ [−1, 1]2 ,
|f (u) − f (v)| 6 K |u − v|α . For tn = (n + 1/2)−1 π −1 we have f (tn ) = (−1)n tn . So

|f (tn ) − f (tn−1 )| = tn + tn−1 = π −1 (n + 1/2)−1 + (n − 1/2)−1 ∼ 2π −1 n−1 .
Since tn − tn−1 = π −1 (n + 1/2)−1 (n − 1/2)−1 ∼ π −1 n−2 it results that |f (tn ) − f (tn−1 )| ∼

|tn − tn−1 |1/2 and hence that α > 1/2.
We now prove that f is indeed Lipschitz 1/2. If u and v have same sign and |1/u − 1/v| > 1
then there exists C > 0 with
|f (u) − f (v)| 6 |f (tn ) − f (tn−1 )| 6 C |tn − tn−1 |1/2 6 C |u − v|1/2 .
If u and v of same sign with |1/u − 1/v| 6 1 and |u| > |v|. It result that |u − v| 6 u2 , and
|f (u) − f (v)| 6 |u − v| + |v|(| sin u−1 | − | sin v −1 |) 6 |u − v| + |v|(v −1 − u−1 |.
Since |u − v| 6 1
|u − v|
|f (u) − f (v)| 6 |u − v|1/2 + 6 2|u − v|1/2 .
|u|
If u and v have different signs, and |f (u)| > |f (v)|, since |u − v| 6 2 it results that
√
|f (u) − f (v)| 6 2|f (u)| 6 2|u| 6 2|u − v| 6 2 2|u − v|1/2 .
It results that for any (u, v) ∈ [−1, 1]2 there exists C 0 > 0 with
|f (u) − f (v)| 6 C 0 |u − v|1/2
and hence that f is uniformly Lipschitz 1/2 on [−1, 1].

Exercise 6.2. a) The definition of f being α Lipschitz can be written
f (x) = f (u) + a(u)(x − u) + K(x, u)(x − u)α (2)
with |K(x, u)| uniformly bounded in both x and u, so
[f (x) − f (u)]/(x − u) = a(u) + K(x, u)(x − u)α−1
which implies that f has derivatives at u with a(u) = f 0 (u), so (2) is re-written
f (x) = f (u) + f 0 (u)(x − u) + K(x, u)(x − u)α (3)
Then summing (2) for x = a + d, x = a and x = a − d shows that
(f (a + d) − 2f (a) + f (a − d))/d = O(dα−1 )
then applying (3) for x = a and both u = a − d and u = a + d shows that
|f 0 (a − d) − f 0 (a + d)|/d = O(dα−1 )
25
this shows that f 0 is Lipschitz α − 1 and that this is uniform with the same bound.
b) If f (t) = t2 cos(1/t) then f 0 (t) = 2t cos(1/t) − sin(1/t). One has |f (t) − f (0)| 6 t2 so that f
is 2-Lipschitz at 0, but sin(1/t) is discontinuous at 0 of f 0 is not Lipschitz 1 at 0.
Exercise 6.3. One can take a hat function that is zero outsize [−1, 1], that is linear over [−1, 0]
and [0, 1], with f (0) = 1. It is 1-Lipschitz. We verify that |fˆ(ω)| = | sin(ω/2)|2 |ω|−2 so (6.4) is
not satisfied.
Exercise 6.4. a) One has, for f (t) = cos(ω0 t), and using Plancherel formula

t−u
Z
1 1
W f (u, s) = √ √ ψ f (t) dt
s s s
√
Z
1
= s ψ̂(sω)e−iuω (δω0 (ω) + δ−ω0 (ω))dω
2
√
= sψ̂(sω0 ) cos(uω0 ),
because ψ̂(−ω) = ψ̂(ω) because ψ(t) is symmetric.

b) One has,
∂W f (u, s) kπ
= 0 ⇐⇒ u = uk = .
∂u ω0
If ψ has p vanishing moments then |ψ̂(ω)| = O(|ω|p ) so
√
|W f (uk , s)| 6 s |ψ(sω0 )| = O(sp+1/2 ).
Exercise 6.5. For f (t) = |t|α , one has, after the change of variable t → t/s,
Z Z
1
W f (u, s) = √ |t|α ψ(t/s − u/s)dt = sα+1/2 |t|α ψ(t − u/s)dt = W f (u/s, 1).
s
If ψ(t) is an antisymmetric wavelet, ψ(−t) = ψ(t), since f (t) is even the wavelet transfrom is
antisymmetric and hence W f (0, s) = 0 for all s. In this case we thus can not derive the Lipschitz
regularity at t = 0 from W f (0, s).
Exercise 6.6. One has |f (t)| 6 |t|α so f is α-Lipschitz. One has for t > 0
f 0 (t) = αtα−1 sin(1/tβ ) + βtα−β−1 cos(1/tβ ),
so f 0 is α − β − 1-Lipschitz.
The signal can be written f (t) = a(t) cos θ(t) with a(t) = |t|α and θ(t) = |t|−β . The instanta-
neous frequency is θ0 (t) = β |t|−β−1 . For such amplitudes and instantaneous frequencies one can
apply the wavelet ridge calculations. Equation (4.109) shows that the larges wavelet coefficients
are along ridge curves (u, s(u)) defined by
s(u) = ηθ0 (u)−1 = η β −1 |u|β+1
where η is the center wavelet frequency.

Since f (t) is Lipschitz α at 0, we derive from (6.20) in Theorem 6.4 that equation (6.21) is
satisfied for α0 = α. One can not decrease α0 because at the ridge locations, (4.108) proves that
|W f (u, s)| = O(s1/2 a(u)) = O(s1/2 |u|α ).
26
Exercise 6.7. a) We proved in exercice 6.5 that if f (t) = |t|α then W f (u, s) = sα+1/2 W f (u/s, 1).
It results that the complex phase Φ(u, s) of W f (u, s) satisfies Φ(u, s) = Φ(u/s, 1). The lines of
constant phase are thus along curves (u, s) which satisfy u/s = cst, and which converge to 0
when s goes to 0.
b) A modulus maxima point u0 is an extrema of W f (u, s) = f ? ψ̃s (u) with ψ̃s (t) = s−1/2 ψ(−st)
and thus satisfies ∂W f (u0 , s)/∂u = 0. But
∂W f (u, s)
= −sW 1 f (u, s) = −sf ? ψ̃s1 (u)
∂u
with ψ 1 (t) = ψ 0 (t). The modulus maxima computed with ψ are thus zeros of a wavelet transform
computed with ψ 0 . Let us now consider the analytic wavelet transform W a f (u, s) = f ? ψ̃sa (u)
computed with the analytic complex wavelet
ψ a (t) = ψ 0 (t) + iH(ψ 0 )(t)
where H(ψ 0 ) is the Hilbert transform of ψ 0 . The modulus maxima of W f (u, s) correspond to
points (u, s) where W a f (u, s) has a real part equal to zero and hence a phase equal to π. The
modulus maxima points thus correspond to points of constant phase (equal to π) on this analytic
complex wavelet tranform.
+∞
√
u−t
Z Z
1
W f (u, s) = √ ψ dt = s ψ(t)dt
s 0 s −u/s
so that
∂W f
(u, s) = 0 ⇐⇒ ψ(−u/s) = 0.
∂u
Since ψ is continuous and orthogonal to polynomials of degree less than p − 1, it has at least p
Q exist less than p − 1 zero crossings {tk }k
zeros. Indeed, if this were not the case, there would
where ψ changes of sign at tk . Denoting P (t) = k (t − tk ) which is a degree less than p, P ψ is
non zero and of constant sign, so hf, P i =
6 0, which is a contradiction.
Rt
Exercise 6.10. If f (t) = 0 dµ∞ (t) is a Cantor devil’s staircase then its singularities are located
on the support of the Cantor positive measure dµ∞ . Let t0 be on the support of dµ∞ . By
construction, explained page 245, for any p > 0 there exists an integer q such that t0 = q3−p ,
and either (q + 1)3−p or (q − 1)3−p is also on the support of dµ∞ . Suppose that it is the first
case (the second one is treated similarly). The constrution shows that
Z q3−p Z (q+1/9)3−p Z (q+2)3−p
dµ∞ (t) = 0 , dµ∞ (t) > 0 , dµ∞ (t) = 0 . (4)
(q−1)3−p q3−p (q+1)3−p
If ψ = −θ0 then according to (6.32)

Z
W f (u, s) = s(f 0 ? θ̃s )(u) = s s−1/2 θ(s−1 (t − u)) dµ∞ (t) . (5)
Let K be such that θ has a support included in [−K, K]. To simplify explanations we suppose that
θ > 0 for |t| < K. If s−1
p = K3
p+1
then θ(s−1
p (t − u)) has a support included in [−3
−p−1 −p−1
,3 ].
It results from (4) and (5) that
W f ((q − 1/2)3−p , sp ) = 0 , W f (q3−p , sp ) > 0 , W f ((q + 3/2)3−p , sp ) = 0 .
27
So necessarily, |W f (u, sp )| has at least one local maxima at position up in the interval [(q −
1/2)3−p , (q + 3/2)3−p ] and since t0 = q3−p
lim up = t0 .
p→+∞
Exercise 6.14. a) One has ψ̂(ω) = ||ω||2 θ̂(ω) so that

+∞ +∞
2j+1
||ω||2
X X
F (ω) = |ψ(2j ω)|2 = 24j e−2 .
j=−∞ j=−∞
One has X X
|ψ̂(2j ω)|2 6 24j = C < +∞.
j60 j60
Using the fact that 2j+1 6 1 + (2j + 1) log(2), one has

X X 2 X 2
|ψ̂(2j ω)|2 6 e4j log(2)−(1+(2j+1) log(2))||ω|| 6 C 0 ej(4 log(2)−2||ω|| log(2)) .
j>0 j>0 j<0
This shows that for 0 6 ||ω||2 6 α < 2, F (ω) 6 B and also

X 2j+1
F (ω) > 24j e−2 α
= A > 0.
j
This proves that for 0 6 ||ω||2 6 α, A 6 F (ω) 6 B. Using the fact that F (2ω) = F (ω), this
bound is extended to all ω ∈ R2 .
b) For ψ = ∆θ, where ∆ is the Laplacian, one has
W f (u, 2j ) = 0 ⇐⇒ ∆(f ? θj ) = 0,
so this is a multiscale crossing of the Laplacian edge detector.

Suppose that the image can locally be approximated by a straight-edge profile
f (x1 , x2 ) = ρ(x1 cos α − x2 sin α) (6)
where ρ(t) is a monotonous function with an inflection point at t = 0. A direct computation

shows that if ψ is the Laplacian of a Gaussian θ then W f (u, s) = 0 when u is an inflection point
of
f ? θs (x) = ρs (x1 cos α − x2 sin α)
which is located along a straight edge curve of angle α, x1 cos α − x2 sin α = t0 with ρ00s (t0 ) = 0.
It thus corresponds to the position of modulus maxima point computed with the two partial
derivative wavelets ψ 1 and ψ 2 .
Suppose that f has a curved edge, which can be modeled with an angle α which has a slow
variation as a function of (x1 , x2 ). Then the location of the zero-crossings of W f (u, s) are not
identical to the modulus maxima of the first derivative wavelets because of second order terms
that are not identical in both wavelet transforms. The zero-crossings are not exactly located at
the inflection points of f ? θs (x) but close to these points if the curvature is small.
28
√
Exercise 6.15. One has, denoting ψs (t) = ψ(t/s)/ s,
E(W B(u1 , s)W B(u2 , s))

= E([BH ? ψ̄s (u1 )][BH ? ψ̄s (u2 )])
ZZ
= E(BH (t)BH (t0 ))ψ̄s (u1 − t)ψ̄s (u2 − t0 )dtdt0
ZZ
= σ2 (|t|2H + |t0 |2H − |t − t0 |2H )ψ̄s (u1 − t)ψ̄s (u2 − t0 )dtdt0
ZZ
= −σ 2
|t − t0 |2H ψ̄s (u1 − t)ψ̄s (u2 − t0 )dtdt0
R
where we have use the fact that ψs = 0 to derive the last equality. Performing the successive
changes of variables t − t0 → t and u2 − t0 → t0 , one gets
Z
E(W B(u1 , s)W B(u2 , s)) = −σ 2 |t|2H ψ̄s ? ψs (u1 − u2 − t)dt
which gives the result after the change of variable t0 /s → t0 in the convolution.
6 Chapter 7
Y ω
φ̂(ω) = ĥ j+1 .
2
j>0
For each j > 0 and ` ∈ Z, h(ω/2 ) has a zero of order p at location 2π(2j (2` + 1)). Remarking
j+1
that any k ∈ Z∗ can be written k = 2j (2` + 1) with j > 0 and ` ∈ Z, one sees that φ̂(ω) has a
zero of order p at each ω = 2kπ, k 6= 0.
b) One has, using Theorem 3.1
!
X
q inω 1 dq X inω
A(ω) = n φ(n)e = q q φ(n)e
n
i dω n
!
1 dq X 1 X (q)
= q q φ̂(ω + 2nπ) = q φ̂ (ω + 2nπ)
i dω n
i n
One thus have, for ω = 0,

Z
X 1 X (q) 1 (q)
nq φ(n) = A(0) = φ̂ (2nπ) = φ̂ (0) = tq φ(t)dt,
n
iq n iq
since φ̂(q) (2nπ) = 0 for n 6= 0.

P
Exercise 7.2. Let A(t) = n φ(t − n). Its Fourier transform satisfies, using Theorem 2.4, in
the sense of distributions
X X X
Â(ω) = φ̂(ω) e−inω = φ̂(ω)2π δ(ω − 2kπ) = 2π φ(2kπ)δ(ω − 2kπ).
n n n
Using Exercise 7.1, one has that φ̂(2kπ) = 0 for k 6= 0, and φ̂(0) = 1 so that Â(ω) = 2πδ(ω) and
hence A(t) = 1.
29
Exercise 7.3. We choose m to be even so that φ̂m is a symmetric function
X 1
φ̂m (ω)−2 = 1 + .
(1 + 2kπ/ω)2m+2
k6=0
For 0 < ω 6 a < π, one has |2nπ/ω| > |n|(1 + ε) with 0 < ε < 1 so that
X 1
|φ̂m (ω)−2 | 6 1 + 2 .
n>1
(n(2 + ε) − 1)(2m+2)
We split the sum into 1 6 n 6 1/ε and 1/ε < n. For the first part, the sum is finite and one has
1/ε
X 1
−→ 0
n=1
(n(2 + ε) − 1)2m+2
when m −→ +∞. For the second part, one has

Z +∞
X 1 X 1 1
6 6 dx −→ 0
(n(2 + ε) − 1)2m+2 (2n)2m+2 1/ε (2x)2m+2
n>1/ε n>1/ε
when m −→ +∞. This shows that φ̂m (ω) −→ 1 for m ∈ (−π, π). A similar derivation shows
φ̂m (ω) −→ 0 for |ω| > π, so that φ̂m → φ̂ almost everywhere. Using dominated convergence and
Plancherel formula, this shows that ||φm − φ|| → 0.
Exercise 7.4. a) If K = 2, then Supp(φ) ⊂ [0, 1] and
φ(t) = m[0]φ(2t) + m[1]φ(2t − 1),
so that
φ(0.ε0 ε1 · · · εi ) = m[0]φ(ε0 .ε1 · · · εi ) + m[1]φ((ε0 − 1).ε1 · · · εi ).
If ε0 = 1, then φ(ε0 .ε1 · · · εi ) = 0 and if ε0 = 0 then φ((ε0 − 1).ε1 · · · εi ) = 0 so that one has
φ(0.ε0 ε1 · · · εi ) = m[ε0 ]φ(0.ε1 · · · εi )
and by recursion one obtains
φ(0.ε0 ε1 · · · εi ) = m[ε0 ] · · · m[εi ]φ(0).
b) If t = 0.ε0 ε1 . . . is a dyadic number, then there exists some K such that εk = 0 for k > K.
For m[0] > 1, if φ(0) 6= 0, this implies that φ(t) is infinite.
c)
Exercise 7.5. a) The recursion formula is written over the Fourier domain as
1
φ̂k+1 (ω) = √ ĥ(ω/2)φ̂k (ω/2)
2
so that, using Plancherel formula
Z
1
ak+1 [n] = |ĥ(ξ/2)|2 |φ̂k (ξ/2)|2 einξ dξ.
4π
30
Using the Poisson summation formula, Theorem 2.4, one has the following equality, in the sense
of distributions
Z !
1 2 2
X
−in(ω−ξ)
âk+1 (ω) = |ĥ(ξ/2)| |φ̂k (ξ/2)| e dξ
4π n
1X
= |ĥ(ω/2 − nπ)|2 |φ̂k (ω/2 − nπ)|2
2 n
b) If φk → φ, then ak → a where a[n] = hφ(t), φ(t − n)i. If |ĥ(ω)| is bounded, then P is a

bounded linear operator on L2 (R). Hence P ak → P a and thus P a = a, which means that a is
an eigenvector of P for the eigenvalue 1.
One has
k
Y
φ̂k (ω) = 2−1/2 ĥ(2−p ω)φ̂0 (ω/2k )
p=1
k
Since φ̂0 (ω/2 ) → 1 when k → +∞ and φ̂k converges to φ̂, one has
Y
φ̂(ω) = 2−1/2 ĥ(2−p ω).
p>1
Exercise 7.6. a) Since f ∈ VL , one has

X
f (x) = hf, φL,k iφL,k
k
X
b[n] = aL [k]2−L/2 φ(k − n)
k
so that b = 2−L/2 aL ? φd where φd [n] = φ(n).

b) Using Theorem 3.1, one has
X
φ̂d (ω) = φ̂(ω + 2kπ).
k
The filter φ−1

d whose Fourier transform is 1/φ̂d (ω) is stable if φ̂d (ω) > A > 0.
d) If {φ̃L,n }n is an interpolation basis of VL for the points 2L n, then
X X
f= b[n]φ̃L,n = aL [n]φL,n .
n n
The change from b to aL is stable if and only if the basis {φ̃L,n }n is stable.
Exercise 7.7. a) A computation similar to the proof of Theorem 7.11 shows that the recon-
struction property is equivalent to having
ˆ (ω) + ĝ(ω + π)g̃ˆ(ω + π) = 0
ĥ(ω + π)h̃
and
ˆ (ω) + ĝ(ω)g̃ˆ(ω) = 2e−ilω .
ĥ(ω)h̃
31
One verifies that for a given h, the proposed filters satisfy the first condition, and that the second
condition is equivalent to
ĥ2 (ω) − ĥ2 (ω + π) = 2e−ilω .
b) Denoting z = e−iω , this last condition is re-written as
X X
( h[k]z k )2 − ( (−1)k h[k]z k )2 = 2z l .
k k
One verifies that in the left hand side of this equation, all terms of even degrees cancel out, so
that ` is necessarily odd.
√
c) For the Haar filter, 2ĥ(ω) = 1 + e−iω , so that
ĥ2 (ω) − ĥ2 (ω + π) = (1 + z)2 /2 − (1 − z)2 /2 = 2z
which corresponds to the quadrature mirror filter condition with ` = 1.

Exercise 7.8. A Daubechies wavelet ψj,n with p vanishing moments has a support of size
2j (2p − 1). One thus concentrates on the scales 2j such that 2j (2p − 1) < mink |τk − τk−1 | such
that each wavelet intersects only one discontinuity. Since the wavelets are translated by 2j with
support of size 2j (2p − 1), each singularity generates exactly 2p − 1 non zero coefficients. One
should thus choose p = q + 1. If p < q + 1, then there are 2−j coefficients at each scale, that can
be all non zero.
Exercise 7.9. a) Denoting u = 1[0,+∞) and v = 1(1,+∞) , one has θ1 = u − v. Denoting f (k) the
convolution k times of f with itself, one has
m+1
X m + 1 (k) (m−k)
θm = (u − v)(m+1) = (−1)k u v .
k
k=0
One has u(2) = [t]+ , u(3) = ([t]+ )2 /2, and more generally, by induction, one proves that
([t]+ )k−1 ([t − k]+ )k−1

u(k) (t) = and v (k) (t) = .
(k − 1)! (k − 1)!
One then verifies that
1
u(i) ? v (j) (t) = ([t − j]+ )i+j+1 ,
(i + j + 1)!
so that
m+1
1 X m+1
θm (t) = (−1)k ([t − j]+ )m .
m! k
k=0
b) One has, using the Fourier convolution theorem
|θm (ω)| = |θ0 (ω)|m+1 = |2 sin(ω/2)/ω|m+1
One has using (7.9)

X
Bm = max |θ̂m (ω + 2kπ)|2
ω
k
X 1
= max 2m | sin(ω/2)|2m+2 .
ω |ω + 2kπ|2m+2
k
32
By comparison with an integral, one obtains
Z +∞
X 1 dx
Bm > | sin(ω/2)| > | sin(ω/2)|
|(ω/(2π) + k)π|2m+2 ω/(2π) (πx)2m+2
k>0
(2m + 1)/π
= | sin(ω/2)| −→ +∞.
(ω/2)2m+1
when m → +∞ as long as 2 < ω < 2π.

Exercise 7.10. This proof is quite technical, and we detail a proof inspired by “Unimodular
Wavelets for L2 and the Hardy Space of H 2 ” by Young-Hwa Ha, Hyeonbae Kang, Jungseob Lee,
and Jin Keun Seo.
If {ψj,n }j,n is an orthogonal basis of L2 (R), then for all f ∈ L2 (R), using Parseval formula
1 X ˆ
fˆ = hf , ψ̂j,n iψ̂j,n .
2π j,n
This corresponds to, for all ξ,
2j X
Z
j
fˆ(ξ) = fˆ(ω)ψ̂ ∗ (2j ω)ψ(2j ξ)e−ik2 (ξ−ω) dω.
2π j,n
One has, using Poisson formula, Theorem 2.4, in the sense of distribution with respect to the
variable ω, the equality
X j 2π X
e−ik2 (ξ−ω)
= δ(ω − ξ + 2−j 2kπ),
n
2j
k
and thus X
fˆ(ξ) = fˆ(ξ − 2−j 2kπ)ψ̂ ∗ (2j ξ − 2kπ)ψ̂(2j ξ)
j,n
and hence X
fˆ(ξ)(1 − θ(ξ)) = fˆ(ξ − 2−j 2kπ)ψ̂ ∗ (2j ξ − 2kπ)ψ̂(2j ξ)
j,k6=0
where X
θ(ξ) = |ψ̂(2j ξ)|2 .
j
For ξ0 > 0 (the same derivation applies for ξ0 < 0) and for some ε > 0 small enough, we use
fˆ = 1I with I = [ξ0 − επ, ξ0 + επ], and integrate with respect to ξ to obtain
Z X Z
|θ − 1| 6 |ψ̂(2j ξ)| |ψ̂(2j ξ − 2kπ)|dξ,
I |k2−j |<ε I
where we have used the fact that fˆ(ξ − 2−j 2kπ) 6= 0 on I only if |k2−j | < ε. Using Cauchy-
Schwartz, and a change of variable 2j ξ → ξ one obtains
 
Z X X
|θ − 1| 6 2−j Aj (ξ)  Bj,k (ξ) ,
I 2−j <ε |k|<ε2j
33
where Z Z
2 2 2
Aj (ξ) = |ψ̂(ξ)| dξ and Bj,k (ξ) = |ψ̂(ξ + 2kπ)|2 dξ.
2j I 2j I
We note that, if |k2−j | < ε, then all the intervals 2j I + 2kπ are included in 2j [ξ0 − 3επ, ξ0 + 3επ],
so that Z 2j (ξ0 +3επ)
X
j
Bj,k (ξ) 6 2ε2 |ψ̂(ξ)|2 dξ.
|k|<ε2j 2j (ξ0 −3επ)
If fallows that Z X Z 2j (ξ0 +3επ)

|θ − 1| 6 2ε |ψ̂(ξ)|2 dξ.
I 2−j <ε 2j (ξ 0 −3επ)
If ε is small enough with respect to ξ0 , ε < ξ0 /(12π), then all the intervals 2j [ξ0 − 3επ, ξ0 + 3επ]
are disjoint and thus
1 +∞
Z Z
1
|θ − 1| 6 |ψ̂(ξ)|2 dξ.
|I| I π (ξ0 −3επ)/ε
If ψ̂ is continuous, then the left part of this inequality tends to θ(ξ0 ) − 1 whereas the right part
tends to 0 when ε → 0. This shows that θ(ξ0 ) = 0 for all ξ0 6= 0. When ψ̂ ∈ L2 (R), this results
holds only for almost all points ξ0 6= 0 that are Lebesgue points of θ̂.
To show that the condition X
|ψ̂(2j ω)|2 = 1 (7)
j
is not sufficient, let us note that the requirement that {ψ(t − n)}n is an orthogonal system is
written X
|ψ̂(ω + 2kπ)|2 = 1. (8)
k
We note that if ψ̂(ω) satisfies (7) and (8), then ψ̂(ω/2) still satisfies (7) but not (8) anymore,
take for instance
ψ̂(ω) = 1[−2π,−π]∪[π,2π] .
Exercise 7.11. If ψ̂ = 1I , is an indicator function, ψj,k is an orthogonal basis if it satisfies

X X
|ψ̂(2j ω)|2 = 1 and |ψ̂(ξ + 2kπ)|2 = 1.
j k
This means that one has the disjoint unions

[ [
R= 2j I = (I + 2kπ).
j k
One verifies that this is the case for the set
I/ω = [−32, −28] ∪ [−7, −4] ∪ [4, 7] ∪ [28, 32].
where in the following we denote ω = π/7. The base approximation space is
V0 = Span(ψj,n )j<0,n∈Z .
34
A function φ ∈ V0 thus satisfies
[
Supp(φ̂) ⊂ (2j I) ⊂ J
j<0
where
J/ω = [−16, −14] ∪ [−8, −7] ∪ [−4, 4] ∪ [7, 8] ∪ [14, 16].
If φ ∈ V0 , then the function X
θ(ω) = |φ̂(ω + 2kπ)|2 ,
k
S
is supported in k (J + 2kπ), and one verifies that
!
[
(J + 2kπ) ∩ [4ω, 6ω] = ∅
k
which implies that θ(ω) = 0 on [4ω, 6ω]. This means that {φ(t − n)}n cannot be an orthogonal
basis of V0 because otherwise θ(ω) = 1 for all ω.
Exercise 7.12. The Coiflet condition is written over the Fourier domain as
dk φ(ω)
∀ 0 < k < p, (0) = 0.
dω k
One has √
2φ̂(2ω) = ĥ(ω)φ̂(ω),
and taking the k th derivative of this expression at ω = 0 leads to the equivalent condition that
dk h(ω)
∀ 0 < k < p, (0) = 0,
dω k
which corresponds to the following conditions on the coefficients of h
X
nk h[n] = 0.
n
Exercise 7.13. Using the same derivation as in Exercise 7.12, the discrete signal ψj has p
vanishing moments if and only if
dk
∀ 0 6 k < p, ψ̂j (0) = 0.
dω k
Taking k derivative of
j−L−2
Y
ψ̂j (ω) = ĝ(2j−L−1 ω)H(ω) where H(ω) = ĥ(2p ω),
p=0
shows that this is indeed the case if

dk dk
∀ 0 6 k < p, ĝ(0) = ĥ(π) = 0
dω k dω k
which is equivalent to the continuous wavelet ψ having p vanishing moments, see Theorem 7.4.
35
Exercise 7.14. a) Since ψ has p > 1 vanishing moments, one can decompose
ĥ(ω) = (eiω + 1)P (e−iω )/2
where P is a polynomial (see (7.91)), so that
ĥ1 (ω) = 2ĥ(ω)/(eiω + 1) = P (eiω)
is a polynomial. It is obvious that 2(eiω − 1)ĝ(ω) is a polynomial in e±iω .

b) One has
ĝ(ω/2) Y ĥ(ω/2p )
ψ̂(ω) = √ √ .
2 p>2 2
One has the following equality
Y 1 + eiω/2j
iω = (eiω − 1) ,
2
p>1
and since the Fourier transform of ψ 0 satisfies ψ̂ 0 (ω) = iω ψ̂(ω), one can write
ĝ1 (ω/2) Y ĥ1 (ω/2p )

ψ̂ 0 (ω) = √ √ ,
2 p>2 2
with
2ĥ(ω)
ĥ1 (ω) = and ĝ1 (ω) = 2(eiω − 1)ĝ(ω).
eiω + 1
c) The derivative coefficients are obtained by replacing (h, g) by (h1 , g1 ) in the pyramid algorithm.
Exercise 7.15. a) Denoting ψar = Re(ψa ), and using the result of Exercise 2.4, one has
2ψ̂ar (ω) = ψ̂a (ω) + ψ̂a∗ (−ω) = ψ̂(ω)|ĥ(ω/4 − π/2)|2 + ψ̂(−ω)∗ |ĥ(−ω/4 − π/2)|2
Since ψ is a real wavelet, ψ̂(−ω)∗ = ψ̂(ω) and since h is also real, |h(−ω)| = |h(ω)|, which leads
to
2ψ̂ar (ω) = ψ̂(ω) |ĥ(ω/4 − π/2)|2 + |ĥ(ω/4 + π/2)|2 = 2ψ̂(ω)
where we have used the fact that |ĥ(ω)|2 + |ĥ(ω + π)|2 = 2 because h is a quadrature filter. This
proves that Re(ψa ) = ψ.
c) Denoting g1 as the finite filter such that ĝ1 (ω) = ĥ(ω)|ĥ(ω − π/2)|2 , one has to switch from
the ordinary computation of the detail coefficients
dj+1 = (aj ? ḡ) ↓ 2
where a ↓ 2 is the sub-sampling by 2 operator, to the following computation

dj+1 = [(aj−1 ? ḡ1 ) ↓ 2] ? ḡ ↓ 2.
d) One can compute the coefficients aj [n] = hf, φj,n1 (x1 )φj,n2 (x2 )i from aj−1 by applying fil-
tering by h̄[n1 ]h̄[n2 ] and subsampling along each direction. Then the detail coefficients d1j [n] =
1
hf, ψj,n i are computed by filtering/sub-sampling aj−2 two times by h̄ along the x2 direction and
36
then by filtering/sub-sampling by ḡ1 and then by ḡ. Similar computations allow one to obtain
the other details coefficients dkj [n] = hf, ψj,n
k
i for k = 2, 3, 4.
e) We denote by
ψ [1] (x) = ψ(x1 )φ(x1 ), ψ [2] (x) = φ(x1 )ψ(x2 ) and ψ [3] (x) = ψ(x1 )ψ(x2 )
the classical 2D wavelets. One has for a real signal f ,

[1] 1 [2] 2
hf, ψj,n i = Re(hf, ψj,n i), hf, ψj,n i = Re(hf, ψj,n i),
and
[3] 3 4
hf, ψj,n i = Re(hf, ψj,n i) + Re(hf, ψj,n i).
This implies that
1 [1] [2]
|hf, ψj,n i|2 > |hf, ψj,n i|2 , 2
|hf, ψj,n i|2 > |hf, ψj,n i|2 ,
and
3 1 1 [3]
|hf, ψj,n i|2 + |hf, ψj,n
4
i|2 > 3
(hf, ψj,n 4
i + hf, ψj,n i)2 > |hf, ψj,n i|2 .
2 2
This shows that X
k 1 X [k]
|hf, ψj,n i|2 > |hf, ψj,n i| = ||f ||2 /2,
2
j,n,k j,n,k
and hence the frame is redundant.

The reverse inequality is more technical, see [108], Appendix A, for a proof. It uses the fact
that if a function ψ satisfies |ψ̂(ω)| = O((1 + ||ω|||s ) with s > 1/2 and |ψ̂(ω)| = O(||ω||α ) with
α > 0, then {ψj,n }j,n is a stable frame of its span. It is easy to show that the almost analytic
wavelets ψ̂ k (ω) satisfy these decay conditions.
Exercise 7.16. a) One notes that φ1 = 1[0,1) satisfies the scaling equation
φ1 (t) = φ1 (2t) + φ1 (2t − 1).
whereas φ1 (t) = 1[0,1) (t)(2t − 1) satisfies

1
φ2 (t) = (φ2 (2t) + φ2 (2t − 1) − φ1 (2t) − φ1 (2t − 1)) .
2
One has hφ1 , φ2 i = 0 so that {φ1 (t − n), φ2 (t − n)}n is an orthogonal basis of the function that
are affine on each interval [n, n + 1).
b) One needs to apply the Gram-Schmidt orthogonalization process to
{φ1 , φ2 , t2 1[0,1) (t), t3 1[0,1) (t)}
to obtain
{φ1 , φ2 , ψ1 , ψ2 }.
They correspond to the Legendre polynomials on [0, 1), and
{φ1 (t − n), φ2 (t − n), ψ1 (t − n), ψ2 (t − n)}n∈Z
is an orthogonal basis of the functions that are polynomials of degree 3 on each interval [n, n+1).
Since they are orthogonal to φ1 , φ2 , they are orthogonal to polynomial of degree 1 on their
support, and hence they have two vanishing moments (and ψ2 has 3 vanishing moments).
37
Z 1 XZ 1 XZ 1
αrepl (t)β repl (t)dt = α(t − 2k)β repl (t)dt + α(2k − t)β repl (t)dt
0 k 0 k 0
XZ −2k+1 XZ 2k
= α(t)β rep (t)dt + α(t)β rep (t)dt
k −2k k 2k−1
Z +∞
= α(t)β repl (t)dt
−∞
X X
= hα(t), β(t − 2k)i + hα(t), β(2k − t)i = 0.
k k
b) We treat the case the case where φ, φ̃ are symmetric about 1/2 and ψ, ψ̃ are ansi-symmetric
about 1/2, which implies ψ(1−t) = −ψ(t). Since the folded signals are 2-periodic, one only needs
to consider, at a scale 2j with j < 0, indexes 0 6 n < 2−j+1 . One thus has, if (j, n) 6= (j 0 , n0 ),
hψj,n (t), ψ̃j 0 ,n0 (t − 2k)i = hψj,n , ψ̃j 0 ,n0 +21−j0 k i = 0
hψj,n (t), ψ̃j 0 ,n0 (2k − t)i = −hψj,n , ψ̃j 0 ,n0 +(1−2k)2−j0 i = 0
which implies, using question a), that
repl
hψj,n , ψjrepl
0 ,n0 i = 0,
and similarly with the other inner products with scaling functions.
For a given function f ∈ L2 ([0, 1]), we extend it by zero outside [0, 1] and decompose it in
the wavelet basis of L2 (R)
X X repl
f= hf, ψj,n iψ̃j,n + hf, φJ,n iψ̃J,n
j6J,n n
so that for t ∈ [0, 1],

X repl repl
X
f (t) = f repl (t) = hf repl , ψj,n iψ̃j,n + hf repl , φrepl repl
J,n iφ̃J,n .
j6J,n n
Exercise 7.18. a) In the following, we denote z = e−iω , and denote
R(z)R(z −1 )
p̂(ω) = P (z) = .
Q(z)Q(z − 1)
pk z k neces-
P
The perfect reconstruction property reads P (z) + P (−z) = 2, so that P (z) = k∈Z
sarily satisfies p2k = 0 for k 6= 0. Hence one can decompose
C(z 2 ) C(z 2 ) + zD(z 2 ) R(z)R(z −1 )
P (z) = 1 + z = =
2
D(z ) 2
D(z ) Q(z)Q(z −1 )
where C and D are polynomials with no root in common. Thus by identification 2D(z 2 ) =
R(z)R(z −1 ) + R(−z)R(−z −1 ), which corresponds to
2|r̂(ω)|2
p̂(ω) =
|r̂(ω)|2 + |r̂(ω + π)|2
38
where r(ω) = R(z).
b) The constraint on r is rewritten R(z −1 ) = z k−1 R(z), so that
R(z)R(−z −1 ) = −R(−z)R(z −1 )
and hence
|r̂(ω) + r̂(ω + π)|2 = (R(z) + R(−z))(R(z −1 ) + R(−z −1 ))

= R(z)R(z −1 ) + R(−z)R(−z −1 ) = |r̂(ω)|2 + |r̂(ω + π)|2
which implies the factorization
2|r̂(ω)|2
p̂(ω) =
|r̂(ω) + r̂(ω + π)|2
c) One can choose

√ (1 + z)5
ĥ(ω) = 2 .
(1 + z)5 + (1 − z)5
One verifies that
(1 + z)5 + (1 − z)5 = 2 + 20z 2 + 10z 4
q √
whose roots are ±i 1 ± 5/5.
Exercise 7.19. a) Let hnew and h̃new be defined as
hnew [n] = (h[n] + h[n − 1])/2 and (h̃new [n] + h̃new [n − 1])/2 = h̃[n]
so that
1 + eiω ˆ ˆ 2
ĥnew (ω) = ĥ(ω) and h̃new (ω) = h̃(ω) .
2 1 + eiω
ˆ
Sine ĥnew (ω)h̃ ˆ
new (ω) = ĥ(ω)h̃(ω), the biorthogonality is conserved.
If h and h̃ have p and p̃ vanishing moments, then
ˆ (ω) = (eiω + 1)p̃ P̃ (eiω )
ĥ(ω) = (eiω + 1)p P (eiω ) and h̃
where P and P̃ are polynomials. This shows that hnew has p + 1 vanishing moments, whereas
h̃new has p̃ − 1 vanishing moments.
b) One has the following matrix expression for the balancing operations
! !
ĥnew (ω) â(ω) 0 ĥ(ω)
ˆ = ˆ (ω) where â(ω) = (1 + eiω )/2.
h̃ new (ω)
0 1/â(ω) h̃
Denoting A = â(ω), and using the following decomposition of a scaling matrix
1 A − A2

A 0 1 0 1 A−1 1 0
= ,
0 1/A 0 1 −1/A 1 0 1 1 1
one sees that the balancing operation can be implemented with 4 lifting steps.
c) We start by the original filters
h = [1] and h̃ = [−1, 0, 9, 16, 9, 0, −1]/16,
39
after one step of balancing, one obtains
h1new = [1, 1]/2 and h̃1new = [−1, 1, 8, 8, 1, −1]/8,
after one other step of balancing, one obtains
h2new = [−1, 2, 6, 2, −1]/4 and h̃2new = [1, 2, 1]/4,
which corresponds to the 5/3 biorthogonal wavelets.

Exercise 7.21. After a simple factorization, applying an odd-length filter of size K such that
h[n] = h[K − n − 1] requires K − 1 additions and (K − 1)/2 multiplications.
The direct application 5/3 filters requires 6 additions and 3 multiplications (times N for a
signal of length N ), whereas the lifting implementation requires 4 additions and 4 multiplications.
The direct application 9/7 filters requires 14 additions and 7 multiplications (times N for a
signal of length N ), whereas the lifting implementation requires 8 additions and 4 multiplication.
Exercise 7.22. For an interpolation wavelet, one has ψ(x) = φ(2x−1) and hence ĥ(ω) = e−iω /2.
One verifies that φ̃ = δ and
g̃ˆ(ω) = 2ĝ(ω + π)e−iω
satisfy the biorthogonality relations, so that

X
k k+1
ψ̃(t) = (−1) h[k]δ t − .
2
k∈Z
For interpolation of order 4, one has
h = [−1, 0, 9, 16, 9, 0, −1]/16.
One verifies that X

∀q < 4, (−1)k h[k]k q = 0
k
so that the dual wavelet has 4 vanishing moments.

Exercise 7.23. Denoting φ0 the Daubechies orthogonal wavelet with p vanishing moments, one
has φ̂(ω) = |φ̂0 (ω)|2 . Following Y. Meyer in “Wavelets with Compact Support”, one has the
following formula Z ω
(2p − 1)!
|φ̂0 (ω)|2 = 1 − sin2p−1 (t)dt
[(p − 1)!]2 22p−1 0
which is an even trigonometric polynomial of order 2p − 1. This can be explicitly computed as
2 X p
(−1)k−1 cos((2k − 1)ω)

2 1 1 (2p − 1)!
|φ̂0 (ω)| = +
2 2 (p − 1)!4p−1 (p − k)!(p + k − 1)!(2k − 1)
k=1
by expanding sin2p−1 (t). When p → +∞, one has for |ω| =

6 π/2,
1 2 X (−1)k−1
|φ̂0 (ω)|2 −→ + cos((2k − 1)ω),
2 π 2k − 1
k>1
which is the Fourier series expansion of 1[−π/2,π/2] (ω) for ω ∈ [−π, π]. This thus shows that
|φ̂0 (ω)|2 −→ 1[−π/2,π/2] (ω),
40
and hence φ(t) converges in L2 (R) toward sin(πt)/(πt).
Exercise 7.24. We follow the proof of Theorem 7.22. For each t = 2j n + h where |h| 6 2j , we
write
|f (t) − PVj f (t)| 6 |f (t) − f (t − h)| + |PVj f (t) − PVj f (t − h)|.
Since f is Lipshitz α regular,
|f (t) − f (t − h)| 6 Cf |h|α 6 Cf 2αj
where Cf is the Lipshitz constant. One also has

∞
X
f (2j (n + 1)) − f (2j n) θj,h (t − n)

PVj f (t) − PVj f (t − h) =
n=−∞
where
∞
X
Φj (t − h − 2j k) − Φj (t − 2j k) .

θj,h =
k=1
As in the proof of the proof of Theorem 7.22, since φ has exponential decay,
+∞
X
|θj,h (t − n)| 6 Cφ
n=−∞
and hence
|PVj f (t) − PVj f (t − h)| 6 Cφ max |f (2j (n + 1)) − f (2j n)| 6 Cφ Cf 2αj .
n
Exercise 7.25. Let φ(x) be a 1D interpolation function. We define the 1D transforms τ0 (x) = 0
and τ1 (x) = 2x − 1. For each set {εk }pk=1 of p binary values εk ∈ {0, 1}, we define
p
Y
ψ ε (x1 , . . . , xp ) = φ(τεk (xk )).
k=1
where we denote 0 6 ε < 2p the number whose binary expansion is {εk }pk=1 . Then ψ 0 is an
interpolating scaling function, and {ψ ε }0<ε<2p defines 2p − 1 interpolating wavelets that can be
used to analyze continuous functions defined on Rp .
Exercise 7.28. We do the derivation here in 1D. The foveated signal T f (x) is obtained from
f (x) by a spatially varying convolution with a filter scaled by t around position t, g(·/t)/t, where
g is a symmetric smooth low pass function
Z +∞
dt
T f (x) = f (t)g (x/t − 1) .
−∞ t
The operator is written over the wavelet domain as

X
Tf = Kj,j 0 ,n,n0 hf, ψj,n iψj 0 ,n0
j,j 0 ,n,n0
where ZZ
dt
Kj,j 0 ,n,n0 = hT ψj,n , ψj 0 ,n0 i = ψj,n (t)ψj 0 ,n0 (x)g (x/t − 1) dx.
t
41
It is shown in “Wavelet Foveation” by Chang, Mallat and Yap that if g and ψ have compact
support, with ψ regular with vanishing moments, then Kj,j 0 ,n,n0 decay fast when |j −j 0 | or |n−n0 |
increases. Together with the fact that Kj,j 0 ,n,n0 depends only on j − j 0 , this shows that one can
approximate the operator as a diagonal one
X
Tf = λn hf, ψj,n iψj,n
j,n
One can show that if g is of class C α , then λn = O(|n|−(α+1) ).
7 Chapter 8
Exercise 8.1. For A = 2j−L , one has that {θj [m − nA]}n is orthonormal if and only if
hθj [m], θj [m − nA]i = θj ? θ̃j [An] = δ[n].
Taking the Fourier transform of this equality, and using exercise 3.20, one has that this is
equivalent to
A−1
X
A−1 |θ̂j (A−1 (ω − 2kπ))|2 = 1. (9)
k=0
With a similar derivation, for B = 2j+1−L , the families {θj+10 1

[m − Bn]}n and {θj+1 [m − Bn]}n
are orthogonal sets of vectors orthogonal to each other if and only if
( PB−1
B −1 k=0 θ̂j0 (B −1 (ω − 2kπ))θ̂j1 (B −1 (ω − 2kπ))∗ = 0,
PB−1
B −1 k=0 |θ̂js (B −1 (ω − 2kπ))|2 = 1, ∀ s = 0, 1.
Similarly to the proof of Theorem 8.1, one verifies these relationships using (9) and
0 1
θ̂j+1 (ω) = θ̂j (ω)ĥ(Aω) and θ̂j+1 (ω) = θ̂j (ω)ĝ(Aω).
Exercise 8.3. Each node in the tree has 2p children. Fixing two quadrature mirror filters
(h0 , h1 ), for each 0 6 s < 2p , one defines the filters
p−1
Y
hs [n] = hsi [ni ]
i=0
where (si )i is the binary expansion of s. The computation of the wavelet packet coefficients is
performed by the following filtering process from the top to the bottom of the tree
p
2 k+s
dj+1 [n] = dkj ? hs [2n].
The number of scale is log2 (N )/d, the number of nodes per scale is 2pj , the number of coefficients
per node is N/2jp , so that the number of coefficients per scale is N . The complexity per scale is
O(KN ) where K is the length of the filter, so that the overall complexity is O(KN log(N )).
√
Exercise 8.4. There are N log(N ) horizontal and vertical atoms, so N log(N )2 atoms √ in total.
One
√ applies
√ the wavelet packet decomposition algorithm along each of the N rows in
O(K N log( N )) operations per row. Then one applies the same process to the column. The
overall complexity is thus O(KN log(N )) operations.
42
Exercise 8.5. a) Denoting α = (1 − i)/2, one has
2iπ
gk = αek + ᾱēk where ek [m] = e N km .
Thus one has
hgk , g` i = hαek , αe` i + hαek , ᾱe` i + hᾱek , αe` i + hᾱek , ᾱe` i

= 2|α|2 δ[k − `] + (ᾱ2 + α2 )δ[k + `].
One conclude by noticing that 2|α|2 = N and α2 = −ᾱ2 .

b) One has
hf, gk i = ᾱhf, ek i + αhf, ek i = ᾱfˆ[k] + αfˆ[k]
Exercise 8.6. One considers the following 2-periodic function


 f (t) if 0 6 t < 1,
−f (−t) if − 1 6 t < 0,

f˜(t) =

 f (2 − t) if 1 6 t < 2,
−f (t + 2) if − 2 6 t < −1.

Since f˜ is antisymmetric, one can expand orthogonally

X
f˜(t) = bk sin(πkt/2)
k
and the other symmetries of f˜ with respect to 1 and −1 implies that b2k = 0.
Since
b2k+1 = hf˜, sin(π(k + 1/2)t)i = 2hf˜, sin(π(k + 1/2)t)i
√
one sees that { 2 sin(π(k + 1/2)t)}k is an orthonormal basis of L2 [0, 1].
For a discrete signal f ∈ CN , performing anti-symmetry about −1/2 and symmetry about
N − 1/2 and −N + 1/2 shows that
r π
2
gk [n] = sin (k + 1/2)(n + 1/2)
N N
is an orthogonal basis of CN .
Exercise 8.7. One considers the following 2-periodic function

˜ f (t) if 0 6 t < 1,
f (t) =
−f (−t) if − 1 6 t < 0.
Since f˜ is antisymmetric, one can expand orthogonally

X
f˜(t) = bk sin(πkt)
k
and since
bk = hf˜, sin(πkt)i = 2hf˜, sin(πkt)i
√
one sees that { 2 sin(πkt)}k is an orthonormal basis of L2 [0, 1].
43
For a discrete signal f ∈ CN , performing anti-symmetry about −1/2 shows that
r π
2
gk [n] = sin k(n + 1/2)
N N
is an orthogonal basis of CN .
Exercise 8.9. The Meyer wavelet ψ is written ψ̂(ω) = g(ω)eiπω . A Meyer wavelet ψj,n (x) =
2−j/2 ψ(x/2j − n) satisfies
ψ̂j,n (ω) = 2j/2 b(2j ω) cos(2π(k/1/2)2j ω) − i sin(2π(k/1/2)2j ω)

for a windowing function that verifies some compatibility conditions, so that

hf, ψj,n i = hfˆ, ψ̂j,n i = 2j/2 hfˆ, gj,k
1
i − i2j/2 hfˆ, gj,k
2
i
1 2
where {gj,k }j,k is a lapped cosine basis and {gj,k }j,k is a lapped sine basis.
The Meyer wavelet coefficients can be computed by applying an O(N log(N )) FFT followed
by log(N ) lapped transform, and thus the complexity of this algorithm is O(N log(N )2 ). The
computation of the filter bank wavelet algorithm with FFT would take O(N log(N )) operation,
but would not provide as accurate results.
Exercise 8.12. a) An admissible tree is a an admissible binary tree (called root tree) with a
collection of admissible binary trees indexed by the leafs of the tree (called leaf trees).
b) We denote as Cj,k the number of admissible double trees with a root tree of depth at most j
and with leaf trees of depth at most k. One has
2 2
Cj,j = Cj−1,j + 1 > Cj−1,j
and C0,j = Bj where Bk is the number of admissible binary tree, so that
j−1 j−1
Cj,j > (Bj )2 > 2(j−1)2 .
Similarly to the proof of Theorem 8.2,
j−1
1X i
log2 (Cj+1,j+1 ) 6 2 log(Cj,j+1 ) + 1/4 6 2j log2 (C1,j+1 ) + 2.
4 i=0
Since j j
5 5
2
C1,j+1 = Bj+1 + Bj+1 6 242 + 222 ,
one obtains
5 j 5 j 1
log2 (Cj+1,j+1 ) 6 2j log2 (2 4 2 + 2 2 2 ) + 2j .
4
8 Chapter 9
Exercise 9.1. a) One has X
||f − fN ||2 = || hf, gm ig̃m ||2 .
m>N
Using the frame property, and denoting (A, B) the frame bounds, one has
X X
A |hf, gm i|2 6 ||f − fN ||2 6 B |hf, gm i|2 .
m>N m>N
44
Hence
+∞
X X X
AC 6 N 2s−1 εl (N, f ) 6 BC where C = |hf, gm i|2 N 2s−1 .
N =0 m>0 N >0
The following is similar to the proof of Theorem 9.1.

b) We denote X
f˜M = |hf, gmk i|g̃mk ,
k<M
and fM the best M term approximation. One has, using the frame property
X
||f − fM ||2 6 ||f − f˜M ||2 6= || |hf, gmk i|g̃mk ||2
k>M
X X
6B |hf, gmk i|2 = O( k −2s ) = O(M 1−2s ).
k>M k>M
Exercise 9.4. An M -term approximation of the multi-channel signal is defined as

X
fΛ = (fk,Λ )k where fk,Λ = hfk , gm igm ,
m∈Λ
where Λ is a set of M = |Λ| indexes. One has

X X
||f − fΛ ||2 = Am where Am = |hfk , gm i|2 .
m∈Λ
/ k
The best support Λ that minimize the approximation error is thus the one that selects the M
largest values of Am .
Exercise 9.5. One has, for N = 2−j
fN (t) = PVj (f )(t) = f ? Φ(N t)
where Φ is a smooth low frequency function. R

We assume that Φ is smooth and compactly supported in [−K/2, K/2], and that Φ = 1,
||Φ||∞ 6 C. One then checks that the approximation error is localized in [−K/2, K/2]/N
Z 1
|f (t) − f ? Φ(N t)|2 dt 6 2CKN −1 .
0
Exercise 9.6. a) For a smooth function, one has

Z x Z x
|f (x)| = | f 0 (x)dx + f (0)| 6 |f 0 (x)|dx + |f (0)| 6 T V (f ) + |f (0)| < +∞.
0 0
This result caries over to arbitrary function by approximating by a smooth function.

b) We consider f (x) = ||x||−α φ(x) where φ is some smooth localizing function, that is 1 inside
the disc of radius 1. One has, for ||x|| 6 1
x
∇f (x) = −α .
||x||α+2
45
Using a polar change of variables,
Z Z Z 1
dx dr
||∇f (x)||dx = C + α = C + 2πα
||x||α+1 0 rα
where C account for the total variation outside the unit disk. One can see that for 0 < α < 1, f
is of bounded variation, but f is not bounded.
Exercise 9.7. Formally, we define
Z
g(t) = H(ω, t)dω where H(ω, t) = (iω)p fˆ(ω)eiωt .
R
One has
|H(ω, t)| 6 |ω|p |fˆ(ω)|
Using Cauchy-Schartz inequality, one has
Z Z !1/2
|ω|p |fˆ(ω)|dω 6 |ω|2(p−s) dω ||f ||Sob(s)
|ω|>η |ω|>η
R
which is bounded if s > p + 1/2. So using classical Theorem of derivation under the sign , this
α
shows that f is C and f (α) = g.
Exercise 9.8. One can verify numerically that Fourier and orthogonal polynomial approxima-
tions behave similarely. An orthogonal polynomial of degree p has p oscillations and is similar to a
sinusoid function. The linear N -term approximation with orthogonal polynomial of a Cα function
is better that the error produced by a Taylor approximation, and is thus ||f − fN ||2 = O(N −2α )
M
X
||f − fM ||2 6 (tk+1 − tk )∆ = ∆.
m=0
We use the fact that any bounded variation function can be decomposed as f = f1 − f2 where
f1 , f2 are increasing functions. Without loss of generality, we assume f1 (0) = f2 (0) = 0 and
f1 (1) = f2 (1) = 1.
For each function fi , we build a set Ti = {tik = fi−1 (i/M )}M −1
k=0 of points which guarantees
−1
that fi varies of at most ∆ = 1/M over each [tk , tk+1 ]. Defining {tk }2M
i i
k=0 = T1 ∪ T2 guarantees
that f varies of less than ∆ = 1/M over each [tk , tk+1 ]. One thus has ||f − fM ||2 = O(M −1 ).
Exercise 9.11. We consider a function f such that |hf, ψj,n i| = 2js . For dyadic N = 2J , and
ordering wavelet coefficients from coarse to fine scales, εn [N ] = εl [N ].
For α < s − 1/2, one has
XX X X
|hf, ψj,n i|2 2−2jα = 2−j 2−2αj 22sj = 22j(s−α−1/2) < +∞
j60 n j j60
and Theorem 9.4 shows that f is in W α .

Exercise 9.12. b) Denoting | supp(ψ)| = C = 8, one has for the CK wavelets whose support
intersect a singularity at scale 2j
|hf, ψj,n i| 6 ||f ||∞ ||ψ||1 2j/2 ,
46
the other coefficients being zero. One has
εl [M ] 6 2CK||f ||2∞ ||ψ||21 M −1 .
Since there is only KC non-zero coefficients per scale, the non-linear approximation selects all
non-zero coefficients corresponding to wavelets whose support intersects a singularity at scale 2j
up to scale −J = M/(KC)
X
εn [M ] 6 CK||f ||2∞ ||ψ||21 2j 6 2CK||f ||2∞ ||ψ||21 2J = 2CK||f ||2∞ ||ψ||21 ω −M ,
j6J
where ω = 2−1/(KC) .
k k
X 1 X
argmin |f [n] − a|2 = f [n].
a∈R `−k+1
n=` n=`
b) We call Vp,` the set of signals supported on [0, . . . , `], that assumes less than p different values.
Any signal fk ∈ Vp,k can be decomposed as
fk = f` + a1[`,k]
where f` ∈ Vp−1,`−1 and a ∈ R. One thus has
min ||f − fk ||2[0,k] = min min ||f − f` ||2[0,`−1] + min ||f − a||2[`,k] ,
fk ∈Vp,k `∈[0,k−1] f ∈Vp,`−1 a∈R
which gives the result.

An algorithm can computes εp,k for increasing p and k, and Σp,k which are the locations of
discontinuities in the optimal signal in Vp,k .
One initializes:
• For all k, ε1,k = c1,k and Σ1,k = ∅.
• For all p, εp,1 = 0 and Σp,0 = {0, . . . , 0} (p times).
For p = 2, . . . , K, for k = 1, . . . , N − 1, compute
εp,k = min εp−1,` + c`,k

`∈[0,k−1]
and denoting `∗ such that εp,k = εp−1,`? + c`? ,k , update
Σp,k = Σp−1,`? ∪ {`? }.
After running this algorithm, fK,N has discontinuities in ΣK,N . The main numerical cost is the
computation of the cp,k , which takes O(N 2 ) for each k, so the overall complexity is O(KN 2 ).
Exercise 9.14. a) Performing a first order approximation of the phase near point 2j n, and
assuming that the amplitude is nearly constant over the support of size 2j , one gets
f (t) ≈ a(2j n) exp iφ(2j n) + iφ0 (2j n)(t − 2j n) .

One thus has
|hf, ψj,n i| ≈ |a(2j n)2−j/2 hψ(t/2j − n), exp(iφ0 (2j n)(t − 2j n))i|
47
and thus performing the change of variable t/2j − n 7→ t/2j − n in the inner product, one obtains
|hf, ψj,n i| ≈ a(2j n)2j/2 |ψ̂(2j φ0 (2j n))|.
b) For f (t) = sin(1/t), one has φ(t) = 1/t, a = 1, φ0 (t) = −1/t2 , and thus
|hf, ψj,n i| ≈ 2j/2 |ψ(2−j n−2 )|.
The `p norm of the coefficients reads

X X
|hf, ψj,n i|p ≈ |2jp/2 |ψ(2−j n−2 )|p . (10)
j60,06n62−j j60,06n62−j
The wavelet function is close to a band pass filter (Shannon wavelet), so |ψ̂(ω)| is nearly constant
equal to A > 0 in some interval [C1 , C2 ], and small outside. As an approximation, we can thus
considers that for each j the sum in (10) is well approximated by restricting it to indexes n that
satisfies
−1/2 −j/2 −1/2 −j/2
C1 6 2−j n−2 6 C2 =⇒ C2 2 6 n 6 C1 2 .
For each j, the number of elements in the sum (10) is thus approximately C2−j/2 for some
constant C, so that X X
|hf, ψj,n i|p ≈ CA 2j/2(p−1) . (11)
j60,06n62−j j60
This `p norm is thus finite if and only if p > 1.

c) Theorem 9.10 tells us that for all p < 1, εn [M ] = O(M 1−2/p ), so that εn [M ] = O(M −α ) for
any α < 1.
48

Wavelet Tour Corrections

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Wavelet Tour Corrections

Uploaded by

Copyright:

Available Formats

Correction of the exercises from the book

A Wavelet Tour of Signal Processing

Exercise 2.4. One has

where Conj(a) = a∗ is the complex conjugate. The same computation leads to

fˆi (ω) = [fˆ(ω) − fˆ∗ (−ω)]/2.

Exercise 2.5. One has Z

where h is a piecewise linear hat function with h(0) = 2.

By integration by parts, one gets an ordinary differential equation

fa,u (t) = eiat g(t − u) + e−iat g(t + u).

We verify that σω (fa,u ) increases proportionally to u. Its Fourier transform is

fˆa,u (ω) = e−iuω ĝ(ω − a) + eiuω ĝ(ω + a)

and cn is the Fourier coefficient

One thus has

fˆn (ω ± 2nω0 ) = 2ĝ(ω)1[(−1±2n)ω0 ,(1±2n)ω0 ] (ω).

The change of variable ω ± 2nω0 → ω and summing for n and −n gives

fˆn (ω) = [ĝ(ω − 2nω0 ) + ĝ(ω + 2nω0 )]ĥ(ω),

For φ = λ1[a,b] , |φ0 | = λδa + λδb and hence ||φ||V = 2λ.

f (x) = f0 (x) + A u(cos(θ)x1 + sin(θ)x2 )

f ? hξ (x) = f ? gξ (x) (1)

Exercise 3.3. a) The function f is an interpolation function if and only if

Taking the inverse Fourier transform leads to

Exercise 3.5. The orthogonal projection of f on Us is defined by

||Lf ||∞ 6 ||f ||∞ ||h||1 .

one has using Theorem 3.1

with the kernel obtained by inverse Fourier transform formula

Exercise 3.10. a) One has f˜(ns) = f ? φs where φs = 1[−s/2,s/2] .

One has, using (7.21),

3 sin2 (ω/2) ĥ(ω)

Exercise 3.13. a) Let h[0] = 1, h[−1] = −1 and h[n] = 0 for n ∈

||f ||V ||f ||V N ||f ||V

since sin(x) > 2x/π for |x| 6 π/2.

If h is a low-pass filter, then g is a high-pass filter.

where the second line is obtained by change of variable n − p → p in the summation.

and one verifies that |a∗k

b) {h[n − m]}m is an orthogonal basis if and only if for all m, m0

hh[n − m], h[n − m0 ]i = h ? h[m − m0 ] = δ[n − m0 ].

Let hp (ω) = (1 − ae−iω )−p . Observe that

and one thus has

Exercise 3.20. a) One has

This shows that

b) If for ω ∈ [−π, π], ŷ(ω) = 0 for ω ∈

ŷ(ω) = x̂(M ω)φ̂M (ω) where φ̂M (ω) = M 1[−π/M,π/M ] (ω).

Exercise 3.21. a) One has, using the change of variable r = n − pN ,

Using the sampling Theorem 3.1, one gets

ĝ[k] = (fˆa [k] + fˆa [−k]∗ )/2 = fˆ[k],

= hek1 , e ihek2 , e i = δ[k1 − k10 ]δ[k2 − k20 ].

Exercise 3.24. We define the sub-images

∀ 0 6 k1 , k2 < L, ∀ 0 6 n1 , n2 < M, fk [n] = f [n + kM ].

48N log2 (M ) 6 2N M 2 ⇔ M 2 / log2 (M ) > 24

which we found numerically to be M > 9.

KN (log(N1 ) + log(N2 ) + log(N3 )) = KN log(N )

and thus, expanding h0 (t)h∗ (t),

and the second integral vanishes because g 0 is an odd function.

Taking the finite discrete Fourier transform of this relation leads to

and hence the result.

Sf [m, l] = e−i2πlm/N f ? gl [m] with gl [n] = g[−n] e−i2πln/N .

which implies P P Φ = P Φ so that P is a projector on Im(P ), which is exactly, by Theorem 4.2

Exercise 4.6. One has

Taking the Fourier transform leads to

Using the fact that |ψ̂s (ω)|2 = s|ψ̂(sω)|2 and

leads to the result.

|(iω)p fˆ(ω)eizω | 6 |ω|p |fˆ(ω)|e−ρω ,

is holomorphic on Ωρ for all ρ > 0.

s2 |θ10 (t)| = s2 |θ200 (t)| = s2 ab2 1.

Exercise 5.22. a) One has