Professional Documents
Culture Documents
Pencho Petrushev
Department of Mathematics, University of South Carolina, Columbia, SC 29208
e-mail: pencho@math.sc.edu
1. Introduction
We consider the problem of recovering a function f from a blurred (by a linear
. It is important to note that,
operator) and noisy version of f : Y = Kf + W
30
31
in general, for a problem like this there exists a basis which is fully adapted
to the problem, and as a consequence, the inversion remains stable; this is the
Singular Value Decomposition (SVD) basis. The SVD basis, however, might be
difficult to determine and handle numerically. Also, it might not be appropriate
for accurate description of the solution with a small number of parameters.
Furthermore, in many practical situations, the signal exhibits inhomogeneous
regularity, and its local features are particularly interesting to recover. In such
cases, other bases or frames (in particular, localized wavelet type bases) might
be much more appropriate to represent the object at hand.
Our goal is to devise an inversion procedure which has the advantages of
localization and multiscale analysis of wavelet representations without losing
the stability and computability of the SVD decompositions. To this end we
utilize the construction (due to Petrushev and his co-authors) of localized frames
(termed needlets) built upon particular bases - here the SVD bases. This
construction uses a Calder
on type decomposition combined with an appropriate
quadrature (cubature) formula. It has the big advantage of producing frames
which are close to wavelet bases in terms of dyadic properties and localization,
but because of their compatibility with the SVD bases provide stable and easily
computable schemes.
NEED-D is an algorithm combining the traditional SVD inversion with an
appropriate thresholding technique in a well chosen new basis. It enables one to
approximate the targeted functions with excellent rates of convergence for any
Lp loss function, and over a wide range of Besov spaces.
Our main idea is by combining the thresholding algorithm with SVD-based
frames to create an effective and practically feasible algorithm for solving the
inverse problem described above. The properties of the localized frame to be
constructed depend on the underlying SVD basis. We will consider two different
behaviors, the first corresponds to a wavelet behavior in the sense that the
properties of the system are equivalent (as far as we are concerned) to the
properties of a true wavelet basis. This case typically arises in the deconvolution
setting. In the second case, the properties of the frame may differ from wavelet
bases and truly depend on the SVD basis at hand (hence on the operator K).
We will explore in detail a case typically arising when the SVD basis is a Jacobi
polynomial basis. It is illustrated by the Wicksell problem. We show that our
scheme is capable of achieving rates of convergence which are optimal in the L2
case (to the best of our knowledge, for the Wicksell problem this is the only
case studied up to now). For other Lp norms we obtain interesting rates of
convergence, which are new in the literature.
We also give a simulation study for the Wicksell problem which shows that the
NEED-D algorithm applied in combination with SVD based frames is valuable
since it outperforms other standard algorithms in almost all situations.
The paper is organized in the following way: the second section introduces
the model, the classical SVD methods, and the two basic examples considered
in this paper, i.e. the deconvolution and Wicksell problems. The third section
introduces the needlet construction, gives some basic properties of needlets and
introduces the NEED-D algorithm. The fourth section explores its properties in
32
the wavelet scenario. The main motivation for the NEED-D algorithm is given
there after. The fifth section is devoted to the results in a Jacobi scenario. The
sixth section is devoted to simulation results. The proofs of the main results
from sections 45 are given in sections 78, respectively. The last section is
an appendix which contains the definition and basic properties of the Jacobi
needlets and the associated Besov spaces.
2. Inverse Models
Suppose H and K are two Hilbert spaces and let K : H 7 K be a linear
operator. The standard linear ill-posed inverse problem consists in recovering a
good approximation f of the solution f of
g = Kf
(1)
(2)
(W , h)K form random Gaussian vectors (centered) with marginal variance kgk2K ,
khk2K , and covariance (g, h)K (with the obvious extension when one considers k
functions instead of 2).
Equation (2) means that for any g K, we observe Y (g) := (Y , g)K =
(Kf, g)K + (g), where (g) N (0, kgk2), and Y (g), Y (h) are independent
random variables for orthogonal functions g and h.
2.1. The SVD Method
Under the assumption that K is compact, there exist two orthonormal bases
(SVD bases) (ek ) of H and (gk ) of K, and a sequence (bk ), bk 0 as k ,
such that
K Kek = b2k ek ,
KK gk = b2k gk ,
Kek = bk gk ;
N
X
k=0
b1
k hY , gk iek ,
K gk = bk ek
33
where N = N () has to be properly selected. This SVD method is very attractive theoretically and can be shown to be asymptotically optimal in many situations (see Mathe and Pereverzev [23] together with their non linear counterparts
Cavalier and Tsybakov [6], Cavalier et al [5], Tsybakov [33], Goldenschluger and
Pereverzev [16], Efromovich and Koltchinskii [12]). It also has the big advantage of performing a quick and stable inversion of the operator. However, it has
serious limitations: First, the SVD bases might be difficult to determine and
handle numerically. Secondly, while these bases are fully adapted to describe
the operator K, they might not be appropriate for accurate description of the
solution with a small number of coefficients. Also in many practical situations,
the signal has inhomogeneous regularitiy, and its local features are particularly
interesting to recover. In such cases, other bases (in particular, localized wavelet
type bases) are much more suitable for representation of the object at hand.
In the last ten years, various nonlinear methods have been developed, especially in the direct case with the objective of automatically adapting to the
unknown smoothness and local singular behavior of the solution. In the direct
case, one of the most attractive methods is probably wavelet thresholding, since
it allies numerical simplicity to asymptotic optimality on a large variety of functional classes such as Besov or Sobolev spaces.
To apply this approach to inverse problems, Donoho [10] introduced a waveletlike decomposition, specifically adapted to the operator K (Wavelet-VagueletteDecomposition) and utilized a thresholding algorithm to this decomposition. In
Abramovich and Silverman [1], this method was compared with the similar
vaguelette-wavelet decomposition. Other wavelet schemes should be mentioned
here, such as the ones from Antoniadis and Bigot [3], Antoniadis & al [4], Dicken
and Maass [9], and especially for the deconvolution problem, Penski & Vidakovic
[29], Fan & Koo [13], Kalifa & Mallat [19], Neelamani & al [27]. Later on Cohen et al [7] introduced an algorithm combining a Galerkin inversion with a
thresholding algorithm.
The approach developed here was greatly influenced by these works.
2.1.1. Deconvolution
The deconvolution problem is probably one of the most famous inverse problems,
giving rise to a great deal of investigations, specially in signal processing, and
has an extensive bibliography. In the deconvolution problem, we consider the
following operator: Let in this case H = K be the set of square integrable periodic
functions, with the standard L2 [0, 1] norm, and consider
Z 1
f H 7 Kf =
(u t)f (t)dt H,
(3)
0
34
In this case simple calculations show that the SVD bases ek and gk both
coincide with the Fourier basis. The singular values correspond to the Fourier
coefficients of the function :
bk = k .
(4)
2.1.2. Wicksells problem
Another typical example is the following classical Wicksells problem [34]. Suppose a population of spheres is embedded in a medium. The spheres have radii
that may be assumed to be drawn independently from a density f . A random
plane slice is taken through the medium and those spheres that are intersected
by the plane furnish circles which radii are the points of observation Y1 , . . . , Yn .
The unfolding problem is then to determine the density of the sphere radii from
the observed circle radii. This problem also arises in medicine, where the spheres
might be tumors in an animals liver (see Nyshka et al [28]), as well as in numerous other contexts (biological, engineering, etc.) see for instance Cruz-Orive
[8].
The difficulty of estimating the target function is well illustrated by figure 1.
The Wicksell operator has a smoothing effect, thus the local variations of the
target function become almost invisible in the case of observations corrupted
by noise. (Also compare the blurred and noised observations in figure 6 to the
target functions of figure 4.)
0.5
0.05
0.5
0.05
0.1
0.05
0
0.5
0.1
0.05
0.5
0.1
0.5
Fig 1. Heavisine function, its image by the Wicksell operator without and with gaussian noise
with rsnr = 5
Following Wicksell [34] and Johnstone and Silverman [18], the Wicksells
problem corresponds to the following operator:
H = L2 ([0, 1], d), d(x) = (4x)1 dx,
K = L2 ([0, 1], d), d(x) = 4 1 (1 y 2 )1/2 dy,
and
35
Here Pk0,1 is the kth degree Jacobi polynomial of type (0, 1) and Uk is the second
type Chebishev polynomial of degree k. The singular values are
bk =
(1 + k)1/2 .
16
(5)
Hk ,
k=0
where the Hk s are finite dimensional spaces. For simplicity, we assume that H0
is reduced to the constants.
Let (eki )i=1,...,lk be an orthonormal basis of Hk . Then the orthogonal projector
Lk onto Hk takes the form
Z
Lk (f )(x) =
f (y)Lk (x, y)d(y), f L2 (Y, ),
Y
where
Lk (x, y) =
lk
X
i=1
(6)
36
Then
a2 (/2j ) = 1,
j0
|| 1.
(7)
a2 (k/2j )Lk
k0
2j1 <k<2j+1
j (f )
in L2 (Y, ).
(8)
j=0
J
X
j = L0 +
J X
X
j=0
j=0
a2 (k/2j )Lk =
(k/2J+1 )Lk
(9)
and hence
kf L0 (f )
J
X
j (f )k2 =
j=0
l2J+1
l2J
kLl (f )k2 +
2J l<2J+1
kLl (f )k2 0 as
k
M
m=0
Hm .
J ,
37
We make two additional assumptions which will enable us to discretize decomposition (8) from Proposition 1:
(a)
f Kk , g Kl = f g Kk+l .
Xk
j 0.
(11)
(12)
and hence
Now, by (10)
j (x, y) =
X2j+2 2
which implies
Z
Z
j (x, y)f (y)d(y) =
j f (x) =
Y X
2j+2 2
X2j+2 2
Z
p
p
Mj (x, )
f (y) Mj (y, )d(y).
(13)
We are now prepared to introduce the desired frame. Let Zj = X2j+2 2 for j 0
and Z1 = X0 . We define the frame elements (needlets) by
p
(14)
j, (x) = Mj (x, ), Zj , j 1.
Proposition 2. The family (j, )Zj ,j1 is a tight frame for L2 (Y, ).
38
Proof. As
f = lim L0 (f ) +
J
we have
J
X
j (f )
j=0
J
X
hj (f ), f i.
kf k = lim hL0 (f ), f i +
2
j=0
But by (15)
hj (f ), f i =
Zj
hf, j, ihj, , f i =
Zj
|hf, j, i|2 ,
j 0,
150
100
50
50
100
150
1.0
0.8
0.6
0.4
0.2
0.0
0.2
0.4
0.6
0.8
1.0
39
In the case of the unit sphere in Rd+1 , where Hk are the spaces of spherical
harmonics, the following localization property of the needlets is established in
Narcowich, Petrushev and Ward [25, 26]: For any k N there exists a constant
Ck such that:
Ck 2dj/2
.
|j ()|
[1 + 2j arccos < , >]k
In the case of Jacobi polynomials on [1, 1], the localization of the needlets
proved in Petrushev, Xu [31] takes the form: For any k N there exists a
constant Ck such that
|j (cos )|
Ck 2j/2
p
,
(1 + 2j | arccos |)k w, (2j , cos )
|| ,
Zj
P
i
with fi = (f, ei )H
Using Parsevals identity, we have j = (f, j )H = i fi j
i
and j = (j , ei )H . If we put Yi = (Y , gi )K , then
X
fj ej , K gi )H + i = bi fi + i ,
Yi = (Kf, gi )K + i = (f, K gi )K + i = (
j
J
X
X
j=1 Zj
t(j )j ,
40
with
(16)
(17)
Here is a tuning parameter of the method which will be properly selected later
on. Notice that the thresholding depends on the resolution level j through the
constant j which will also be specified later on, and the same with regard to
the upper level of details J. Notice also that, in this paper, we concentrated on
hard thresholding, whereas various other kind of thresholdings could be used,
likely giving comparable results at least theoretically.
We will particularly focus on two situations (corresponding to the two examples discussed above). In the first case (see subsection 4), the needlets have very
nice properties and behave exactly like wavelets. This is for instance the case
of the deconvolution, where the SVD basis is the Fourier basis. However, more
complicated problems e.g. the Wicksells problem exhibit more delicate concentration properties for the needlets giving rise to different behaviors in terms of
rates of convergence for the estimators.
4. NEED-D in wavelet scenario
In this section, we assume that the needlet system has the following properties:
For any 1 p < , there exist positive constants cp , Cp , and Dp such that
Card Zj C2j ,
j( p
2 1)
(18)
j( p
2 1)
kj kpp
cp 2
Cp 2
,
X
X
u j kpp Dp
|u |p kj kpp , for any collection (u ).
k
Zj
(19)
(20)
Zj
s
define the space B,r
as the collection of all functions f with f =
P We P
such
that
j0
Zj j j
1
s
kf kB,r
:= k(2j[s+ 2 ] k(j )Zj kl )j0 klr < ,
s
B,r
(M )
s
kf kB,r
M.
and
(21)
(22)
i
X j
]2 C22j , j 0.
[
b
i
i
2
(23)
41
where
1 p
s
, if s ( + )( 1)
s + + 1/2
2
1
1 p
s 1/ + 1/p
, if
s < ( + )( 1).
=
s + + 1/2 1/
2
=
m 2m i2m+1 1
i
j
]2 .
bi
Hence, m j,
42
i 2
[j
] c22(jm) .
2m i2m+1 1
i
This means that the energy of j
decays exponentially for i 2j , which reveals
the role of the Littlewood Paley decomposition in the previous construction,
replacing the exponential discrepancy by a cut-off.
The following proposition establishes a kind of converse property: The construction of needlet systems always implies that condition (23) is satisfied in the
regular case.
i
Proposition 3. If (j, ) is a frame such that {i : j
6= 0} is contained in a
j
j
j2 :=
i
X j
]2 C22j .
[
b
i
i
i
Proof. Since the elements of an arbitrary frame are bounded in norm and j
6=
j
j
0 only for C1 2 i C2 2 , we have
i
X j
[
]2 C22j kj, k2 C 22j .
b
i
i
43
fulfilled by the needlet system (Jacobi needlets) constructed using the Jacobi
polynomials (Pk )k0 . The proofs are given in the appendix.
We will consider two sets of conditions. The first one (which only depends on
) is the following:
Card Zj 2j ,
X
kj kpp Cp 2jp/2 2j(p2)(1+) ,
Zj
Xj
j, kp C(
Xj
(25)
1
,
j, p 6= 2 +
+ 1/2
(26)
(27)
s
e,r
We define the space B
as the collection of all functions f on [1, 1] with
representation
X X
f=
j j
j1 Zj
such that
kf kBes
,r
X
:= k(2js (
|j, | kj, k )1/ )j1 klr < ,
and
e s (M ) kf k s M.
f B
,r
er
B
(28)
(29)
p
log 1/
+1
p )+
and
( 2
bi i ,
2
1+2
2 J = t
+1
p )+ }]
1
> .
2
where
(i) if p < 2 +
1
+1/2 ,
then
=
(ii) if p > 2 +
1
+1/2 ,
p
log(1/)]p ,
s
;
s + + 12
then
=
s
.
s + + ( + 1)(1 p2 )
44
Remarks :
1
, the rate obtained here is the usual one, and can
1. In the case p < 2 + +1/2
1
be proved to be minimax (see [35]). The case p > 2 + +1/2
introduces a
new rate of convergence.
2. Conditions (25)(27) enabled us to estimate the rates of convergence of
our scheme, whenever the index of the Besov space is the same as the
index of the loss function (p = ). In the sequel, we will study the case
where p and are independently chosen. This requires, however, some
additional assumptions. 3
c2
j(p2)(+1)
(p2)(+1/2)
k ()
kj kpp
(30)
j(p2)(+1)
(p2)(+1/2)
C2
k ()
k () = 2 k() < 2
j1
(31)
2 J = t
and
2 16p[1 + 4{(
+1
+1
)+ (
)+ }]
2
p
2
p
1
> .
2
s
e,r
Then for f B
(M ) with s > max{,} { 12 2( + 1)( 21
1
1
1)( p ) 0}, we have
and
1
)
p
log(1/)]p ,
a = max{a(), a()} 2
2( +
(32)
with
(s) =
(s, ) =
45
s
,
s + + 12
s 2(1 + )( 1 p1 )
s + + 2(1 + )( 12 1 )
and,
a() =
I{p = 0}
(+ 12 )(p)
(2)(+1/2)1
X
i
Yi
ei ,
bi
46
In the simulations a first SVD estimator with projection weights was used:
(
i = 1 if i N,
i = 0 if i > N,
where the parameter N was fitted for each setting so as to minimize the root
mean square error (RM SE) of the estimator.
We also use the SVD estimator developed in Cavalier and Tsybakov [6], which
is completely adaptive with a data driven choice of the filter and thus much more
convenient than the former in practice. The values of i are constant in blocs
Ij = [j1 , j 1] with limits 0 = 1 and J = N + 1 determined further:
2
= 1 j (1+j )
if i Ij , j = 1, . . . J,
i
kY k2(j)
+
= 0
if i > N,
i
where:
Yi
Yi = ,
bi
j =
kY k2(j) =
maxiIj b2
i
P
2 ,
b
iIj i
Yi2 ,
j2 = 2
iIj
bi2 ,
iIj
if j = 0,
j = 1
j =
if j = 1,
j = j1 + (1 + )j1
if j = 2, . . . , J,
Pm
2 3
where J is large enough such that: J > max{m : i=1 b2
}.
i
47
0.5
1.5
2.5
60
40
20
20
40
1
0.7
4
1
2
1
4
0.7
Fig 3. Examples of needlets: 7,10 , 7,40 , 7,80 and 7,120 (from left to right)
p
i
ej,
= a(i/2j1 )Pi () bj, ,
thus the estimated coefficients of f in the frame are very easy to compute.
6.2. Parameters of the simulation
We consider the four commonly used target functions f represented in figure 4,
and three levels of noise corresponding to three values of the root signal to
Bumps
Heavisine
48
Doppler
1
0.5
0.5
0.5
0.5
0
0
0
0.5
0.5
0
0.5
0.5
0.5
0.5
noise ratio of K(f ): rsnr {3, 5, 7}. The discretization resolution level is setto
n = 1024, and the constant in the thresholds of NEED-D is set to = 0.75 2.
The estimation error is evaluated by a Monte Carlo approximation of several
Lp () losses:
L1 is computed as the average over 20 runs of
1
n
n
P
i=1
s
|f ( ni ) f( ni )|/( 4i
n ).
1
n
n
P
(f ( ni ) f( ni ))2 /( 4i
n ).
i=1
0.8
0.8
0.6
0.4
0.2
0
200
400
200
400
200
400
0.8
0.2
200
400
200
400
200
400
0.5
0.6
0.3
0.4
0.4
0.4
0.2
0.2
0
200
400
0.8
0.1
0.3
0.2
0
200
400
0.4
0.6
200
400
0.8
0.2
0.8
0.6
0.3
0.4
0.6
0.4
0.2
0.2
0
0.3
0.2
0.4
0.6
0.4
0.4
0.1
0.8
0.5
0.6
0.2
49
200
400
0.1
0.4
0.2
0
200
400
200
400
0.2
Fig 5. Value of the mean square error of the non adaptive SVD estimator (y-axis) for each
value of N (x-axis) for rsnr = 7 to rsnr = 3 (from top to bottom) and for the target function
Blocks, Bumps, Heavisine and Doppler (from left to right)
of the interval, which is given lesser importance by errors measured with the
weight (x) = 1/(4x), for x ]0, 1].) This order of comparison is confirmed by
the lower values of L1 and RM SE for NEED-D than for SVD in all the settings
(see tables 1 and 2).
Blocks
Bumps
Heavisine
Doppler
Adaptive SVD
NEED-D
low
med
high
low
med
0.0399 0.0465 0.0591 0.0347 0.0404
0.0258 0.0295 0.0361 0.0180 0.0206
0.0248 0.0299 0.0401 0.0205 0.0254
0.1002 0.1085 0.1230 0.0858 0.0909
Table 1
Error L1 for each target, each noise level and each estimator
low
0.0452
0.0324
0.0257
0.1032
SVD
med
0.0495
0.0388
0.0305
0.1138
high
0.0677
0.0463
0.0402
0.1307
high
0.0511
0.0270
0.0321
0.1007
0.1
0.05
0.05
0
0.05
0.1
0.1
0
0
0.5
0.5
0.1
0.5
0.1
0.5
0.5
0
0.5
0
0
0.5
0.5
0.5
0.5
0.5
0
0.5
0
0
0.5
0.5
0.5
0.5
0.5
0
0.5
50
0
0
0.5
0.5
0.5
0.5
0.5
0.5
0.5
Fig 6. From top to bottom: observed data, NEED-D estimator, adaptive SVD estimator and
non adaptive SVD estimator for high noise (rsnr=3)
7. Proof of Theorem 1
In this proof, C will denote an absolute constant which may change from one
line to the other.
Blocks
Bumps
Heavisine
Doppler
Adaptive SVD
low
med
high
0.0665 0.0743 0.0900
0.0453 0.0508 0.0617
0.0266 0.0317 0.0418
0.1042 0.1114 0.1258
Table 2
Error L2 for each target, each noise level and each
low
0.0714
0.0489
0.0278
0.1092
SVD
med
0.0790
0.0577
0.0327
0.1200
high
0.0959
0.0706
0.0422
0.1378
low
0.0606
0.0378
0.0235
0.0969
NEED-D
med
0.0673
0.0416
0.0288
0.0999
estimator
high
0.0816
0.0523
0.0379
0.1071
51
J
X
X
j=1 Zj
(t(j ) j )j kpp + k
XX
j>J Zj
j j kpp }
s
The term II is easy to analyse, as follows: Since f belongs to B,r
(M ), using
standard embedding results (which in this case simply follows from direct com1
p1 )+
s(
(M ), for
j>J Zj
1
1
p
)+
s(
+1/2
Bounding the term I is more involved. Using the triangular inequality together
with H
older inequality, and property (20) for the second line, we get
I 2p1 J p1
J
X
Ek
j=1
2p1 J p1 C
J
X
Zj
j=1 Zj
(t(j ) j )j kpp
E|t(j ) j |p kj kpp .
j=1 Zj
E|t(j ) j |p kj kpp
J
X
X
j=1 Zj
E|t(j ) j |p kj kpp I{|j | t j } + I{|j | < t j }
J
X
X
: Bb + Bs + Sb + Ss.
P i
= i i bj
is a gaussian rani
i
P j
2
2
dom variable centered, and with variance
i [ bi ] , we have using standard
P
If we notice that j j = i
Yi bi fi i
j
bi
52
P{|j j | t j } 2 /8
2
Hence,
Bb
Ss
J
X
X
j=1 Zj
J
X
X
j=1 Zj
jp p kj kpp I{|j |
t j }
2
And,
Bs
J
X
X
j=1 Zj
Zj
J
X
2jp(+ 2 ) p
/16
j=1
/16
Now, if we remark that the j are necessarily all bounded by some constant
s
(depending on M ) since f belongs to B,r
(M ), and using (19),
Sb
J
X
X
j=1 Zj
J
X
X
j=1 Zj
J
X
j=1
2 j 2
/8
2
8
/8
I{|j | 2t t j }
p
(2+1)
It is easy to check that in any cases if 2 16p the terms Bs and Sb are
smaller than the rates announced in the theorem.
If we recall that:
r
1
t = log
53
J
X
2j(p+ 2 1)
j=1
J
X
I{|j |
|j |z [t j ]z
Zj
p
2j(p+ 2 1)
j=1
Ct pz
Zj
J
X
2j[(pz)+ 2 1]
j=1
Zj
t j }
2
|j |z
J
X
2j( 2 1)
j=1
Zj
C[t ]pz
J
X
|j |z jpz [t ]pz
p
2j((pz)+ 2 1)
j=1
Zj
|j |z
So in both cases we have the same bound to investigate. We will write this
bound on the following form (forgetting the constant):
I + II = t pz1 [
j0
X
2j[(pz1 )+ 2 1]
j=1
+ t pz2 [
|j |z1 ]
Zj
J
X
2j[(pz2 )+ 2 1]
j=j0 +1
Zj
|j |z2 ]
The constants zi and j0 will be chosen depending on the cases, with the only
constraint p zi 0.
Notice first, that we only need to investigate the case p , since when
s
s
p , Br
(M ) Bpr
(M ).
Let us first consider the case where s ( + 21 )( p 1), put
q=
p(2 + 1)
2(s + ) + 1
and observe that on the considered domain, q and p > q. In the sequel it
will be useful to observe that we have s = ( + 21 )( pq 1). Now, taking z2 = ,
we get:
J
X
X
p
|j | ]
2j[(p)+ 2 1]
II t p [
j=j0 +1
Now, as
Zj
p
1
p
1
1
+ ( 1) = s +
2q
q
2
and
54
|j | = 2j(s+ 2 ) j
Zj
s
with (j )j lr (this last thing is a consequence of the fact that f B,r
(M )
and item (5)), we can write:
X
1
2jp(1 q )(+ 2 ) j
II t p
j=j0 +1
Ct
1
p j0 p(1
q )(+ 2 )
The last inequality is true for any r 1 if > q and for r if = q. Notice
p
12 ). Now if we choose j0 such that
that = q is equivalent to s = (2 + 1)( 2
p
1
2j0 q (+ 2 ) t 1 we get the bound
t pq
which exactly gives the rate announced in the theorem for this case.
As for the first part of the sum (before j0 ), we have, taking now z1 = qe, with
1
P
P
1
q e
qe , so that [ 21j Zj |j |e
] q [ 21j Zj |j | ] , and using again (7),
I
j0
X
X
q
q
q )+ p
2 1]
|j |e
]
t pe
[
2j[(pe
1
j0
X
q
t pe
[
q
t pe
j0
X
Zj
eq
eq X
q )+ p
2 ][
2j[(pe
|j | ] ]
Zj
e
q
1
q
2j[(+ 2 )p(1 q )] je
e
q
q j0 [(+ 21 )p(1 q )]
Ct pe
2
Ct pq
The last two lines are valid if qe is chosen strictly smaller than q (this is possible
since q).
p
Let us now consider the case where s < (2 + 1)( 2
21 ), and choose now
q=p
1
2
s+ +
1
2
1
p
1
2
1
1
)
p(s+ p
and q
1
s++ 21
p
p
1
1
= 2q q + ( q
=
1).
55
s
Hence taking z1 = and using again the fact that f belongs to B,r
(M ),
I t
j0
X
X
p
|j | ]
2j[(p)+ 2 1]
[
j0
X
t p
Ct
This is true
2j[(+ 2 p ) q (q)] j
p
p j0 [(+ 21 p1 ) q (q)]
since + 21 p1
1
1
j0 p
q (+ 2 p )
If we now take 2
Zj
J
X
j=j0 +1
q
Ct pe
q )+ 2 1]
2j[(pe
j=j0 +1
Zj
q
|j |e
]
e
q
1
1 p
q )]
2j[(+ 2 p ) q (qe
zj
q )]
q j0 [(+ 2 p ) q (qe
Ct pe
2
1
Ct pq
2.
<p
,k()<2j1
<p
,k()<2j1
,k()2j1
,k()2j1
56
Zj
X
X
(
|j |p kj, kpp )1/p (
|j | kj, k )1/
If p, clearly
X
(
|j |p kj, kpp )1/p (
,k()<2j1
,k()<2j1
|j | kj, kp )1/
But
Hence,
X
(
,k()<2j1
,k()<2j1
2p1 {Ek
=: I + II
J
X
X
j=1 Zj
(t(j ) j )j kpp + k
XX
j>J Zj
j j kpp }
j>J
If p, if we put =
2
1+2 ,
Zj
j>J Zj
e s (M ),
using f B
p,r
II C2Jsp = Ctsp
+[
j>J ,k()2j1
:=
II() + II().
|j |p kj kpp )1/p ]p }
57
s
ep,r
Now, using (4), and f B
(M ),
II() C[
2js 2j2(+1)(1/1/p) ]p
j>J
2p1 J p1
J
X
Ek
j=1
2p1 J p1 C
p1 p1
J
X
Zj
j=1 Zj
(t(j ) j )j kpp
E|t(j ) j |p kj kpp
C[I() + I()]
j=1 ,k()<2j1
J
X
E|t(j ) j |p kj kpp
j=1 ,k()<2j1
J
X
j=1 ,k()<2j1
E|t(j ) j |p kj kpp I{|j | t j } + I{|j | < t j }
E|j j |p kj kpp I{|j | t j }
t j } + I{|j | < t j }
2
2
+|j |p kj kpp I{|j | t j } I{|j | 2t j } + I{|j | < 2t j }
I{|j |
: Bb + Bs + Sb + Ss
P i
= i i bj
is a gausi
i
P
)2 , we have using
sian random variable centered, and with variance 2 i ( bj
i
standard properties of the gaussian distribution, for any q > 0:
P
If we notice, as before, that j j = i
Yi bi fi i
j
bi
58
i
X j
)2 ]q/2 sq jq q C2jq q
(
b
i
i
r
2
1
P{|j j | log j } c /8
2
E|j j |q sq [2
Hence,
r
1
log j }
2
j=1 ,k()<2j1
r
J
X
X
1
p
p
|j | kj kp I{|j | < 2 log j }
j1
j=1
J
X
Bb
Ss
jp p kj kpp I{|j |
,k()<2
And,
r
1
2p 1/2
j=1 ,k()<2j1
r
1
kj kpp I{|j | < log j }
2
r
J
X
X
1
2p j c
kj kp I{|j | < log j }
2
j1
j=1
J
X
,k()<2
c p
J
X
/16
2jp
j=1
Zj
kj kpp
p 2 /16 J(p+(p/2)(p2)(1+))
using (26). Now, if we remark that the j are necessarily all bounded by some
s
ep,r
constant M , since f B
(M ),
Sb
J
X
j=1 ,k()<2j1
|j |p kj kpp P{|j j |
r
1
1
2 log j }I{|j | 2 log j }
r
J
X
X
2
1
|j |p kj kpp c /8 I{|j | 2 log j }
j1
j=1
,k()<2
/8
J
X
X
j=1 Zj
2
8
kj kpp
2J(p/2(p2)(1+))
59
It is easy to check that in any cases for 2 large enough, the terms Bs and
Sb are smaller than the rates announced in the two theorems.
Now we focus on the bounds of Bb and Ss. Let q [0, p], we always have:
p
J
X
j=1 ,k()<2j1
J
X
j=1
jp kj kpp I{
,k()<2j1
|j |
t }
j
2
jp kj kpp |j |q
(j t /2)q
p (t /2)q
J
X
j=1 ,k()<2j1
jpq kj kpp |j |q
And
J
X
j=1
,k()<2j1
J
X
(2t j )pq
j=1
(2t )pq
,k()<2j1
J
X
j=1 ,k()<2j1
|j |q kj kpp
jpq kj kpp |j |q .
J
X
j=1 ,k()<2j1
(t j )pq kj kpp |j |q ,
then we use (30) and we separate as before the bound obtained in two terms A
and B with some parameters j0 , z1 and z2 determined later, depending on the
cases:
A :=
j0
X
(t j )pz1 2j(p2)(+1)
j=1
B :=
J
X
(t j )pz2 2j(p2)(+1)
j=j0 +1
|j |z1 k (p2)(+1/2)
|j |z2 k (p2)(+1/2) .
,k()<2j1
,k()<2j1
60
j0
X
j=1
j0
X
(t j )p 2j(p2)(+1)
k (p2)(+1/2)
,k()<2j1
j=1
1
C(t j0 )p 2j0 (p/2)(p2)(+1) (log )I(p =0) .
And by treating B as was done previously with the term II(), we obtain:
B
J
X
j=j0 +1
2j(p2)(+1)
,k()<2j1
1
|j |p k (p2)(+1/2)
C2j0 p[s2(+1)( p )+ ] .
1/[s++ 12 ]
, which
2
)]
1/[s++(+1)(1
p
s++(+1)(1 2 )
A + B Ct
, which
1
(log )I(p =0) .
In the other cases: p < and (p2)(+1/2) > 1, or p > and (p2)(+1/2) <
(p2)(+1/2)1
, which satisfies:
1, let us set q = (2)(+1/2)1
pq =
2( + 1)( p)
,
( 2)( + 1/2) 1
and q =
( + 1/2)( p)
,
( 2)( + 1/2) 1
,k()<2j1
q
1 = ( 2)( + 1/2),
|j |q k (p2)(+1/2) . We define:
61
s
ep,r
Using H
older inequality, (30), and the fact that f B
(M ), we have:
X
|j |q k (p2)(+1/2)
,k()<2j1
Zj
|j |q k 1 k 2
X
,k()<2j1
C2jsqj
|j | k (2)(+1/2) ] [
2
q
k 1 ]1
,k()<2j1
q
(2)(+1)
2
q
]1
,k()<2j1
= C2j(p2)(+1) 2j(sq+
pq
2 )
j 1 .
(p 2)( + 1) sq
and
2
= 1.
q
A =
j=1
J
X
(t j )pq 2j(p2)(+1)
j=j0 +1
J
X
k (p2)(+1/2)
,k()<2j1
p j0 (p/2)(p2)(+1)
(t j0 ) 2
B
(t j )p 2j(p2)(+1)
,
,k()<2j1
(t j )pq 2j(sq+
C[
pq
2 )
|j |q k (p2)(+1/2)
]J 1
j=j0 +1
pq
2 )
q
1
(log )1 .
2
1/[s++(+1)(1
)]
A + B Ct
s2(+1)( 1 1 )
p
s++(+1)(1 2 )
q
1
(log )1 ,
, which yields:
62
, which yields:
s
p
s++ 1
2
q
1
(log )1 .
A + B Ct
j0
X
(t j )pq 2j(sq+
pq
2 )
]J 1
j=1
pq
2 )
q
1
(log )1 ,
J
X
j=j0 +1
C2
2j(p2)(+1)
,k()<2j1
1
1
p
)+ ]
j0 p[s2(+1)(
1/[s++ 21 ]
|j |p k (p2)(+1/2)
, which yields:
A + B Ct
s
s++ 1
2
q
1
(log )1 ,
2
1/[s++(+1)(1
)]
s2(+1)( 1 1 )
p
p
s++(+1)(1 2 )
A + B Ct
, which yields:
q
1
(log )1 .
j0
X
tpq
j 1 + C2j0 p[s2(+1)( p )+ ]
j=1
q
1
1
1
Ctpq
(log )2 + C2j0 p[s2(+1)( p )+ ] .
We proceed exactly like in the previous case, and we obtain the same rate
whether p or < p:
A + B Ct
s
s++ 1
2
q
1
(log )2 .
63
We can sum up all the results for Bb and Ss (and thus on I()) very simply:
if 2( + 1)( 1 p1 ) < s and s[1 (p 2)( + 1/2)] p(2 + 1)( + 1)( 1 p1 )
then:
1
1
p
Bb + Ss Ct
s+2(+1)( )
p
s++(+1)(1 2 )
1
(log )a ,
Bb + Ss Ct
s
s++ 1
2
1
(log )a ,
Let a() be as in 3.1 with the additional condition: a() > c > 0 for 3/4
7/4. Note that supp a [1/2, 2]. We define as in 3.1
X
,
j (x, y) =
a(k/2j ),
k (x)k (y).
k
64
.
2j
(33)
We set
Xj = { : = 1, 2, . . . , 2j }.
Let n denote the space of all polynomials of degree less than n. As is well
known [32] the zeros of the Jacobi polynomial P2j serve as knots of the Gaussian
quadrature which is exact for all polynomials from 2j+1 1 , that is,
Z
X
P d, =
bj, P ( ), P 2j+1 1 ,
I
Xj
where the coefficients bj, > 0 are the Christoffel numbers [32] and bj,
2j , (2j ; ) with
, (2j ; x) := (1 x + 22j )+1/2 (1 + x + 22j )+1/2 .
We now define the Jacobi needlets by
p
j, (x) = bj, 2j (x, ),
= 1, 2, . . . , 2j ; j 0,
Hence
kj, k2 1.
(34)
Notice that (34) cannot be an equality since otherwise the needlet system (j, )
would be an orthonormal basis and this is impossible since
Z
Xp
X
bj, j, =
bj, L2j (x, ) = L2j (x, y)d(x) = 0.
We now recall the two main results from [31] which will be essential steps in
our development.
Theorem 4. For any l 1 there exists a constant Cl > 0 such that
1
2j/2
|j, (cos )| Cl p
j
j
, (2 , cos ) (1 + 2 |
l ,
2j |)
0 .
(35)
65
Obviously
, (2j ; cos ) = (2 sin2 (/2) + 22j )+1/2 (2 cos2 (/2) + 22j )+1/2 .
(36)
1
2j(1+)
,
l (2j + 1)+1/2
(1 + 2j |
|)
2j
0 /2.
(37)
2j(1+)
1
,
l
j
j
(1 + 2 | 2j |) (2 ( ) + 1)+1/2
/2 . (38)
1/21/p
2j
.
, (2j ; )
j1
<2
kj, kp Cp
kj, kp Cp
2j(+1)
+1/2
12/p
2j(+1)
(2j + 1)+1/2
12/p
(39)
(40)
j1
, cp
2j(+1)
+1/2
12/p
k j, kp Cp
2j(+1)
+1/2
12/p
66
2j1 < 2j ,
12/p
12/p
2j(+1)
2j(+1)
cp
k j, kp Cp
(1 + (2j ))+1/2
(1 + (2j ))+1/2
or equivalently
1/21/p
2j
.
(41)
, (2j ; )
A critical role in the proof of this theorem will play the following proposition.
kj, kp
for
c N 1 c N 1 ,
N 2.
(42)
Proof. The proof will rely on the well known asymptotic representation of Jacobi
polynomials (sf. [32, Theorem 8.21.12, p. 195]): For any constants c > 0 and
>0
cos
sin
Pn, (cos )
2
2
(n + + 1) 1/2
= N
J (N ) + 1/2 O(n3/2 )
(43)
n!
sin
for cn1 , where N = n + ( + + 1)/2 and J is the Bessel function.
Further, using the well known asymptotic identity
1/2
2
J (z) =
cos(z + ) + O(z 3/2 ), z ( = /2 /4), (44)
z
one obtains (sf. [32, Theorem 8.21.13, p. 195])
1/2
1/2
cos
Pn, (cos ) = (n)1/2 sin
{cos(N +)+(n)1 O(1)}
2
2
(45)
for cn1 cn1 .
,
As is well known the Jacobi polynomials Pk, and Pk+1
have no common
zeros and hence it suffices to prove (42) only for sufficiently large N . Also,
Pk, (x) = (1)k Pk, (x) and therefore it suffices to prove (42) only in the
case c N 1 /2.
Denote by FN () the left-hand side quantity in (42). Then by (45), applied
with c = 1/2, it follows that
FN () N 1 21
2N
1
X
k=N
c N 1 21
2N
1
X
k=N
cos2 (k + h()) c 21 (N )2 ,
67
sin N
N
N
N
+
cos((3N 1) + 2h)
1
.
2
2 sin
2
2N
4
cos2 (k + h) =
k=N
Therefore,
21 (c /4 c (N )2 ) (c /8)21
c, (N ; ) for c N 1 /2,
FN ()
(46)
Choose so that =
FN () c1
2N
1
X
k=N
= c1 2 N
[J ((k + ))]2 c2 N 1
2N
1
X
k=N
= c1 2 N
N
1
X
j=0
1 h (k + ) i2
J
c2 N 1
N
N
i2
1 h j
J
c2 N 1 .
++
N
N
N
Obviously, the last sum above involves only values of the Bessel function J ()
for c c (2 + ) and hence uniformly in [c , c ]
N 1
1
NX
i2
1 h j
i2 X 1 h j
J
J
++
+
0,
N
N
N
N
N
j=0
j=0
N .
The
R 1 2second sum above can be viewed as a Riemann sum of the integral
J (( + 1))d, which is a continuous function of [c , c ] and hence
0
R1
min[c ,c ] 0 J2 (( + 1))d e
c > 0. Consequently, for sufficiently large N
FN ()
2 (e
cc1 N/2 cN 1 ) c2 N
c, (N ; )for c N 1 c N 1 .
From this and (46) it follows that (42) holds for sufficiently large N and this
completes the proof of Proposition 5.
68
1/2 ,
Proof of Theorem 6. We first note that (sf. [32]) ,
Pk (x) and
k (x) k
hence
X
2
kj, k22 = bj,
a2 (k/2j )(,
k (cos j, ))
2j2 <k<2j
c, (2j ; )
2j2 <k<2j
c, (2 ; )
7 j
3 j
4 2 k 4 2
Observe also that there exists a constant c > 0 such that c /2j j,
c /2j , = 1, 2, . . . , 2j . We now employ Proposition 5 and (34) to conclude
that
(47)
0 < c kj, k2 1.
We need to establish only the lower bound in Theorem 6. Recall first the
upper bound from Theorem 5
kj, kp Cp
1/21/p
2j
,
j
, (2 ; )
0 < p .
(48)
1/21/p
2j
.
j
, (2 ; )
(49)
The case p = is similar. In the case 0 < p < 2, we have using (47)
p
0 < c kj, k22 kj, kpp kj, k2p
ckj, kp
1p/2
2j
,
j
, (2 ; )
which implies (49). The lower bound estimates in Theorem 6 follow by (49).
9.3. Bounding for the norm of a linear combination of needlets
Our goal is to prove estimate (27), which we record in the following theorem:
Theorem 7. Let 0 < p < . There exists a constant Ap > 0 such that for any
collection of numbers { : = 1, 2, . . . , 2j }, j 0,
j
2
X
=1
j, kpLp (, )
Ap
2
X
=1
| |p kj, kpLp (, ) .
(50)
and if p = ,
k
2j
j, kL (, ) A sup | |kj, kL (, ) .
(51)
=1
69
j, k sup | |kj, k k
j,
k
kj, k
j,
k C
kj, k
p
, (2j , )
2j/2
and hence
k
p
X
, (2j , )
j,
1
p
k sup C
j , cos ) (1 + |2j |)l
kj, k
(2
p
, (2j , )
1
p
sup
j
j
, (2 , cos ) (1 + |2 |)l
X
1
sup
<
j |)l1/2
(1
+
|2
(52)
70
Lp (, )
X
1/r
Cp,r
|fk |r
Lp (, )
Using that M1 |f |s = (Ms f )s one easily infers from above that the following
p
}, then
maximal inequality holds: If 0 < p, r < and 0 < s < min{p, r, +1
for any sequence of functions (fk ) on [1, 1]
X
1/r
(Ms fk )r
Lp (, )
X
1/r
C
|fk |r
k
Lp (, )
(53)
with h =
1/2
2j
,
j
, (2 ; )
x [1, 1],
= 1, 2, . . . , 2j , j 0.
(54)
Obviously, (Ms 1I )(x) = 1I (x) for x I . Let x [1, 1]\ I and set cos = x,
[0, ]. Then
[(Ms 1I )(x)]s
1 +1
|I |
|x |
|x |
1
sin 2 (j,+1 j,1 ) sin 21 (j,+1 + j,1 )
sin 12 | j, | sin 21 ( + j, )
2j j,
.
| j, |( + j, )
Using that j, c 2j for some constant c > 0, one easily verifies the inequality
1
j,
.
1
+ j,
(2 + c )(1 + 2j | j, |)
(Ms 1I )(cos )
c
,
(1 + 2j | j, |)2/s
[0, ],
71
2
X
=1
j, kpLp (, ) c
2
X
=1
| |p kH kpLp (, ) .
(55)
1/p
Straightforward calculation show that k1I kLp (, ) 2j , (2j ; )
and
hence, using Theorem 6,
kH kLp (, )
1/21/p
2j
kj, kLp (, ) .
, (2j ; )
2
X
=1
2
X
Ap (
Ap kf kLp(, )
=1
(56)
=1
2 Z
X
j
2
X
=1
2
X
|j, (x)|
|f |(x){
kj, k kj, k1 }d(x) CC1 C kf k1
kj, k
=1
12j
sup
12j
|f |(x)|j, (x)|d(x)kj, kL (, )
p/q
|f |p |j, |dkj, k1
72
Hence
=1
|hf, j, i|
2 Z
X
j
2
X
kj, kpp
=1
p/q
|f | {
2
X
=1
p/q
|f |p {
2
X
|j, (x)|
p/q
p/q
kj, k kj, k1 kj, kpp }d(x) CC C1 Cpp kf kpp
k
k
j,
=1
p/q
p/q
j (f ) =
2
X
hf, j, ij,
=1
j
kj (f )kLp (, )
Cp (
2
X
=1
Cp kj1 f
+ j f + j+1 f kLp (, )
(by convention 1 (f ) = 0)
This claim is a simple consequence of the previous corollary, as
n u Xj , hf, j, i = hj1 f + j f + j+1 f, j, i
9.4. Besov Space
For f Lp (, ) let us define:
n N, En (f, p) = inf kf P kp
P n
Clearly
kf kp E0 (f, p) E1 (f, p)....
(57)
73
s
Definition 2. Besov space Bp,q
.
s
Let 1 p , 0 < q , 0 < s < . The space Bp,q
is defined as the
set of all the functions f Lp such that:
1
(ns En (f, p))q )1/q < , if q < ,
n
n=1
s
kf kBp,q
= kf kp + (
s
(resp. kf kBp,
= kf kp + sup ns En (f, p) < )
n1
s
As n En (f, p) is not increasing, we have an equivalent norm kf kBp,q
:
X
s
kf kBp,q
kf kp + ( (2js E2j (f, p))q )1/q , if q < ,
j=0
s
(resp. kf kBp,
= kf kp + sup 2js E2j (f, p))
j0
j (f ) =
2
X
hf, j, ij,
=1
then
1.
X
s
kf kBp,q
kf kp + ( (2js kj (f )kp )q )1/q
(58)
j=0
2.
s
If p < , kf kBp,q
2
X
X
kf kp + { (2js (
|hf, j, i|p kj, kpp )1/p )q }1/q
=1
j=0
(59)
X
(2js (sup |hf, j, i|kj, k ))q }1/q
j=0
(60)
m=j+1
74
km kp .
Therefore, if
km kp m 2ms , lq (N),
we have
E2j (f, p)
m 2ms = 2js
m 2|mj|s = j 2js ,
m=j+1
m=j+1
kj (f )kp
Cp (
2
X
=1
Cp (kj1 f kp
+ kj f kp + kj+1 f kp ).
X
j=0
js
(2 (
2
X
=1
s
(which is, by Theorem 7, obviously defined in Lp ) f Bp,q
even though not
necessarily
j, = hf, j, i.
Aj =
2
X
=1
j, j, , kAj kp Cp (
2
X
=1
75
References
[1] F. Abramovich and B. W. Silverman. Wavelet decomposition approaches to
statistical inverse problems. Biometrika, 85(1):115129, 1998. MR1627226
[2] Kenneth F. Andersen and Russel T. John. Weighted inequalities for vectorvalued maximal functions and singular integrals. Studia Math., 69(1):1931,
1980/81. MR0604351
[3] A. Antoniadis and J. Bigot. Poisson inverse models. 2004. Preprint Grenoble.
[4] A. Antoniadis, J. Fan, and I. Gijbels. A wavelet method for unfolding
sphere size distributions. The Canadian Journal of Statistics, 29:265290,
2001. MR1840708
[5] L. Cavalier, G. K. Golubev, D. Picard, and A. B. Tsybakov. Oracle inequalities for inverse problems. Ann. Statist., 30(3):843874, 2002. MR1922543
[6] Laurent Cavalier and Alexandre Tsybakov. Sharp adaptation for inverse
problems with random noise. Probab. Theory Related Fields, 123(3):323
354, 2002. MR1918537
[7] Albert Cohen, Marc Hoffmann, and Markus Rei. Adaptive wavelet
Galerkin methods for linear inverse problems. SIAM J. Numer. Anal.,
42(4):14791501 (electronic), 2004. MR2114287
[8] L.M. Cruz-Orive. Distribution-free estimation of sphere size distributions
from slabs showing overprojections and truncations, with a review of previous methods. J. Microscopy, 131:265290, 1983.
[9] V. Dicken and P. Maass. Wavelet-Galerkin methods for ill-posed problems.
J. Inverse Ill-Posed Probl., 4(3):203221, 1996. MR1401487
[10] David L. Donoho. Nonlinear solution of linear inverse problems by waveletvaguelette decomposition. Appl. Comput. Harmon. Anal., 2(2):101126,
1995. MR1325535
[11] R. J. Duffin and A. C. Schaeffer. A class of nonharmonic Fourier series.
Trans. Amer. Math. Soc., 72:341366, 1952. MR0047179
[12] Sam Efromovich and Vladimir Koltchinskii. On inverse problems with
unknown operators. IEEE Trans. Inform. Theory, 47(7):28762894, 2001.
MR1872847
[13] J. Fan and J.K. Koo. Wavelet deconvolution. IEEE Transactions on Information Theory, 48(3):734747, 2002. MR1889978
[14] C. Fefferman and E. M. Stein. Some maximal inequalities. Amer. J. Math.,
93:107115, 1971. MR0284802
[15] M. Frazier, B. Jawerth, and G. Weiss. Littlewood paley theory and the
study of functions spaces. CMBS, 79, 1991. AMS. MR1107300
[16] Alexander Goldenshluger and Sergei V. Pereverzev. On adaptive inverse
estimation of linear functionals in Hilbert scales. Bernoulli, 9(5):783807,
2003. MR2047686
[17] Iain M. Johnstone, Gerard Kerkyacharian, Dominique Picard, and Marc
Raimondo. Wavelet deconvolution in a periodic setting. J. R. Stat. Soc.
Ser. B Stat. Methodol., 66(3):547573, 2004. MR2088290
[18] Iain M. Johnstone and Bernard W. Silverman. Discretization effects in
76