on shape-preserving probabilistic wavelet approximators
TRANSCRIPT
![Page 1: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/1.jpg)
This article was downloaded by: [McMaster University]On: 18 November 2014, At: 08:18Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,37-41 Mortimer Street, London W1T 3JH, UK
Stochastic Analysis and ApplicationsPublication details, including instructions for authors and subscription information:http://www.tandfonline.com/loi/lsaa20
On shape-preserving probabilistic waveletapproximatorsLubomir Dechevsky a & Spiridon Penev ba Institute of Mathematics & Informatics , Technical University , Sofia, Bulgaria, 1156b School of Mathernatics Department of Statistics , The University of New South Wales ,Sydney, 2052, AustraliaPublished online: 03 Apr 2007.
To cite this article: Lubomir Dechevsky & Spiridon Penev (1997) On shape-preserving probabilistic wavelet approximators,Stochastic Analysis and Applications, 15:2, 187-215, DOI: 10.1080/07362999708809471
To link to this article: http://dx.doi.org/10.1080/07362999708809471
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) containedin the publications on our platform. However, Taylor & Francis, our agents, and our licensors make norepresentations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of theContent. Any opinions and views expressed in this publication are the opinions and views of the authors, andare not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon andshould be independently verified with primary sources of information. Taylor and Francis shall not be liable forany losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoeveror howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use ofthe Content.
This article may be used for research, teaching, and private study purposes. Any substantial or systematicreproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in anyform to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://www.tandfonline.com/page/terms-and-conditions
![Page 2: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/2.jpg)
STOCHASTIC ANALYSIS AND APPLICATIONS, 15(2), 187-2 15 (1997)
ON SHAPE-PRESERVING PROBABILISTIC WAVELET
APPROXIMATORS
Lubomir Dechevsky
Institute of Mathematics R- Informatics Technical Cniversity, Sofia
1156 Sofia, Bulgaria
Spiridon Penev
School of Mathematics Department of Statistics
The University of Kew South \Vales Sydney 2052. Australia
ABSTRACT \Ve introduce a general class of shape-preserving wavelet approsimating opera-
tors (approximators) which transform cumulative distribution functions and densi- ties into functions of the same type. Our operators can be considered as a general- ization of the operators introduced by Anastassiou and Yu [I]. Further, we extend the consideration by studying the approximation properties for the whole variety of L,-norms, 0 < p 5 m. In [l] the case p = o is discussed. Using the properties of integral moduli of smoothness, we obtain various approximation rates under no (or minimal) additional assumptions on the functions to be approximated. These assumptions are in terms of the function or its Riesz potential belonging to certain ho~nogeneous Besov. Triebel-Lizorkin. Sobolev spaces, the pace BT., of functions ~v i th bounded Wiener-Young p-variation, etc.
0. PRELIMISARIES
Copyright G 1997 by Marcel Dekker. Inc.
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 3: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/3.jpg)
188 DECHEVSKYANDPENEV
is well-known that L, is a quasi-Banach space for 0 < p < m ([2], [IS]) with em-
bedding constant in the quasi-triangle inequality c, = max{l,2$- ') : 1 1 f f glL,II < c,((l f iL,JJ -+ JlglL,ll). and, in particular. L p is a Banach space for 1 5 p 5 x?. Fur-
ther we denote by L1,i,, = { f : f l c € L1(C) for every compact C C R) xvhere f l c is the restriction o f f on C.
The spaces of lo call^) absolutely continuous functions are defined. as usual.
by .ACloc == {f E Ll,i ,c : 3 f ' ( the derivative of f ) Lebesgue almost o l e r \nhe re and
""j is the r/-th j' E I,l,i,l). AC = .AC(R) = {f E ACI,>, : f' E L1). f (" ' = - dxu
derivative of a univariate function. For a 2 O.[a] is the integer part of n. For
f E Ll,i,,.p E N, 0 < h < x. the Steklov's function (Steklov-means) j,,,,, is defined
by (see [12].[13]):
t .>c.
V f denotes the usual variation of a function f in j-x,fx). \ lk shall also use -
b
the general notation V f for the variation in [a. b ) . -x < (1 5 6 < x. E'os a
+ x
1 5 p < x? V f = I/p f is the IViener-hung variation of f . Denote ljy 131; = -= P
the space of functions with a bounded LYiener-Young variation 1;. Note tha t B l ;
is the space of functions xvith bounded variation in the usual sense.
It is easily seen tha t BV;!" BL;lq for 1 5 p 5 q 5 x. The space BT$'" is a
seminornied space, whileas BVp is a quasi-seminormed abelian group with imbedding
constant in the quasi-triangle inequality c, = Zp-l, 1 < p < x. Here and in the se-
quel "A (I B", when applied to quasi-seminormed spaces .A and B , has the meaning
of continuous imbedding (see, e.g. [2],[15]). For such A and B . A n B is the inter-
section of the two spaces with quasi-seminorm 11 f Id fi BII = n~ax{ll fi.-lii. l i flB11).
and A + B is the sum of the two spaces with quasi-seminorm
inf l l f I A + = f = f o + f , . f 0 E A . f l tB (IlfolAii + IlfllBll)
(see [2], [15]). For t E (0 ,cc) Peetre's Ii-functional (e.g., [I l l) is defined by Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 4: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/4.jpg)
PROBABILISTIC WAVELET APPROXIMATORS 189
The homogeneous Sobolev spaces are defined by ~@F(R) = {f E AClOc : 1 1 f l ~ ; l l =
Il.f('")lLPll < x) , 1 5 p 5 x, p E N. Their inhomogeneous analogs are defined by
\Ve introduce the homogeneous Besov and Triebel-Lizorkin spaces B;, and
F;,, s E R, 0 < p , q 5 x respectively, as well as their inhomogeneous analogs
B,&, F;,. Our notations are the same as in [2]. [15]. [lG]. For equivalent definitions
and properties (imbeddings, dual spaces, interpolation) we refer also to these. as
well as [ll]. [8]. Here we only make few useful notes for the sake of the reader's
orientation which cannot be explicitly found in the references but are collections
or implications of known facts to be found in these references. If the homogeneous . .
spaces B;,, FpSq are defined using Tl'iener-Paley's theory (via Peetre's function (see
121. [lj]) or Calderon's function (see [6]) as a basis for their atomic decomposition).
then a factorization is carried out modulo polynomials orthogonal to the concretely
chosen Calderon's function (with respective modification for Peetre's function). The
factor-spaces obtained are independent of the concrete choice of Calderon's or Pee-
tre's function and are quasi-Banach spaces (Banach spaces for 1 < minip. q ) < x).
It is often collvenient t o consider the elements of these spaces as functions rather
than as equivalent classes of functions and, for a fixed choice of Calderon's (Pee- . .
tre's) function. consider B;,. F;, to be (quasi) seminorrned. (Recall that a similar
situation arises with L,-spaces).
For s > mas{;. l } - 1. respectively s > max{ l P ' 1 . 1 ) - 9 1, the Besov. respectively
Triebel-Lizorkin spaces admit equivalent (quasi-)norms via finite differences and
functional moduli of smoothness. The above restrictions on s are essential and
are related to the fact tha t for these ranges of parameters the Besov and Triebel-
Lizorkin spaces are contained in Ll,i ,c. Outside these ranges the Besov and Triebel-
Lizorkin spaces contain generalized functions which are not regular. Other important
ranges of parameters are .c > b. respectively s > max{b, a ) . Then each element
(equivalence class) of the Besov, respectively Triebel-Lizorkin space is (contains) a
continuous function.
The inhomogeneous versions B;,, FpSq are quasi- Banach spaces (Banach spaces
for 1 < m i n { ~ , ~ ) < m). It is important tha t for s > maxi; , 1) - 1 : B,S, =
L , n B;,, F h ~9 LP n F,",. For p = q F;, = Bi9, F;q = B& (with equicalent quasi- norms). If p # q. s E R then the Besov and Triebel-Lizorkin spaces are essentially
diverse (see e.g. [ l j ] ) .
;In enormous variety of well-known function spaces can be identified as 110-
mogenous or inhomogeneous Besot or 'Triebel- Li~orkin spaces for specific values of
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 5: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/5.jpg)
190 DECHEVSKY AND PENEV
p,q and s . For an orientation we refer t o [6] and [ E l . Two (quasi-semi)normed
spaces are meant here to be identical if they are isomorphic (or. even simpler. coin-
cide as sets and have equivalent (quasi-semi)norms). Here we mention fea relevant
examples only:
11 E N. For p = 1: x' W F and are not Triebel-Lizorkin. or Besov spaces. The
same refers to L p . 0 < p 1 or p = x. For the definition ant1 relevant properties of
the Riesz potential I S f . s E R, we refer to [ 2 ] . \Ye note that it exists in any case for
f that may be of interest t o us. The following property will be of interest t o us (see
[2]. [IS]): IIIS f IL,II is an equivalent norm in F ; ~ for 1 < p < x. s 6 R. hloreolrer.
if o < q := cx: s , g E R then 1 1 1 " j ~ ; ~ ( l is an equivalent (quasi-)norm on F;:' for
0 < p < x, and I I I S ~ ~ B , " , / J is an e q u i d e n t (quasi-)norm on for 0 < p 5 x ( the lifting property).
Let h > 0, LL E -Y. Denote by ;,(f, h ) , . 1 5 p 5 x the iu t eg~a l p-modulus of
smoothncw of f . The latter is defined b j :
l e e A ( = f + t - f ) : A ( 2 ) = A A:-' f ( z ) ) . 1, = 2.3.
The moduli of smoothness (which for 11 = .x, are usually referred to as moduli of
continuity). are a basic error estimation tool in approximation theory. For some more
details see [12], [ 1 3 ] , [14]. Their properties related to the spaces BT.',. 1.i7, B,",: F;", etc. , will be used essentially later in the text. Here we only note that (see. e.g..
[ lo]? [14:1) for 1 I p < m xjw,(f,t), = ot( l ) , t -+ O+ if and only if f E L,. For
p = m :~,(f , t) , = o t ( l ) , t + O+ if and only if f is continuous. Besides (see [7]:
[12]. [13]), for 1 I p 5 m, Steklov's function f,,h is related to f by:
These bounds are useful for a variety of purposes. In particular, they allow a
very convenient and precise error estimation technique to be developed which, in
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 6: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/6.jpg)
PROBABILISTIC WAVELET APPROXIMATORS 191
particular, can be used in proving that the moduli of smoothness are equivalent t o
appropriate A'-functionals ( [ 7 ] ) .
Formally, the above definition of the integral moduli of smoothness for 1 5 p 5 rn can be also extended for the case 0 < p < 1 (see, e.g., [12]). However, for
0 < p < 1, the moduli so defined have much worse properties as an error estimation
tool. In particular, the inequalities about Steklov's function f,,), and its derivative
fir2 are generally not true if 0 < p < 1. It is important to mention here that the
quantity
is an equivalent (quasi)norm in B ; ~ if s > max (i. 1) - 1. na logous ly
l/q L + [ ( ( ) ) is an equivalent (quasi- )norm in B;q if s >
max {i, 1) - 1 (see [ l j ] ) . (Readers who use the source 1121 should take our remark
in consideration).
In view of the last observations for the case 0 < p < 1. we consider the following
definition of integral moduli which holds for 0 < p 5 x , p E N:
This definition. while preserving the useful properties of the previous definition
for 1 _< p 5 w. has an essential adbantage. To begin with. for 1 5 p 5 cc it is
equitalent t o the old one (see the proof of Theorem 2.1.1. ii) belou). Next. similar
to the prelious definition (see [15]), for ~ , ( f , 6 ) ~ it holds in case 0 < p = q < 1 I?
1. - 1 < s < p tha t [ s f p ) ] is an equivalent quasinorm in
F& = B,S, (which equally implies &,( f . S)? 5 c(p.p, S ) S ' f l ~ ; ~ j l for all f E B;,).
Finally, for this definition a version of the inequalities involving Steklov's func-
tion and its derivatives continues to hold true now also for the case 0 < p < 1
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 7: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/7.jpg)
192 DECHEVSKY AND PENEV
(compare with (3.25) and (3.25') in the sequel of the present paper.
We assume that the reader is familiar with the basic notions of wavelets theory.
For references on unconditional bases, atomic decompositions and wavelets. we refer
1. INTRODUCTION
This is the first paper of a sequence of several papers whose objective is as
follows:
1. To introduce a general class of shape-preserving wavelet approximating oper-
ators (approximators) which transform cumulative distribution functions and
densities into functions of the same type.
2. To analyze the approximation properties of these operators under minimal as-
sumptions about the regularity of the cumulative distribution function (cdf)/density.
3. To introduce, as data-dependent versions of these approximators, shape-preserving
wavelet estimators with an a priori prescribed smoothness properties.
4. To evaluate the risk of the estimators.
5. To analyze the asymptotic optimality of the estimator class considered
In the pioneering paper [l] the following operator is considered (modified nota-
tions):
where F is a cdf and pk, is defined via p as in (2.1.1). It is shown in [I] that , if 3 is
supported in [ -a ,a ] and satisfies (2.1.2-5), then the class of cdf's is invariant under
the action of Tk . It is proved further that for continuous cdf F
Moreover, if, additionally, p satisfies the condition
then the order of approximation increases from 1 to 2 and
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 8: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/8.jpg)
PROBABILISTIC WAVELET APPROXIMATORS 193
IITk(F) .- FILII I w ( F , 21-ka)m (1.3)
A model example of an admissible q satisfying (1.2) is also given there.
I t is remarkable that the explicit construction of the monotone operators Tk
produces orthogonal wavelets only in very restricted cases of the choice of +. In
particular, we observe that , because of the monotonicity, T k ( F ) cannot simultane-
ously be an orthogonal wavelet expansion and a continuous function. In other words,
it is possible to combine monotonicity of the wavelet approximation of a cdf with
certain regularity requirements about the image cdf (e.g., continuity or smoothness
of a certain order) only if the orthogonality of the wavelet is lost (and replaced by
almost-orthogonality).
In the present paper we develop the idea of using almost- orthogonal wavelets
further by considering (see (2.1.1))
where 9 and y do not necessarily coincide. We show that , if p satisfies (2.1.2-5) and
y satisfies (2.1.2) and (2.1.6), t h m Ah is cdf-preserving . Kote that the conditions
(2.1.2,6) are weaker than (2.1.2-5) because, as shown in [I], (2.1.3) implies (2.1.6).
In general, Ak is non-orthogonal, although (1.4) includes some orthogonal wavelet
operators, too. We extend the consideration to estimating ( (Ak(F) - FIL,(( , 0 < p 5 m. noting that all cases 0 < p 5 m are equally important for stochastic applications.
Our estimates for IIAk(F) - FIL,j are both an extension and an improvement of
(1.1) and (1.3)
IlAk(F) - FIL,II i cl(a)lla,,,l ~ , ( l w . l ( ~ , 2 ' - ~ a ) , t c2(a,p. y. $ ) 0 4 ( ~ . 2 ~ - ~ a ) , ,
(1 .5) for 1 5 p 5 m. and for 0 < p < ffi the estimate is the same, with an inessential
modification. Here
Note that the condition a,,, = 0 Lebesgue a.e. on R is an essentially weaker t 2c
condition than 3 = zi.. J ~ ( C - J ) = ( (which implies (see [I]). r g ( r ) d r = j=-w
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 9: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/9.jpg)
1 94 DECHEVSKYANDPENEV
The rich variety of properties of the moduli of smoothness implies various ap-
proximation rates under no (or minimal) additional assumptions about F which are formulated in corollaries 2.1.1-15. For example, under no additional assumpt,ions
(i.e., F being a cdf is the only assumption) we obtain that
-4dditional assumptions on F are in terms of F or its Riesz potential (derivative,
fractional derivative) belonging to certain well-known function spaces: the homoge-
nous Besov or Triebel-Lizorkin spaces, the Sobolev spaces (including the limit cases
p = l , ~ , when Sobolev spaces are not in the Triebel-Lizorkin scale), the spaces
BV, of functions with bounded Wiener-Young p-variation, etc.
As far as densities f are concerned, we show that , if both p and II, satisfy (2.1.2)
and (2.2.1), then Ak preserves simultaneously both cdf's and densities.
The estimation of / \Ak( f ) - flLpjj is carried out along the same lines as the one
of J IAk(F) - FIL,II. In particular, (1.5) holds true, with F replaced by f .
The corollaries (under additional assumptions) about f are similar to the ones
about a cdf. However, if no additional assumptions are made about f . then there
are essential differences in the rates of convergence (due to the different "natural"
properties of a cdf F and a density f - more precisely F E BVl, respectively f E L1).
Thus the analog of (1.6) is only (IAk(f) - flLlli = ok( l ) , k -i +w.
2. &LAIN RESULTS
2.1 Approximation of the CDF
Let F be a cumulative distribution function (cdf). Consider the wavelet oper-
ator "at the k-th resolution level"
+ m
< f , >= lw F(t)ykj( t )dt , where ykj(x) = 2k /2p(2kz - j) :
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 10: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/10.jpg)
PROBABILISTIC WAVELET APPROXIMATORS
+ 00 V p < m , p - right continuous - 05
There exists b E ( -a ,a ) such that cp is non- decreasing on (-m,b] and non-
increasing on [b, tee,) (2.1.5)
$ satisfies (2.1.2) ; $ E L1(R) and $(x)dx = 1 (2.1.6)
The operators A k ( F ) be the main objects of our investigation. They will be used
for approximating the cdf F. Operators of the same type when applied to a density
f , will be used later in Section (2.2), for density approximation. The function II, will
have to satisfy some additional conditions then.
The construction of the operator Ak(E) is inspired by the pioneering work of
Anastassiou and Yu [I] who consider the partial case 9 = $. Our generalization
makes it possible t o distinguish the essential differences of the roles of p and y in
the approximation process. In particular, $ may be selected in a much more general
class of functions than q. Note that (2.1.2,3) imply that p E L, , iipiLmll I: 1 and
that (see [I] p. 255) (2.1.3) implies v(x)dx = 1, but the opposite implication
is not true.
Denote E k ( F ) = A k ( F ) - F.
Our main result on estimating the rate of approximation of F by A k ( F ) is given
in the following
Theorem 2.1.1 Let i l k , ~ , y satisfy (2.1.1-6). Denote
i) For 1 5 p 5 cc, there exist constants cl > O , q > 0 such that
where cl = c l l ( 2 a j . c1 = c; . ua . 2 ' l p . a l / p , c ; , c ~ being absolute positive constants.
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 11: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/11.jpg)
196 DECHEVSKY AND PENEV
p' E 11, co] is such that $ + $ = 1
ii) For 0 < p < 1 there exist c 3 , ~ 4 , ~ 5 > 0 such that
where c3 = a-p .2-p(l+ 22p), q = va . ap. 2(1 f 23p), cs = 2aP(2P-1 + v,(l + 2 3 ~ ) ) .
Let us note that in view of the importance in applications and especially in view
of the future statistical applications (see Dechevsky and Penev [5]) the quasi-Banach
case 0 < p < 1 is also included in Theorem 2.1.1.
A set of important corollaries can be derived from the main result.
Corollarv 2.1 . I . Under the conditions of Theorem 2.1.1, let o,,+(C) = 0 V ( E W. Then
ii) For 0 < p < 1
i.e. the rate of approximation is evaluated from above by the second and third
integral modulus of smoothness. Therefore, one is interested in finding couples of
functions p and $ for which a,,$(x) 0 holds.
Remark 2.1.1. In [I] convergence of second order of the approximation was shown
under the assumption
Our Theorem 2.1.1. shows that (2.1.7) can be replaced by the essentially weaker
requirement
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 12: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/12.jpg)
PROBABILISTIC WAVELET APPROXIMATORS 197
Note that (2.1.8) is an essentially less restrictive condition than (2 .1 .7) even in the
special case y - Q. Indeed. (see [I], p. 264) the property (2.1.7) implies that + n
L r p ( r ) d r = r p ( r ) d r = 0 and (2.1.8) will hold for 9 5 Q. On the other 1: hand there exist y's for which (2.1.8) is true but not (2.1.7).
Examples illustrating this fact are given below.
Example 2.1 . I .
i) Case p ( x ) E $(x). Let us define
~ $ 1 - c o . x E ( - 1 t c o , c o l t co - a: , 2 E (co, 1 t co]
, elsewhere
where co # 0 with 0 < / co J < i. Now it is easily seen that 9 satisfies all requirements t o o
(2.1.2-5), ( - p ( j - () = co # 0 holds but , yet. (2.1.8) holds true and there is j=-CC
a second order convergence of the approximation.
i i ) Case p ( x ) + f:[xj. Define p as in i) and let
) = ( 1 2 E ( - 1 + co. 1 t co) 0 elsewhere
Then (2.1.8) is again fulfilled, and there is a second order convergence of the ap-
proximation.
Rcmark 2 . 1 . 2 The problem of explicitly finding all appropriate p and q) for which
(2.1.8) holds can be approached from the Fourier-transform side. This \\ill not be
discussed In details here.
R c r n c ~ ~ l : 2.1 .i \Ye note another case of G satisfying (2.1.6) which has interesting
extremal propeltles with respect t o the embedding constants in 'Theoren1 2.1 1 and
colollary 2.1.1. If A c [--a, a] is measurable, with Lebesgue measule /A1 = 1 ( a 2 $)
and cl = y,, the indicator function of the interl~al A. then apparently i.lL,iJl = + x
L , ! L * I I = lm V ( T ) ~ T = 1 for every p : 1 < p _< r: (even 0 < p 5 x' if we assume
the natural convention p' = x for 0 < p _< 1 j. Moreover, the inf { ~ I L ' L , ~ 1 : Q
satisfies (2.1.2.6) ) = 1, 0 < p 5 x is attained exactly for 9 = 1,. Since this
nlinirliizes the embedding constant, we have a hint for choosing suitable functions
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 13: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/13.jpg)
198 DECHEVSKYANDPENEV
$(z). Note also that choosing a discontinuous ~ ( x ) does not effect the smoothness
properties of l ik (F) , as long as we can take y # ~. Besides, example 2.1.1. ii) can be
easily modified to display a +(x) = x, (2) which, in combination with appropriate
(smooth or non- smooth) p(x) satisfies 2.1.8.
The following assertion holds:
Lemma 2.1.1. The operators Ak(F) are shape-preserving for cdf, i.e., if F is a cdf,
then Ak(F) is also a cdf.
Remark 2.1.1. It is worth noting that Lemma 2.1.1 is the only assertion which makes
use of the condition (2.1.5) for ~ ( x ) . We would also like to point out that our lemma
2.1.1. shows that the condition about continuity of the cdf F which is imposed in
[I] is in fact not necessary for the operators Ak(F) to be shape- preserving.
There follow important corollaries of Theorem 2.1.1 and of Corollary 2.1.1 which
are obtained by utilizing the properties of the moduli of smoothness. It is convenient
to divide these corollaries into two classes depending on whether IIu,,$IL,H = 0 or
not. In all corollaries F is understood to be a cdf. In corollaries 2.1.2-7 we assume
ll~v,*lLmll > 0. Corollaru 2.1.2. There exists c > 0 independent of F , such that ( ( E k ( F ) / L 1 / / < ~ . 2 - ~ .
Corollaru 2.1.3. Let 1 5 p < co. Then, there exists c > 0, independent of F, such
that
Corollaru 2.1.1. If F E .4C(R), i.e. F is absolutely continuous, and, additionally,
f = F' E L, , 1 < p 5 co, then there exists c > 0, independent of f . such that
Corollarz, 2.1.5. Let F E B,", , 1 5 p < co , 0 < s < 1. Then, there exists c > 0,
independent of F, such that
Corollarv 2.1.6. Let F E gip , < p < 1 , - 1 < s < 1. Then, there exists c > 0,
independent of F , such that
Remark 2.1.5. Since every cdf F E ~ ; k . 1 _< p 5 co, with no loss of generality we --
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 14: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/14.jpg)
PROBABILISTIC WAVELET APPROXIMATORS 199
may assume tha t instead of s : 0 < s < 1, only s : 0 < < s < 1 is required. This makes sense for 1 < p < m.
R ~ r n a r k 2.1.6. The imbedding B ; , ~ C B,",,~ , 0 < q1 5 43 5 X , shows tha t the
estimate in Corollary 2.1.5 is the best possible in the Besov scale. i.e. replacing
11. !B,S,!( by 11. ~ B ; J for some other q E ( 0 , x ) is true but would only make (2.1.10)
rougher. For the same reason estimates in terms of the inhomo~eneous Besob spaces
B,S, = L , n B,", C B,", , 1 < p < m. P > 0 are rougher. The same applies for the
case 0 < p < 1 where B i p = L p n B&, C B i p , s > - 1.
Corollarv 2.1.7. Let 1 < p 5 K , < s < 1 and assume tha t F E B,", . Then, for every q : p 5 q < x. there exists c independent of F. such tha t
Yext. in corollaries 2.1.8-15 u e assume lia,,,lL,li = 0 .
In this case corollaries 2.1.2-7 hold true with embedding constants which are
smaller than if llu,,,,lL,ll = 0 was not fulfilled
Cbrollary 2.1.8. In Corollary 2.1.5 s may be taken 0 < s < 2 (or, with no loss
of generality for cdf. $ < s < 2 ) . Moreover. for 1 5 s < 2, Inequality (2.1.10)
is equivalent t o the statement: there exists c > 0. independent of F. such tha t
! j E k ( F ) I L p ! ! 5 c.2-"!! f I B & ~ /I, where f = F' is the density.
Corollnru 1.1.9. In Corollary 2.1.6 one can take $ < p < 1 and $ - 1 < s < 2 .
Corollnru 2.1.10. In Corollary 2.l. ' i . if s = 2: (2.1.11) is modified to the statement
that there exists c > 0. independent of F , such that
~ h c r e f = 1:' is t,lle density.
C'orollnr.71 2 .1 .1 1. Let F E ~i~i. 1 < 11 < X. Then, for every q : p 5 q 5 x there
cxists r. > 0. independent of 6'. such that I E I ; ( F ) i L , l < c . 2 - " ' - ~ + ~ ' ~ 1 ~ i ~ ~ / l
(-'omllaru 2.1.12. Let I E 1Cr(IW). with density f = P'. Thcn. for cvcry 11 : I < p 5 x thcre exists c. i n d e p e ~ ~ d e n t of l'-. such that
IIEklFJILpIl
i'oivl1nr.11 2.1.13. Let I P F E U1.k . q : p < q < x there exists c > 0 such
< C . 2 - k 4 (f, 21-A[ l )y -
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 15: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/15.jpg)
200 DECHEVSKY AND PENEV
Corollaru 2.1.11. For 1 < p < ce , 5 s < 1. Let I S F E L1,EOc. Then, for every
q : p < q < w there exists c > 0, independent of F, such that
Corollaru 2.1.15. If F E AC(R), f = F' is the density and f E BVl, then there +03
exists c > 0, independent of F, such that ( I E k ( F ) ( L l j ( 5 ~ . 2 - ' ~ V f -co 1
2.2 Densitv Approximation
Let F be an absolutely continuous cdf, with a density f(x) = F 1 ( x ) . Consider
A k ( f ) defined as in (2.1.1). with F replaced by f . Function p is now supposed to
satisfy (2.1.2, 3) but with (2.1.3) loosened:
p(r - j ) E I Lebesgue a.e. x E R (2.2.1) +-.a
The function $ ( x ) is now supposed to have the same properties as p, i.e. (2.1.2)
and (2.2.1). These conditions imply that p, t,h E Ll n L,. Besides (see [ I ] ) , (2.2.1) +m
also implies l_ Theorem 2.2.1. Theorem 2.1.1 and Corollary 2.1.1 continue to hold true if the cdf -- F be replaced by a density function f . Corollaries 2.1.2-15 continue to hold true
with f instead of F, if all conditions of the type 5 s be replaced by 0 < s.
Lemma 2.2.1. Let p,$ satisfy (2.1.2) , (2.2.1) . Then A k ( f ) is a density. i f f is such. -- i.e., the operator Ak is shape-preserving for densities.
3. PROOFS
Proof of Theorem 2.1.1
i ) First consider the case 1 _< p _< co +oo
We consider the homogeneous Sobolev space I&':. Since F is a cdf, i.e. V F =
1 < co, it follows that F E L1 + I@: for every p E N. This means that
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 16: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/16.jpg)
PROBABILISTIC WAVELET APPROXIMATORS
where 1 1 . IL, + W:)) is any of the equivalent norms
or w.,( f , h),-the integral p-modulus of smoothness of f, h > 0. It is well known
that for F E L, f there exists the Steklov function Fu,hrp E -V, h > 0 (see
Petrushev and Popov [12, Section 31 and the Preliminaries for the explicit definition
of F,J,). This function has derivatives F::; , u = 1 , 2 . . . . , p and satisfies follouing
relations:
where c,,, are positive constants. There exist explicit estimates from above for
those constants (see [12], [13]). For a given cdf F we shall concentrate on the
Steklov functions Fk := F u , h for p = 2. h = 21-ka. Let us note that Fk may be. or
may be not a cdf. By llinkowski's inequality:
IIEk(F)ILpII I lIEk(F - J-~)lLpll + IIEkJFk)ILpII (3.3)
U e shall estimate from above each of the summands in the right hand side JRHS)
of (3.3). ITsing (2.1.3). we have
+cs
Hence. for any function f J x ) n e can write: f ( x ) = 2-"2pi,,[x) f ( n ). whence
+ s Since L ' ( T ) ~ T = I , one can ~vr i te 2-$ =< 1 , ~ ' ~ ~ > and a substitution in (3 .4 ) .i,, leads to
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 17: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/17.jpg)
202 DECHEVSKY AND PENEV
Application of the above equality for f(x) := F(x)-Fk(x) leads to following estimate
from above of the first summand in (3.3):
Note that we used the fact that the non-zero summands in the infinite sum are only those for which the summation index j (2% - a, 2% + a). Since these are a finite
number of summands, we can utilize for 1 5 p < m, p : $ f $ = 1 and for arbitrary
numbers a,, i = 1 ,2 , . . . , 12 the inequality
LVe can easily see that
Hence the RHS in (3 .6) can be further bounded from above by
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 18: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/18.jpg)
PROBABILISTIC WAVELET APPROXIMATORS 203
Further application of Holder's inequality and change of variables r = 2k t - j yields
Hence
Sote . however, tha t Ilpks ILmii = 2"' IlpiLmll.
.inother fact easy to observe is that
- Hence. if we introduce the constant t := V,P' . 2 k l p l w l ~ p , / . llplLmil, we can further
evaluate the RHS of (3.8) from above and by Fubini's theorem arrive a t
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 19: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/19.jpg)
204 DECHEVSKY AND PENEV
Our goal is to evaluate the bound from above in terms of the second integral modulus
of smoothness. Using the property
w i f , h ) p I2llflL,ll . h > 0
of the integral modulus and property (3.1) with p = 2. we obtain from (3 .9) for
f = F - Fk that
Now we start with the evaluation of the second summand in the RHS of (3.3). For
f = F k r similar to the estimate of IIEk(F - F k ) l L p J J , we have
(3.11)
Under the assumption that w z ( F , 21-ka)p < oo, it follows that f = Fk E L p , w;, 147;.
Hence f admits a local Taylor expansion of second order with an integral remainder,
i.e.
f ( t ) == f (x) + ( t - x ) f l ( x ) + ( t - x)' ( 1 - 0 ) f "(x + O(t - x))dO , x, t E R d' Hence
2-k(j+a) 2-" ( 2 - j ) )
2 - k ( 3 - a )
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 20: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/20.jpg)
PROBABILISTIC WAVELET APPROXIMATORS 205
(3.12) In the expression in the large brackets we change the variables by introducing T =
2k t - j . Utilizing $(r)d~ = 1 and the property (2.1.3) for y, this expression 1: can be easily simplified. For the second expression in (3.12) we use the fact that
Zkx - a < < < 2kx f a imposes that 1x - ti 5 21-ka in the range of integration of
t . This helps to further estimate this expression from above. Finally we obtain by
Fubini's theorem
J - a
Applying Minkowski's inequality in (3.11) and using (3.13) leads t o
But on applying (3.2) with v = 1,h = 21-k. a , we get:
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 21: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/21.jpg)
DECHEVSKYANDPENEV
ci > 0 being an absolute constant.
We note on passing that apparently
(3.16)
Further, on applying generalized Minkowski's inequality, we have
(3.17)
Repeating the same steps as in evaluating IIEk(F - Fk)ILplI after (3.6), the RHS of
(3.17) can be bounded from above by
By changing variables t into r = t - z and analogously to the estimation of the RHS
in (3.8) one can estimate (3.18) from above by
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 22: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/22.jpg)
PROBABILISTIC WAVELET APPROXIMATORS 207
Putting c* into (3.19), utilizing again inequality (3.2), the fact that f" = F[ and
$ + b = l , w e g e t
where i;; is an absolute constant.
From (3.10), (3.15) and (3.20) we get with ck = 2 + $ the final evaluation
ii) Kow we shall consider the case 0 < p < 1. The definition of the integral
p-modulus of smoothness given in the preliminary section for 0 < p < 1 can be
shown to be more general and to work equally well also for the case 1 5 p 5 co. To see this, let us consider
also for p E [I , m] and compare it to the usual definition in case p E [ I . m]:
On the one hand, L ~ , ( f , 6 ) ~ 5 u ~ , ( f , S ) ~ is obvious (take k = 1. X = 0. bl = p).
In the opposite direction, it is easily seen that w.,(f,6), 5 c;llflL,jl V f E L, and
j , ( f ,&) , 5 c;6,ll f l ~ p " l l V f E ~kp" hold wherefrom (u., 5 cG, follows easily from
the imbedding in the respective side in the equivalence
G,( f .6)p .- I i (6@, f : L,,$,")
(compare the preliminaries section).
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 23: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/23.jpg)
208 DECHEVSKY AND PENEV
Note, however, tha t for 0 < p < 1, w, and G, are not equivalent. As a version
of the generalized Minkowski inequality is unavailable here, w, is a much better
estimation tool for 0 < p < 1. In particular, the estimation technique using Steklov-
means intermediate approximation is available for w, but not for G,.
Steklov's function f,,h(z) is given in the Preliminaries. The properties (3.1)
and (3.2) are known to hold true for 1 < p 5 cc (see [12, Section 31). For 0 < p < 1.
(3.1) and (3.2) fail to hold for G, as defined in (3.23). Kext, we show that (3.1)
and (3.2) do have their analogs for 0 < p < 1, if w, is defined by (3.22). Let us see
first that for 0 < p < 1 for any Lebesgue measurable function f ( x ) and p E N there
exists a constant ~ ( p ) such that:
holds true. Changing the varialbe Q1 to cu = $ z:=l Ox and utilizing the inequality
we obtain
I PI I J1 I A : ~ ~ ( . ) I ~ ~ I L ~ I I I ~ w , ( i , h) , -1
which proves (3.24).
Next we shall show that for 0 < p < 1 and for w, as defined in (3.22) the following inequalities hold true:
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 24: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/24.jpg)
PROBABILISTIC WAVELET APPROXIMATORS 209
Indeed, for u = 1 , 2 , . . . , u - 1 the derivative of the Steklov function can be written
as ([12, p. 511:
Hence
P - 1 - ,!, I L - V
" 1 1 Ih . . J h / n ~ P - * ) h / P / ( ~ + - c 0n)idOl . . , d o u - u ~ ~ p ~ ~ ~ 5 ( : ) p (5) 0
X=O 0 I-" ,=I
5 h-"? (:)' (A) " [I, l h lh l ~ L - , ~ ) h , , f ( ~ ) l d Q ~ . . . ~ ~ p - u ~ L p l l p + X=O I-" -
Since '-*h 5 h and x:Ir 0 , 5 ( g - v ) h holds, changing variables 8, to I-"
p, = B,/h and evaluating both integrals in the last expression from above yields
from which (3.25) follows with
For Y = p . the derivative fby? is given by
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 25: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/25.jpg)
210 DECHEVSKYANDPENEV
and the proof of (3.25') is analogous to the one of (3.25) (but simpler).
Inequalities (3.24), (3.25) and (3.25') help us to proceed in case ii) (0 < p < 1)
along the same lines as in part i) of the proof. The same steps are repeated until one
obtains (3.7). Note, however, that in the case 0 < p < 1 the inequality laYl 5 p= 1
l l P ( a ) holds SO that in (3.7) we have the factor 1 instead of v:Ip'. The next
step in the evaluation is the following :
(3.28)
Changing the variable t to CY = a-12k-1(t - x) in (3.28), we arrive a t
Note, however, that for x E [2-k( j - a), 2-"(j + a)], it follows that
[ ( j - 2 k x - a ) / ( 2 a ) , ( j - 2 k x + a ) / ( 2 a ) ] E [-1,1].
Hence
Like in case i), we substitute f = F - Fk (Fk being the Steklov function Fp,t, for
p = 2, h = 21-ka) into (3.29) to get Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 26: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/26.jpg)
PROBABILISTIC WAVELET APPROXIMATORS 211
Proceeding like in the proof of (3 .25) , we obtain from the last inequality
l iEk(F)lLpl lP 5 IIEk(F - Fk)ILpllP $. IIEk(Fk)lLpllP ( 3 . 3 1 )
holds. We shall evaluate IIEk(Fk)ILpIIP from above.
Analogously to the respective evaluation in case i), we have
Ysing ( 3 . 2 5 ) (with the constant c, := c ( 2 , l , p ) = (1 +2?p) ' /p from ( 3 . 2 6 ) ) we obtain
( 3 . 3 2 )
Proceeding like in ( 3 . 2 8 ) , we have:
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 27: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/27.jpg)
212 DECHEVSKY AND PENEV
(3.33)
But applying (3.27) for p = 2, h = 21-ka, one can write
Hence the RHS of (3.33) can be further evaluated from above by the same method
we applied to get (3.29). Changing the variable to a = a-' . 2k-1(t - a ) we obtain
easily the following bound from above:
Note that we also used the property w,(F, 6 1 ) p 5 w,(F, 6")p , 6' 5 6" when deriving
(3.34).
Now it remains to insert the estimate from (3.34) into (3.32) to get:
Substitut,ing the last result and the result from (3.30) in (3.31), one finally gets the
estimate in part ii) of Theorem 2.1.1. Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 28: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/28.jpg)
PROBABILISTIC WAVELET APPROXIMATORS 213
Proof o f Lemma 2.1.1.
This assertion is proved in Anastassiou and Yu [I] for the case p (x ) = $(I). In
the case p(x) + $(x) the proof goes along the same lines, using properties (2.1.2) of
$(x) only. Since we are imposing the same conditions on the function p(x) as in [I],
the only condition to be checked is that for fixed resolution level k , the coefficients
{cj):=y, with
form a non-decreasing sequence. But this is an immediate consequence of the fact
that $(x) 2 0 has a compact support and that F ( x ) is non-decreasing, since cj is
* tco aIso equal to c j = 2--
a L ~ ( 2 - ~ ( t + j)$(t)dt. Let us note that the properties
lim Ak(F)(x) = 1 and lim Ak(F)(x) = 0 are an immediate consequence of the x-m 2--02
fact that if p = 1, then the right hand side in (3.21) is bounded for any cdf F (see
the proof of corollary 2.1.2 below).
Proof o f Corollarw 2.1.2. It is immediate consequence from theorem 2.1.1 and the -t-02
fact that any cdf F E BV*. Using the properties wl(F, h)l < V F and u k ( F , h)l 5 -m 1
2wkW1(F, h)l (see (12, Section 3.11) leads to the desired estimate.
Proof o f Corollaru 2.1.3. It follows from Corollary 2.1.2, from the fact that BVl c BV, for 1 < p < cc and from (v,F)lIp 5 VIF.
Proof o f Corollarv 2 . l . l . The proof utilizes the properties wl(F, h)p < hll fILpJl and
wk(F, h)p 5 2i3k-1(Fl h )p ,p E [ I , m] of the integral p-modulus of smoothness ([12],
Section 3.11)
Proof o f Corollary 2.1.5. It follows from the property wl(F, h), < EhSl( Fj ~ , & l l for
some positive constant E (see [15]).
Proof o f Corollarw 2.1.6. I t follows from the property dl(F, h), < ~ h ~ j l ~ ~ B ; , l l for
4 < p < 1, - 1 < s < 1 for some positive constant ? (see Preliminaries).
Proof o f Corollarv 2.1.7. This is a consequence of the embeddings B,S, c B;;,, for
Proof o f Corollaru 2.1.8. This follows from the fact that F E B;, with 1 5 s < 2
iff f E B,S;';'.
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 29: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/29.jpg)
214 DECHEVSKYANDPENEV
Proof of Corollaru 2.1.9, 2.1.10. 2.1.11. Omitted since similar to 2.1.6, 2.1.7.
Proof of Corollaru 2.1.1 2. Follows from the property w ~ ( F , h)p 5 hwk-l( f , h)p , k 2 2 ([12, Section 3.11).
Proof of Corollaru 2.1.13. By the equivalence of w,( f , 21-ka) to 1 < ( ( 2 l - ~ a ) ~ , f ; L,,
w;), p = 1,2, we obtain ((Ek(g)IL,II I c(lglLqlI for d l g E Lq, and llEdg)lL,ll I ~ ( 2 ~ - ~ a ) f i l l ~ j ~ [ l I for all g E w:. Recalling that (see Preliminaries) Lq = F$, W: =
k g , taking p = 2 and using the complex interpolation method (see [2],[15]), we
interpolate between the estimates and obtain J(Ek(g)(LqII 5 c(21-ka)s+~llg(~d$311,
I(Ek(g)IL,II 5 ~ ( 2 ' - ~ a ) ~ .1<(2'-~a, F; ~;?,k:Szfl) = S
Using the lifting property !(ISF~F:~!( ( ( F I ~ , " ~ + ~ I I (see Preliminaries) we obtain
and the proof follows from the respective property of the integral modulus.
Proof of (?orollaru 2.1.1.4. Follows from Corollary 2.1.13 and the imbedding theorem
about Triebel- Lizorkin spaces with different metric index ([15]).
Proof of Corollaru 2.1.15. Follows from Corollary 2.1.12 and the property wl( f , h ) l _<
h f. -03 1
Proof of Lemma 2.2.1. The condition Ak(f)(x) 2 0 almost everywhere is obvious
since y ( x ) , $(x) are non-negative. Further, on applying Fubini's theorem, we have
t m
Changing the variable x to ( = 2*z-j, applying Fubini again and using y(()d( =
1, one obtains
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014
![Page 30: On shape-preserving probabilistic wavelet approximators](https://reader035.vdocuments.us/reader035/viewer/2022073013/5750a7d01a28abcf0cc3dfd2/html5/thumbnails/30.jpg)
PROBABILISTIC WAVELET APPROXIMATORS
ACKNOWLEDGEMENT
This work was supported by the Australian Research Council. I t was initiated during a visit of the first author a t the Department of Statistics, University of New South Wales.
REFERENCES
1. G.A. Anastassiou, and X.M. Yu, Monotone and probabilistic wavelet approxi- mation, Stochastic analysis and applications, lO(3) (1992), 251-264.
2. J . Bergh, and J. Lofstrom, "Interpolation spaces. An introduction," Grundl. der Math. Wiss., 223, Springer, Berlin, Heidelberg, New York, 1976.
3. C.K. Chui, L'An Introduction to Wavelets," Academic Press, Boston, 1992.
4. I. Daubechies, "Ten Lectures on Wavelets," SIAM, Philadelphia, 1992.
5. L.T. Dechevsky, and S.I. Penev, On shape- preserving wavelet estimators of cumulative distribution functions and densities, preprint No. 11, Department of Statistics, School of Mathematics, The University of New South Wales (1993).
6. M. Frazier, B. Jawerth, and G. Weiss, "Littlewood-Paley theory and the study of function spaces," AMS, Providence, R.I., 1991.
7. H. Johnen, and K . Scherer, On the equivalence of the Ii-functional and moduli of continuity and some applications, in Tonstructive theory of functions of several variables, Oberwolfach '76" (W. Schempp, and K. Zeller, Eds), Lect. Notes in Math. 571, pp. 119-140, Springer, Berlin Heidelberg New York, 1977.
8. G. Kerkyacharian, and D. Picard, Density estimation in Besov spaces, Statistics €4 Probability Letters. 13 (1992), 15-24.
9. Y. Meyer, "Ondelettes e t op6rateurs I," Hermann, Paris, 1990.
10. SA4. Nikol'skii, "Approximation of functions of several variables and imbedding theorems," Springer, Berlin New York, 1975.
11. J . Peetre, "New thoughts on Besov spaces," Duke University Math. Series, Durham, 1976.
12. P.P. Petrushev, and V.A. Popov, "Rational approximation of real functions," Cambridge University Press, Cambridge, 1987.
13. B1. Sendov, and V.A. Popov, "Averaged moduli of smoothness," Bulgarian Academy of Science, Sofia, 1983.
14. A.F. Timan, "Theory of approximation of functions of a real variable", Hindus- tan Publ. Corp., Delhi, 1966.
15. H. Triebel, "Theory of function spaces." Monographs in Mathematics 78, Birk- hauser, Basel, 1983.
16. H. Triebel "Theory of function spaces 11," hlonographs in Mathematics 84, Birkhauser, Basel, 1992.
Dow
nloa
ded
by [
McM
aste
r U
nive
rsity
] at
08:
18 1
8 N
ovem
ber
2014