Wavelet bases in generalized Besov spaces

J. Math. Anal. Appl. 304 (2005) 198–211 www.elsevier.com/locate/jmaa Wavelet bases in generalized Besov spaces ✩ Alexandre Almeida Departamento de Matemática, Universidade de Aveiro, 3810-193 Aveiro, Portugal Received 22 June 2004 Available online 23 November 2004 Submitted by William F. Ames Abstract In this paper we obtain a wavelet representation in (inhomogeneous) Besov spaces of generalized smoothness via interpolation techniques. As consequence, we show that compactly supported wavelets of Daubechies type provide an unconditional Schauder basis in these spaces when the integrability parameters are finite.  2004 Elsevier Inc. All rights reserved. Keywords: Besov spaces; Generalized smoothness; Real interpolation; Wavelet representation; Wavelet bases 1. Introduction Wavelets have many applications into mathematics and other areas, such as engineering and physics. For instance, wavelet bases are used in the numerical resolution of some PDE’s with the advantage of providing fast and efficient algorithms. Concerning functions spaces, wavelet bases give us the possibility of describing their elements in terms of basic and simple “building blocks.” In general, an important point is that we can characterize the original (quasi-)norm by means of certain sums involving the wavelet coefficients. On the other hand, wavelet bases can be quite useful to study some intrinsic questions related to ✩ This research was partially supported by Unidade de Investigação “Matemática e Aplicações” of Universidade de Aveiro, through Programa Operacional “Ciência, Tecnologia e Inovação” (POCTI) of the Fundação para a Ciência e a Tecnologia (FCT), co-financed by the European Community fund FEDER. E-mail address: [email protected]. 0022-247X/$ – see front matter  2004 Elsevier Inc. All rights reserved. doi:10.1016/j.jmaa.2004.09.017 A. Almeida / J. Math. Anal. Appl. 304 (2005) 198–211 199 functions spaces. Recently, for example, they were successfully used to estimate entropy numbers of compact embeddings between weighted spaces (see [12] for details). Motivated by a recent work of Triebel on wavelet bases in function spaces, we deal with wavelet representations in Besov spaces with generalized smoothness. In [16] it was proved, in particular, that compactly supported wavelets of Daubechies type form an uns . Our main aim is to extend conditional Schauder basis in the “classic” Besov spaces Bpq φ this result to the “generalized” Besov spaces Bpq (cf. Definition 2), showing that the same wavelet system also provides an unconditional Schauder basis in these spaces. We would like to remark that function spaces of generalized smoothness have applications in other fields such as probability theory and stochastic processes (see [7]). We realize that it is possible to get the result without repeating the approach suggested in [16]. Hence, instead of making use of all that powerful tools (atomic decompositions, local means, maximal functions, duality theory), we try mainly to take advantage of the classic case by means of suitable interpolation techniques. We would like to remark that interpolation tools were recently used by Caetano (see [3]) in order to get subatomic representations of Bessel potential spaces modelled on Lorentz spaces from the corresponding ones for the usual spaces Hps . As long as wavelet bases literature is concerned, we refer to [6,14,22] and, of course, to [16] as well as to other references therein. Close to this matter we also mention [10], where wavelet decompositions of Besov spaces were studied in a multiresolution analysis framework. This paper is structured as follows. In Section 2 we give the definition of Besov spaces of generalized smoothness and compare them to other well-known function spaces. In this section we also discuss some interpolation properties which will play a key role later on. Section 3 is devoted to the wavelet representation of Besov spaces. For convenience, we contextualize the problem recalling what is already done in the “classic” case, and then we formulate our main result as well as some of its consequences. We will follow standard notation or it will be properly introduced whenever needed. 2. Generalized Besov spaces 2.1. Preliminaries We shall consider standard notation as follows. Let Rn be the n-dimensional Euclidean space and Zn the usual lattice of all points with integers components (n ∈ N). For 0 < p
∞, Lp (Rn ) denotes the well-known quasi-Banach space with respect to the Lebesgue measure, quasi-normed by  1/p  
  n 
f (x)p dx
f  L p R
= , Rn with the usual modification if p = ∞. Let C(Rn ) be the space of all complex-valued uniformly continuous bounded functions in Rn and let, for r ∈ N,        C r Rn = f ∈ C Rn : D α f ∈ C Rn , |α|
r , (1) 200 A. Almeida / J. Math. Anal. Appl. 304 (2005) 198–211 normed by
 r  n 

f  C R
= |α|
r S(Rn )
α   

D f  L ∞ R n
. By we denote the Schwartz space of all rapidly decreasing and infinitely differentiable functions on Rn , and by S ′ (Rn ) its topological dual, that is, the space of all tempered distributions. If ϕ ∈ S(Rn ), then Fϕ (or ϕ̂) stands for the Fourier transform of ϕ,  (Fϕ)(ξ ) = (2π)−n/2 e−ixξ ϕ(x) dx, ξ ∈ Rn , (2) Rn F −1 ϕ ϕ∨) whereas (or denotes its inverse Fourier transform, given by the right-hand side of (2) with i in place of −i. Both the Fourier transform and its inverse are extended to S ′ (Rn ) in the usual way. Let ϕ0 ∈ S(Rn ) be such that  ϕ0 (x) = 1 if |x|
1 and supp ϕ0 ⊂ x ∈ Rn : |x|
2 . (3) Putting   and ϕj (x) := ϕ1 2−j +1 x , ϕ1 (x) := ϕ0 (x/2) − ϕ0 (x) then and  supp ϕj ⊂ x ∈ Rn : 2j −1
|x|
2j +1 , x ∈ Rn , j ∈ N, (4) j ∈ N, ∞ ϕj (x) = 1, x ∈ Rn . j =0 Hence {ϕj }j ∈N0 forms a dyadic smooth resolution of unity. We recall that, for s ∈ R, 0 < p
∞, 0 < q
∞, the usual Besov and Triebel–Lizorkin spaces are defined as the collection of all f ∈ S ′ (Rn ) such that 1/q ∞ and
 s  n 

f  B R
:= pq

 s  n 


f  F R
:=

pq
2 j =0   
q
(ϕj fˆ)∨  Lp Rn
j sq
1/q ∞ 2 j =0 j sq   q (ϕj fˆ) (·) ∨
   n
 Lp R


(5) (6) (with the usual modification if q = ∞ and p < ∞ in the F -case) are finite, respectively. They are quasi-Banach spaces and are independent of the system {ϕj }j ∈N0 chosen according to (3) and (4) (with equivalent quasi-norms). We refer to [18] for a systematic theory on these spaces. It is well known that these scales contain some classic spaces as special cases. For instance,  n   s R = Hps Rn , s ∈ R, 1 < p < ∞, Fp,2 A. Almeida / J. Math. Anal. Appl. 304 (2005) 198–211 201 are the fractional Sobolev spaces (they are the classic Sobolev spaces when s ∈ N0 ) and  n   0 R = hp Rn , 0 < p < ∞, (7) Fp,2 are the local (or inhomogeneous) Hardy spaces introduced by Goldberg (see [11]). In the sequel, we need to deal with some sequence spaces into a general context as follows. Let E be a quasi-normed space, I a countable set and 0 < q
∞. We denote by ℓq (I, E) the “sequence” spaces of all E-valued families a ≡ {ai }i∈I such that a | ℓq (I, E) is finite, where  1/q

q
a | ℓq (I, E)
:= , 0 < q < ∞, (8) ai | E i∈I and


a | ℓ∞ (I, E)
:= sup ai | E (9) i∈I define quasi-norms. If the set I is clear from the context, we shall omit it. Besides, we may omit E from the notation if E = C. 2.2. Definition and basic properties Roughly speaking, we obtain Besov spaces of generalized smoothness replacing the usual regularity index s in (5) by a certain function with given properties. We consider a sufficiently wide class of such functions, which allows us to cover many cases in the literature. Definition 1. We say that a function φ : (0, ∞) → (0, ∞) belongs to the class B if it is continuous, φ(1) = 1, and φ(ts) < ∞, s>0 φ(s) φ̄(t) := sup t ∈ (0, ∞). We refer to [5,13] for more details concerning this class. For a function φ ∈ B, the Boyd upper and lower indices αφ̄ and βφ̄ are then well-defined, respectively, by αφ̄ = lim t→+∞ log φ̄(t) log t log φ̄(t) t→0 log t and βφ̄ = lim with −∞ < βφ̄
αφ̄ < +∞. φ If E is a quasi-normed space, 0 < q
∞ and φ ∈ B, one can consider the spaces ℓq (E) of all sequences {aj }j ∈N0 such that {φ(2j ) aj }j ∈N0 ∈ ℓq (E), equipped with the quasinorms  · | ℓq (E) according to (8) and (9) (with I = N0 ). When φ(t) = t s , t ∈ (0, ∞), φ s ∈ R, we simply write ℓsq (E) instead of ℓq (E) for short. n Let {ϕj }j ∈N0 ⊂ S(R ) be a system with the following properties:  supp ϕ0 ⊂ ξ ∈ Rn : |ξ |
2 ; (10)  n j −1 j +1
|ξ |
2 , j ∈ N; (11) supp ϕj ⊂ ξ ∈ R : 2 202 A. Almeida / J. Math. Anal. Appl. 304 (2005) 198–211   sup D α ϕj (ξ )
cα 2−j |α| , ξ ∈Rn j ∈ N0 , α ∈ Nn0 ; (12) ∞ ϕj (ξ ) = 1, ξ ∈ Rn . (13) j =0 Definition 2. Let {ϕj }j ∈N0 be a dyadic resolution of unity with the properties (10)–(13) φ above. For φ ∈ B, 0 < p
∞, and 0 < q
∞, we define Bpq (Rn ) as being the class of φ all f ∈ S ′ (Rn ) such that {(ϕj fˆ)∨ }j ∈N0 ∈ ℓq (Lp (Rn )) with  φ   n 

 φ  n 

  ℓ L p R
.
f  B R
:=
(ϕj fˆ)∨ q pq j ∈N 0 These spaces were studied by Merucci (see [13]) as a result of real interpolation with function parameter between Sobolev spaces and then by Cobos and Fernandez in [5]. Such as in the classic case according to (5), they are quasi-Banach spaces and are independent of the system {ϕj }j ∈N0 chosen, up to equivalent quasi-norms. We point out that the spaces s (Rn ) can be obtained as a particular case of the spaces B φ (Rn ) by taking φ(t) = t s , Bpq pq t ∈ (0, ∞), s ∈ R. In general, we are only dealing with functions spaces on Rn . Hence, from now on, we s as shall omit the Rn in their notation. For convenience, we will refer to the spaces Bpq classic Besov spaces. Besov spaces with generalized smoothness have been considered and studied by many authors in different contexts. We refer to the paper [7] for historical remarks and literature concerning this subject. In [7] we can also find a general and unified approach for these spaces, as well as the counterpart for the Triebel–Lizorkin scale. As far as Besov spaces σ by replacing φ(2j ) by σ , are concerned, it is possible to define generalized spaces Bpq j j ∈ N0 , in Definition 2, where σ is a certain admissible sequence of positive real numbers in the sense of [7]:



   ℓq (Lp )
< ∞ , (14) B σ = f ∈ S ′ :
f  B σ
:=
σj (ϕj fˆ)∨ pq pq j ∈N0 where σ ≡ {σj }j ∈N0 satisfies the condition d0 σj
σj +1
d1 σj , ∀j ∈ N0 , (15) for some d0 , d1 > 0. The definition given in [7] is even more general: it is introduced a fourth parameter N ≡ {Nj }j ∈N0 related to generalized resolutions of unity, namely, allowing different sizes for the support of the involved functions. We restrict ourselves here to the standard decomposition, that is, with N = {2j }j ∈N0 . Some “other” generalized spaces of Besov type were introduced by Edmunds and (s,Ψ ) Triebel. They are usually denoted by Bpq and are defined as in (5) with 2j sq Ψ (2−j )q j sq in place of 2 . The parameter Ψ here represents a perturbation on the smoothness index s, and, of course, it fulfills certain conditions. We refer to [15] for a systematic study (s,Ψ ) on spaces Bpq . (s,Ψ ) As it was remarked in [7], the spaces Bpq are covered by the general formulation (14), by taking σj = 2j s Ψ (2−j ), j ∈ N0 . Since we have φ̄(1/2)−1 φ(2j )
φ(2j +1 )
φ φ̄(2)φ(2j ), j ∈ N0 , the spaces Bpq , φ ∈ B, are also a particular case of the spaces defined A. Almeida / J. Math. Anal. Appl. 304 (2005) 198–211 203 φ in (14). However, we would like to point out that is enough to consider the spaces Bpq . This fact may be justified by the following result, which was suggested to us by Caetano. Proposition 3. Let σ be an admissible sequence in the sense of (15) and 0 < p, q
∞. Then there exists a function φσ ∈ B such that σ φσ = Bpq . Bpq Proof. Let σ be admissible. First, we remark that one can always assume σ0 = 1 without loss the generality. In fact, the sequence σ ′ defined as σ0′ = 1 and σj′ = σj , j ∈ N, is ′ σ = Bσ . equivalent to σ , so Bpq pq We can construct a function φσ ∈ B as follows: φσ (t) = σj +1 −σj 2j σ0 , (t − 2j ) + σj , t ∈ [2j , 2j +1 ), j ∈ N0 , t ∈ (0, 1) (cf. [4, Section 2.2]). Hence, φσ (2j ) = σj for all j ∈ N0 and we get the result. ✷ Taking into account this proposition, from now on we will only deal with Besov spaces from Definition 2. Such as in the classic case (cf. [18, pp. 47, 48]) one proves the following φ embeddings related to the spaces Bpq . Proposition 4. (i) Let φ ∈ B, 0 < p
∞, 0 < q
∞. Then  n     φ R ֒→ S ′ Rn . S Rn ֒→ Bpq (ii) Let φ ∈ B, 0 < p
∞, 0 < q0
q1
∞. Then  n  n φ φ R . R ֒→ Bpq Bpq 1 0  φ(2j ) (iii) Let φ, ψ ∈ B, 0 < p
∞, 0 < q0 , q1
∞. If ψ(2 j)     φ ψ Rn . Rn ֒→ Bpq Bpq 1 0 j ∈N0 ∈ ℓmin{q1 , 1} , then As usually, the symbol “֒→” above indicates that the corresponding embedding is continuous. Property (iii) is important, in particular, to derive Lemma 6 bellow. 2.3. Interpolation with function parameter φ As it was referred before, the spaces Bpq , φ ∈ B, can be obtained from real interpolation between Sobolev spaces with an appropriate function parameter. Interpolation of this kind fits well into these generalized Besov spaces framework if the function parameter belongs to the same class B. We refer to the papers [5,13] for the notation and basic properties concerning interpolation. In [5] several interpolation results were obtained for the φ spaces Bpq in the Banach case (1
p, q
∞). The approach followed there was based on interpolation properties of sequence spaces. Those properties were then transferred to 204 A. Almeida / J. Math. Anal. Appl. 304 (2005) 198–211 φ the spaces Bpq , by means of the so-called method of retraction and co-retraction (cf. [1, p. 150] and [19, p. 22], for example). Briefly, let E be a quasi-Banach space, φ ∈ B and 0 < q0 , q1 , q
∞. Taking into account [5, Theorem 5.1 and Remark 5.4], one can write  s  ℓq00 (E), ℓsq11 (E) γ ,q = ℓφq (E) (16) if s0 , s1 ∈ R with s1 < βφ̄
αφ̄ < s0 and s0 t s0 −s1 γ (t) =  1  , φ t s0 −s1 t ∈ (0, ∞). φ (17) φ It is possible to show that Bpq is a retract of ℓq (Lp ) if p  1 by constructing certain applications (retraction and co-retractions) based on the Fourier transform. But this does not work if 0 < p < 1. However, as it was remarked in [5, Remark 5.4], some of the interpolation results obtained hold in the quasi-Banach case as well. We do not intend to go into too many details, but we give here a brief description how this question in the general case can be dealt with. Following [17, Theorem 2.2.10], one can prove the result bellow. Proposition 5. Let f ∈ S ′ , 0 < p < ∞ and {ϕj }j ∈N0 satisfying the conditions (10)–(13). Then F −1 (ϕj Ff ) ∈ Lp if and only if F −1 (ϕj Ff ) ∈ hp , j ∈ N0 . Moreover, there are constants c1 , c2 > 0 independent of f and j such that 





c1
F −1 (ϕj Ff )  Lp


F −1 (ϕj Ff )  hp

c2
F −1 (ϕj Ff )  Lp
. Note that hp is the local Hardy space from (7). Using these estimates, one can replace Lp by hp in Definition 2 when p < ∞ (note that hp = Lp if 1 < p < ∞). With this change, one avoids the mentioned troubles caused by the Fourier transform. On the other φ φ hand, we can prove that Bpq is a retract of ℓq (hp ): if {ϕj }j ∈N0 ⊂ S is a system with the properties (10)–(13), then ∞ 1 F −1 (ϕ̃j Ffj ), R{fj }j ∈N0 := j =0 φ with ϕ̃j = ϕj +r , r=−1 φ is a retraction from ℓq (hp ) to Bpq and Sf := {F −1 (ϕj Ff )}j ∈N0 is the corresponding coretraction. We remark that R is well-defined with the help of the following lemma, which can be proved following similar techniques as in [23, Theorem 3.6], Lemma 6. Let φ ∈ B, 0 < p
1, 0 < q
∞. Assume that {gj }j ∈N0 ⊂ S ′ fulfills the conditions   supp Fg0 ⊂ x: |x|
2 and supp Fgj ⊂ x: 2j −1
|x|
2j +1 , j ∈ N. If φ(2j )gj | ℓq (hp ) < ∞, then ∞  gj converges in S ′ . j =0 Hence, taking into account the remarks above, it is possible to get the result bellow, which will play a crucial role in next section. A. Almeida / J. Math. Anal. Appl. 304 (2005) 198–211 205 Proposition 7. Let φ ∈ B, 0 < p
∞, and 0 < q0 , q1 , q
∞. Assume s0 , s1 ∈ R satisfy s1 < βφ̄
αφ̄ < s0 and γ as in (17). Then  s  φ 0 , B s1 Bpq pq1 γ ,q = Bpq . 0 (s,Ψ ) Proposition 7 shows, in particular, that spaces Bpq mentioned in Section 2.2 can be obtained by interpolation of classic Besov spaces with a suitable function parameter. This fact was already observed in [2]. 3. Wavelet representation of Besov spaces The aim of this section is to obtain wavelet representations for the generalized Besov spaces under consideration. We will make use of the system considered in [16] and follow the same notation. Let Lj = L = 2n − 1 if j ∈ N and L0 = 1. It is known that, for any r ∈ N, there are real compactly supported functions ψ0 ∈ C r , ψ l ∈ Cr , l = 1, . . . , L, (18) α ∈ Nn0 , |α|
r, (19) with  x α ψ l (x) dx = 0, Rn such that  j n/2 l ψj m : j ∈ N0 , 1
l
Lj , m ∈ Zn 2 (20) with ψjl m (·) = ψ0 (· − m), ψ l (2j −1 · − m), j = 0, m ∈ Zn , l = 1, j ∈ N, m ∈ Zn , 1
l
L, (21) is an orthonormal basis in L2 . As mentioned in [16], an example of such a system of functions is the (inhomogeneous) Daubechies wavelet basis (see, for example, [6,14,22] for further information). Wavelets with the properties above are sufficiently good to provide unconditional bases in many classical spaces. For instance, it was known that the mentioned Daubechies system forms an unconditional Schauder basis in the Sobolev spaces Hps if 1 < p < ∞, r > |s|, s if 1
p, q < ∞, r > |s|. These two examples show, in and in the Besov spaces Bpq particular, that the smoothness required on the wavelets in (18) should be large enough, depending on the regularity of the functions that we pretend to represent. This fact can also be observed in the sequel. 3.1. The classic case The main aim in [16] was to extend the results above about Sobolev spaces and some s and F s . For convenience, we recall here the main Besov spaces to the entire scales Bpq pq 206 A. Almeida / J. Math. Anal. Appl. 304 (2005) 198–211 result related to Besov spaces. Let I = {(l, j, m): j ∈ N0 , 1
l
Lj , m ∈ Zn } and I ′ = {(l, j ): j ∈ N0 , 1
l
Lj }. Theorem 8. Let s ∈ R, 0 < p
∞, 0 < q
∞, and   2n n + −s . r(s, p) := max s, p 2 (22) s if and only if it can be (i) Assume r ∈ N with r > r(s, p) and let f ∈ S ′ . Then f ∈ Bpq represented as λlj m ψjl m f= s with λ ∈ bpq , (23) (l,j,m)∈I t if t < s. Moreover, the repreunconditional convergence in S ′ and in any space Bpu sentation (23) is unique:   (24) λ = λ(f ) with λlj m (f ) := 2j n f, ψjl m . s onto bs Furthermore, f → {2j n f, ψjl m }(l,j,m)∈I defines an isomorphic map of Bpq pq and 

 s


f  B
∼
λ(f )  bs
(25) pq pq (equivalent quasi-norms). s (ii) In addition, if max(p, q) < ∞, then (23) with (24) converges unconditionally in Bpq l s and {ψj m }(l,j,m)∈I is an unconditional Schauder basis in Bpq . s is the space of all complex-valued sequences λ ≡ {λl } Here, bpq j m (l,j,m)∈I such that q/p 1/q    l p
 s
λ 
λ  b
:= < ∞, 2j (s−n/p)q pq jm (l,j )∈I ′ m∈Zn with standard modifications if p = ∞ and/or q = ∞. Remark 9. This result was extended to weighted function spaces of Besov and Triebel– Lizorkin type very recently in the paper [12]. s , ψ l ∈ C r , should be properly interRemark 10. The symbol f, ψjl m in (24), f ∈ Bpq jm l preted since the functions ψj m are not in S in general. As remarked in [16], it makes sense if r > −s + σp (with σp := max(n/p − n, 0)), which is covered by condition r > r(s, p). In fact, in that case, one can interpret f as an element of the dual of a space to which ψjl m belongs to. 3.2. The generalized case The proof of Theorem 8 was based on atomic decompositions, characterizations by local means, and duality theory (we refer to [20,21] for details on these properties). An A. Almeida / J. Math. Anal. Appl. 304 (2005) 198–211 207 important point there was that the wavelets considered were simultaneously atoms and kernels of those local means. It was also commented in [16] the possibility of getting a similar result in the context of other scales of function spaces. To do that, it would be enough to have the same tools available. However, as we mentioned before, we will not follow this approach. Instead, we will consider a scheme based on interpolation techniques in order to take advantage of the already known wavelet decompositions for the classic case. Let φ ∈ B, 0 < p
∞, 0 < q
∞. For our purposes we need to introduce the seφ quence spaces bpq , consisting of all complex-valued sequences λ ≡ {λlj m }(l,j,m)∈I such that the quasi-norm    
 φ
 l p q/p 1/q   j  −j n/p q
λ  b
:=   λj m φ 2 2 (26) pq (l,j )∈I ′ m∈Zn (with the usual modifications if p = ∞ and/or q = ∞) is finite. When φ(t) = t s , t ∈ φ s defined in [16]. We would like to (0, ∞), s ∈ R, then bpq coincides with the space bpq remark that sequence spaces with this structure were introduced by Frazier and Jawerth in [8,9] in connection with atomic decompositions of (classic) Besov and Triebel–Lizorkin spaces and they have been used afterwards by many authors. The interpolation property bellow will be very useful in proving our main result. Proposition 11. Let φ ∈ B and 0 < p, q, q0 , q1
∞. If s0 , s1 are real numbers fulfilling s1 < βφ̄
αφ̄ < s0 , then we have   s φ 0 , b s1 bpq pq1 γ ,q = bpq , 0 where γ is defined as in (17). s s1 0 Proof. Firstly, we note that spaces bpq 0 and bpq1 form an interpolation couple since they φ s1 , for example. We can interpret bpq as the sequence are both continuously embedded in bp∞ φ1 space ℓq (ℓp (Zn )) where φ1 (t) := φ(t) t −n/p , t ∈ (0, ∞). In fact, the index l does not bring any trouble. It is not hard to see that formula (16) remains valid for these spaces. On the other hand, the Boyd indices of φ1 are given by βφ1 = βφ̄ − n p and αφ1 = αφ̄ − n . p Taking σi = si − n/p (i = 0, 1), then we have σ0  1  and γ1 (t) := t σ0 −σ1 /φ1 t σ0 −σ1 = γ (t), σ1 < βφ1
αφ1 < σ0 Hence, attending to formula (16), we have  σ   n  σ   n     ℓq00 ℓp Z , ℓq11 ℓp Z = ℓφq 1 ℓp Zn , γ ,q s φ s1 0 that is, (bpq 0 , bpq1 )γ ,q = bpq . ✷ t ∈ (0, ∞). 208 A. Almeida / J. Math. Anal. Appl. 304 (2005) 198–211 s and r is a Lemma 12. Let s ∈ R, 0 < p < ∞, and 0 < q
∞. If {λlj m }(l,j,m)∈I ∈ bpq  l l natural number such that r > max(s, σp − s), then (l,j,m)∈I λj m ψj m converges uncons if q < ∞ and in any B t with t < s, if q = ∞. ditionally in Bpq pq Proof. First, we assume that q < ∞. From properties (18), (19), and (21), we see that, for each l, the functions 2−j (s−n/p) ψjl m are 1r -atoms (j = 0) or (s, p)r,r -atoms (j ∈ N) according to [20, Definition 13.3], ignoring constants which are independent of ℓ, j , and m. Using the Atomic Decomposition Theorem (cf. [20, pp. 75–76]), we arrive at the conclusion that there exists c > 0 such that the estimate
q
    l p q/p

l  s
l j (s−n/p)q  
λj m ψj m  Bpq

c λj m 2 (27)
l,j (l,j,m)∈K m holds for all finite subsets K of I (the sums on the right-hand side run over all indices (l, j ) and m such that (l, j, m) ∈ K). From this estimate and from the summability of the two families of positive real numbers involved in (26), we conclude that the partial sums on the s , left-hand side of (27) constitute a generalized Cauchy sequence in the complete space Bpq thus converge in this space. Now, let t < s and q = ∞. We reduce this case to the previous one by using the atomic s t with ֒→ bpu decomposition result as before (with t in place of s) and remarking that bp∞ 0 < u < ∞. ✷ In the sequel, we formulate our main result related to wavelet representation and some of its consequences. Theorem 13. Let φ ∈ B, 0 < p < ∞, and 0 < q
∞. Consider the system {ψjl m }(l,j,m)∈I as previously. Then there exists r(φ, p) such that, for any r ∈ N with r > r(φ, p), the following holds: φ Given f ∈ S ′ , then f ∈ Bpq if and only if it can be represented as λlj m ψjl m f= φ with λ ∈ bpq (28) (l,j,m)∈I (unconditional convergence in S ′ ). Moreover, the “wavelet coefficients” λlj m are uniquely determined by   (29) λlj m = λlj m (f ) := 2j n f, ψjl m , (l, j, m) ∈ I. Further, 

 φ


f  B
∼
λ(f )  bφ
pq pq (equivalent quasi-norms), where λ(f ) ≡ {λlj m (f )}(l,j,m)∈I . Proof. Step 1. Assume that f ∈ S ′ can be represented as λlj m ψjl m f= (l,j,m)∈I (unconditional convergence in S ′ ) (30) A. Almeida / J. Math. Anal. Appl. 304 (2005) 198–211 209 φ for some λ ∈ bpq . Let s0 , s1 ∈ R. Attending to Lemma 12, we conclude that the operator s s s1 s1 0 0 T : bp,1 + bp,1 −→ Bp,1 + Bp,1 , given by µlj m ψjl m Tµ= (unconditional convergence in S ′ ), (l,j,m)∈I is well-defined and linear if r > max(r(s0 , p), r(s1 , p)), for example, where r(si , p) (i = 0, 1) is given by (22). Moreover, by Theorem 8 one concludes that the restriction of T to si si is a bounded linear operator into Bp,1 . Choosing s0 , s1 above such that s1 < each bp,1 βφ̄
αφ̄ < s0 and attending to the interpolation property and to Propositions 7 and 11, we φ arrive at the conclusion that the restriction of T to bpq is also a bounded linear operator φ φ into Bpq . Thus, f ∈ Bpq and


 φ

 φ

f  B
=
T λ  B

c
λ  b φ
pq pq pq for some c > 0 independent of λ and f . φ Step 2. Now, let f ∈ Bpq . Assume that s0 , s1 , and r fulfill the same conditions as in Step 1. Consider the operator s s s1 s1 0 0 S : Bp,1 + Bp,1 −→ bp,1 + bp,1 defined by      S g = λ(g) := 2j n g0 , ψjl m + g1 , ψjl m (l,j,m)∈I , (31) si where g = g0 + g1 with gi ∈ Bp,1 , i = 0, 1. Theorem 8 (and Remark 10) shows that S si is a bounded linear operator is well-defined, it is linear and its restriction to each Bp,1 si into bp,1 . Taking into account the interpolation property as previously, one concludes that φ φ the restriction of S to Bpq is a bounded linear operator into bpq as well. Therefore, 



 φ


Sf  b
=
λ(f )  bφ

c
f  B φ
, (32) pq pq pq φ where c > 0 does not dependent on f . So, λ(f ) ∈ bpq and hence λlj m (f )ψjl m g := (33) (l,j,m)∈I φ (unconditional convergence in S ′ ) belongs to the space Bpq by Step 1. But Theorem 8 once again allows us to conclude that T S is the identity operator, so g = f . But we have (by Step 1)


 φ

f  B

c
λ(f )  bφ
, (34) pq pq c > 0 independent of f . Therefore, equivalence (30) follows from estimates (32) and (34). φ It remains to show that representation (28) is unique. We do this next. Suppose that f ∈ Bpq admits the representation λlj m ψjl m f= (l,j,m)∈I φ (unconditional convergence in S ′ ). with λ ∈ bpq 210 A. Almeida / J. Math. Anal. Appl. 304 (2005) 198–211 φ φ s s s1 s1 s1 s1 0 0 Since Bpq ֒→ Bp,1 + Bp,1 ֒→ Bp,1 and bpq ֒→ bp,1 + bp,1 ֒→ bp,1 (note that s0 > s1 ), s1 then f ∈ Bp,1 has the representation λlj m ψjl m f= s1 with λ ∈ bpq (unconditional convergence in S ′ ), (l,j,m)∈I which is unique by Theorem 8. The proof of the theorem is completed. ✷ Remark 14. We can choose s0 and s1 close enough to αφ̄ and βφ̄ , respectively, and take   2n n + − βφ̄ r(φ, p) := max αφ̄ , p 2 in Theorem 13. In fact, in that case, is possible to choose ε > 0 such that r is greater than n 2n n βφ̄ , 2n p + 2 − βφ̄ + ε, αφ̄ + ε, and p + 2 − αφ̄ . On the other hand, one can take s0 and s1 such that βφ̄ − ε < s1 < βφ̄
αφ̄ < s0 < αφ̄ + ε. In this way, we have r > max(r(s0 , p), r(s1 , p)) as required in the proof above. Remark 15. We would like to remark also that one did not make a direct use of duality φ results about the spaces Bpq . According to (31), we have defined the symbol f, ψjl m in (29) as the sum of two quantities interpreted as in Remark 10. Of course, taking into φ s1 ), one can take account the choice of s0 and s1 made above (which implies Bpq ֒→ Bp,1 f0 = 0 and f1 = f , so f, ψjl m can be evaluated as described in that remark. Corollary 16. Let φ, p, and q be as in Theorem 13. If r ∈ N is large enough and q < ∞, φ then {ψjl m }(l,j,m)∈I forms an unconditional Schauder basis in Bpq . Proof. Attending to Theorem 13, all we need to do is to check that the series in (28) φ converges unconditionally in Bpq (if p, q < ∞). We proceed as in the first part of the proof of Lemma 12: observe that φ(2j )−1 2j n/p ψjl m are 1r -N -atoms (l = 1, j = 0) or (σ, p)r,r -N -atoms (j ∈ N) according to [7, Definition 4.4.1], with σ = {φ(2j )}j ∈N0 and N = {2j }j ∈N0 . Hence, it is possible to use the Atomic Decomposition Theorem from [7, Section 4.4.2], in order to get the counterpart of estimate (27), that is,

q   

 l p q/p   j  −j n/p q l  φ
l
  λj m ψj m  Bpq

c λj m φ 2 2
(l,j,m)∈K m l,j with c > 0 independent of K. To do that we have to assume that r > r(φ, p) satisfies also the conditions mentioned in that theorem restricted to our particular case. We conclude now as in Lemma 12. ✷ Corollary 17. Let φ, p, q, and r as in Theorem 13. Then    I : f −→ 2j n f, ψjl m (l,j,m)∈I φ φ establishes a topological isomorphism from Bpq onto bpq (interpreted as in Remark 15). A. Almeida / J. Math. Anal. Appl. 304 (2005) 198–211 211 Proof. This result follows at once from the properties of the operators T and S studied in the proof of Theorem 13. ✷ Acknowledgment We thank Professor António Caetano for suggesting the topic and for many valuable comments and the constant interest taken during the preparation of this paper. References [1] J. Bergh, J. Löfström, Interpolation Spaces: An Introduction, Springer-Verlag, Berlin, 1976. [2] A. Caetano, S. Moura, Local growth envelopes of spaces of generalized smoothness: the sub-critical case, Math. Nachr. 273 (2004) 43–57. [3] A.M. Caetano, Subatomic representation of Bessel potential spaces modelled on Lorentz spaces, in: J.A. Sampaio Martins Anniversary Volume, Textos Mat. Univ. Coimbra, Sér. B 34 (2004) 37–47. [4] A.M. Caetano, W. Farkas, Local growth envelopes of Besov spaces of generalized smoothness, preprint, Univ. Aveiro. [5] F. Cobos, D.L. Fernandez, Hardy–Sobolev Spaces and Besov Spaces with a Function Parameter, Lecture Notes in Math., vol. 1302, Springer-Verlag, Berlin, 1988, pp. 158–170. [6] I. Daubechies, Ten Lectures on Wavelets, CBMS–NSF Regional Conf. Ser. in Appl. Math., SIAM, Philadelphia, 1992. [7] W. Farkas, H.-G. Leopold, Characterizations of function spaces of generalized smoothness, Ann. Mat. Pura Appl., in press. [8] M. Frazier, B. Jawerth, Decomposition of Besov spaces, Indiana Univ. Math. J. 34 (1985) 777–799. [9] M. Frazier, B. Jawerth, A discrete transform and decompositions of distribution spaces, J. Funct. Anal. 93 (1990) 34–170. [10] G. Garrigós, A. Tabacco, Wavelet decompositions of anisotropic Besov spaces, Math. Nachr. 239/240 (2002) 80–102. [11] D. Goldberg, A local version of real Hardy spaces, Duke Math. J. 46 (1979) 27–42. [12] D. Haroske, H. Triebel, Wavelet bases and entropy numbers in weighted function, preprint Math/Inf/01/04, University of Jena, Germany, 2004. [13] C. Merucci, Applications of Interpolation with a Function Parameter to Lorentz, Sobolev and Besov Spaces, Lecture Notes in Math., vol. 1070, Springer-Verlag, Berlin, 1984, pp. 183–201. [14] Y. Meyer, Wavelets and Operators, Cambridge Univ. Press, Cambridge, UK, 1992. [15] S. Moura, Function spaces of generalised smoothness, Dissertationes Math. 398 (2001) 1–87. [16] H. Triebel, A note on wavelet bases in function spaces, in: Proc. Orlicz Centenary Conference “Function Spaces 7,” Poznan, Banach Center Publ., 2003, in press. [17] H. Triebel, Spaces of Besov–Hardy–Sobolev Type, Teubner, Leipzig, 1978. [18] H. Triebel, Theory of Function Spaces, Birkhäuser, Basel, 1983. [19] H. Triebel, Interpolation Theory, Function Spaces, Differential Operators, Barth, Heidelberg, 1995. [20] H. Triebel, Fractals and Spectra, Birkhäuser, Basel, 1997. [21] H. Triebel, The Structure of Functions, Birkhäuser, Basel, 2001. [22] P. Wojtaszczyk, A Mathematical Introduction to Wavelets, Cambridge Univ. Press, Cambridge, UK, 1997. [23] M. Yamazaki, A quasihomogeneous version of paradifferential operators I, Boundedness on spaces of Besov type, J. Fac. Sci. Univ. Tokyo, Sect. IA Math. 33 (1986) 131–174.