We investigate the estimation of a multiplicative separable regression function from a bidimensional nonparametric regression model with random design. We present a general estimator for this problem and study its mean integrated squared error (MISE) properties. A wavelet version of this estimator is developed. In some situations, we prove that it attains the standard unidimensional rate of convergence under the MISE over Besov balls.

1. Motivations

We consider the bidimensional nonparametric regression model with random design described as follows. Let (Yi,Ui,Vi)i∈ℤ be a stochastic process defined on a probability space (Ω,𝒜,ℙ), where
(1)Yi=h(Ui,Vi)+ξi,i∈ℤ,(ξi)i∈ℤ is a strictly stationary stochastic process, (Ui,Vi)i∈ℤ is a strictly stationary stochastic process with support in [0,1]2, and h:[0,1]2→ℝ is an unknown bivariate regression function. It is assumed that 𝔼(ξ1)=0, 𝔼(ξ12) exists, (Ui,Vi)i∈ℤ are independent, (ξi)i∈ℤ are independent, and, for any i∈ℤ, (Ui,Vi) and ξi are independent. In this study, we focus our attention on the case where h is a multiplicative separable regression function: there exist two functions f:[0,1]→ℝ and g:[0,1]→ℝ such that
(2)h(x,y)=f(x)g(y).
We aim to estimate h from the n random variables: (Y1,U1,V1),…,(Yn,Un,Vn). This problem is plausible in many practical situations as in utility, production, and cost function applications (see, e.g., Linton and Nielsen [1], Yatchew and Bos [2], Pinske [3], Lewbel and Linton [4], and Jacho-Chávez [5]).

In this note, we provide a theoretical contribution to the subject by introducing a new general estimation method for h. A sharp upper bound for its mean integrated squared error (MISE) is proved. Then we adapt our methodology to propose an efficient and adaptive procedure. It is based on two wavelet thresholding estimators following the construction studied in Chaubey et al. [6]. It has the features to be adaptive for a wide class of unknown functions and enjoy nice MISE properties. Further details on wavelet estimators can be found in, for example, Antoniadis [7], Vidakovic [8], and Härdle et al. [9]. Despite the so-called “curse of dimensionality” coming from the bidimensionality of (1), we prove that our wavelet estimator attains the standard unidimensional rate of convergence under the MISE over Besov balls (for both the homogeneous and inhomogeneous zones). It completes asymptotic results proved by Linton and Nielsen [1] via nonadaptive kernel methods for the structured nonparametric regression model.

The paper is organized as follows. Assumptions on (1) and some notations are introduced in Section 2. Section 3 presents our general MISE result. Section 4 is devoted to our wavelet estimator and its performances in terms of rate of convergence under the MISE over Besov balls. Technical proofs are collected in Section 5.

2. Assumptions and Notations

For any p≥1, we set
(3)𝕃p([0,1])={v:[0,1]⟶ℝ;∥v∥p=(∫01|v(x)|pdx)1/p<∞}.
We set
(4)eo=∫01f(x)dx,e*=∫01g(x)dx,
provided that they exist.

We formulate the following assumptions.

There exists a known constant C1>0 such that
(5)supx∈[0,1]|f(x)|≤C1.

There exists a known constant C2>0 such that
(6)supx∈[0,1]|g(x)|≤C2.

The density of (U1,V1), denoted by q, is known and there exists a constant c3>0 such that
(7)c3≤inf(x,y)∈[0,1]2q(x,y).

There exists a known constant ω>0 such that
(8)|eoe*|≥ω.

The assumptions (H1) and (H2), involving the boundedness of h, are standard in nonparametric regression models. The knowledge of q discussed in (H3) is restrictive but plausible in some situations, the most common case being (U1,V1)~𝒰([0,1]2) (the uniform distribution on [0,1]2). Finally, mention that (H4) is just a technical assumption more realistic to the knowledge of eo and e* (depending on f and g, resp.).
3. MISE Result

Theorem 1 presents an estimator for h and shows an upper bound for its MISE.

Theorem 1.

One considers (1) under (H1)–(H4). One introduces the following estimator for h (2):
(9)h^(x,y)=f~(x)g~(y)e~1{|e~|≥ω/2},
where f~ denotes an arbitrary estimator for fe* in 𝕃2([0,1]), g~ denotes an arbitrary estimator for geo in 𝕃2([0,1]), 1 denotes the indicator function,
(10)e~=1n∑i=1nYiq(Ui,Vi),
and ω refers to (H4).

Then there exists a constant C>0 such that
(11)𝔼(∬01(h^(x,y)-h(x,y))2dxdy)≤C(1n𝔼(∥g~-geo∥22)+𝔼(∥f~-fe*∥22)hinda+𝔼(∥g~-geo∥22∥f~-fe*∥22)+1n).

The form of h~ (9) is derived to the multiplicative separable structure of h (2) and a ratio-type normalization. Other results about such ratio-type estimators in a general statistical context can be found in Vasiliev [10].

Based on Theorem 1, h^ is efficient for h if and only if f~ is efficient for fe* and g~ is efficient for geo in terms of MISE. Even if several methods are possible, we focus our attention on wavelet methods enjoying adaptivity for a wide class of unknown functions and having optimal properties under the MISE. For details on the interests of wavelet methods in nonparametric statistics, we refer to Antoniadis [7], Vidakovic [8], and Härdle et al. [9].

4. Adaptive Wavelet Estimation

Before introducing our wavelet estimators, let us present some basics on wavelets.

4.1. Wavelet Basis on [0, 1]

Let us briefly recall the construction of wavelet basis on the interval [0,1] introduced by Cohen et al. [11]. Let N be a positive integer, and let ϕ and ψ be the initial wavelets of the Daubechies orthogonal wavelets db2N. We set
(12)ϕj,k(x)=2j/2ϕ(2jx-k),ψj,k(x)=2j/2ψ(2jx-k).
With appropriate treatments at the boundaries, there exists an integer τ satisfying 2τ≥2N such that the collection 𝒮={ϕτ,k(·),k∈{0,…,2τ-1};ψj,k(·);j∈ℕ-{0,…,τ-1},k∈{0,…,2j-1}}, is an orthonormal basis of 𝕃2([0,1]).

Any v∈𝕃2([0,1]) can be expanded on 𝒮 as
(13)v(x)=∑k=02τ-1ατ,kϕτ,k(x)+∑j=τ∞∑k=02j-1βj,kψj,k(x),x∈[0,1],
where αj,k and βj,k are the wavelet coefficients of v defined by
(14)αj,k=∫01v(x)ϕj,k(x)dx,βj,k=∫01v(x)ψj,k(x)dx.

4.2. Besov Balls

For the sake of simplicity, we consider the sequential version of Besov balls defined as follows. Let M>0, s∈(0,N), p≥1 and r≥1. A function v belongs to Bp,rs(M) if and only if there exists a constant M*>0 (depending on M) such that the associated wavelet coefficients (14) satisfy
(15)2τ(1/2-1/p)(∑k=02τ-1|ατ,k|p)1/p+(∑j=τ∞(2j(s+1/2-1/p)(∑k=02j-1|βj,k|p)1/p)r)1/r≤M*.
In this expression, s is a smoothness parameter and p and r are norm parameters. For a particular choice of s, p, and r, Bp,rs(M) contains the Hölder and Sobolev balls (see, e.g., DeVore and Popov [12], Meyer [13], and Härdle et al. [9]).

4.3. Hard Thresholding Estimators

In the sequel, we consider (1) under (H1)–(H4).

We consider hard thresholding wavelet estimators for f~ and g~ in (9). They are based on a term-by-term selection of estimators of the wavelet coefficients of the unknown function. Those which are greater to a threshold are kept; the others are removed. This selection is the key to the adaptivity and the good performances of the hard thresholding wavelet estimators (see, e.g., Donoho et al. [14], Delyon and Juditsky [15], and Härdle et al. [9]).

To be more specific, we use the “double thresholding” wavelet technique, introduced by Delyon and Juditsky [15] then recently improved by Chaubey et al. [6]. The role of the second thresholding (appearing in the definition of the wavelet estimator for βj,k) is to relax assumption on the model (see Remark 6).

Estimator f~ for fe*. We define the hard thresholding wavelet estimator f~ by
(16)f~(x)=∑k=02τ-1α^τ,kϕτ,k(x)+∑j=τj1∑k=02j-1β^j,k1{|β^j,k|≥κC*λn}ψj,k(x),
where
(17)α^τ,k=1an∑i=1anYiq(Ui,Vi)ϕτ,k(Ui),
where an is the integer part of n/2,
(18)β^j,k=1an∑i=1anWi,j,k1{|Wi,j,k|≤C*/λn},Wi,j,k=Yiq(Ui,Vi)ψj,k(Ui),
where j1 is the integer satisfying (1/2)an<2j1≤an, κ=2+8/3+24+16/9, C*=(2/c3)(C12C22+𝔼(ξ12)), and
(19)λn=lnanan.Estimator g~ for geo. We define the hard thresholding wavelet estimator g~ by
(20)g~(x)=∑k=02τ-1υ^τ,kϕτ,k(x)+∑j=τj2∑k=02j-1θ^j,k1{|θ^j,k|≥κ*C*ηn}ψj,k(x),
where
(21)υ^τ,k=1bn∑i=1bnYan+iq(Uan+i,Van+i)ϕτ,k(Van+i),
Where an is the integer part of n/2, bn=n-an,
(22)θ^j,k=1bn∑i=1bnZan+i,j,k1{|Zan+i,j,k|≤C*/ηn},Zan+i,j,k=Yan+iq(Uan+i,Van+i)ψj,k(Van+i),
Where j2 is the integer satisfying (1/2)bn<2j2≤bn, κ*=2+8/3+24+16/9, C*=(2/c3)(C12C22+𝔼(ξ12)), and
(23)ηn=lnbnbn.Estimator for h. From f~ (16) and g~ (20), we consider the following estimator for h (2):
(24)h^(x,y)=f~(x)g~(y)e~1{|e~|≥ω/2},
where
(25)e~=1n∑i=1nYiq(Ui,Vi)
and ω refers to (H4).

Let us mention that h~ is adaptive in the sense that it does not depend on f or g in its construction.

Remark 2.

Since f~ is defined with (Y1,U1,V1),…,(Yan,Uan,Van) and g~ is defined with (Yan+1,Uan+1,Van+1),…,(Yn,Un,Vn), thanks to the independence of (Y1,U1,V1),…,(Yn,Un,Vn), f~ and g~ are independent.

Remark 3.

The calibration of the parameters in f~ and g~ is based on theoretical considerations; thus defined, f~ and g~ can attain a fast rate of convergence under the MISE over Besov balls (see [6], Theorem 6.1]). Further details are given in the proof of Theorem 4.

4.4. Rate of Convergence

Theorem 4 investigates the rate of convergence attains by h^ under the MISE over Besov balls.

Theorem 4.

We consider (1) under (H1)–(H4). Let h^ be (24) and let h be (2). Suppose that

f∈Bp1,r1s1(M1) with M1>0, r1≥1, either {p1≥2 and s1∈(0,N)} or {p1∈[1,2) and s1∈(1/p1,N)},

g∈Bp2,r2s2(M2) with M2>0, r2≥1, either {p2≥2 and s2∈(0,N)} or {p2∈[1,2) and s2∈(1/p2,N)}.

Then there exists a constant C>0 such that
(26)𝔼(∬01(h^(x,y)-h(x,y))2dxdy)≤C(lnnn)2s*/(2s*+1),
where s*=min(s1,s2).

The rate of convergence (lnn/n)2s*/(2s*+1) is the near optimal one in the minimax sense for the unidimensional regression model with random design under the MISE over Besov balls Bp,rs*(M) (see, e.g., Tsybakov [16], and Härdle et al. [9]). Thus Theorem 4 proves that our estimator escapes to the so-called “curse of dimensionality.” Such a result is not possible with the standard bidimensional hard thresholding wavelet estimator attaining the rate of convergence (lnn/n)2s/(2s+d) with d=2 under the MISE over bidimensional Besov balls defined with s as smoothness parameter (see Delyon and Juditsky [15]).

Theorem 4 completes asymptotic results proved by Linton and Nielsen [1] investigating this problem for the structured nonparametric regression model via another estimation method based on nonadaptive kernels.

Remark 5.

In Theorem 4, we take into account both the homogeneous zone of Besov balls, that is, {p1≥2 and s1∈(0,N)}, and the inhomogeneous zone, that is, {p1∈[1,2) and s1∈(1/p1,N)}, for the case f∈Bp1,r1s1(M1) and the same for g∈Bp2,r2s2(M2). This has the advantage to cover a very rich class of unknown regression functions h.

Remark 6.

Note that Theorem 4 does not require the knowledge of the distribution of ξ1; {𝔼(ξ1)=0 and the existence of 𝔼(ξ12)} is enough.

Remark 7.

Let us mention that the phenomenon of curse of dimensionality has also been studied via wavelet methods by Neumann [17] but for the multidimensional Gaussian white noise model and with different approaches based on anysotropic frameworks.

Remark 8.

Our study can be extended to the multidimensional case considered by Yatchew and Bos [2], that is, f:[0,1]q1→ℝ and g:[0,1]q2→ℝ; q1 and q2 denoting two positive integers. In this case, adapting our framework to the multidimensional case (q1 dimensional Besov balls, q1 dimensional (tensorial) wavelet basis, q1 dimensional hard thresholding wavelet estimator,… see, e.g, Delyon and Juditsky [15]), one can prove that (9) attains the rate of convergence (lnn/n)2s*/(2s*+q*), where s*=min(s1,s2) and q*=max(q1,q2).

5. Proofs

In this section, for the sake of simplicity, C denotes a generic constant; its value may change from one term to another.

Proof of Theorem <xref ref-type="statement" rid="thm1">1</xref>.

Observe that
(27)h^(x,y)-h(x,y)=f~(x)g~(y)e~1{|e~|≥ω/2}-f(x)g(y)=1e~(f~(x)g~(y)-f(x)g(y)e~)1{|e~|≥ω/2}-f(x)g(y)1{|e~|<ω/2}.
Therefore, using the triangular inequality, the Markov inequality, (H1), (H2), (H4), {|e~|<ω/2}∩{|e*eo|≥ω}⊆{|e~-e*eo|≥ω/2}, and again the Markov inequality, we get
(28)|h^(x,y)-h(x,y)|≤2ω|f~(x)g~(y)-f(x)g(y)e~|+|f(x)||g(y)|1{|e~|<ω/2}≤C(|f~(x)g~(y)-f(x)g(y)e~|+1{|e~-e*eo|≥ω/2})≤C(|f~(x)g~(y)-f(x)g(y)e~|+|e~-e*eo|).
On the other hand, we have the decomposition
(29)f~(x)g~(y)-f(x)g(y)e~=f(x)e*(g~(y)-g(y)eo)+g(y)eo(f~(x)-f(x)e*)+(g~(y)-g(y)eo)(f~(x)-f(x)e*)+f(x)g(y)(e*eo-e~).
Owing to the triangular inequality, (H1) and (H2), we have
(30)|f~(x)g~(y)-f(x)g(y)e~|≤C(|g~(y)-g(y)eo|+|f~(x)-f(x)e*|h+|g~(y)-g(y)eo||f~(x)-f(x)e*|+|e~-e*eo|).
Putting (28) and (30) together, we obtain
(31)|h^(x,y)-h(x,y)|≤C(|g~(y)-g(y)eo|+|f~(x)-f(x)e*|h+|g~(y)-g(y)eo||f~(x)-f(x)e*|+|e~-e*eo|).
Therefore, by the elementary inequality:(a+b+c+d)2≤8(a2+b2+c2+d2), (a,b,c,d)∈ℝ4, an integration over [0,1]2 and taking the expectation, it comes
(32)𝔼(∬01(h^(x,y)-h(x,y))2dxdy)≤C(𝔼(∥g~-geo∥22)+𝔼(∥f~-fe*∥22)h+𝔼(∥g~-geo∥22∥f~-fe*∥22)+𝔼((e~-e*eo)2)).
Now observe that, owing to the independence of (Ui,Vi)i∈ℤ, the independence between (U1,V1) and ξ1, and 𝔼(ξ1)=0, we obtain
(33)𝔼(e~)=𝔼(Y1q(U1,V1))=𝔼(h(U1,V1)q(U1,V1))+𝔼(ξ1)𝔼(1q(U1,V1))=∬01f(x)g(y)q(x,y)q(x,y)dxdy=(∫01f(x)dx)(∫01g(y)dy)=e*eo.
Then, using similar arguments to (33), (a+b)2≤2(a2+b2), (a,b)∈ℝ2, (H1), (H2), (H3), and 𝔼(ξ12)<∞, we have
(34)𝔼((e~-e*eo)2)=𝕍(e~)=1n𝕍(Y1q(U1,V1))≤1n𝔼((Y1q(U1,V1))2)≤2n𝔼((h(U1,V1))2+ξ12(q(U1,V1))2)≤2c32(C12C22+𝔼(ξ12))1n=C1n.
Equations (32) and (34) yield the desired inequality:
(35)𝔼(∬01(h^(x,y)-h(x,y))2dxdy)≤C(𝔼(∥g~-geo∥22)+𝔼(∥f~-fe*∥22)hind+𝔼(∥g~-geo∥22∥f~-fe*∥22)+1n).

Proof of Theorem <xref ref-type="statement" rid="thm2">4</xref>.

We aim to apply Theorem 1 by investigating the rate of convergence attained by f~ and g~ under the MISE over Besov balls.

First of all, remark that, for γ∈{ϕ,ψ}, any integer j≥τ and any k∈{0,…,2j-1}.

Using similar arguments to (33), we obtain
(36)𝔼(1an∑i=1anYiq(Ui,Vi)γj,k(Ui))=𝔼(Y1q(U1,V1)γj,k(U1))=𝔼(h(U1,V1)q(U1,V1)γj,k(U1))+𝔼(ξ1)𝔼(γj,k(U1)q(U1,V1))=∬01f(x)g(y)q(x,y)γj,k(x)q(x,y)dxdy=(∫01f(x)γj,k(x)dx)(∫01g(y)dy)=∫01(f(x)e*)γj,k(x)dx.

Using similar arguments to (34) and ∥γj,k∥22=1, we have
(37)∑i=1an𝔼((Yiq(Ui,Vi)γj,k(Ui))2)=𝔼((Y1q(U1,V1)γj,k(U1))2)an≤2𝔼((h(U1,V1))2+ξ12(q(U1,V1))2(γj,k(U1))2)an≤2c3(C12C22+𝔼(ξ12))𝔼((γj,k(U1))2q(U1,V1))an=2c3(C12C22+𝔼(ξ12))∬01(γj,k(x))2q(x,y)q(x,y)dxdyan=2c3(C12C22+𝔼(ξ12))∥γj,k∥22an=C*2an,

with C*2=(2/c3)(C12C22+𝔼(ξ12)).

Applying [6, Theorem 6.1] (see the Appendix) with n=μn=υn=an, δ=0, θγ=C*, Wi=(Yi,Ui,Vi),
(38)qi(γ,(y,x,w))=yq(x,w)γ(x)
and f∈Bp1,r1s1(M1) (so fe*∈Bp1,r1s1(M1e*)) with M1>0, r1≥1, either {p1≥2 and s1∈(0,N)} or {p1∈[1,2) and s1∈(1/p1,N)}, we prove the existence of a constant C>0 such that
(39)𝔼(∥f~-fe*∥22)≤C(lnanan)2s1/(2s1+1)≤C(lnnn)2s1/(2s1+1),
when n is large enough.

The MISE of g~ can be investigated in a similar way: for γ∈{ϕ,ψ}, any integer j≥τ and any k∈{0,…,2j-1}.

We show that
(40)𝔼(1bn∑i=1bnYan+iq(Uan+i,Van+i)γj,k(Van+i))=∫01(g(x)eo)γj,k(x)dx.

We show that
(41)∑i=1bn𝔼((Yan+iq(Uan+i,Van+i)γj,k(Van+i))2)≤C*2bn,

with always C*2=(2/c3)(C12C22+𝔼(ξ12)).

Applying again [6, Theorem 6.1] (see the Appendix) with n=μn=υn=bn, δ=0, θγ=C*, Wi=(Yi,Ui,Vi),
(42)qi(γ,(y,x,w))=yq(x,w)γ(w)
and g∈Bp2,r2s2(M2) with M2>0, r2≥1, either {p2≥2 and s2∈(0,N)} or {p2∈[1,2) and s2∈(1/p2,N)}; we prove the existence of a constant C>0 such that
(43)𝔼(∥g~-geo∥22)≤C(lnbnbn)2s2/(2s2+1)≤C(lnnn)2s2/(2s2+1),
when n is large enough.

Using the independence between f~ and g~ (see Remark 2), it follows from (39) and (43) that
(44)𝔼(∥g~-geo∥22∥f~-fe*∥22)=𝔼(∥g~-geo∥22)𝔼(∥f~-fe*∥22)≤C(lnnn)4s1s2/(2s1+1)(2s2+1).
Owing to Theorem 1, (39), (43) and (44), we get
(45)𝔼(∬01(h^(x,y)-h(x,y))2dxdy)≤C(1n𝔼(∥g~-geo∥22)+𝔼(∥f~-fe*∥22)+𝔼(∥g~-geo∥22∥f~-fe*∥22)+1n)≤C((lnnn)2s2/(2s2+1)+(lnnn)2s1/(2s1+1)+(lnnn)4s1s2/(2s1+1)(2s2+1)+1n)≤C(lnnn)2s*/(2s*+1),
with s*=min(s1,s2).

Theorem 4 is proved.

Appendix

Let us now present in detail [6, Theorem 6.1] which is used two times in the proof of Theorem 4.

We consider a general form of the hard thresholding wavelet estimator denoted by f^H for estimating an unknown function f∈𝕃2([0,1]) from n independent random variables W1,…,Wn:
(A.1)f^H(x)=∑k=02τ-1α^τ,kϕτ,k(x)+∑j=τj1∑k=02j-1β^j,k1{|β^j,k|≥κϑj}ψj,k(x),
where
(A.2)α^j,k=1υn∑i=1nqi(ϕj,k,Wi),β^j,k=1υn∑i=1nqi(ψj,k,Wi)1{|qi(ψj,k,Wi)|≤ςj},ςj=θψ2δjυnμnlnμn,ϑj=θψ2δjlnμnμn,κ≥2+8/3+24+16/9 and j1 is the integer satisfying
(A.3)12μn1/(2δ+1)<2j1≤μn1/(2δ+1).
Here, we suppose that there exist

n functions q1,…,qn with qi:𝕃2([0,1])×Wi(Ω)→ℂ for any i∈{1,…,n},

two sequences of real numbers (υn)n∈ℕ and (μn)n∈ℕ satisfying limn→∞υn=∞ and limn→∞μn=∞,

such that, for γ∈{ϕ,ψ},

any integer j≥τ and any k∈{0,…,2j-1},
(A.4)𝔼(1υn∑i=1nqi(γj,k,Wi))=∫01f(x)γj,k(x)dx.

there exist two constants, θγ>0 and δ≥0, such that, for any integer j≥τ and any k∈{0,…,2j-1},
(A.5)∑i=1n𝔼(|qi(γj,k,Wi)|2)≤θγ222δjυn2μn.

Let f^H be (A.1) under (A1) and (A2). Suppose that f∈Bp,rs(M) with r≥1, {p≥2 and s∈(0,N)} or {p∈[1,2) and s∈((2δ+1)/p,N)}. Then there exists a constant C>0 such that
(A.6)𝔼(∥f^H-f∥22)≤C(lnμnμn)2s/(2s+2δ+1).Conflict of Interests

The author declares that there is no conflict of interests regarding the publication of this paper.

LintonO. B.NielsenJ. P.A kernel method of estimating structured nonparametric regression based on marginal integrationYatchewA.BosL.Nonparametric least squares estimation and testing of economic modelsPinskeJ.LewbelA.LintonO.Nonparametric matching and efficient estimators of homothetically separable functionsJacho-ChávezD.LewbelA.LintonO.Identification and nonparametric estimation of a transformed additively separable modelChaubeyY. P.ChesneauC.DoostiH.Adaptive wavelet estimation of a density from mixtures under multiplicative censoring2014, http://hal.archives-ouvertes.fr/hal-00918069AntoniadisA.Wavelets in statistics: a review (with discussion)VidakovicB.HärdleW.KerkyacharianG.PicardD.TsybakovA.VasilievV. A.One investigation method of a ratios type estimatorsProceedings of the 16th IFAC Symposium on System IdentificationJuly 2012Brussels, Belgium16CohenA.DaubechiesI.VialP.Wavelets on the interval and fast wavelet transformsDeVoreR.PopovV.Interpolation of Besov spacesMeyerY.DonohoD. L.JohnstoneI. M.KerkyacharianG.PicardD.Density estimation by wavelet thresholdingDelyonB.JuditskyA.On minimax wavelet estimatorsTsybakovA. B.NeumannM. H.Multivariate wavelet thresholding in anisotropic function spaces