A Construction of C 1 -Wavelets on the Two-Dimensional Sphere

Transcription

1 Applied and Computational Harmonic Analysis 10, 1 26 (2001) doi: /acha , available online at on A Construction of C 1 -Wavelets on the Two-Dimensional Sphere Ilona Weinreich Department of Mathematics and Technology, Rhein Ahr Campus Remagen, University of Applied Sciences Koblenz, Südallee 2, D Remagen, Germany weinreich@rheinahrcampus.de Communicated by Wim Sweldens Received June 17, 1998; revised September 4, 2000 In this paper a construction of C 1 -wavelets on the two-dimensional sphere is presented. First, we focus on the construction of a multiresolution analysis leading to C 1 -functions on S 2. We show refinability of the constructed tensor product generators. Second, for the wavelet construction we employ a factorization of the refinement matrices which leads to refinement matrices characterizing complement spaces. With this method we achieve an initial stable completion. A desired stable completion can be gained by lifting the initial stable completion. The result is a biorthogonal wavelet basis leading to C 1 -functions on the sphere Academic Press Key Words: multiresolution analysis; wavelets; exponential splines; tensor product splines; wavelets on surfaces. 1. INTRODUCTION During the past several years there has been great interest in wavelet constructions on closed manifolds. The two-dimensional sphere has to be considered as an example of central importance for many applications. Some of the common applications are the representation and manipulation of geographical data, the illumination algorithms in computer graphics, and the solution of integral or more general operator equations as well as medical problems because of the sphere-like structure of organs. The S 2 is a simple manifold from the mathematical point of view. Nevertheless there are remarkable difficulties to transfer the global concepts of classical wavelet theory, e.g., in R 2, to general manifolds. The first approach for the construction of wavelets on the sphere has been presented by Dahlke et al. [10]; here a characterization of C 1 -functions on S 2 (and S 3 ) with a wavelet representation is performed. The basic idea goes back to an approach using splines proposed by Schumaker and Traas [25] where the sphere is parameterized by polar coordinates. They employ a tensor product of trigonometric and polynomial splines for /01 $35.00 Copyright 2001 by Academic Press All rights of reproduction in any form reserved.

2 2 ILONA WEINREICH scattered data approximation. In [10] tensor products are employed, too. There, E-splines (exponential splines) [12, 21] are connected with interval wavelets [2]. A characteristic feature of an approach which yields C 1 -functions is the reproduction of trigonometric functions. This property is ensured with the help of appropriate E-splines. Before summarizing the contents of the present work we sketch some other approaches which also treat the sphere S 2. Schröder and Sweldens [22] start with a triangulation of the sphere. Simple wavelets, e.g., Haar wavelets, are constructed on the triangles. With the lifting scheme one obtains smoother wavelets on the sphere. This approach is very useful in practical areas, e.g., for the compression of topographic data as well as in computer graphics. This approach is of great interest especially in connection to new results from Cohen et al. [6]. They establish stability of these wavelets. In [6] finite element wavelets with compactly supported duals are obtained by the lifting scheme on triangulations starting with an initial triangulation by uniform refinements. With a special treatment of the exceptional points a minimal Sobolev smoothness of the dual system can be shown. The results can be applied to the sphere by using a triangulation from an initial icosahedron or dodecahedron. With tensor products of interpolatory trigonometric wavelets and polynomial wavelets Potts and Tasche [19] construct continuous wavelets on S 2.HeretheC 1 -conditions are satisfied approximately. A similar idea with spherical harmonics is performed by Potts et al. [20]. They achieve smooth but as in [19] globally supported wavelets. In [18] Lyche and Schumaker present a construction of L-spline wavelets with compact support. An explicit multiresolution analysis for the sphere is not treated. Constructions of Freeden, Schreiner, and Windheuser are based on spherical harmonics and are applied to geodetical problems; see, e.g., [15, 27]. Another approach coming from CAGD, by Dahmen and Schneider [14], is based on a representation of the sphere (or more general manifolds) as a union of parametric patches which join at least continuously. On the single patches biorthogonal wavelets which satisfy certain transfer conditions are constructed. This approach is applied successfully for the treatment of integral equations. A further recent construction from Göttelmann [17] is based on polar coordinates. On a reduced grid with appropriately adapted splines a stable wavelet basis is obtained. An advantage of that approach in comparison with the present paper is the stability in a scale of Sobolev spaces H s ( ), 0 s<1. Nevertheless, our approach seems to be simpler; it is not necessary to modify the grid. The central concern of the present paper is the construction of a multiresolution analysis as well as a wavelet basis of locally supported C 2 -functions on the sphere. Beginning with a parameterization of the sphere as already employed in [10] we reduce the problem to considering appropriate wavelets on the square [0, 1] 2 for the moment. The fulfillment of side conditions (i.e., the reproduction of constant and trigonometric functions) must be guaranteed for obtaining C 1 -functions. For that reason tensor products are considered which are similar to those in [10]. The first component function is again an E-spline and the second component an interval wavelet. We show that the so-constructed functions are refinable. We obtain a multiresolution of L 2 ([0, 1] 2 ) and therefore a correspondence to C 1 -functions on S 2 by coordinate transformation. This approach allows a construction of a biorthogonal basis, the existence of which is crucial for stability considerations. The wavelet construction is based on a generalization of a matrix factorization method from Dahmen and Micchelli [13] to nonstationary splines. This method yields a so-called

3 SMOOTH WAVELETS ON THE SPHERE 3 initial stable completion. A suitable stable completion corresponding to the biorthogonal system is obtained with a method from Carnicer et al. [4]. This method is very similar and algebraically equivalent to the lifting scheme invented by Sweldens [23] which has been developed independently. The lifting scheme provides a relationship between different multiresolution analyses with the same scaling function. It enables the construction of wavelets for a particular application using the degrees of freedom after fixing the biorthogonality relations. For details see [24]. We proceed as follows: In Section 2 we introduce nonstationary multiresolution analysis with E-splines. In Section 3 we present our approach for the sphere. We consider a parameterization of the sphere by polar coordinates. The claim for C 1 -continuity makes things more difficult. The basis functions must represent not only constant but also trigonometric functions. We develop a multiresolution analysis which satisfies the corresponding conditions. Refinability is explicitly shown. In Section 4 we construct wavelets as so-called initial stable completion which can be mapped to each desired stable completion, especially the one induced by the biorthogonal system, with the aid of the method introduced by Carnicer et al. [4]. For the constructed basis, L 2 -stability can be shown, if a biorthogonal basis exists. 2. BASIC CONCEPTS: E-SPLINES AND INTERVAL WAVELETS In this section we introduce two concepts of multiresolution analysis which we are going to employ for the construction of a spherical multiresolution analysis. On the one hand, this is a nonstationary multiresolution analysis with E-splines and, on the other hand, a construction for a bounded interval. First, we consider a nonstationary multiresolution analysis where the generator depends on the refinement level. A typical example arises in connection with exponential splines [12, 21]. DEFINITION 2.1. Let µ C N be fixed and X a (d N)-matrix where x 1,...,x N R d are the columns of X. The E-spline C µ ( X) of dimension d is defined by f(x)c µ (x X)dx = e µ u f(x u) du (2.1) R d [0,1] N for all f C(R d ). If we choose f(x)= e iξx in Definition 2.1 we obtain the Fourier transform of C µ ( X) Formula (2.2) implies Ĉ µ (ξ X) Ĉ µ/2 ( ξ 2 X) = 1 2 N Ĉ µ (ξ X) = N ( 1 e (iξ x l +µ l ) ) iξ x l. (2.2) + µ l l=1 N (1 + e (iξ/2 xl +µ l /2) ) =: 1 2 aµ (z), z = e iξ/2. (2.3) l=1

4 4 ILONA WEINREICH This means that C µ ( X) is refinable with symbol a µ (z) = k aµ k zk and satisfies a twoscale relation of the form C µ (x X) = a µ k C µ/2(2x k X), (2.4) k Z d where the mask {a µ k } k Z is given by the coefficients of the symbol a µ (z). Let the spaces S j be defined by S j := span{c µ/2 j (2 j k X) k Z d }; (2.5) then Formula (2.4) implies that {S j } j Z is a sequence of nested spaces; i.e., S 1 S 0 S 1. (2.6) For Re µ = 0 the spaces {S j } j Z form a (nonstationary) multiresolution analysis. This has been shown by de Boor et al. [9]. In the following lemma we summarize some useful properties of E-splines. For the general theory we refer to [12, 21]. LEMMA 2.1. If we consider X as the indexed set of column vectors of the matrix X and X as the number of columns in X then the following holds: (i) The translates {C µ ( k X)} k Z d are linearly independent if and only if X is unimodular and µ y µ y / 2πiZ\{0} for all y, y. (ii) Let Y be a maximal linearly independent subset of X with Y =d. Then the following formula is valid: k L(Y ) e uy k C µ (x k X) = Ĉµ(iu y X) e u y x det Y (2.7) with u y = Y T µ y, L(Y ) ={Yl l Z}. For our application we employ several special properties of the E-splines C µ ( X), which can be formulated as conditions concerning the parameters µ = (µ 1,...,µ N ).We consider the univariate E-splines matrix X = (1,...,1). With the following special choice of µ we obtain the desired representation of constant and trigonometric functions with E-splines. Remark 2.1. Let φ j = 2 j/2 C p (2 j ) = 2 j/2 µ/2 j l Z C µ/2 j (2j 2 j l) be a periodic E-spline with parameter vector µ = (0,iν, iν). Then for ν R and j>log 2 (2/π)ν there exist coefficients K j, L j k, Mj k such that k=0 k=0 k=0 K j φ j (2 j x k) = 1, (2.8) L j k φj (2 j x k) = cos(νx), (2.9) M j k φj (2 j x k) = sin(νx). (2.10)

5 SMOOTH WAVELETS ON THE SPHERE 5 This is a consequence of (2.7). Next we describe a multiresolution analysis on the interval. The starting point for the construction of biorthogonal wavelets on the interval [0, 1] is a multiresolution analysis derived from B-splines; see [1, 5, 11]. Here φ j,k := φ(2 j x k) are the dilates and translates of the centralized cardinal B-spline of order d; i.e., φ(x) := N d (x + d/2 ). According to [11] we define j := {φ j,k : k j }, φ j,k := { φ X j,k, X {L,R}, k X j, φ [j,k] = 2 j/2 φ(2 j k), k 0 j. (2.11) The index set j is divided into sets L j, R j,and 0 j, where the former two are associated with the left and right border of the interval and the latter corresponds to the interior functions which do not overlap the border. In our context the explicit form of the edge functions φj,k L and φr j,k is not needed, so for details we refer to [11] or Appendix A.1. In the following we suppose that j is large enough such that supp φ [j,k] [0, 1] for k 0 j. More precisely, we demand j log 2 ( l + l 2 1) + 1 =: j 0, (2.12) where l, l 2 depend on the support of the dual function φ; see Appendix A.1. We consider the unit sphere 3. MULTIRESOLUTION ANALYSIS ON THE SPHERE S 2 := { x R 3 x 2 2 = x x2 2 + x2 3 = 1}. (3.1) Our aim is to construct a stable wavelet basis on S 2. For this, we take spherical coordinates (for technical reasons with transformation on the unit square) ρ: [0, 1] 2 S 2 cos(2πx)sin(πy) (x, y) sin(2πx)sin(πy). (3.2) cos(πy) In the following we identify S 2 with the unit square [0, 1] 2. Of course not every (differentiable) function on [0, 1] 2 corresponds to a (differentiable) function on S 2.The transformation ρ maps each of the edges {(x, 0) x [0, 1]} and {(x, 1) x [0, 1]} of the parameter domain to one point on the sphere, namely the north and the south pole. We formulate conditions for functions on [0, 1] 2 which guarantee that the corresponding functions on S 2 are differentiable. A function f is continuous on S 2 if the periodicity condition f(0,y)= f(1,y), 0 y 1, (3.3)

6 6 ILONA WEINREICH holds and if there exist constants S S, S N such that f(x,0) = S S, (3.4) f(x,1) = S N, 0 x 1 (3.5) are satisfied. Moreover, the function f is differentiable on S 2 if there exist constants A S, B S, A N, B N such that f f (0,y)= (1,y), x x 0 y 1, (3.6) f y (x, 0) = A S cos(2πx)+ B S sin(2πx), 0 x 1, (3.7) f y (x, 1) = A N cos(2πx)+ B N sin(2πx), 0 x 1, (3.8) hold. Conditions (3.3) (3.6) are obvious because of the chosen parameterization. For the derivation of the last two conditions (3.7), (3.8) we refer to [16]. THEOREM 3.1 [16]. A function f represents a closed surface of the class C 1 (i.e., at each point of f the tangential plane of f varies continuously) if f and the partial derivatives f/ x and f/ y are continuous and satisfy the conditions (3.3) (3.8) with appropriate A S, B S, A N, B N. In the following we construct a wavelet basis of L 2 ([0, 1] 2 ) that satisfies the conditions (3.3) (3.8). As generators we choose tensor products of univariate generators of a multiresolution analysis of L 2 ([0, 1] 2 ), where we denote bivariate generators with bold φ and univariate with φ. We consider the tensor product φ j,o k = φ j φ j,k2, (3.9) where the two component functions are defined as follows. The first component φ j := 2 j/2 C p µ/2 j (2 j ) = 2 j/2 l Z C µ/2 j (2j 2 j l ) is a periodized E-spline with C µ ( X) according to (2.1). We choose X = (1,...,1) and µ 1 = 0, µ 2 = 2πi, µ 3 = 2πi. So the conditions for a multiresolution analysis are satisfied for j>2; see [9, 10]. The second component is a dilated and translated version of a centered B-spline φ j,k2 = 2 j/2 N d (2 j k 2 + d 2 ) of order d with support in the interior of [0, 1]. The indices k 2 with the property suppφ j,k2 [0, 1] belong to the set 0 j. In the following let j j n with j 0 according to (2.12). The functions φ j,o k satisfy supp φ j,o k [0, 1] (0, 1). (3.10) On the edges y = 0andy = 1 we consider fixed linear combinations of several generators. Therefore we need to know which of the functions φ j,k2, k 2 X j, X {L,R}, defined in [11] do not vanish for y = 0, y = 1.

7 SMOOTH WAVELETS ON THE SPHERE 7 Remark 3.1. Letφ be a spline of degree d 1. For the generators φ j,k, k X j, X {L,R}, corresponding to (2.11) one has φ j,l d+r (0) = φ j,2 j l l(d)+d r (1) = { 0 forr = 1,...,d 1, 2 j/2 for r = 0. (3.11) For the derivatives i.e., the one-sided limits for r = 0 of the generators one has { (φ j,l d+r ) 0 forr 1, (0) = (φ j,2 j l l(d)+d r ) (1) = 2 j 3/2 for r = 1. (3.12) On the border y = 0of[0, 1] 2 we consider linear combinations of all φ j and the function φ j,l d, which does not vanish at y = 0. We define the combined generator (which we call edge generator in the following) φ j,l := λ j,l φ j φ j,l d. (3.13) Further edge generators φ j,l cos, φ j,l sin are defined as linear combinations of all φ j and the function φ j,l d+1, the derivative of which does not vanish for y = 0. Therefore let φ j,l cos := λ j,l cos φ j φ j,l d+1, (3.14) φ j,l sin := λ j,l sin φ j φ j,l d+1. (3.15) Analogously, we define the edge generators for the right border y = 1of[0, 1] 2 φ j,r := λ j,r φ j φ j,2 j r, (3.16) φ j,r cos := λ j,r cos φ j φ j,2 j r 1, (3.17) φ j,r sin := λ j,r sin φ j φ j,2 j r 1, (3.18) with φ j,k2, k 2 R j according to [11] and τ := l d + l(d). Hereagaind is the order of the employed B-spline and l, l(d) depend on its support; see Appendix A.1. For X {L,L cos,l sin,r,r cos,r sin } the coefficients λ j,x will be chosen such that the conditions (3.3) (3.8) for C 1 -continuity on S 2 are satisfied (see Theorem 3.2 below) for φ j,x. Furthermore, the normalization condition φ j,x L2 ([0,1] 2 ) = 1 (3.19) is required. The periodicity conditions (3.3) and (3.6) obviously hold because of the periodicity of φ j for all φ j,x. Moreover, functions which do not overlap the borders

8 8 ILONA WEINREICH y = 0, y = 1 satisfy (3.4) and (3.5). For the functions φ j,x, X {L cos,l sin,r cos,r sin } Lemma 3.1 provides φ j,l cos (x, 0) = φ j,l sin (x, 0) = φ j,r cos (x, 1) = φ j,r sin (x, 1) = 0. Special consideration is necessary for φ j,l and φ j,r. They are supposed to reproduce a constant. Furthermore, they must satisfy the normality conditions (3.19) and therefore φ j,l (x, 0) = 2 j/2, φ j,r (x, 1) = 2 j/2. (3.20) For interior functions (3.7) and (3.8) are valid with A S = B S = A N = B N = 0. Especially, one has y φj,l (x, 0) = y φj,r (x, 1) = 0. The remaining functions φ j,l cos, φ j,l sin, φ j,r cos,andφ j,r sin should satisfy y φj,l cos (x, 0) = y φj,r cos (x, 1) = 2 3j/2 cos(2πx), y φj,l sin (x, 0) = y φj,r sin (x, 1) = 2 3j/2 sin(2πx). (3.21) THEOREM 3.2. By coordinate transformation ρ (see (3.2)) the generators φ j,x, X {L,L cos,l sin,r,r cos,r sin },leadtoc 1 -functions on S 2, and holds iff the coefficients are φ j,x L2 ([0,1] 2 ) = 1 λ j,l = λ j,r = 2 j/2 K j, (3.22) λ j,l cos λ j,l sin = λ j,r cos = 2 j/2 L j, (3.23) = λ j,r sin = 2 j/2 M j, (3.24) for each = 0,...,2 j 1 and j j 0 where K j, L j, and M j are determined by (2.8), (2.9), and (2.10). Proof. For the norms we have 2 φ j,l L2 ([0,1] 2 ) = λ j,l φ j φ j,l d L2 ([0,1] 2 ) 2 = K j φ j (2 j x ) φ j,l d L2 ([0,1] 2 ) = 1 φ j,l d L2 ([0,1] 2 ) = 1, 2 φ j,l cos L2 ([0,1] 2 ) = L j φ j (2 j x ) φ j,l d+1 L2 ([0,1] 2 )

9 SMOOTH WAVELETS ON THE SPHERE 9 = cos φ j,l d+1 L2 ([0,1] 2 ) = 1, 2 φ j,l sin L2 ([0,1] 2 ) = M j φ j (2 j x ) φ j,l d+1 L2 ([0,1] 2 ) = sin φ j,l d+1 L2 ([0,1] 2 ) = 1. In the same way, we obtain φ j,x L2 ([0,1] 2 ) = 1forX {R,R cos,r sin }. To prove the C 1 -continuity conditions (3.3) (3.8) for all φ j,x, X I φ, we still have to consider (3.20) and (3.21). We show (3.20) for φ j,l. First, with φ j,l (x, 0) = λ j,l φ j 2j 1 (x)φ j,l d (0) = 2 j/2 K j φ j (2 j x ) = 2 j/2 we obtain (3.20). Condition (3.21) arises from 2 y φj,l cos (x, 0) = λ j,l cos φ j (x)(φ j,l d+1 ) (0) = 2 j/2 L j 2 j/2 φ j (2 j x )2 3j/2 = 2 3j/2 cos(2πx). For φ j,l sin as well as φ j,r cos and φ j,r sin the argumentation is analogous. The values of the generators on the interval edges are calculated as follows: 2 y φj,l cos (0, 0) = L j φ j ( )2 3j/2 = 2 3j/2 cos(0)2 3j/2, 2 y φj,l sin ( 1 4, 0) = M j φ j (2 j 1 4 )2 3j/2 = 2 3j/2 sin(2π 1 4 ) = 23j/2, y φj,r cos (0, 1) = y φj,r sin ( 1 4, 1) = 23j/2. So (3.3) (3.8) are valid for all φ j,x, X {L,L cos,l sin,r,r cos,r sin }, which means that the corresponding functions give rise to smooth functions on the sphere. In the following we choose the generators φ j,x, X {L,L cos,l sin,r,r cos,r sin }, according to Theorem 3.2 such that they represent C 1 -functions on S 2. The inner functions φ j,o k from (3.9) satisfy the conditions for C 1 -continuity automatically, because their value on the border of the interval [0, 1] is zero. The same is true for those φ j,o k, k = (,k 2 ), with k 2 / 0 j and k 2 l d, l d + 1, 2 j τ 1, 2 j τ. Let the following index set Ij o be associated with these functions, which do not overlap the borders of the interval I o j := {k Z2 = 0,...,2 j 1, k 2 j \{l d,l d + 1, 2 j τ 1, 2 j τ}}. By Sj diff we denote the space spanned by all edge generators φ j,x, X I φ, and by all φ j,o k Ij o.by diff j we denote the following basis of Sj diff diff j := {φ j,l, φ j,l cos, φ j,l sin, φ j,r, φ j,r cos, φ j,l sin, φ j,o k k Ij o }. (3.25) k,

10 10 ILONA WEINREICH The first question is whether the so-constructed spaces are nested and therefore may give rise to a multiresolution analysis. We have to check whether S diff j Sj+1 diff (3.26) is valid. Therefore, we show that φ j,l, φ j,l cos, φ j,l sin, φ j,r, φ j,r cos,andφ j,r sin as well as, k Ij o, are refinable. φ j,o k Remark 3.2. The inner functions φ j,o k, k Ij o, are obviously refinable because they are a tensor product where both component functions are refinable. The first component function, the E-spline, is nonstationary refinable; see (2.4). We show the refinability of φ j,l and φ j,l cos. The arguments for the remaining edge generator are analogous. THEOREM 3.3. Let φ j,l, φ j,l cos, φ j,l sin Sj diff, j j 0, be defined as in (3.13) (3.15) with coefficients λ j,l, λ j,l cos, and λ j,l sin as in Theorem 3.2. The functions satisfy the refinement equations φ j,l = 2 1/2 φ j+1,l + s, (3.27) φ j,l cos = 2 3/2 φ j+1,l cos + φ j,l sin = 2 3/2 φ j+1,l sin + s I j+1 γ L j,s φj+1,o γ Lcos j,s s I j+1 γ Lsin j,s s I j+1 φ j+1,o s, (3.28) φj+1,o s, (3.29) with γj,s L 1,s 2 = 2 1/2 λ j+1,l σ s2,0, γ L cos j,s 1,s 2 = 2 1/2 λ j+1,l cos σ s2,1, γ L sin j,s 1,s 2 = 2 1/2 λ j+1,l sin σ s2,1. Here s I j+1 := {s Z 2 s 1 = 0,...,2 j+1 1,s 2 = l,...,2l + l 2 2} and (3.30) { αs2,r for s 2 = l,...,2l + l 1 1, σ s2,r = 2 r+1/2 β s2,r for s 2 = 2l + l 1,...,2l + l 2 2 (3.31) with α s2,r, asin(a.45) and β s2,r as in (A.51). Proof. of First we show (3.27). To do so we have to represent φ j,l as a linear combination 2 j+1 1 φ j+1,l = λ j+1,l φs j+1 1 φ j+1,l d s 1 =0 and inner functions φ j+1,o k. Using the refinement equation φ j 2 j+1 1 = 2 1/2 a j s 1 2 φs j+1 1 (3.32) s 1 =0

11 SMOOTH WAVELETS ON THE SPHERE 11 for the periodic E-splines φ j as well as Eq. (A.51) for φ j,l d we obtain φ j,l = λ j,l φ j φ j,l d = = ( 2j+1 1 λ j,l s 1 =0 2 j+1 ( s 1 =0 ) 2l+l 2 1/2 a j s 1 2 φs j (φ 1/2 j+1,l d + ) λ j,l 2 1 a j s 1 2 s 1 =0 2 j+1 ( 1 2l+l s 2 =l φs j+1 1 φ j+1,l d s 2 =l σ s2,0φ j+1,s2 ) ) λ j,l a j s σ s2,0 φs j+1 1 φ j+1,s2. (3.33) Expression (3.33) corresponds to the refinement equation (3.27) if and only if λ j,l 2 1 a j s 1 2 = 2 1/2 λ j+1,l (3.34) is valid and γ L j,s is of the form (3.30). According to Theorem 3.2, φj,l corresponds to a C 1 -function on S 2 if the equation λ j,l = 2 j/2 K j holds, where the K j are the coefficients from the representation of the unity (2.8). With the refinement equation (3.32) for E-splines in (2.8) we obtain This gives 2j+1 1 K j s 1 =0 2 j+1 ( 1 2 = s 1 =0 a j s 1 φ j+1 (2 j+1 x 2 s 1 ) K j a j s 1 2 )φ j+1 (2 j+1 x s 1 ) = 1. and the relation K j+1 = K j a j s 1 2 (3.35) λ j,l a j s = 2 (j/2+1) K j a j s 1 2 = 2 (j/2+1) K j+1 = 2 1/2 λ j+1,l for the coefficients λ j,l. Therefore (3.34) and the refinement equation (3.27) with γ L j,s according to (3.30) are valid. An application of the refinement equations (A.51) and (3.32) together with the representation formula for the cosine with E-splines yields the proof

12 12 ILONA WEINREICH of (3.29). The refinement equation (3.29) is valid iff λ j,l cos 2 2 a j s 1 2 = 2 3/2 λ j+1,l cos s 1. (3.36) The coefficients of φ j,l cos are chosen according to Theorem 3.2 and we obtain λ j,l cos = 2 j/2 L j,wherel j are the coefficients of the cosine representation (2.9) with E-splines. Similar to the calculation of K j, we receive For the coefficients λ j,l cos Ls j+1 1 = L j a j s 1 2. (3.37) we obtain the relation λ j,l cos a j r = j/2 L j a j s 1 2 = 2 (j+1)/2 2 3/2 L j+1 s 1 = 2 3/2 λ j+1,l cos s 1. We have shown that (3.36) is valid. Thus the statement follows. The proof of the generator φ j,l sin is analogous to that of φ j,l cos by employing the corresponding sine representation formula with E-splines. For the dual multiresolution analysis we consider two-dimensional functions based on a tensor product of dual E-splines and dual B-splines. Dual functions corresponding to E-splines are constructed as generalizations of the stationary case [3] in [26, Chap. 3]. The biorthogonal B-splines on the interval are built in [11]. In this case a biorthogonalization procedure is performed for obtaining the desired biorthogonal system. A similar procedure is necessary in the present case. Under the assumption that the corresponding biorthogonalization matrix is nonsingular we obtain a dual basis j of diff j. For details we refer to [26]. There a biorthogonal basis is constructed such that diff j, j = I (3.38) is valid. In the present paper we do not need that basis explicitly, but assume its existence. 4. WAVELETS ON S 2 BY STABLE COMPLETIONS For the construction of the wavelet spaces we employ the concept of stable completions which has been introduced by Carnicer et al. in [4]. The principal idea is to start with a refinement matrix M and to complete it with a matrix M such that the matrix M j = (M, M ) is uniformly bounded and invertible. More precisely, we consider a dense subset S of closed nested subspaces of S j of a Hilbert space H with a stable basis j of S j.let

13 SMOOTH WAVELETS ON THE SPHERE 13 j S j+1 be a set of functions spanning a complement of S j in S j+i.theset{ j j } is uniformly stable iff j T = T j+1 M (4.1) for M [l 2,( j ), l 2 ( j+1 )] such that M j = (M, M ) [l 2 ( j j ), l 2 ( j+1 )] (4.2) is uniformly bounded and invertible. This means that there exists a matrix G j [l 2 ( j+1 ), l 2 ( j j )] which can be represented as and has the following properties G j = ( G G j M j = M j G j = I (4.3) M j, G j =O(1), j N 0. (4.4) Condition (4.3) is equivalent to ( ) ( ) G M G M I 0 =, G M G M 0 I M G + M G = I. (4.5) In particular, the reconstruction property G ) T j+1 = T j G + T j G follows. Each matrix M with the properties mentioned above is called the stable completion of M. In [13], a method for how to obtain a stable completion for B-splines with a factorization method of the refinement matrix is developed. Once given a so-called initial stable completion we can obtain each desired stable completion of M by the method described in [4]. For example, if j ={ φ j,k k j } is a biorthogonal basis of j, i.e., which is also refinable and stable, which means φ j,k, φ j,k =δ kk, k j,k j, (4.6) T j = T j+1 M (4.7) and c k 2 2 c k φ j,k, (4.8) k j k j H then the induced projectors Q n satisfy a commutator property (see [8]) which is essential for the global stability of the basis { j } and { j }. The following proposition describes how to achieve another stable completion M which corresponds to the complements (Q j+1 Q j )S j+1 from an initial stable completion ˇM of M.

14 14 ILONA WEINREICH PROPOSITION 4.1 [4]. Let { j }, { j }, M, and M be associated as above and ˇM be a stable completion of M and let Ǧ j = ˇM 1 j.then is a stable completion as well and G j = M 1 j The sets form biorthogonal systems M := (I M M T ) ˇM (4.9) is of the form ( M T ) G j =. (4.10) Ǧ j := M T j+1, j := Ǧ, j+1 (4.11) j, j =I, j, j = j j =0 (4.12) such that (Q j+1 Q j )S j+1 = W j, (Q j+1 Q j ) S j+1 = W j. (4.13) With these results we are prepared to construct bases for the complement spaces W diff j = S diff j+1 Sdiff j. (4.14) As a basis of Sj diff we again consider diff j as defined in (3.25). We want to characterize the wavelet basis of the complement spaces Wj diff by stable completions; see [4]. For this, we first construct an initial stable completion. Our proceeding is based on a matrix factorization method of Dahmen and Micchelli [13]. The generators of {Sj diff } j j0 are tensor products of two one-dimensional generators except for the edge generators φ j,l, φ j,l cos, φ j,l cos, φ j,r cos, φ j,r sin, φ j,r. Hence we need to modify the procedure only concerning these edge functions. In [26] we showed that under certain conditions (which are satisfied in our context) the matrix factorization method is applicable also for E-splines. Let M E be the E-spline refinement matrix, M E a stable completion, and GE, GE the reconstruction matrices, with G E ME = I (2j ), G E ME = 0, G E ME = 0, GE ME = I (2j ), M E GE + ME GE = I 2j+1, M E j, GE j =O(1). Let M B be the refinement matrix of the interval B-spline. In [11] an initial stable completion M B is constructed. According to Proposition 4.6 [11] there exist matrices M B, GB, GB with

15 SMOOTH WAVELETS ON THE SPHERE 15 G B MB = I # j, G B MB = 0, G B MB = 0, GB MB = I (2j ), M B GB + MB GB = I # j +1, Mj B, GB j =O(1). For the construction of matrices corresponding to diff j we first consider the reduced refinement matrix M B,red, which is the internal (# j+1 4) (# j 4) submatrix of M B. For this submatrix a stable completion can easily be constructed, because it corresponds to the inner functions of the multiresolution analysis of the interval [0, 1]. Finally we obtain matrices M B,red, G B,red, G B,red which constitute a stable completion of M B ; i.e., G B,red M B,red = I # j, G B,red M B,red = 0, (4.15) G B,red M B,red = 0, G B,red M B,red = I (2j ). For a more detailed presentation we refer to Appendix A.2. To construct an initial stable completion for the sphere let the refinement matrix M for the generators of Sj diff be given by φ j,l φ j+1,l φ j,l cos φ j+1,l cos φ j,l sin φ j+1,l sin φ j,o 0,l d+2 φ j+1,o 0,l d+2. =: M T φ j,o., (4.16) 2,2 j φ j+1,o τ 2 2 j+1 1,2 j+1 τ 2 φ j,r cos φ j+1,r sin φ j,r sin φ j+1,r cos φ j,r φ j+1,r where we order the interior generators lexicographically with k = (,k 2 ) I o j. φ j,o k = φ j φ j,k2 (4.17) Remark 4.1. Note that the size of the refinement matrix M is (2 j+1 (2 j+1 σ 4) + 6) (2 j (2 j σ 4) + 6) with σ := 2l + l(d) 2d 1. Proof. The generators in the interior consist of tensor products (4.17), where all 2 j translates of the E-spline component φ j and # j 4 = 2 j + 2d + 2l l(d) = 2 j σ 4φ j,k2 (all, except the four with value or derivative value zero 0 or 1) are considered. All together one has 2 j (2 j σ 4) + 6 basis elements including the six edge generators φ j,l, φ j,l cos, φ j,l sin, φ j,r cos, φ j,r sin, φ j,r at the jth level.

16 16 ILONA WEINREICH With α = 2 1/2 and β = 2 3/2 the matrix M has the representation (4.18) with submatrices N j,l, N j,r, M, where N j,l results from the refinement equations (3.27), (3.29), and N j,r from corresponding equations for the right edge generators φ j,r, φ j,r. Finally, M is the Kronecker product of the involved refinement matrices for the one-dimensional generators; i.e., M = ME MB,red, (4.19) where M E is the refinement matrix of the periodic E-spline and MB,red is defined as in (A.56). In view of the construction of an initial stable completion of M, we first look for a matrix G such that Regarding G M = I (2j (2 j σ 4)+6). (4.20) (4.21)

17 SMOOTH WAVELETS ON THE SPHERE 17 where we obtain 1 α Ñ j,l = N j,l 0 β 0, β 1 β Ñ j,r = N j,r 0 β α, (4.22) For the construction of a stable completion of M we define the matrices Ǧ, ˇM, Ǧ as follows: G := GE GB,red, (4.23) G d := (4.24) Ǧ = G d H j,m, (4.25) M := (ME Mred,ME Mred,ME Mred ), (4.26) (4.27)

18 18 ILONA WEINREICH (4.28) where we denote the interior (3 2 2 j 2 j (σ + 4)) (2 j+1 (2 j+1 σ 4) + 6)-block with G. Note that ˇM = H 1 j,m ˇM and Ǧ = Ǧ H j,m. (4.29) Altogether, the preceding constructions lead to a stable completion of M. This is stated in the following theorem. The proof can be done straightforward by employing the onedimensional components of the involved matrices, from which the desired properties are known. For details we refer to Appendix A.3. THEOREM 4.1. The matrix ˇM (4.27) is a stable completion of M (4.18); i.e., (i) (ii) Ǧ M = I 2j (2 j σ 4)+6 (4.30) Ǧ ˇM = 0 (4.31) Ǧ M = 0 (4.32) Ǧ ˇM = I 3 22j 2 j (σ +4) (4.33) M Ǧ + ˇM Ǧ = I (2j+1 (2 j+1 σ 4)+6) (4.34) (iii) M j, G j =O(1). (4.35) Let diff j be defined according to (3.25) and j a dual basis. For the construction of a biorthogonal wavelet basis we define the projectors Q j f = f, φ j,o k φ j,o k + f, φ j,x φ j,x (4.36) k Ij o X {L,L,R,R} Q j f = f,φ j,o k φ j,o k + f,φ j,x φ j,x. (4.37) k Ij o X {L,L,R,R} Remark 4.2. If both of the sets diff j, j are l 2 -stable, then the sequences of linear projectors defined by (4.36) and (4.37) are uniformly bounded operators. The

19 SMOOTH WAVELETS ON THE SPHERE 19 biorthogonality (3.38) implies Q l Q j = Q l, Q l Q j = Q l (4.38) for l j. By M we denote the refinement matrix corresponding to j and write the refinement equation as ( j ) T = ( j+1 ) T M. (4.39) The construction of the biorthogonal wavelet bases is a direct consequence of Proposition 4.1. COROLLARY 4.1. Let diff j be defined by (3.25) and j a dual basis. Let M be the refinement matrix (4.18) and M a refinement matrix according to (4.39).Let ˇM be an initial stable completion corresponding to Theorem 4.1.Then is a stable completion of M as well. The sets M := (I M ( M ) T ) ˇM (4.40) j diff := M T diff j+1, j := Ǧ T j+1 (4.41) build biorthogonal systems j diff, j =1 j diff, j = diff j, j =0. Complement spaces are given by (Q j+1 Q j )Sj+1 diff = W j diff (Q j+1 Q j ) S j+1 = W j. Under the assumption that the biorthogonalization matrix for the construction of the biorthogonal system j is invertible and has a uniform bounded condition number one can show the Riesz basis property for L 2 (S 2 ) by verifying the Jackson and Bernstein estimates following arguments developed by Dahmen [7]. For details see [26]. In the present paper we sketched some of the results of [26]. First, we have focused on the construction of the multiresolution analysis leading to C 1 -functions on the sphere. We showed refinability of the constructed combined tensor product generators. Second, we employed a matrix factorization method for the wavelet construction. With this method we achieved an initial stable completion. A desired stable completion can be gained by the method of Carnicer et al. [4]; see Proposition 4.1 in this paper. The result is a biorthogonal wavelet basis leading to C 1 -functions on the sphere. A.1. Multiresolution Analysis on an Interval APPENDIX Here we describe a multiresolution analysis on the interval. The starting point is a multiresolution analysis derived from B-splines. To obtain moment conditions for the

20 20 ILONA WEINREICH wavelets one also needs a reproduction of polynomials for the dual generator. We consider cardinal B-splines N d of order d with supp N d =[0,d]. Let the generator φ be a centered B-spline φ(x):= N d (x + ) d 2 (A.42) with support [ d 2, d 2 ] =: [l 1,l 2 ]. We have the symmetry property φ(x + l(d)) = φ( x),wherel(d) = d mod 2. It is well known that B-splines generate a multiresolution analysis. Dual functions to φ are φ = d, d, φ with d d and d + d even, as constructed in [3] such that φ, φ( k) =δ 0k. (A.43) An important property of B-splines is their approximation quality which depends on the reproduction of polynomials. B-splines of order d are exact of degree d 1; i.e., x r = k Z ( ) r, φ( k) φ(x k) (A.44) for all r = 0,...,d 1. The duals φ are exact of degree d 1. Following [11], the idea for constructing a multiresolution analysis on the interval [0, 1] is to keep all functions φ [j,k] = 2 j/2 φ(2 j k), with support in [0, 1]. For retaining the optimal approximation it is necessary to take functions which overlap the borders of [0, 1] into consideration also. By defining α L j,m,r := α m,r := ( ) r, φ( m) and α R j,m,r := α 2 j m l(d),r, (A.45) for r = 0,...,d 1, the relations α j,m,r L φ [j,m](x) = 2 j(r+1/2) x r, m Z (A.46) α j,m,r R φ [j,m](x) = 2 j(r+1/2) (1 x) r, r = 0,...,d 1 (A.47) m Z are valid. As in [1, 5] these relations can be employed to construct basis functions on the interval [0, 1]. Again following [11] we have supp φ =[l 1 d + 1,l 2 + d 1]=:[ l 1, l 2 ]. Let l > l 1, l 2. With l := l ( d d) the index sets L j := {l d,...,l 1}, 0 j := {l,...,2 j l l(d)}, R j := {2j l + 1 l(d),...,2 j l + d l(d)} are defined. The generators on the right and left border of [0, 1] are defined by φ L j,l d+r := φ R j,2 j l+d l(d) r := l 1 m= l j l 1 1 m=2 j l l(d)+1 α m,r φ [j,m] [0,1], α j,m,r R φ [j,m] [0,1] (A.48) (A.49)

21 SMOOTH WAVELETS ON THE SPHERE 21 for r = 0,...,d 1, and for the functions in the interior of [0, 1], i.e., for k 0 j, we write φ j,k := φ [j,k]. On the basis of this background we define j := {φ j,k : k j }, where φ j,k := { φ X j,k, k X j,x {L,R} φ [j,k], k 0 j. (A.50) In [11] refinement equations for the generators φj,l d+r L and φr are proved. j,2 j l+d l(d) r For the left border, for instance, we have ( ) 2l+l 1 1 φj,l d+r L = 2 (r+1/2) φj+1,l d+r L + 2l+l 2 2 α m,r φ [j+1,m] + β m,r φ [j+1,m] m=l m=2l+l 1 (A.51) with β L m,r = 2 1/2 l 1 q= (m l 2 )/2 A.2. Construction of Stable Completion α q,r a m 2q for r = 0,...,d 1. For the construction of matrices corresponding to diff j we need reduced refinement matrices M B,red, M B,red, G B,red, G B,red which are defined below. We eliminate all nondiagonal elements from the first two columns of M B.Letml d, m l d+1 be the first two and m 2j r 1, m 2j r the last two columns of M B and let e k be the kth unit vector of length # j+1.wedefine m r := diag( 2 1/2, 2 3/2, 1,...,1, 2 3/2, 2 1/2 )m r for r = l d, l d + 1, 2 j τ 1, 2 j τ. Moreover, we define the elimination matrix H M, which eliminates all entries of M B in the first two and last two columns apart from the diagonal: H M := ( 2 1/2 m l d, 2 3/2 m l d+1,e 3,...,e # j+1 2, 2 3/2 m 2j τ 1, 2 1/2 m 2j τ ), H M R (# j+1) (# j+1 ).ForM B, := H MM B by MB,red we denote, the inner (# j+1 4) (# j 4)-block of M B,. Thus one gets (A.52)

22 22 ILONA WEINREICH Similarly we proceed with G B and define G B, := GB H 1 M. From G B MB = I # j it follows that G B, MB, = I # j. (A.53) Dividing G B, by G B,red, we obtain similarly to MB, and denoting the inner (# j 4) (# j+1 4)-block (A.54) Finally, (A.53) implies G B,red M B,red = I # j 4. (A.55) For our construction we need the fact that the first and the last d rows of M B as well as the first and the last d columns of G B are identically zero; see [11, Chap. 4.1]. For the spherical construction we always choose d>2because we want to represent differentiable functions. Next, reduced matrices M B,red, G B,red are considered. The matrix M B,red arises from M B by neglecting the first two and the last two rows (A.56)

23 SMOOTH WAVELETS ON THE SPHERE 23 and G B,red is obtained from G B by neglecting the first and last two columns (A.57) For the reduced matrices we obviously have A.3. Proof of Theorem 4.1 G B,red M B,red = I # j, G B,red M B,red = 0, G B,red M B,red = 0, G B,red M B,red = I (2j ). Proof. (i) We have Ǧ M = G d Md. Consider the interior blocks of Gd and M d : G M = (GE GB,red )(M E MB,red ) = (G E ME ) (GB,red M B,red ) = I (2j ) I (2j σ 4) = I (2j (2 j σ 4)). Relation (4.30) follows directly. The matrix M E is a stable completion of ME which yields G E ME = 0. GB,red M B,red = 0 follows from formula (4.15) for the reduced refinement matrices of the B-splines. Taking into account the rules for computation with the Kronecker product we obtain (G E GB,red )(M E MB,red ) = 0, (G E GB,red )(M E MB,red ) = 0, (G E GB,red )(M E MB,red ) = 0, and therefore G M = 0. Equations (4.29) and (4.25) yield the assertion (4.31). With G E ME = 0andGB,red M B,red = 0 we receive (G E GB,red )(M E MB,red ) = 0, (G E GB,red )(M E MB,red ) = 0, (G E GB,red )(M E MB,red ) = 0,

24 24 ILONA WEINREICH and therefore G M = 0. Equation (4.22) and again (4.29) yield (4.32). The remaining assertion (4.33) follows analogously to (4.30) (4.32). (ii) We transform the left-hand side of (4.34) as follows: M Ǧ + ˇM Ǧ = H 1 j,m Md Gd H j,m + H 1 ˇM j,m Ǧ H j,m = H 1 j,m (Md Gd + ˇM Ǧ )H j,m. Again we consider the interior blocks M, G.Wehave M G = ME GE MB,red G B,red. (A.58) For the interior blocks M, G the following is true M G = ME GE MB,red G B,red Therefore, we have + M E GE MB,red G B,red + M E GE MB,red G B,red. M G + M G = ME GE (MB,red G B,red + M B,red G B,red ) + M E GE (MB,red G B,red + M B,red G B,red ) = I 2j+1 I 2j+1 σ 4 = I 2j+1 (2 j+1 σ 4). The claim is proved for the interior blocks. Clearly, the identity (4.34) follows. (iii) For the interior blocks the statement is clear because of the tensor product structure where for each of the components the statement is already proved. We extend the matrix M with six zero rows and G with six zero columns, respectively, and obtain ˇM (see (4.27)) and Ǧ (see (4.28)). For these matrices the uniform boundedness follows directly from the uniform boundedness of the component functions. The matrices M, G are extended by six rows (six columns, respectively) with unit vectors (or an appropriate multiple) to M d (see (4.22)), Gd (see (4.24)). This procedure does not change anything with respect to the uniform boundedness. Finally, we obtain M and G from M d, Gd by multiplication by H j,m, respectively H 1 j,m. It remains to show H j,m, H 1 j,m =O(1). We consider the -norm. With (4.21) we have 1 H j,m Ñ j,l + Ñ j,r + 1. (A.59) It suffices to give an estimate of the entries of Ñ j,l, Ñ j,r from above. W.l.o.g. consider Ñ j,l ; the first column consists of the coefficients γj,s L of the refinement equation (3.27). In the second column there are the coefficients γ L cos of the refinement equation (3.29) and in the third column γ L sin of (3.29). We have γj,s L 1,s 2 = 2 1/2 λ j+1,l σ s2,0, γ L cos j,s 1,s 2 = 2 1/2 λ j+1,l cos σ s2,1, andγ L sin j,s 1,s 2 = 2 1/2 λ j+1,l sin σ s2,1. Applying Theorem 3.2 given λ j,l = 2 j/2 K j, λ j,l cos = 2 j/2 L j, and λ j,l sin = 2 j/2 M j for the coefficients, where the quantities K j, L j, M j are bounded independent

25 SMOOTH WAVELETS ON THE SPHERE 25 from j; see [26]. The quantities σ s2,0, σ s2,1 come from the refinement equations for the interval B-splines. According to [11] they are bounded (independent of j). H j,m 2 = O(1) follows and assertion (4.35) with it. ACKNOWLEDGMENTS This work partially represents results of the author s PhD thesis which has been performed under the advice of W. Dahmen. I thank him for his support. REFERENCES 1. L. Andersson, N. Hall, B. Jawerth, and G. Peters, Wavelets on closed subsets of the real line, in Topics in the Theory and Applications of Wavelets, (L. L. Schumaker and G. Webb, Eds.), pp. 1 61, Academic Press, Boston, C. K. Chui and E. Quak, Wavelets on a bounded interval, in Numerical Methods of Approximation Theory, (D. Braess and L. L. Schumaker, Eds.), pp. 1 24, Birkhäuser, Basel, A. Cohen, I. Daubechies, and J. Feauveau, Bi-orthogonal bases of compactly supported wavelets, Comm. Pure Appl. Math. 45 (1992), J. M. Carnicer, W. Dahmen, and J. M. Peña, Local decomposition of refinable spaces, Appl. Comput. Harmon. Anal. 3 (1996), A. Cohen, I. Daubechies, and P. Vial, Wavelets on the interval and fast wavelet transforms, Appl. Comput. Harmon. Anal. 1(1) (1993), A. Cohen, L.-M. Echeverry, and Q. Sun, Finite element wavelets, Preprint University Pierre et Marie Curie, Paris, W. Dahmen, Stability of multiscale transformations, J. Fourier Anal. Appl. 4 (1996), W. Dahmen, Multiscale analysis, approximation, and interpolation spaces, in Approximation Theory VIII, (C. K. Chui and L. L. Schumaker, Eds.), pp , World Scientific, Singapore, C. de Boor, R. A. DeVore, and A. Ron, On the construction of multivariate (pre)wavelets, Constr. Approx. 9 (1993), S. Dahlke, W. Dahmen, E. Schmitt, and I. Weinreich, Multiresolution analysis and wavelets on S 2 and S 3, Numer. Funct. Anal. Optim. 16 (1995), W. Dahmen, A. Kunoth, and K. Urban, Biorthogonal spline wavelets on the interval stability and moment conditions, Appl. Comput. Harmon. Anal. 6 (1999), W. Dahmen and C. A. Micchelli, On multivariate E-splines, Adv. in Math. 76 (1989), W. Dahmen and C. A. Micchelli, Banded matrices with banded inverses, II: Locally finite decomposition of spline spaces, Constr. Approx. 9 (1993), W. Dahmen and R. Schneider, Wavelets on manifolds I: construction and domain decomposition, SIAM J. Math. Anal. 31 (1999), W. Freeden and M. Schreiner, Orthogonal and nonorthogonal multiresolution analysis, scale discrete and exact fully discrete wavelet transform on the sphere, Constr. Approx. 14 (1998), R. H. J. Gmelig Meyling and P. R. Pfluger, B-spline approximation of a closed surface, IMA J. Numer. Anal. 7 (1987), J. Göttelmann, Locally supported wavelets on manifolds with applications to the 2D sphere, Appl. Comput. Harmon. Anal. 7 (1999), T. Lyche and L. L. Schumaker, L-spline wavelets, in Wavelets: Theory, Algorithms, and Applications, (C. K. Chui, L. Montefusco, and L. Puccio, Eds.), pp , Academic Press, San Diego, D. Potts and M. Tasche, Interpolatory wavelets on the sphere, in Approximation Theory VIII, (C. K. Chui and L. L. Schumaker, Eds.), pp , World Scientific, Singapore, 1995.

26 26 ILONA WEINREICH 20. D. Potts, G. Steidl, and M. Tasche, Kernels of spherical harmonics and spherical frames, in Advanced Topics in Multivariate Approximation, (F. Fontanella, K. Jetter, P.-J. Laurent, Eds.), World Scientific, Singapore, A. Ron, Exponential box splines, Constr. Approx. 4 (1988), P. Schroeder and W. Sweldens, Spherical wavelets: Efficiently representing functions on the sphere, in Computer Graphics Proceedings (SIGGRAPH 95), pp , W. Sweldens, The lifting scheme: A custom-design construction of biorthogonal wavelets, Appl. Comput. Harmon. Anal. 3 (1996), W. Sweldens, The lifting scheme: A construction of second generation wavelets, SIAM J. Math Anal. 2 (1997), L. L. Schumaker and C. Traas, Fitting scattered data on spherelike surfaces using tensor products of trigonometric and polynomial splines, Numer. Math. 60 (1991), I. Weinreich, Biorthogonale Wavelets auf der Sphäre, Shaker, Aachen, Germany, U. Windheuser, Sphärische Wavelets: Theorie und Anwendungen in der Physikalischen Geodäsie, Dissertation, Univ. Kaiserslautern, 1995.