A spherical Shepard operator constructed using the least squares method

Abstract

We consider the problem of approximating scattered datasets on the sphere by using a modified form of the Shepard operator, improved by its combination with polynomials obtained using the least squares method. The theoretical results are supported by several numerical experiments. The effectiveness of the operator in solving real-data problems is demonstrated through numerous practical applications, such as the approximation of atmospheric pressure, temperature, wind speed and precipitation.

Authors

Teodora Cătinaş
Faculty of Mathematics and Computer Science, Babeş-Bolyai University, Cluj-Napoca, Romania

Andra Malina
Faculty of Mathematics and Computer Science, Babeş-Bolyai University, Cluj-Napoca, Romania
Tiberiu Popoviciu Institute of Numerical Analysis, Romanian Academy, Cluj-Napoca, Romania

Keywords

Approximation of scattered data; interpolation on the sphere; Shepard operator; error estimations; least squares polynomials

Paper coordinates

A. Malina, T. Cătinaș, A spherical Shepard operator constructed using the least squares method, Numer. Algorithms, https://doi.org/10.1007/s11075-026-02544-z, accepted.

PDF

??

About this paper

Journal

Numerical Algorithms

Publisher Name

Springer Link

Print ISSN
1572-9265

 

Online ISSN

1017-1398

Paper (preprint) in HTML form

A spherical Shepard operator constructed using the least squares method

A spherical Shepard operator constructed using the least squares method

Abstract

We consider the problem of approximating scattered datasets on the sphere by using a modified form of the Shepard operator, improved by its combination with polynomials obtained using the least squares method. The theoretical results are supported by several numerical experiments. The effectiveness of the operator in solving real-data problems is demonstrated through numerous practical applications, such as the approximation of atmospheric pressure, temperature, wind speed and precipitation.

keywords
Approximation of scattered data, interpolation on the sphere, Shepard operator, error estimations, least squares polynomials

1 Introduction

One of the most important tools in data interpolation problems is the method introduced by Shepard in [23], together with many refinements of the method presented, for example, in [14], [15], [16], [18], [21], [22], [28].

The problem of approximating large sets of scattered data on the sphere is an active research subject due to its high applicability in practical problems where the sphere is considered to be a model of the Earth (see, e.g., [17], [19], [20]). These reasons led to the study of different spherical operators by means of Shepard-like methods. Some papers that address this problem are, for example, [1], [5]–[10], [13], [26], [27].

Let S2 denote the unit sphere in ℝ3, centered at the origin,

S2={𝐱=(x,y,z)∈ℝ3:x2+y2+z2=1},

and let 𝒳⊂S2 be a set of nodes 𝐱i, i=1,…,N, such that the values fi=f⁡(𝐱i), i=1,…,N, of an unknown function f:S2→ℝ are available on 𝒳.

The classical spherical Shepard operator is defined for a point 𝐱∈S2 as

Sμ⁢(𝐱)=∑i=1NAi,μ⁢(𝐱)⁢fi, (1)

with the basis functions Ai,μ given by

Ai,μ⁢(𝐱)=(g⁡(𝐱,𝐱i))−μ∑k=1N(g⁡(𝐱,𝐱k))−μ, (2)

for μ>0 and g representing the geodesic distance between 𝐱 and 𝐱i, i.e.,

g⁡(𝐱,𝐱i)=arccos⁡(𝐱⋅𝐱i), (3)

with ′′⋅′′ denoting the dot product of 𝐱 and 𝐱i.

One drawback of the operator in this form is its global nature, which can lead to low accuracy. To address this problem, a local variant has been introduced, which ensures that only the interpolation nodes that are located in the neighborhood of the evaluation point influence the construction of the operator. In this sense, the so-called modified spherical Shepard operator is defined by [13]

S⁡(𝐱)=∑i=1Nw¯i⁢(𝐱)⁢fi, (4)

with

w¯i⁢(𝐱)=wi⁢(𝐱)∑k=1Nwk⁢(𝐱). (5)

The weights wi are defined as

wi⁢(𝐱)=[(Riw−g⁡(𝐱,𝐱i))+Riw⁢g⁢(𝐱,𝐱i)]μ, (6)

considering Riw a radius of influence about the node i and g the geodesic distance (3) between 𝐱 and 𝐱i.

It is important to note that the basis functions w¯i are nonnegative and form a partition of unity, i.e.,

w¯i(𝐱)≥0,for all i=1,…,N,and ∑i=1Nw¯i(𝐱)=1.

To increase the degree of exactness of the spherical operator and its accuracy, the function values fi can be replaced by the values of an interpolation operator (see, e.g., [1], [6]) or by those of a radial basis function (see, e.g., [5], [10], [13]).

The purpose of this paper is to introduce a new local operator constructed using least squares polynomials obtained in the tangent planes at the given nodes. Unlike other methods proposed before, such as those in [1] or [6], this operator does not require partial derivatives in its construction. Moreover, although the interpolation properties are not preserved by using the least squares method, the advantages of this technique in terms of robustness to noisy or scattered data are well known in the literature.

The construction of the operator and the theoretical study of the approximation error are detailed in Section 2. Numerical tests performed on different datasets are presented in Section 3. The practical applicability of this class of operators is demonstrated through real-data problems involving the approximation of temperature, atmospheric pressure, wind speed, and precipitation. These applications are presented in Section 4.

2 A modified version of the Shepard operator constructed using the least squares technique

In the first part of this section, we propose a technique for constructing a new local spherical Shepard operator combined with least squares polynomials. The method consists of two steps. First, for each node 𝐱i, i=1,…,N, we construct a local polynomial approximation of total degree d in the tangent plane at 𝐱i, using the least squares method. To construct this polynomial, only the nearest neighbors of 𝐱i are considered. Second, for each point 𝐱∈S2, we define the Shepard operator using these local polynomials. The section ends with theoretical results regarding the approximation error of the proposed operator.

2.1 Construction of local least squares polynomials

Let 𝒳={𝐱i,i=1,…,N} be the set of given nodes. For each node 𝐱i∈𝒳, denote by pid the local polynomial of total degree d, constructed using the least squares method in the tangent plane at 𝐱i. The construction of pid involves the following steps:

  1. 1.

    Let 𝒫i be the set of neighbors for 𝐱i and initialize it with the closest m nodes to 𝐱i, where m=(d+22) is the dimension of the bivariate polynomial space ℙd⁢(ℝ2), to ensure that the approximation polynomial can be constructed.

  2. 2.

    Project all the nodes in 𝒫i onto the tangent plane at 𝐱i. Following some ideas from [12], the projection of a point 𝐱j∈𝒫i onto the tangent plane at 𝐱i is obtained in the following manner:

    1. (a)

      Construct the normal vector η𝐱i at 𝐱i. Since 𝐱i∈S2, it follows that η𝐱i=𝐱i;

    2. (b)

      Construct two orthonormal vectors {ei1,ei2} in the tangent plane at 𝐱i, i.e., vectors perpendicular to η𝐱i, such that ei1×ei2=𝐱i:

      ei1 =γ‖γ‖,where ⁢γ={ez×𝐱i,ez=(0,0,1),if ⁢ez∦𝐱i,ey×𝐱i,ey=(0,1,0),otherwise, 
      ei2 =𝐱i×ei1;
    3. (c)

      Obtain the coordinates 𝐮j=(uj,vj) of the projection of the point 𝐱j onto the tangent plane at 𝐱i as 𝐮j=ϕi⁢(𝐱j), where the function ϕi⁢(𝐱) is given as in [12], i.e.,

      ϕi⁢(𝐱)=[(𝐱−𝐱i)⋅ei1,(𝐱−𝐱i)⋅ei2]T. (7)

    Note that in the above writing, ′′⋅′′ represents the inner product in ℝ3 and ′′×′′ represents the cross product in ℝ3.

  3. 3.

    Construct the matrix Vi of the polynomial basis of degree d:

    Vi=[1u1v1u12u1⁢v1v12⋯u1d⋯v1d⋱⋱1um~vm~um~2um~⁢vm~vm~2⋯um~d⋯vm~d]∈ℝm~×m, (8)

    where m~=|𝒫i|≥m and the columns correspond to all monomials uα⁢vβ with α+β≤d.

  4. 4.

    If the condition number κ⁡(Vi)>κmax, where κmax is a fixed value, add to 𝒫i the next nearest node and repeat Steps 2-4, until the conditioning criterion is met, ensuring in this way numerical stability.

  5. 5.

    Construct the local polynomial pid as pid⁢(𝐱)=Pid⁢(ϕi⁢(𝐱)), with

    Pid⁢(𝐮)=∑α+β≤dciα,β⁢uα⁢vβ,for ⁢𝐮=(u,v). (9)

    The coefficients ciα,β,α+β≤d, are found by solving the following least squares problem:

    ciα,β=argminc∈ℝm∥Vic−f𝒫i∥2, (10)

    where f𝒫i represents the values of the function f at the nodes in 𝒫i.

  6. 6.

    Store the radius of influence of 𝐱i, i.e., Riw:=maxj∈𝒫i⁡g⁡(𝐱i,𝐱j), to subsequently use it in the construction of the modified Shepard operator.

2.2 Construction of the modified Shepard operator

For an evaluation point 𝐱∈S2, the modified Shepard operator combined with a least squares polynomial is defined as

S⁢Pd⁢(𝐱)=∑i=1Nw¯i⁢(𝐱)⁢pid⁢(𝐱), (11)

with pid⁢(𝐱) representing the local least squares polynomial associated with the node 𝐱i and given by (9), and w¯i⁢(𝐱) being the basis function defined in (5).

This method ensures that only the closest nodes to the evaluation point 𝐱 have an influence on the construction of the operator, and thus on the approximation results.

Remark 1.

For the preprocessing phase (construction of the local polynomials), the computational cost for each node is about 𝒪⁡(N⁢log⁡N), since we have to compute the distances to all the nodes, sort them, and solve a least squares system of small dimension on the resulting local patch. In total, the cost for all N nodes at this phase is 𝒪⁡(N2⁢log⁡N). For evaluating the Shepard operator at one point, the cost is about 𝒪⁡(N), so the cost for K evaluation points is 𝒪⁡(K⁢N). The overall computational cost of the proposed algorithm is 𝒪⁡(N2⁢log⁡N+K⁢N). This cost can be reduced by replacing the brute-force nearest-neighbor search with a spatial data structure suited to the sphere, which we believe represents a natural direction for future work to improve the method’s efficiency (see, e.g., [7] – [10]).

2.3 Error estimations

Theorem 1.

Let f∈Cd+1⁢(S2), 𝒳={𝐱1,…,𝐱N} be a set of given nodes and denote by h𝒳=sup𝐱∈S2mini=1,…,N⁡g⁡(𝐱,𝐱i) the fill distance and by q𝒳=12⁢mini≠j⁡g⁡(𝐱i,𝐱j) the separation distance. Assuming that:

  1. (A1)

    ∃α>0 such that h𝒳≤α⁢q𝒳 (quasi-uniformity of the nodes),

  2. (A2)

    Each matrix Vi,i=1,…,N, given in (8) has full-column rank,

  3. (A3)

    κ⁡(Vi)≤κmax, for every i=1,…,N,

  4. (A4)

    ∃β>0 such that R:=maxi=1,…,N⁡Riw≤β⁢h𝒳,

then, for the Shepard operator S⁢Pd introduced in (11) there exists a constant C>0, depending only on α,β,κmax,d and the geometry of S2, such that, for every 𝐱∈S2

|S⁢Pd⁢(𝐱)−f⁡(𝐱)|≤C⁢h𝒳d+1⁢‖f‖Cd+1⁢(S2). (12)
Proof.

Let 𝐱∈S2. For any f∈Cd+1⁢(S2), based on the partition of unity property of the basis functions w¯i and their nonnegativity, we have that

|S⁢Pd⁢(𝐱)−f⁡(𝐱)| =|∑i=1Nw¯i⁢(𝐱)⁢pid⁢(𝐱)−f⁡(𝐱)⁢∑i=1Nw¯i⁢(𝐱)|=|∑i=1Nw¯i⁢(𝐱)⁢(pid⁢(𝐱)−f⁡(𝐱))|
≤∑i=1N|w¯i⁢(𝐱)⁢(pid⁢(𝐱)−f⁡(𝐱))|=∑i=1Nw¯i⁢(𝐱)⁢|pid⁢(𝐱)−f⁡(𝐱)|. (13)

We need an estimate for |pid⁢(𝐱)−f⁡(𝐱)|, with i fixed. To this end, let B⁡(𝐱i,Riw) be the neighborhood of the node 𝐱i that has been used in the construction of the least squares polynomial pid in Subsection 2.1. Let ϕi:B⁡(𝐱i,Riw)⊂S2→ℝ2, ϕi⁢(𝐱)=[(𝐱−𝐱i)⋅ei1,(𝐱−𝐱i)⋅ei2]T be the projection function defined in (7) and set Di:=ϕi⁢(B⁡(𝐱i,Riw))⊂ℝ2 the projection onto the tangent plane at 𝐱i of the neighborhood B⁡(𝐱i,Riw) corresponding to 𝐱i.

According to [12], the mapping ϕi is a homeomorphism, so there exists ϕi−1:Di→S2. Defining the maps Fi:=f∘ϕi−1 and Pid:=pid∘ϕi−1 on Di, we have that

|pid⁢(𝐱)−f⁡(𝐱)|=|Pid⁢(ϕi⁢(𝐱))−Fi⁢(ϕi⁢(𝐱))|=|Pid⁢(u)−Fi⁢(u)|, (14)

for u=(u,v)∈Di, with 𝐮:=ϕi⁢(𝐱).

The polynomial Pid is the least squares polynomial corresponding to Fi. Let L be the least squares operator in Di. From (A2), it follows that the set Ai:=ϕi⁢(𝒫i), which contains the nodes used in the construction of the least squares polynomial, is unisolvent for ℙd⁢(ℝ2) (i.e., the only polynomial q∈ℙd⁢(ℝ2) with q⁡(𝐚)=0, for all 𝐚∈Ai, is the zero polynomial). Then, according to [4], L is a linear operator and it reproduces polynomials of degree ≤d, hence:

Pid=L⁡(Fi) =L⁡(Td⁢Fi+Rd⁢Fi)=L⁡(Td⁢Fi)+L⁡(Rd⁢Fi)=Td⁢Fi+L⁡(Rd⁢Fi),

where Td⁢Fi denotes the Taylor polynomial of degree d around the origin (because ϕi⁢(𝐱i)=0) and Rd⁢Fi is the corresponding remainder term.

Therefore,

|Pid⁢(u)−Fi⁢(u)| =|Td⁢Fi⁢(u)+L⁡(Rd⁢Fi)⁢(u)−Td⁢Fi⁢(u)−Rd⁢Fi⁢(u)| (15)
=|(L⁡(Rd⁢Fi)−Rd⁢Fi)⁢(u)|=|(L−I)⁢(Rd⁢Fi)⁢(u)|
≤‖L−I‖∞⁢|Rd⁢Fi⁢(u)|≤(1+‖L‖∞)⁢|Rd⁢Fi⁢(u)|,

where ‖L‖∞ is the induced operatorial norm [11].

For the Taylor remainder, we have the following bound [24]

|Rd⁢Fi⁢(u)|≤C1⁢‖u‖2d+1⁢‖Fi‖Cd+1⁢(Di).

Since the geodesic distance g can be expressed in terms of the Euclidean distance d, i.e., g⁡(𝐱,𝐲)=2⁢arcsin⁡d⁡(𝐱,𝐲)2, and, conversely, d⁡(𝐱,𝐲)=2⁢sin⁡g⁡(𝐱,𝐲)2, after an asymptotic expansion, it follows that the difference between the two distances can be very small for sufficiently close points (for more details, one can refer to [1, Section 5.1.]). Based on this observation and on assumption (A4), we get that

|Rd⁢Fi⁢(u)| ≤C1⁢(Riw)d+1⁢‖Fi‖Cd+1⁢(Di)≤C1⁢Rd+1⁢‖Fi‖Cd+1⁢(Di) (16)
≤C1⁢βd+1⁢h𝒳d+1⁢‖Fi‖Cd+1⁢(Di).

By the multivariate Faà di Bruno formula, any partial derivative of Fi:=f∘ϕi−1 of order α (|α|≤d+1) is a linear combination of the partial derivatives of f evaluated at ϕi−1⁢(u) multiplied by products of bounded partial derivatives of ϕi−1.

Hence, we can find a constant C2>0 depending only on d, the sphere’s curvature and R such that

‖Fi‖Cd+1⁢(Di)≤C2⁢‖f‖Cd+1⁢(S2). (17)

If we substitute (16) and (17) in (15), we get that

|Pid⁢(u)−Fi⁢(u)|≤C1⁢C2⁢(1+‖L‖∞)⁢βd+1⁢h𝒳d+1⁢‖f‖Cd+1⁢(S2). (18)

We now show that ‖L‖∞ is bounded. By (A3), we have that κ⁡(Vi)≤κmax. In particular, we consider κ⁡(Vi)=κ2⁢(Vi)=‖Vi‖2⁢‖Vi+‖2=σmax⁢(Vi)σmin⁢(Vi), where Vi+, σmax, and σmin denote, respectively, the Moore-Penrose pseudoinverse, the largest singular value, and the smallest singular value of Vi. Since ‖Vi‖2≥‖Vi⁢e1‖2, where e1 is the usual canonical vector, and since the first column of Vi∈ℝm~×m has all entries equal to 1, it follows, after some straightforward computations, that σmax⁢(Vi)=‖Vi‖2≥m~≥m. Hence, ‖Vi+‖2=1σmin⁢(Vi)=κ⁡(Vi)σmax⁢(Vi)≤κmaxm=:C~.

Now recall that Pid⁢(𝐮)=∑α+β≤dciα,β⁢uα⁢vβ, which can be written as Pid⁢(𝐮)=ν⁡(𝐮)⁢ci, for ν⁡(𝐮)=(uα⁢vβ)α+β≤d and ci=(ciα,β)α+β≤dT. Because the coefficients ci are the solution of the least squares problem (10), they are obtained as ci=Vi+f𝒫i, so Pid(𝐮)=ν(𝐮)Vi+f𝒫i. Since sup𝐮∈Di‖ν⁡(𝐮)‖2 is bounded, general norm inequalities and norm equivalence yield ‖L‖∞=sup‖f𝒫i‖∞≤1sup𝐮∈Di|Pid⁢(𝐮)|≤η⁢‖Vi+‖2≤η⁢C~, for some η>0. Letting C3:=1+η⁢C~, relation (18) becomes

|Pid⁢(u)−Fi⁢(u)|≤C1⁢C2⁢C3⁢βd+1⁢h𝒳d+1⁢‖f‖Cd+1⁢(S2).

Finally, denoting C:=C1⁢C2⁢C3⁢βd+1, recalling the equality in (14) and returning to (13), we obtain that

|S⁢Pd⁢(𝐱)−f⁡(𝐱)| ≤∑i=1Nw¯i⁢(𝐱)⁢|pid⁢(𝐱)−f⁡(𝐱)|≤C⁢h𝒳d+1|f|∑i=1NCd+1⁢(S2)⁡w¯i⁢(𝐱)
=C⁢h𝒳d+1⁢‖f‖Cd+1⁢(S2),

which completes the proof. ∎

3 Numerical results on test functions

We begin this section by recalling the definitions of Halton and Leja points, which serve as nodes in our numerical experiments performed on various test functions.

3.1 Halton points

The Halton points are pseudo-random points defined in [25]. Their construction is based on the expansion using a prime base p of a natural number k,

k=a0+a1p+…+ampm,ai∈{0,1,…,p−1},i=0,…,m,

and on the function

αp⁢(k)=a0p+a1p2+…+ampm+1.

When p=2, one obtains the sequence α2⁢(k), for k=0,1,2,…, which is called the Van der Corput sequence.

The spherical Halton points are obtained by applying the following mappings to two p-adic Van der Corput sequences with different prime bases:

(αp1⁢(k),αp2⁢(k))↦(ϕ,t)↦(1−t2⁢cos⁡ϕ,1−t2⁢sin⁡ϕ,t).

The first mapping represents a linear scaling to a cylindrical domain, ϕ∈[0,2⁢π), t∈[−1,1], and the second one is a z-preserving radial projection from the unit cylinder C={𝐱=(x,y,z):x2+y2=1,|z|≤1} to the unit sphere S2, as emphasized in [25].

3.2 Leja points

Leja points are known to be suitable for polynomial interpolation and were extensively studied in [2]. We first recall some notions and definitions that appear in the construction of these points.

Let K⊂ℝd be a compact set and K~ a compact subset of K. Consider also an admissible mesh of K, that is, a sequence of finite discrete subsets 𝒜n of K such that the following polynomial inequality

‖p‖K≤C⁢‖p‖𝒜n,∀p∈ℙnd⁢(K),

holds for a constant C>0. Here, ℙnd⁢(K) represents the space of polynomials in d variables of total degree at most n on K, ‖p‖K=sup𝐱∈K|p⁡(𝐱)| and the cardinal of 𝒜n grows at most polynomially in n. Another way to construct good low-cardinality meshes for polynomial approximation, according to [4], is to use weakly admissible meshes, which use a non-constant C=Cn that increases at most polynomially in n.

Letting M:=dim(ℙnd⁢(K)) and 𝐩=(pi)1≤i≤M be an ordered set of basis polynomials of ℙnd⁢(K), the sequence {𝐱i∈K~:i=1,…,M} of Leja points is defined as follows (see, e.g., [2], [3]):

𝐱1 =arg⁡max𝐱∈K~⁢|p1⁢(𝐱)|,
𝐱k =arg⁡max⁡det𝐱∈K~⁡|V⁡(𝐱1,…,𝐱k−1,𝐱,p1,…,pk)|,k≥2.

As noted in [2], a computationally less expensive approach to obtain the points is by choosing K~=𝒜n, an admissible mesh, with n sufficiently large, resulting in the so-called discrete Leja points. To construct these points, the authors proposed in [2] a greedy algorithm for discrete maximization on (weakly) admissible meshes, using the LU factorization with partial (row) pivoting.

3.3 Numerical results

To illustrate the theoretical results, we compute the approximation errors of the operator S⁢P2 introduced in (11), considering a second-degree least squares polynomial, i.e., pi2. According to (9), this polynomial has the form

pi2⁢(𝐱)=Pi2⁢(𝐮)=ci0,0+ci1,0⁢u+ci0,1⁢v+ci2,0⁢u2+ci1,1⁢u⁢v+ci0,2⁢v2,

where 𝐮=(u,v)=:ϕi⁢(𝐱) denotes the coordinates of the projection of 𝐱∈S2 onto the tangent plane at 𝐱i and ciα,β,α+β≤2, are obtained by solving the least squares problem (10), for each i=1,…,N.

We use the following test functions (see, e.g., [15], [16], [21], [22]):

f1⁢(x,y,z)=34⁢exp⁡[−(9⁢x−2)2+(9⁢y−2)2+(9⁢z−2)24]+34⁢exp⁡[−(9⁢x+1)249−(9⁢y+1)210−(9⁢z+1)210]
+12⁢exp⁡[−(9⁢x−7)2+(9⁢y−3)2+(9⁢z−5)24]−15⁢exp⁡[−(9⁢x−4)2−(9⁢y−7)2−(9⁢z−5)2],
f2⁢(x,y,z)=[1.25+cos⁡(5.4⁢y)]⁢cos⁡(6⁢z)6+6⁢(3⁢x−1)2,
f3⁢(x,y,z)=13⁢exp⁡[−8116⁢((x−0.5)2+(y−0.5)2+(z−0.5)2)],
f4⁢(x,y,z)=13⁢exp⁡[−814⁢((x−0.5)2+(y−0.5)2+(z−0.5)2)],
f5⁢(x,y,z)=110⁢(exp⁡(x)+2⁢exp⁡(y+z)),
f6⁢(x,y,z)=sin⁡x⁢sin⁡y⁢sin⁡z,
f7⁢(x,y,z)=sin⁡(x+y)+sin⁡x⁢z,
f8⁢(x,y,z)=110⁢(9⁢x3−2⁢x2⁢y+3⁢x⁢y2−4⁢y3+2⁢z3−x⁢y⁢z).

We consider two sets of nodes, namely the Halton nodes and the Leja points, whose construction was described at the beginning of this section. For each node type, we use sets of N=500, 1000, 2000 nodes and evaluate the operator on 2000 and 5000 spiral points, offering in this way a fair representation of the sphere. The maximum absolute errors (em⁢a⁢x), mean absolute errors (em⁢e⁢a⁢n) and root mean square errors (er⁢m⁢s) are displayed in Tables 1 and 2 for the Halton-nodes case and in Tables 3 and 4 for the Leja-nodes case. Both choices of nodes deliver good numerical results, even if the number of nodes is less than or equal to the number of evaluation points. As expected, in most cases, a greater number of nodes produces better approximation results.

f 500 nodes 1000 nodes 2000 nodes
f1 em⁢a⁢x 0.1188 0.0743 0.0615
em⁢e⁢a⁢n 0.0019 0.0011 7.8425⁢e−04
er⁢m⁢s 0.0082 0.0050 0.0037
CPU (s) 0.14 0.32 1.03
f2 em⁢a⁢x 0.0362 0.0209 0.0172
em⁢e⁢a⁢n 0.0034 0.0019 0.0013
er⁢m⁢s 0.0059 0.0033 0.0023
CPU (s) 0.15 0.33 1.04
f3 em⁢a⁢x 0.0109 0.0044 0.0023
em⁢e⁢a⁢n 3.0674⁢e−04 1.5489⁢e−04 8.8493⁢e−05
er⁢m⁢s 8.8938⁢e−04 4.5581⁢e−04 2.4750⁢e−04
CPU (s) 0.16 0.32 1.09
f4 em⁢a⁢x 0.0248 0.0259 0.0129
em⁢e⁢a⁢n 3.8952⁢e−04 2.2571⁢e−04 1.6275⁢e−04
er⁢m⁢s 0.0019 0.0012 8.0082⁢e−04
CPU (s) 0.16 0.33 1.03
f5 em⁢a⁢x 0.0013 5.5842⁢e−04 3.5511⁢e−04
em⁢e⁢a⁢n 1.4802⁢e−04 7.6085⁢e−05 4.2507⁢e−05
er⁢m⁢s 2.2550⁢e−04 1.1491⁢e−04 6.2839⁢e−05
CPU (s) 0.24 0.32 1.03
f6 em⁢a⁢x 0.0033 0.0015 7.3800⁢e−04
em⁢e⁢a⁢n 6.0406⁢e−04 3.1387⁢e−04 1.7396⁢e−04
er⁢m⁢s 7.8209⁢e−04 3.9783⁢e−04 2.2142⁢e−04
CPU (s) 0.14 0.32 1.02
f7 em⁢a⁢x 0.0037 0.0021 0.0011
em⁢e⁢a⁢n 6.4486⁢e−04 3.3421⁢e−04 1.9402⁢e−04
er⁢m⁢s 8.5561⁢e−04 4.4443⁢e−04 2.5311⁢e−04
CPU (s) 0.15 0.33 1.10
f8 em⁢a⁢x 0.0054 0.0028 0.0012
em⁢e⁢a⁢n 9.3223⁢e−04 4.7057⁢e−04 2.7175⁢e−04
er⁢m⁢s 0.0012 6.0794⁢e−04 3.4115⁢e−04
CPU (s) 0.15 0.35 1.18
Table 1: Approximation errors for S⁢P2, Halton nodes and 2000 spiral evaluation points.
f 500 nodes 1000 nodes 2000 nodes
f1 em⁢a⁢x 0.1188 0.0778 0.0680
em⁢e⁢a⁢n 0.0019 0.0011 7.8949⁢e−04
er⁢m⁢s 0.0083 0.0050 0.0037
CPU (s) 0.22 0.42 1.29
f2 em⁢a⁢x 0.0441 0.0265 0.0171
em⁢e⁢a⁢n 0.0034 0.0019 0.0012
er⁢m⁢s 0.0059 0.0033 0.0023
CPU (s) 0.25 0.43 1.24
f3 em⁢a⁢x 0.0120 0.0045 0.0027
em⁢e⁢a⁢n 3.0439⁢e−04 1.5374⁢e−04 9.0694⁢e−05
er⁢m⁢s 8.9169⁢e−04 4.4729⁢e−04 2.5413⁢e−04
CPU (s) 0.23 0.44 1.20
f4 em⁢a⁢x 0.0318 0.0275 0.0155
em⁢e⁢a⁢n 3.8881⁢e−04 2.2783⁢e−04 1.5855⁢e−04
er⁢m⁢s 0.0020 0.0013 7.9346⁢e−04
CPU (s) 0.22 0.43 1.20
f5 em⁢a⁢x 0.0013 5.7466⁢e−04 4.3314⁢e−04
em⁢e⁢a⁢n 1.4781⁢e−04 7.5524⁢e−05 4.1854⁢e−05
er⁢m⁢s 2.2667⁢e−04 1.1244⁢e−04 6.2536⁢e−05
CPU (s) 0.22 0.43 1.19
f6 em⁢a⁢x 0.0033 0.0016 9.3993⁢e−04
em⁢e⁢a⁢n 6.0016⁢e−04 3.1091⁢e−04 1.7115⁢e−04
er⁢m⁢s 7.8026⁢e−04 3.9393⁢e−04 2.1741⁢e−04
CPU (s) 0.22 0.49 1.21
f7 em⁢a⁢x 0.0040 0.0021 0.0012
em⁢e⁢a⁢n 6.4050⁢e−04 3.3176⁢e−04 1.9044⁢e−04
er⁢m⁢s 8.5642⁢e−04 4.4089⁢e−04 2.5162⁢e−04
CPU (s) 0.25 0.43 1.24
f8 em⁢a⁢x 0.0054 0.0027 0.0014
em⁢e⁢a⁢n 9.2436⁢e−04 4.6507⁢e−04 2.6901⁢e−04
er⁢m⁢s 0.0012 6.0337⁢e−04 3.4149⁢e−04
CPU (s) 0.22 0.44 1.29
Table 2: Approximation errors for S⁢P2, Halton nodes and 5000 spiral evaluation points.
f 500 nodes 1000 nodes 2000 nodes
f1 em⁢a⁢x 0.0699 0.0520 0.0472
em⁢e⁢a⁢n 0.0013 8.9501⁢e−04 7.7724⁢e−04
er⁢m⁢s 0.0052 0.0037 0.0034
CPU (s) 0.12 0.36 1.03
f2 em⁢a⁢x 0.0319 0.0182 0.0148
em⁢e⁢a⁢n 0.0028 0.0018 0.0013
er⁢m⁢s 0.0049 0.0030 0.0023
CPU (s) 0.14 0.32 1.05
f3 em⁢a⁢x 0.0062 0.0034 0.0022
em⁢e⁢a⁢n 2.2741⁢e−04 1.2912⁢e−04 8.4569⁢e−05
er⁢m⁢s 6.1633⁢e−04 3.6050⁢e−04 2.3224⁢e−04
CPU (s) 0.13 0.41 1.02
f4 em⁢a⁢x 0.0270 0.0212 0.0162
em⁢e⁢a⁢n 3.6189⁢e−04 2.6144⁢e−04 1.8832⁢e−04
er⁢m⁢s 0.0018 0.0013 9.5403⁢e−04
CPU (s) 0.15 0.32 1.03
f5 em⁢a⁢x 9.3668⁢e−04 5.7557⁢e−04 2.5473⁢e−04
em⁢e⁢a⁢n 1.0918⁢e−04 5.7835⁢e−05 3.3449⁢e−05
er⁢m⁢s 1.6455⁢e−04 8.5514⁢e−05 4.8278⁢e−05
CPU (s) 0.14 0.32 1.02
f6 em⁢a⁢x 0.0031 0.0014 5.5184⁢e−04
em⁢e⁢a⁢n 4.5373⁢e−04 2.3727⁢e−04 1.3355⁢e−04
er⁢m⁢s 5.8780⁢e−04 3.0202⁢e−04 1.7091⁢e−04
CPU (s) 0.18 0.31 1.13
f7 em⁢a⁢x 0.0030 0.0013 7.5843⁢e−04
em⁢e⁢a⁢n 4.8235⁢e−04 2.5471⁢e−04 1.4400⁢e−04
er⁢m⁢s 6.3871⁢e−04 3.3232⁢e−04 1.8738⁢e−04
CPU (s) 0.16 0.35 1.11
f8 em⁢a⁢x 0.0033 0.0021 9.1786⁢e−04
em⁢e⁢a⁢n 7.1797⁢e−04 3.7992⁢e−04 2.2248⁢e−04
er⁢m⁢s 9.1565⁢e−04 4.8712⁢e−04 2.7571⁢e−04
CPU (s) 0.17 0.32 1.09
Table 3: Approximation errors for S⁢P2, Leja nodes and 2000 spiral evaluation points.
f 500 nodes 1000 nodes 2000 nodes
f1 em⁢a⁢x 0.0744 0.0640 0.0630
em⁢e⁢a⁢n 0.0013 8.8234⁢e−04 7.7311⁢e−04
er⁢m⁢s 0.0052 0.0036 0.0033
CPU (s) 0.21 0.44 1.27
f2 em⁢a⁢x 0.0318 0.0209 0.0154
em⁢e⁢a⁢n 0.0028 0.0017 0.0013
er⁢m⁢s 0.0049 0.0030 0.0023
CPU (s) 0.20 0.43 1.30
f3 em⁢a⁢x 0.0066 0.0036 0.0022
em⁢e⁢a⁢n 2.2611⁢e−04 1.3025⁢e−04 8.3659⁢e−05
er⁢m⁢s 6.1565⁢e−04 3.6707⁢e−04 2.2996⁢e−04
CPU (s) 0.20 0.43 1.26
f4 em⁢a⁢x 0.0282 0.0211 0.0168
em⁢e⁢a⁢n 3.5994⁢e−04 2.6283⁢e−04 1.8504⁢e−04
er⁢m⁢s 0.0018 0.0013 9.4753⁢e−04
CPU (s) 0.21 0.43 1.30
f5 em⁢a⁢x 9.5052⁢e−04 6.7559⁢e−04 2.5406⁢e−04
em⁢e⁢a⁢n 1.0945⁢e−04 5.7708⁢e−05 3.4160⁢e−05
er⁢m⁢s 1.6427⁢e−04 8.5355⁢e−05 4.8880⁢e−05
CPU (s) 0.20 0.46 1.30
f6 em⁢a⁢x 0.0031 0.0014 5.7748⁢e−04
em⁢e⁢a⁢n 4.5544⁢e−04 2.3817⁢e−04 1.3384⁢e−04
er⁢m⁢s 5.8891⁢e−04 3.0449⁢e−04 1.7060⁢e−04
CPU (s) 0.20 0.44 1.28
f7 em⁢a⁢x 0.0029 0.0016 9.4177⁢e−04
em⁢e⁢a⁢n 4.8338⁢e−04 2.5490⁢e−04 1.4527⁢e−04
er⁢m⁢s 6.4003⁢e−04 3.3344⁢e−04 1.8777⁢e−04
CPU (s) 0.20 0.44 1.34
f8 em⁢a⁢x 0.0034 0.0022 8.8787⁢e−04
em⁢e⁢a⁢n 7.1403⁢e−04 3.8294⁢e−04 2.2537⁢e−04
er⁢m⁢s 9.1400⁢e−04 4.8881⁢e−04 2.7775⁢e−04
CPU (s) 0.21 0.44 1.28
Table 4: Approximation errors for S⁢P2, Leja nodes and 5000 spiral evaluation points.

We also plot the mean absolute errors (em⁢e⁢a⁢n), considering N=500, 1000, 2000, 3000, 4000, 5000 nodes (see Halton case in Figure 1 and Leja case in Figure 2) to evaluate the operator S⁢P2 on 10000 spiral points.

Refer to caption
Figure 1: Semilogy plot of em⁢e⁢a⁢n for S⁢P2 considering N=500,1000,2000,…,5000 Halton nodes and 10000 spiral evaluation points.
Refer to caption
Figure 2: Semilogy plot of em⁢e⁢a⁢n for S⁢P2 considering N=500,1000,2000,…,5000 Leja nodes and 10000 spiral evaluation points.

We now present, in Figure 3, the log-log plots of the root mean square error (er⁢m⁢s) and the maximum error (em⁢a⁢x) versus the fill distance h𝒳, obtained using N=1000,2000,4000,8000 Halton nodes and all eight test functions, evaluated at K=105 spiral points. For reference, the black dashed line indicates the theoretical cubic trend.

Refer to caption
(a) er⁢m⁢s
Refer to caption
(b) em⁢a⁢x
Figure 3: Log-log plots of the er⁢m⁢s and em⁢a⁢x of the proposed operator S⁢P2.

Next, we compare the operator S⁢P2, introduced in (11) with several operators from the literature: the classical spherical Shepard operator Sμ given in (1), the modified variant S defined in (4), the Shepard-Bernoulli operator SB1 introduced in [6], and the augmented Shepard operators ST⁢P⁢S and SI⁢M⁢Q, constructed using the thin-plate spline and inverse multiquadric spherical radial basis functions (SBFs) [5].

The Shepard-Bernoulli operator SBm of order m is defined in Eq. (4.4) in [6]. Here, we consider the case m=1, since higher orders also require the partial derivatives of the function. Consider the Delaunay triangulation T of the sphere and K the number of triangles that form the triangulation. We let 𝒯i,i=1,…,K, denote a triangle from the triangulation, with vertices 𝐱i1,𝐱i2,𝐱i3, such that each node 𝐱i,i=1,…,N, is a vertex of at least one triangle, i.e.,

⋃𝑖⁢{i1,i2,i3}={1,…,N}.

Consider also for a point 𝐱, its barycentric coordinates λ1,λ2,λ3 relative to 𝒯i, which are defined as

λ1=𝒜⁡(𝐱,𝐱i2,𝐱i3)𝒜⁡(𝐱i1,𝐱i2,𝐱i3),λ2=𝒜⁡(𝐱i1,𝐱,𝐱i3)𝒜⁡(𝐱i1,𝐱i2,𝐱i3),λ3=𝒜⁡(𝐱i1,𝐱i2,𝐱)𝒜⁡(𝐱i1,𝐱i2,𝐱i3),

with 𝒜⁡(𝐱,𝐲,𝐳) representing the signed area of the triangle with vertices 𝐱,𝐲,𝐳.

The Shepard-Bernoulli operator of order 1 is given as

SB1⁢[f]⁢(𝐱)=∑i=1KΦi,μ⁢(𝐱)⁢B1𝒯i⁢[f]⁢(𝐱),for all ⁢𝐱∈S2, (19)

with

B1𝒯i⁢[f]⁢(𝐱)=λ1⁢f⁢(𝐱i1)+λ2⁢f⁢(𝐱i2)+λ3⁢f⁢(𝐱i3),

and

Φi,μ(𝐱)=∏j=13(g⁡(𝐱,𝐱ij))−μ∑k=1K∏j=13(g⁡(𝐱,𝐱kj))−μ,i=1,…,K.

The augmented spherical radial basis function operators ST⁢P⁢S and SI⁢M⁢Q are defined in Eq. (2.10) and Eq. (2.12) in [5], respectively, as

SS⁢B⁢F⁢(𝐱)=∑i=1Nw¯i⁢(𝐱)⁢Zi⁢(𝐱). (20)

Here, the augmented local zonal basis function Zi is of the following particular form

Zi⁢(𝐱)=∑j∈Iiaji⁢ψ⁢(rj)+Ai⁢x+Bi⁢y+Ci⁢z,

for rj=2⁢sin⁡g⁡(𝐱,𝐱j)2 and with ψ being defined in the thin-plate spline case as ψ⁡(r)=r2⁢log⁡r and in the inverse multiquadric case as ψ(r)=(r2+c2)−1/2.

The index set Ii contains the nZ nearest neighbors of the node 𝐱i and the coefficients aji,Ai,Bi,Ci are obtained by imposing general interpolation conditions and constraints from the radial basis function theory. For more details, one can consult [5] and the references therein.

We considered 600 spiral evaluation points and two sets of 1000 nodes: Halton nodes and Leja nodes. The maximum absolute errors (em⁢a⁢x), mean absolute errors (em⁢e⁢a⁢n), root mean square errors (er⁢m⁢s), and CPU times are presented in Table 5 for the Halton-nodes case and in Table 6 for the Leja-nodes case. We can observe that the proposed operator outperforms the classical and modified variants of the Shepard operator, as well as the variants based on spherical radial basis functions, in terms of approximation accuracy. It is also faster than both SBF-augmented variants. Compared with the Bernoulli variant, the errors of the two operators are, in most cases, of the same order, while the operator introduced in this paper requires less CPU time. Nonetheless, the Bernoulli variant is constructed somewhat differently from the least squares spherical Shepard operator, since it relies on the Delaunay triangulation of the sphere and different basis functions.

f S⁢P2 Sμ S SB1 ST⁢P⁢S SI⁢M⁢Q
f1 em⁢a⁢x 4.2844⁢e−02 2.6165⁢e−01 1.5717⁢e−01 1.8910⁢e−02 6.9179⁢e−02 1.6153⁢e−01
em⁢e⁢a⁢n 9.4263⁢e−04 9.3958⁢e−03 4.3841⁢e−03 4.7674⁢e−04 1.8020⁢e−03 8.0074⁢e−03
er⁢m⁢s 3.6525⁢e−03 2.1956⁢e−02 1.2305⁢e−02 1.8505⁢e−03 4.8309⁢e−03 1.6126⁢e−02
CPU(s) 0.35 0.01 0.02 1.14 2.62 1.37
f2 em⁢a⁢x 2.5482⁢e−02 1.8188⁢e−01 1.1647⁢e−01 1.2530⁢e−02 4.5354⁢e−02 1.0276⁢e+00
em⁢e⁢a⁢n 2.0188⁢e−03 2.0971⁢e−02 1.1184⁢e−02 9.0765⁢e−04 3.7924⁢e−03 2.1668⁢e−02
er⁢m⁢s 3.6406⁢e−03 3.5917⁢e−02 2.0733⁢e−02 1.5974⁢e−03 7.3353⁢e−03 5.4345⁢e−02
CPU(s) 0.39 0.02 0.02 1.14 2.38 1.41
f3 em⁢a⁢x 3.7241⁢e−03 1.2053⁢e−01 9.4589⁢e−02 1.9209⁢e−03 2.8877⁢e−02 3.0628⁢e−02
em⁢e⁢a⁢n 1.4247⁢e−04 7.0648⁢e−03 2.9859⁢e−03 9.9969⁢e−05 1.1603⁢e−03 3.3556⁢e−03
er⁢m⁢s 3.8954⁢e−04 1.5743⁢e−02 8.5102⁢e−03 2.7497⁢e−04 2.3679⁢e−03 5.6322⁢e−03
CPU(s) 0.39 0.02 0.02 1.13 2.37 1.39
f4 em⁢a⁢x 1.9779⁢e−02 1.0216⁢e−01 7.1479⁢e−02 6.9621⁢e−03 3.4464⁢e−02 7.4249⁢e−02
em⁢e⁢a⁢n 3.0814⁢e−04 2.1346⁢e−03 8.4473⁢e−04 9.5668⁢e−05 5.2074⁢e−04 1.9190⁢e−03
er⁢m⁢s 1.5788⁢e−03 7.9525⁢e−03 4.8161⁢e−03 4.9552⁢e−04 2.2541⁢e−03 6.0742⁢e−03
CPU(s) 0.37 0.02 0.02 1.12 2.35 1.38
f5 em⁢a⁢x 6.1773⁢e−04 1.3363⁢e−01 5.3874⁢e−02 1.6490⁢e−03 3.4252⁢e−02 1.4907⁢e−01
em⁢e⁢a⁢n 7.0624⁢e−05 3.2469⁢e−02 9.8482⁢e−03 1.5609⁢e−04 1.3501⁢e−02 6.0934⁢e−03
er⁢m⁢s 1.0261⁢e−04 4.1852⁢e−02 1.3729⁢e−02 2.4934⁢e−04 1.5205⁢e−02 1.0357⁢e−02
CPU(s) 0.38 0.02 0.02 1.13 2.33 1.39
f6 em⁢a⁢x 1.5756⁢e−03 8.2954⁢e−02 5.7864⁢e−02 2.1282⁢e−03 1.3111⁢e−02 1.5033⁢e−01
em⁢e⁢a⁢n 3.0928⁢e−04 2.3689⁢e−02 1.1339⁢e−02 3.5138⁢e−04 2.2287⁢e−03 5.4558⁢e−03
er⁢m⁢s 4.0285⁢e−04 3.0526⁢e−02 1.5355⁢e−02 5.0863⁢e−04 3.0883⁢e−03 1.2103⁢e−02
CPU(s) 0.37 0.02 0.02 1.11 2.51 1.43
f7 em⁢a⁢x 1.7103⁢e−03 3.7667⁢e−01 1.7851⁢e−01 7.6343⁢e−03 9.6883⁢e−03 5.2388⁢e−01
em⁢e⁢a⁢n 3.2319⁢e−04 1.1697⁢e−01 3.9476⁢e−02 6.7000⁢e−04 1.9147⁢e−03 1.2812⁢e−02
er⁢m⁢s 4.2941⁢e−04 1.4386⁢e−01 5.2219⁢e−02 1.0975⁢e−03 2.6391⁢e−03 3.5361⁢e−02
CPU(s) 0.40 0.02 0.02 1.12 2.38 1.39
f8 em⁢a⁢x 2.4550⁢e−03 3.1742⁢e−01 1.7672⁢e−01 9.9741⁢e−03 2.1091⁢e−02 1.2385⁢e−01
em⁢e⁢a⁢n 4.6359⁢e−04 6.3610⁢e−02 2.2056⁢e−02 6.0279⁢e−04 2.8750⁢e−03 8.9861⁢e−03
er⁢m⁢s 5.9780⁢e−04 8.5717⁢e−02 3.2149⁢e−02 9.2897⁢e−04 3.9295⁢e−03 1.5182⁢e−02
CPU(s) 0.40 0.02 0.02 1.09 2.33 1.38
Table 5: Comparison of the proposed operator with existing spherical Shepard-type operators for N=1000 Halton nodes and K=600 spiral evaluation points.
f S⁢P2 Sμ S SB1 ST⁢P⁢S SI⁢M⁢Q
f1 em⁢a⁢x 5.0652⁢e−02 2.4746⁢e−01 1.4213⁢e−01 1.6510⁢e−02 5.7299⁢e−02 3.5551⁢e−01
em⁢e⁢a⁢n 9.9952⁢e−04 9.3787⁢e−03 5.0177⁢e−03 3.8139⁢e−04 2.0716⁢e−03 7.4838⁢e−03
er⁢m⁢s 4.0867⁢e−03 2.1379⁢e−02 1.2393⁢e−02 1.5296⁢e−03 4.5680⁢e−03 1.9321⁢e−02
CPU(s) 0.36 0.01 0.02 1.17 2.35 1.48
f2 em⁢a⁢x 1.9069⁢e−02 1.5996⁢e−01 1.0394⁢e−01 7.4687⁢e−03 4.3738⁢e−02 8.2345⁢e−01
em⁢e⁢a⁢n 1.6324⁢e−03 2.0469⁢e−02 1.1104⁢e−02 7.6351⁢e−04 3.9890⁢e−03 1.6106⁢e−02
er⁢m⁢s 2.8269⁢e−03 3.3385⁢e−02 1.9272⁢e−02 1.2860⁢e−03 7.0424⁢e−03 5.2110⁢e−02
CPU(s) 0.38 0.01 0.02 1.19 2.42 1.41
f3 em⁢a⁢x 3.4055⁢e−03 1.3058⁢e−01 6.9567⁢e−02 1.8677⁢e−03 1.8362⁢e−02 5.7758⁢e−01
em⁢e⁢a⁢n 1.1625⁢e−04 7.1136⁢e−03 3.0338⁢e−03 9.0613⁢e−05 1.2617⁢e−03 7.1505⁢e−03
er⁢m⁢s 3.3757⁢e−04 1.6239⁢e−02 7.3130⁢e−03 2.6373⁢e−04 2.0301⁢e−03 2.5132⁢e−02
CPU(s) 0.39 0.02 0.02 1.11 2.33 1.39
f4 em⁢a⁢x 2.4744⁢e−02 1.3839⁢e−01 9.5982⁢e−02 6.6942⁢e−03 3.6230⁢e−02 1.0602⁢e+00
em⁢e⁢a⁢n 2.6532⁢e−04 2.2372⁢e−03 1.4267⁢e−03 1.1091⁢e−04 6.5942⁢e−04 1.0961⁢e−02
er⁢m⁢s 1.5503⁢e−03 9.3145⁢e−03 5.8678⁢e−03 6.1772⁢e−04 2.1021⁢e−03 4.5921⁢e−02
CPU(s) 0.41 0.02 0.02 1.12 2.37 1.40
f5 em⁢a⁢x 4.5504⁢e−04 1.4185⁢e−01 6.1697⁢e−02 8.8337⁢e−04 3.0269⁢e−02 1.2052⁢e−01
em⁢e⁢a⁢n 5.4911⁢e−05 3.2267⁢e−02 1.0120⁢e−02 1.0895⁢e−04 1.4404⁢e−02 8.2109⁢e−03
er⁢m⁢s 8.0683⁢e−05 4.1651⁢e−02 1.4093⁢e−02 1.7115⁢e−04 1.5559⁢e−02 1.1553⁢e−02
CPU(s) 0.39 0.02 0.02 1.11 2.35 1.45
f6 em⁢a⁢x 8.6110⁢e−04 7.7001⁢e−02 5.3266⁢e−02 1.7731⁢e−03 9.7749⁢e−03 2.4354⁢e−01
em⁢e⁢a⁢n 2.1054⁢e−04 2.3635⁢e−02 1.1758⁢e−02 2.2980⁢e−04 2.1163⁢e−03 1.2332⁢e−02
er⁢m⁢s 2.7008⁢e−04 3.0205⁢e−02 1.5590⁢e−02 3.2567⁢e−04 2.8165⁢e−03 1.8844⁢e−02
CPU(s) 0.36 0.02 0.02 1.14 2.36 1.44
f7 em⁢a⁢x 1.1749⁢e−03 3.4849⁢e−01 1.4162⁢e−01 2.9405⁢e−03 1.0245⁢e−02 2.0577⁢e−01
em⁢e⁢a⁢n 2.3674⁢e−04 1.1610⁢e−01 3.6175⁢e−02 4.1571⁢e−04 1.9228⁢e−03 9.5847⁢e−03
er⁢m⁢s 3.0839⁢e−04 1.4119⁢e−01 4.6922⁢e−02 5.9389⁢e−04 2.5098⁢e−03 1.9008⁢e−02
CPU(s) 0.39 0.02 0.02 1.14 2.40 1.38
f8 em⁢a⁢x 1.6532⁢e−03 2.5218⁢e−01 1.1544⁢e−01 2.3640⁢e−03 1.6385⁢e−02 3.1483⁢e−01
em⁢e⁢a⁢n 3.7162⁢e−04 6.4308⁢e−02 2.2937⁢e−02 4.0944⁢e−04 3.2699⁢e−03 8.7628⁢e−03
er⁢m⁢s 4.7946⁢e−04 8.5110⁢e−02 3.2308⁢e−02 5.7436⁢e−04 4.3267⁢e−03 2.1019⁢e−02
CPU(s) 0.45 0.02 0.03 1.21 2.42 1.48
Table 6: Comparison of the proposed operator with existing spherical Shepard-type operators for N=1000 Leja nodes and K=600 spiral evaluation points.

4 Applications to real-data problems

To illustrate the practical benefits of the method introduced in this paper, we apply the proposed operator to reconstruct meteorological data. We use a dataset downloaded from https://www.kaggle.com/datasets/jakubkrasuski/global-weather-data-conditions-for-209579-cities, which contains current weather conditions for 209579 cities around the world (the dataset used is from February 2025). As noted by the authors, the dataset includes numerous measurements of meteorological variables such as temperature, pressure, humidity, wind characteristics, cloudiness, and precipitation. We randomly selected N sample points from the dataset as nodes and reconstructed the corresponding values at K evaluation points. Although the nodes were randomly selected, we ensured that they remained sufficiently uniformly distributed such that no geographic regions were left unsampled. However, in cases where certain evaluation points were located far away from all other nodes (i.e., no neighboring node lay within a radius of 10o), we treated them as outliers and excluded them from the error analysis, since an accurate reconstruction is not possible in the absence of nearby supporting data. The percentage of discarded points was very small (at most 0.9% for N=1000 nodes and at most 0.2% for N=2000 nodes), and it did not meaningfully influence the results. Below, we present different practical applications of the Shepard operator S⁢P2 introduced in (11). The operator is constructed using local quadratic polynomials pi2, as described in (9), and various choices for the number of nodes and evaluation points. In each case, we compute the mean absolute error (em⁢e⁢a⁢n) and the root mean square error (er⁢m⁢s).

4.1 Pressure approximation

We first consider the problem of reconstructing atmospheric pressure values in different zones on Earth. The dataset contains values between 966 hPa and 1061 hPa. The errors are displayed in Table 7 and a graphical representation of the actual and approximated values in the case of N=2000 nodes and K=200000 evaluation points is presented in Figure 4. The relative error in all the cases is 0.2%.

N 1000 2000
K 5⋅104 105 2⋅105 5⋅104 105 2⋅105
em⁢e⁢a⁢n ( hPa) 2.02 2.03 2.04 1.91 1.92 1.92
er⁢m⁢s ( hPa) 2.74 2.76 2.76 2.63 2.65 2.65
Table 7: Atmospheric pressure approximation errors for various choices of nodes (N) and evaluation points (K).
Refer to caption
(a) Real values.
Refer to caption
(b) Approximated values.
Figure 4: K=200000 atmospheric pressure values reconstructed using S⁢P2 and N=2000 nodes.

4.2 Temperature approximation

We next consider the problem of reconstructing temperature values on the globe. The minimum temperature in the dataset is −41.21 °C and the maximum temperature is 44.16 °C. The errors are displayed in Table 8, and a graphical representation of the actual and approximated values in the case of N=2000 nodes and K=200000 evaluation points is presented in Figure 5. The relative error in all the cases is 0.9%.

N 1000 2000
K 5⋅104 105 2⋅105 5⋅104 105 2⋅105
em⁢e⁢a⁢n ( °C) 2.5468 2.5562 2.5521 2.4761 2.4825 2.4765
er⁢m⁢s ( °C) 3.4817 3.5090 3.4902 3.3195 3.3378 3.3218
Table 8: Temperature approximation errors for various choices of nodes (N) and evaluation points (K).
Refer to caption
(a) Real values.
Refer to caption
(b) Approximated values.
Figure 5: K=200000 temperature values reconstructed using S⁢P2 and N=2000 nodes.

4.3 Wind speed approximation

We now investigate the problem of wind speed approximation. The values are given in  m/s, with a range between 0 m/s and 20.06 m/s. Table 9 displays the approximation errors and Figure 6 shows the graphical representation of the real and approximated values.

N 1000 2000
K 5⋅104 105 2⋅105 5⋅104 105 2⋅105
em⁢e⁢a⁢n ( m/s) 1.4564 1.4586 1.4612 1.4521 1.4548 1.4586
er⁢m⁢s ( m/s) 1.8773 1.8827 1.8844 1.8602 1.8649 1.8682
Table 9: Wind speed approximation errors for various choices of nodes (N) and evaluation points (K).
Refer to caption
(a) Real values.
Refer to caption
(b) Approximated values.
Figure 6: K=200000 wind speed values reconstructed using S⁢P2 and N=2000 nodes.

4.4 Rainfall approximation

The last practical application of the operator introduced in (11) that we study in this paper is the approximation of rainfall amounts (precipitation) in mm measured in an hour. We considered only the non-zero values, which left us with 24513 values between 0.1 mm/h and 47.32 mm/h. Table 10 presents the approximation errors for different choices of the number of nodes (N) and evaluation points (K).

N 1000 2000
K 104 1.5⋅104 2⋅104 104 1.5⋅104 2⋅104
em⁢e⁢a⁢n ( mm/h) 0.7624 0.7615 0.7687 0.7471 0.7510 0.7563
er⁢m⁢s ( mm/h) 1.5024 1.5073 1.5149 1.4648 1.5060 1.5038
Table 10: Precipitation quantities approximation errors for various choices of nodes (N) and evaluation points (K).

Finally, we study the behavior of the operators presented above in reconstructing the data associated with the phenomena discussed in this section. The operators investigated are the following: the Shepard-type operator S⁢P2 proposed in this paper and defined in (11), the classical Shepard operator Sμ given in (1), the modified variant S defined in (4), the Shepard-Bernoulli operator of order 1, SB1, defined in (19), and the variant combined with the thin-plate spline spherical radial basis function, ST⁢P⁢S, from (20). We consider K=104 evaluation points and use sets of N=1000 and N=2000 nodes. The corresponding results are presented in Table 11 and Table 12, respectively.

We can observe that, for wind speed and precipitation, the best approximation results are given by the simple operators, classical Sμ and modified S. For pressure, the Shepard-Bernoulli operator SB1 achieves the lowest errors among all the operators considered. For temperature, the results are mixed, but S⁢P2 achieves the smallest root mean square error in both tables. Overall, S⁢P2 and SB1 are comparable, with S⁢P2 performing better for wind speed and, in most cases, for precipitation, while SB1 performs better for pressure and, with respect to the mean absolute error, for temperature. However, S⁢P2 has an advantage in terms of computational time. The proposed operator outperforms ST⁢P⁢S in the majority of the cases considered. Besides requiring less computational time than the combined operators SB1 and ST⁢P⁢S in these experiments, the proposed operator has an advantage over higher-order Shepard-Bernoulli operators: its local polynomials, even when a higher polynomial degree is used, can be constructed from observed function values alone, without requiring partial derivatives, which are generally unavailable in meteorological and other observational datasets. Nevertheless, the numerical results in the preceding two sections show that the dataset influences the outcome, as is known to happen in this kind of problem.

S⁢P2 Sμ S SB1 ST⁢P⁢S
Pressure ( hPa) em⁢e⁢a⁢n 2.01 1.94 1.66 0.95 13.42
er⁢m⁢s 2.75 3.49 3.02 1.87 34.69
CPU(s) 1.96 0.27 0.30 19.55 37.24
Temperature ( °C) em⁢e⁢a⁢n 2.51 2.75 2.35 2.37 4.84
er⁢m⁢s 3.43 4.47 3.75 4.15 11.75
CPU(s) 2.31 0.27 0.31 19.19 37.68
Wind speed ( m/s) em⁢e⁢a⁢n 1.43 1.13 1.12 1.89 1.34
er⁢m⁢s 1.84 1.54 1.54 3.91 2.06
CPU(s) 1.89 0.27 0.28 19.81 39.01
Precipitation ( mm/h) em⁢e⁢a⁢n 0.76 0.48 0.48 0.89 1.19
er⁢m⁢s 1.49 1.04 1.04 2.84 3.79
CPU(s) 2.93 0.33 0.41 21.14 41.31
Table 11: Comparison of spherical Shepard-type operators on real-data applications, for N=1000 nodes and K=104 evaluation points.
S⁢P2 Sμ S SB1 ST⁢P⁢S
Pressure ( hPa) em⁢e⁢a⁢n 1.91 1.72 1.49 0.78 14.84
er⁢m⁢s 2.65 3.11 2.72 1.57 39.82
CPU(s) 7.10 0.55 1.21 62.03 105.81
Temperature ( °C) em⁢e⁢a⁢n 2.46 2.50 2.19 1.98 4.91
er⁢m⁢s 3.30 4.16 3.60 3.59 12.38
CPU(s) 8.73 0.65 1.02 74.28 243.15
Wind speed ( m/s) em⁢e⁢a⁢n 1.43 1.07 1.05 1.56 1.10
er⁢m⁢s 1.83 1.49 1.47 2.89 1.64
CPU(s) 3.83 0.66 0.53 44.39 85.33
Precipitation ( mm/h) em⁢e⁢a⁢n 0.74 0.43 0.43 0.70 0.66
er⁢m⁢s 1.45 0.94 0.94 2.22 1.56
CPU(s) 7.15 0.84 0.78 48.29 97.40
Table 12: Comparison of spherical Shepard-type operators on real-data applications, for N=2000 nodes and K=104 evaluation points.

References

  • [1] Allasia, G., Cavoretto, R., De Rossi, A.: Hermite-Birkhoff interpolation on scattered data on the sphere and other manifolds. Appl. Math. Comput. 318, 35–50 (2018)
  • [2] Bos, L., De Marchi, S., Sommariva, A., Vianello, M.: Computing multivariate Fekete and Leja points by numerical linear algebra. SIAM J. Numer. Anal. 48(5), 1984–1999 (2010)
  • [3] Bos, L., De Marchi, S., Sommariva, A., Vianello, M.: On multivariate Newton interpolation at discrete Leja points. Dolomites Res. Notes Approx. 4, 15–20 (2010)
  • [4] Calvi, J.P., Levenberg, N.: Uniform approximation by discrete least squares polynomials, J. Approx. Theory 152, 82–100 (2008)
  • [5] Cătinaş, T., Malina, A.: Spherical interpolation of scattered data using least squares thin-plate spline and inverse multiquadric functions. Numer. Algor. 97, 1397–1414 (2024)
  • [6] Cătinaş, T., Malina, A.: Spherical Shepard-Bernoulli operator. J. Appl. Math. Comput. 71, 1267–1283 (2025)
  • [7] Cavoretto, R., De Rossi, A.: Fast and accurate interpolation of large scattered data sets on the sphere. J. Comput. Appl. Math. 234, 1505–1521 (2010)
  • [8] Cavoretto, R., De Rossi, A.: Numerical comparison of different weights in Shepard’s interpolants on the sphere. Appl. Math. Sci. 4, 3425–3435 (2010)
  • [9] Cavoretto, R., De Rossi, A.: Spherical interpolation using the partition of unity method: An efficient and flexible algorithm. Appl. Math. Lett. 25, 1251–1256 (2012)
  • [10] Cavoretto, R., De Rossi, A.: Achieving accuracy and efficiency in spherical modelling of real data. Math. Methods Appl. Sci. 37, 1449–1459 (2014)
  • [11] Cheney, E.W., Light, W.A.: A Course in Approximation Theory, vol. 101. American Mathematical Soc. (2009)
  • [12] Davydov, O., Schumaker, L.L.: Scattered data fitting on surfaces using projected Powell-Sabin splines. In: Martin, R., Sabin, M., Winkler, J. (eds) Mathematics of Surfaces XII. Mathematics of Surfaces 2007. Lecture Notes in Computer Science, vol 4647. Springer, Berlin, Heidelberg (2007)
  • [13] De Rossi, A.: Spherical interpolation of large scattered data sets using zonal basis functions. In: Daehlen, M., Morken, K., Schumaker, L. (eds.) Mathematical Methods for Curves and Surfaces Trømso 2004, Nashboro Press, 125–134 (2005)
  • [14] Farwig, R.: Rate of convergence of Shepard’s global interpolation formula. Math. Comp. 46, 577–590 (1986)
  • [15] Franke, R.: Scattered data interpolation: tests of some methods. Math. Comp. 38, 181–200 (1982)
  • [16] Franke, R., Nielson, G.: Smooth interpolation of large sets of scattered data. Int. J. Numer. Meths. Engrg. 15, 1691–1704 (1980)
  • [17] Le Gia, Q., Sloan, I., Wendland, H.: Multiscale analysis in Sobolev spaces on the sphere. SIAM J. Numer. Anal. 48, 2065–2090 (2010)
  • [18] Little, F.: Convex combination surfaces. In: Barnhill, R.E., Boehm, W. (eds) Surfaces in Computer Aided Geometric Design, Amsterdam: North-Holland, 99–107 (1983)
  • [19] Nielson, G., Ramaraj, R.: Interpolation over a sphere based upon a minimum norm network. Comput. Aided Geom. Design 4, 41–57 (1987)
  • [20] Renka, R.J.: Interpolation of data on the surface of a sphere. ACM Trans. Math. Software 10, 417–436 (1984)
  • [21] Renka, R.J., Cline, A.K.: A triangle-based C1 interpolation method. Rocky Mountain J. Math. 14, 223–237 (1984)
  • [22] Renka, R.J.: Multivariate interpolation of large sets of scattered data. ACM Trans. Math. Software 14, 139–148 (1988)
  • [23] Shepard, D.: A two dimensional interpolation function for irregularly spaced data. In: Proceedings of the 1968 23rd ACM National Conference, 517–523. ACM (1968)
  • [24] Wendland, H.: Scattered Data Approximation. Cambridge University Press (2004)
  • [25] Wong, T.T., Luk, W.S., Heng, P.A.: Sampling with Hammersley and Halton Points. J. Graph. Tools, 2(2), 9–24 (1997)
  • [26] Zerroudi, B., Tayeq, H., El Harrak, A.: Effective interpolation of scattered data on a sphere through a Shepard-like method. Math. Model. Comput. 10(4), 1174–1186 (2023)
  • [27] Zerroudi, B., Tayeq, H., El Harrak, A.: Scattered data interpolation on the 2-dimensional surface through Shepard-like technique. Math. Model. Comput. 11(1), 277–289 (2024)
  • [28] Zuppa, C.: Error estimates for moving least square approximations. Bull. Braz. Math. Soc. 34(2), 231–249 (2003)

[1] Allasia, G., Cavoretto, R., De Rossi, A.: Hermite-Birkhoff interpolation on scattered data on the sphere and other manifolds. Appl. Math. Comput. 318, 35–50 (2018)

[2] Bos, L., De Marchi, S., Sommariva, A., Vianello, M.: Computing multivariate Fekete and Leja points by numerical linear algebra. SIAM J. Numer. Anal. 48(5), 1984–1999 (2010)

[3] Bos, L., De Marchi, S., Sommariva, A., Vianello, M.: On multivariate Newton interpolation at discrete Leja points. Dolomites Res. Notes Approx. 4, 15–20 (2010)

[4] Calvi, J.P., Levenberg, N.: Uniform approximation by discrete least squares polynomials, J. Approx. Theory 152, 82–100 (2008)

[5] Cătinaș, T., Malina, A.: Spherical interpolation of scattered data using least squares thin-plate spline and inverse multiquadric functions. Numer. Algor. 97, 1397–1414 (2024)

[6] Cătinaș, T., Malina, A.: Spherical Shepard-Bernoulli operator. J. Appl. Math. Comput. 71, 1267–1283 (2025)

[7] Cavoretto, R., De Rossi, A.: Fast and accurate interpolation of large scattered data sets on the sphere. J. Comput. Appl. Math. 234, 1505–1521 (2010)

[8] Cavoretto, R., De Rossi, A.: Numerical comparison of different weights in Shepard’s interpolants on the sphere. Appl. Math. Sci. 4, 3425–3435 (2010)

[9] Cavoretto, R., De Rossi, A.: Spherical interpolation using the partition of unity method: An efficient and flexible algorithm. Appl. Math. Lett. 25, 1251–1256 (2012)

[10] Cavoretto, R., De Rossi, A.: Achieving accuracy and efficiency in spherical modelling of real data. Math. Methods Appl. Sci. 37, 1449–1459 (2014)

[11] Cheney, E.W., Light, W.A.: A Course in Approximation Theory, vol. 101. American Mathematical Soc. (2009)

[12] Davydov, O., Schumaker, L.L.: Scattered data fitting on surfaces using projected Powell-Sabin splines. In: Martin, R., Sabin, M., Winkler, J. (eds) Mathematics of Surfaces XII. Mathematics of Surfaces 2007. Lecture Notes in Computer Science, vol 4647. Springer, Berlin, Heidelberg (2007)

[13] De Rossi, A.: Spherical interpolation of large scattered data sets using zonal basis functions. In: Daehlen, M., Morken, K., Schumaker, L. (eds.) Mathematical Methods for Curves and Surfaces Trømso 2004, Nashboro Press, 125–134 (2005)

[14] Farwig, R.: Rate of convergence of Shepard’s global interpolation formula. Math. Comp. 46, 577–590 (1986)

[15] Franke, R.: Scattered data interpolation: tests of some methods. Math. Comp. 38, 181–200 (1982)

[16] Franke, R., Nielson, G.: Smooth interpolation of large sets of scattered data. Int. J. Numer. Meths. Engrg. 15, 1691–1704 (1980)

[17] Le Gia, Q., Sloan, I., Wendland, H.: Multiscale analysis in Sobolev spaces on the sphere. SIAM J. Numer. Anal. 48, 2065–2090 (2010)

[18] Little, F.: Convex combination surfaces. In: Barnhill, R.E., Boehm, W. (eds) Surfaces in Computer Aided Geometric Design, Amsterdam: North-Holland, 99–107 (1983)

[19] Nielson, G., Ramaraj, R.: Interpolation over a sphere based upon a minimum norm network. Comput. Aided Geom. Design 4, 41–57 (1987)

[20] Renka, R.J.: Interpolation of data on the surface of a sphere. ACM Trans. Math. Software 10, 417–436 (1984)

[21] Renka, R.J., Cline, A.K.: A triangle-based C1 interpolation method. Rocky Mountain J. Math. 14, 223–237 (1984)

[22] Renka, R.J.: Multivariate interpolation of large sets of scattered data. ACM Trans. Math. Software 14, 139–148 (1988)

[23] Shepard, D.: A two dimensional interpolation function for irregularly spaced data. In: Proceedings of the 1968 23rd ACM National Conference, 517–523. ACM (1968)

[24] Wendland, H.: Scattered Data Approximation. Cambridge University Press (2004)

[25] Wong, T.T., Luk, W.S., Heng, P.A.: Sampling with Hammersley and Halton Points. J. Graph. Tools, 2(2), 9–24 (1997)

[26] Zerroudi, B., Tayeq, H., El Harrak, A.: Effective interpolation of scattered data on a sphere through a Shepard-like method. Math. Model. Comput. 10(4), 1174–1186 (2023)

[27] Zerroudi, B., Tayeq, H., El Harrak, A.: Scattered data interpolation on the 2-dimensional surface through Shepard-like technique. Math. Model. Comput. 11(1), 277–289 (2024)

[28] Zuppa, C.: Error estimates for moving least square approximations. Bull. Braz. Math. Soc. 34(2), 231–249 (2003)

Related Posts

No results found.