A direct prediction of the shape parameter—A purely scattered data approach☆
Introduction
In this paper the approximated functions lie in a space called Bσ, as in the following definition. Definition 1.1 For any σ > 0, the class of band-limited functions f in L2(Rn), the family of square-integrable functions, iswhere denotes the Fourier transform of f.
This function space looks small. In fact, it plays only an intermediate role in the interpolation. Via the Bσ functions, all functions in the Sobolev space can be interpolated by the multiquadrics or inverse multiquadrics, as will be explained further in the paper.
The radial function we adopt iswhere ‖x‖ is the Euclidean norm of x ∈ Rn, Γ is the classical gamma function, and β, c are constants. Note that this definition is slightly different from the one mentioned in the abstract. The definition (1) will simplify its Fourier transform and the presentation of our central theorem. The function h(x) in (1) is conditionally positive definite (c.p.d.) of order where means the smallest integer greater than or equal to . For further details we refer the reader to Madych and Nelson [1] and Wendland [2].
For any set of data points where is a subset of Rn and yj are real or complex numbers, we can always find an interpolant of the formwhere p(x) is a polynomial in to be determined, and cj are coefficients to be chosen, as long as X is a determining set for . Thus for . The requirement that X determine means that two polynomials in this polynomial space are exactly the same as long as they are equal at . Interested readers can find these in Madych and Nelson [1].
Although we are interested only in scattered data, our criteria of choosing c are developed from a core theorem which involves a simplex scheme with evenly spaced data points, as mentioned in the abstract. Therefore it is necessary to make a brief description of the evenly spaced scheme.
Let Tn denote an n-simplex in Rn. Then T1 is just a line segment, T2 is a triangle, and T3 is a tetrahedron with four vertices. The exact definition can be found in Fleming [3].
Let be the vertices of Tn. Then any point x ∈ Tn can be written as a convex combination of the vertices:The numbers are called the barycentric coordinates of x.
For any n-simplex Tn, the evenly spaced points of degree l are those points whose barycentric coordinates are of the formObviously, the number of evenly spaced points of degree l is . It’s proven in Bos [4] that such points do form a determining set for .
Before entering our core theorem, some ingredients must be explained. Each function of the form (1) induces a function space called native space. Also, there is a seminorm ‖f‖h for each . These can be found in Luh [5], [6], [7], Madych and Nelson [1], [8] and Wendland [2]. The constants ρ and Δ0, which are usually very small positive numbers for low dimensions, in the theorem are determined by n and β. We omit their complicated definitions and refer the reader to Luh [9].
The following theorem is just our core theorem. We omit its complicated proof and take it directly from [9]. Theorem 1.2 Let h be as in (1). For any positive number b0, let and . For any n-simplex Tn of diameter r satisfying (note that ), if holds for all x ∈ Tn and 0 < δ ≤ δ0, where s(x) is defined as in (2) with the evenly spaced points of degree l in Tn satisfying . The constant αn denotes the volume of the unit ball in Rn, and 0 < λ′ < 1 is given bywhich only in some cases mildly depends on the dimension n.
Remark. Note that as the degree l of the evenly spaced data points increases, the number δ will decrease, making the upper bound in (3) small. Hence δ can be regarded in spirit as the well-known fill distance. It is natural to ask what will happen if one regards δ completely the same as the fill distance. If so, the requirement that the centers be evenly spaced in the simplex can be dropped, making this theorem much more useful. In fact, this is just the central idea of this paper.
Section snippets
Criteria of choosing c
The number b0 in Theorem 1.2 controls the diameter of the domain. The upper bound in (3) is greatly related to the choice of c. In Luh [9] (3) is successfully transformed into a pleasant and lucid form which shows the influence of c explicitly. There are three cases: (i)β > 0 and n ≥ 1, (ii) β < 0 and or and (iii) and .
For (i) and (ii), we havewhere d0 is a small (for low dimensions) constant independent of c, σ, and f, and MN(c) is a
Experiments
We provide three sets of experiments here. Although we concern ourselves mainly with the scattered data setting, as a comparison, the evenly spaced data setting is also tested for one-dimensional experiments. As for the two-dimensional experiment, a purely scattered data setting is adopted. Moreover, we totally get rid of the parameter δ defined in Theorem 1.2 and choose c according to the MN curves, considering the moving direction of the bottoms of the curves as the number of data points
The failures of the MN curve approach
As is well known, Newton’s method of root-finding may fail whenever there are horizontal or nearly horizontal tangent lines. Similarly, our approach may also fail whenever there are nearly horizontal zones on the MN curves. Let us see a few MN curves first. If we further decrease the parameter δ in Fig. 1, Fig. 2, Fig. 3, Fig. 4, Fig. 5 of Section 3, nearly horizontal zones will appear, near the bottoms of the MN curves, as shown in Fig. 9, Fig. 10, Fig. 11. Also, note that, for the same δ, the
Summary
We are satisfied with the performance of the MN curve approach to finding the optimal value of the shape parameter, both in the evenly spaced and purely scattered data settings. Although this approach was presented by the author, the foundation built by W.R. Madych and S.A. Nelson plays an important role. Based on this foundation, the author eventually presented a practically useful theory. Hence we name the crucial function MN function, in honor of their outstanding contributions. It is
References (13)
Bounding the Lebesgue function for lagrange interpolation in a simplex
J Approx Theory
(1983)The mystery of the shape parameter III
Appl Comput Harmon Anal
(2016)- et al.
Bounds on multivariate polynomials and exponential error estimates for multiquadric interpolation
J Approx Theory
(1992) - et al.
Multivariate interpolation and conditionally positive definite function
Approx Theory Appl
(1988) Scattered data approximation
(2005)Functions of several variables
(1977)
Cited by (0)
- ☆
This research is supported by the Taiwanese Ministry of Science and Technology with project number 108-2115-M-126-004.