Schaback, R. (1993): Comparison of radial basis function interpolants. In: K. Jetter and F. Utreras, eds., From CAGD to Wavelets, World Scientific, pp. 293-- 305.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....X i=1 i s(y i ) Gamma1) m 0 1 N(s) X i;j=1 i j OE(ky i Gamma y j k) 1.12) is strictly positive, if satisfies (1.5) for n = N(s) and any m m 0 . Thus (1.11) is indeed a semi inner product. Then (hs; si) 1=2 is a semi norm on the space A OE;m with null space Pi m . Schaback [8] shows that the solution of (1.6) can be characterized as follows. Theorem 1 Let OE be any radial basis function from (1.2) and let m be chosen such that m m 0 . Given are points x 1 ; x n in IR d that satisfy (1.9) and values f 1 ; f n in IR. Let s be the radial function of ....

R. Schaback. Comparison of radial basis function interpolants. In K. Jetter and F. Utreras, editors, Multivariate Approximations: From CAGD to Wavelets, pages 293--305. World Scientific, Singapore, 1993.

.... from a larger space F to F ( Omega Gamma3 one can ask for more natural topologies on F ( Omega Gamma that are directly related to the representation (1) Such a topology actually exists for Omega = IR d , dating back to early papers of Madych and Nelson [4,5,6] and conveniently described in [8]. In short, it uses (3) to define js f j 2 : N X j;k=1 c j c k (x j Gamma x k ) 4) as a seminorm on functions of the form (1) Such a seminorm arises straightforwardly from a bilinear form, and one can take a Hilbert space closure to generate an appropriate function space ....

.... j 2 : N X j;k=1 c j c k (x j Gamma x k ) 4) as a seminorm on functions of the form (1) Such a seminorm arises straightforwardly from a bilinear form, and one can take a Hilbert space closure to generate an appropriate function space F ( Omega Gamma2 This has been done by various authors [1,2,4,5,6,8,9] in different ways, using three basic techniques: 1. variational inequalities, 2. generalized convolutions, 3. generalized Fourier transforms. This paper serves to give a unified theory and to show that these approaches are equivalent. Up to now, the dependence of the spaces F ( Omega Gamma ....

Schaback, R., Comparison of Radial Basis Function Interpolants, in Multivariate Approximation: From CAGD to Wavelets, K. Jetter and F. I. Utreras, (eds.), World Scientific, Singapore, 1993, 293--305.

....preserves polynomials of degree k Gamma 1. Since his interpolants are all special cases of the ones just described, the polynomial preservation property of U , that is, Up = p for all p 2 Pi k Gamma1 , is clear. Another approach, taken by Madych and Nelson [7] Powell [10] and Schaback [12] uses a pointwise error estimator. This estimator involves an expression which Schaback calls the power function. Both Powell and Schaback compute this power function in some sense. Their estimates coincide with those of Duchon, although one should note that Schaback is interested in a much wider ....

Schaback, R. A comparison of radial basis function interpolants, in Multivariate approximation -- from CAGD to wavelets, eds., Jetter, K. and F. Utreras, World Scientific (London), 1993, 293--305.

....splines are a member of the class. Our approach is via the work of Golomb and Weinberger [2] together with some interpretations of that work by Meinguet [8] These are fine papers which are still well worth reading. In the course of our work, we will establish results of Powell [10] and Schaback [12]. We begin with a normed linear space X . Let fl 1 ; fl m be linear information functionals on X . The intention is that for a given f 2 X , the information fl i (f) ff i , i = 1; m is known. From this information it 9 is desired to compute a value fl(f) where fl is another ....

....p i 0 3 X j=1 fi j a j 1 A = 3 X j=1 fi j p i (a j ) fi i ; and so, fl(w) 2 Gamma2 3 X i=1 fi i OE(kx Gamma a i k 2 ) 3 X i;j=1 fi i fi j OE(ka i Gamma a j k) 2 = Phi(fi) One can also find a detailed account of results similar to Theorem 3. 1 in a paper by Schaback [12]. In this paper the level of generality is much greater than that of Powell. Schaback considers many more examples of radial functions rather than just looking at thin plate splines in IR 2 . However, the error formula given is similar to that of Theorem 3.1. Our final contribution in this ....

R. Schaback (1994), "Comparison of Radial Basis Function Interpolants", (preprint), Universitat Gottingen.

....F ) where vL is the representer of the linear functional L in the native space and the distance is taken in the norm of this space. In this form the hypercircle inequality was given in [6] The term jLyj, for special choice of the functional L, is sometimes called the power function (Schaback [18]) Our aim is to give estimates for jL(y)j in the hypercircle inequality. This is done by eliminating a number of Fourier coefficients and finding bounds on the tails by means of appropriate assumptions on the kernel . If we employ the expansion (6) of y, then we obtain by orthogonality jLyj = ....

Schaback, R., Comparison of radial basis function interpolants, in: Multivariate Approximation: From Theory to Software, (K. Jetter and F. I. Utreras, eds.), 293--305, World Scientific, Singapore 1995.

....2 (IR d ) ae f fi fi fi fi Z IR d j f ( j 2 (1 k k 2 ) k d 1 oe : However, if decays exponentially, e.g. for Phi(x) e Gammaffkxk 2 , then F consists of C 1 functions. Further details can be found in W. Madych and S. Nelson [9] 10] N. Dyn [6] and R. Schaback [15]. Numerical observations and theoretical results have revealed that the error and the sensitivity, described by P (x) and Gamma1 , seem to be intimately related. In particular, there is no case known where the error and the sensitivity are both reasonably small. There is a dichotomy: Either ....

....explicitly written as P 2 (x) Phi(0) Gamma 2 N X j=1 u j (x) Phi(x j Gamma x) N X j=1 N X k=1 u j (x)u k (x) Phi(x j Gamma x k ) where u 1 (x) uN (x) are the Lagrange basis functions for interpolation, i.e. 1. 6) equals s f = N X j=1 f(x j )u j (see e.g. Schaback [15]) We now set x 0 : x and add a first row and column to the N Theta N matrix A in (1.2) to get a (N 1) Theta (N 1) matrix A x . With u x : 1; Gammau 1 (x) Gammau N (x) T 2 IR N 1 we then have P 2 (x) u T x A x u x and simply use (1.3) for A x in the form (x)kflk ....

R. Schaback. Comparison of radial basis function interpolants. In Multivariate Approximation. From CAGD to Wavelets, pages 293--305. K. Jetter and F. Utreras, editors; World Scientific, London, 1993.

.... errors and condition numbers can be found in [31] After some practical experience we found the following guidelines to be useful: a) First fix a function OE by consideration of smoothness requirements for s in (2) This is a crucial step, because from a theoretical point of view (see [8, 9, 20, 28]) each OE is an optimal choice for a specific space of smooth objects. b) If N is small, say N 200 for a standard 1994 workstation, then don t care about compact supports. Depending on smoothness requirements, try thin plate splines or multiquadrics, but for the latter make sure that the ....

Schaback, R., Comparison of Radial Basis Function Interpolants, Multivariate Approximations: From CAGD to Wavelets, K. Jetter and F. Utreras, (eds), World Scientific, London, 1993, 293--305.

....there is a useful connection between this bound and more familiar bounds involving the notion of power function . If ffi q is the Dirac functional at q 2 M, the power function dist Phi (ffi q ; U) for q 2 M arises in many publications and there are various papers proving upper bounds for it; see [15, 16]. If we take s n=2 in our setting, then ffi q is in H Gammas (M) H Phi . Consequently, we may set w = ffi q in (3.20) Using this in connection with (3.20) then gives us the standard pointwise error bound j(f v Gamma f v ) q)j dist Phi (ffi q ; U)kf v Gamma f v k Phi (3.21) for ....

R. Schaback. Comparison of radial basis function interpolants. In Multivariate Approximation. From CAGD to Wavelets, pages 293--305. K. Jetter and F. Utreras, editors; World Scientific, London, 1993.

....the approximants are formed by (1.1) with another, possibly different function Phi 1 . The corresponding seminorms will be j j 0 and j j 1 . Error bounds are known so far only for interpolants with Phi 0 = Phi 1 , and for Phi 0 6= Phi 1 there are some interesting numerical observations (see [8]) ffl For F 0 F 1 the Phi 1 interpolants seem to have more or less the same error on the larger space F 0 as the optimal Phi 0 interpolants (quasi optimality) ffl For F 0 F 1 the Phi 1 interpolants seem to behave better on F 0 than on F 1 (supercon vergence) The results of ....

....function c 0 (M) decreases to zero while C 01 (M) does not decrease. Thus Lemma 4.3 There is a positive constant c depending only on d; 0 , and 1 , such that for all 0 c we have an M( with C 01 (M( Delta c 0 (M( 4:4) and M( 1 for 0. 2 From the literature (see e.g. [8]) we cite Lemma 4.4 Given Phi 1 with 1 and m 1 , there is an error bound of the form jf(x) Gamma s f;X (x)j jf j 1 Delta P 1;X (x) for all functions in the native space F 1 and interpolants s f;X to f by functions (1.1) on sets X = fx 1 ; xN g ae IR d with (1.4) The power ....

R. Schaback. Comparison of radial basis function interpolants. In Multivariate Approximation. From CAGD to Wavelets, pages 293--305. K. Jetter and F. Utreras, editors; World Scientific, London, 1993.

....add the space IP d m of d variate polynomials of order not exceeding m to the interpolating functions. Interpolation is uniquely possible under the requirement If p 2 P d m satisfies p(x i ) 0 for all x i 2 X then p = 0; 1) and if Phi is conditionally positive definite of order m (see e.g. [8]) Definition 1. A function Phi : IR d IR with Phi( Gammax) Phi(x) is conditionally positive definite of order m on IR d , if for all sets X = fx 1 ; xNX g ae IR d with NX distinct points and all vectors ff : ff 1 ; ff NX ) 2 IR NX with NX X j=1 ff j p(x j ) ....

....interpolation in the scattered data case. x2. Spaces for Radial Basis Functions Each conditionally positive definite function Phi allows two constructions of an inner product space of functions. Both constructions are based on ideas of Madych and Nelson [4, 5] but we adopt the terminology of [8] here and omit details. The algebraic approach introduces a space F Phi by direct reference to the conditional positive definiteness of Phi of order m. Functions of the form f ff (x) N X j=1 ff j Phi(x Gamma x j ) 3) with (2) are in F Phi and have a norm kf ff k 2 Phi = P N j;k=1 ....

[Article contains additional citation context not shown here]

Schaback, R., Comparison of radial basis function interpolants, in Multivariate Approximations: From CAGD to Wavelets, K. Jetter and F. Utereas (eds.), World Scientific Publishing Co., 1993

No context found.

Schaback, R. (1993): Comparison of radial basis function interpolants. In: K. Jetter and F. Utreras, eds., From CAGD to Wavelets, World Scientific, pp. 293-- 305.

No context found.

R. Schaback. Comparison of radial basis function interpolants. In K. Jetter and F. Utreras, editors, Multivariate Approximations: From CAGD to Wavelets, pages 293--305. World Scientific, Singapore, 1993.

No context found.

Schaback, R. (1993): Comparison of radial basis function interpolants. In: K. Jetter and F. Utreras eds., From CAGD to Wavelets, pp. 293--305, World Scientific.

No context found.

Schaback, R. (1993) "Comparison of radial basis function interpolants", in Multivariate Approximation: From CAGD to Wavelets, K. Jetter and F.I. Utreras, eds., World

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC