No description available.

In this thesis, we integrate the particular solutions of singularly perturbed partial differential equations into radial basis collocation method to solve two kinds of boundary layer problem.

Thesis (Ph. D.)--Texas A&M University, 2006. / "May 2006." "Major subject: Mathematics." Vita. Includes bibliographical references (p. 67-70). Also available online.

In approximation theory, three classical types of results are direct theorems, Bernstein inequalities, and inverse theorems. In this paper, we include results about radial basis function (RBF) approximation from all three classes. Bernstein inequalities are a recent development in the theory of RBF approximation, and on Rd, only L2 results are known for RBFs with algebraically decaying Fourier transforms (e.g. the Sobolev splines and thin-plate splines). We will therefore extend what is known by establishing Lp Bernstein inequalities for RBF networks on Rd. These inequalities involve bounding a Bessel-potential norm of an RBF network by its corresponding Lp norm in terms of the separation radius associated with the network. While Bernstein inequalities have a variety of applications in approximation theory, they are most commonly used to prove inverse theorems. Therefore, using the Lp Bernstein inequalities for RBF approximants, we will establish the corresponding inverse theorems. The direct theorems of this paper relate to approximation in Lp(Rd) by RBFs which are perturbations of Green's functions. Results of this type are known for certain compact domains, and results have recently been derived for approximation in Lp(Rd) by RBFs that are Green's functions. Therefore, we will prove that known results for approximation in Lp(Rd) hold for a larger class of RBFs. We will then show how this result can be used to derive rates for approximation by Wendland functions.

Abstract The representation of 3D Computer Graphics has been studied for a long time. Most 3D object models are obtained by 3D scan systems. These kinds of data are not only very huge, but also have a lot of redundancy. It consumes a large mount of time and resources. For this reason, how to represent the object efficiently is always an important issue. The purpose of this study is to present the objects by implicit functions. Different with the presentation of polygon mesh, implicit function is very compact to present objects because that the mathematical form of object can be obtained in different data forms. The implicit function used is Radial Basis Function, and then the BSP tree is used to partition the object to reduce the amount of computing. We also use the compression of progressive mesh to decrease the storage and the computing time. In addition, the object can be rendered according to the sampling points on each implicit surface.

Radial basis functions (RBFs) are probably best known for their applications to scattered data problems. Until the 1990s, RBF theory only involved functions that were scalar-valued. Matrix-valued RBFs were subsequently introduced by Narcowich and Ward in 1994, when they constructed divergence-free vector-valued functions that interpolate data at scattered points. In 2002, Lowitzsch gave the first error estimates for divergence-free interpolants. However, these estimates are only valid when the target function resides in the native space of the RBF. In this paper we develop Sobolev-type error estimates for cases where the target function is less smooth than functions in the native space. In the process of doing this, we give an alternate characterization of the native space, derive improved stability estimates for the interpolation matrix, and give divergence-free interpolation and approximation results for band-limited functions. Furthermore, we introduce a new class of matrix-valued RBFs that can be used to produce curl-free interpolants.

No description available.

RBFs are useful in scientific computing. In this thesis, we are interested in the positions of collocation points and RBF centers which causes the matrix for RBF interpolation singular and ill-conditioned. We explore the best bases by minimizing error function in supremum norm and root mean squares. We also use radial basis function to interpolate shifted data and find the best basis in certain sense. In the second part, we solve ill-posed problems by radial basis collocation method with different radial basis functions and various number of bases. If the solution is not unique, then the numerical solutions are different for different bases. To construct all the solutions, we can choose one approximation solution and add the linear combinations of the difference functions for various bases. If the solution does not exist, we show the numerical solution always fail to satisfy the origin equation.

Thesis (Ph.D.)--Tufts University, 2004. / Adviser: Fulton B. Gonzalez. Submitted to the Dept. of Mathematics. Includes bibliographical references (leaves 72-73). Access restricted to members of the Tufts University community. Also available via the World Wide Web;

In this thesis, we propose a new class of learning models, namely the additive generalized radial basis function networks (AGRBFNs), for general nonlinear regression problems. This class of learning machines combines the generalized radial basis function networks (GRBFNs) commonly used in general machine learning problems and the additive models (AMs) frequently encountered in semiparametric regression problems. In statistical regression theory, AM is a good compromise between the linear model and the nonparametric model. In order for more general network structure hoping to address more general data sets, the AMs are embedded in the output layer of the GRBFNs to form the AGRBFNs. Simple weights updating rules based on incremental gradient descent will be derived. Several illustrative examples are provided to compare the performances for the classical GRBFNs and the proposed AGRBFNs. Simulation results show that upon proper selection of the hidden nodes and the bandwidth of the kernel smoother used in additive output layer, AGRBFNs can give better fits than the classical GRBFNs. Furthermore, for the given learning problem, AGRBFNs usually need fewer hidden nodes than those of GRBFNs for the same level of accuracy.