Interpolation By Radial Basis Function (RBF) By: Reihane Khajepiri , Narges Gorji Supervisor: Dr.Rabiei 1 1. Introduction • Our problem is to interpolate the following tabular function: Where the nodes ๐ฅ๐ ∈ ๐and ๐ข(๐ฅ๐ ) ∈ โ. • The interpolation function has the form ๐ ๐ข(๐ฅ) โ ๐ ๐ฅ = ๐ผ๐ ๐๐ (๐ฅ) , ๐=1 such that ๐๐ (๐ฅ)’s are the basis of a prescribed ndimensional vector space of functions on ๐, i.e., ๐ค =< ๐1 , ๐2 , … , ๐๐ >. 2 • This is a system of ๐ linear equations in ๐ unknowns. It can be written in matrix form as ๐ด๐ถ = ๐, or in details as ๐1 (๐ฅ1 ) ๐2 (๐ฅ1 ) ๐๐ (๐ฅ1 ) ๐ผ1 ๐ข(๐ฅ1 ) โฏ ๐1 (๐ฅ2 ) ๐2 (๐ฅ2 ) ๐๐ (๐ฅ2 ) ๐ผ2 ๐ข(๐ฅ2 ) = โฎ โฎ โฑ โฎ โฎ ๐1 (๐ฅ๐ ) ๐2 (๐ฅ๐ ) โฏ ๐๐ (๐ฅ๐ ) ๐ผ๐ ๐ข(๐ฅ๐ ) • The ๐ × ๐ matrix ๐ด appearing here is called the interpolation matrix. • In order that our problem be solvable for any choice of arbitrary ๐ข(๐ฅ๐ ), it is necessary and sufficient that the interpolation matrix be nonsingular. • The ideal situation is that this matrix be nonsingular for all choices of ๐ distinct nodes ๐ฅ๐ . 3 • Classical methods for numerical solution of PDEs such as finite difference, finite elements, finite volume, pseudo-spectral methods are base on polynomial interpolation. • Local polynomial based methods (finite difference, finite elements and finite volume) are limited by their algebraic convergence rate. • MQ collection methods in comparison to finite element method have superior accuracy. 4 • Global polynomial methods such as spectral methods have exponential convergence rate but are limited by being tied to a fixed grid. Standard" multivariate approximation methods (splines or finite elements) require an underlying mesh (e.g., a triangulation) for the definition of basis functions or elements. This is usually rather difficult to accomplish in space dimensions > 2. • RBF method are not tied to a grid but to a category of methods called meshless methods. The global non polynomial RBF methods are successfully applied to achieve exponential accuracy where traditional methods either have difficulties or fail. • RBF methods are generalization of Multi Quadric RBF , MQ RBF have a rich theoretical development. 5 2. Literature Review • RBF developed by Iowa State university, Rolland Hardy in 1968 for scattered data be easily used in computations which polynomial interpolation has failed in some cases. RBF present a topological surface as well as other three dimensional shapes. • In 1979 at Naval post graduate school Richard Franke compared different methods to solve scattered data interpolation problem and he applied Hardy's MQ method and shows it is the best approximation & also the matrix is invertible. 6 • In 1986 Charles Micchelli a mathematician with IBM proved the system matrix for MQ is invertible and the theoretical basis began to develop. His approach is based on conditionally positive definite functions. • In 1990 the first use of MQ to solve PDE was presented by Edward Kansa. • In 1992 spectral convergence rate of MQ interpolation presented by Nelson &Madych. 7 • Originally, the motivation for two of the most common basis meshfree approximation methods (radial basis functions and moving least squares methods) came from applications in geodesy, geophysics, mapping, or meteorology. • Later, applications were found in many areas such as in the numerical solution of PDEs, artificial intelligence, learning theory, neural networks, signal processing, sampling theory, statistics ,finance, and optimization. 8 Remaking Images By RBF 9 • It should be pointed out that meshfree local regression methods have been used independently in statistics for more than 100 years. • Radial Basis Function "RBF" interpolate a multi dimensional scattered data which easily generalized to several space dimension & provide spectral accuracy. So it is so popular 10 3. RBF Method 3.1 Radial Function • In many applications it is desirable to have invariance not only under translation, but also under rotation & reflection. This leads to positive definite functions which are also radial. Radial functions are invariant under all Euclidean transformations (translations, reflections & rotations) • Def: A function ๐ท: โ ๐ → โ is called radial provided there exist a univariate function ๐: [0, ∞) → โ such that ๐ท(๐ฅ) = ๐ (๐) where r = ๐ฅ And . is some norm on โ ๐ , usually the Euclidean norm. 11 • For a radial function ๐ท : ๐ฅ1 = ๐ฅ2 ๏ ๐ท(๐ฅ1 ) = ๐ท(๐ฅ2 ) ๐ฅ1 , ๐ฅ2 ∈ โ๐ . By radial functions the interpolation problem becomes insensitive to the dimension s of the space in which the data sites lie. Instead of having to deal with a multivariate function ๐ท (whose complexity will increase with increasing space dimension s) we can work with the same univariate function ๐for all choices of s. 12 3.2 Positive Definite Matrices & Functions • Def : A real symmetric matrix A is called positive semidefinite if its associated quadratic form is non negative For c= ๐ ๐ ๐=1 ๐=1 ๐๐ ๐๐ ๐ด๐๐ [๐1 , … , ๐๐ ]๐ ∈ ๐ ๐ ≥0 (1) • If the only vector c that turns (1) into an equality is the zero vector , then A is called positive definite. • An important property of positive definite matrices is that all their eigenvalues are positive, and therefore a positive definite matrix is non-singular (but certainly not vice versa). 13 • Def: A real-valued continuous function ๐ท is positive definite on ๐ ๐ if & only if ๐ ๐=1 ๐ ๐=1 ๐๐ ๐๐ ๐ท(๐ฅ๐ − ๐ฅ๐ ) ≥ 0 (2) For any N pairwise different points ๐ฅ1 , ๐ฅ2 , … , ๐ฅ๐ ∈ ๐ ๐ & c = [๐1 , … , ๐๐ ]๐ • The function ๐ท is strictly positive definite on ๐ ๐ if the only vector c that turns (2) into an equality is the zero vector. 14 3.3 Interpolation of scattered data • One dimensional data polynomial &fourier interpolation has the form: S(x) = ๐๐ ๐๐ (๐ฅ) ๐๐ : basis function {๐ฅ๐ } : are distinct ๐(๐ฅ๐ ) = ๐๐ → ๐๐๐ก๐๐๐๐๐๐ ๐๐ • For any set of basis function ๐๐ (๐ฅ) independent of data points & sets of distinct data points {๐ฅ๐ } such that the linear system of equation for determining the expansion coefficient become non-singular {Haar theorem} • Def: An n-dimensional vector space Γ of functions on a domain ๐ is said to be a Haar space if the only element of Γwhich has more than n-1 roots in ๐ is the zero element. 15 • Theorem1. Let Γ have the basis {๐1 , ๐2 , … , ๐๐ }. These properties are equivalent: a) Γ is a Haar space. b) det(๐๐ (๐ฅ๐ )) ≠ 0 for any set of distinct points ๐ฅ1 , ๐ฅ2 , … , ๐ฅ๐ ∈ ๐. Any basis for a Haar space is called Chebyshev system. • A Haar space is a space of functions that guarantees invertibility of the interpolation matrix. • Some examples of Chebyshev systems on โ: 1)1, ๐ฅ, ๐ฅ 2 , … , ๐ฅ ๐ 2)๐ ๐1 ๐ฅ , ๐ ๐2 ๐ฅ , … , ๐ ๐๐ ๐ฅ , (๐1 < ๐2 < โฏ < ๐๐ ) 3)1, ๐๐๐ โ๐ฅ, ๐ ๐๐โ๐ฅ, … , ๐โ๐ โ๐๐ฅ, ๐ ๐๐โ๐๐ฅ 4)(๐ฅ + ๐1 )−1 , (๐ฅ + ๐2 )−1 , … ,(๐ฅ + ๐๐ )−1 , 0 ≤ ๐1 ≤ โฏ ≤ ๐๐ 16 3-4 development of RBF method • Hardy present RBF which is linear combination of translate of a single basis function that is radially symmetric a bout its center. ๐(๐ฅ) = ๐๐ |๐ฅ − ๐ฅ๐ | S (๐ฅ๐ ) = ๐๐ ๏ ๐๐ is determined • Problems in Continuously differentiability of S ๐ฅ result to the following form: S ๐ฅ = ๐๐ ๐2 + ๐ฅ − ๐ฅ๐ 2 c≠0 Which is a linear combination of translate of the hyperbola basis function. 17 Accurate representation of a topographic profile in more than one dimensional space S ๐ฅ, ๐ฆ = ๐๐ ๐ฅ − ๐ฅ๐ 2 + ๐ฆ − ๐ฆ๐ 2 S ๐ฅ๐ , ๐ฆ๐ = ๐๐ S ๐ฅ, ๐ฆ = ๐๐ S ๐ฅ, ๐ฆ = ๐2 + ๐ฅ − ๐ฅ๐ 2 + ๐ฆ − ๐ฆ๐ 2 ๐๐ ๐ 2 + ||๐ฑ − ๐ฑ๐ ||2 S ๐ฅ๐ , ๐ฆ๐ = ๐๐ ๏ ๐ด๐ = ๐ 18 3.5 RBF categories: • We have two kinds of RBF methods 1. Basis RBF 2. Augmented RBF • RBF basic methods: n distinct data points {๐ฅ๐ } & corresponding data values {๐๐ }, • The basic RBF interpolant is given by ๐ ๐ฅ = ๐๐๐ ๐๐ ๐ ๐ฅ − ๐ฅ๐ ๐ด๐ = ๐ = ๐( ๐ฅ๐ − ๐ฅ๐ ) • Micchelli gave sufficient conditions for φ(r) to guarantee that the matrix A is unconditionally nonsingular. ๏ RBF method is uniquely solvable. 19 Def: A function ๐ โถ 0, ∞ → โ is completely monotone on[0,∞] if : 1. ๐๐๐ถ[0, ∞) 2. ๐๐๐ถ ∞ (0, ∞) 3. −1 ๐ ๐๐ (๐) ≥ 0; r>0 ; 20 Types of basis function: • Infinitely smooth RBF • Piecewise smooth RBF 21