Generalized Bochner’s Theorem for Radial Function

Wu Zongmin

DOI: https://doi.org/10.1007/bf02837010

1997-01-01

Approximation Theory and Its Applications

Abstract:A radial function Φ(x) can be expressed by its generator ϕ(·) through Φ(x)=ϕ(‖x‖). The positive de finite of the function Φ plays an important role in the radial basis interpolation. We can naturally use Bochner’s Theorem to check if Φ is positive de finite. This requires however a n-dimensional Fourier transformation and it is not very easy to calculate. Furthermore in a lot of cases we will use ϕ for spaces of various dimensions too, then for every fixed n we need do the Fourier transformation once to check if the function is positive definite in the n-dimensional space. The completely monotone function, which is discussed in [4], is positive definite for arbitrary space dimensions. With this technique we can very easily characterize the positive definite of a radial function through its generator. Unfortunately there is only a very small subset of radial function which is completely monotone. Thus this criterion excluded a lot of interesting functions such as compactly supported radial function, which are very use ful in application. Can we find some conditions (as the completely monotone function) only for the 1-dimensional Fourier transform of the generator ϕ to characterize a radial function Φ, which is positive definite in n-dimensional (fixed n) space? In this paper we defined a kind of incompletely monotone function of order α, for α=0, 1/2, 1, 3/2, 2, … (we denote the function class by ICM), in this sence a normal positvie function is in ICM0; a positive monotone decreasing function is in ICM1 and a positive monotone decreasing and convex function is in ICM2. Based on this definition we get a generalized Bochner’s Theorem for radial function: If1-dimensional Fourier transform of the generator of a radial function can be written as\(F{}_1\varphi (t) = \tilde F(\frac{{t^2 }}{2})\), then corresponding radial function Φ(x) is positive definite as a n-variate function iff\(\tilde F\) is an incompletely monotone function of order α=(n-1)/2 (or simply\(\tilde F \in ICM_{\frac{{n - 1}}{2}} \).