Fast summation of radial functions on the sphere (Q2505233)

Let a function \( K \in L^1 ([-1, 1] ) \) be given and for fixed \( \eta \) on the sphere \( {\mathcal S}^2 := \{{\mathbf x} \in \mathbb R^3 : \| {\mathbf x }\|_2 = 1 \}\subset \mathbb R^3 \) consider the \(\eta \)-zonal function \[ K({\pmb \eta}\;\cdot ): {\mathcal S}^2 \rightarrow \mathbb R , \quad {\pmb \xi} \mapsto K ({\pmb \eta} \cdot {\pmb \xi}). \] For example, a radial function \( \phi \) corresponds to a zonal one by \( K({\pmb \eta } \, \cdot ) = \phi (\| {\pmb \eta } - \cdot \|_2). \) The paper is concerned with the following problem. Given \( D, L \in \mathbb N \), a set of arbitrary source nodes \( Y := \{{\pmb \eta}_e \in {\mathcal S}^2: l = 0, \ldots , L-1\}\), and a vector \( {\mathbf b} := (b_l)^{L-1}_{l=0} \in \mathbb R^L \) evaluate the sum \[ f({\pmb \xi}) := \sum^{L-1}_{l=0} b_l K (\eta_l \cdot \xi) \] on a set of arbitrary target nodes \( X := \{\xi_d \in {\mathcal S}^2 : d = 0, \dots , D-1 \}. \) A simple structured approximate algorithm with arithmetic complexity \( {\mathcal O} (D + L) \) is presented to solve this problem. It is based on the nonequispaced fast spherical Fourier transform introduced by \textit{S. Kunis} and \textit{D. Potts} [Comput. Appl. Math. 161, 75--98 (2003; Zbl 1033.65123)] and can be easily adapted to different kernels \(K\). For the construction of the fast algorithm, a truncated series expansion of \( K ({\pmb \eta} \cdot {\pmb \xi })\) in spherical harmonics is used yielding a separation of source and target nodes. Error estimates are obtained and numerical examples are presented.