Highly effective stable evaluation of bandlimited functions on the sphere (Q261861)

This paper is on the fast and accurate evaluation of a spherical polynomial given by its coefficients in the standard basis. The paper first investigates the trigonometric needlet operator \[ \Phi_N f(x):=\frac{1}{M}\sum_{\xi\in\mathcal{X}} \mathcal{K}_N(x-\xi)f(\xi), \] where \(\mathcal{X}=\mathcal{X}_M:=\{\xi_k=2\pi k/M : k=0,\dots,M-1\}\) and \(\mathcal{K}_N(x):=1+2\sum_{n=1}^\infty \varphi(n/N)\cos(nx)\) for some smooth cut-off function \(\varphi\) supported on \([0,1+\tau]\) with \(\varphi(x)\equiv 1\) for \(x\in[0,1]\). By the localization property of \(\mathcal{K}_N\), \(\Phi_N f(x)\) can be well-approximated by its truncated version \[ \Phi_{N,\delta} f(x) = \frac{1}{M}\sum_{\xi\in\mathcal{X}, \rho(x,\xi)\leq \delta} \mathcal{K}_N(x-\xi)f(\xi). \] The paper then turns to investigate the best possible \(\delta\) for which the error \(\|f-\Phi_{N,\delta}f\|_\infty\leq \varepsilon \|f\|_\infty\) is achieved for a fixed input precision \(\varepsilon\). Once \(\delta\) is fixed, the kernel \(\mathcal{K}_N(x)\) can then be approximated for \(x\in[0,\delta]\) and a fast evaluation of \(\Phi_{N,\delta}f(x)\) with \(f(\xi), \xi\in\mathcal{X}_M\) given for \(x\in\mathcal{Z}\) can be implemented with computational cost \(\mathcal{O}(N \ln (1/\varepsilon)+|\mathcal{Z}|\ln(1/\varepsilon))\). When only coefficients of the spherical polynomials \(f\) are given instead of \(f(\xi), \xi\in\mathcal{X}_M\), the paper uses the fast Fourier transform (FFT) to precompute \(f(\xi), \xi\in\mathcal{X}_M\) and the computational cost of \(\Phi_{N,\delta}f(x)\), \(x\in \mathcal{Z}\), is increased by \(\mathcal{O}(N\ln N)\). For a fixed precision \(\varepsilon\) and oversampling constant \(\sigma\), the paper then compares their algorithms with two other methods: the spline interpolation based on Lagrange interpolation and NFFT with the Kaiser-Bessel window function. Using the tensor product approach and spherical coordinates, the paper shows that one can approximate spherical polynomials \[ f(\theta,\lambda) = \frac{1}{(2\pi)^2} \int_0^{2\pi}\int_0^{2\pi} \mathcal{K}_N(\theta-\theta')\mathcal{K}_N(\lambda-\lambda')f(\theta',\lambda')d\lambda'd\theta', \] by \[ \Phi_{N,\delta}^2f(\theta,\lambda):=\frac{1}{4KL}\sum_{0\leq k\leq K, \rho(\theta,\theta_k)\leq \delta}\sum_{0\leq \ell\leq 2L, \rho(\lambda,\lambda_\ell)\leq \delta}\mathcal{K}_N(\theta-\theta_k)\mathcal{K}_N(\lambda-\lambda_\ell)f(\theta_k,\lambda_\ell), \] which can be evaluated fast using the same approach of its 1D version. The paper discusses the tensor product approach by Lagrange interpolation and the nonequispaced fast spherical Fourier transform. No numerical comparison among the three methods is provided.