Riemann Localisation on the Sphere

Yu Guang Wang, Ian H. Sloan, Robert S. WomersleyJournal of Fourier Analysis and Applications 2016 · DOI · arXiv

Abstract

This paper investigates the Riemann localisation property on $d$-dimensional spheres ($\mathbb{S}^d$). While localisation is a foundational property of Fourier series on the circle (1D), its extension to spherical geometries reveals a significant dimensional divide. We demonstrate that while the property holds for the 2D sphere, it fails for spheres in three or more dimensions under standard Fourier-Laplace partial sums. However, we prove that localisation can be restored in any dimension by applying a sufficiently smooth filter to the Fourier coefficients. These findings have significant implications for local signal processing and multiresolution analysis on spherical manifolds.

1. Introduction

In classical harmonic analysis on a 1D circle, the Riemann localisation principle states that the behavior of a Fourier series partial sum at a point depends only on the behavior of the function in an arbitrarily small neighborhood of that point. Specifically, if a function is zero on an interval, its Fourier partial sums converge to zero within that interval. This property is intuitive for local signal processing, as it suggests that distant noise or signals do not interfere with local approximations.

As research moves toward complex data structures in geophysics and atmospheric modeling, understanding whether this property holds on $d$-dimensional spheres ($\mathbb{S}^d$) becomes critical. This study rigorously examines the conditions under which localisation occurs, identifying a sharp transition in behavior as the spatial dimension increases.

2. Theoretical Framework and Methodology

The study utilizes a rigorous analytical approach centered on Fourier-Laplace series. Functions on the sphere are represented using spherical harmonics, which serve as the spherical analogues to the trigonometric basis functions used in 1D Fourier analysis.

2.1 Local Convolution and Dirichlet Kernels

To test the localisation property, the authors define a "local convolution." This involves analyzing the Fourier approximation of a function that has been "zeroed out" in a specific neighborhood. The core of the mathematical proof relies on deriving asymptotic estimates for Dirichlet kernels and Jacobi polynomials. These estimates describe the decay and oscillation of the approximation kernels as the degree of the polynomial ($L$) approaches infinity.

2.2 Filtered Kernels

Recognizing the limitations of standard partial sums, the authors introduce a filter function $g$. This function modifies the Fourier coefficients to create a "filtered" approximation. The smoothness of this filter, denoted by $\kappa$, is a primary variable in determining the convergence rates of the resulting series.

3. The Dimensionality Gap

The most striking finding of this research is the "magic" of dimension $d=2$. The study identifies a sharp divide in the behavior of standard Fourier-Laplace partial sums:

  • The 2D Case ($d=2$): For the 2D sphere (the geometry of the Earth's surface), the Riemann localisation property holds for sufficiently smooth functions (specifically those in the Sobolev space $W^s_p$ with $s \geq 2$). In this case, the $L_p$-norm of the local convolution converges to zero at a rate of $O(L^{-1/2})$.
  • Higher Dimensions ($d \geq 3$): For spheres in three or more dimensions, standard localisation fails. The authors provide a lower bound demonstrating that the $L_p$-norm either grows or remains bounded away from zero. This failure occurs because the "tails" of the Dirichlet kernel do not decay rapidly enough in higher dimensions, allowing distant function values to exert a persistent influence on local approximations.

4. Restoring Localisation via Filtering

While standard Fourier methods fail in high dimensions, the authors prove that localisation is not fundamentally impossible. By applying a smooth filter $g$ to the Fourier coefficients, localisation can be restored for a sphere of any dimension.

4.1 Smoothness Requirements

The effectiveness of the filter is directly tied to its smoothness. The authors establish that the filtered local convolution converges to zero at a rate of: $$O(L^{-(\kappa - d/2 + 3/2)})$$ This formula reveals a "dimension penalty" of $d/2$. To achieve successful localisation, the filter's smoothness level $\kappa$ must be high enough to overcome this penalty. Essentially, as the dimension of the sphere increases, the filter must become progressively smoother to suppress the interference from distant points.

5. Conclusion and Practical Implications

This research clarifies a long-standing question in spherical harmonic analysis. The discovery that localisation fails for $d \geq 3$ under standard conditions explains why certain local approximation techniques may struggle in higher-dimensional data spaces.

For practitioners in geopotential modeling and signal processing, these results provide a theoretical roadmap for handling "missing observation" problems. By selecting a filter with appropriate smoothness, researchers can ensure that local data reconstructions remain isolated from distant noise, regardless of the dimensionality of the manifold. Future work remains to identify optimal filter shapes for specific geophysical applications and to explore the limits of localisation for functions with lower levels of regularity.

Summary updated

BibTeX
@article{wang2016riemann,
  author = {Yu Guang Wang and Ian H. Sloan and Robert S. Womersley},
  title = {{Riemann Localisation on the Sphere}},
  journal = {Journal of Fourier Analysis and Applications},
  year = {2016},
  doi = {10.1007/s00041-016-9496-4},
}

Loading related content...

Figure