Super-Resolution on the Two-Dimensional Unit Sphere

Frank Filbir, Kristof Schröder, Anna VeselovskaarXiv 2021 · arXiv

Abstract

This paper addresses the problem of recovering high-resolution "spiky" signals—modeled as atomic measures—from a limited set of low-frequency spherical harmonic observations. By framing the recovery as a Total Variation (TV) minimization problem, the authors establish a rigorous mathematical framework for exact recovery on the two-dimensional unit sphere. The study introduces a dual certificate constructed via Jackson kernels and provides both theoretical bounds and numerical evidence for the separation distance required between spikes to ensure uniqueness and stability.

Introduction

In diverse scientific fields such as astrophysics, medical imaging, and global topography, data is naturally collected on a spherical surface. Often, the underlying signal of interest consists of discrete, localized sources—such as stars in a sky or sound sources in a room. However, physical measuring instruments typically act as low-pass filters, blurring these "spikes" into coarse, low-resolution maps.

The challenge of super-resolution (or de-convolution) is to mathematically "zoom in" and recover the exact locations and intensities of the original spikes using only a finite number of low-frequency moments. While this problem has been extensively studied for Euclidean domains (like the circle or the torus), the geometry of the sphere introduces unique topological hurdles. Most notably, the "Hairy Ball Theorem" prevents the existence of a continuous, non-vanishing vector field on the sphere, which complicates the construction of the mathematical tools—specifically dual certificates—needed to prove recovery uniqueness.

Problem Formulation and Methodology

The authors treat the super-resolution task as a Total Variation (TV) minimization problem. This approach serves as the continuous-space analog of the $\ell_1$-minimization used in compressed sensing to find sparse solutions in discrete settings.

1. Dual Certificates and Jackson Kernels

To prove that TV minimization identifies the "true" signal, one must construct a dual certificate. This is a specific polynomial that peaks exactly at the spike locations (reaching a value of 1) and remains strictly below 1 everywhere else. The authors utilize the classical Jackson kernel, derived from the Fejér kernel, to build this certificate. This choice is strategic: the Jackson kernel's explicit form allows for sharp, localized bounds on its derivatives, which is essential for handling the sphere's curvature and geodesic distances.

2. Numerical Optimization

The paper proposes two primary paths for solving the recovery problem:

  • Semidefinite Programming (SDP): By applying the "Bounded Real Lemma," the authors transform the infinite-dimensional TV minimization into a finite-dimensional SDP. This allows the problem to be solved using standard interior-point methods, providing a path to exact recovery.
  • Discretization: Alternatively, the problem can be projected onto a fine grid and solved as a basis pursuit problem. To mitigate the "leakage" that occurs when a spike does not fall exactly on a grid point, the authors employ kernel density estimation to refine the final spike locations.

Key Insights: Geometry and Separation

A central theme of the research is the impact of spherical geometry on signal recovery. Unlike the torus, where translation is uniform, the sphere requires a point-wise analysis of the interpolation problem due to the lack of a continuous basis for the tangent space.

Furthermore, the authors bridge the gap between theoretical requirements and numerical reality regarding the separation distance. For the proof of exact recovery to hold, spikes must be separated by a minimum distance. While the theoretical derivation provides a conservative bound of $19.2\pi/N$ (where $N$ is the maximum degree of the spherical harmonics), numerical experiments suggest that the actual required separation is much lower, with successful recovery observed at constants $\nu$ between $1.15\pi$ and $1.33\pi$.

Results and Performance

  • Exact Recovery: In noise-free simulations using only 49 moments ($N=6$), the algorithm successfully recovered spike locations and intensities with an error margin of less than $10^{-7}$.
  • Robustness to Noise: The framework is not limited to idealized data. By incorporating Tikhonov-type regularization into the SDP formulation, the authors demonstrated that the method degrades gracefully when low-frequency data is corrupted by deterministic noise.
  • Empirical Validation: Using synthetically generated test measures consisting of Dirac spikes with random locations and amplitudes, the authors confirmed that the recovery error ($\epsilon_x$ for location and $\epsilon_c$ for intensity) remains negligible as long as the separation condition is met.

Limitations and Future Directions

Despite its mathematical elegance, the SDP approach faces high computational costs as the number of moments ($N$) increases, primarily due to the expanding size of the resulting matrices. Additionally, the discretization method remains sensitive to "basis mismatch," where spikes falling between grid points can cause artifacts, though refinement techniques help mitigate this.

Future work may involve exploring alternative solvers to handle singularities at the poles, where spherical coordinates can become problematic for standard optimization routines. Extending these results to higher-dimensional spheres or more complex manifold geometries represents a natural progression of this framework.

Summary updated

BibTeX
@misc{filbir2021super,
  author = {Frank Filbir and Kristof Schröder and Anna Veselovska},
  title = {{Super-Resolution on the Two-Dimensional Unit Sphere}},
  year = {2021},
  eprint = {2107.10762},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2107.10762},
}
Figure