Abstract
We present a novel method for the accurate reconstruction of non-bandlimited finite rate of innovation (FRI) signals on a sphere. By utilizing an annihilating filter-based approach, we demonstrate the ability to recover the parameters of Dirac functions from a finite set of spherical harmonic (SH) observations. Our method significantly improves numerical stability through better conditioning of the underlying mathematical systems. Numerical experiments show that the proposed technique outperforms existing state-of-the-art methods by a factor of $10^7$ or more in reconstruction accuracy while requiring the same or fewer samples.
Introduction
Signals defined on a spherical domain are prevalent in diverse fields, including diffusion MRI, acoustics, and cosmology. Often, these signals are sparse, consisting of a finite number of localized "spikes" or Dirac deltas. In signal processing, these are categorized as Finite Rate of Innovation (FRI) signals because they are characterized by a finite set of parameters (amplitudes and locations) despite having infinite frequency content.
Reconstructing these signals from limited measurements is a classic inverse problem. While theoretically possible, previous methods have struggled with numerical ill-conditioning. As the number of Diracs ($K$) increases, the sensitivity of the reconstruction to small numerical errors grows exponentially, leading to a rapid decline in precision. This paper addresses this bottleneck by introducing a more stable algebraic framework for parameter estimation.
Methodology
The proposed method adapts the 1D annihilating filter technique to the geometry of the sphere. The core innovation lies in the rearrangement of spherical harmonic coefficients into a structured matrix that exhibits superior numerical properties.
- Coefficient Transformation: Spherical harmonic coefficients are algebraically transformed into a new sequence, $d_{pm}$, which follows a sum-of-exponentials model.
- Annihilating Filter Construction: A matrix $\mathbf{Z}$ is constructed from these transformed coefficients. The null-space of $\mathbf{Z}$ provides the coefficients of a filter that "annihilates" the signal. The roots of this filter directly correspond to the longitudinal ($\phi$) and colatitudinal ($\theta$) coordinates of the Diracs.
- Amplitude Recovery: Once the locations are identified, the amplitudes ($\alpha$) are recovered by solving a linear Vandermonde system.
By optimizing the structure of $\mathbf{Z}$, the method achieves a required bandlimit of $L \approx K + \sqrt{K}$, matching the efficiency of the most advanced existing algorithms while providing vastly superior precision.
Experimental Results
To validate the method, we conducted extensive synthetic experiments in MATLAB, comparing our approach against two leading techniques (referred to as [6] and [14] in the literature). We tested signals with $K$ ranging from 2 to 20, averaging results over 1,000 independent trials for each configuration.
As shown in the figure above, the proposed method achieves a massive gain in precision. While existing methods see their mean-squared error (MSE) for location ($E_\theta, E_\phi$) and amplitude ($E_\alpha$) rise toward $10^0$ as $K$ increases, the proposed method remains stable, often maintaining errors as low as $10^{-10}$ to $10^{-25}$. This represents an improvement of at least seven orders of magnitude ($10^7$) over the state-of-the-art.
Conclusion
This work provides a robust framework for reconstructing sparse signals on the sphere. By focusing on the numerical conditioning of the annihilating filter, we have enabled high-precision recovery that was previously unattainable for larger numbers of Diracs. Future work will explore the robustness of this method in the presence of measurement noise and its application to real-world datasets like diffusion MRI fiber orientation mapping.
Summary updated