Weighted Group Lasso for a static EEG problem

We investigate the weighted Group Lasso formulation for the static inverse electroencephalography (EEG) problem, aiming at reconstructing the unknown underlying neuronal sources from voltage measurements on the scalp. By modelling the three orthogonal dipole components at each location as a single coherent group, we demonstrate that depth bias and orientation bias can be effectively mitigated through the proposed regularization framework. On the theoretical front, we provide concise recovery guarantees for both single and multiple group sources. Our numerical experiments highlight that while theoretical bounds hold for a broad range of weight definitions, the practical reconstruction quality, for cases not covered by the theory, depends significantly on the specific weighting strategy employed. Specifically, employing a truncated Moore-Penrose pseudoinverse for the involved weighting matrix gives a small Dipole Localization Error (DLE). The proposed method offers a robust approach for inverse EEG problems, enabling improved spatial accuracy and a more physiologically realistic reconstruction of neural activity.

💡 Research Summary

This paper presents a comprehensive investigation into the application of the Weighted Group Lasso regularization framework for solving the static inverse electroencephalography (EEG) source localization problem. The core challenge addressed is the ill-posed nature of reconstructing underlying neuronal current sources from scalp potential measurements, which traditionally suffers from depth bias (misattributing deep sources to superficial ones) and orientation/angle bias (reconstructed dipoles artificially aligning with the computational grid axes).

The authors’ key innovation is to model the three orthogonal dipole moment components (x, y, z) at each candidate brain location as a single, coherent group. The proposed optimization problem minimizes a cost function comprising a data fidelity term (∥BAx - Bb∥²) and a weighted group sparsity penalty (α Σ_g ∥(BA)_g x_g∥). Here, matrix B introduces flexibility for weighting. Using a group ℓ2-norm penalty promotes group-level sparsity, meaning all components within a group are either jointly selected or set to zero. This enforces rotational invariance, effectively eliminating the orientation bias inherent in component-wise ℓ1 penalties.

On the theoretical front, the paper provides concise and transparent recovery guarantees. Theorem 3.2 proves that for a single active group source, under a linear independence assumption (Assumption 3.1), the solution to the group pursuit problem (min Σ_g ∥C_g x_g∥ subject to Cx = Cx*) is unique and matches the true source. Theorem 3.4 extends this to multiple group sources, demonstrating exact recovery under a condition where the influences of different active groups do not overlap on each other’s support sets (Assumption 3.3). Furthermore, Theorem 3.5 shows that for a single group source, the solution to the regularized Group Lasso problem is a scaled version of the true source.

The numerical experiments, conducted using the SEREEGA toolbox and the ICBM-NY head model, reveal a critical insight that bridges theory and practice. While the theoretical recovery guarantees hold for a broad range of weight matrix (B) definitions, the practical reconstruction quality—especially for complex scenarios not covered by the theory—is highly sensitive to the specific choice of B. The authors empirically demonstrate that employing a truncated Moore-Penrose pseudoinverse of the lead field matrix A for B yields the best performance, quantified by a small Dipole Localization Error (DLE). This choice is interpreted as providing numerical stability by truncating small singular values while simultaneously compensating for the signal attenuation of deeper sources, thereby mitigating depth bias.

In summary, this work makes a dual contribution: it rigorously establishes that the group sparsity structure of Weighted Group Lasso fundamentally addresses the orientation bias in EEG inverse problems, and it empirically identifies that a physics-informed weighting strategy (via the truncated pseudoinverse for B) is crucial for achieving high spatial accuracy and physiologically realistic reconstructions in practice. The method offers a robust framework that improves upon traditional sparse recovery techniques for neuroimaging.