When we record brain activity using Electroencephalography (EEG) or Magnetoencephalography (MEG), we are essentially looking at a blurry shadow cast by a complex, three-dimensional light show. The fundamental challenge of neuroimaging is to work backward from these surface measurements to pinpoint the exact anatomical origins of neural activity. This is known as the inverse problem, and it is mathematically "ill-posed."
The Dimensionality Gap
The core of the problem lies in a massive discrepancy in information. As Feng Liu et al. and Jair Montoya-Martínez et al. explain, the number of potential internal brain source locations (often thousands of voxels) vastly exceeds the number of available scalp sensors (usually dozens to a few hundred).
Ali Hashemi et al. describe this as a "dimensionality gap." Because of this imbalance, the system is underdetermined: an infinite variety of internal activity patterns can produce the exact same external measurements. As Christoph Bledowski et al. note, this mathematical ambiguity prevents researchers from definitively localizing the anatomical origins of even well-known signals like the P300 wave without additional assumptions.
The Physical Smearing of Neural Signals
If the dimensionality gap is a mathematical hurdle, volume conduction is the physical one. As electrical signals travel from the brain through the cerebrospinal fluid, the skull, and the scalp, they encounter varying levels of resistance. Scott Makeig et al. describe this as "spatial smearing," where activity from multiple brain regions blends together before reaching a single electrode.
This effect is worsened by the directional properties of head tissues. Hans Hallez et al. show that the skull is anisotropic—meaning electricity travels through it differently depending on the direction. This causes signals to attenuate and spread, making it difficult to distinguish between two nearby sources. To the sensors, these distinct events appear as large, overlapping clusters of activity, a phenomenon the authors of Strategies for automatic generation of information processing pathway maps call "point spread."
Constraining the Infinite Solution Space
To solve an ill-posed problem, researchers must impose constraints—mathematical "rules" that narrow down the infinite possibilities to the most biologically plausible ones.
One common strategy is sparsity. The intuition is that at any given moment, only a small fraction of the brain is likely responsible for a specific task. Daniel Strohmeier et al. introduced the irMxNE framework, which uses non-convex penalties to eliminate "ghost" sources that often plague traditional imaging. Similarly, Michael Lim et al. proposed a "group lasso" approach, which groups activity into functionally defined regions of interest to produce more focal estimates.
Another approach involves supervised learning. By incorporating task labels directly into the reconstruction process, Feng Liu demonstrated that it is possible to filter out the "noise" of spontaneous brain activity and isolate the precise neural patterns associated with specific cognitive states.
Integrating Geometric and Multimodal Priors
Modern methods increasingly rely on the physical reality of the brain to guide the math. Song Wang et al. developed Geometric-informed Basis Functions (GBFs), which use the actual folds and curves of a subject's cortical surface as a mathematical prior. This turns blurry scalp signals into sharp, biologically grounded maps.
Researchers are also fusing different types of data to cancel out individual weaknesses. For example:
- MEG + fMRI: Beige Jerry Jin et al. used transformer-based models to fuse the high temporal resolution of MEG with the spatial precision of fMRI.
- Scalp + Intracranial EEG: Shihao Yang et al. integrated surface recordings with deep-brain electrodes to reduce "signal leakage" and better map how information flows between regions.
- Optimal Transport: Hicham Janati et al. developed Minimum Wasserstein Estimates (MWE), which uses Optimal Transport to account for the fact that while two people’s brains are shaped differently, their functional activations should be "spatially close."
By treating the inverse problem not just as a data challenge, but as a synthesis of geometry, physics, and statistics, these methods are moving neuroimaging closer to a true, high-resolution window into the thinking brain.
Key Foundational Work
Independent Component Analysis of Electroencephalographic Data (1995) — contributed 2 concept elements to this direction.