Adaptive spatial filters: the LCMV beamformer
MNE and its relatives are fixed linear inverses — the operator \mathbf{G} ignores the data's covariance structure. A beamformer is different: it designs a spatial filter per location and from the data itself, tuned to pass a target source at unit gain while suppressing everything else. It is the same idea as a steerable, adaptive antenna array pointed at one location in the brain.
\mathbf{w}(\mathbf{r}) = \frac{\mathbf{C}^{-1}\mathbf{l}(\mathbf{r})}{\mathbf{l}(\mathbf{r})^{\top}\mathbf{C}^{-1}\mathbf{l}(\mathbf{r})} \qquad\text{s.t.}\quad \mathbf{w}^{\top}\mathbf{l}=1The scalar LCMV (linearly-constrained minimum-variance) beamformer: minimize output variance wᵀCw subject to unit gain on the target lead field l(r); C is the sensor data covariance.
A beamformer builds a custom spatial filter for each brain location that lets that spot's signal through at full strength while actively suppressing everything else, using the data's own covariance to know what to cancel. Like a directional microphone aimed at one voice in a crowd.
- \mathbf{w}(\mathbf{r})
- The filter weights aimed at location \mathbf{r}.
- \mathbf{l}(\mathbf{r})
- The lead field (forward pattern) of the target source.
- \mathbf{C}
- The sensor data covariance — what the filter learns to suppress.
- \mathbf{w}^{\top}\mathbf{l}=1
- The unit-gain constraint keeping the target undistorted.
Scan \mathbf{r} across the brain and the beamformer output power \hat{P}(\mathbf{r})=(\mathbf{l}^{\top}\mathbf{C}^{-1}\mathbf{l})^{-1} becomes a source image. Because it uses the data covariance, an LCMV beamformer can achieve sharper, more focal localization than MNE for well-separated sources — but that advantage rests on a strong assumption.
DICS: beamforming in the frequency domain
For oscillatory activity and connectivity, the beamformer is recast in the frequency domain as DICS (dynamic imaging of coherent sources): replace the time-domain covariance \mathbf{C} with the cross-spectral density matrix at a frequency of interest, and the same LCMV algebra returns source power — and coherence between source pairs — band by band. This is a natural fit for BCI, where the informative features are often oscillatory (mu/beta motor rhythms).
Connectivity in source space and the leakage trap
A major reason to localize at all is to estimate how brain regions communicate. But source leakage (guide 4) is a landmine: because two estimated time courses share leaked signal from a common source, they exhibit zero-lag spurious coupling that perfectly mimics genuine connectivity. Naive coherence in source space is not to be trusted.
The standard defenses exploit the fact that linear leakage is instantaneous (zero phase lag). Metrics such as the imaginary part of coherency, or the phase-lag index, deliberately discard zero-lag interactions and keep only time-delayed — and therefore more plausibly physiological — coupling. Explicit leakage correction (e.g. orthogonalizing signals before computing connectivity) is a routine, necessary step, not an optional refinement.
MEG, EEG, and the OPM revolution
How accurate is any of this? With good sensor coregistration and a realistic head model, localization errors for well-focal, superficial sources are typically on the order of a centimeter — degrading for deep, weak, or multiple simultaneous sources. MEG generally localizes tangential superficial sources more accurately (no skull distortion), while EEG uniquely captures radial sources; a combined EEG+MEG forward model outperforms either alone. Speak in these qualitative ranges — exact error numbers depend heavily on montage, model and source depth.
The frontier hardware is the optically pumped magnetometer (OPM). Unlike cryogenic SQUID-MEG with its fixed helmet and centimeters of standoff, OPMs are small, room-temperature sensors that sit directly on the scalp — closer sensors mean larger signals and finer spatial sampling, and the arrays can adapt to any head size, including children and moving participants. On-scalp MEG promises a genuine step-change in non-invasive source resolution.
Open problems and how to keep yourself honest
Source imaging remains an inference under strong priors, and its open problems are real: recovering deep sources (hippocampus, brainstem) against the null space; resolving multiple correlated sources; propagating forward-model uncertainty (conductivity, coregistration) into honest source-level error bars; and validating against a ground truth we rarely possess. None of these is fully solved, and progress is as much about discipline as about new algorithms.
- Simulate a known source through your exact pipeline — if you cannot recover a planted dipole, do not trust the real result.
- Report the resolution: point-spread and cross-talk functions for your regions of interest quantify exactly what your setup can and cannot separate.
- Validate against independent modalities — intracranial EEG, direct cortical stimulation, lesions, or fMRI — wherever they are available.
- Quantify forward uncertainty: perturb the skull conductivity and electrode positions and check that your conclusions survive the perturbation.
Used this way — as a hypothesis-generating tool whose priors and limits you state out loud — source localization turns blurred scalp maps into genuine, testable claims about where and how the brain is active. That honest discipline, more than any single algorithm, is what separates graduate-level source analysis from pretty pictures.