When two brain regions fire in lockstep, with no time delay between them, electroencephalography (EEG) and magnetoencephalography (MEG) see something that looks exactly like an artifact. Every scalp sensor picks up a blurred mixture of many cortical regions. Mapping sensor data back onto the cortex ("source localization") is ill-posed, so the recovered activity of one region leaks into the estimates of its neighbours. That leakage is instantaneous. The standard remedy is to strip every zero-lag correlation out of the reconstructed signals before measuring connectivity. It removes leakage, and it also removes any genuine zero-lag coupling, such as coupling carried by reciprocal cortico-cortical or thalamo-cortical loops. Current pipelines are therefore blind to a whole class of interaction by construction, and they are blind in every frequency band and at every moment in time.
We propose to stop discarding leakage and to model it instead. The forward model, meaning the physics that says how each cortical source projects to the sensors, fully determines how much zero-lag correlation a single uncoupled source produces. Our approach fits band-specific, state-dependent source coupling matrices directly to the sensors' band-limited cross-spectra through that forward model. Zero-lag correlation that the physics can explain is attributed to leakage. Anything left over is attributed to coupling, but only for region pairs whose lead fields are distinct enough for the question to have an answer at all. The main deliverable is a coupling estimator plus a per-head-model identifiability map, together with a clear test of when this beats orthogonalization and when it does not.
Core Claim
Existing leakage-robust connectivity methods remove every instantaneous interaction. The multivariate power-envelope test of Measuring functional connectivity in MEG: A multivariate approach insensitive to linear source leakage orthogonalizes the test voxel against the seed voxel after beamforming. The dynamic instantaneous amplitude correlation (IAC) pipeline of Tracking dynamic brain networks using high temporal resolution MEG measures of functional connectivity does the same pairwise orthogonalization before tracking networks over time. Both papers motivate this proposal: they show that leakage correction is necessary, and they show the cost of their particular fix.
Our claim is that zero-lag coupling is recoverable in a covariance-domain generative model, for a specific and computable subset of region pairs. We write $\Sigma_{f,k}$ for the sensor cross-spectral matrix in frequency band $f$ during network state $k$, $L$ for the lead field (sensors × cortical parcels), $S_{f,k}$ for the source cross-spectral matrix, and $N$ for the noise covariance, estimated from empty-room recordings. The model is
$$\Sigma_{f,k} = L\, S_{f,k}\, L^{\top} + N .$$
We estimate $S_{f,k}$, constrained to be positive semidefinite, by maximizing a Whittle (frequency-domain Gaussian) likelihood of the observed band-limited sensor cross-spectra. We add a group-sparsity penalty on the off-diagonal entries of $S_{f,k}$, and the penalty strength is allowed to differ by band.
The real part of an off-diagonal entry of $S_{f,k}$ is the zero-lag coupling. The imaginary part is lagged coupling. Leakage never appears as a parameter, because it lives entirely in $L(\cdot)L^{\top}$.
The states $k$ address the static-connectivity problem. A small number of discrete states is inferred from the time-resolved sensor cross-spectra. This follows the logic of Modelling discrete states and long-term dynamics in functional brain networks, but the emission model is our forward-model covariance rather than a downstream model of pre-extracted parcel time series.
Research Program
Identifiability analysis. The first component is analytical and comes before any estimation. The map from source coupling to sensor covariance is linear in $S$. For a parcel pair $(i,j)$, the zero-lag coupling is identifiable only if the sensor pattern it produces, $\ell_i \ell_j^{\top} + \ell_j \ell_i^{\top}$ (where $\ell_i$ is column $i$ of $L$), is not well approximated by patterns in the span of the diagonal power terms and the other coupling terms. We compute this pair's residual norm after projection. From it we build an identifiability map for each head model, for example the New York Head used in SEREEGA: Simulating Event-Related EEG Activity. We build it for MEG, for EEG, and for combined MEG+EEG, and at several parcellation resolutions.
We expect adjacent parcels to be essentially unidentifiable, which honestly concedes that orthogonalization is the right tool there. We expect distant pairs, especially interhemispheric homologues, to be identifiable. The estimator reports coupling only where this map clears a threshold that we set on held-out simulations.
The estimator. The second component is a scalable fitting procedure. Working with $M \times M$ cross-spectra, rather than $M \times T$ time series, makes the cost independent of recording length. This matters because the joint state-space methods are bottlenecked by it. A state-space model of cross-region dynamic connectivity in MEG/EEG was restricted to two regions of interest. Solving large-scale MEG/EEG source localisation and functional connectivity problems simultaneously using state-space models scales to more regions but models broadband autoregressive lags.
We will use a parcel-level lead field with patch-summed columns, at 68 to 400 parcels. States and coupling matrices are updated by alternating two steps. The first step fixes the $S_{f,k}$ and assigns time windows to states. The second step fixes the assignments and runs proximal gradient updates on each $S_{f,k}$. The group lasso machinery follows Modeling Sparse Connectivity Between Underlying Brain Sources for EEG/MEG, which jointly estimated demixing and stationary autoregressive coupling. We differ from that work in three ways: we fix the demixing to the physical lead field, we add band and state structure, and we target zero-lag rather than lagged coupling.
Frequency-dependent sparsity. The third component tests whether sparsity should depend on frequency. Low-frequency rhythms are thought to engage broad networks and high-frequency rhythms focal patches (Spectral Changes in Cortical Surface Potentials during Motor Movement). We therefore compare a single shared penalty against a penalty tuned separately for each band. We treat this as an ablation, not a headline.
Baselines
The naive reference is the standard two-step pipeline: a minimum-norm or beamformer inverse, parcel averaging, band-pass filtering, and zero-lag correlation of amplitude envelopes, with no leakage correction, computed in sliding windows. It will show massive spurious coupling between neighbours. It fixes the floor on false positives.
The strongest existing competitor is orthogonalized IAC with recurrence-based state segmentation, as in Tracking dynamic brain networks using high temporal resolution MEG measures of functional connectivity. It is the one published method in the file that is leakage-robust, band-specific, and dynamic at the same time. That paper reports that IAC and wavelet coherence reach their accuracy plateau at an SNR of about 10 dB in a two-node Jansen-Rit simulation. It also reports cross-metric subnetwork correlations of R = 0.75 to 0.95 on HCP resting MEG (N=89). No reported result was found on Lacuna for zero-lag coupling recovery by any method.
We will run this baseline on identical simulated and real data, with the same parcellation, bands, and noise covariance. We will tune its window and threshold settings with the same number of simulation-based trials as our penalty strength. Beating it means two things. On lagged edges, it must match within confidence intervals. On zero-lag edges in identifiable pairs, it must reach a clearly higher edge-detection area under the curve (AUC). The baseline scores zero-lag edges near chance by design, so a win there is expected. The real test is that we must not raise the false-positive rate on neighbouring pairs that are uncoupled.
The third baseline is the one-step time-varying autoregressive state-space model of A state-space model of cross-region dynamic connectivity in MEG/EEG. Its innovation covariance can in principle absorb zero-lag coupling. Because of its scaling limits, we will run it only in small-region simulations. It is the closest method to ours in treating leakage through the forward model, so it tests whether our covariance-domain formulation adds anything beyond "one-step" modelling.
The key ablation removes the forward-model leakage term. We fit the same state and sparsity model to parcel-level cross-spectra computed after a minimum-norm inverse. This isolates whether explaining leakage in sensor space is the active ingredient.
Why not just…?
Why not just orthogonalize, or use lag-only metrics such as the imaginary part of coherency or the phase lag index? These methods have zero sensitivity to instantaneous coupling by construction, and that is exactly the class of interaction we target. The orthogonalization used in Measuring functional connectivity in MEG: A multivariate approach insensitive to linear source leakage is right for adjacent parcels, and our identifiability map says so explicitly. The comparison is decided by simulations that plant pure zero-lag edges: our method must detect them, and lag-only metrics cannot.
Why not just compute a leakage null from the resolution matrix of a two-step inverse and flag correlations that exceed it? This is the real threat. The resolution matrix of a two-step inverse predicts how much correlation a lone source would produce, and excess over that prediction flags coupling. When only one or two sources are active, this may work nearly as well as our method. It is essentially our model with the coupling fitted after inversion rather than before. We will implement it as a baseline alongside the ablation above.
Our expectation is that it degrades when several coupled sources are active at once. In that case the null needs source powers that are themselves distorted by the inverse, a distortion documented by the amplitude bias results in The Iterative Reweighted Mixed-Norm Estimate for Spatio-Temporal MEG/EEG Source Reconstruction. If it matches our method across network sizes, we will report that the simpler procedure suffices, and the contribution narrows to the identifiability map and the state model.
Why not just learn connectivity on sensors with a graph neural network? Sensor-space graphs encode volume conduction directly. They also generalize poorly across subjects: GNN4EEG: A Benchmark and Toolkit for Electroencephalography Classification with Graph Neural Network reports 94.5% intra-subject two-class accuracy against 31.3% for the best cross-subject nine-class model. Such a graph also produces no statement about which cortical regions are coupled, which is the question here.
Proposed Experiments
Simulations. Ground-truth simulations will use the New York Head in SEREEGA, with MEG-like and EEG-like sensor sets. Networks of 4 to 20 Jansen-Rit neural masses will generalize the two-node setup in the IAC paper. Some edges will be driven by a shared thalamic-like input or by fast reciprocal coupling, which yields near-zero lag. Other edges will carry conduction delays of 10 to 30 ms. Coupling will switch between two or three states with known dwell times, and noise will span 0 to 20 dB SNR.
Metrics are computed separately for zero-lag and lagged edges. They are edge-detection AUC, the false-positive rate among uncoupled neighbouring pairs, the Pearson correlation between estimated and true zero-lag coupling strengths on true edges, and state-assignment accuracy. We report everything as a function of the identifiability score, so that we can check whether the map predicts where recovery succeeds.
Real data. On HCP resting MEG, the question is whether estimated zero-lag interhemispheric coupling of homologous regions in the alpha and beta bands reproduces across the two scan sessions. On the MIND somatosensory data, we look for band-specific coupling between contralateral and ipsilateral secondary somatosensory cortex after median-nerve stimulation. Only amplitude sequences have been reported there so far (M/EEG source localization with multi-scale time-frequency dictionaries). On the Nottingham MEGUK data, we look at state-specific zero-lag networks.
Decisive check. The most informative test uses the simultaneous MEG and EEG of the Wakeman and Henson face dataset. MEG and EEG have different lead fields, so their leakage patterns differ. Genuine zero-lag coupling between identifiable pairs should agree across the two modalities fitted separately, and leakage-driven artifacts should not. We will compare cross-modal agreement for our estimates, for unorthogonalized two-step correlation, and for the resolution-matrix null. Two further sanity checks will be run. The first is the pre-stimulus test from the state-space literature: ventral-stream coupling should not appear before face onset. The second uses empty-room recordings, where the estimator must report no coupling.
The proposal fails if any of three things happens: zero-lag estimates in identifiable pairs disagree between MEG and EEG as badly as the unorthogonalized baseline, the resolution-matrix null matches our method in multi-source simulations, or the false-positive rate on neighbouring pairs rises above that of the orthogonalized baseline.
References
- Measuring functional connectivity in MEG: A multivariate approach insensitive to linear source leakage
- Tracking dynamic brain networks using high temporal resolution MEG measures of functional connectivity
- Modelling discrete states and long-term dynamics in functional brain networks
- SEREEGA: Simulating event-related EEG activity
- A state-space model of cross-region dynamic connectivity in MEG/EEG
- Solving large-scale MEG/EEG source localisation and functional connectivity problems simultaneously using state-space models
- Modeling Sparse Connectivity Between Underlying Brain Sources for EEG/MEG
- Spectral Changes in Cortical Surface Potentials during Motor Movement
- The Iterative Reweighted Mixed-Norm Estimate for Spatio-Temporal MEG/EEG Source Reconstruction
- GNN4EEG: A Benchmark and Toolkit for Electroencephalography Classification with Graph Neural Network
- M/EEG source localization with multi-scale time-frequency dictionaries