Do White-Matter Conduction Delays Matter Above Alpha? A Leakage-Matched Test in MEG

Signals take time to travel along the brain's white-matter cables. A 10 cm fiber carries a signal in roughly 10–20 ms. For a slow 10 Hz rhythm, that delay is a modest fraction of one oscillation cycle. For a 40 Hz rhythm, it is most of a cycle. A careful modeling study found something surprising: a whole-brain model that ignores travel time entirely matches resting-state magnetoencephalography (MEG, which records the brain's magnetic fields at millisecond resolution) at least as well as a model with realistic, fiber-length-dependent delays. We call this the "delay paradox." That study looked only at the alpha rhythm and only at one hemisphere. We propose to test whether the paradox survives at faster rhythms, where physics says delays should matter most.

We also propose to test whether the paradox may be partly a measurement artifact. MEG source estimates blur neighboring regions together, which creates spurious zero-lag synchrony. A model with no delays produces exactly that kind of zero-lag synchrony. If the empirical data are compared to clean simulated sources, the zero-delay model gets an unfair advantage. We will push every simulated brain through the same blurring as the real data. Because MEG cohorts with matched diffusion imaging are small and MEG phase measures are unreliable within a single subject, the study is built around held-out subjects, test–retest noise ceilings and a power analysis. The power analysis asks how many subjects are needed before "the models are equivalent" becomes a meaningful statement rather than a failure to detect a difference.

Core Claim

We make three claims. First, if axonal delays shape large-scale synchrony, a model's advantage from using delays should grow with oscillation frequency. The expected phase offset between regions $i$ and $j$ is $\Delta\phi_{ij} = 2\pi f L_{ij}/v$. Here $f$ is the band's center frequency, $L_{ij}$ is the tract length estimated from diffusion tensor imaging (DTI, which reconstructs fiber pathways from the direction of water diffusion), and $v$ is conduction velocity. Second, the zero-delay advantage reported at alpha should shrink or reverse once model and data share the same spatial leakage. Third, frequency and leakage matching together may still leave the delay models indistinguishable. In that case we should be able to show whether the data lack the reliability needed to separate the models, or whether topology genuinely dominates across bands.

This is not a new classifier. It is a falsifiable test of a specific published result, extended along the two axes that result left open, and evaluated under small-cohort constraints.

Relation to Prior Work

The motivating result is Biophysical network models of phase-synchronization in MEG resting-state. It coupled Wilson-Cowan oscillators (paired excitatory and inhibitory populations) over Human Connectome Project diffusion connectomes across 74 left-hemisphere regions. It fit alpha-band phase synchronization in a 23/23 train/test split and found that zero delays matched the data (ρ = 0.49, RMSE = 0.04) as well as distance-dependent delays did. We reuse its model family and its split-sample design. We differ from it in four ways: we test delta through gamma, we include both hemispheres and their callosal connections (the longest delays), we match leakage between model and data, and we add lag-sensitive metrics.

Two papers constrain the design. How reliable are MEG resting-state connectivity metrics? showed several things about MEG metrics. Uncorrected coherence and amplitude-envelope correlation reach group repeatability near 0.99 because they partly measure shared leakage. Leakage-corrected amplitude correlation stays near 0.95. Phase-lag metrics fall to about 0.90 or below and have poor single-subject reliability. This tells us that any model comparison must report a noise ceiling and must not rely on single-subject phase metrics. Harmoni showed that non-sinusoidal rhythms create spurious harmonic interactions. This matters because beta-band synchrony can be a harmonic of the arch-shaped mu rhythm (a sensorimotor rhythm near 10 Hz), which would make an apparent beta result really an alpha result.

A third line of work points in the opposite direction. Pairwise maximum entropy model explains the role of white matter structure in shaping emergent co-activation states found that intracranial interaction weights predicting DTI connections were nearly identical across theta through high-gamma in five epilepsy patients. We take this seriously: frequency-invariant, topology-only coupling is our null hypothesis. That analysis, however, binarized band power and so discarded phase, which is where delays would appear. The fMRI harmonic studies do not constrain this question, because hemodynamic smoothing removes millisecond timing. These are Distributed harmonic patterns of structure-function dependence orchestrate human consciousness and A Graph Signal Processing Perspective on Functional Brain Imaging.

Research Program

We will use HCP Young Adult participants with both resting-state MEG (three sessions) and diffusion MRI. The previous studies drew subsets of 46 and 61 subjects from this pool. We will build bilateral connectomes and tract-length matrices from DTI tractography. MEG will be source-reconstructed with a linearly constrained minimum-variance (LCMV) beamformer, a standard spatial filter for estimating source activity from sensor data. We will filter into delta, theta, alpha, beta and gamma bands. As a secondary analysis, we will also use individualized alpha and beta peaks, because canonical band boundaries are conventions that can split an individual's peak across bins.

For each band we will compute three empirical targets. The first is leakage-corrected amplitude-envelope correlation. The second is the weighted phase lag index, a phase-synchrony measure that ignores zero-lag coupling. The third, which delays specifically predict, is the matrix of mean phase lags between region pairs.

We will simulate four model variants, all with each node tuned to oscillate in the target band: zero delay, constant delay, distance-dependent delay, and distance-dependent delay with velocity $v$ fit per band. The key ingredient is the leakage-matched readout. Each simulated source time series is projected to the sensors through the subject's own lead field (the mapping from brain sources to sensor readings). Sensor noise is added at the empirical level. The data are then reconstructed with the identical beamformer, and the identical metric is computed. Global coupling and velocity are fit only on training subjects. All comparisons are reported on held-out subjects. Session-to-session agreement of the empirical matrices defines the noise ceiling.

For small-sample honesty, we will also estimate the minimum detectable difference in model fit as a function of the number of subjects, using subject-level bootstrap resampling.

Baselines

The naive reference point has two parts. The first is structure alone: the raw DTI connectivity matrix correlated with each band's empirical matrix. The second is a distance-only null: Euclidean inter-region distance predicting the same matrix. Distance-only is important because both leakage and synchrony fall off with distance. Any model that fails to beat it has not demonstrated anything about white matter. The delay model beats this reference only if it exceeds both, on held-out subjects, beyond the bootstrap interval.

The strongest existing method is the zero-delay Wilson-Cowan model from Biophysical network models of phase-synchronization in MEG resting-state. Its reported alpha result is ρ = 0.49 and RMSE = 0.04 against empirical phase synchronization. No reported result found on Lacuna for this model family in the beta or gamma band. We will first rerun it in their left-hemisphere alpha setting to confirm we recover roughly that number before extending it. It receives exactly the same subjects, parcellation, parameter-search grid and compute budget as the delay variants. The delay variants have one extra free parameter, velocity, so we will compare them by held-out fit rather than training fit. Beating the zero-delay model means a distance-delay advantage in held-out correlation that grows with frequency and is significant in at least beta and gamma. It must also appear in the phase-lag target, where the zero-delay model should do worst.

The ablation that isolates our key ingredient repeats every comparison with the leakage-matched readout switched off, comparing clean model sources directly to beamformed data. If the zero-delay advantage appears only without leakage matching, the original paradox is at least partly a measurement artifact.

Why not just…?

Why not just rerun the published model at other frequency bands? That is the first step, but by itself it cannot settle the question. If the zero-delay model's alpha advantage comes from shared zero-lag leakage, the same bias will favor it at every band, and simply rerunning would reproduce the paradox for the wrong reason. The leakage-matching ablation is what shows whether this happens.

Why not just regress empirical phase lags on tract length, with no model at all? This is a real threat. It is cheaper, and a clear lag–length slope rising with frequency would be strong evidence by itself, so we include it as a primary model-free test. Its limitation is that distant regions often communicate over multi-step paths, and leakage also depends on distance, so a slope cannot tell direct delay from indirect topology. The network model supplies the counterfactual, namely the same topology with delays removed. If the regression and the model disagree, we will report both and treat the regression slope as the more conservative claim.

Why not just use communication measures on the connectome, such as shortest-path length or search information, instead of simulating oscillators? These measures capture topology without any notion of oscillation frequency. They therefore predict the same coupling at every band, making them a clean embodiment of the frequency-invariance hypothesis suggested by the maximum entropy iEEG study. We fit them as an additional baseline. If they match the delay model at every band, the delay story fails.

Decisive Checks

The central test is the interaction between band and delay: held-out fit advantage of distance-delay over zero-delay, plotted against $f$. The theory predicts a monotone increase with the steepest gain in gamma, and a fitted velocity within a physiologically plausible range that is consistent across bands. A velocity that must change drastically from band to band would count against a single-mechanism delay account.

Three controls follow. First, beta results must survive Harmoni-based removal of alpha harmonics. Second, the interhemispheric subset of connections, which have the longest tracts, should show the largest delay advantage. Third, every effect is reported relative to the test–retest noise ceiling, together with the bootstrap-estimated number of subjects needed to detect it.

We commit in advance to three possible outcomes. If the advantage is flat across bands with leakage matched and the noise ceiling well above the model fits, topology dominates and the paradox is real. If it is flat but model fits already sit at the ceiling, the data cannot tell the models apart at this sample size. If the advantage rises with frequency, delays matter above alpha, and fiber length must enter models of fast MEG synchrony.