Moving the Edge of a Perceptual Moment

The brain may not see the world as a smooth movie. A long line of work suggests that vision chops time into short chunks, roughly one per cycle of the 10 Hz "alpha" brain rhythm. Two events that fall inside the same chunk are merged and their order is lost. Two events that straddle a chunk boundary are kept apart. The evidence so far is mostly correlational: the phase of the rhythm (where in its cycle it is) when a stimulus arrives predicts what is perceived. The causal evidence uses stimulation that overrides the brain's own rhythm instead of steering it. We propose a direct test of the strong version of the claim. A precisely timed magnetic pulse should shift a person's ongoing alpha rhythm forward or backward by an amount we can predict in advance. If chunk boundaries are set by that rhythm, the moment at which two flashes start being seen in the correct order should move by the same amount, measured in milliseconds.

Here is the idea in two sentences. Measure how each person's alpha rhythm responds to a pulse delivered at different points in its cycle, then deliver pulses in real time at the points that should advance, delay, or leave the rhythm unchanged. If perceptual chunks follow the brain's phase, the timing of order judgments should shift exactly as the phase-shift model predicts. If they follow the stimulator's clock, or the click and tap of the coil, they should not.

Core Claim

Our claim is that alpha phase defines the boundaries of visual temporal windows, which is a stronger and more testable statement than "alpha phase correlates with perception." The test is a pre-registered quantitative prediction. A transcranial magnetic stimulation pulse (TMS, a brief magnetic field from a coil on the scalp that induces current in cortex) delivered at pre-pulse phase $\phi_p$ shifts the ongoing rhythm by $\Delta\phi(\phi_p)$. That function is the individual's phase response curve (PRC), the map from "when in the cycle you push" to "how much the cycle moves." The behavioral boundary should then move by

$$\Delta t_{\text{boundary}}(\phi_p) = -\frac{\Delta\phi(\phi_p)}{2\pi f_i},$$

where $f_i$ is that participant's individual alpha frequency. The formula is the definition of phase converted into time at the person's own clock speed. Pulses at different phases produce the same coil click, scalp sensation, and electrical artifact, yet they predict different boundary shifts, including roughly zero shift at phases where the PRC crosses zero. The comparison of interest is therefore between physically identical interventions that differ only in their predicted effect on the internal rhythm. This is as close to the counterfactual rung of the causal ladder as a human experiment can reach.

How This Sits Against Prior Work

Three studies motivate the proposal and also set its bar. The Phase of Ongoing EEG Oscillations Predicts Visual Perception showed that pre-stimulus 6–10 Hz phase predicts flash detection, explaining at least 16% of variability. That result is purely observational and concerns detection rather than temporal parsing. The Phase of Ongoing Oscillations Mediates the Causal Relation between Brain Excitation and Visual Perception delivered TMS at random times and sorted trials by phase afterward. It found that phosphene perception (seeing a flash of light caused by the pulse itself) varies by about 15% between opposite alpha phases. That is an intervention on excitability, not on temporal windows, and the pulse was used as a probe rather than as a phase-mover. Occipital Alpha-TMS causally modulates Temporal Order Judgements comes closest. Its 5-pulse 10 Hz trains made temporal order judgment accuracy oscillate at 10 Hz (p = 0.00006 against a million surrogates). The single-stimulus report rate stayed at 5.4%, which argues against a pure visibility account. However, a forced fixed-frequency train cannot tell apart two possibilities. In one, perception follows the endogenous rhythm that the train entrained. In the other, perception follows the train itself. That study also did not tailor frequency to individuals. Our design differs on exactly these points. We use a single pulse that perturbs rather than replaces the endogenous rhythm, phase is targeted prospectively in real time, frequency is individualized, and the outcome is a predicted displacement rather than the presence of a rhythm.

Two papers supply the method. Non-invasive suppression of essential tremor via phase-locked disruption of its temporal coherence introduced the endpoint-corrected Hilbert transform (ecHT), a real-time phase estimator that cut phase error from about 47° to 9°. It used ecHT only for motor tremor, and we reuse it to time pulses to visual alpha. Response of Neuronal Populations to Phase-Locked Stimulation showed in Parkinsonian rats that coupled-oscillator models predict how phase-locked pulses change rhythm amplitude from the shape of the PRC (R = −0.84). We borrow that modeling stance for sensory cortex, where it has not been applied. TMS-induced phase resets depend on TMS intensity and EEG phase acts as a constraint. In motor cortex, high-intensity pulses near the peak fully invert the rhythm, and auditory sham did not reproduce the parametric resetting. This tells us intensity must be chosen to produce graded, phase-dependent shifts rather than a full reset, which would flatten the PRC and remove our contrast.

One paper appears to contradict the premise. An Entrained Rhythm's Frequency, Not Phase, Influences Temporal Sampling of Speech found that the rate of a preceding tone sequence, not the phase at target onset, shifted speech categorization. We take this seriously rather than set it aside. Our design separates the two accounts in vision. Phase-gating predicts a shift in boundary location that follows $\Delta\phi$. A clock-rate account predicts a change in window width that follows any post-pulse change in instantaneous frequency, with no systematic location shift. Both quantities are measured on every trial. That study also lacked any neural recording, so its null on phase may reflect failed entrainment. We verify phase in EEG directly.

Causality is Key for Interpretability Claims to Generalise supplies the lens. It found that 53.5% of interpretability claims exceed the causal rung of their evidence. We state up front that the design reaches strong Rung 2 and approaches Rung 3 only for one specific counterfactual: same physical pulse, different predicted phase consequence. It does not establish that alpha phase is the only pathway to temporal parsing.

Research Program

The work proceeds in three sessions per participant, about 24 healthy adults, with concurrent 64-channel EEG and occipital TMS. All procedures follow standard TMS safety screening.

Session one characterizes the person's rhythm. We first find the individual alpha peak frequency from resting posterior EEG. A spectral analysis then checks that the occipital rhythm is a genuine peak rather than a shoulder of theta or beta, because the standard 8–12 Hz band boundaries can put a fixed 10 Hz target at the edge of someone's alpha. We then deliver several hundred single pulses at random times, at an intensity just below phosphene threshold. Offline, with non-causal filtering that is more accurate than anything possible in real time, we estimate each pulse's pre-pulse phase and the resulting phase shift 150–300 ms later. Fitting these gives a smooth individual PRC. From it we choose four target phases per person: maximal advance, maximal delay, and the two zero crossings, which act as matched null pulses.

Sessions two and three are the test. An ecHT pipeline, with its bandpass centered on the individual alpha frequency, triggers pulses at the four target phases in random interleaved order. After each pulse, two Gabor patches appear in opposite hemifields, separated by a short onset asynchrony titrated to about 75% order-judgment accuracy. The pair onset is swept from 50 to 350 ms after the pulse in 10 ms steps. Participants report which patch came first and whether they saw one or two stimuli. For each pulse-phase condition, accuracy as a function of pair onset is fitted with a cosine at the individual frequency. The fitted behavioral phase gives the boundary location, and fitted width parameters give window width.

The primary analysis compares observed boundary displacements across the four conditions with the PRC-predicted displacements from session one. That prediction is fixed before sessions two and three. The trial-level analysis asks what order-judgment correctness tracks. One candidate is the EEG-measured alpha phase at pair onset, estimated on post-pulse data with artifact removal. The other is the time since the pulse. Under phase-gating, EEG phase should carry the variance and pulse latency should add nothing once phase is known. Under a pulse-driven transient or sensory artifact account, the reverse holds. Alpha power at pair onset is included as a covariate throughout, because power and phase are known to make separable contributions (12% versus at least 16% in Busch et al.).

Baselines

The naive reference point is no stimulation. Participants perform the identical order-judgment task with pairs presented at random times, and we bin trials by endogenous pre-stimulus alpha phase after the fact. This is the Rung 1 design of Busch et al., applied to temporal order instead of detection. It runs in a separate block with the same trial count as one active condition. The proposal beats it only if the active conditions show phase-predicted displacement that the observational block cannot, by construction, show. The block also tells us how large the purely correlational phase effect on order judgments is, which we expect to be smaller than the 16% detection swing that study reported.

The strongest existing method is open-loop rhythmic entrainment as in Chota et al.: 5-pulse trains at a fixed 10 Hz, with pair onset swept afterward. Its reported result is a significant 10 Hz periodicity in contralateral order-judgment accuracy (p = 0.00006 versus 1,000,000 surrogates) and a 5.4% single-stimulus report rate. No reported result found on Lacuna for the effect size (peak-to-trough accuracy modulation) of that periodicity, so we will compute it ourselves on our replication. We run this condition with the same participants, coil position, intensity, total pulse count matched by reducing trial number, and the same onset sweep and cosine fit. Beating it means two things. The single-pulse closed-loop conditions must produce displacements predicted from the individual PRC, which the train design cannot test because it fixes the phase. Behavioral periodicity must also be locked to the measured EEG phase better than to train timing, meaning a larger trial-level variance share for EEG phase.

Three ablations isolate the key ingredient. The first is random-time single pulses with post-hoc phase binning, the Dugué et al. design. It is matched for total trials and pulses and isolates the value of prospective closed-loop timing. The second runs the full closed-loop design with a fixed 10 Hz bandpass in place of the individual alpha frequency, which isolates individualization. The third is an auditory-click-plus-scalp-tap sham, delivered at the same four phases with the coil tilted away. It reproduces the peripheral sensations, isolates cortical from cross-modal resetting, and follows the sham logic of Thies et al..

Why not just…?

Why not just bin endogenous phase during the task and skip stimulation? That stays at Rung 1. Any phase–behavior link could reflect an unobserved thalamic driver of both, a concern the c-WWI work itself raises about pre-perceptual power changes (VanRullen et al.). We run this block as the naive baseline precisely to show what it cannot say. It cannot produce a prediction that differs between physically identical interventions.

Why not just use random-time pulses and sort by phase offline, as Dugué et al. did? This is the real threat, and we do not argue it away. Offline phase estimates are more accurate than real-time ones, and in principle enough random trials fill the same pulse-phase by pair-onset grid. The advantage of closed-loop delivery is trial efficiency. It concentrates trials at the four phases where the PRC makes divergent predictions instead of spreading them uniformly around the cycle, and it lets the onset sweep be dense within each condition. The matched-budget ablation settles the question. If post-hoc binning recovers the same displacements with comparable confidence intervals, we will report that the closed loop is a convenience, not a necessity. The scientific claim does not depend on the answer.

Why not just entrain with visual flicker or tone trains instead of TMS? Bosker et al. found no phase effect from tone precursors and no evidence that they entrained cortex. A visual flicker would also share the visual channel with the Gabor patches and contaminate the judgment itself. More fundamentally, a rhythmic stimulus imposes its own clock, the same confound we raise against train-TMS. The auditory-plus-tactile sham tests the cross-modal version directly at matched phases. If that sham reproduces the predicted shifts, the conclusion becomes "phase sets the boundary, and any resetting input moves it." That would still support phase-gating while ruling out a focal occipital claim.

Why not just model order judgments as continuous evidence accumulation? A drift-diffusion account predicts accuracy that rises smoothly with onset asynchrony and has no reason to depend on post-pulse onset in a phase-specific way. Chota et al.'s 10 Hz periodicity already argues against it. Our design sharpens the test, because the accumulation account predicts identical curves for advance, delay, and null pulses. We fit an accumulation model with a pulse-locked transient drift change as the competing model and compare it by cross-validated likelihood.

Decisive Checks

The central test is pre-registered. Across participants, observed boundary displacements should correlate positively with PRC-predicted displacements. Advance and delay pulses should differ in boundary location by the predicted amount, about 20–40 ms if PRC shifts of roughly a quarter cycle materialize. This difference should be significant while the two null-phase pulses do not differ from each other. Failure has a clear reading. If identical-artifact pulses at different phases yield the same behavioral shift, the effect is pulse-driven, not phase-driven, and the strong discrete-window claim is weakened. If location stays put but width tracks post-pulse frequency, the clock-rate account carried over from Bosker et al. wins in vision too.

Two manipulation checks gate interpretation. ecHT phase error at the trigger must fall near the 9° reported by Schreglmann et al., and we report our own error against offline estimates. Single-stimulus reports must stay low, comparable to the 5.4% of Chota et al., so that order effects cannot be explained by missed patches. A secondary analysis tests whether contralateral and ipsilateral boundaries move in anti-phase, as Chota et al.'s hemispheric push-pull implies. Because alpha travels across cortex as a wave (Lozano-Soldevilla & VanRullen), we also ask whether displacement scales with the phase lag between the stimulated site and each hemifield's cortical representation. A lumped single-phase model would miss this structure, and observing it would be the first sign that closed-loop control needs to treat cortex as a sheet rather than a point.