Inter-trial phase coherence (ITPC) asks a simple question of brain recordings: when the same stimulus is presented many times, does the brain's rhythm at a given frequency land at the same point in its cycle each time? It ignores how strong the rhythm is and scores only how consistently its timing lines up. That makes it a natural feature for decoding models that must work on people they were never trained on, because the timing of a response to a stimulus is plausibly more shared across brains than its raw amplitude. The catch is that ITPC is never zero, even for pure noise. Its floor depends on how many trials are averaged. We argue that it also depends on how many of those trials are actually independent. That number differs by frequency band and by experimental design. Adjacent trials in a block, or slow rhythms that span several trials, share phase for reasons unrelated to the stimulus. The result is inflated coherence that looks like signal and can carry over into cross-subject decoders.
We propose to replace the nominal trial count with an effective trial count, estimated separately for each band from the pre-stimulus data of each recording. We then use it to put ITPC from different bands, subjects, and designs on a common calibrated scale. The proposal makes two testable claims. First, this calibration should remove the spurious 8–14 Hz coherence advantage that block designs show over randomized designs. Second, once the floor is removed, the question left open in the literature can be answered cleanly: do multi-band phase features transfer to unseen subjects better than spectral-power features when only a few subjects and trials are available? The deliverable is an estimator, a re-analysis of existing auditory and visual datasets, and cross-subject learning curves that separate true phase-locking from bookkeeping artifacts.
Core Claim
For $N$ trials with unit phase vectors $z_k = e^{i\theta_k(f,t)}$ at frequency $f$ and latency $t$, write $R = \frac{1}{N}\sum_k z_k$ for the trial-averaged phase vector. ITPC is $|R|$. With no stimulus locking, averaging over all pairs of trials gives
$$ \mathbb{E}\,|R|^2 \;=\; \frac{1}{N}\Big(1 + 2\sum_{\ell=1}^{N-1}\big(1-\tfrac{\ell}{N}\big)\,\mathrm{Re}\,\rho_\ell(f)\Big) \;\equiv\; \frac{1}{N_{\mathrm{eff}}(f)}, $$
where $\rho_\ell(f) = \mathbb{E}[z_j \bar z_{j+\ell}]$ is the phase correlation between trials $\ell$ positions apart in presentation order. The equation follows directly from expanding the squared magnitude of a mean.
The familiar $1/N$ floor that analytic bias corrections subtract is the special case $\rho_\ell = 0$. The claim is that $\rho_\ell$ is not zero in practice. Three things push it away from zero: narrowband filters that ring across short inter-trial intervals, regular stimulus timing, and slow state drift within blocks. The effect is strongest in low bands and in block designs. We estimate $\rho_\ell(f)$ from pre-stimulus phase at matched latencies, where no stimulus locking can exist. The calibrated feature is ITPC$^2$ expressed relative to the band-specific floor $1/N_{\mathrm{eff}}(f)$.
This is what differs from existing corrections. The standard fixes subtract the independent-trial floor or shuffle trial labels, and both assume exchangeable trials. The file documents a case where that assumption demonstrably fails: Overestimated performance of auditory attention decoding caused by experimental design in EEG recordings reports higher 8–14 Hz ITPC at Cz and C3 under block than under non-block design.
Relation to Prior Work
Three papers motivate the proposal by showing that phase coherence is informative but fragile. Subject-Independent Classification of Japanese Spoken Sentences by Multiple Frequency Bands Phase Pattern of EEG Response During Speech Perception computes cross-trial phase coherence to speech envelopes in four bands for 10 listeners. It reaches 44.0% subject-independent accuracy against 33.3% chance by combining bands. However, it does not ask how much of that phase structure is floor, and trial counts differ across subjects and bands. The block-versus-non-block study ( Overestimated performance of auditory attention decoding caused by experimental design in EEG recordings, duplicate record; Overestimated performance of auditory attention decoding caused by experimental design in EEG recordings; Overestimated performance of auditory attention decoding caused by experimental design in EEG recordings) shows that the design itself inflates both ITPC and deep-decoder accuracy. STANet falls from about 92% to about 29%. That paper diagnoses the problem but proposes no correction for existing block-design data. The Relationship between Trial-by-Trial Variability and Oscillations of Cortical Population Activity separates ITPC from alpha-beta power suppression in 23 MEG subjects. It finds that power, not phase, explains 86% of individual differences in variability quenching. We use its dataset as the case where calibration should leave a genuine phase-versus-power dissociation intact.
Two papers constrain the design. Linear stimulus reconstruction works on the KU Leuven audiovisual, gaze-controlled auditory attention decoding dataset shows that a shrinkage-regularized linear backward model stays above 80% under leave-one-subject-out evaluation. The block-design study shows the same model is stable across designs. Any phase-feature decoder must therefore at least match a linear model, not just beat chance. Multivariate Phase Synchrony and Auditory Attention Decoding for Speech and Music reports the opposite ranking, with deep models at about 80% against about 65% for linear models in 35 listeners. That contradiction is part of what our learning curves are meant to resolve.
The method borrows in spirit from Uncovering the organization of neural circuits with Generalized Phase Locking Analysis. That work replaces permutation testing of phase-locking with an analytic null for spike-field coupling. We likewise derive the null analytically, but we target human cross-trial coherence and serially dependent trials.
Finally, the power-feature pipelines define the comparison class. GNN4EEG uses band-wise differential entropy, which measures band power on a log scale, and collapses from intra-subject to cross-subject evaluation. CodeBrain and Brant encode waveform shape and band power but not cross-trial phase. We do not propose a new foundation model. The question we answer comes first: whether calibrated phase is worth adding to such models at all.
Research Program
The first stage is estimation and validation in simulation. We generate synthetic EEG with a known phase-locked component in chosen bands, embedded in noise whose rhythms follow a realistic 1/f spectrum. We then vary three things: trial count from 10 to 200, inter-trial interval, and block structure, meaning slow drift in phase or state shared within blocks. We check that the pre-stimulus estimate of $\rho_\ell(f)$ recovers the true $N_{\mathrm{eff}}$. We also check that calibrated ITPC has a correct false-positive rate under the null where uncorrected, analytically debiased, and label-shuffled ITPC do not. We record the failure mode where it matters: dependence that exists after the stimulus but not before it, such as a stimulus-induced state change. In that case the baseline estimate is biased, and we quantify by how much.
The second stage is the confound test on real data. On the block and non-block scalp and ear-EEG recordings from the block-design study, we compute band-resolved ITPC four ways: raw, analytically debiased, label-shuffle normalized, and design-calibrated. The prediction is that only the design-calibrated version removes the block-design excess at 8–14 Hz while preserving coherence in bands where the stimulus truly drives phase. On the MEG visual dataset, calibration should leave the published phase-power dissociation unchanged, which guards against overcorrection.
The third stage is cross-subject learning curves. On the Japanese sentence data, the KU Leuven audiovisual attention data, and the music-and-speech attention data (subject to access), we train decoders on $n_{\mathrm{sub}}$ subjects and $n_{\mathrm{trial}}$ trials per subject, subsampled over a grid. We test on held-out subjects. Every feature family gets the same decoder families: a regularized linear model and a small nonlinear model. The features are band-wise power or differential entropy, raw multi-band ITPC, and calibrated multi-band ITPC. The object of study is the shape of the cross-subject accuracy curve: its intercept at few subjects, its slope, and the gap between within-subject and cross-subject performance. We also report which bands carry the transferable signal after calibration.
Baselines
The naive reference point is chance: 33.3% on the three-sentence task, and the corresponding class-balanced chance for each attention task. Alongside chance we run raw multi-band ITPC fed to the same decoders. Raw ITPC is what the Japanese sentence study effectively used, and it shows how much of the reported phase signal survives calibration. Calibrated features count as meaningful only if they beat chance with confidence intervals from leave-one-subject-out resampling. Calibration counts as "working" if it removes spurious gains without destroying real ones. Higher accuracy alone is not the criterion.
The strongest existing competitor on attention data is regularized linear stimulus reconstruction. This backward model maps multichannel EEG to the attended speech envelope with shrinkage. Linear stimulus reconstruction works on the KU Leuven audiovisual, gaze-controlled auditory attention decoding dataset reports "peaking at over 80% accuracy" under leave-one-subject-out evaluation with 120 s windows, and 85–86% subject-specific accuracy with 60 s windows. The block-design study reports 80–90% for 1 s windows under both designs. We rerun it with the same subject splits, window lengths, and the same budget for inner cross-validation of regularization as the phase decoders. Calibrated-ITPC decoders beat it only if they match or exceed its leave-one-subject-out accuracy at small $n_{\mathrm{sub}}$, or reach its accuracy with fewer subjects. If they do neither, we report that phase features are dominated on this task.
On the sentence task, the strongest reported method is multi-band template matching from Subject-Independent Classification of Japanese Spoken Sentences by Multiple Frequency Bands Phase Pattern of EEG Response During Speech Perception, at 44.0% subject-independent accuracy. The subject-dependent SVM reaches 55.2%. We reimplement template matching and the SVM on identical folds.
For power features, the reference is band-wise differential entropy with per-subject z-scoring, following the feature construction of GNN4EEG. That paper reports 31.3% for RGNN on 9-class cross-subject FACED. FACED lacks repeated trials of the same stimulus within a subject, so we use that figure only to illustrate the power-feature cross-subject gap, not as a head-to-head target.
We also run STANet on the block and non-block data, with a reported ~92% versus ~29%, to confirm that we can reproduce the inflation. No reported result was found on Lacuna for any calibrated or debiased ITPC decoder on these datasets.
Two ablations isolate the key ingredient. The first is the textbook analytic debiasing, which is our estimator with $\rho_\ell$ forced to zero. The second uses a single $\rho_\ell$ pooled across bands. The first tests whether trial dependence matters at all. The second tests whether the band-specificity matters.
Why not just…?
Why not just subtract the standard $1/N$ bias or use a pairwise debiased estimator? This is the most serious threat, and it is our first ablation. If trials are close to independent, the textbook correction and ours coincide, and the proposal reduces to a benchmarking study. The block-versus-non-block ITPC difference reported in Overestimated performance of auditory attention decoding caused by experimental design in EEG recordings is exactly the situation in which independence fails. Stage two settles the matter directly: if analytic debiasing already erases the block-design excess, we will say so and drop the dependence modeling.
Why not just shuffle trial labels to build a permutation null? Shuffling preserves the marginal phase distribution but destroys the ordering of trials. It therefore rebuilds the independent-trial floor and misses the shared phase between adjacent trials that inflates block designs. We include it as a baseline in stage two. It should behave like the analytic correction, not like ours.
Why not just use linear stimulus reconstruction and skip phase features? On continuous-speech attention tasks this may simply be the right answer. The linear model's robustness across designs and unseen subjects is well documented in Linear stimulus reconstruction works on the KU Leuven audiovisual, gaze-controlled auditory attention decoding dataset Linear stimulus reconstruction works on the KU Leuven audiovisual, gaze-controlled auditory attention decoding dataset and Overestimated performance of auditory attention decoding caused by experimental design in EEG recordings Overestimated performance of auditory attention decoding caused by experimental design in EEG recordings. However, linear reconstruction needs a continuous stimulus to regress against. It does not apply to event-related paradigms such as the visual MEG data, and it does not say which bands carry the transferable signal. We treat it as the bar to clear on attention data, and we report plainly if it is not cleared.
Why not just collect randomized, non-block data going forward? New studies should do this. It does nothing for the many existing block-design datasets, or for pooled corpora that foundation models are trained on, where design cannot be changed after the fact. Calibration is a correction for data that already exists.
Decisive Checks
The proposal succeeds or fails on three observations. First, in simulation, design-calibrated ITPC must hold nominal false-positive rates across trial counts and block structures where the alternatives do not. Second, on the block and non-block recordings, the 8–14 Hz block excess must shrink to within noise after calibration but not after analytic debiasing. At the same time, the MEG phase-power dissociation must survive. Third, on the cross-subject learning curves, calibrated multi-band ITPC must show a smaller within-to-cross-subject gap than per-subject normalized power features at small subject counts. If the second check fails, dependence is not the problem and we fall back to reporting clean learning curves. If the third fails while the second succeeds, we will have shown that much of phase coherence's apparent transfer advantage was floor. That is still a result the field needs before building phase features into larger models.
References
- Overestimated performance of auditory attention decoding caused by experimental design in EEG recordings
- Subject-Independent Classification of Japanese Spoken Sentences by Multiple Frequency Bands Phase Pattern of EEG Response During Speech Perception
- Overestimated performance of auditory attention decoding caused by experimental design in EEG recordings
- The Relationship between Trial-by-Trial Variability and Oscillations of Cortical Population Activity
- Linear stimulus reconstruction works on the KU Leuven audiovisual, gaze-controlled auditory attention decoding dataset
- Multivariate Phase Synchrony and Auditory Attention Decoding for Speech and Music
- Uncovering the organization of neural circuits with Generalized Phase Locking Analysis
- GNN4EEG: A Benchmark and Toolkit for Electroencephalography Classification with Graph Neural Network
- CodeBrain: Bridging Decoupled Tokenizer and Multi-Scale Architecture for EEG Foundation Model
- Brant: Foundation Model for Intracranial Neural Signal
- Linear stimulus reconstruction works on the KU Leuven audiovisual, gaze-controlled auditory attention decoding dataset