acceptodds
Under review as a conference paper at ICLR 2027

High-Resolution Dynamic Functional Connectivity Generation with Graph-Variate Flow Matching

Abstract

High-resolution dynamic functional connectivity (DFC) is attractive for studying rapidly evolving brain-network interactions, but it is difficult to estimate and generate reliably because short temporal windows produce noisy and often low-rank covariance estimates. Graph-Variate Dynamic (GVD) connectivity addresses this problem by modulating fast instantaneous interactions with a stable trial-level support, suppressing spurious fluctuations while emphasizing persistent and informative connections. We further show that this Hadamard construction lifts low-rank instantaneous connectivity from the positive-semidefinite to the positive-definite cone, enabling high-resolution connectivity trajectories to remain on the SPD manifold without additive ridge regularisation or post-hoc projection. Building on this structure, we introduce GVD-CFM, a class-conditional generative model designed specifically for high-resolution dynamic connectivity generation. Each trial is represented as a sequence of SPD GVD matrices on a product Riemannian manifold and mapped through a global log-Euclidean diffeomorphism and an invertible temporal DCT basis. A Transformer-based conditional flow models all spectral modes jointly, allowing the complete high-resolution trajectory to be generated non-autoregressively in Euclidean coordinates while preserving exact correspondence with valid SPD connectivity sequences. Because the full DCT basis is retained, the learned representation also defines a temporal basis expansion that can be decoded on denser temporal grids without retraining.Across multiple EEG motor-imagery datasets, GVD-CFM achieves the strongest overall performance across held-out distributional fidelity, preservation of temporal dynamics, and synthetic-to-real classification, while remaining computationally efficient relative to strong raw-signal and direct GVD-space generative baselines. These results establish GVD-CFM as a framework for generating realistic, temporally coherent, high-resolution brain-network trajectories while preserving manifold structure and supporting resolution-flexible decoding from a single trained model.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG, Flow Matching, Dynamic Functional Connectivity

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_J6QBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces GVD CFM for generating high resolution EEG dynamic connectivity trajectories. Key experiments are that it shows that graph variate matrices remain PSD when local covariances are rank deficient; transformer based flow model enabled generating complete connectivity trajectories; and the approach is evaluated with five motor imagery EEG datasets compared to several generative baselines.

Critical strengths

The paper’s aim of modelling complete trajectories in spectral log Euclidean coordinates and to produce more realistic DFC has good motivation. The empirical evidence for spectral representation is sound and the evaluation across datasets and baselines are broad. The GVD baselines share the representation and decoding framework which isolates the effect of the generative model.

Critical weaknesses
  1. It appears strong classification can be achieved with static support (W) alone (Appendix F2; AUC .822). This may suggest that discriminative information is largely in the static connectivity, with modest additional power from the dynamics. More importantly, this contribution is not isolated in the generated data. Could authors quantify or discuss the additional value of temporal dynamics, e.g., through comparing W only and full trajectory representations in the same synthetic and real classification?

  2. Some baselines (e.g., JET and EEGGAN) perform at chance level in the classification and show poor Eva F1. The configurations are described in the Appendix A8, but the hyperparameter tuning and model selection are unmentioned. Could the authors clarify how baselines were tuned and if the configurations were adapted to the evaluated datasets?

  3. Writing: Some acronyms (e.g., CFM, DCT) were not properly defined at the first mention. The differences between resolution flexible decoding and the recovery of higher frequency dynamics could be clearer when the approach is first introduced; generally the introduction and motivation would benefit from a clearer explanation of practical significance; likewise, methods section would benefit from improved clarity.

Questions for authors

How is B=100 and B=25 chosen for the main experiments and for flexible decoding, respectively? How sensitive is the approach to these choices in generative quality?

3 comments
AuthorsOpenReview ↗

Rebuttal to reviewer

We thank the reviewer for their constructive comments which lead to changes in the manuscript. All new changes are in blue in the revised paper.

W1 (Static support versus dynamics). We agree with the reviewer that the static support carries a large part of the class information. The submission already separates the two on real data in the appendix (App. F.3, Table 11; CAS AUC, dataset-balanced over five datasets):

RepresentationB=8B=8B=32B=32B=128B=128
Stable support with real dynamics (GVD)0.8410.8400.837
Support only (static)0.8220.8220.822
Dynamics only (no support)0.8160.7650.756
Support with Gaussian dynamics0.5120.5270.574

We can see that neither part is completely sufficient on its own. The static support alone reaches 0.822, the dynamics without the support degrade as the windows shorten, and a support modulated by Gaussian dynamics falls to chance. It is the synergy of the stable support with the real dynamics that gives the best performance at every resolution. Following the review we repeated the static-versus-full comparison on generated data (App. F.4, Table 11). Static structure alone gives 0.802 AUC on real data and 0.765 when the classifier is trained on generated data, against 0.856 and 0.795 for the full trajectory, so the generated data preserve this ordering. The generated dynamics are also validated by the temporal metrics of Table 2, which remove each trial's temporal mean (temporal agreement 0.910, dynamic-energy ratio 0.953).

W2. (Baseline tuning and selection). We thank the reviewer for raising this; the settings are now given in full in App A.9, and no test data were used for any selection. JET and EEGGAN use the author implementations with their release settings, adapted only to the channel count and trial length of each dataset. To test whether tuning changes their results, we also tuned both again where two learning rates per model were trained for 100 pilot epochs on 85% of the training trials, the configuration was selected on the remaining 15% by a discrepancy over band powers, channel scale and correlations, and the selected model was retrained for 1,000 epochs. The tuned CAS AUC is 0.526 for EEGGAN and 0.496 for JET, against 0.524 and 0.495 in Table 1, so tuning does not change the result.

The submission already included GVD-cVAE and GVD-DDPM trained on the GVD targets. We have now added three further time-series generators trained on the same targets with fixed settings on every dataset (Tables 1 and 2), so the comparison no longer rests on the raw-EEG baselines:

ModelEvaGeM F1 ↑\uparrowCAS AUC ↑\uparrowTemp. corr. ↑\uparrowDyn. energy →1\to 1Rel. GVD-FID ↓\downarrow
GVD-CFM0.6130.7900.9100.9531.022
Diffusion-TS0.0070.7740.8450.4780.771
TimeGAN0.2190.6490.2600.4151.130
TimeVAE0.0010.6420.7670.1880.863

GVD-CFM has the best overall performance, with the highest EvaGeM F1 on all five datasets and the highest CAS AUC on four. Diffusion-TS obtains a lower FID, but generates about half of the real dynamic energy.

W3. (Writing). We have addressed the terminology and wording issues raised by the reviewer in the paper. CFM and DCT are now defined at first use. The introduction states that "high resolution" refers to temporal resolution, and that decoding on a finer grid interpolates the learned finite-bandwidth trajectory and does not recover faster dynamics. It also names the practical use case, training data for connectivity-based decoders.

Q. (Choice of BB) B=100B = 100 is the training resolution. B=25B = 25 is used only in the decoding experiment, where a coarse grid allows upsampling by up to 16 times to be tested. To answer the question on sensitivity we ran an additional sweep, training the full model at four values of BB on all five datasets (App. D.6, Table 8):

Training BBCAS AUC ↑\uparrowEvaGeM F1 ↑\uparrowRel. GVD-FID ↓\downarrow
250.7940.5471.008
500.8050.6090.980
1000.8030.6200.998
2000.7780.6181.037

We can see that the method is not sensitive between B=50B = 50 and B=100B = 100; thus B=100B = 100 is a good compromise point and not necessarily a uniquely optimal one.

We hope this addresses the reviewer's concerns, we are happy to discuss further and that they will consider updating their score.

AuthorsOpenReview ↗

Response to Reviewer

We thank the reviewer again for their review. Since we are approaching the end of the discussion period, we ask if there is any feedback for us or any updates to the score you would like to make based on our rebuttal. Thank you.

Thank you very much for your responses. I appreciate the additions of baseline comparisons and sensitivity analysis. Additional discriminative value of temporal dynamics is still uncertain. All things considered (including concerns on clarity), I would like to maintain my original ratings. Many thanks to authors' for their efforts and time.

Reviewer_VKwUOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a generative model for dynamic EEG functional connectivity. First, the authors compute a stable whole-trial correlation matrix and short-window covariance matrices. They combine them through a Hadamard product to obtain SPD connectivity matrices, even when the short-window covariances are rank deficient. They then map these matrices into Euclidean coordinates using the matrix logarithm and apply a DCT across time to represent the connectivity trajectory in temporal frequency modes. A flow-matching model is trained to predict the distribution of these DCT coefficients. During generation, the coefficients are transformed back using the inverse DCT and matrix exponential, producing synthetic SPD connectivity trajectories. The authors evaluate the generated data in terms of distributional fidelity, temporal dynamics and downstream classification performance. They also show that the DCT representation allows decoding trajectories at finer temporal grids.

Critical strengths

S1. The Hadamard product result for GVD windows is elegant and practically important.

S2. Non-autoregressive, class-conditional generation of complete SPD trajectories with resolution-flexible decoding is a compelling design with potential beyond EEG.

S3. The paper is well-structured and mathematically careful.

Critical weaknesses

W1. Although the method guarantees that each generated matrix is SPD, to my understanding this does not necessarily mean that the complete sequence corresponds to a realizable GVD trajectory. In particular, the GVD construction imposes cross-window consistency constraints, such as ∑b=1B∣Ib∣TΔb=W⊙W.\sum_{b=1}^{B}\frac{|I_b|}{T}\Delta_b = W\odot W., which are not explicitly enforced during generation.

W2. To my understanding, the GVD representation depends on instantaneous signal amplitudes, with its diagonal entries corresponding directly to local channel power. Consequently, temporal changes in GVD matrices may reflect amplitude fluctuations rather than changes in the underlying inter-channel coupling. Although the paper evaluates temporal fidelity, it is unclear how much of the reported performance is attributable to genuine connectivity dynamics.

Questions for authors

Q1. Could the authors quantify whether generated trajectories satisfy the constraints raised in weakness 1 and assess whether violations affect distributional fidelity or downstream performance?

Q2. Could the authors include a power-only or diagonal-only baseline, or evaluate whether off-diagonal GVD dynamics provide additional information after controlling for channel-power variations?

Q3. The authors mention fMRI as a potential extension. Given the substantially lower temporal resolution and hemodynamic smoothing of fMRI BOLD signals, would the DCT representation retain the same advantages in temporal decorrelation? Furthermore, since the number of ROIs can exceed the available time points, how would the method handle a rank-deficient whole-trial support matrix?

5 comments
AuthorsOpenReview ↗

Rebuttal to reviewer

We thank the reviewer for their careful reading and for their precise technical points. They are correct and are now addressed in the paper. All new changes are in blue.

W1 / Q1 (Cross-window consistency). We agree with the reviewer. The constraint is now stated as Proposition 3 (App. B.3). With wbw_b the fraction of trial samples in window bb,

∑b=1Bwb Δb=W⊙W,\sum_{b=1}^{B} w_b \, \Delta_b = W \odot W ,

so the weighted temporal mean MM of a real trajectory is entrywise nonnegative with unit diagonal. We audited both conditions on BNCI2014_001 with 1,296 generated trials (Table 4):

Real, held outGeneratedGenerated, corrected
Mean diagonal error of MM1.0×10−71.0 \times 10^{-7}0.0681.1×10−161.1 \times 10^{-16}
Trials with a negative mean edge0%0.077%0.077%
CAS AUC0.7950.797
Rel. GVD-FID1.0101.010
EvaGeM F10.8620.857
Temporal-correlation agreement0.9140.914

We can see that generated trajectories violate the unit-diagonal condition by about 7%, and nonnegativity in one trial of 1,296. We then give an exact correction: each window is replaced by SΔbSS \Delta_b S, where SS is diagonal with entries 1/Mii1/\sqrt{M_{ii}}. This preserves positive definiteness and the power-normalized matrix of every window, and needs no real data. After the correction no metric changes by more than 0.007, so the violation does not affect fidelity or downstream performance.

We agree that full realisability of a GVD sequence is stronger than this; it requires a correlation matrix WW with W⊙W=MW \odot W = M and Δb⊘W⪰0\Delta_b \oslash W \succeq 0 for every window, which GVD-CFM does not enforce. Following the review we therefore developed an optional variant, Exact GVD-CFM, which satisfies it by construction (App. B.3, Table 4). Essentially, it generates one standardized signal U=TLQHU = \sqrt{T} L Q H per trial with W=LL⊤≻0W = LL^\top \succ 0, and computes every window from it, so the constraint holds identically and every window remains SPD. Across 25,251 generated trials on five datasets and three seeds we find no violations (maximum error 2.4×10−152.4 \times 10^{-15}). We can see that exact realisability comes at some cost in fidelity, with a CAS AUC of 0.747 and EvaGeM F1 of 0.423 against 0.790 and 0.613 for GVD-CFM. We therefore retain GVD-CFM as the main model and provide Exact GVD-CFM for cases where realisability is required.

W2 / Q2. (Power versus coupling). We thank the reviewer for this suggestion and ran the power-only baseline together with a control for channel power (App. F.4, Table 11). We use Rb=Db−1/2ΔbDb−1/2R_b = D_b^{-1/2} \Delta_b D_b^{-1/2}, with DbD_b the diagonal of Δb\Delta_b, which removes the direct scaling of each entry by channel power.

RepresentationReal →\to real AUCSynthetic →\to real AUC
Power only (diagonal)0.5570.526
Power-normalized off-diagonal trajectory0.6980.643
Static power-normalized structure0.8020.765
Full GVD trajectory0.8560.795

We can see that channel power alone is close to chance, so the results are not explained by amplitude. The off-diagonal structure remains informative once power is removed, on both real and generated data. After removing the temporal mean from the power-normalized trajectories, real and generated data still agree in lag-ACF (0.995) and dynamic energy (ratio 0.974). In source space, a power-normalized precentral-postcentral interaction contrast is also reproduced after the cue (r=0.550r = 0.550, FDR q=0.0158q = 0.0158; App. E.1) in BNCI2014_001.

Q3. (fMRI). Both points are interesting and are now discussed in App. B.3. The DCT chart remains invertible for any trajectory length, but its decorrelation of BOLD signals is untested and grounds for future work. When WW is rank deficient, Proposition 2 does not apply as stated. A shrinkage support Wγ=(1−γ)W+γIW_\gamma = (1-\gamma) W + \gamma I with 0<γ≤10 < \gamma \le 1 could be used to restore positive definiteness,

λmin⁡(Wγ⊙Jb) ≥ γ min⁡iqi >0,\lambda_{\min}(W_\gamma \odot J_b) \ \ge\ \gamma \, \min_i q_i \ > 0,

where qiq_i is the window power of channel ii. This is an explicit regularization and gives up the ridge-free property, so the empirical claims of the paper are now restricted to EEG.

We also note that three time-series generators trained in the GVD space were added to Tables 1 and 2. GVD-CFM has the best overall performance, with the highest EvaGeM F1 on all five datasets (0.613 against at most 0.219) and the highest CAS AUC on four.

We hope this addresses the reviewer's concerns and that they will consider updating their score if we have suitably answered their questions.

Thank you for the detailed response and the new experiments. Proposition 3, the power controls and exact GVD-CFM are useful additions, but some concerns remain, especially regarding W2/Q2.

Concretely, the power-only baseline is helpful, but App. F.4 (l. 3064-3066) reports that adding normalized dynamics to power + static lowers AUC from 0.658 to 0.639. Unless I am misreading it, this answers Q2 negatively, yet the text concludes the opposite and the rebuttal does not mention it. The MLP numbers are also hard to interpret (power + static scores below static alone, 0.658 vs. 0.802); a regularized linear model with paired CIs over all five datasets would help. The RbR_b lag-ACF and energy results in the rebuttal are not in the revised PDF and should be added, together with temporal-correlation agreement. Power normalization also does not address field spread (zero-lag sensor correlations change with source power even when coupling is fixed), and the new ERD/ERS evidence is itself a power measure. As it stands, the claims should be limited to realistic GVD trajectories rather than coupling dynamics beyond channel power.

AuthorsOpenReview ↗

Response to Reviewer

We sincerely thank the reviewer for their careful attention to our paper. Your feedback has improved it significantly, and we are happy that the new experiments were useful.

Regarding App. F.4, the missing explanation was that the 0.658 to 0.639 comparison used separately extracted features concatenated before classification, not the full GVD representation. Specifically, these were power + static versus power + static + normalised dynamics. We have updated the manuscript to make this explicit and discuss the result further. Our original discussion focused on normalised off-diagonal trajectories outperforming power alone and the synergy of the GVD matrix being the highest performing one, but we should also have clearly explained these concatenation controls and their negative result.

We have thus repeated a regularised linear model with paired CIs over all five datasets as suggested. The confidence intervals use 10,000 paired participant-bootstrap resamples. This also includes the concatenated results for full clarity, the main results are as follows with the entire comparison in another comment and also will be in the updated manuscript.

RepresentationReal → realSynthetic → real
Power only0.576 [0.559, 0.596]0.553 [0.528, 0.576]
Normalized off-diagonal trajectory0.757 [0.722, 0.790]0.713 [0.680, 0.746]
Static normalized structure0.787 [0.753, 0.822]0.769 [0.738, 0.804]
Full GVD trajectory0.794 [0.768, 0.825]0.781 [0.750, 0.814]
Power + static0.709 [0.682, 0.739]0.701 [0.669, 0.730]
Power + static + normalized dynamics0.685 [0.656, 0.714]0.638 [0.613, 0.663]

The normalised off-diagonal trajectory performs significantly better than power alone in both settings, and full GVD has the highest mean AUC. Concatenation of normalized dynamics in this case does not help performance. Full GVD combines these components differently through its multiplicative construction and matrix-log representation. Our results show that this joint representation is useful, while simply concatenating the separately extracted components does not provide the same benefit and can even worsen performance.

The motivation for GVD follows GVSA and FAST. GVSA uses stable connectivity information to weight instantaneous signal interactions (Smith et al., 2019). FAST develops a related approach using a cohort-level support and shows its usefulness for studying transient EEG differences in simulations and Alzheimer’s disease datasets (Roy et al., 2024). Our work builds on this idea by generating complete GVD trajectories. A static representation can be useful for classification, but it cannot describe how those trajectories change over time, which is what our generator aims to reproduce.

We also agree with your concern about field spread. We therefore ran controlled Gaussian-source simulations with eight sources mixed into 22 sensors, using 48 fixed mixing matrices across three seeds. We varied source powers while keeping source correlation fixed, and separately varied source correlation while keeping source powers fixed. The table below shows the noiseless population results, with confidence intervals across mixing matrices.

Source manipulationObservationNormalised-GVD temporal RMS [95% CI]
Powers and source correlations remain constantFixed mixing into sensors0
Powers vary; source correlation remains fixed at zero or 0.3Unmixed sources0
Powers vary; source correlation remains fixed at zeroFixed mixing into sensors0.0279 [0.0267, 0.0291]
Powers vary; source correlation remains fixed at 0.3Fixed mixing into sensors0.0326 [0.0312, 0.0341]
Source correlation varies; source powers remain fixedFixed mixing into sensors0.0314 [0.0291, 0.0336]

(The RMS values are small because they measure fluctuations in normalised edge weights averaged across all sensor pairs)

Without mixing, changing source powers alone produces no normalised-GVD dynamics. After mixing, it does, even when the sources remain uncorrelated. Changing source correlation at fixed source power also produces dynamics. Stationary population controls have zero temporal variation. We found the same qualitative pattern with observation noise and in 7,680 sampled trials using the paper’s 512-sample, 100-window setup.

These controls support your point that normalising sensor power does not remove field spread. GVD dynamics can reflect both source-power changes and source-correlation changes, so we agree that they should not automatically be interpreted as changes in physiological coupling independent of power. We will make this clear in the paper and describe ERD/ERS agreement specifically as evidence of power-modulation fidelity.

Reviewer_6EexOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The author proposes GVD-CFM, a generative model based on conditional flow matching, for synthesizing high-resolution dynamic functional connectivity (DFC) trajectories from EEG data. The paper combines flow matching, DCT, and log-Euclidean mapping together, which technically makes sense. But I think this manuscript has some issues at both the methodological and neurophysiological levels. Firstly, the author's mathematical assumption of solving rank deficiency is equivalent to adding a 'strong' static constraint to dynamic connections, which conflicts with the word 'dynamic' itself. Secondly, the laboratory design is not very reasonable.

Critical strengths

As the connectivity trajectories are generated in the DCT frequency domain, the model can directly decode the results to any precision time grid without retraining (Sec. 6.5). This is a very practical feature.

Critical weaknesses
  1. My main concern is: Based on Proposition 2 (Schur product rank boosting): using a stable global experimental correlation matrix $W $ to modulate the instantaneous covariance JbJ_b of rank deficit, i.e., W⊙JbW \odot J_b. Mathematically, this can indeed guarantee SPD. However, the problem is that after multiplying a static global matrix element by element, any connections that are not obvious in the experimental average WW will have their instantaneous changes suppressed. In my understanding, the essence of dynamic functional connectivity often lies in transient network states that are short-lived and therefore washed out by trial averaging. The proposed operation will weaken or even erase them. In current manuscript, the author has not provided sufficient evidence to show that the generated trajectory reflects real brain dynamics.

  2. The choice of comparison method is a bit unreasonable. JET, EEGGAN, cVAE, DDPM are trained to optimize the reconstruction or generation of raw time series waveforms, not to focus on optimizing the fidelity of downstream derived GVD manifolds. Punishing the original EEG model based on low EvaGeM or GVD-FID scores is unfair, as they have never been optimized for SPD manifold geometry. To show superiority in DFC generation, the author should compare it with state-of-the-art dynamic SPD-based or timing graph generation models specifically designed for connection matrices.

  3. The log-Euclidean mapping used in the model requires $d(d+1)/2 $ coordinates. This method is feasible under 22 channels, but it can lead to severe dimensionality explosion when dealing with modern high-density EEG (such as 256 channels, single-window features increasing to 32896 dimensions). In this context, the computational cost of running the joint spectrum transformer on multiple windows will be unbearable. But this work claims "high resolution", this architecture inherently lacks spatial scalability.

  4. Sec. 6.4 only compares the average GVD amplitude of macroscopic brain regions. This static statistical matching cannot prove the physiological significance of the model in "high-resolution temporal dynamics" (i.e. the core contribution of this article). Lack of validation of deep indicators such as source space localization or task induced ERD/ERS time accuracy, claims that it can generate "realistic brain networks" are a little bit overstated.

Questions for authors
  1. Why do you use log-Euclidean? Have you tested the Affine-Invariant Riemannian Metric (AIRM), which has better robustness to physiological noise and scaled covariance across subjects.
4 comments
AuthorsOpenReview ↗

Rebuttal to reviewer

We thank the reviewer for their detailed comments. The review led to four new experiments, three new baselines and greatly improved the quality of the paper. All new changes are in blue in the updated manuscript.

W1. (Static support and transients) We agree with the reviewer on this point and ran a separate experiment to quantify it (App. F.5, Table 12). Detection AUC for 0.5 s covariance transients:

EdgeGVDPearson window
Supported (trial-level connectivity present)0.8650.694
Weak (no trial-level connectivity)0.6610.849

We can see that GVD improves detection on supported edges, and that its performance drops on unsupported edges. We hypothesise that the supported edges are the important ones. This is the premise of the GVSA and FAST line of work (Smith et al., 2017, 2019; Roy et al., 2024): edges with weak long-term connectivity are dominated by noise in short windows, so weighting by the long-term support retains the reliable dynamics. Our real-data sweep supports this (Table 10 in the paper); as the windows shorten, the unsupported window covariance falls from 0.816 to 0.690 AUC while GVD remains at 0.83 to 0.84. GVD is therefore a complement to window correlation and not a replacement, and Section 4.1 and the limitations now state that transients on weak edges are under-represented.

W2. ( New baselines) We thank the reviewer for this suggestion. The submission already included two generators trained directly on the GVD targets, GVD-cVAE and GVD-DDPM. Following the review we added three further time-series generators, trained on the same log-SPD GVD coordinates with the same decoder and evaluator (Tables 1 and 2):

ModelEvaGeM F1 ↑\uparrowCAS AUC ↑\uparrowTemp. corr. ↑\uparrowDyn. energy →1\to 1Rel. GVD-FID ↓\downarrow
GVD-CFM0.6130.7900.9100.9531.022
Diffusion-TS0.0070.7740.8450.4780.771
TimeGAN0.2190.6490.2600.4151.130
TimeVAE0.0010.6420.7670.1880.863

GVD-CFM has the best overall performance; it obtains the highest EvaGeM F1 on all five datasets, the highest CAS AUC on four, and the dynamic energy closest to real data. Diffusion-TS obtains a lower FID and a higher CAS AUC on Shin2017A (0.711 against 0.672). However, we can see that a low FID can come with a collapse of other important metrics. Diffusion-TS and TimeVAE have an EvaGeM F1 close to zero and generate 0.478 and 0.188 of the real dynamic energy. FID alone is therefore not the best measure of quality for connectivity trajectories, and Section 6.1 now states this.

W3 (Scalability) "High resolution" refers to temporal resolution, which the introduction now states. Each of the BB temporal modes is one token, so attention costs O(B2h)O(B^2 h) and does not depend on the number of channels. We measured a full training update on one H100 at B=100B = 100 and batch size 32 (App. F.2): 43.7 ms and 0.54 GiB at 22 channels, and 44.7 ms and 4.16 GiB at 256 channels. Our approach is therefore computationally valid up to 256 channels. We also note that additional EEG channels show diminishing returns in the literature, source localization plateaus beyond about 64 channels (Lantz et al., 2003; Sohrabpour et al., 2015), and connectivity estimation gains are very small from 128 to 256 (Allouch et al., 2022).

W4. (Physiology) We agree that static regional averages are not sufficient, and added an ERD/ERS timing analysis in source space on BNCI2014_001 (Sec. 6.4, App. E.1, Figure 9). We trained separate mu (8-13 Hz) and beta (13-30 Hz) models on source-domain GVD trajectories. Sources were estimated with sLORETA on the fsaverage template head model (three-layer BEM, standard electrode positions) and averaged into ten parcels: bilateral precentral, postcentral, paracentral, superior frontal and pericalcarine cortex. This is an approximate reconstruction, since it uses template anatomy and 22 electrodes without individual MRI. For mu-band left-hand imagery, the generated and held-out post-cue ERD/ERS curves correlate at r=0.888r = 0.888 in right precentral and r=0.805r = 0.805 in right postcentral cortex (FDR q=0.0063q = 0.0063), which is the contralateral sensorimotor site. Agreement is weaker in the beta band. We therefore consider this as regional plausibility and not localization accuracy, and have removed "realistic brain networks" from the abstract.

Q1 (AIRM) The log-Euclidean metric is the pullback of a Euclidean metric through the matrix logarithm. This is what reduces Riemannian flow matching to Euclidean flow matching with SPD outputs (Proposition 1). This makes the log euclidean metric appealing. The AIRM can also be considered but because AIRM is curved, each step then requires matrix square roots and logarithms, which the log-Euclidean metric avoids.

We hope this addresses the reviewer's concerns and that they will consider updating their score.

AuthorsOpenReview ↗

Q1 follow-up on AIRM

We thank the reviewer for suggesting an empirical AIRM comparison. For completeness we spent some time developing an AIRM variant of GVD-CFM. We trained an AIRM conditional flow-matching baseline on BNCI2014_001 using the same GVD targets and Transformer architecture, with three seeds and the same number of epochs.

Each trial contains 100 windows and 253 symmetric coordinates per window. We obtain the SPD prior by decoding standard Gaussian samples from the log-svec/DCT chart.

For each source-target SPD window pair, suppressing the window index for clarity, define

H=log⁡(X0−1/2X1X0−1/2).H=\log\left(X_0^{-1/2}X_1X_0^{-1/2}\right).

The exact affine-invariant geodesic is

Xτ=X01/2exp⁡(τH)X01/2.X_\tau=X_0^{1/2}\exp\left(\tau H\right)X_0^{1/2}.

Its analytic target velocity is

Uτ=dXτdτ=X01/2exp⁡(τH)HX01/2.U_\tau=\frac{dX_\tau}{d\tau}=X_0^{1/2}\exp\left(\tau H\right)H X_0^{1/2}.

The network predicts a symmetric whitened tangent for each window from the current complete log-svec/DCT trajectory, flow time and class. The corresponding matrix-space velocity is

Vθ,b=Xτ,b1/2Sθ,bXτ,b1/2.V_{\theta,b}=X_{\tau,b}^{1/2}S_{\theta,b}X_{\tau,b}^{1/2}.

Under AIRM, the squared norm of a tangent vector is

∥U∥X,AI2=∥X−1/2UX−1/2∥F2.\Vert U\Vert_{X,\mathrm{AI}}^2=\Vert X^{-1/2}UX^{-1/2}\Vert_F^2.

We therefore define the whitened target velocity as

Sτ,b⋆=Xτ,b−1/2Uτ,bXτ,b−1/2.S_{\tau,b}^{\star}=X_{\tau,b}^{-1/2}U_{\tau,b}X_{\tau,b}^{-1/2}.

The intrinsic AIRM flow-matching loss becomes

LAI=E[1Bm∑b=1B∥Sθ,b−Sτ,b⋆∥F2].\mathcal{L}_{\mathrm{AI}}=\mathbb{E}\left[\frac{1}{Bm}\sum_{b=1}^{B}\Vert S_{\theta,b}-S_{\tau,b}^{\star}\Vert_F^2\right].

This equals the AIRM squared error between the predicted and target matrix-space velocities. The complete orthonormal DCT and Frobenius-isometric svec allow the same loss to be computed in spectral output coordinates.

Sampling uses the exponential-map update

Xk+1,b=Xk,b1/2exp⁡(hSθ,b)Xk,b1/2.X_{k+1,b}=X_{k,b}^{1/2}\exp\left(hS_{\theta,b}\right)X_{k,b}^{1/2}.

This update preserves positive definiteness. We use 200 sampling steps with a step size of 1/200, requiring 200 network evaluations per trajectory. This matches the evaluation count of the paper's 50-step RK4 sampler.

The GVD-CFM column reproduces the BNCI2014_001 results already reported in Tables 16 and 18. Entries are mean ± standard deviation over three generator seeds.

MetricGVD-CFM in paperAIRM
CAS AUC ↑0.789 ± 0.0150.7894 ± 0.0094
Temporal correlation ↑0.920 ± 0.0060.9362 ± 0.0069
Lag-ACF agreement ↑0.999 ± 0.0000.9985 ± 0.0003
Dynamic-energy ratio, ideal 10.995 ± 0.0230.8716 ± 0.0201
EvaGeM F1 ↑0.864 ± 0.0270.4383 ± 0.0764
Relative GVD-FID ↓1.006 ± 0.0140.9605 ± 0.0047

AIRM achieves essentially the same mean classification AUC, higher temporal correlation and lower relative GVD-FID. The reported GVD-CFM better preserves dynamic energy and achieves substantially higher EvaGeM F1.

We hope the reviewer will find this interesting and we have included it in the appendix and mentioned it as potential for future work.

I would like to thank the authors for their detailed rebuttal. However, my fundamental concern (W1) that the static support matrix suppresses true instantaneous dynamics remains a critical limitation. Since this conceptual issue is inherent to the method and cannot be fully resolved, I will maintain my current score. Thanks again.

We thank the reviewer for considering our response. We agree that support weighting can attenuate genuine transients on weakly supported edges. However, we respectfully believe that calling this a fundamental conceptual problem overlooks both the motivation for this representation and also the existing published evidence that it can improve detection of transient activity in cases where other approaches fail.

We conjecture that a static support does not imply static dynamics. The support determines how strongly each edge contributes, while the instantaneous interactions determine its temporal evolution. This distinction is central to Smith et al. (2019), Graph-Variate Signal Analysis. That study evaluated the approach beyond its mathematical construction where GVD detected a single correlated pair among 8,128 possible edges, improved tracking of a moving simulated spatial pattern relative to the tested graph-signal approaches, and more robustly distinguished resting-state from task-related EEG activity than the tested conventional dynamic-connectivity measures. These results provide empirical support for using stable dependencies to analyse changing signals.

More directly, Filter Average Short-Term (FAST) demonstrates that fixed support can retain useful transient information. Roy et al. (2024) used a connectivity filter fixed across participants, yet detected differences between simulated noisy ERPs where the comparison methods failed. In two Alzheimer’s-related MCI cohorts, the method also identified reproducible task-specific differences in the P300 interval. Importantly, its Methods section explicitly acknowledges the same limitation raised here, potentially meaningful short-term interactions between weakly connected regions may be lost. The motivation is therefore an explicit trade-off, supported by simulations and task-related EEG results, rather than an assumption that all transient interactions survive weighting.

There is also a reason to avoid treating every short-window fluctuation as ground truth. for example Leonardi and Van De Ville (2015) showed analytically that sliding-window connectivity can exhibit fluctuations even when the underlying signal relationship is fixed. Although their application concerns fMRI, it illustrates the broader estimation problem where greater apparent temporal variability does not necessarily mean more faithful recovery of underlying dynamics. This does not prove that GVD is always preferable, but it explains why unrestricted transient sensitivity is not sufficient on its own.

Our additional experiment directly quantifies the resulting trade-off:

Transient locationGVD detection AUCWindowed Pearson AUC
Supported edges0.8650.694
Weakly supported edges0.6610.849

We report both outcomes because we do not assume that weakly supported edges are unimportant. Nevertheless, the improvement on supported edges shows that weighting can improve transient detection in one regime while reducing sensitivity in another.

Our contribution is to learn and generate complete sensor-level GVD trajectories, with their distributional and temporal properties, within this already established representation. We do not claim exhaustive recovery of all EEG transients or identification of physiological coupling independent of power. We therefore agree that weak-edge attenuation limits what GVD represents, but respectfully disagree that its being inherent makes the generation framework conceptually invalid. The relevant question is whether the representation offers a useful, clearly stated trade-off which the existing literature and our additional controls provide evidence that it does.

Regardless, we thank the reviewer again for their engagement during this review process.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.