acceptodds
Under review as a conference paper at ICLR 2027

Support Is Free, Bandwidth Is Not: Basis-Parameterised Convolutions for Generative EEG.

Abstract

Functional convolutions write a kernel as coefficients over a fixed basis sampled on an -point grid, so their parameter count does not depend on the support . We show that the ratio buying that free support is also a band-limit: a Fourier-basis kernel cannot represent frequencies above , and Legendre and B-spline kernels follow the same scaling (slope 1.00–1.04, ) with basis-specific constants. We test what this costs on raw four-channel consumer EEG, with a flow-matching model whose velocity predictor is built from these kernels and held at a 64–128 Hz cutoff against a 40 Hz signal band. Across six checkpoints per model its samples carry 1.21±0.60 times real beta power (13–30 Hz) and 1.63±0.95 times real gamma (30–40 Hz), against 0.052 and 0.020 for a functional VAE on the same substrate, with no overlap in either band (). A discrete-kernel twin of the velocity predictor is not distinguishable from it in either band over three paired seeds, so this fidelity is not what basis parameterisation buys. What it buys is an analysable band-limit, and spending that headroom on longer support raises held-out loss monotonically at every basis budget while parameter count explains none of it (). A linear aggregate 0–40 Hz spectral error ranks the VAE better on the same twelve checkpoints and does not separate the two models (), because two bins below 0.5 Hz account for at least 94% of it. Over 30 seeds, a gaze regressor trained only on the flow model's samples at 5× real volume shows no confirmed degradation against one trained on real data, while the VAE's samples drive it to zero correlation. A 1 Hz high-pass shows that this task reads sub-1 Hz ocular signal. Support is free in parameters; it is paid for in bandwidth.

Primary area:
learning on time series and dynamical systems
Keywords:
functional convolutions, functional data analysis, basis expansion, band-limited kernels, spectral bias, flow matching, generative models for time series, EEG synthesis, evaluation of generative models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_fBmYOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates the application of basis-parameterized convolutions to EEG signal generation, focusing on the relationships between the convolutional kernel's receptive field, parameter count, and frequency bandwidth. The authors developed an EEG generation model named FuncFlow based on Flow Matching and compared it with FuncVAE and traditional discrete convolution models.

Critical strengths
  1. The paper conducts a systematic study on the trade-off between parameter efficiency and frequency representation capability in basis-parameterized convolutions; the theoretical analysis and experimental design offer valuable insights.
Critical weaknesses
  1. The research scope is narrow, and the level of innovation is limited; the work primarily combines existing basis-parameterized convolutions with the Flow Matching framework rather than proposing a novel convolution operator or generative learning algorithm.
  2. The experimental validation is also limited in scope, covering only the EEG modality.
  3. The model fails to demonstrate a clear performance advantage in practice (experiments show no significant difference in high-frequency signal fidelity between FuncFlow and traditional discrete convolution models).
  4. The paper is also somewhat difficult to follow; for instance, there is no dedicated section for related work, no main figure illustrating the proposed method, and the core contributions are not prominently highlighted.
Questions for authors

see above

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

This paper investigates the trade-off between parameter efficiency and spectral expressivity in basis-parameterised convolutions for generative EEG modeling. The authors show that although the parameter count is independent of kernel support length, the ratio between support length and basis count imposes an effective bandwidth constraint. They introduce FuncFlow, a flow-matching generative model built from functional convolutions, and evaluate its ability to preserve the spectral characteristics of raw EEG. The experimental results show that FuncFlow preserves substantially more beta- and gamma-band power than functional VAE baselines, while achieving comparable high-frequency fidelity to a discrete-convolution counterpart. Through analytical derivations, basis-projection experiments, and controlled support-length sweeps, the authors demonstrate that increasing kernel support at a fixed basis budget reduces spectral expressivity and degrades held-out flow-matching performance. Overall, the work provides empirical evidence and practical guidance for balancing receptive-field size, parameter efficiency, and spectral representation capacity.

Critical strengths
  1. A meaningful characterization of the trade-off between parameter efficiency and spectral expressivity. The paper provides a clear analytical characterization of how the support-to-basis ratio constrains the effective bandwidth of basis-parameterised convolutions. The projection experiments across different basis families and the controlled support-length sweeps offer complementary evidence for this relationship. This provides useful design guidance beyond the specific EEG generation setting.
  2. Comprehensive evaluation with informative controls and frequency-aware analysis. The paper evaluates the proposed approach from multiple perspectives, including spectral fidelity, discrete-convolution comparisons, parameter-efficiency experiments, and downstream gaze regression. In particular, the discrete-kernel baseline helps distinguish improvements associated with the generative framework from those attributable to basis parameterisation. The analysis also demonstrates why aggregate spectral errors can be misleading when low-frequency components dominate the evaluation.
Critical weaknesses
  1. Unclear practical implications of the theoretical bandwidth constraint Although the paper establishes a theoretical bandwidth constraint on basis-parameterised convolutional kernels, its implications for EEG generation and downstream decoding remain insufficiently demonstrated. In particular, the generated spectra do not directly follow the predicted bandwidth limitation, and the paper does not provide direct evidence that reduced kernel bandwidth necessarily leads to the loss of task-relevant EEG information or degraded decoding performance. Therefore, the practical significance of the proposed bandwidth-based design rule remains unclear.
  2. Limited practical advantages over conventional convolutions The practical advantages of basis parameterisation over conventional convolutions remain insufficiently established. The discrete-convolution baseline achieves comparable high-frequency fidelity, while the canonical FuncFlow model provides only a modest reduction in parameter count. Furthermore, increasing kernel support does not improve held-out performance and introduces additional bandwidth constraints without reducing the computational complexity of direct convolution. These findings raise questions about when and why basis parameterisation should be preferred over conventional convolutional architectures.
  3. Limited generalizability and insufficient validation on EEG-specific tasks The experimental validation is limited to a single four-channel consumer EEG dataset and one downstream gaze-regression task. Moreover, the downstream task primarily relies on sub-1 Hz ocular signals, providing limited evidence that the generated EEG preserves task-relevant neural information. Although the theoretical analysis may apply to broader classes of signals, the effectiveness of the proposed approach on high-density EEG, larger datasets, and diverse decoding tasks remains unverified. This limits the demonstrated generalizability and practical relevance of the findings to broader EEG applications.
Questions for authors
  1. Relationship between theoretical bandwidth constraints and downstream performance The paper demonstrates that the theoretical kernel bandwidth constraint does not directly translate into reduced high-frequency power in generated EEG. Under what conditions does this constraint become a practical bottleneck for EEG generation or decoding? Can the authors provide evidence that reducing kernel bandwidth leads to the loss of task-relevant information or degraded downstream performance, rather than merely affecting velocity-field approximation and training loss?
  2. Practical benefits compared with conventional convolutions Given that the discrete-convolution baseline achieves comparable spectral fidelity and that increasing kernel support does not improve held-out performance, under what conditions does basis parameterisation offer meaningful advantages over conventional convolutions? Can the authors demonstrate benefits in terms of parameter efficiency, computational cost, generation quality, or long-range temporal modeling under comparable experimental settings?
  3. Generalizability and scalability to broader EEG applications To what extent do the observed bandwidth–performance trade-offs generalize to other EEG datasets, high-density recordings, and downstream tasks such as motor imagery, ERP classification, or sleep staging? In particular, can the proposed architecture effectively model spatial dependencies in multi-channel EEG and scale to larger datasets and longer recordings while maintaining its claimed benefits?
Reviewer_CmDBOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This is a purely computational paper on how a particular type of neural network layer is built. They chose to use EEG as the test case largely for its convenience. However, given that the network layer was designed specific for EEG, and little in the paper addresses generation of neural signals, I find framing it as "for generative EEG" in the title a bit contrived. I also multiple aspects of EEG analysis and validation problematic.

Critical strengths

The EEG data is a good choice for testing the method.

Critical weaknesses

The EEG framing in the title feels contrived. The high-frequency activity being preserved is mostly not brain activity--dry electrodes over the forehead and temples would be contaminated by gamma-range muscle artifact. The gaze being predicted by activity below 1Hz, which could be mostly eye-movement potentials, not neural signals.

Questions for authors

Can you examine gaze-decoding performance and signal amplitude per channel? If the signal is ocular, it should concentrate at the frontal sites, with opposite polarity left vs right during horizontal pursuit. Do the synthetic data reproduce this pattern?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.