Frequency-Coupled Connectivity Learning

• Frequency-coupled connectivity learning is a framework that jointly models within-frequency (WFC) and cross-frequency (CFC) interactions to capture multi-scale neural coordination.
• It leverages graph-based methods and hybrid architectures—such as GATs, Transformers, and adaptive decomposers—to integrate local and global coupling features for improved task performance.
• Empirical results from EEG and fMRI studies demonstrate that combining WFC and CFC yields notable improvements in diagnostic accuracy and model robustness over single-band approaches.

Frequency-coupled connectivity learning denotes a family of methods in which connectivity is learned jointly from within-frequency coupling (WFC) and cross-frequency coupling (CFC), rather than from a monolithic signal or from frequency bands treated independently. In the cited literature, the term appears in graph-based EEG emotion recognition, resting-state fMRI brain disorder classification, EEG-based Alzheimer’s disease classification, and bio-inspired sparse neural architectures. Across these settings, the common premise is that neural communication is intrinsically multi-frequency: WFC captures synchronization within the same band, whereas CFC captures interactions across bands or frequencies, and their joint use yields a richer representation of neural coordination than either component alone (Wang et al., 29 Apr 2025, Xun et al., 6 Nov 2025, Klepl et al., 2022, Hays, 21 Feb 2026).

1. Conceptual scope and defining formulations

The central distinction in frequency-coupled connectivity learning is between connectivity estimated within a single oscillatory scale and connectivity estimated across scales. In the EEG literature, WFC is typically operationalized as same-band synchronization, whereas CFC is used for cross-band oscillatory interactions. In the fMRI literature, the same distinction is recast over learned or predefined frequency sub-bands of BOLD signals, with intra-band functional connections supplemented by explicit cross-band coupling. In bio-inspired dynamical systems, the phrase extends further: connectivity updates are driven by frequency coherence, so structural learning is coupled to oscillatory synchronization.

A recurring motivation is methodological rather than purely descriptive. Several papers argue that prior methods either use fixed/predefined frequency bands, restrict analysis to within-band interactions, or simply concatenate/fuse bands without explicitly modeling inter-band physiology. Frequency-coupled connectivity learning is introduced precisely to avoid those weaknesses and to produce connectivity representations that are both frequency-aware and interaction-aware (Xun et al., 6 Nov 2025, Klepl et al., 2022).

Paper	Signal/modality	Core coupling formulation
(Wang et al., 29 Apr 2025)	Multi-channel EEG	Joint WFC/CFC brain graphs with GAT + Transformer
(Xun et al., 6 Nov 2025)	Resting-state fMRI	Unified adjacency from intra-band and cross-band blocks
(Klepl et al., 2022)	EEG in Alzheimer’s disease	Cross-bispectrum networks for WFC and CFC
(Hays, 21 Feb 2026)	Sparse oscillatory architecture	Synchronization-gated Hebbian structural updates

A common misconception is that frequency-coupled learning is equivalent to stacking bandwise features. The cited fMRI work explicitly contrasts its unified graph construction with methods that merely concatenate or fuse band features, and the EEG works similarly treat CFC as an interaction structure to be learned, not as an auxiliary descriptor (Xun et al., 6 Nov 2025, Wang et al., 29 Apr 2025).

2. EEG formulations: jointly identifying within- and cross-frequency coupled brain networks

In EEG-based formulations, the typical pipeline begins by decomposing the signal into frequency bands and constructing connectivity graphs from coupling measures. In DB-GNN, EEG is filtered into five bands, $\delta$ (1–3 Hz), $\theta$ (3–8 Hz), $\alpha$ (8–12 Hz), $\beta$ (12–30 Hz), and $\gamma$ (30–48 Hz), written as

$$X' = \Psi_k(X_i), \quad k \in \{\delta,\theta,\alpha,\beta,\gamma\}.$$

WFC is measured by Phase Locking Value: $$\mathrm{PLV}_{ij}^k = \left| \frac{1}{N}\sum_{n=1}^{N} \exp\left(i\left(\phi_i(E_n)-\phi_j(E_n)\right)\right)\right| ,$$ and CFC is measured by Modulation Index in entropy/KL-divergence form,

$$\mathrm{MI} = \frac{\mathrm{KL}(U,X)}{\log N} = \frac{\log N - H(p)}{\log N}, \qquad H(p) = -\sum_{j=1}^{N} p(j)\log p(j).$$

These measures generate 5 WFC graphs and 10 CFC graphs, for 15 coupling graphs in total; adjacency matrices are thresholded so that the brain-network density stays around 20% (Wang et al., 29 Apr 2025).

DB-GNN then applies a dual-branch architecture. Its local perception branch runs Graph Attention Networks on each of the 15 graphs individually and aggregates their outputs by average pooling to obtain a local representation $H^{\text{local}} \in \mathbb{R}^{C \times d}$. This branch is intended to preserve fine-grained, localized coupling patterns. The global perception branch, the Prior information-based graph transformer module (PiGTM), is designed to capture holistic interactions across WFC and CFC networks. The global branch maps node features to query, key, and value tensors and computes attention with prior coupling guidance: $$E = \frac{QK^\top}{\sqrt{d_k}} \odot A + F - \mathrm{MLP}(M),$$ where $\theta$0 encodes coupling priors from PLV/MI and $\theta$1 is a learnable scalar matrix derived from coupling degrees. The Transformer output is

$\theta$2

followed by residual, LayerNorm, and feed-forward operations.

The paper’s defining claim is not simply that local and global branches are both useful, but that prior coupling information is integrated into Transformer inference to prevent overfitting on small EEG datasets. This is coupled with multi-scale graph contrastive learning: $\theta$3 where node-level positives are the same EEG node encoded by the global and local branches, and graph-level positives are the same brain-network sample across the two branches. The full objective is

$\theta$4

The resulting system learns local coupling specificity, global coupling coordination, and cross-branch agreement in a single end-to-end framework (Wang et al., 29 Apr 2025).

3. Adaptive frequency-coupled connectivity learning in fMRI

In Ada-FCN, frequency-coupled connectivity learning is the graph-construction module that turns adaptively decomposed fMRI signals into a single unified functional network containing both within-band functional connections among ROIs and cross-band coupling between different learned frequency sub-bands. The model first decomposes each ROI time series $\theta$5 through an Adaptive Cascade Decomposer. For decomposition level $\theta$6,

$\theta$7

with $\theta$8 implemented by a dilated 1D convolution plus LeakyReLU and $\theta$9 by a separate 1D convolution without activation. Stacking all sub-bands gives $\alpha$0 (Xun et al., 6 Nov 2025).

Connectivity learning then proceeds in two parts. First, for each decomposed band $\alpha$1, the model computes a Pearson correlation matrix

$\alpha$2

and applies band-specific dynamic thresholding: $\alpha$3 The $\alpha$4 sparsified intra-band matrices are combined via the Kronecker direct sum into a block-diagonal adjacency $\alpha$5.

Second, cross-band coupling is modeled by Dual-Projection Bilinear Attention. For two distinct bands $\alpha$6 and $\alpha$7,

$\alpha$8

followed by

$\alpha$9

These off-diagonal blocks form the global cross-band adjacency $\beta$0, regularized by a sparsity penalty

$\beta$1

The final unified adjacency is

$\beta$2

where $\beta$3 is a learnable scaling factor.

This unified graph is passed to Unified-GCN: $\beta$4 with node features constructed by stacking the per-band correlation matrices,

$\beta$5

The graph therefore contains ROI-specific nodes replicated across frequency bands, block-diagonal intra-band edges, and off-diagonal cross-band edges. The paper’s contribution is specifically to learn this as a single graph over frequency-expanded ROI nodes rather than to treat frequency bands as separate channels of evidence (Xun et al., 6 Nov 2025).

4. Nonlinear spectral and dynamical variants

A distinct but related formulation appears in EEG-based Alzheimer’s disease classification using cross-bispectrum. That work proposes a frequency-coupled functional connectivity framework in which conventional within-band measures are augmented by nonlinear cross-frequency measures. Five canonical EEG bands are used: $\beta$6 Hz, $\beta$7 Hz, $\beta$8 Hz, $\beta$9 Hz, and $\gamma$0 Hz. Cross-spectrum is used as a linear within-frequency baseline, with

$\gamma$1

while cross-bispectrum is used to capture nonlinear within-frequency and cross-frequency coupling: $\gamma$2 and

$\gamma$3

This yields 5 within-frequency networks on the diagonal and 20 cross-frequency networks off-diagonal, for 25 band-pair measures in total. Surrogate thresholding with 200 surrogate signals per channel pair and quantile-based sparsification are used to suppress spurious coupling (Klepl et al., 2022).

A more mechanistic extension appears in Hebbian-Oscillatory Co-Learning, where frequency-coupled connectivity learning is formulated as synchronization-gated structural plasticity in a sparse oscillatory architecture. The model couples Resonant Sparse Geometry Networks and Selective Synchronization Attention. Fast dynamics are Kuramoto-like: $\gamma$4 with Gaussian compatibility kernel

$\gamma$5

and global order parameter

$\gamma$6

Structural updates are gated by coherence. The hard-gated rule is

$\gamma$7

and the smooth version is

$\gamma$8

Sparse neighborhoods are defined in the Poincaré ball by

$\gamma$9

so coherence is computed over sparse neighborhoods whereas the global $$X' = \Psi_k(X_i), \quad k \in \{\delta,\theta,\alpha,\beta,\gamma\}.$$0 gates plasticity. In this formulation, frequency-coupled connectivity learning means that connectivity consolidation occurs only when sufficient phase coherence signals a meaningful computational pattern (Hays, 21 Feb 2026).

5. Empirical performance and ablation evidence

The empirical record is strongest in task-specific classification settings. On the SEED emotion recognition dataset, evaluated in a subject-dependent setting with 15 subjects, 3 emotion classes, and an 80% train / 20% test split, DB-GNN reports testing accuracy of $$X' = \Psi_k(X_i), \quad k \in \{\delta,\theta,\alpha,\beta,\gamma\}.$$1 and an F1-score of $$X' = \Psi_k(X_i), \quad k \in \{\delta,\theta,\alpha,\beta,\gamma\}.$$2, reaching the state-of-the-art performance. Ablation results indicate that removing the global branch hurts performance substantially, removing prior coupling from the Transformer reduces performance, and removing contrastive learning further degrades performance. The reported interpretation is that global + local learning, prior-guided attention, and multi-level contrastive learning are all critical to stable joint perception of WFC and CFC (Wang et al., 29 Apr 2025).

For resting-state fMRI brain disorder classification, Ada-FCN reports superior performance over listed baselines on both ADNI and ABIDE. The full model reaches 79.68% accuracy and 75.30% AUROC on ADNI, and 77.89% accuracy and 77.62% AUROC on ABIDE. The ablation study shows that removing dynamic thresholding hurts performance the most; removing sparsity loss and removing diversity loss also reduce accuracy. The reported scores are 79.68 ± 1.65 for the full model on ADNI versus 77.35 ± 1.22 without DT, 79.33 ± 1.29 without sparsity loss, and 76.14 ± 1.66 without diversity loss; on ABIDE, 77.89 ± 1.52 for the full model versus 76.53 ± 1.95 without DT, 76.26 ± 1.14 without sparsity loss, and 73.29 ± 1.61 without diversity loss. The number of decomposition levels is set to $$X' = \Psi_k(X_i), \quad k \in \{\delta,\theta,\alpha,\beta,\gamma\}.$$3, which gave the best performance on both datasets (Xun et al., 6 Nov 2025).

In EEG-based Alzheimer’s disease classification, the cross-bispectrum framework shows that fusing multiple frequency-coupling networks improves accuracy and that bispectrum-based fusion consistently outperforms spectrum-based fusion. The best reported result is 77.6% accuracy, 74.2% sensitivity, and 81.5% specificity for bispectrum at the top 20% edges; the top 10% bispectrum model reaches 77.0% accuracy, 81.4% sensitivity, and 74.0% specificity. By comparison, the best fused cross-spectrum model reports 73.0% accuracy. The study therefore argues that cross-frequency coupling itself carries discriminative information and that nonlinear frequency coupling reveals disease-related structure beyond simple within-band synchrony (Klepl et al., 2022).

Across these applications, a plausible common implication is that frequency-coupled learning tends to be most useful when classification depends on multiscale coordination rather than on a single band-specific pattern.

6. Interpretation, related findings, and open issues

Several broader lines of work help contextualize frequency-coupled connectivity learning. In oscillatory neural networks, cross-frequency coupling implemented through subharmonic injection locking is reported to increase memory capacity and enable error-free pattern retrieval, whereas retrieval fails without CFC when load increases. The model links theta–gamma coupling, discrete $$X' = \Psi_k(X_i), \quad k \in \{\delta,\theta,\alpha,\beta,\gamma\}.$$4-state associative memory, and Hebbian synaptic storage through a phasor neural network / phasor associative memory framework. This suggests a computational rationale for why cross-frequency interactions may matter not only as biomarkers but also as mechanisms for information storage and pattern completion (Bybee et al., 2022).

Related work on human learning shows that functional connectivity reorganizes differently from activity over time. Using wavelet transform coherence on fMRI data, one longitudinal study reports that activity reorganization is earlier and more local, whereas connectivity reorganization is later and more spatially distributed; connectivity also persists into rest as “sticky connectivity.” Although this work is not itself a formulation of frequency-coupled connectivity learning, it suggests that methods emphasizing network-level coordination are compatible with broader observations that learning and performance increasingly depend on distributed connectivity structure rather than only on regional activity (Bertolero et al., 2018).

Open issues are explicit in the dynamical literature. In HOC-L, the stability proof uses the smooth sigmoid gate rather than the hard indicator gate; extending the analysis to the discontinuous case would require nonsmooth tools such as Clarke subdifferentials or Filippov solutions. The use of a single global order parameter may be too coarse for heterogeneous architectures, and neighborhood-level gating is suggested as a more selective alternative. The Lyapunov proof guarantees dynamical stability but not task-level generalization. The decay term intentionally erases structure during long incoherent periods, which may be undesirable in some applications. Broader or multimodal frequency settings would require more advanced mean-field tools such as the Ott–Antonsen framework (Hays, 21 Feb 2026).

Another unresolved distinction concerns band specification. Some methods rely on canonical frequency bands, as in EEG emotion recognition and EEG Alzheimer’s disease classification, whereas Ada-FCN argues that predefined bands may not be optimal and instead learns task-relevant sub-bands for each brain region. This suggests that “frequency-coupled connectivity learning” is not a single algorithmic template but a design principle: learn connectivity from multiple oscillatory scales while explicitly modeling how those scales interact.