A mathematical perspective on edge-centric brain functional connectivity

Functional connectivity (FC) captures statistical dependence between BOLD signals. While static FC (computed over entire scans) is widely used and linked to cognition and pathology, time-varying FC typically uses dynamic models or sliding windows, with concerns about temporal blurring and statistical artefacts. Framewise approaches (e.g., CAPs and point-process analyses) and more recent edge-centric methods examine instantaneous cofluctuations between all regional pairs, revealing that a small subset of frames with high root-sum-of-squares (RSS) cofluctuations can approximate static nFC and contribute disproportionately to network modularity. However, prior work has cautioned that apparent dynamics can arise from sampling variability or simple null models. This paper addresses the central question: to what extent do current edge-centric measures provide information beyond static nFC? The authors develop a mathematical framework showing that many hallmark edge-centric findings arise under a static, temporally i.i.d. Gaussian null that preserves spatial correlations (the observed nFC) but ignores temporal order. They test these predictions on HCP fMRI data to assess how much of edge-centric structure is explained by nFC alone.

The paper reviews: (a) established static FC methods and associations with cognitive and clinical factors; (b) time-varying FC via dynamic models and sliding windows and associated controversies about extracting structure from noise; (c) single-frame methods including CAPs and point-process analyses showing that few high-activity frames reproduce seed-based FC; (d) edge-centric approaches that unwrap FC into edge time series and define edge FC (eFC), reporting replicability and overlapping communities; (e) prior warnings that minimal null models can reproduce dynamic FC findings; (f) observations that high-amplitude RSS events appear in temporally uncorrelated synthetic data when spatial correlations are preserved, raising the question of what edge-centric metrics add beyond nFC.

Analytic framework and null model: Define z-scored BOLD time series zi(t) for N parcels across T frames. Edge time series for parcels i and j: cij(t)=zi(t)zj(t). Edge functional connectivity (eFC) is the normalized inner product between edge time series. The static null hypothesis: i.i.d. multivariate Gaussian samples Z(t) ~ N(0,R), where R is the empirical nFC (node-node correlation); temporal correlations are ignored but spatial correlations preserved. Under Gaussianity, fourth-order moments factor via Isserlis' theorem, and ergodicity links ensemble expectations to sample estimates.

Key derivations:

• eFC from nFC: The expected product of edge pairs yields E[Cij Ckl] = rij rkl + rik rjl + ril rjk. Normalizing leads to an analytic eFC entry: eFCjk,lm = (rjk rlm + rjl rkm + rjm rkl) / (sqrt(1+2 rjk^2) sqrt(1+2 rlm^2)). Thus the eFC is fully determined by nFC.
• Edge and node community structure: Define an L1 distance between eFC rows to approximate clustering proximity. Summing edge distances from two nodes to common targets yields a node distance proportional to (1 − rij)^2, implying edge-cluster similarity between nodes is predictable from nFC without computing eFC or performing k-means.
• RSS equivalence: The RSS of edge time series at frame t equals the squared Euclidean norm of the BOLD vector: RSS(t)=||Z(t)||^2 (exact when all i≤j edges included). This reframes RSS as overall BOLD amplitude.
• Eigen-decomposition and alignment: With R=UΛU^T and whitening W(t)=R^{-1/2}Z(t), RSS(t) ≈ (1/√2) Σ_i λi ||W(t)||^2 cos^2 θi(t). Large RSS arises when Z(t) aligns with leading eigenvectors (especially u1), explaining high similarity of high-RSS frames to nFC and their high modularity.
• RSS null distribution: Under the null, RSS is a generalized chi-square (sum of independent Gamma components weighted by nFC eigenvalues). The MGF M_RSS(s)=∏_i (1 − √2 λi s)^{-1} shows subexponential tails, predicting sporadic large events.
• Binary edge null: Thresholded edges Cjk(t)=1 if zi and zj share sign follow Bernoulli(pjk) with pjk = 1/2 + (1/π) arcsin(rjk). Time-averaging the binary edges recovers nFC due to ergodicity and E[Bernoulli]=p.
• CAPs linkage: Conditioning on a seed k at value zk gives E[Z(t)|Zk=zk]=zk Rk (k-th nFC column). Correlation between the BOLD frame and the seed’s nFC column is proportional to |zk|, predicting CAPs’ core observation that high seed-activity frames resemble seed-based FC.

Empirical evaluation: HCP 100 unrelated subjects (minimally preprocessed, ICA-FIX denoised), Schaefer-200 parcellation, with and without global signal regression (GSR). For each subject, compute empirical eFC, edge communities (k-means), RSS statistics, eigenvector alignments, binary edge averages, and CAPs-like analyses. Compare empirical quantities to analytic predictions from nFC and to simulations drawn from the static Gaussian null. Statistical tests included Pearson correlations between predicted and empirical matrices, Kolmogorov–Smirnov tests for RSS distribution convergence, and similarity analyses across percentiles of RSS or seed activity.

• Analytic eFC from nFC: The closed-form eFC expression (Eq. 1) closely matches empirical eFC. Across 100 HCP subjects, Pearson r between predicted and empirical eFC: mean r=0.93 (without GSR) and r=0.88 (with GSR).
• Edge communities from nFC: Clustering the predicted eFC reproduces empirical edge community labels exactly for 84% of 19,900 edges (74% with GSR). At node level, predicted edge-cluster similarity correlates strongly with empirical: r=0.96 (0.95 with GSR). Using nFC alone (without constructing eFC), predicted node similarity achieves r=0.76 with empirical edge-cluster similarity.
• High-RSS frames explain nFC: A small fraction of frames with largest RSS reproduces most nFC variance and network modularity. Ordering frames by RSS, top-percentile averages exhibit high similarity to full nFC in both empirical data and null simulations; matching of event timing is unnecessary.
• RSS distribution shaped by nFC eigenvalues: Under the null, RSS is a sum of Gamma components weighted by R’s eigenvalues. The largest eigenvalues govern the heavy tail, explaining prevalent large events. Empirically, the null distribution approximates the empirical RSS increasingly well with more frames; using all 1200 frames, the null could not be rejected for 58% of participants at 5% significance and for 90% after Bonferroni correction (two-sided KS test).
• Spatial patterns of high activity: High-RSS frames align with leading nFC eigenvectors. The leading eigenvector captures the dominant spatial mode (default-mode/control vs sensorimotor/attention anticorrelation). As more high-RSS frames are averaged, additional principal components contribute, accounting for near-complete nFC reconstruction from few frames.
• Binary edge series: Averaging binarized edge time series yields an almost perfect approximation to nFC (mean r=0.98). Analytically, P(Cjk=1)=1/2 + (1/π) arcsin(rjk), explaining the strong correspondence and constraints imposed by nFC.
• CAPs linkage: Frames selected by high seed activity correlate strongly with the seed’s nFC column; averaging top seed-activity frames rapidly approaches the seed-based FC, reproducing foundational CAPs results within the null framework.
• Modular influence: Larger functional modules enable larger nFC eigenvalues, which increase the magnitude and frequency of large RSS events; disrupting modular structure reduces RSS peaks, aligning with reported empirical findings.

The study shows mathematically and empirically that many hallmark edge-centric metrics—eFC structure, overlapping communities, dominance of high-RSS frames, spatial patterns during large cofluctuations, and properties of binarized edges—are largely determined by static nFC and can be reproduced under a temporally i.i.d. Gaussian null model that preserves spatial correlations. This addresses the research question by demonstrating that current edge-centric findings do not, by themselves, necessitate temporally structured neural dynamics and may reflect static second-order statistics. The results underscore the need to focus on dynamic measures that truly exploit temporal ordering (e.g., cross-subject synchrony during movies, peak-to-peak intervals, dwell times and transitions in state models) which cannot be replicated by static nulls. The link between modular organization and nFC spectrum provides a mechanistic explanation for the presence or absence of large cofluctuation events when modular structure is altered. While static null models can reproduce static clustering outcomes, they do not capture temporal transitions, leaving room for edge-centric approaches to contribute when combined with dynamic analyses and model-based frameworks bridging structure and function.

This work provides a mathematical foundation for edge-centric analysis, deriving eFC, edge communities, RSS behavior, spatial activity modes, binarized edge properties, and CAPs findings directly from static nFC. Empirical validation on HCP data confirms that a static Gaussian null model explains the majority of edge-centric features, suggesting that many current edge-centric metrics are predominantly static. The main contributions are: (1) a closed-form mapping from nFC to eFC; (2) analytic links between communities and nFC; (3) identification of RSS as BOLD amplitude with a heavy-tailed distribution governed by nFC eigenvalues; and (4) unifying interpretations connecting edge-centric events, CAPs, and principal components of nFC. Future research should prioritize dynamic measures that leverage temporal structure (e.g., event timing, synchronization, transition dynamics), develop model-based approaches to relate structure and function using temporally unfolded dependence measures, and explore efficient frame-selection strategies for data compression without loss of predictive power.

• The primary null hypothesis assumes temporally i.i.d. multivariate Gaussian processes with covariance equal to the empirical nFC, which does not reflect the known autocorrelation and non-Gaussian features of BOLD signals. Thus, findings show sufficiency of static models to reproduce edge-centric features but do not negate the existence of temporal neural dynamics.
• Many derivations rely on ergodicity and Gaussian fourth-moment factorization; deviations from these assumptions may alter predictions.
• Sliding-window-specific dynamics (ordering of windows, dwell times, transition probabilities) are not addressed by the presented static analyses.
• Empirical tests are based on HCP “100 unrelated” dataset with specific preprocessing (ICA-FIX) and Schaefer-200 parcellation; generalizability to other datasets, preprocessing choices, and parcellation scales, while likely, is not exhaustively demonstrated.
• Community detection outcomes depend on stochastic clustering algorithms (e.g., k-means); while distances predict high agreement, exact assignments can vary with initialization and parameters.