Human cortex organizes dynamic co-fluctuations along the sensorimotor-association axis Authors: De-Zhi Jin, Changsong Zhou, Xi-Nian Zuo, Joshua Faskowitz, Ting Xu, Ye He Published: 2026 ABSTRACT Human functional brain networks, originating from coherent fluctuations in brain activity, are organized along a hierarchical axis. However, how this stable hierarchical architecture emerges from time-varying co-fluctuations remains unknown. Existing dynamic analyses have primarily focused on high-amplitude co-fluctuations, largely overlooking the contribution of lower-amplitude activity. Here, we investigated the amplitude-dependent configurations of regional co-fluctuation. We found that these patterns were hierarchically aligned with the sensorimotor-association (SA) axis: sensorimotor networks are preferentially expressed during high-amplitude co-fluctuations, associative systems prevailed during intermediate amplitudes, and limbic system preferentially engaged in low-amplitude states. This amplitude-stratified hierarchy underwent developmental refinement from childhood to adulthood and adaptively reconfigured under naturalistic stimuli. Replicated across four independent datasets including 7 T fMRI, these findings uncover a fundamental principle whereby the brain’s hierarchical architecture is actively preserved through a structured, amplitude-dependent cascade of functional co-fluctuations. Our framework bridges dynamic coordination and stable architecture, demonstrating how the brain balances external processing with internal cognition through amplitude-stratified interactions. Introduction The human brain is organized into large-scale functional networks, a principle reproducibly identified through functional magnetic resonance imaging (fMRI)1–3. These networks exhibit spatial and temporal coherent fluctuations (co-fluctuation) in blood-oxygen-level-dependent (BOLD) signals of fMRI–collectively termed functional connectivity (FC)–which is critical for supporting specialized cognitive functions4–6. Understanding the organization of these networks has been a central focus in neuroscience, particularly as their disruption is linked to many neurological and psychiatric disorders7–9. A key discovery is that these functional networks are arranged along a hierarchical axis, from unimodal sensorimotor cortices to transmodal association areas10,11. This hierarchical framework represents a fundamental principle of brain organization, improving our understanding of how the brain supports a spectrum of functions from immediate sensorimotor tasks to complex cognition. While static FC analyses have provided invaluable insights, this time-averaged view belies the brain’s intrinsically dynamic nature. Brain function arises from continuous, moment-to-moment fluctuations in neural activity. A growing body of evidence has demonstrated that studying time-varying patterns of brain interactions unveils critical dynamical structures with methods such as sliding-window correlations and coactivation patterns (CAPs)12–26. At the system level, studies have revealed that the brain transitions nonrandomly between recurring, hierarchically organized network configurations27–30. Yet, it remains poorly understood how these transient, moment-to-moment co-fluctuations across the whole brain are organized to support the stable, hierarchical architecture of brain networks. Specifically, it is unclear whether the relative configuration of regional interactions remains constant across different spontaneous ebb and flow of whole-brain activity levels. Investigating this question requires tracking dynamic network interactions. Traditional dynamic methods, such as sliding-window correlations or CAPs, could not capture frame-wise interaction across all regions. Conversely, recent edge-centric frameworks evaluate co-fluctuation time series to capture single-frame FC across regions31–33. This framework interprets static FC as the temporal aggregation of moment-to-moment co-fluctuations34, providing unique sensitivity to capture transient changes in functional networks35. However, studies utilizing this approach have primarily focused on sparse, high-amplitude co-fluctuation “events”32,34,36–38, which are identified as the dominant drivers of static FC32. The low-amplitude fluctuations are frequently dismissed as susceptible to noise. This assumption overlooks the spontaneous dynamics of co-fluctuation, as emerging work suggests that different amplitudes may inform distinct patterns of brain networks39–42. Raw co-fluctuation values are inherently scaled by the global amplitude of the system, the fine-grained regional configurations during lower-amplitude periods would be obscured by the reduced signal magnitude. This makes it difficult to determine whether different amplitude regimes are characterized by distinct patterns of regional co-fluctuations. To address this, we sought to explore how regional co-fluctuation configurations reorganize across the different global amplitudes to support stable, hierarchical brain organization. We introduce a metric termed the co-fluctuation score, calculated as the ratio of regional to global co-fluctuation amplitude. Rather than focusing on absolute interaction strength (aggregated from edge timeseries), which is dominated by global fluctuations, this score quantifies the proportional engagement of individual regions relative to the whole-brain co-fluctuation at each moment. This approach allows us to characterize the shifting landscape of regional contributions across high, intermediate, and low-amplitude states. Applying this framework to resting-state fMRI, our analysis reveals a previously unrecognized dynamic dimension of the brain’s hierarchical organization: intrinsic co-fluctuation patterns are systematically stratified by their amplitude. We show that as global amplitude shifts, the relative configuration of regional interactions changes along the sensorimotor-association (SA) axis. High-amplitude states are preferentially expressed by sensorimotor systems tuned to external processing, whereas mid-to-low amplitude states preferentially engage association and limbic systems that facilitate internal cognition. This organizational pattern exhibits developmental refinement from childhood to adulthood and reconfigures adaptively during naturalistic stimulation. Validated across multiple independent datasets including 7 T fMRI, our findings demonstrate that spontaneous brain activities interact within a structured, amplitude-dependent framework that hierarchically transit from sensorimotor to association dominance, revealing a fundamental mechanism by which intrinsic, time-varying interactions are organized to preserve the brain’s hierarchical architecture. Results Static FC between two brain regions can be temporally unwrapped into a timeseries of co-fluctuations (edge timeseries), derived from the frame-wise product of z-scored BOLD signals (Fig. 1a). At each timepoint, the magnitude of global and regional co-fluctuation is quantified using the root sum of squares (RSS) of the corresponding edge timeseries. Notably, global co-fluctuation amplitude is not directly related to the frame displacement, global signal, respiratory variation, and heart rate variability (Supplementary Fig. 1). To enhance signal-to-noise ratio and facilitate cross-subject alignment of amplitudes, we partitioned time points into 20 bins based on the frame-wise global RSS (5% of total time points per bin) for each subject. Global and regional RSS values are subsequently averaged within each bin. The regional co-fluctuation score is defined as the ratio of regional RSS to global RSS for each amplitude bin. This metric quantifies proportional co-fluctuation level of a given region relative to the whole-brain, thereby enabling a direct comparison of functional topographies across amplitudes (Fig. 1b). We calculated co-fluctuation scores for each of the 200 brain regions defined by Shaefer parcellation43 for each subject from 100 unrelated subjects of the Human Connectome Project (HCP) dataset (Fig. 1c, see “Materials and Methods” for details).Fig. 1Schematic illustration for amplitude-stratified co-fluctuation analysis.a Computing edge-wise co-fluctuation amplitudes. For each pair of brain regions, their edge timeseries (ETS, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${E}_{i,j}$$\end{document}Ei,j) is derived from the element-wise product of z-scored BOLD timeseries (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${z}_{i}$$\end{document}zi and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${z}_{j}$$\end{document}zj). At each timepoint, global or regional co-fluctuation amplitudes are computed as the root sum of squares (RSS) of ETS across all edges (RSS GLOBAL) or across edges connected to a given region (RSS REGION). b Amplitude-stratified co-fluctuation scoring. Timepoints are ranked by their global RSS values and partitioned into 20 bins (5 % of timepoints/bin). Within each bin, global and regional RSS values are averaged. The regional co-fluctuation score is defined as the ratio of regional RSS to global RSS from a specific bin. c Two analytical strategies of co-fluctuation scores. First, model amplitude-dependent trends via generalized additive models. Second, compare spatial patterns across global co-fluctuation amplitude bins. Generated using MATLAB R2023a and R version 4.3.2. Source data are provided as a Source Data file. BOLD: Blood Oxygen Level Dependent. Schematic illustration for amplitude-stratified co-fluctuation analysis. a Computing edge-wise co-fluctuation amplitudes. For each pair of brain regions, their edge timeseries (ETS, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${E}_{i,j}$$\end{document}Ei,j) is derived from the element-wise product of z-scored BOLD timeseries (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${z}_{i}$$\end{document}zi and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${z}_{j}$$\end{document}zj). At each timepoint, global or regional co-fluctuation amplitudes are computed as the root sum of squares (RSS) of ETS across all edges (RSS GLOBAL) or across edges connected to a given region (RSS REGION). b Amplitude-stratified co-fluctuation scoring. Timepoints are ranked by their global RSS values and partitioned into 20 bins (5 % of timepoints/bin). Within each bin, global and regional RSS values are averaged. The regional co-fluctuation score is defined as the ratio of regional RSS to global RSS from a specific bin. c Two analytical strategies of co-fluctuation scores. First, model amplitude-dependent trends via generalized additive models. Second, compare spatial patterns across global co-fluctuation amplitude bins. Generated using MATLAB R2023a and R version 4.3.2. Source data are provided as a Source Data file. BOLD: Blood Oxygen Level Dependent. Hierarchies of amplitude-stratified co-fluctuation dynamics across cortical networks To better understand how the relative engagement of regional co-fluctuations dynamically change under different global co-fluctuation amplitudes, we delineated the trends of co-fluctuation scores changing with global amplitudes by employing Generalized Additive Models (GAMs) for each brain region (Fig. 2a).Fig. 2Amplitude-dependent trends of co-fluctuation scores.a Amplitude-dependent co-fluctuation score dynamics for a representative region of 100 subjects modeled via generalized additive models (GAMs). b Co-fluctuation score trends were grouped for eight canonical functional networks (VIS-visual, SMN-somatomotor, DAN-dorsal attention, SVAN-salience/ventral attention, LIMBIC-limbic, CONT-control, DMN-default mode, TP-temporal parietal). c Primary amplitude effects. Trends of regional co-fluctuation scores are colored by partial R2, with each line representing a single brain region. d Second amplitude effects. Trends of regional co-fluctuation scores are colored by the averaged second-order derivatives. Only the top and bottom 50 regions for both the primary and second amplitude effects are shown to clearly depict their distinct trends. e Scatter plot of the primary and second amplitude effects colored by eight functional networks. The lines represent the contours of the highest density regions of each network generated by two-dimensional density estimation with ggdensity package in R. f Primary amplitude effects correlate with T1w/T2w myelination map (left) and the difference map of normalized parvalbumin and somatostatin (PVALB-SST) expression (right). Pearson’s correlation coefficients (r) and corresponding two-sided spin-test p values (pspin) are reported in each panel (n = 200). The gray shaded area indicates the 95% confidence interval of the fitted line. Generated using R version 4.3.2. Source data are provided as a Source Data file. Amplitude-dependent trends of co-fluctuation scores. a Amplitude-dependent co-fluctuation score dynamics for a representative region of 100 subjects modeled via generalized additive models (GAMs). b Co-fluctuation score trends were grouped for eight canonical functional networks (VIS-visual, SMN-somatomotor, DAN-dorsal attention, SVAN-salience/ventral attention, LIMBIC-limbic, CONT-control, DMN-default mode, TP-temporal parietal). c Primary amplitude effects. Trends of regional co-fluctuation scores are colored by partial R2, with each line representing a single brain region. d Second amplitude effects. Trends of regional co-fluctuation scores are colored by the averaged second-order derivatives. Only the top and bottom 50 regions for both the primary and second amplitude effects are shown to clearly depict their distinct trends. e Scatter plot of the primary and second amplitude effects colored by eight functional networks. The lines represent the contours of the highest density regions of each network generated by two-dimensional density estimation with ggdensity package in R. f Primary amplitude effects correlate with T1w/T2w myelination map (left) and the difference map of normalized parvalbumin and somatostatin (PVALB-SST) expression (right). Pearson’s correlation coefficients (r) and corresponding two-sided spin-test p values (pspin) are reported in each panel (n = 200). The gray shaded area indicates the 95% confidence interval of the fitted line. Generated using R version 4.3.2. Source data are provided as a Source Data file. From low to high global amplitude, the changing patterns of co-fluctuation scores showed systematic variations across eight canonical functional networks (Fig. 2b). Regions within somatomotor (SMN), visual (VIS) and dorsal attention (DAN) networks exhibited near-linear increases in co-fluctuation scores with rising global amplitudes. Regions within limbic network (LIMBIC) showed a downward trend as global co-fluctuation amplitudes increased. Regions in the control (CONT) and default mode (DMN) networks showed mixed trajectories, with the majority of regions exhibiting an inverted U-shaped trend. In addition, regions within salience/ventral attention network (SVAN) exhibited multiple profiles. These patterns delineate three amplitude-dependent coupling regimes: the primary sensorimotor systems driving global co-fluctuations through linear amplification; higher-order cognitive systems balancing co-fluctuation with peaks under the intermediate amplitudes; and limbic systems engaging more during low-amplitudes states and decoupling during high-amplitude states. GAM analysis identified significant smooth terms in 195 of 200 regions (Benjamini-Hochberg corrected p < 0.05; Supplementary Fig. 7c), indicating a strong association between co-fluctuation score and global RSS. We quantified these effects through partial R2 (linear) and the averaged second derivatives (quadratic) of the trajectories, namely the primary and second amplitude effects (Fig. 2c, d). The distribution of the primary and second amplitude effects (Fig. 2e) revealed three regimes of amplitude-dependent co-fluctuation reorganization, including the limbic regions, higher-order (mean centile: x = 27.0%, y = 24.0%) and primary systems (mean centile: x = 74.5%, y = 72.0%). It further confirmed three types of amplitude-dependent co-fluctuation reorganization. The primary and second amplitude effects may reflect a spatial-temporal organizing principle of dynamic functional architecture. We further explored the correlates of regional heterogeneity of the primary amplitude effects with neuroanatomical and transcriptomic profiles. Previous works suggest cortical myelination and the relative presence of parvalbumin (PVALB) and somatostatin (SST) reflect hierarchical functional differences across cortex44–47. Their regional heterogeneity may be related to the amplitude effects. Consistent with this prediction, significant positive correlations of the primary amplitude effects with cortical myelination (r = 0.61, pspin = 1 × 10−4, n = 200; Fig. 2f), as well as with relative densities of PVALB-SST (r = 0.74, pspin = 9 × 10−5, n = 200; Fig. 2f) were observed. Amplitude-dependent spatial configuration of co-fluctuation scores along the SA axis To further explore how the spatial organization of regional co-fluctuation scores changed with respect to different global co-fluctuation amplitudes, we computed group-averaged spatial maps of co-fluctuation scores for each amplitude bin (Fig. 3a, the full set of 20 brain maps is shown in Supplementary Fig. 2). Distinct spatial configurations emerged across global amplitudes. High-, intermediate-, and low-amplitude bins revealed spatial patterns of sensorimotor, associative and limbic systems respectively. The spatial similarities between these maps revealed two clusters with dissociation between high-amplitude bins and mid-to-low amplitude bins (Fig. 3b). Neurosynth meta-analysis further demonstrated functional segregation of these states. High-amplitude maps of co-fluctuation scores predominantly correlated with sensorimotor function, while intermediate-amplitude maps aligned with high-order cognitive domains like memory and language. Low-amplitude maps mainly linked to social and emotion functions (Fig. 3c).Fig. 3Amplitude-dependent spatial configuration of co-fluctuation scores.a Group-averaged maps of co-fluctuation scores for high, intermediate and low amplitude bins. Sensorimotor regions dominate high-amplitude states, whereas associative regions peak at intermediate amplitudes, and limbic system prevail during low-amplitude states. b The spatial similarities of co-fluctuation score maps across 20 bins (Pearson’s correlation coefficients, n = 200), revealing two clusters. c Neurosynth meta-analysis of top 20% regions per bin decoded with 23 topic terms identifies amplitude-stratified functions (Pearson’s correlation coefficients, two-sided; BH-FDR corrected, *p < 0.005). d Sensorimotor-association (SA) axis alignment. Spearman’s correlation (n = 200) between co-fluctuation score maps and SA-rank map transitions from high amplitudes to low amplitudes. Spearman’s correlation coefficients (r) and corresponding two-sided spin-test p values (pspin) are reported in Supplementary Table 5. Generated using MATLAB R2023a and R version 4.3.2. Source data are provided as a Source Data file. Amplitude-dependent spatial configuration of co-fluctuation scores. a Group-averaged maps of co-fluctuation scores for high, intermediate and low amplitude bins. Sensorimotor regions dominate high-amplitude states, whereas associative regions peak at intermediate amplitudes, and limbic system prevail during low-amplitude states. b The spatial similarities of co-fluctuation score maps across 20 bins (Pearson’s correlation coefficients, n = 200), revealing two clusters. c Neurosynth meta-analysis of top 20% regions per bin decoded with 23 topic terms identifies amplitude-stratified functions (Pearson’s correlation coefficients, two-sided; BH-FDR corrected, *p < 0.005). d Sensorimotor-association (SA) axis alignment. Spearman’s correlation (n = 200) between co-fluctuation score maps and SA-rank map transitions from high amplitudes to low amplitudes. Spearman’s correlation coefficients (r) and corresponding two-sided spin-test p values (pspin) are reported in Supplementary Table 5. Generated using MATLAB R2023a and R version 4.3.2. Source data are provided as a Source Data file. Next, we assessed the spatial alignment of these spatial maps with the sensorimotor-association (SA) axis, which represents a fundamental hierarchy of human cortex followed by multiple neurobiological properties10,11,48,49. As illustrated in Fig. 3d, the spatial similarities between co-fluctuation score maps and the SA axis exhibited systematic amplitude dependence (Supplementary Table 5). Notably, the two identified clusters exhibited opposing patterns along the SA axis. The highest-amplitude map (top 5% bins) of co-fluctuation scores was strongly aligned with the sensorimotor-dominated SA axis (r = −0.69, pspin = 9 × 10−5, n = 200). This relationship weakened progressively with decreasing amplitude, reaching non-significance at bin 75–80%. Beyond this turning point, the relationship reversed, with co-fluctuation score maps increasingly aligning with the association-dominated SA axis and peaked at mid-amplitudes (at bin 50-55%, r = 0.79, pspin = 9 × 10−5, n = 200). These amplitude-dependent shifts suggest that dynamic co-fluctuation across amplitudes is constrained by the SA axis, with switching between three systems: sensorimotor regions driving global co-fluctuation at high amplitudes, association regions engaging at mid-amplitudes, and limbic regions prevailing during low-amplitude states. This dynamic alignment may underscore the hierarchical and amplitude-dependent organization of brain functional networks. Developmental refinement of opposing co-fluctuation patterns along the SA axis The adult findings revealed amplitude-dependent configurations of co-fluctuation scores along the SA axis. To investigate how these patterns develop during childhood and adolescence, we analyzed the HCP development dataset (6-18 years old) by averaging co-fluctuation score maps across participants within the same age groups. The two clusters identified in adults were detectable as early as age 6 and persisted across development (Fig. 4a). Importantly, the distinction between the two clusters became more pronounced with age, as indicated by increasingly divergent distributions of the correlations between co-fluctuation score maps of different amplitudes (Fig. 4b). Furthermore, the amplitude-dependent alignment of co-fluctuation score maps with the SA axis exhibited progressive convergence toward adult profiles, particularly at the trough (high amplitude) and peak (intermediate amplitude), reaching adult-equivalent values by late adolescence (rage18, adult = 0.98, rage6, adult = 0.72, Spearman’s rho, n = 20; Fig. 4c, d). These results suggest that spatial patterns of regional co-fluctuation become more defined and specialized throughout development.Fig. 4Developmental refinement of amplitude-dependent co-fluctuation hierarchy.a Correlations of co-fluctuation score maps across 20 amplitude bins from 6 to 18 years old (Pearson’s correlation coefficient, n = 200). b Distributions of correlations of co-fluctuation score maps across 20 bins, revealing increasing dissociation of two clusters from 6 to 18 years old. c Trends of amplitude-dependent alignment between co-fluctuation score maps and SA rank map approach adult-like pattern (gray line) from 6 to 18 years old (Spearman’s correlation coefficients, n = 200). To facilitate visualization, lines were smoothed, and the original figure is provided in Supplementary Fig. 4. d The similarities between SA ranks and co-fluctuation score maps of individuals at high and intermediate amplitude bins. The dark red lines represent GAM fits, with age included as a smooth term. Each point represents an individual participant (n = 530). e, f Left: Age-related effects on co-fluctuation scores at high and intermediate amplitude bins quantified using t-statistics from linear mixed-effects models (two-sided, BH-FDR corrected p < 0.005). Right: SA ranks predict age-related effects on co-fluctuation scores at high and intermediate amplitude bins. Pearson’s correlation coefficients (r, n = 200) and corresponding two-sided spin-test p values (pspin) reported. The gray shaded area indicates the 95% confidence interval of the fitted line. Generated using MATLAB R2023a and R version 4.3.2. Source data are provided as a Source Data file. Developmental refinement of amplitude-dependent co-fluctuation hierarchy. a Correlations of co-fluctuation score maps across 20 amplitude bins from 6 to 18 years old (Pearson’s correlation coefficient, n = 200). b Distributions of correlations of co-fluctuation score maps across 20 bins, revealing increasing dissociation of two clusters from 6 to 18 years old. c Trends of amplitude-dependent alignment between co-fluctuation score maps and SA rank map approach adult-like pattern (gray line) from 6 to 18 years old (Spearman’s correlation coefficients, n = 200). To facilitate visualization, lines were smoothed, and the original figure is provided in Supplementary Fig. 4. d The similarities between SA ranks and co-fluctuation score maps of individuals at high and intermediate amplitude bins. The dark red lines represent GAM fits, with age included as a smooth term. Each point represents an individual participant (n = 530). e, f Left: Age-related effects on co-fluctuation scores at high and intermediate amplitude bins quantified using t-statistics from linear mixed-effects models (two-sided, BH-FDR corrected p < 0.005). Right: SA ranks predict age-related effects on co-fluctuation scores at high and intermediate amplitude bins. Pearson’s correlation coefficients (r, n = 200) and corresponding two-sided spin-test p values (pspin) reported. The gray shaded area indicates the 95% confidence interval of the fitted line. Generated using MATLAB R2023a and R version 4.3.2. Source data are provided as a Source Data file. To identify key brain regions associated with the development of regional co-fluctuation patterns, we examined age-related effects on regional co-fluctuation scores at high (90–100%) and intermediate (40–60%) amplitudes separately. At high amplitudes, regions within the sensorimotor and visual networks exhibited significant age-related increases in co-fluctuation scores, whereas regions within the default mode and control networks showed significant age-related decreases (Fig. 4e; and Supplementary Table 1). Conversely, these regions displayed the opposite developmental effects at intermediate amplitudes (Fig. 4f; and Supplementary Table 2). These findings suggest that the development of spatial differentiation of co-fluctuation scores unfolds along the SA axis in opposite directions, ultimately shaping the hierarchical organization of functional networks. It highlights the developmental mechanisms underlying the functional organization of the cerebral cortex. State-dependent modulation of regional co-fluctuation patterns Previous section reveals the long-term changes of two opposite spatial patterns of regional co-fluctuation. Whether these patterns can change in short-term remains unclear. To address this, we investigated state modulation by comparing resting-state and movie-watching conditions in the HCP 7 T fMRI dataset. The trends of co-fluctuation scores across the global amplitudes in both conditions were replicated in 7 T dataset compared to previous results from 3 T dataset (Fig. 5a).Fig. 5Naturalistic stimuli modulate spatial patterns of co-fluctuation scores.a Amplitude-dependent trends of co-fluctuation scores of seven canonical functional networks under resting-state and movie-watching conditions, derived from HCP 7 T fMRI dataset. b Spatial alignments of SA ranks and co-fluctuation scores across amplitude bins in four test-retest scans during resting-state and movie-watching conditions (Spearman’s correlation coefficients, n = 200). Lines were smoothed using GAM model. c Significant differences in regional co-fluctuation scores between resting-state and movie-watching conditions at high-amplitude bins (left) and intermediate-amplitude bins (right), assessed using paired two-sided t-tests (n = 182). Regions surviving BH-FDR correction (p < 0.05) are shown. Generated using MATLAB R2023a and R version 4.3.2. Source data are provided as a Source Data file. Naturalistic stimuli modulate spatial patterns of co-fluctuation scores. a Amplitude-dependent trends of co-fluctuation scores of seven canonical functional networks under resting-state and movie-watching conditions, derived from HCP 7 T fMRI dataset. b Spatial alignments of SA ranks and co-fluctuation scores across amplitude bins in four test-retest scans during resting-state and movie-watching conditions (Spearman’s correlation coefficients, n = 200). Lines were smoothed using GAM model. c Significant differences in regional co-fluctuation scores between resting-state and movie-watching conditions at high-amplitude bins (left) and intermediate-amplitude bins (right), assessed using paired two-sided t-tests (n = 182). Regions surviving BH-FDR correction (p < 0.05) are shown. Generated using MATLAB R2023a and R version 4.3.2. Source data are provided as a Source Data file. However, naturalistic stimulation attenuated the alignment of co-fluctuation scores with the SA axis particularly at the high and middle amplitudes (Fig. 5b), suggesting stimulus-driven hierarchy rebalancing. Regional specificity analysis unveiled the underlying basis of this misalignment (Fig. 5c; Supplementary Tables 3 and 4). Sensory-processing regions (parieto-occipital/superior temporal cortices) showed significantly reduced co-fluctuation scores at high-amplitudes and enhanced that at mid-amplitudes in movie-state compared to resting-state. Conversely, high-order functioning regions (prefrontal/cingulate cortices) displayed mirror-symmetric patterns, evidencing co-fluctuation competition. Notably, the primary visual cortex exhibited a distinct state-dependent pattern. At high amplitudes, its co-fluctuation score was significantly higher in the movie-watching state, likely reflecting visual stimuli-driven processing. At middle amplitudes, its co-fluctuation score was significantly higher during the resting state, indicating suppression of the co-fluctuation of the primary visual cortex during movie-watching. Collectively, these findings demonstrate rapid reconfiguration of co-fluctuation patterns between brain states, particularly in response to naturalistic stimuli, suggesting a dynamic modulation of functional organization that rebalance external sensory processing against internal cognitive demands. Sensitivity and replication analyses We have done extended analyses to make sure our findings are reliable and replicable. First, addressing the contentious methodological issue of noise regression, we demonstrated that the frame displacement (FD), global signal (GS), respiratory variation (RV) and heart rate variability (HRV) were not directly related to global RSS (Supplementary Fig. 1). Critically, our findings are robust to the exclusion of these noises or not (Supplementary Fig. 1 and 12). Meanwhile, temporal signal-to-noise ratio (tSNR), mean FD, mean GS, mean RV, and mean HRV showed significant covariate effects on co-fluctuation scores in specific brain regions (Supplementary Fig. 11), suggesting region-dependent nuisance influences and underscoring the importance of covariate control, particularly head motion. Second, nonlinear models provided a better fit than linear models for most regions (Supplementary Fig. 7b) when modeling the relationship between the co-fluctuation score and global RSS. Fixed-effect and mixed-effect GAM specifications produced similar results (Supplementary Fig. 7a). Third, the overall trajectory shapes and the primary amplitude effect remained consistent across GAM models fitted with different number of basis dimensions (k = 3, 5 and 10; Supplementary Fig. 18 and 19). Fourth, sensitivity analyses varying the number of bins confirmed the robustness of the results. In the HCP3T dataset, findings were highly consistent across 10, 20, and 40 bins (Supplementary Fig. 4c and d). Given the shorter scan duration in the HCP development dataset compared to the HCP dataset, we repeated the analysis with 10 bins on developmental dataset. The age-related findings remained consistent with those observed using 20 bins (Supplementary Fig. 4a). Next, we replicated the main findings across independent datasets: 7 T resting-state data recapitulated 3T-derived findings (Fig. 5a; and Supplementary Fig. 6), while independent developmental dataset replicated age-related findings (see details in Materials and Methods, and results in Supplementary Fig. 5). Meanwhile, given the relatively larger head motion in children, we evaluated the robustness of the main results in the HCPD dataset using two FD thresholds (0.2 mm and 0.3 mm, Supplementary Fig. 17). In the HCP3T dataset, we assessed robustness across cortical parcellations with different construction principles (Supplementary Fig. 16): the functionally derived Schaefer200×17 and Schaefer400×17 atlases (200 and 400 parcels across 17 canonical networks) and the multimodal Glasser360 atlas (360 cortical areas defined using structural, functional, and myelin-based features). Finally, to test the robustness of the proposed co-fluctuation score, we designed several variants accounting for amplitude and ratio scaling (see Supplementary Methods and Supplementary Fig. 9). The first four variants use ratio-based formulations: Methods 1–2 share a common denominator across regions, whereas Methods 3–4 exclude each region’s own contribution from the denominator. Method 6 applies z-scoring to regional RSS values to rule out artificial anticorrelation effects caused by the scaling, and Method 7 directly uses regional RSS values, providing the most direct test of the opposing spatial patterns captured by the co-fluctuation score. Results were highly consistent across these approaches. By contrast, a fifth variant based on global regression of regional RSS yielded a uniform spatial pattern across regions, suggesting that global regression obscures amplitude-dependent cortical organization (Supplementary Fig. 9). Despite the artificial anticorrelation effects potentially introduced by the proportional scaling, it would not affect the spatial correlation with the SA-axis mathematically. Its simplicity and stability make it suitable for capturing amplitude-dependent co-varying spatial patterns. Phase randomization further indicates that the co-fluctuation score captures intrinsic hierarchical dynamics (see details in Supplementary methods and Supplementary Fig. 10). This convergent evidence across MRI protocols, preprocessing strategies, datasets, and methods substantiates that the observed regional co-fluctuation patterns reflect intrinsic neurobiological profiles rather than methodological artifacts. Discussion Our study reveals an amplitude-stratified cortical dynamic architecture where spontaneous co-fluctuation patterns reorganized between functional systems along the sensorimotor-association axis (Fig. 6). Moving beyond conventional focus on high-amplitude events, our work demonstrate that dynamic co-fluctuation patterns are organized depending on amplitudes by conserved hierarchical principles: sensorimotor regions are preferentially expressed during high-amplitude global co-fluctuations states, while associative/limbic cortices preferentially engage during mid-low amplitude regimes. This dual-dominance co-fluctuation patterns evolve during development, becoming more aligned with adult-like functional organization. Furthermore, the state-dependent modulation of these patterns under naturalistic stimulation demonstrates their adaptive reconfiguration capacity corresponding to cognitive and perceptual demands. This amplitude-stratified framework bridges transient dynamic interaction with stable cortical hierarchy, offering insights into how intrinsic hierarchical functional networks are dynamically organized.Fig. 6A schematic representation of the strong association between static hierarchy and dynamic co-fluctuations across high, middle and low co-fluctuation amplitudes.From static to dynamic, the hierarchy between sensorimotor and association regions remains evident, with two opposing co-fluctuation score patterns anchored at high and middle-to-low amplitudes, respectively. This underscores that dynamic co-fluctuations are intrinsically configured along the SA axis and are balanced between internal and external processing. Generated using R version 4.3.2. A schematic representation of the strong association between static hierarchy and dynamic co-fluctuations across high, middle and low co-fluctuation amplitudes. From static to dynamic, the hierarchy between sensorimotor and association regions remains evident, with two opposing co-fluctuation score patterns anchored at high and middle-to-low amplitudes, respectively. This underscores that dynamic co-fluctuations are intrinsically configured along the SA axis and are balanced between internal and external processing. Generated using R version 4.3.2. Dynamic stratification of cortical interactions Our work introduced amplitude-dependent state stratification as organizational principle of dynamic networks. Brain activity is ever-changing, involving dynamic interactions across functional networks that adapt us to the surrounding changing world. While conventional functional connectivity captures time-averaged coordination between regions, our amplitude-resolved decomposition demonstrates how transient co-fluctuations assemble into dynamic network states39. Co-fluctuation provides complementary information beyond traditional FC metrics36,40,50,51. However, prior research predominantly focused on high-amplitude co-fluctuations32,34,52, neglecting low-amplitude counterparts that constitute the majority of brain activity16,51. By implementing co-fluctuation score normalization, we mitigate the amplitude confound and enable direct comparison of regional co-fluctuation patterns across different amplitude states, thereby providing a unified characterization from a systematic perspective. Our findings revealed that regional co-fluctuations had systematically changing patterns with global amplitude. The primary and secondary amplitude effects exhibited a hierarchical distribution of three systems53,54. Sensorimotor systems exhibited near-linear amplification of coupling with rising global amplitudes, implying their role in driving global co-fluctuation55. High-order systems, especially DMN and CONT, followed inverted U-shaped trends, peaking at intermediate amplitudes. Limbic system demonstrated amplitude-dependent decoupling, potentially reflecting their role in dominating low-amplitude co-fluctuation. These findings delineate a tripartite amplitude-dependent reorganization of functional networks, in which optimally orchestrated reconfiguration maximizes the capacity of information preprocessing56 and enables transient functional transitions with minimal metabolic cost57. Amplitude-mediated co-fluctuation transitions along cortical hierarchy Our discovery of amplitude-dependent transition in regional co-fluctuation patterns unveils how the SA axis orchestrates the dynamic reconfiguration of functional interaction. This axis represents a critical organizational principle of the human cortex10,11,48,49, characterized by a hierarchy from primary sensorimotor areas to high-order association areas. It conserved across the structure46, function10, development58, and evolution59 of human cerebral cortex. The functional dynamics also follow the cortical hierarchy27,28,60–62. Recent dynamic modeling studies demonstrate that incorporating SA-axis topology recapitulates empirical dynamic FC55,63. Hierarchical dynamics may represent an intrinsic organizational principle of brain function that forming the stable hierarchical architecture41,64,65. While prior work identified temporal transitions between recurring functional states66–69, our amplitude-resolved analysis reveals these represent meta-states of a continuum governed by SA-axis dynamics from high to middle-low amplitudes27. High-amplitude states were dominated by co-fluctuation of sensorimotor systems, whereas mid-low amplitude phases engaged interaction of association/limbic systems. Esfahlani et al. (2020) found similar S-A pattern with our findings at high amplitude32, but without investigating other amplitudes. The dynamic transition of co-fluctuation implies that, under the overarching constraint of the SA axis, complementary and flexible co-fluctuation patterns exist within the human cerebral cortex. High-amplitude states facilitate processing rapid-changing external information, whereas mid-low amplitude phases may process internal information including complex cognition and emotion, collectively reflecting the spatiotemporal dynamics of information processing in the brain70. The two opposing co-fluctuation modes organized along the SA axis and shifted across the amplitude reveal that the human brain may sustain its function through a dynamic equilibrium71—one that confers flexibility to meet complex tasks while guaranteeing the stability of its intrinsic hierarchical architecture72. Developmental refinement of hierarchical dynamics Previous studies have proposed that the development of functional connectivity across cortices unfolds along SA-axis hierarchy, with increasing segregation between sensorimotor and association regions48,49,58,73–78. Here, we showed that the two opposing co-fluctuation patterns gradually became more aligned with the SA axis from childhood to adolescence–sensorimotor networks strengthening high-amplitude coupling while association cortices enhance mid-low amplitude synchronization. It exemplifies how SA-axis hierarchy scaffolds the development of brain dynamics. The developmental pattern of co-fluctuation scores at high amplitudes is consistent with existing research findings48,49,75,77–79. However, the development of co-fluctuation scores at middle-to-low amplitudes is less discussed. Previous research has suggested that system segregation is crucial for cognitive29, language80, and task performance81 processes. Our findings suggest that these opposing co-fluctuation patterns may facilitate information transmission processes between bottom-up sensory processing and top-down cognitive control, supporting increasingly complex task demands during development82,83. These insights could aid in identifying atypical developmental patterns, such as those seen in neurodevelopmental disorders84,85. Stimulus-driven co-fluctuation hierarchy reorganization Naturalistic stimulation induced amplitude-dependent co-fluctuation patterns decoupling with SA axis. During naturalistic stimulation, the dynamic hierarchy adaptively attenuates but does not vanish, reaffirming its role as an inherent organizational principle86. This attenuation further reflects strengthened inter-system coupling87, enabling more efficient processing and integration of contextual information. During movie-watching, sensorimotor regions exhibited suppressed co-fluctuation at high amplitudes and increased co-fluctuation at middle amplitudes, while association regions showed reverse tuning. This stimulus-driven reorganization of co-fluctuation reflects adaptive functional dynamics88,89, where external demands transiently override intrinsic hierarchy constraints87,88,90,91. Such constraint flexibility may enable rapid transitions between perceptive and complex cognitive processes to meet varying environmental challenges87,92–96. While our findings establish amplitude-stratified hierarchy as a core principle of dynamic functional architecture, several key frontiers warrant exploration. First, whether the dynamic hierarchy revealed in this study represents a fundamental organizational principle that is phylogenetically conserved remains to be further investigated10,97, which could help elucidate its origins and functional significance. Second, our study focused solely on conscious states. Investigating unconscious states, such as sleep98,99 or anesthesia100,101, may further elucidate the arousal-related dynamic organization of functional interactions. Third, the study examined only healthy populations. Whether psychiatric or neurodevelopmental disorders manifest atypical dynamics requires further exploration12,13. Four, the neurobiological mechanisms driving amplitudes transitions were unexplored. Future research incorporating biophysical models may help uncover these mechanisms55,63,65,102. Five, how anatomical structure gives rise to the complex functions remains poorly understand103–105. How structural connectome shapes co-fluctuation patterns is still an open question, which would advance our understanding of structure and function coupling. Finally, it should be noted that this study employed the proportional scaling as the measure of regional co-fluctuation score, a method that can potentially introduce artificial anti-correlation effect. Although our sensitivity analyses support the effectiveness and robustness of this approach, this methodological consideration warrants attention, and caution is advised when interpreting results in future studies. Our study unveils that the human cortex employs amplitude-dependent hierarchy to organize dynamic co-fluctuations. High-amplitude states are driven by interactions among sensorimotor regions, while mid-low amplitude regimes recruit co-fluctuation of association/limbic cortices. These opposing patterns are developmentally refined and adaptively reorganized by naturalistic stimuli, establishing the SA axis as fundamental infrastructure of brain dynamic. By resolving amplitude-dependent transitions of co-fluctuation patterns, we provide a framework for understanding how transient coordination assemble into stable functional architectures. Methods Participants We analyzed four independent resting-state fMRI (rs-fMRI) datasets. Three of the datasets are from Human Connectome Project (HCP), including 3 T and 7 T adults rs-fMRI datasets106 as well as 3 T rs-fMRI developmental dataset107. The fourth dataset is from the Chinese Color Nest Project (CCNP), which is served as an independent dataset to assess the reproducibility of the developmental findings108. 3 T rs-fMRI data were obtained from HCP S1200 release and included a subset of 100 unrelated adult participants (54% females; mean age = 29.11 ± 3.67 years; age range = 22−36 years). The 7 T movie-watching fMRI data were used to examine differences between resting-state and naturalistic stimulation conditions. 7 T rs-fMRI dataset consists 182 healthy subjects (60% female; mean age = 27.53 ± 3.26 years; age range = 22−36 years). The participants underwent four 3 T or 7 T rs-fMRI, with eyes open with relaxed fixation on a cross. The HCP study was approved by the Washington University Institutional Review Board, and all participants provided informed consent to the study protocols and procedures. Human connectome project development dataset was used for developmental analyses. This dataset includes 652 subjects (53.9% females; mean age = 14.5 ± 4.1 years; age range = 5−22 years). Each participant older than 8 years underwent a total of 26-minute rs-fMRI scanning across four sessions. For the younger group (5−7 years), rs-fMRI scanning was divided into six sessions, with a total duration reduced to 21 mins. In this study, rs-fMRI data acquired from participants aged 6–18 years were analyzed, totaling 2139 sessions from 530 individuals (53.2% females) after quality control. All participants were instructed to stay still and awake, and blink normally while looking at the fixation crosshair on the screen. The study procedures were approved by the central Institutional Review Board at Washington University in St. Louis, and written informed consent was obtained from parents of participants under 18 years. The Chinese Color Nest Project dataset included a total of 479 participants. Each participant undergoes one, two or three visits, finally composed of a total of 864 visits. Each visit included two rs-fMRI scans. All participants were instructed to stay still and awake, and blink normally while looking at the fixation crosshair on the screen. After quality control, a total of 734 sessions were included for the subsequent analyses, comprising 403 from baseline (50.1% females; mean age = 10.66 ± 3.14 years, age range = 6−18.9 years), 216 from 2-th visit (52.7% females; mean age = 11.97 ± 2.76 years, age range = 7.6−18.9 years) and 115 from 3-th visit (55.6% females; mean age = 12.72 ± 2.51 years, age range = 8.9−18.9 years). CCNP was approved by the Institutional Review Board of the Institute of Psychology, Chinese Academy of Sciences. Prior to conducting the research, written informed consent was obtained from one of the participants’ legal guardians, and written assent was obtained from the participants. Sex information was obtained from the original datasets. For adult participants, sex was determined based on self-report. For children and adolescents, sex was reported by a parent or legal guardian. MRI acquisition For the HCP 3 T dataset, rs-fMRI images were acquired with a multiband gradient-echo echo-planar imaging (EPI) sequence (run duration = 14.33 mins, TR = 720 ms, TE = 33.1 ms, flip angle = 52°, FOV = 208 × 180 mm2, slice thickness = 2 mm, 72 slices, 2-mm isotropic voxel resolution, multiband factor = 8). Physiological data were recorded at 400 Hz using a respiration bellow placed around the chest and a pulse oximeter attached to the fingertip to measure respiration and pulse, respectively. All data were collected on a customized Siemens 3 T Connectome Skyra scanner with a standard 32-channel head coil. 7 T Rs-fMRI images were acquired with a multiband gradient-echo EPI sequence (run duration = 16 mins, TR = 1000 ms, TE = 22.2 ms, flip angle = 45°, FOV = 208 × 208 mm2, slice thickness = 1.6 mm, 85 slices, 1.6-mm isotropic voxel resolution, multiband factor = 5). There are 4 movie-watching scans in total, with each 2 scans were collected after rs-fMRI scanning in session 1 and session 4. It was acquired with the same scanning protocols as rs-fMRI, except for the duration (movie 1/2/3/4 = 15.35 mins/15.3 mins/15.25 mins/15.02 mins). All participants were scanned on a Siemens Magnetom 7 T MRI scanner equipped with a Nova Medical 32-channel receive head coil. For the HCP Development dataset, rs-fMRI data were acquired using a multiband gradient-echo EPI sequence (TR = 800 ms, TE = 37 ms, flip angle = 52°, FOV = 208 × 180 mm2, isotropic voxel size = 2 mm). All data were collected on Siemens 3 T Prisma scanner. The CCNP dataset were acquired at two imaging sites. For CKG sample, rs-fMRI images were acquired with EPI sequence (run duration = 7 min 45 s, TR = 2500 ms, TE = 30 ms, flip angle = 80°, 38 slices). For PEK sample, rs-fMRI images were acquired with a gradient-echo EPI sequence (run duration = 6 mins, TR = 2000 ms, TE = 30 ms, flip angle = 90°, 33 slices). All data were collected on either Siemens 3 T Trio Tim or GE 3 T Discovery MR750 scanners. MRI data preprocessing All HCP and HCPD brain images were preprocessed using the ‘minimal preprocessing pipelines’109. The preprocessing of rs-fMRI images includes distortion and motion correction, alignment to respective T1w image with one-spline interpolation, intensity bias correction and alignment to common space using a multi-model surface registration. Then the BOLD timeseries was linearly detrended, band-pass filtered (0.008−0.08 Hz), and 36-parameters confounds regressed including six head-motion parameters, mean signals of cerebrospinal fluid (CSF), white matters (WM), mean global signal, the derivates of above nine regressors, and the squares of above 18 parameters. A functional parcellation designed to optimize both local gradient and global similarity measures of the fMRI signals was used to define 200 regions of the cerebral cortex43. For HCP dataset, timepoints with framewise displacement > 0.2 mm were excluded, and for HCPD dataset, we removed timepoints with framewise displacement > 0.3 mm. Respiration variability (RV) and heart rate variability (HRV) were computed from the physiological recordings42. Specifically, the temporal standard deviation of the respiratory trace and the mean beat-to-beat interval were calculated within 6-s sliding windows centered on each TR. CCNP dataset was preprocessed using Connectome Computation System (CCS)110. All preprocessing scripts are available on github (https://github.com/zuoxinian/CCS). Rs-fMRI data preprocessing included the following steps: (1) dropping the first 10 s (5 TRs) for the equilibrium of the magnetic field; (2) correcting head motion; (3) slice timing; (4) despiking for the time series; (5) estimating head motion parameters; (6) aligning functional images to high resolution T1 images using boundary-based registration; (7) mitigating nuisance effects such as ICA-AROMA-derived, CSF and white matter signals; (8) removing linear and quadratic trends of the time series; (9) projecting volumetric time series to fsaverage5 cortical surface space; and (10) 6-mm spatial smoothing. Timepoints with framewise displacement > 0.3 mm were excluded and sessions with at least 60% reserved timepoints were included in this study. Estimating time-varying co-fluctuation Functional connectivity (FC) is commonly calculated as the Pearson’s correlation coefficient between BOLD timeseries, which is mathematically equal to the average of temporal co-fluctuation between two regions. Therefore it can be decomposed into a series of moment-by-moment co-fluctuations, which is called edge timeseries (ETS)31,32,34. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${E}_{i,j}$$\end{document}Ei,j is defined as the element-wise product between the z-scored BOLD timeseries \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${z}_{i}(t)$$\end{document}zi(t) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${z}_{j}(t)$$\end{document}zj(t) from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{ROI}}}_{i}$$\end{document}ROIi and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{ROI}}}_{j}$$\end{document}ROIj (Fig. 1a). Each element of ETS denotes the instantaneous co-fluctuation between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{ROI}}}_{i}$$\end{document}ROIi and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{ROI}}}_{j}$$\end{document}ROIj. It is worth noting that the mean of all elements in the edge timeseries is exactly equal to the corresponding FC between two ROIs. Therefore, ETS can be considered as the decomposition of FC. Estimating co-fluctuation score We calculated the root sum squares (RSS) of the co-fluctuation values at each time point in ETS across different edges to quantify the co-fluctuation amplitude. The global co-fluctuation amplitude reflects the instantaneous co-fluctuation across all pairs of cortical regions at a given timepoint, while the co-fluctuation amplitude of single region captures the co-fluctuation between that region and all other brain regions. To compare the configurations of regional co-fluctuation across different global amplitudes, we proposed a measure–co-fluctuation score, which indexes the relative co-fluctuation of a specific region. It can also be considered as the contribution of regional co-fluctuation to the global co-fluctuation. Firstly, we sorted the timepoints into 20 bins (5% timepoints per bin) based on their global co-fluctuation amplitudes, ranging from low to high. Then we calculated the mean global and regional co-fluctuation amplitudes separately within each bin. The ratio of regional to global co-fluctuation is defined as the co-fluctuation score (Eq. 1).1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{Cofluctuation}}\,{\rm{Score}}}_{i,m}=\frac{{{\rm{mean}}}_{t\in {T}_{m}}({{\rm{RSS}}\_{\rm{REGION}}}_{i,t})}{{{\rm{mean}}}_{t\in {T}_{m}}({{\rm{RSS}}\_{\rm{GLOBAL}}}_{t})}$$\end{document}CofluctuationScorei,m=meant∈Tm(RSS_REGIONi,t)meant∈Tm(RSS_GLOBALt) ii represents the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i-{th}$$\end{document}i−th region, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{m}}$$\end{document}m represents the\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\,m-{th}$$\end{document}m−th bin and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${T}_{m}$$\end{document}Tm means the timepoints within \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m-{th}$$\end{document}m−th bin. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{RSS}}\_{\rm{REGION}}}_{i,t}$$\end{document}RSS_REGIONi,t represents the RSS for the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i-{th}$$\end{document}i−th region at timepoint tt. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{RSS}}\_{\rm{GLOBAL}}}_{t}$$\end{document}RSS_GLOBALt represents the RSS of whole-brain at timepoint tt. This procedure generates a co-fluctuation score matrix (N_ROIs × N_Bins) for each subject, with each element representing the contribution of a specific region to the whole-brain co-fluctuation amplitude at a given amplitude bin. The subject-level co-fluctuation scores were averaged to generate group-level co-fluctuation scores at each bin. Delineating the changes of co-fluctuation scores across global amplitudes To explore how relative regional co-fluctuation varies across different amplitudes, we delineated the trajectory of co-fluctuation score across amplitudes for each region. We employed generalized additive models (GAMs) to flexibly model linear or non-linear relationship between regional co-fluctuation scores and global co-fluctuation amplitudes, using ‘mgcv’ package in R. GAM was fitted with co-fluctuation score as the dependent variable, the mean of global co-fluctuation amplitude within each bin (z-scored) as a smooth term. To mitigate potential confounding effects, age, sex, temporal signal-to-noise ratio (tSNR), mean framewise displacement (mFD), mean global signal (mGS), mean respiration variability (mRV) and mean heart rate variability (mHRV) were modeled as linear covariates (Eq. 2). mFD, mGS, mRV and mHRV were calculated within each bin. Because the physiological data were only available in the HCP3T, the covariates only included sex, age, tSNR, mean FD, and mean GS in other datasets.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{Cofluctuation}}\,{\rm{Score}} \sim {\rm{smooth}}\,\left({\rm{global}}\right)\,+\sum _{i=1}^{7}{\beta }_{i}{C}_{i}$$\end{document}CofluctuationScore~smoothglobal+ ∑i=17βiCiwhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C}_{i}$$\end{document}Ci denotes the seven covariates. The models were fitted separately for each region using thin plate regression splines as the basis set of the smooth term and the restricted maximal likelihood (REML) approach for smoothing parameter selection. The degree of penalization was determined automatically within the REML framework, allowing the model to control the effective degrees of freedom of the smooth term. The basis dimension k was set to 5 after the evaluation analyses based on k-check and the Akaike Information Criterion (AIC). Specific procedures and the results of sensitivity analysis were provided in the supplementary materials (Supplementary Table 6, Supplementary Fig 18 and 19). The GAM smooth term produces a spline that represents the trajectory of how regional co-fluctuation contributions to global co-fluctuation vary across global amplitudes. The effect size of each spline (the primary amplitude effect) was quantified as the difference in partial R2 between the full and a nested, reduced model with no smooth term. Statistical significance was assessed by the analysis of variance (ANOVA) to compare the full and nested models. The direction of primary amplitude effect was determined by the sign of the averaged first derivatives of the smooth function, from which we generated a brain map of the primary amplitude effect. To further elucidate the shape of the trajectories, we calculated the averaged second-order derivatives of each trajectory, a measure we termed the second amplitude effect. Positive values indicate a U-shaped profile, while negative values correspond to an inverted U-shape pattern. The curvature of the fitted trajectories was computed as:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k\left(x\right)=\frac{\left|{f}^{{\prime} {\prime} }(x)\right|}{{(1+{(1+{f}^{{\prime} }(x))}^{2})}^{3/2}}$$\end{document}kx=f″(x)(1+(1+f′(x))2)3/2which quantifies the rate of change of the slope along the curve. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f}^{{\prime} }(x)$$\end{document}f′(x) is the first derivative (slope), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f}^{{\prime} {\prime} }(x)$$\end{document}f″(x) is the second derivative, k(x)k(x) is the curvature at position \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{x}}$$\end{document}x. In addition to the fixed-effects model, we fitted mixed-effects GAMs including subject-level random intercepts to account for between-subject variability. To evaluate whether nonlinear effects improved model fit, we compared GAMs with simpler linear models using a likelihood ratio test implemented in anova (model_linear, model_gam, test = “Chisq”) in R. Neurobiological underpinnings To probe the neurobiological basis of the amplitude effect, we examined cortical myelination and microcircuit organization. Cortical myelination provides a canonical structural gradient of cortical hierarchy against which amplitude-stratified co-fluctuation patterns can be evaluated. Cortical distributions of parvalbumin (PVALB) and somatostatin (SST) interneuron markers closely track regional variation in fMRI variability, reflecting differences in inhibitory microcircuit composition across the cortex45. Pearson’s correlation coefficients were calculated between the amplitude effect and two neurobiological markers: myelination and gene expression. Spin testing was used to establish statistical significance. For cortical myelination analysis, we used the averaged T1w/T2w ratio map released by the HCP111. No additional preprocessing was performed beyond averaging the myelin values within each region based on the Schaefer200x1743. For gene expression analysis, microarray expression data and accompanying metadata were obtained from the Allen Human Brain Atlas (AHBA)112. Group-averaged expression values for 16,088 brain specific genes across 200 (Schaefer200x17) cortical areas were used for the analyses. Finally, the difference between z-normalized gene expression values of parvalbumin (PVALB) and somatostatin (SST) was calculated. Meta-analysis of co-fluctuation scores For each amplitude bin, we select the 20% brain regions with highest co-fluctuation score and then project to the 2-mm volumetric MNI152 standard space. The 20 brain volumes were input to meta-analysis (Neurosynth113) and z-statistic is used to represent the association between regions and terms. Similar to previous study10, 23 topic terms were selected within our analysis, and the “reading” term was removed because it did not capture any significant relationships at each bin. To identify cognitive measures significantly associated with the co-fluctuation score, we applied Benjamini-Hochberg multiple-comparison correction within each bin. The relationship between co-fluctuation scores and the SA axis The sensorimotor-association axis (SA axis) represents a cortical hierarchy that varies from unimodal (sensory) regions to transmodal (association) cortex48. The cortical anatomical46, functional10 and evolutionary59 organizations all exhibited similar patterns as SA axis. We investigated whether the co-fluctuation scores at different amplitudes consistently align with the SA axis48. Specifically, we averaged co-fluctuation score map at each amplitude bin across subject as group-level co-fluctuation score. Then we correlated the SA rank (z-scored) map with the group-level co-fluctuation score map at each amplitude bin. For the HCPD dataset, we averaged co-fluctuation score maps across subjects within each age group (ranging from 6–18 year) to assess whether the alignment of co-fluctuation scores with the SA axis changes progressively during development. For better visualization, the trajectories of the similarity between SA ranks and co-fluctuation scores across bins were smoothed by fitting with GAM for each age group. Raw trajectories are illustrated in Supplementary Fig. 4. Spin Permutation Test Spatial permutation testing was performed using 10,000 spin rotations to generate null distributions. Specifically, regional maps were first projected to the fslr32k surface space. Spherical rotations were then applied to the surface maps while preserving hemispheric boundaries and spatial contiguity. Following each rotation, vertex-level data were re-parcellated to the regional level to generate null regional maps, from which the null distribution of correlation values was constructed. The spin_test function provided by the ENIGMA-2.0.0 toolbox was used to implementing above procedures114. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Transparent Peer Review file Supplementary Information Reporting Summary Transparent Peer Review file Source data Source Data Source Data