Cognition emerges from phase dynamics of intrinsic coordination

Curation statements for this article:
  • Curated by eLife

    eLife logo

    eLife Assessment

    This study offers a valuable investigation into how diverse cognitive states might arise from shifts in the relative phase of a fixed set of large-scale propagating modes in fMRI, rather than from reconfigurations of their functional architecture. However, the evidence to indicate that phase features decode 23 tasks better than classical activation-based features is incomplete and would benefit from a comparison against more stringent null models, a more detailed consideration of potential confounds, and a stronger grounding in existing literature. This work will be of interest to cognitive neuroscientists working on the dynamics of large-scale brain modes.

This article has been Reviewed by the following groups

Read the full article

Abstract

The brain generates diverse cognitive states while maintaining a stable functional architecture, a duality that remains difficult to reconcile. Prevailing views assume that flexible cognition necessitates correspondingly flexible architecture with different tasks demanding distinct reconfigurations, leaving the coexistence of stability and flexibility unexplained. Here we introduce the intrinsic network flow (INF) framework as a complementary view. This framework is built on temporally coordinated signal flows across brain networks that constitute a universal scaffold, stable across diverse cognitive states and common across individuals. We show that a wide range of task-evoked activation patterns can be reconstructed by modulating only the temporal phase alignment of these flows, whose fixed structure determines functional connectivity topology, gradients, and large-scale networks, thereby preserving these properties across task states and reconciling flexibility with stability. This situates resting-state and task-state dynamics within a unified framework and suggests a generative relationship from flow-like dynamics to the full landscape of resting-state and task-state phenomena. Crucially, phase information achieved 89.9% accuracy in classifying 23 cognitive tasks, outperforming amplitude- or activation-based markers. These findings reframe task-evoked activation and deactivation as interference among concurrent flows. Together, our work indicates that a key variable for cognition is when intrinsic dynamics align in time, not where or how much the brain activates.

Article activity feed

  1. eLife Assessment

    This study offers a valuable investigation into how diverse cognitive states might arise from shifts in the relative phase of a fixed set of large-scale propagating modes in fMRI, rather than from reconfigurations of their functional architecture. However, the evidence to indicate that phase features decode 23 tasks better than classical activation-based features is incomplete and would benefit from a comparison against more stringent null models, a more detailed consideration of potential confounds, and a stronger grounding in existing literature. This work will be of interest to cognitive neuroscientists working on the dynamics of large-scale brain modes.

  2. Reviewer #1 (Public review):

    Summary:

    Song et al. introduce an Intrinsic Network Flow (INF) framework, which combines the power of dynamic mode decomposition (DMD) with dimensionality reduction from ICA. They describe brain signals through INF modes reflecting the amplitudes and relative temporal ordering between different INs, relating these INF modes to the principal functional connectivity gradient, and finally showing that the dynamic amplitudes and phases have different properties related to tasks and subject specificity. The most compelling result is that tasks are best discriminated by the relative phases of INF modes rather than their amplitudes, highlighting that task dynamics in the brain may depend on a stable mode organization while modulating the phases of these modes.

    Strengths:

    Using DMD on INs to gain a tractable insight into the spatiotemporal dynamics of the brain is a clear advantage of this work. The authors rightly point out that spatial modes and temporal dynamics are not often unified, and therefore this INF framework nicely addresses that limitation. The framework also accommodates the often neglected subcortex and cerebellum, which is an asset.

    The training/testing split combined with the replication analysis offers another compelling strength: the findings are stable and likely to generalize among healthy young adults. This is also bolstered by the analysis relating INF modes to the principal gradients of FC, which show that different principal gradients can be viewed as different modes with specific phase relationships between the component INs.

    In general, the main argument of the work is convincing: that the brain has stable modes with internal phase relationships (y) that manifest in macroscopic dynamical activity with between-mode phase offsets (θ) collectively constructively/destructively interfering to produce observed brain activity.

    Weaknesses:

    Although this main argument is convincing, the study's framing of INs as the primitives for brain-level activity is not adequately justified. ICA aims to identify latent sources (here, INs) in the data, but taking these as representations of the data from which generative arguments can be made is not presently justified. Several prior works by Braga and colleagues have shown that resting state networks can become more subject-specific and decouple when the single-subject data is sufficiently sampled (Braga et al., 2017, Neuron, among others). This raises a question about whether group-level INFs performing similarly to single-subject INFs may simply be due to the group-level INs that enter the single-subject DMD. Fully subject-specific analyses would be beneficial, but even in HCP it remains unclear whether the INs that could be identified are truly as subject-specific as some of the above-referenced works show they can be in principle.

    Along a similar point, the AR(1) null-model analysis seems like a relatively weak null, and alternative comparison or null models could better justify the primacy of INs. For example, INs could be spatially randomized (the authors already do spatial randomization for their correlational statistics) and then used in DMD to identify modes with similar between-mode phase (θ) relationships. Or more generally, if the dimensionality reduction was in the form of parcels, as commonly is the case in connectomics, do similar DMD modes appear? The authors in the discussion position their findings in relationship to geometric modes and connectome harmonics, which in principle could also enter a DMD algorithm to identify analogous INF modes. The DMD algorithm is what creates the spatiotemporal modes, and the basis for defining y(t) to enter DMD has many possible choices. Additionally, at least one previous analysis did analyze spatiotemporal structure: they used the eigenmodes of a complex Laplacian with propagation delays corresponding to phases between brain regions, and they showed that these modes themselves relate to INs (Xie et al., 2021, NeuroImage). This appears much more generative than the present article. If the authors want to claim the ICA-based INs as the decisive generators, a competition between these alternative choices would decide that result and significantly strengthen the generality of the article's main argument.

    Along this point, the authors spend considerable space in the article describing the correspondence to the principal functional gradients, yet the heart of their argument, as I understood it, is about the primacy of phase relationships in whole-brain dynamics. True, these INF modes now recontextualize the principal gradients in terms of phases, but the most convincing piece is the 23-task classification, which is relegated to the final panel of the final figure. Moreover, other convincing analyses, such as the heritability analysis and its relationship to INF amplitude, are not shown in the main text and instead placed in the supplemental figure. Related to the above comment about alternative bases to define DMD, could the relationship between principal functional gradients and the INF modes be a result of defining those very modes by INs which we already know correspond to principal gradients?

    None of these weaknesses, however, appear to directly undermine the results that the authors do present. They merely suggest that the specific interpretations the authors want to make may outrun what their results can presently support.

  3. Reviewer #2 (Public review):

    Summary:

    The authors address how a stable functional architecture supports flexible neural dynamics. They propose that brain activity measured with fMRI is a superposition of a small number of large-scale propagating modes, or "flows". Each flow is a "wave-like" pattern that engages brain networks in a fixed order, and these flows are the same across people and across rest and task. In this view, different cognitive states arise not from reconfiguring brain networks but from changing the relative timing, or phase, of these fixed flows: which flows are in step at a given moment determines which regions appear active. The strength, or amplitude, of each flow is instead proposed to reflect stable individual traits. To identify the flows, the authors divide fMRI signals into 27 large-scale networks, fit a single linear model describing how activity in these networks evolves from one time point to the next, and take the oscillatory components of this model (13 pairs of complex eigenvectors, via dynamic mode decomposition) as the flows. Using modes fitted to HCP resting-state data, the authors report that the principal flow aligns with the first two cortical FC gradients and with subcortical and cerebellar gradients, and that the next three flows capture lateralized interactions among the default mode, executive and salience networks. Modulating only the flow phases reconstructs activation maps from 23 tasks, and FC predicted from the modes matches task-state FC. Phase features decode the 23 tasks with 89.9% accuracy, compared with 76.6% for network activation and 28.4% for flow amplitude; flow amplitudes are heritable and modestly associated with general cognitive ability.

    Strengths:

    (1) The question is important, and a compact description linking resting-state organization, task activation and individual differences is interesting.

    (2) The analyses use large, public datasets (HCP, a multi-domain task battery and an audiovisual speech dataset), with held-out prediction and an independent replication sample.

    (3) The decoding result is the most novel contribution. Phase features separate conditions whose activation maps are nearly identical, which, if robust, indicates that the timing of large-scale propagation carries state information that activation amplitude does not.

    (4) The separation of group-level modes from subject-level amplitudes provides a common coordinate system for comparing individuals.

    Weaknesses:

    (1) Relation to prior work:

    The stability-flexibility problem and the role of relative phase are central to coordination dynamics and metastability theory (Tognoli & Kelso, Neuron 2014), and connectome models have shown that dynamic phase flows coexist with stable time-averaged FC (Roberts et al., Nat. Commun. 2019). The correspondence between FC gradients, networks, the global signal and a few propagating patterns was also established previously with complex PCA on HCP data (Bolt et al., Nat. Neurosci. 2022). Readers should view the gradient and network results as a replication with a different decomposition.

    (2) Mechanism:

    The framework is descriptive: task phases are estimated from data, not generated by a model. The work therefore does not show how phase alignments are selected or sustained, and claims that cognition "emerges from" phase dynamics should be read as compatibility rather than demonstration.

    (3) Baselines and null models:

    The reconstruction is not compared with the ceiling set by projecting maps onto the 27 networks, or with amplitude-only and random-phase reconstructions. The decoding compares feature sets of different sizes. The only dynamical null is AR(1), and the gain over it during tasks is small.

    (4) Internal consistency:

    FC is argued to be invariant because phases decorrelate over time, yet the task analysis assumes phases held in specific alignments within blocks. These assumptions need to be reconciled.

    (5) (Possible) Physiological confounds:

    HCP data were analyzed without nuisance or global signal regression. The principal flow likely overlaps with the global, arousal-linked signal, whose apparent sequence partly reflects vascular transit delays. Recent work in mice shows that the leading propagating mode of infra-slow cortical activity, identified with the same DMD approach, can be reconstructed from pupil diameter alone, indicating that it largely reflects arousal dynamics (Raut et al., Nature 2025). This caveat extends to the heritability and cognition associations of flow amplitude, and to the interpretation of the flows as separate, cognitively specific processes.

    Overall, the work offers a useful descriptive decomposition and an intriguing decoding result, but its broader theoretical claims are not yet supported and should be read in the context of the prior literature on similar questions.

  4. Reviewer #3 (Public review):

    The manuscript presents a potentially useful description of BOLD fMRI data that seeks to link resting-state organization with task-related activity. I hope that I can help improve the manuscript with my suggestions below, which follow a summary of the main results.

    By applying PCA/ICA/DMD to resting-state data, the authors derive a set of spatiotemporal modes they call intrinsic network flows (INFs). They then present two main results suggesting that these modes provide a useful basis for representing task activity. First, they reconstruct static, group-level activation maps using phases extracted from those same maps while replacing coefficient amplitudes with their mean resting-state values. These reconstructions achieve a mean spatial correlation of 0.73 across 23 conditions. Second, they extract INF amplitude and phase time courses and use GLMs to obtain run-averaged, condition-level classification features. Phase features outperform amplitude and intrinsic-network activation features in leave-one-participant-out testing, with a similar advantage in a second dataset. The independent resting-state basis, held-out-participant testing, and second-dataset replication are strengths. However, I am not yet convinced that these analyses establish the authors' stronger interpretation that cognitive states arise through retiming a shared repertoire of intrinsic flows.

    In the static-map reconstruction, changing a coefficient's phase changes the mixture of a mode's fixed real and imaginary spatial components; no temporal trajectory is reconstructed. Fixing amplitudes imposes a meaningful constraint, but successful reconstruction establishes representational capacity rather than temporal generation. I recommend comparisons with reconstructions retaining both task-derived amplitudes and phases, to quantify the information lost by fixing amplitudes, and with phase configurations exchanged between tasks, to distinguish task-specific reconstruction from recovery of shared spatial structure.

    Both reconstruction and classification should also be compared with alternative complex-valued bases spanning the same spatial subspace, with matched dimensionality and scaling. A clean comparison could use instantaneous spatial projection for both the original and alternative bases, with resting-state reference amplitudes recomputed for each basis. This would test whether the particular organization identified by DMD is advantageous, separately from the contribution of the temporally informed coefficient estimator.

    In classification, the phase time courses are mathematically equivalent to magnitude-normalized real and imaginary coefficients. Their advantage may therefore reflect a useful nonlinear transformation that reduces irrelevant magnitude variability, without uniquely implicating retiming. I recommend comparisons with unnormalized real and imaginary coefficients and with activation features normalized for instantaneous magnitude before GLM estimation, rather than simply standardized across time. Moreover, coefficient estimation incorporates three adjacent time points, whereas an analogous operation is not described for the activation baseline. This issue concerns the comparison with network activation, since INF phase and amplitude share the same estimator; its contribution should be assessed by comparing instantaneous and temporally informed estimates and matching temporal information across representations.

    Finally, an informative control would apply the pipeline to simulated data containing conventional task-evoked responses added to stationary, temporally correlated background activity, without task-dependent retiming. Simulations should approximately match the relevant spatial covariance, temporal structure, and signal-to-noise characteristics. This would test whether ordinary activation changes can produce a similar phase advantage. Task-locked phase trajectories should also be evaluated against this null model, since an additive response can itself change the phase of a complex coefficient without retiming the background activity.

    Until these alternatives are distinguished, the findings support useful task descriptors more directly than a temporal mechanism. If these concerns remain unresolved, I recommend emphasizing the framework's descriptive utility and presenting task-dependent retiming as a hypothesis rather than a demonstrated mechanism of cognition.

  5. Author response:

    We thank the editors and the three reviewers for their careful and constructive reading. We are glad that all three reviewers regard the central empirical findings as valuable, and we are grateful for their insightful comments. In this provisional response, we address central concerns raised in the reviews and outline the plans for the revision.

    Reviewer #1 (Public review):

    Although this main argument is convincing, the study's framing of INs as the primitives for brain-level activity is not adequately justified. ICA aims to identify latent sources (here, INs) in the data, but taking these as representations of the data from which generative arguments can be made is not presently justified. Several prior works by Braga and colleagues have shown that resting-state networks can become more subject-specific and decouple when the single-subject data is sufficiently sampled (Braga et al., 2017, Neuron, among others). This raises a question about whether group-level INFs performing similarly to single-subject INFs may simply be due to the group-level INs that enter the single-subject DMD. Fully subject-specific analyses would be beneficial, but even in HCP it remains unclear whether the INs that could be identified are truly as subject-specific as some of the above-referenced works show they can be in principle.

    We thank the reviewer for this point. We would first like to clarify that our generative description concerns the flows rather than the INs: the framework specifies how observed brain activity and its diverse features underlying cognition (e.g., stable FC topology and diverse task activation) can arise from fixed spatiotemporal structures (flows) (as well as how this common structure is expressed differently across individuals). We do not propose INs as generative primitives of brain activity, although we do not exclude this possibility. Rather, in our framework, INs serve as spatial coordinates, i.e., dominant coherent spatial patterns that efficiently describe the large-scale dynamics, in terms of which these common coordination structures are estimated and described. We agree that defining flows in terms of INs may have suggested that INs are proposed as generative primitives. We will address this point more explicitly where INs are first introduced.

    Regarding subject-specific network organization, we would clarify that the subject-specific INF modes in the predictive comparison were estimated entirely within each subject: INs were identified by ICA applied to that subject’s data alone, and DMD (the dynamic coherence modeling step in our framework) was then performed on the resulting time courses. The comparison therefore does not reflect group-level INs imposed on single-subject DMD. We apologize for the lack of clarity and will make this procedure explicit in the manuscript. Group-level modes nonetheless predicted individuals’ BOLD signals better than these fully subject-specific modes. This result indicates that the coordination structure is stable enough to be shared across individuals, since group-level modes would otherwise reduce rather than improve prediction. This does not, of course, rule out individual-specific flow structure that may emerge with denser within-subject sampling, and we thank the reviewer for prompting us to make the procedure explicit and to discuss this point in the revised manuscript.

    Along a similar point, the AR(1) null-model analysis seems like a relatively weak null, and alternative comparison or null models could better justify the primacy of INs. For example, INs could be spatially randomized (the authors already do spatial randomization for their correlational statistics) and then used in DMD to identify modes with similar between-mode phase (θ) relationships. Or more generally, if the dimensionality reduction was in the form of parcels, as commonly is the case in connectomics, do similar DMD modes appear? The authors in the discussion position their findings in relationship to geometric modes and connectome harmonics, which in principle could also enter a DMD algorithm to identify analogous INF modes. The DMD algorithm is what creates the spatiotemporal modes, and the basis for defining y(t) to enter DMD has many possible choices. Additionally, at least one previous analysis did analyze spatiotemporal structure: they used the eigenmodes of a complex Laplacian with propagation delays corresponding to phases between brain regions, and they showed that these modes themselves relate to INs (Xie et al., 2021, NeuroImage). This appears much more generative than the present article. If the authors want to claim the ICA-based INs as the decisive generators, a competition between these alternative choices would decide that result and significantly strengthen the generality of the article's main argument.

    Following from our response above, we do not regard INs as decisive generators. Thus, the relevant point is whether the flow structure is robust to the choice of spatial basis (e.g., parcel). We addressed this in previous work (Song et al., 2025), in which DMD applied to parcel-level time series (Glasser cortical parcellation with the Cole-Anticevic partition for subcortical and cerebellar regions; 716 ROIs) yielded major modes highly similar to those reported here. Moreover, in that work, in addition to AR(1) null model, DMD outperformed both the linear model and the nonlinear manifold model that had shown the best performance in a recent systematic evaluation of macroscopic brain dynamics models (Nozari et al., 2024). Having established this previously, in this manuscript, we focus on the predictive comparison simple here, to examine how the flows relate to established features of functional organization and task-related dynamics. We now recognize that these previous comparisons are worth to briefly summarize in the revised manuscript.

    As the reviewer notes, geometric eigenmodes and connectome harmonics could also serve as inputs to DMD; they provide a spatial Fourier basis onto which time series can be projected. Because DMD is invariant to the choice of basis, we would expect the dominant INF modes to be recovered when a sufficient number of geometric eigenmodes or connectome harmonics are used. In the Discussion, our comparison with these approaches concerned the decompositions themselves, not excluding their use as alternative inputs to DMD.

    Finally, we thank the reviewer for pointing to Xie et al. (2021), whose complex-Laplacian eigenmodes provide a structurally derived counterpart to our data-driven flows, and we will discuss its relationship to the INF framework in the revised manuscript.

    Along this point, the authors spend considerable space in the article describing the correspondence to the principal functional gradients, yet the heart of their argument, as I understood it, is about the primacy of phase relationships in whole-brain dynamics. True, these INF modes now recontextualize the principal gradients in terms of phases, but the most convincing piece is the 23-task classification, which is relegated to the final panel of the final figure. Moreover, other convincing analyses, such as the heritability analysis and its relationship to INF amplitude, are not shown in the main text and instead placed in the supplemental figure. Related to the above comment about alternative bases to define DMD, could the relationship between principal functional gradients and the INF modes be a result of defining those very modes by INs which we already know correspond to principal gradients?

    We thank the reviewer for this helpful suggestion regarding the structure of the manuscript. We agree that the task classification and heritability analyses are important to our argument. Following the reviewer’s suggestion, we will restructure the Results to give these analyses greater emphasis, moving the heritability analysis to the main text, while condensing the description of the gradient correspondence.

    Regarding the last comment, the principal gradients are likely well represented in the subspace spanned by INs. However, which pattern within this subspace emerges as the principal mode (i.e., INF mode 1) is determined by the temporal dynamics rather than the INs themselves since the coherence modeling step (STEP 2; i.e., group-level DMD) weights the INs according to their coordinated temporal evolution. The correspondence is therefore unlikely to reflect mere spatial configuration of INs. Consistent with this, the principal mode obtained from our previous parcel-level DMD also aligned with the functional gradients (Song et al., 2025), indicating that the gradient correspondence mainly arises from the temporal dynamics.

    Reviewer #2 (Public review):

    Relation to prior work:

    The stability-flexibility problem and the role of relative phase are central to coordination dynamics and metastability theory (Tognoli & Kelso, Neuron 2014), and connectome models have shown that dynamic phase flows coexist with stable time-averaged FC (Roberts et al., Nat. Commun. 2019). The correspondence between FC gradients, networks, the global signal and a few propagating patterns was also established previously with complex PCA on HCP data (Bolt et al., Nat. Neurosci. 2022). Readers should view the gradient and network results as a replication with a different decomposition.

    We thank the reviewer for pointing us to this literature, which we should have engaged with and will incorporate in the revision. As the reviewer notes, our work is closely related to these studies, in particular, the phase-based perspective they established. We would like to clarify how the question we address relates to and is distinct from theirs. Coordination dynamics identifies relative phase as the key variable through which neural ensembles coordinate, emphasizing that coordination patterns form and dissolve continually (Tognoli & Kelso, 2014), and connectome-based modeling has shown that continually reconfiguring wave patterns coexist with a stable time-averaged FC (Roberts et al., 2019). These works concern the ongoing dynamics in coordination, and the latter one showed their relation to time-averaged structure.

    The duality we address is of a related but different kind: it concerns two time-averaged quantities, one of which is largely invariant across tasks (FC topology, i.e., inter-regional correlation) while the other varies substantially (task-evoked activation, i.e., the average response). This coexistence is a long-standing puzzle in neuroimaging, because the prevailing activation-based view identifies task-specific regions, which would in turn require task-specific interactions among them. Our framework addresses this through the phase dynamics of a fixed set of flows, offering a view complementary to this prevailing view underlying task-related analysis.

    In addition, phase here refers to the global phase of each flow, which sets the timing of the pattern as a whole, rather than to the relative phase between individual neural ensembles emphasized in coordination dynamics. As the reviewer’s comment indicates, however, the two are closely related if an INF mode is interpreted as a coordination pattern that recurrently forms across the brain. In the revised manuscript, we will situate our work in relation to coordination dynamics and metastability, state this distinction explicitly, and clarify the contribution of INF. We thank the reviewer for the opportunity to make this connection.

    Regarding Bolt et al. (2022), we agree that our gradient and network results converge with theirs and will state this more clearly in the revision. We see this convergence as reassuring, but in the context of our study, these results also play a further role. Beyond providing a parsimonious description of resting-state organization, the same fixed structure that carries these resting-state features also yields task-evoked activation patterns through phase modulation alone, which is how diverse task states can arise without disrupting intrinsic organization. The correspondence with resting-state features thus serves as the link between resting-state organization and task-evoked activity. We will make this role of the gradient and network results explicit in the revision.

    Mechanism:

    The framework is descriptive: task phases are estimated from data, not generated by a model. The work therefore does not show how phase alignments are selected or sustained, and claims that cognition "emerges from" phase dynamics should be read as compatibility rather than demonstration.

    We thank the reviewer for this point, and we agree with the distinction being drawn. We make no claim about what generates the flows themselves, or about what selects or sustains a particular phase alignment. Accordingly, “emerges from” should be read in a more limited sense: the features tied to task-related cognition, namely task-evoked activation patterns and the information that discriminates cognitive states, can be accounted for by a single generative principle, which is “phase modulation of a fixed set of flows.” We will define this usage explicitly in the Introduction and Discussion, and state that the mechanisms governing phase selection remain to be established.

    Baselines and null models:

    The reconstruction is not compared with the ceiling set by projecting maps onto the 27 networks, or with amplitude-only and random-phase reconstructions. The decoding compares feature sets of different sizes. The only dynamical null is AR(1), and the gain over it during tasks is small.

    We thank the reviewer for these suggestions, which we will adopt. For reconstruction, we agree that comparison with the ceiling set by the 27-network projection and random-phase reconstruction will clarify how much of the attainable and chance reconstruction accuracy is achieved by phase alone, and we will report it in the revision.

    For decoding, the aim of the analysis was to assess how much task information each representation carries, and the dimensionality of each feature set follows from the representation itself. Nonetheless, we agree that a comparison with amplitude at matched dimensionality would be more directly interpretable: we already observed that using only the cosine or only the sine components (13 features each) yielded substantially higher accuracy than amplitude, which we will report this in full in the revision.

    Regarding AR(1), we agree the margin during task is small, and we will present it with that caveat. Two points bear on its interpretation. First, this analysis was intended to test whether modes estimated from resting-state data retain predictive value during task performance, not to maximize the gain over a null. Second, the size of the margin is specific to the MDTB data, which were band-pass filtered at 0.008–0.15 Hz; under this filtering, lag-1 autocorrelation is high and AR(1) is already a strong baseline (R2 = 0.828). The task-state gain, although small, was consistent across participants (R2 = 0.832 versus 0.828, p = 1.4 × 10-6, N = 18, paired t-test). We will make this dependence on preprocessing explicit and will note that the central claims of the manuscript rest on the reconstruction and decoding analyses rather than on the size of this prediction margin.

    Internal consistency:

    FC is argued to be invariant because phases decorrelate over time, yet the task analysis assumes phases held in specific alignments within blocks. These assumptions need to be reconciled.

    We thank the reviewer for raising this point, which we agree deserves explicit clarification. The two assumptions are compatible once task activation is understood as a time-averaged quantity rather than a fixed spatial configuration. In our framework, the same intrinsic flows remain in place and continue to evolve during task performance. The “specific alignment” captured by a task activation map reflects a systematic bias in the phase distribution over the course of the task block, i.e., a tendency toward certain alignments rather than phases held fixed. As derived in Supplementary Result 3, the correlation structure is primarily determined by the fixed relative lags between regions and by the mode amplitudes. Thus, as long as the flows continue to evolve, the phase-dependent contributions to the correlation largely integrate out, and FC topology is preserved. We will clarify this distinction in the revised manuscript.

    (Possible) Physiological confounds:

    HCP data were analyzed without nuisance or global signal regression. The principal flow likely overlaps with the global, arousal-linked signal, whose apparent sequence partly reflects vascular transit delays. Recent work in mice shows that the leading propagating mode of infra-slow cortical activity, identified with the same DMD approach, can be reconstructed from pupil diameter alone, indicating that it largely reflects arousal dynamics (Raut et al., Nature 2025). This caveat extends to the heritability and cognition associations of flow amplitude, and to the interpretation of the flows as separate, cognitively specific processes.

    We thank the reviewer for this point. We note that the HCP data were denoised with ICA-FIX, although global signal regression was not applied. We agree that arousal and other physiological factors likely contribute to flow dynamics, and we view this as compatible with the framework rather than as a confound of its main argument. In the manuscript, we primarily consider that phase alignment reflects task-related cognitive input, but it does not exclude contributions from other physiological influences, such as arousal or respiratory signals. The central claim, that task states are characterized by phase modulation of a stable set of flows, holds regardless of the (additional) sources that drive this modulation. We also note that Raut et al. (2025) interpret arousal-related dynamics not as mere noise confounding the observation but as a process that actively organize/accompany with this large-scale spatiotemporal structure. We will discuss this point, including Raut et al. (2025), more thoroughly in the revised manuscript.

    Reviewer #3 (Public review):

    In the static-map reconstruction, changing a coefficient's phase changes the mixture of a mode's fixed real and imaginary spatial components; no temporal trajectory is reconstructed. Fixing amplitudes imposes a meaningful constraint, but successful reconstruction establishes representational capacity rather than temporal generation. I recommend comparisons with reconstructions retaining both task-derived amplitudes and phases, to quantify the information lost by fixing amplitudes, and with phase configurations exchanged between tasks, to distinguish task-specific reconstruction from recovery of shared spatial structure.

    We thank the reviewer for this precise characterization, which we accept. The reconstruction analysis was designed to test representational sufficiency, i.e., whether 13 phase values suffice to represent task activation maps, rather than to reconstruct temporal trajectories. We will state this scope more explicitly in the main text. We agree that quantifying the information lost by fixing amplitudes would be informative, and we will report this in the revision. We will also report reconstructions with phase configurations exchanged between tasks, which will help distinguish task-specific reconstruction from the mere recovery of shared spatial structure.

    Both reconstruction and classification should also be compared with alternative complex-valued bases spanning the same spatial subspace, with matched dimensionality and scaling. A clean comparison could use instantaneous spatial projection for both the original and alternative bases, with resting-state reference amplitudes recomputed for each basis. This would test whether the particular organization identified by DMD is advantageous, separately from the contribution of the temporally informed coefficient estimator.

    We thank the reviewer for this suggestion, which offers a direct way to isolate the contribution of the flow organization we identified. We will compare reconstruction and classification between the INF basis and alternative complex-valued bases spanning the same spatial subspace, with matched dimensionality and scaling. This comparison will more clearly demonstrate whether our INF modes confer an advantage beyond the representation itself.

    In classification, the phase time courses are mathematically equivalent to magnitude-normalized real and imaginary coefficients. Their advantage may therefore reflect a useful nonlinear transformation that reduces irrelevant magnitude variability, without uniquely implicating retiming. I recommend comparisons with unnormalized real and imaginary coefficients and with activation features normalized for instantaneous magnitude before GLM estimation, rather than simply standardized across time. Moreover, coefficient estimation incorporates three adjacent time points, whereas an analogous operation is not described for the activation baseline. This issue concerns the comparison with network activation, since INF phase and amplitude share the same estimator; its contribution should be assessed by comparing instantaneous and temporally informed estimates and matching temporal information across representations.

    We thank the reviewer for this insightful point. We agree that the effect of this nonlinear transformation should be tested to strengthen our argument. We will include this analysis in the revision. We also agree on the need to match temporal information across representations, which would further strengthen our point. We will compare classification performance between phase and network activation features under matched temporal information.

    Finally, an informative control would apply the pipeline to simulated data containing conventional task-evoked responses added to stationary, temporally correlated background activity, without task-dependent retiming. Simulations should approximately match the relevant spatial covariance, temporal structure, and signal-to-noise characteristics. This would test whether ordinary activation changes can produce a similar phase advantage. Task-locked phase trajectories should also be evaluated against this null model, since an additive response can itself change the phase of a complex coefficient without retiming the background activity.

    We thank the reviewer for this thoughtful suggestion. We would like to clarify that our claim is not that task responses are non-additive. As the reviewer rightly notes, an additive response can itself be described as a phase modulation. Rather, our claim is that phase modulation of a fixed set of resting-state flows is sufficient to describe how task-related features are generated. In light of this, we believe the most informative null (simulation) is one in which task states involve patterns not organized by the resting-state flows. This is directly tested by the comparison with alternative complex bases spanning the same subspace, which the reviewer helpfully suggests above.

    References

    Beckmann, C. F., Mackay, C. E., Filippini, N., & Smith, S. M. (2009). Group comparison of resting-state FMRI data using multi-subject ICA and dual regression. Neuroimage, 47(Suppl 1), S148.

    Bolt, T., Nomi, J. S., Bzdok, D., Salas, J. A., Chang, C., Thomas Yeo, B., Uddin, L. Q., & Keilholz, S. D. (2022). A parsimonious description of global functional brain organization in three spatiotemporal patterns. Nature neuroscience, 25(8), 1093-1103.

    Calhoun, V. D., Adali, T., Pearlson, G. D., & Pekar, J. J. (2001). A method for making group inferences from functional MRI data using independent component analysis. Human brain mapping, 14(3), 140-151.

    Erhardt, E. B., Rachakonda, S., Bedrick, E. J., Allen, E. A., Adali, T., & Calhoun, V. D. (2011). Comparison of multi‐subject ICA methods for analysis of fMRI data. Human brain mapping, 32(12), 2075-2095.

    Kutz, J. N., Brunton, S. L., Brunton, B. W., & Proctor, J. L. (2016). Dynamic mode decomposition: data-driven modeling of complex systems. SIAM.

    Nozari, E., Bertolero, M. A., Stiso, J., Caciagli, L., Cornblath, E. J., He, X., Mahadevan, A. S., Pappas, G. J., & Bassett, D. S. (2024). Macroscopic resting-state brain dynamics are best described by linear models. Nature biomedical engineering, 8(1), 68-84.

    Raut, R. V., Rosenthal, Z. P., Wang, X., Miao, H., Zhang, Z., Lee, J.-M., Raichle, M. E., Bauer, A. Q., Brunton, S. L., & Brunton, B. W. (2025). Arousal as a universal embedding for spatiotemporal brain dynamics. Nature, 647(8089), 454-461.

    Raut, R. V., Snyder, A. Z., Mitra, A., Yellin, D., Fujii, N., Malach, R., & Raichle, M. E. (2021). Global waves synchronize the brain’s functional systems with fluctuating arousal. Science advances, 7(30), eabf2709.

    Roberts, J. A., Gollo, L. L., Abeysuriya, R. G., Roberts, G., Mitchell, P. B., Woolrich, M. W., & Breakspear, M. (2019). Metastable brain waves. Nature communications, 10(1), 1056.

    Song, Y., Kim, P. S., Philip, B. A., & Kim, T. (2025). Large-scale Signal Propagation Modes in the Human Brain. Neuroimage, 318, 121357.

    Tognoli, E., & Kelso, J. S. (2014). The metastable brain. Neuron, 81(1), 35-48.

    Xie, X., Cai, C., Damasceno, P. F., Nagarajan, S. S., & Raj, A. (2021). Emergence of canonical functional networks from the structural connectome. Neuroimage, 237, 118190.

  6. Cognition emerges from phase dynamics of intrinsic coordination https://www.biorxiv.org/content/10.64898/2026.03.26.714488v1

  7. Cognition emerges from phase dynamics of intrinsic coordination https://www.biorxiv.org/content/10.64898/2026.03.26.714488v1