New fluctuation-dissipation test maps hidden memory effects in brain dynamics
A new computational framework called FDTest uses violations of a generalized fluctuation-dissipation theorem to measure memory effects in brain dynamics, offering a way to quantify how much a region’s future activity depends on past signals from the wider network. The method targets a long-standing problem in neuroscience: brain activity is never purely local, yet the field still needs principled tools to separate a region’s intrinsic behavior from the influence of distributed interactions. In this study, the authors treat memory not as a vague metaphor but as a measurable dynamical property. When a regional signal depends on history embedded in the network, the system no longer behaves as if each node is acting in isolation.
Wakefulness carried stronger memory effects than deep sleep.
Why memory effects matter for whole-brain organization
The central idea rests on a familiar but difficult question in systems neuroscience: how does the brain shift between segregated processing within individual regions and integrated processing across the network? The FDTest is built to detect those hidden dependencies by measuring departures from fluctuation-dissipation behavior in multidimensional systems. In practical terms, the method asks whether spontaneous fluctuations and the system’s response to perturbation remain aligned in a way expected from simpler, memoryless dynamics. When they do not, the mismatch can reveal that past network activity is still shaping current behavior.
This is not just a mathematical exercise. In a biological network as dense and recurrent as the human brain, past states are constantly carried forward through ongoing synaptic and circuit-level interactions. A region that appears locally driven may still be receiving delayed influence from elsewhere, and that history can alter how the system responds to new input. By formalizing this logic, FDTest provides a way to quantify local memory effects in a framework that is compatible with whole-brain modeling.
Whole-brain models linked memory structure to connectivity
The first part of the work applied the FDTest to whole-brain models fitted to human neuroimaging data. The result was straightforward and biologically plausible: the brain’s memory structure reflected its underlying connectivity. Highly connected regions showed stronger memory effects, consistent with the idea that hubs and integrative nodes sit in a denser web of recurrent influences. More isolated regions behaved more independently, with weaker evidence of past network information shaping their activity.
This pattern matters because connectivity is not merely an anatomical map. Dense interregional coupling creates the conditions for accumulation, delay, and feedback, all of which can leave a memory trace in ongoing dynamics. A region embedded in a heavily connected network cannot be understood only by its immediate state at a single time point. Its future depends on where it has been, and on what the rest of the system has been doing around it.
Wakefulness showed richer interregional dependence than deep sleep
The individualized analysis separated subjects during wakefulness and deep sleep, and the state contrast was consistent across the models: memory effects were stronger when participants were awake. In the sleep condition, the network appeared less integrated and more locally driven, with reduced interregional dependence. The implication is not that the sleeping brain is inactive. It is that the architecture of influence changes, and the system moves toward a configuration in which current activity is less shaped by accumulated network history.
That shift fits the broader physiology of sleep, where conscious experience recedes and global integration weakens. If wakefulness depends on coordinated exchanges across distributed nodes, stronger memory effects may reflect exactly that richer coupling. During deep sleep, by contrast, weaker memory effects suggest a brain that is less responsive to wide-network history and more dominated by local dynamics. The study positions this difference as a dynamical marker of brain state, not simply a byproduct of altered vigilance.
The biological logic is direct. Recurrent interactions can preserve information across time, and when those interactions become less influential, the system should show a reduced memory signature. FDTest turns that expectation into a measurable quantity.
A new marker for hidden structure in neural dynamics
The value of the method lies in what it can reveal beyond static summaries of connectivity. Standard network measures can describe who is connected to whom, but they do not fully capture how those connections shape temporal dependence in the activity itself. FDTest is designed to probe the hidden structure of neural dynamics by testing whether the present can be explained without accounting for network-fed history. In a system as adaptive as the brain, that distinction is critical.
Because the approach was applied to both whole-brain models and individualized wake-sleep comparisons, the findings support more than one level of interpretation. At the model level, memory effects tracked connectivity. At the state level, memory effects were stronger in wakefulness than in deep sleep. Those two observations point in the same direction: integration across the network leaves a measurable signature in temporal dynamics, and that signature changes with consciousness-related state.
The work also benefits from practical transparency. The code is openly available on Zenodo, and the sleep dataset is publicly accessible through GitHub, making the framework reproducible for groups interested in extending the analysis to other brain states or other kinds of large-scale dynamical systems. The paper was published as an uncorrected proof, so the core message should be read with the usual caution attached to a proof stage publication. Even so, the conceptual advance is clear: memory effects can be formalized as a testable property of brain dynamics rather than treated as an informal description.
A remaining limitation is built into the scope of the article itself. The framework is presented through computational modeling and state comparison, not through direct clinical validation in patient populations. Future work will need to determine how robust the FDTest is across other datasets, how it behaves in disease, and whether it can distinguish among different pathological disruptions of network integration. For now, the study supplies a mathematically grounded way to ask a sharper question of brain activity: how much of the present is still being carried by the past?