Weight geometry governs functional memory in complex systems
Elkaïoum M. Moutuou, Habib Benali · 2026
Read Weight geometry governs functional memory in complex systems by Elkaïoum M. Moutuou, Habib Benali (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-25826.txt.
Introduction
Does a network remember its past because of how it is wired, or because of how strongly its links pull? Elkaïoum Moutuou and Habib Benali test that question in 2026 across thirty four real networks, from gene regulation and neural circuits to ecology, society, and transport. A thermodynamic account of information flow measures how deeply earlier interactions stay woven into present behaviour. Shuffle the link strengths while keeping the wiring fixed, and the memory shallows. Real weights organise memory at greater depth than random ones, in every domain studied.
The memory patterns collapsed onto just four recurrent dynamical organisations, a strikingly low dimensional order. Targeted null models then separated the ingredients. Weight geometry governs memory depth. Middle scale structure shapes how memory spreads across scales. Directionality tunes how sensitive the cascade is to damage. The comparison doubles as a test for whether a network's weights carry genuine functional meaning.
Catalog record
| CREATOR | Elkaïoum M. Moutuou, Habib Benali |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 108 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.25826
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Weight geometry governs functional memory in complex systems Elkaïoum M. Moutuou, Habib Benali Abstract Complex systems, from gene regulatory networks to neural circuits and transportation infrastructures, exhibit rich functional behaviour that topology alone does not capture. Here we show that functional memory exhibits a universal organisational regularity: in every biological, ecological, social, and technological domain studied, real interaction strengths organise memory at greater hierarchical depth than random weight assignment on the same topology, across thirty-four networks spanning several orders of magnitude in size and density. Using a thermodynamic description of multiscale information flow, we quantify how memory is distributed across path lengths and show that functional memory organisation collapses onto four recurrent dynamical organisations, revealing an intrinsically low-dimensional structure. Comparing each network against null models that selectively perturb weighted transport geometry, mesoscale structure, and directionality reveals that these ingredients contribute distinct and non-equivalent roles: weight geometry systematically governs memory depth, mesoscale structure shapes memory organisation across scales, and directionality modulates the sensitivity of the cascade to structural perturbation. The same comparison provides an operational criterion for whether network weights encode genuine functional interaction structure. These results establish weighted transport geometry as a primary organiser of functional memory and show that weighted interactions carry dynamical structure that binary topology alone cannot recover. I Introduction What principle governs how living, social, and engineered systems store the memory of their past interactions? Network science has traditionally sought the answer in structural properties of the wiring diagram, including degree heterogeneity, modularity, and spectral organisation Simon1991 ; Newman2003 ; Bianconi2023 . This topological perspective has been productive, yet it implicitly assumes that the functional organisation of a system is largely determined by which connections exist. Even minimal systems illustrate the distinction between structure and the dynamics of historical integration. Networks with similar static structural properties can nevertheless compress causal history at markedly different rates, revealing differences invisible to conventional structural observables (Supplementary Section 1). Here we show that this assumption fails in one precise and empirically falsifiable respect: the geometry of interaction strengths consistently shapes how deeply past interactions remain integrated into present dynamics. In directed networks, information propagates locally along directed paths and integrates contributions from progressively earlier interactions. The depth over which past interactions continue to influence a node’s present state is finite and constrained by both structure and physical cost. Structural observables characterise which paths exist, but they do not resolve how path contributions are weighted across temporal depth, how historical influence is suppressed as the system evolves, or how this suppression is distributed across scales. The organisation of functional memory is encoded precisely in these quantities, and they lie beyond the reach of topological analysis alone. Across thirty-four empirical networks spanning gene regulation, protein interaction, neural connectomes, ecological webs, social and information systems, and transportation infrastructure, randomising edge weights while preserving binary topology reduces functional memory depth. The effect holds across all domains and over three orders of magnitude in network size. The mechanism follows directly from the structure of directed walks: functional memory is sustained by long weighted paths, and real systems correlate strong interactions along coherent directed pathways, producing multiplicative amplification of deep contributions that random weight configuration suppresses. Topology determines which paths exist, whereas the geometry of directed transport determines how strongly distant causal histories remain integrated. Figure 1: Thermodynamic cascade of hierarchical functional memory. a , Example directed network used to illustrate the formalism. b , Thermalization cascade generated by progressively suppressing long-memory contributions as the inverse temperature increases. Starting from the infinite-history equilibrium state ( \mathrm{EMD}=\infty , dark red), the effective memory depth \mathrm{EMD}(t) decreases in discrete steps as successive thermalization times \tau_{n} are crossed (vertical dashed lines). Each horizontal band corresponds to the interval [\tau_{n},\tau_{n-1}) during which the depth- n propagator T_{t}^{(n)} provides the effective description of probability flow. Colour progressively shifts from long-memory (red) to short-memory (blue) dynamics. The hierarchical memory spectrum \mu_{l}=(\tau_{l-1}-\tau_{l})/\tau_{0} records how total thermalization time is distributed across depths, while the non-Markovian ratio \mathrm{NMR}=\tau_{1}/\tau_{0} measures the persistence of long-memory relative to the freezing time \tau_{0} . c , Memory flow networks generated by the full propagator T_{t}^{(\infty)} (top row) and by the effective truncated propagator at the dominant active depth (bottom row) at successive macroscopic times along the cascade. Blue edges represent direct Markovian flow, whereas red edges represent indirect multi-hop contributions from long-memory. As thermalization proceeds, deep indirect flows are progressively compressed into shallower effective transitions, producing a systematic coarsening of the flow structure. At late times the dynamics collapse onto persistent local self-loops corresponding to frozen local-memory regime. To quantify these effects, we introduce a thermodynamic framework in which a macroscopic inverse temperature controls the depth over which past interactions contribute to present dynamics, driving a cascade of memory-depth transitions whose infinite-memory limit is identified with an equilibrium flow associated with a Kubo–Martin–Schwinger (KMS) state Moutuou2025a ; Moutuou2025b , the equilibrium reference state of the theory. Three observables derived from this cascade define a signature of functional memory organisation: the effective memory depth, which tracks the instantaneous causal horizon; the hierarchical memory spectrum, which encodes how memory loss is distributed across depth scales; and the memory compression rate, which quantifies how rapidly deep causal history compresses into shallow effective dynamics. Together, these observables resolve distinctions invisible to graph-theoretic analysis alone. Comparing each network against null models that selectively perturb weighted transport geometry, mesoscale structure, and edge directionality reveals that these ingredients contribute distinct and non-equivalent roles: weight geometry systematically governs memory depth, mesoscale wiring shapes memory organisation across scales, and directionality modulates dynamical response without universally determining memory depth. Despite their structural diversity, the networks collapse onto a small number of recurrent organisations of functional memory that remain invisible to topological analysis alone. II Results 1 Functional memory organisation as hierarchical resistance to compression We characterise each network through its memory architecture: the hierarchical organisation of how causal history is retained, distributed, and progressively compressed as the network evolves. Structural observables characterise connectivity and path organisation, but they do not resolve how contributions from different depths are weighted, suppressed, or redistributed across scales. Functional memory organisation is encoded precisely in these quantities. To make memory architecture measurable, we model directed networks as multigraphs whose asymmetric edges encode causal interactions. Edge weights w_{j\to i}=A_{ij}>0 are either continuous positive reals, representing interaction strengths, or positive integers, representing multiplicities of parallel directed interactions; both cases are handled uniformly by the framework below. We introduce a family of n -memory propagators T^{(n)}_{t} : transition matrices of a non-Markovian random walk that, at macroscopic time t , samples directed walks of length at most n , weighted by the energetic cost e^{-\beta(t)|\gamma|} of sustaining causal correlations across depth |\gamma| . These propagators admit the compact representation R^{(n)}_{t}=\sum_{k=0}^{n}e^{-\beta(t)k}A^{k}, (1) with T^{(n)}_{t} obtained by column normalization, where A is the weighted adjacency matrix with entry A_{ij}=w_{j\to i} for each directed edge j\to i (see Methods). As the memory depth increases, the propagators converge toward an infinite-history equilibrium propagator T^{(\infty)}_{t} defined by R^{(\infty)}_{t}:=\lim_{n\to\infty}R^{(n)}_{t}, representing the fully integrated dynamics of the network for \beta(t) exceeding the critical inverse temperature \beta_{c}=\log\rho(A) (Methods). This equilibrium flow is associated with KMS states Moutuou2025a ; Moutuou2025b and provides the thermodynamic reference against