The geometry of weights governs the functional memory of complex systems ✛ ✦
Context
Why do some systems remember the past more deeply than others? Traditionally, network science explained this through topology: number of connections, modularity, spectral properties. However, two networks with the same structure can compress interaction history differently. New research shows that the key role is played by weight geometry — how connection strengths are distributed along directed paths. This discovery challenges the exclusive role of binary architecture.
Methods
For 34 real networks — from gene regulatory to air traffic — the authors constructed a thermodynamic cascade inspired by the ideas of statistical physics. They introduced the concept of inverse temperature, suppressing the contribution of long paths, and computed thermalization times when depth-n memory becomes indistinguishable from equilibrium. They compared real networks with three null models: the geometry fluctuations model (GFM), randomly shuffling weights on fixed topology; the directed configuration model (DCM), destroying mesostructure; and the polarity noise model (PNM), flipping edge directions. For evaluation, they used Wasserstein distance and Gaussian mixtures to classify memory regimes. The formalism was based on Markov chains on directed graphs.
Results
The main empirical result: in 32 out of 34 networks, real weights systematically increase memory depth compared to GFM. The mechanism is simple: strong interactions correlate along directed paths, enhancing the contribution of the distant past, which is destroyed upon randomization. Exceptions — US air traffic (memory determined by hub topology) and one food web with model weights. Directedness of edges modulates memory sensitivity to perturbations: in neural networks it suppresses deep memory, while in food webs it sustains it. Networks fall into four dynamic classes: diffuse (broad memory spectrum, e.g., E. coli), reactive (deep memory with concentrated compression, many food webs and connectomes), stable (deep memory with diffuse compression), and impulsive (memory collapses on a single scale, as in plant photosystem I and US air traffic). Quantitatively, the non-Markovian ratio NMR (measure of long memory retention) for real networks is on average 12% higher than in GFM, and for GFM-dominated systems the difference reaches 0.115.
Implications
The work introduces a measurable concept of functional memory and shows that binary topology does not reconstruct memory organization. Weight geometry emerges as the primary organizer, and directedness as a modulator. This calls for a revision of many network models in biology and engineering where weights are often discarded during analysis. The proposed GFM test serves as an operational criterion: if weight randomization does not alter memory depth, the network's weights carry no functional information (as in modeled carbon flows). From a thermodynamic perspective, memory suppression is analogous to entropy increase, linking memory organization to the second law of thermodynamics.
Future development
Future development of the topic involves moving to temporal networks where weights evolve, and to multilayer structures where different interaction channels can give rise to competing or reinforcing memory regimes. The connection with algorithmic complexity and Bennett's logical depth may explain the low dimensionality of functional memory space.
Impact
The results will influence the analysis of biological networks, the design of resilient infrastructures, and the understanding of cognitive processes where history integration is crucial.
Next steps
Verify the prediction about memory recovery upon complete edge reversal in systems with positive polarity response. Investigate whether the geometric effect persists in networks with variable topology.
Key open problems
The work connects functional memory to the problem of measuring complexity: the thermalization cascade is analogous to logical depth, and its observables are candidates for universal indicators of organization.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors introduce the concept of functional memory, measured by the thermalization time of a thermodynamic cascade, and show on 34 real networks that the geometry of weights, not binary topology, determines memory depth. They compare real networks with null models, including GFM, and identify four dynamic memory classes.
Where its strength is
The strength of the work lies in the operational criterion of the GFM test, which allows checking whether weights carry functional information, and in the systematic empirical analysis linking weight geometry to memory.
Where it could go next
- It might be interesting to see how the introduced concept of functional memory relates to memory in biological systems, where the order of elements, not just their presence, determines function, as in the work on memory not in cell number but in their order.based on: Memory Lies Not in Cell Count but in Cell Order arXiv:2605.17199
- A related approach is where weak agents in networks crystallize inequality through link dynamics; perhaps similar mechanisms could be linked to weight geometry and memory in social or technological networks.based on: How Weak AI Agents Take Over Networks: The Crystallization of Inequality arXiv:2607.01148
- There is a paper in the archive on the geometry of energy release in batteries; it might be interesting to compare how weight geometry governs memory in networks with geometry governing dynamics in other physical systems.based on: A Battery with Character: The Geometry of Energy Delivery arXiv:2607.16645
Worth trying
- Test whether the link between weight geometry and memory holds in networks with time-varying topology, using ideas from the work on memory in cell order.
- Apply the GFM test to networks from the work on weak AI agents to see how inequality affects functional memory.
- Investigate whether measures of energy release geometry can be transferred to the analysis of weight geometry in biological networks.
Why this may matter
This could be important for understanding how biological and technical systems process information, and for developing methods to test the functional significance of weights in network models.
Nearby works in our archive
- How to Weigh the Universe with Color-Coded Tagssimilarity 0.59
- Memory Lies Not in Cell Count but in Cell Ordersimilarity 0.58
- Hidden Symmetry: How Particles Negotiate Their Own Rulessimilarity 0.58
- The Black Hole's Information Backpack: How Memory Shifts the Tone of Gravitational Ringingsimilarity 0.58
- Cosmic Masers Rewrite the Speed of the Universesimilarity 0.58
- How Weak AI Agents Take Over Networks: The Crystallization of Inequalitysimilarity 0.58
- A Battery with Character: The Geometry of Energy Deliverysimilarity 0.57
The archive is sparse around this work: the closest review is only 0.59, against a usual 0.64. It looks like few have drilled nearby.
Original abstract
as written by the authors arXiv:2606.25826
Complex systems, from gene regulatory networks to neural circuits and ecological food webs, exhibit rich functional behaviour that topology alone does not capture. Yet functional complexity remains difficult to quantify independently of structural organisation. Here we introduce a thermodynamic framework in which functional complexity is characterised through the hierarchical organisation of functional memory, quantifying how the influence of past interactions is distributed and progressively compressed across scales. Across thirty-four empirical networks spanning biological, ecological, social, technological, and biophysical systems and several orders of magnitude in size and density, real interaction strengths organise functional memory at greater hierarchical depth than random weight assignment on the same topology in every domain studied. The framework further reveals that functional memory occupies a remarkably low-dimensional space, collapsing onto four recurrent dynamical organisations. Comparisons with null models that selectively perturb weighted transport geometry, mesoscale wiring, and directionality show that these structural ingredients play distinct roles: weighted transport geometry systematically governs memory depth, whereas mesoscale wiring organises memory across scales and directionality modulates the response of the cascade to structural perturbation. The same comparisons provide an operational criterion for determining whether network weights encode functionally meaningful structure beyond topology. These results establish weighted transport geometry as a primary organiser of functional memory and provide a quantitative framework for studying functional complexity in directed weighted networks.
Plant photosystem I and US air traffic — two systems with nothing in common in scale or nature — compress memory equally sharply, as if a giant hub and a molecular machine follow the same principle.
In Peter Watts's novel 'Blindsight,' alien beings lack consciousness, but their distributed networks exhibit complex memory without a center — like the diffuse regimes described here.
Key numbers
- fraction of networks with effect: 32 out of 34
- number of dynamic regimes: 4
- average NMR increase relative to GFM: 12%
- maximum compression peak (US air traffic): 70.4
- number of networks studied: 34