ABSTRACT. Traditional warehouse planning often relies on heuristic design and human intuition, which may not correctly capture the complexity of interactions among multiple operational zones and material flows. As warehouse systems scale, these approaches become limited in identifying efficient layouts and routing strategies increasing time and space complexities.

This study proposes a graph-based framework to represent warehouse operations as a weighted and constrained network. Functional areas such as receiving, storage, production, and shipping are modeled as nodes, while connections between them are represented as edges with weights corresponding to distance, travel time, and handling constraints. This formulation enables the application of graph-theoretic methods to evaluate shortest paths, flow efficiency, and structural properties.

this model supports systematic exploration of alternative layouts and routing strategies that are difficult to assess through manual planning. Preliminary insights indicate that network-based representations can help identify critical nodes and reduce material handling distances.

This work bridges industrial engineering and complex systems by demonstrating how graph-based models can improve warehouse design and operational decision-making. Future work will incorporate dynamic demand and stochastic behavior to capture real-world variability

ABSTRACT. Networked surge systems can exhibit similar aggregate load while differing in how stress is organized across local subsystems. We study California hospital facility-level strain patterns from historical HHS data augmented with a geographic substitute graph. We compare global positional/correlational structural descriptors with event-conditioned descriptors computed on the shock-relevant receiver set. Global PI/CI descriptors add little to out-of-window secondary-shock prediction, whereas event-conditioned PI/CI provides modest but consistent additional signal beyond local load baselines. These results suggest that structural information is not a universal network-level predictor; it becomes useful only when measured at the operational scale where downstream failure is realized.

ABSTRACT. Complex systems often reveal their interaction structure only indirectly, through collective dynamics observed over time. We ask whether persistent homology can recover such structure from short windows of activity, and whether it captures information not available to standard pairwise synchrony measures. As a testbed, we simulate noisy Kuramoto oscillators on undirected graph families across a range of coupling strengths. Trajectories are divided into overlapping windows, converted into frequency-phase point clouds, and summarized using persistence landscapes from H0 (connected components) and H1 (loops) homology groups. Linear support vector machines (SVMs) are trained on phase-locking value (PLV) features, topological features, and their combination, to retrieve the network graph. We find that when combined, persistent homology contributes non-redundant information precisely when PLV is most uncertain: accuracy gains increase monotonically with PLV prediction uncertainty, reaching 20–28% in the highest-uncertainty regime across graph family sets. This identifies a data-driven criterion for predicting when latent interaction structure is recoverable from dynamical topology, with direct implications for topology-aware graph inference from observed dynamics. We further consider extensions to Stuart-Landau oscillators, Vicsek models, and empirical human coordination data, where interaction structure is partially known.

ABSTRACT. Emergence describes how complex, macro-level properties arise from micro-level interactions, despite being absent in the individual components. Formally introduced by E. Hoel [1], this concept is grounded in the framework of causal emergence (CE) and Effective Information (EI). Understanding emergence is critical as it explains how nature self-organises: it is how consciousness arises from neurons, or temperature emerges from molecular dynamics. Hoel's framework proves that scaling a system changes its causal rules, meaning a coarser description can exhibit stronger, more structured causal relationships than its microscopic baseline. The critical mechanism for this scale transition is the coarse-graining function, which maps high-dimensional microscopic states to lower-dimensional macroscopic representations. Finding this optimal function requires searching a combinatorially large space, a task that is computationally daunting. The current state-of-the-art, the Neural Information Squeezer (NIS) [2], uses neural networks for this mapping but scales poorly due to its reliance on brute-force search. To resolve this bottleneck, we replace the brute-force search with a two-stage spectral identification method requiring only a single NIS training. We first project observations onto a PCA signal subspace to discard noise, then identify the optimal dimension via the Shannon effective rank of the pipeline's Jacobian singular spectrum. A bridge identity further connects this spectral approach to Zhang’s SVD-based reversibility measure [3]. Evaluated on six synthetic benchmarks, our spectral estimate identifies the coarse-grained function 5 to 158 times faster than existing methods.

References: [1] Hoel, E. P., Albantakis, L., & Tononi, G. (2013). Quantifying causal emergence shows that macro can beat micro. PNAS, 110(49), 19790-19795. [2] Zhang, J., & Liu, K. (2023). Neural Information Squeezer for causal emergence. Entropy, 25(1), 26. [3] Zhang, J., Tao, R., Leong, K. H., Yang, M., & Yuan, B. (2025). Dynamical Reversibility and A New Theory of Causal Emergence based on SVD. arXiv:2402.15054v6.

ABSTRACT. A realistic representation of firms and households is essential for large-scale economic models. While robust methods exist for generating synthetic household populations, constructing synthetic populations of firms and, crucially, their supply chain linkages remain significantly more challenging. In this work, we introduce a flexible method for generating synthetic supply networks that reproduce both the known properties of firm-level supply networks and the aggregated input-output tables commonly used in macroeconomic models. Our method is computationally efficient, fully reproducible, and relies only on publicly available data, making it straightforward to extend and apply across settings.

ABSTRACT. Organizations increasingly embed AI systems into high-stakes decision workflows, yet the central governance question is not whether to adopt AI but how it enters the workflow. We develop a model of AI-enabled decision making in which a possibly non-Bayesian human optimizes organizational utility with or without AI support. We compare three configurations: the human decides alone (Ignore), the human sees the AI recommendation before deciding (Consult), or the organization delegates to the AI (Delegate).

We first establish a negative benchmark: without verification, advisory use cannot strictly outperform the best corner solution, and with any consultation or override friction it collapses to a corner solution dominated by Ignore or Delegate.

Verification overturns this impossibility. When a costly verification process is available, human-AI disagreement becomes an endogenous signal identifying the cases with the highest posterior error risk. For intermediate verification costs, the optimal policy verifies disagreements but not agreements, making advisory use strictly optimal—not by averaging two opinions, but by using disagreement as the trigger for selective verification. The mechanism neutralizes biased human overrides and concentrates scarce review capacity where accuracy gains are largest.

ABSTRACT. Earthquakes are complex phenomena marked by nonlinear dynamics and long-range spatial and temporal correlations, making them challenging to analyze but essential for improving seismic risk assessment. Complex network theory has become a useful tool for studying seismicity, with two main approaches: networks of events (nodes as individual earthquakes) and networks of epicenters (nodes as spatial cells). Connections are typically defined using either the Visibility Graph (VG) model, based on geometric relations in time-magnitude space, or the Time-Window (TW) model, which links events occurring within a set time frame. Although both models reveal key features like small-world structure and scale-free distributions, they have limitations. The VG model depends heavily on temporal ordering, making it less suitable for epicenter networks, while the TW model ignores earthquake magnitude, missing important physical information. To address this, we propose a hybrid time-window model that incorporates magnitude when forming connections (Fig. 1). Using global earthquake data, we show that this model preserves key network properties across regions and timescales. Centrality measures follow Tsallis q-exponential distributions, indicating non-extensive statistical behavior, and networks display assortative mixing, where highly connected nodes link to similar ones. Geospatial analysis highlights that highly connected nodes correspond to active seismic regions, while some act as bridges between distant areas. Overall, this hybrid approach offers a more comprehensive framework for capturing both local and long-range seismic interactions.

ABSTRACT. In this work, we investigated the topological properties of multilayer earthquake networks using a methodology similar to previous studies, but applied to global seismic events. Shallow (≤70 km) and deep (>70 km) earthquakes were analyzed separately. Temporal multiplex networks were constructed, with each layer representing one year from 2000 to 2019, and links defined by successive temporal connections between events. This approach revealed spatial and temporal patterns not visible in aggregated networks. For shallow earthquakes, regions such as Japan, Sumatra, Tonga, and Chile emerged as hubs, consistent with their high seismicity. Notably, the 2004 Sumatra earthquake was preceded by increased local connectivity and followed by long-range interactions, indicating global effects (Fig. 1) . In contrast, deep earthquake networks showed more stable behavior over time, with activity concentrated in subduction zones, particularly Tonga. The strength distributions followed a power law with an exponential cutoff. An optimal number of layers (3 for shallow and 5 for deep events) enhanced the power law regime, highlighting the advantages of the multilayer framework. The global clustering coefficient decreased with the number of layers for shallow events, suggesting scale invariance, while deep events resembled random networks. The average shortest path length increased linearly with layers in both cases. From 2003 to 2006, shallow networks showed evolving global connectivity, especially around Sumatra following the 2004 event, with increased activity also observed in distant regions like China and Iceland. Deep networks (2000–2005) exhibited little variation, with hubs consistently located in regions such as Fiji, Argentina, and Colombia. Future work will explore alternative edge construction methods to further investigate scale-free and small-world properties in global seismic networks.

ABSTRACT. This work analyzes the resilience of decentralized complex systems through the max k- cut game, played on an undirected graph where each agent chooses a strategy (color) from the available ones. Individual payoff is defined as the count of neighbors with different strategies, effectively modeling heterogeneous partitioning and resource differentiation. The game is integrated in Multi-Agent Reinforcement Learning (MARL), where autonomous agents must learn optimal anti-coordination strategies using only local reward signals. MARL agents often face ’lazy agents’ failures, as decentralized optimization typically converges to configurations that are stable against individual moves but globally inefficient. To cope with this problem, after agents employ a policy gradient-based MARL architecture to maximize individual payoffs and eventually become trapped in a Nash equilibrium from which they can not escape alone, they are allowed to develop a joint agency protocol that gives rise to the spontaneous formation of coalitions to overcome Nash traps. Coalitions of size 3 are formed via random sampling, allowing the system to modify strategies and unlock configurations that are inaccessible through isolated strategic moves. During this phase, we observe temporary alignment of decisions able to save color heterogeneity. The proposed framework was benchmarked across four canonical graph topologies: Erdos-Rényi, Small-World, Regular, and Scale-Free, with an average degree of approximately 4 in all cases. We show that the sequential onset of metastable states in such competitive/coalitional networks can drive the system to escape from suboptimal Nash traps and to improve global efficiency by expanding the strategy space, allowing the emergence of a topologically resilient core.

ABSTRACT. Dyadic interview interactions (spanning journalism, political hearings, legal depositions, clinical settings, and media broadcasts) are a fundamental form of human communication and information exchange. Automatically inferring characteristics of the interviewer, interviewee, and session content from video and behavioral data offers substantial theoretical and practical value. Yet, dyadic interaction is inherently temporal, relational, and continuously evolving, while most existing approaches either aggregate behavior across entire conversations or analyze affective signals in isolation from conversational structure. In this work, we introduce a turn-level, event-driven framework for modeling dyadic emotional coordination using Time-Lagged Cross-Correlation (TLCC). Rather than computing global coordination statistics, our approach applies TLCC within transcript-aligned conversational turns, yielding three interpretable metrics per turn: synchrony (strength of alignment), leadership (direction of influence), and response speed (temporal delay).

We evaluate this framework on a dataset of real-world dyadic interview interactions and demonstrate that emotional coordination is highly nonstationary and emerges as localized events rather than global properties of interaction. Our results show that global TLCC compresses heterogeneous dynamics into misleading summaries, while turn-level analysis reveals substantial temporal variability, asymmetric coordination, and shifting leadership roles within the same conversation. Furthermore, we find that traditional conversational structure, including speaker identity and turn position, is insufficient to explain coordination dynamics.

Overall, these findings suggest that dyadic interaction is best understood as a temporally evolving system in which emotional alignment, influence, and response timing are partially decoupled processes. The proposed framework provides a unified and interpretable representation of interaction dynamics, enabling more precise modeling of engagement, rapport, and conversational effectiveness from micro-level behavioral signals.

ABSTRACT. Entering the post-pandemic era, cities navigate a new normal shaped by hybrid work and space-time flexibility, but its concrete contours remain insufficiently understood. Here we present longitudinal, population-scale evidence across 15 U.S. metropolitan areas, analyzing billions of mobile device records spanning 2019--2024. We find roughly $20\%$ of jobs stay remote, marking a threefold increase from pre-pandemic. Nonetheless, daily vehicle kilometers tra\-veled (VKT) increased despite less commuting. Unpacking this pattern, we identify a behavior shift of shorter, more frequent car trips, persisting after a short-term collapse in transit use. We present an approach to delineate urban functional centrality based on mobility data-informed metrics, capturing the interplay of destination pull and commute anchoring. Notably, the observed restructuring of mobility appears ubiquitous across distinct spatial structures. Decomposition analysis reveals trip frequency drives VKT growth, while compact urban forms and concentrated employment mitigate this increase. Planning efforts seeking VKT reduction need to explore ways to take these trips outside of personal cars.

ABSTRACT. Traffic-induced failures—ranging from packet loss in communication networks to congestion-driven breakdown in transport systems—arise when flows progressively exhaust the edges they traverse. Path percolation captures this mechanism by removing edges along sampled paths between origin–destination pairs. Existing studies focus on locally tree-like networks under shortest-path routing, leaving open how path degeneracy and routing stochasticity affect fragmentation in the clustered networks typical of real systems. Here, we introduce a generalised path percolation framework in which paths are sampled from a temperature-controlled ensemble, interpolating between geodesic and noisy routing.

The main results are two-fold. First, for any finite routing horizon $C$, the process coarse-grains to the mean-field percolation. Routing temperature controls non-universal properties, most notably the percolation threshold $p_c$, through the entropy of the load distribution. This identifies entropy as a control parameter for network robustness under path-based failures. Second, tuning the horizon to the mean-field correlation length, $C = N^{1/3}$ within a source-uniform ensemble, drives the system into a crossover regime with distinct scaling behaviour from that reported in the shortest-path percolation, along with a characteristic divergence of the typical path length before giant-component collapse. These findings clarify how microscopic routing choices shape macroscopic resilience, with direct implications for the design of communication and transportation infrastructure.

ABSTRACT. The Density Classification Task (DCT) in cellular automata (CA) is a well-known model and benchmark for collective decision, whereby binary agents coordinate to decide whether their collective contains more individuals in 0 or 1 state.

It provides a canonical framework for studying how purely local interactions can coordinate to achieve nontrivial collective decisions. In this sense, it is akin to distributed coordination problems like quorum sensing in bacteria or social decision.

However, while in reality those problems face noisy environments, most prior work on the DCT has not studied the effect of noise on task performance, nor the mechanisms that allow collectives to evolve robustness to noise.

Our analysis of top-performing DCT rules shows that robustness to noise is strongly correlated with increased \textit{input redundancy} \cite{challa2024}, which is understood as a form of \textit{canalization} in the context of automata networks \cite{manicka2022}, whereby agent decisions are determined by a small subset of effective inputs while others are ignored.

This leads to a fundamental evolutionary question: \textit{does selection for robustness in noisy environments explicitly drive the evolution of canalization, and how does genomic control of input redundancy facilitate this?}

\textbf{Methods:} We conduct large-scale evolutionary searches of one-dimensional CA rules for the DCT via genetic algorithms, optimizing for robustness to noise (accuracy across 10 noise levels).

We further compared two types of agent genomes for encoding decision rules: with and without explicit encoding of redundancy. Canalization in the logic of automata was measured using the CANA Python package \cite{marcus2025} function for $K_r$.

\textbf{Results \& Conclusion:} Our experiments show that pressure for robustness to noise consistently drives the evolution of agents with significantly higher $K_r$ (Fig: \ref{fig:ranks}), suggesting that redundancy is an adaptive mechanism for robust, nontrivial collective decision-making.

Indeed, we find a distinct advantage for explicit genetic control: agents that explicitly encode redundancy in their decision logic evolve significantly higher $K_r$ and superior performance under noise compared to those where redundancy arises implicitly via genetic variation of decision rules.

This shows that evolution finds better collective decision solutions when genomes explicitly encode the canalization mechanism of input redundancy.

Our study suggests that that the pervasive micro-level input redundancy found in experimentally-validated models of biochemical regulation and signaling in nature \cite{manicka2022}, results from the evolution of gene regulatory networks that yield macro-level robustness in decision-making---and this is better achieved via explicit genetic encoding of micro-level input redundancy.

Similar mechanisms under noise may also be at play in collective decision and intelligence in social systems.

ABSTRACT. Urban decision-making requires coordination across interconnected domains such as housing, transportation, and climate resilience, yet systems understanding remains fragmented across stakeholders which limits the development of holistic solutions. Existing urban digital tools largely emphasize descriptive analysis over fostering collective systems understanding, leaving conclusions about system dynamics uncertain even when abundant data is available, as they rely primarily on associative relationships and modelling choices. Here we introduce a knowledge graph–powered AI platform for collaboratively building causal system diagrams through human–AI interaction, enabling more efficient development of explainable and comprehensive system models. We evaluate the approach through usability pilot studies with researchers across urban planning, transportation, environmental, and sustainability science, along with system dynamics stakeholder mapping exercises in New York City involving public and private stakeholders through interviews and participatory workshops. Our results show that AI-enabled collaborative systems mapping accelerates convergence on shared causal structures and improves understanding of complex feedback processes. By integrating human expertise with AI-assisted knowledge structuring, our approach enables a scalable, domain-agnostic workflow for crowdsourcing and refining collective system intelligence across disciplines. Together, these findings position collaborative causal modelling as a foundation for transparent, adaptive, and integrated urban decision-making driven by shared system intelligence.

ABSTRACT. Biological regulatory networks must often remain well adapted to current conditions while retain ing the capacity to respond rapidly to future environmental change. Here, we study how the archi tecture of noise affects this balance in large-scale simulations of Boolean regulatory networks. We compare noise as an evolvable feature of the system: either absent, controlled by a single global parameter, or distributed across genes as gene-specific noise levels. This distinction allows us to test whether isotropic, system-wide variability and anisotropic, gene-level variability have differ ent consequences for evolvability. We study large Boolean networks with N = 100 genes evolving under changing environmental conditions. Across evolutionary runs, we observe the emergence of a bet-hedging attractor structure: networks evolve a dominant basin of attraction that supports high performance in the prevailing environment, while maintaining accessible alternative attractors that can become advantageous after environmental shifts. Gene-level noise substantially accelerates access to these secondary phenotypes compared with both no evolvable noise and a single global noise parameter. The effect is qualitative as well as quantitative: anisotropic noise does not merely increase stochastic exploration, but appears to channel exploration along phenotypic directions that are useful after environmental change. In contrast, global noise produces broader variability without the same degree of directed access to adaptive alternatives.Finally, we examine how evolved internal noise interacts with externally imposed environmental noise. Environmental noise can steer evolutionary trajectories and modulate long-term evolvabil ity, suggesting that variable environments both promote the evolution of evolvability and provide mechanisms for reshaping it. These results connect noise, attractor landscapes, and bet-hedging in regulatory networks, and highlight gene-level stochasticity as a possible mechanism by which complex adaptive systems maintain flexibility under changing conditions

ABSTRACT. Cardiovascular disease emerges from interactions spanning molecular, physiological, and nutritional scales, yet nutritional research commonly evaluates dietary components as isolated variables. Here, we integrate molecular interaction, epidemiological dependency, and food composition networks to identify collective vitamin effects associated with myocardial infarction, hypothesizing that vitamins act as coordinated perturbations across coupled biological and dietary systems.

At the molecular scale, we identify a statistically significant disease module within the human interactome. Network proximity analyses show that different vitamins preferentially align with distinct cardiovascular endophenotypes, suggesting that vitamins collectively modulate cardiovascular pathology through coordinated regulation of complementary disease processes.

At the population scale, Gaussian graphical models inferred from patient dietary, biomarker, endophenotype, and comorbidity data reveal a dependency structure linking vitamin intake, circulating vitamin levels, pathological endophenotypes, comorbidities, and myocardial infarction risk. We identify synergistic vitamin groups associated with coordinated modulation of phenotypes leading to MI, highlighting collective nutritional effects not observable under reductionist analyses of individual nutrients.

Finally, we show that foods self-organize into clusters with distinct vitamin composition profiles, revealing structures in nutritional space that can be leveraged to derive protective dietary patterns. Together, these results demonstrate how multiscale complex systems frameworks can translate molecular mechanisms into realistic population-level dietary interventions and identify synergistic nutritional strategies to reduce chronic disease risk.

ABSTRACT. The molecular origins of aging of biomolecular condensates, which play a central role in cellular organization, is poorly understood. Here, we use coarse-grained molecular simulations to investigate how RNA sequence and chain connectivity govern condensate aging over extended timescales. Condensates formed by CAG-repeat RNA undergo pronounced aging characterized by progressive dynamical slowing, loss of ergodicity, and the emergence of two distinct relaxation timescales. Aging proceeds heterogeneously in space, giving rise to a dynamically arrested, solid-like core surrounded by a more fluid shell. We demonstrate that aging is driven by sequence-encoded base pairing that favors RNA expansion, alignment and the formation of a dense interchain interaction network. These structural changes lead to increased topological entanglement, stabilizing long-lived conformations and reinforcing dynamic arrest in the condensate interior. Strikingly, a scrambled sequence with identical composition remains largely liquid-like. Our results establish RNA sequence patterning as a key determinant not only of phase separation but also of condensate aging and spatial organization. These findings provide a molecular framework for understanding the persistence and solidification of repeat RNA assemblies observed in diseases and suggest general physical principles by which entangled polymer networks drive aging in biomolecular condensates.

ABSTRACT. Cooperation is often easier to sustain among friends than among strangers. People typically have direct knowledge of their friends' behavior towards one another, whereas they must rely on public information about the reputations of strangers. Here we formulate a mathematical framework that integrates direct reciprocity among friends, based on observed actions, and indirect reciprocity among strangers, based on public reputations. Players engage in local games with a small number of neighbors (friends) and in global games with a large pool of non-neighbors (strangers). Behavior toward non-neighbors may be conditioned only on their public reputations, whereas behavior toward neighbors may be conditioned on both observed actions and public reputations. We show that this combination of direct and indirect reciprocity solves the so-called scoring dilemma---allowing cooperation to persist under a simple norm of judgment, even when cooperation would not survive in a classical model of indirect reciprocity alone. This still leaves the question of how to act towards neighbors when direct observations of their past behavior differ from their public reputations. For example, how should we treat a neighbor who has been personally cooperative but is publicly known to be uncooperative? We find that, in order to maximize cooperation and fitness, whenever there is a conflict between local and global information, it is best to overlook whichever bit of information is negative. However, these forgiving strategies that maximize cooperation can become vulnerable to invasion by unconditional strategies when actions and assessments are highly error-prone. Our results contribute to the growing literature on the interplay between different forms of reciprocity.

ABSTRACT. Radicalisation and political polarisation are typically studied as separate phenomena at seperate levels of analysis - individual cognition, network structure or event dynamics. Existing approaches are typically anchored at one level of analysis — individual cognition, network structure, or temporal event dynamics — and lack a unified mechanism that explains how these levels interact to produce self-sustaining escalation. This project proposes hierarchical prior concentration as the core mechanism: the progressive narrowing of moral belief distributions simultaneously across individual (micro), cluster (meso), and population (macro) levels, with each level constituting and amplifying the others in a co-constitutive dynamic. The mechanism is formalised within a logistic-normal Hierarchical Bayesian Model (HBM) that enables bidirectionele uncertainty propagation across levels, early warning detection of approaching tipping points, and mechanistic validation through agent-based model ablation. The empirical design integrates three databases — a longitudinal Reddit dataset (2020–2024), a focused Israel-Gaza Reddit dataset constituting a natural experiment (October 7, 2023 – January 9, 2024), and a Parliamentary Twitter dataset — with four analytical components: NLP-based semantic measurement, network analysis, Hawkes process temporal modelling, and agent-based simulation. The project advances both a formal theory of how radicalisation and polarisation are the same mechanism at different levels of analysis, and a practical early warning architecture with directly actionable intervention implications.

ABSTRACT. Evolution by natural selection occurs when organisms more adept at living and reproducing in their current environment increase in frequency over time. However, organisms are not just passive inhabitants of their environment — they are part of it, and can directly influence each others’ survival and reproduction. Organisms can also remodel their environment, creating further opportunities to shape biotic and abiotic selection, and thus which traits are beneficial or deleterious (i.e., niche construction).

While studying the evolution of multicellularity using populations of event-driven multi-threaded replicating computer programs (i.e., digital organisms), we observed through detailed case study analysis and evolutionary replay experiments that selection imposed by organisms on one another can lead to the evolution of distinct complex multicellular life histories. Surprisingly, organisms with complex genetic programs were not inherently more fit; instead they had collectively reshaped selection. This effect helped stabilize evolved complexity, where evolution of multi-stage developmental patterns “locked in” selection for organisms with similarly coordinated developmental programs (Figure 1). These findings show how biotic selection can act as a powerful complexity ratchet, and concretely demonstrate a possibly critical role of eco-evolutionary feedback in early evolution of biological multicellularity.

ABSTRACT. Collectives of frog embryos show dramatically enhanced survival and morphogenetic outcomes under chemical and genetic stress at larger group sizes — a phenomenon known as cross-embryo mediated assistance (CEMA) [1]. Here we introduce topoiesis, a mechanism by which a diffusive stress field, collectively produced and shared across the embryo ensemble, accumulates more strongly in larger groups through a simple geometric principle: volume scales faster than surface area, reducing boundary loss and sustaining higher field concentrations in the bulk. This minimal, biophysically calibrated model quantitatively recapitulates the observed survival transition across six group sizes (N = 1–300, 0%→92%) with only two free parameters. Beyond rescue, the field imposes a spatial structure on the collective — a weakly-surviving boundary layer enclosing a strongly-surviving core — constituting what we term a super-collective: an emergent entity with its own spatial body plan that tends to reconstitute following random embryo shuffling at early developmental stages. We further introduce metrics of collective individuality and metabolic efficiency that jointly characterize how group geometry shapes the capacity for self-sustaining, resource-efficient rescue. Taken together, these results reveal hidden simplicity behind collective embryonic behavior: diffusion and geometry alone may be sufficient to generate not only collective survival but also emergent spatial self-organization and measurable group-level identity.

ABSTRACT. The brain’s connectome and the vascular system are examples of physical networks—tangible, web-like objects that exist in real space (not just in our papers). This physical reality means these networks combine a graph structure, describing their topological connectivity, with a physical structure, capturing the shape of all nodes and links. How do we best describe this physical structure? Naturally, we model it as a geometric object, i.e., a manifold embedded in 3D space. To do this, we turn to an unexpected mathematical tool: the framework of covariant closed string field theory, developed in the 1980s. This framework provides an exact correspondence between network-like graphs and smooth surfaces. We show that, as interpreted by this string-theoretical framework, geometric objects acquire network shapes precisely because they tend to minimize their surface areas. We developed both a Riemann surface formulation and a numerical algorithm to simulate this minimization process, finding that it predicts structural features challenging traditional explanations of network formation. Specifically, this minimization predicts the emergence of trifurcations and branching angles that, while defying conventional models such as Steiner graphs, are in excellent agreement with the local tree-like organization of physical networks across diverse domains, from human neurons to corals. We conclude by discussing potential applications of this fundamental discovery, from interpreting structural changes in neurological disorders to designing novel metamaterials.

ABSTRACT. Quantum networks, much like complex systems, rely on multifaceted connections. Every network link is defined by a dual nature: a ``quantum'' connection characterizing its entanglement resource, and a ``classical'' connection dictating the success probability of generating this entanglement. Conventionally, efforts to network these links have focused on manipulating one connection type while ignoring the other. This isolated approach inherently causes the ignored connection to compound its imperfections and quickly vanish at large scales. To overcome this scaling bottleneck, here we harness the concept of interdependency from network science, utilizing basic purification protocols to actively design a coupling between entanglement and success probability. We reveal that this designed interdependency drives the emergence of a joint percolating phase across both quantum and classical connections. This phase demonstrates that, despite the inherent stochasticity, quantum networks can still sustain long-range connectivity at arbitrarily large scales---without demanding the unrealistic perfect entanglement or deterministic generation. Crucially, the designed interdependency is fully modularized, enabling direct experimental demonstration of the percolating phase on IBM’s Heron R3 quantum processor without requiring an impractical infinite-scale network. This experimental evidence establishes rigorous physical principles for future scalable network designs, particularly for interconnections between quantum processors.

ABSTRACT. Introduction. The human brain is a complex and dynamical system that changes throughout the lifetime. Noninvasive imaging techniques, including magnetic resonance imaging (MRI), allow for the measurement of large scale functional and structural connectivity networks in the brain. Beyond trends, the dynamics across long timescales are still poorly understood. Methods. Our research aims to address the question of these long-term dynamics by adapting and applying the stochastic actor-oriented model to functional and structural brain networks derived from MRI data in an aging population. Conclusions. The resulting models of network change suggest that the brain is promoting modularity, local efficiency, and short-range connectivity during aging with more adaptive functional networks than the structural networks. Differences between individuals with and without cognitive impairment suggest a disruption in the maintenance of key brain topology in the progression of cognitive impairment with age. These results contribute to the emerging picture of long-term dynamics in aging.

ABSTRACT. Despite its rich temporal resolution, the Electrocochleogram (EEG) remains underexplored with machine learning approaches that also leverage it for physiological interpretation. Often, approaches to EEG analysis rely on hand-crafted features and statistical pipelines that rarely connect machine learning models’ outputs back to established brain physiology, leaving a persistent gap between computational findings and neurological understanding. We propose a comprehensive deep learning framework designed to bridge this gap, using subject classification as the vehicle through which model behavior can be probed and biologically interpreted. We apply this framework to an EEG dataset recorded during subjects' interaction with a simulated social media platform [1], where subjects are categorized as opinion-changers or non-changers. The framework evaluates a broad range of deep learning architecture choices, spanning temporal, spatial, sequential, and graph neural networks for brain connectomics, across multiple input types. These inputs include broadband EEG, frequency-band-specific signals, and network-level EEG features. For any input-model combination yielding strong classification performance (Figure 1A), we apply Explainable AI (XAI) techniques to identify which characteristics of the input are driving the model's decisions, as illustrated by saliency maps highlighting time windows, most influential to subject-level classification (Figure 1B). These explanations are then interpreted through the lens of established neuroscience: network dynamics, functional connectivity, and known neural signatures of cognitive processing, to assess whether the model's learned representations are neurophysiologically coherent and verifiable against existing knowledge of brain function. The result is a principled, replicable pathway for transforming deep learning classification into biologically meaningful insight from EEG.

ABSTRACT. In complex sociotechnical systems - supply chain optimization, sustainability compliance, real asset investment - decision makers operate under a persistent constraint: the models they rely on represent reality without matching it. In the classical Shannon information-theoretic view, new information is valuable precisely because it sharpens our beliefs about outcome distributions. Yet the most decision-relevant information is often distant, costly, and difficult to obtain. Third-party verification services have emerged across industries as a mechanism for reducing decision risk by providing more accurate, targeted assessments of ground truth. This raises a core question: what is the value of verified data, and does Shannon mutual information remain a useful metric as information value grows more complex? We investigate this through a decision-theoretic framework modeling a risk-perceiving agent choosing whether to import agriculture linked to illegal deforestation, under an imperfect calibrated model over a spatial binary risk grid. We formalize two complementary measures: value of information (VOI) in payoff units, and I(V;A*) - Shannon mutual information between verification observations and optimal post-verification action. We show these measures informatively disagree at knife-edge economic configurations, where verification shifts decisions without improving expected outcomes. This work is actively in progress to develop further.

ABSTRACT. Collective decision-making in complex systems often emerges from the interplay between nonlinear individual dynamics and structured interactions over networks. In many real-world settings—ranging from social opinion formation to coordinated behavior in biological systems—agents must select between competing alternatives, leading to bistable or multistable outcomes. While prior work has extensively studied consensus formation, a fundamental question remains open: how do initial conditions and network structure jointly determine which collective decision is ultimately realized? In this work, we introduce a nonlinear networked dynamical framework for collective decision-making based on a double-well potential function, where each agent evolves under intrinsic bistable dynamics coupled through a communication network. As illustrated in Fig. 1(a), the local decision function is shaped by a decision inertia parameter and an external bias, which together determine the relative stability of competing choices. Agents interact over a directed or undirected graph (Fig. 1(b)), where social influence acts to align individual states, while intrinsic dynamics promote commitment to one of the stable equilibria. The resulting system captures the tension between self-reinforcement and social influence, a hallmark of complex adaptive systems (Fig. 1(c)). Our main contribution is the derivation of closed-form algebraic estimates of the basins of attraction associated with each collective decision state in the high-dimensional networked system. The proposed framework is validated through computational experiments and case studies, demonstrating how variations in system parameters and network structure influence convergence to different decision outcomes.

ABSTRACT. Collective decision-making is a fundamental process in animal groups, where individuals rely on local sensory information and social interactions to respond to environmental uncertainty. In fish schools, such decisions can be expressed as transitions between social cohesion and exploration. Here, we conduct behavioral experiments with zebrafish under bright and dark illumination conditions across different number of individuals using the experimental setup shown in Fig. 1(a). We then investigate how visual input affects locomotor activity and the inferred stochastic decision-making mechanisms underlying transitions between social cohesion and exploration. We quantified visual cues via opacity (field occupancy) and optic flow (relative motion). With visual input, zebrafish exhibited higher swimming activity and shorter exploratory bursts, while its absence led to more dispersed and prolonged exploration. Interestingly, fish triads without visual input exhibited longer exploration times compared to dyads. A data-driven stochastic model of interindividual distance revealed a bistable potential landscape that governs transitions between social cohesion and exploration, together with a state-dependent parabolic diffusion term related to the stochasticity of decision-making, as shown in Fig. 1(b). Visual cues biased the potential towards conspecific proximity, while their absence diminished this effect, promoting prolonged exploration. The diffusion term follows an entropy-like pattern analogous to a coin flip, reaching maximal uncertainty at intermediate distances and forcing individuals to break indecision between proximity and exploration. These findings provide quantitative insight into how group size and visual input shape zebrafish swimming and decision-making behavior, offering a useful foundation for future studies of collective dynamical models and the underlying mechanisms by which the structure of stochastic models influences collective decision-making

ABSTRACT. Privacy preferences are not fixed individual traits, they depend on context and lived experiences. In this study, we analyze 2,912 survey responses from 782 college students collected over seven survey periods during 2023 and 2024. We ask about their usage of social media, the security settings of their accounts, and measure their comfort in sharing personally identifiable information (PII) across 17 different institutional contexts. Compared to past research, we observe a large shift towards private accounts, going from 1/3rd private in 2007 to 2/3rds in 2024, and find that participants' discomfort sharing PII with social media platforms strongly predicts their privacy settings. Beyond social media, we identify a stable ranking of institutional trust, though some institutions, like the police, show high variability reflecting divergent lived experiences. Traditionally marginalized groups and participants having faced adverse childhood experiences show more discomfort with institutions of power, especially in areas where they face greater vulnerability. We argue for context-adaptive privacy settings that recognize institutional relationships and demographic vulnerabilities, moving beyond one-size-fits-all consent frameworks toward contextually appropriate data governance.

ABSTRACT. Tourism growth and rising housing costs are accelerating the displacement of long-term residents from historic city centers, yet the effects of regulatory interventions are difficult to anticipate because economic incentives, neighborhood composition, and individual adaptation co-evolve. Here, we develop a spatial, data-driven agent-based model of residential relocation that links rent-gap incentives, neighborhood-driven social dynamics, and experience-weighted attraction learning. We feed the model with multi-source data on rental burden, the spatial mix of long- and short-term accommodation, and local urban features, and calibrate it to reproduce district-level trends in long-term residency in Lisbon during 2018--2025. The calibrated model serves as a counterfactual testbed to compare policy designs under sustained tourism pressure. In Lisbon, simulations reproduce the observed shift of residents away from the historic center and show that limiting short-term accommodation is the most effective lever for retaining residents, while restricting rent increases alone yields limited benefits in the most pressured districts. Figure 1 shows the distribution of housing types in Lisbon in 2025 as predicted by the model.

ABSTRACT. Cities generate increasing returns through agglomeration, producing growing urban surplus and rising locational rents. We present an agent-based urban model in which two competing strategy types, use-seeking residents and return-seeking investors, compete on a spatial grid where aggregate productivity scales superlinearly with population and rents emerge endogenously through competitive bidding.

The central mechanism is bid asymmetry. Investors capitalize expected price appreciation into bids unconstrained by local income, while resident bids remain bounded by wages, savings, and financing constraints. As urban growth increases expected capital gains, realized appreciation feeds back into future investor purchasing power, widening bid asymmetry and producing a nonlinear regime transition.

% ── Figure wraps right ──────────────────────────────────────── \begin{wrapfigure}{r}{0.39\textwidth} \vspace{-6pt} \centering \includegraphics[width=0.38\textwidth]{fig1_regime_transition.png} \captionof{figure}{\small\textbf{Baseline dynamics} ($\tau_{cg}{=}0.05$, 50 seeds, mean $\pm1$ s.d.). \textbf{(a)} Owner-occupiers vs.\ tenants. \textbf{(b)} Investor ownership share $s(t)$: critical transition to ${\sim}90\%$. \textbf{(c)} Auction win rates: investor advantage widens while resident wins collapse.} \vspace{-8pt} \end{wrapfigure}

Three results emerge from this minimal competitive system. \textit{First}, four distinct social classes: rural owners, urban owner-occupiers, urban tenants, and investors, crystallize endogenously from only two initial strategy types. Class stratification is therefore an emergent feature.

\textit{Second}, investor ownership exhibits a sharp S-shaped transition with a bifurcation near capital gains tax rate $\tau^{*}\!\approx\!0.05$--$0.10$ across a range of parameterizations. The system displays \emph{hysteresis}: once the investor-dominated attractor is established, reducing the control parameter $\tau$ does not restore the owner-occupier regime. Accumulated wealth and elevated reservation prices lock the system into a new basin of attraction.

\textit{Third}, when ownership structure feeds back into urban productivity, investor dominance suppresses wages by approximately 15\%, generating a collective maladaptation in which individually rational investment strategies erode the agglomeration externalities that sustain long-run urban productivity.

These results reframe housing financialization as an attractor-boundary problem in complex adaptive systems characterized by endogenous regime transitions and path-dependent lock-in.

ABSTRACT. Controlling synchronization in complex networks typically requires substantial interventions. Inspired by adversarial attack principles from machine learning, we recently showed [1] that tiny periodic phase kicks can dramatically promote or suppress collective synchronization in the Kuramoto model. However, the theoretical mechanism behind this effect was not well understood. Here we provide an analytical explanation using the Ott–Antonsen (OA) reduction [2]. We derive a closed-form expression for how a single adversarial kick changes the order parameter (R), and show that the key lies in a constant-magnitude contribution that is independent of the current level of synchrony. This R-independent term, acting in concert with critically slow relaxation near the synchronization transition and a positive mean-field feedback cascade, explains why even tiny kicks produce disproportionately large effects on the collective state. The sign of the kick determines the outcome in a strikingly asymmetric way: positive kicks push phases toward the mean, eliminating the synchronization transition and inducing coherence at any coupling strength; negative kicks scatter phases away, rendering the incoherent state absorbing for all coupling strengths. This asymmetry is captured within a single unified framework. We further extend the OA theory to uncorrelated heterogeneous networks with arbitrary degree distributions via the annealed network approximation, and validate it on Erdős–Rényi (ER) and Barabási–Albert (BA) networks. Figure 1 compares OA theory and numerical simulations for all-to-all, ER, and BA networks, providing the first analytical foundation for adversarial control of synchronization in complex networks.

ABSTRACT. Collective origin–destination (OD) flows are macroscopic spatial interaction that emerge from the aggregation of individual movements. These flows are central to understanding urban mobility systems, yet real-world observations are often sparse: only a small fraction of OD interactions is observed. This raises a fundamental question: how can collective mobility interactions be inferred from incomplete observations while remaining consistent with regularities in individual visitation behavior? In this work, we address this question by developing a theory-informed graph learning framework for sparse OD flow imputation. Our approach builds on the universal visitation law [1], which relates visit density to destination attractiveness, travel distance, and individual visiting frequency through an inverse-square form. This law is useful for sparse OD flow inference because it suggests that missing interactions are not arbitrary entries, but hidden components of a mobility system governed by shared visitation regularities. However, its application is limited because key quantities such as attractiveness, travel friction, and frequency spectra are typically latent, context-dependent, and nonlinearly coupled with urban form. We therefore propose the Deep Visitation Model (DVM), which uses the visitation law as a structural prior to learn hidden mobility interactions. DVM represents locations and OD relationships as graph-based node and edge structures enriched by socioeconomic attributes, adjacency, distance, and partially observed mobility information. It learns latent representations of destination attractiveness and travel friction, while auxiliary visitation-law-inspired objectives regularize these representations toward population-based attractiveness and distance-based impedance, allowing the model to remain physically grounded while adapting to nonlinear spatial heterogeneity. Experiments on large-scale mobile positioning data from the Twin Cities Metropolitan Area show that DVM robustly recovers missing OD flows across observation ratios. As shown in Figure 1a, it consistently outperforms the classical visitation-law implementation, the gravity model, and Deep Gravity model (DVM), while ablation results confirm the contribution of both attractiveness- and distance-informed constraints. The spatial showcase in Figure 1b further shows lower reconstruction errors than DGM. This research illustrates how the synthesis of statistical physics and deep learning can advance our understanding and predictability of complex spatial interactions in human mobility.

ABSTRACT. Understanding how human mobility, spatial heterogeneity, and evolving epidemic dynamics interact within networked populations remains a central challenge for epidemic preparedness and forecasting. In this work, we present an integrated framework that combines cellphone mobility-informed metapopulation modeling, hotspot-based intervention analysis, and epidemic forecasting to study epidemic dynamics across heterogeneous spatial networks. First, we develop a county-level mobility-driven metapopulation framework to identify epidemic hotspots, defined as counties that generate disproportionately large statewide outbreaks when acting as outbreak origins. We show that targeted interventions that reduce transmission only within hotspot counties can reduce statewide epidemic burden by 60--90\% without requiring broad statewide lockdowns. Moreover, hybrid strategies combining moderate reductions in transmission with partial mobility restrictions from hotspot counties achieve epidemic control comparable to full hotspot suppression while preserving activity elsewhere. Second, we extend this mobility-informed model to analyze forecasting performance across early, surge, peak, post-peak, decline, and long epidemic phases. These studies demonstrate how integrating mobility networks, adaptive inference, and targeted interventions can improve epidemic preparedness and forecasting in complex interconnected systems. The proposed framework provides scalable, policy-relevant tools for understanding, forecasting, and mitigating the dynamics of emerging infectious diseases.

ABSTRACT. Collective perception and social learning strongly influence how populations respond to external threats and environmental pressures [1]. Motivated by recent advances in opinion dynamics and collective behavioral adaptation on networks [2], we introduce a nonlinear networked framework describing the evolution of protective or cooperative actions through adoption, reinforcement, forgetting, and repeated behavioral adaptation. In contrast to standard awareness models, where precautionary responses rapidly disappear once perceived risk decreases, our framework captures persistent behavioral changes sustained through collective memory and nonlinear social reinforcement. Using a degree-based mean-field reduction inspired by recent reduction approaches in collective dynamics [2], we derive the reduced opinion dynamics ⟨x˙ ⟩ = −γ⟨x⟩ + β˜ ⟨x⟩d(1 − ⟨x⟩), where ⟨x⟩ denotes the average level of collective adoption, γ the forgetting rate, and d the nonlinear reinforcement parameter controlling social learning. Stability and bifurcation analysis reveal that nonlinear reinforcement fundamentally reshapes the system dynamics, generating threshold effects, bistability, hysteresis, and oscillatory regimes. We apply the framework to two representative scenarios: ecological preservation behavior and mitigation responses during epidemic outbreaks. In both cases, adaptive perception produces long-lasting collective responses, allowing populations to maintain elevated precautionary behavior even after the original threat has weakened. Figure 1 illustrates the bifurcation structure of the reduced opinion dynamics under different reinforcement regimes, highlighting abrupt transitions and hysteresis associated with persistent collective adaptation. These results demonstrate how reinforcement and social learning can generate durable behavioral responses beyond immediate risk perception, providing insight into adaptive dynamics in social and biological systems.

ABSTRACT. The opioid epidemic was officially declared a public health crisis in 2017 in the United States. Since then, the number of yearly deaths has stabilized, but has yet to return to pre-2017 levels. Empirical research has increasingly demonstrated that factors such as peer dynamics and the presence of specific personal risk profiles shape patterns of opioid consumption and recovery trajectories. In our work, we use a model inspired by coevolving network systems which incorporate processes of social influence (the adoption of your peers' behavior) and social selection (choosing to associate with others similar to you) to simulate the social contagion of a condition such as opioid misuse. We begin with a random graph G(N, M). Nodes can be in one of two states: drug dependent (D) or drug independent (I). We assign each node in the model a risk factor ri drawn from a beta distribution with parameters α and β. Nodes with higher risk factors will tend towards state D. A probability γ represents the rate of social selection. The model is said to have converged when there are no remaining edges between nodes of different states or it has reached a steady state. A set of approximate master equations telling the rates of change of various categories of nodes were derived using pair and triple approximations under a mean field assumption. These differential equations can be solved numerically for a solution independent of scale, sacrificing some accuracy. Simulations of the state space show that the risk distribution plays the most important role in determining the density of nodes in state D when the model converges, with initial density only having an impact at larger γ or when the risk distribution is centered near 0.5. When D0 is close to 0 (≤ 0.001), which is closer to the practical use cases, we observe bistability, where the model either continues on a regular trajectory or collapses to 0. While this is a simplistic model, it provides us with the most important factors at play. This isolated environment allows us to see how these factors will affect the overall behavior, which may be dampened or exacerbated by other forces in the real world, such as triadic closure and homophily on risk factors.

ABSTRACT. Network structure is typically treated as static, while functional organization is inferred from the dynamics unfolding on it [1]. Here, we bridge this gap by defining edges directly through pairwise interaction dynamics, yielding a functional network in which structure emerges from the underlying processes themselves. Within this framework, both node dynamics and local heterogeneity become intrinsically encoded into the effective interactions. As a minimal realization, we investigate nonlinear random walks with crowding, where transitions are modulated by local occupancy constraints and finite carrying capacities [2,3]. The nonlinear diffusion dynamics are governed by x' i =∑j Lij [1 −(1 −kj/ki)xi]xj, ∀i, where A and D denote the adjacency and degree matrices, respectively, and L = AD−1 − I is the random-walk Laplacian, while xi(t) is the density associated with node i. Linearization around the heterogeneous equilibrium reveals that the Jacobian can be decomposed into a baseline diffusion operator together with a degree-dependent diagonal perturbation. Using first-order spectral perturbation theory, we derive reduced expressions describing the deformation of the spectrum induced by nonlinear crowding. Figure 1 illustrates the resulting spectral deformation in scale-free networks. Peripheral low-degree modes are compressed toward the origin, while hub-associated modes are systematically displaced outward, producing a characteristic nonlinear spectral reorganization. The reduced spectral descriptions accurately capture these deformations and reveal how heterogeneity and crowding jointly reshape the effective dynamical structure of the network. These results suggest a form of dynamical modularity, where nonlinear interactions reorganize the spectrum and reshape the effective connectivity perceived by the dynamics.

ABSTRACT. The hypothesis that biological networks exhibit modular organization is widely accepted, yet there is no consensus on what constitutes a biological ``module”, particularly in a way that captures system dynamics rather than the structure of network associations alone. Typical approaches identify modules as community structure, i.e., subgraphs with dense internal and sparse external associations. However, such structure-only definitions fail to explain how coordinated patterns of activity emerge over time. In contrast, the concept of canalization in automata networks highlights how subsets of variables determine dynamics by buffering redundant inputs to enable reliable outcomes despite perturbations. The Dynamics Canalization Map (DCM) of automata networks provides a compact representation of causal dynamics, representing all non-redundant (effective) logical interactions in a form that allows inference of minimal control sets of variable states (seeds).

Here, we present a computational method to extract and characterize dynamical modules from the DCM. In this formulation, modules are defined over node variables in specific states (s-units), allowing that a given node may participate in different modules depending on the state it takes. Starting from pathway modules—sets of node-state transitions guaranteed to unfold from specific seed conditions—we introduce complex modules, which are maximal and require synergistic interactions among their seeds. This reduces the otherwise exponential space of possible pathway modules while preserving those that capture meaningful dynamical dependencies. We then formulate the organization of these modules as a set cover problem, identifying optimal covers that span all s-units (i.e., each node in each ON/OFF state) in the DCM while minimizing overlap between modules. This leads naturally to a measure of dynamical modularity: the average fraction of unique (non-overlapping) s-units in each module in the cover. In this formulation, larger dynamical modularity corresponds to a network dynamics where modules are more decoupled from the rest of the dynamics, i.e., each module explains a distinct and effective component of the dynamics. The approach resembles Simon's concept of near-decomposability. However, rather than identifying partitions of the set of node variables, we identify sets of variables in specific states. Thus, the same variable can participate in different modules.

We computed the dynamical modularity of three benchmark biological networks: Drosophila Single Cell SPN, Budding Yeast Cell Cycle (see Figure), and Arabidopsis Thaliana. The resulting optimal covers consist of largely decoupled modules with high dynamical modularity scores (0.81–1.0), indicating that the dynamics of these models can be decomposed into largely non-overlapping pathways. This suggests that complex biological dynamics may be organized around a small set of irreducible, functionally independent building blocks, with complex modules serving as the fundamental units through which global system behavior may be constructed.

ABSTRACT. Predicting the dynamical behavior of network systems is of paramount relevance to many technological, biological, and social systems. An outstanding challenge with existing prediction methods is that they are often non-scalable with the network size. Here, we show that most empirical networks of scientific interest are localized, and we explore this property of network locality to predict the behavior of large networks at the cost of small ones. We show that locality is characterized by an information distance induced by the network’s coupling matrix. As illustrated in Fig. 1, the propagation of a disturbance throughout a localized network is necessarily concentrated in induced information neighborhoods around the perturbed nodes. Further, we show that the perturbation propagation times tend to scale as a monotonic function of the information distance. Our results show that network locality is a fundamental property governing the dynamics of complex systems across various disciplines, with both conceptual and practical implications.

ABSTRACT. Real-world networks are intrinsically directed characterized by a high degree of non-normality. This implies that eigenvalues alone do not fully characterize short-time responses to perturbations [1]. Even asymptotically stable systems may exhibit strong transient amplification due to non-orthogonal modal interactions. Here, we investigate instantaneous linear response in directed complex networks and derive localization-based approximations for the maximal transversal reactivity. We consider a one-variable-per-node dynamics written on the adjacency network as x˙ i = fi(xi) + σi∑jAijG(xj), ∀i ⇒ x˙ i = Fi(xi) + σi ∑jLijG(xj), ∀i where L = A − Kin is the directed graph Laplacian and Fi(xi) = fi(xi) + σikini G(xi), so that the degree contribution is absorbed into the effective local dynamics Fi. Linearizing around a homogeneous reference configuration yields δ˙x = J δx, where J is the Jacobian operator. The instantaneous response of a perturbation δx is governed by ℓ(J) = δx⊤H δx, H =(J + J⊤)/2, and the exact maximal reactivity is ω(J) = maxδx∈Ω δx⊤H δx. To obtain reduced predictors, we exploit the spectral localization of the eigenvectors v(m) associated with the Laplacian of the undirected backbone network. The localization reduction identifies, for each mode, a dominant node ηm carrying most of the eigenvector weight, with localization strength sm = maxi(v(m)i )2. This yields ω(single)I,proxy ≈ max m≥2 1\2(λm +kout_ηm − kin_ηm)sm, ω(deg)II,proxy ≈ max m≥2 1/4(kout_ηm − 3kin_ηm)sm. The first proxy preserves the backbone eigenvalue contribution while localizing the directed imbalance term; the second applies the localization reduction also to the spectral contribution itself. Figure 1 compares the exact maximal reactivity with the localization-based proxies across biological and social network families. Agreement improves as the relevant backbone modes become more localized, showing that transient amplification in directed systems can often be predicted from sparse spectral information and local in–out degree imbalance.

ABSTRACT. Complex systems often display highly heterogeneous microscopic dynamics while exhibiting robust collective behavior at larger scales. A central challenge is therefore to derive reduced descriptions that preserve emergent dynamics without explicitly resolving all individual-level interactions. Here, we present a computational framework for coarse-graining complex systems by systematically connecting individual-level rules to effective macroscopic equations.

The framework bridges individual-level models, where dynamics are governed by rules and local interaction conditions, with partial differential equations that describe the collective behavior of the system at coarser spatial and temporal scales. The approach constructs low-dimensional dynamical representations that preserve the dominant collective organization and predictive structure of the original system while substantially reducing computational complexity, provided a suitable set of coarse-graining conditions is satisfied.

As a case study, we apply the framework to marine plankton communities, which are central to ocean food webs, global biogeochemical cycles, carbon sequestration, and oxygen production. The framework, by linking microscopic behavioral rules to macroscopic ecological dynamics across scales, opens the possibility of top-down inference of zooplankton behavioral strategies from large-scale ecological observations and remote sensing data.

ABSTRACT. Climate network analysis has emerged as a powerful framework for studying climate dynamics across multiple spatial and temporal scales. However, most existing studies model climate networks primarily through pairwise interactions between geographical locations, thereby overlooking higher-order interactions among regions. Here, we introduce a novel framework that integrates climate network analysis with topological data analysis (TDA) to investigate interactions among higher-dimensional structures within the climate networks. Using simplices and simplicial complexes, our approach captures higher-order interactions beyond conventional edge-based networks and enables the characterization of climate topology through invariant measures such as Betti numbers, Euler characteristics, entropy, and curvature. Applying this hybrid framework to global spatiotemporal climate fields, such as surface air temperature, we demonstrate how major modes of climate variability, including the Pacific Decadal Oscillation (PDO), Atlantic Multidecadal Oscillation (AMO), and El Niño–Southern Oscillation (ENSO), restructure higher-order connectivity patterns across the climate system. We further show that the interplay between these modes and episodic forcings, such as large volcanic eruptions, contributes to climate regime transitions. Finally, using persistent homology analysis, we assess the robustness of these topological structures across varying network thresholds, thereby providing new insights into the multiscale organization and evolution of the global climate system.

ABSTRACT. Pairwise interactions between nonlinear oscillators can reduce, via phase reduction, to Kuramoto-type phase coupling $\sin(\theta_k - \theta_j)$. For higher-order interactions, however, multiple phase couplings exist---such as $\sin(\theta_k+\theta_l -2\theta_j)$ and $\sin(2\theta_k-\theta_l-\theta_j)$. Since different nonpairwise coupling functions produce qualitatively different dynamics, it is important to understand which phase couplings should be included in coupled phase oscillator models. Here, we establish selection rules for higher-order phase coupling functions. These selection rules, which can be applied without the need of explicit phase reduction, are solely based on the symmetry of the isolated oscillator velocity field and the $n$-body interaction functions. As phase reduction established the mechanistic basis for the Kuramoto model, our results provide a theoretical link between physical models with prescribed symmetries and higher-order phase couplings.

ABSTRACT. Phase reduction is a powerful technique to obtain phase models from highly dimensional oscillatory systems. Starting from pairwise interactions, a first order approximation yields Kuramoto and Winfree-like phase models, while nonpairwise interactions emerge already in the second order. Recently, phase models have been extended to account for nonpairwise interactions, notably, the higher-order Kuramoto model. Is there an intrinsic difference between physical and emergent nonpairwise interactions? And can we exploit the former to even out the effects of the latter? In this work, we exploit a recently developed parametrization method to compute the phase reduction to answer these questions. After a comparison of the new method with the classic Kuramoto-style phase reduction, we show that physical and emergent nonpairwise interactions, despite sharing analogous harmonics, are intrinsically different. Lastly, we adopt a synchronization-engineering approach to hack the non-reduced system to behave as much as possible as a first order Kuramoto-like model (see Fig. 1), paving the way for further exploitations of physical nonpairwise interactions for applications in synchronization engineering.

ABSTRACT. In this paper, I introduce labor market mechanisms familiar from partial equilibrium models, microdata econometric studies, and research in psychology and sociology into an agent-based macroeconomic model. These mechanisms include skill-based worker switching, skill loss during unemployment, and skill-dependence of wages. I show that when a shock occurs, these complexities prevent GDP from fully recovering, at least in the medium term: the GDP gap between the shock and no-shock scenarios persists even five years after the shock, whereas no such difference is observed in the baseline model. I also show that the aggregate shock affects different industries differently, and this effect is statistically significant depending on the level of skills required in a given industry. Finally, I demonstrate that training programs for new workers in just the two most affected industries have a comparable impact to training a much larger number of workers without any industry-specific constraints.

ABSTRACT. Adaptive Traffic Signal Control (ATSC) is vital for urban mobility, aligning with United Nation (UN) Sustainable Development Goal 11 by reducing emissions through intelligent coordination. However, most Multi-Agent Reinforcement Learning (MARL) approaches assume unrealistic full observability and use non-contextualized exploration (e.g., ε-greedy), which destabilizes recurrent encoders in decentralized partially observable environments (Dec-POMDP). We propose LSTM-NoisyNet-Double-DQN-MA, a decentralized ATSC architecture integrating a two-layer LSTM encoder, NoisyLinear layers with factorized noise, prioritized sequence replay, and soft target updates (τ=0.001). Trained in SUMO on a 2x2 grid, our model outperformed five baselines (including MA-PPO and LSTM-A3C). It achieved an Average Queue Length (AQL) of 638 vehicles—a 42.9% reduction over the best baseline—while maintaining 1.8x greater stability (CV=3.46%). Results show strict stochastic dominance (U=0, p<10^(-165)) across all comparisons, demonstrating that combining temporal encoders with off-policy rules effectively bridges the observability gap. On-policy LSTM variants failed, evidencing that recurrent encoders require off-policy learning rules. These results support parametric exploration with sequence replay under partial observability, caution against recurrent PPO/A3C, and provide a reproducible baseline for sim-to-real transitions.

ABSTRACT. This paper analyzes the structure of the Argentine labor market using a complex networks approach that links occupations (ISCO-08) and economic sectors (CAES 1.0). Using microdata from the ENES (2019) and ESAyPP (2021) surveys, we construct a high-resolution bipartite network and its corresponding one-mode projections to uncover patterns of interdependence that are not visible through traditional methods.

The results reveal a modular organization in which industrial activities and social services occupy distinct regions of the network, while commerce acts as a bridging node connecting both spaces. In the occupational network, we observe a clear segmentation between manual and non-manual occupations, together with the presence of transversal occupations that connect multiple economic sectors.

A central finding is that gender segregation has a topological counterpart: occupations with similar gender composition tend to cluster in the network (homophily), indicating that inequality is embedded in the relational structure of the labor market. These results highlight the potential of network analysis to better understand the structural organization of labor markets and their patterns of segmentation.

ABSTRACT. Neuronal synchronization can emerge through various coupling mechanisms, but its expression depends strongly on how these connections are organized. Gap junctions, in particular, can reshape inhibitory synchrony, sometimes reinforcing coherence, other times fragmenting it. Building on the classic Rinzel-Golomb model of the thalamic reticular nucleus (TRN), we extend an inhibitory network to include gap-junction coupling arranged in biologically motivated clustered patterns. In particular, we explore the effects of the size, strength, and spatial distribution of gap-junction clusters on synchronization, and how these effects are modulated by the level of background inhibition. Across conductance regimes, weak electrical coupling can transiently destabilize synchrony, while stronger or more extensive clustering promotes coherence or dampens oscillations. These results suggest that the spatial organization of electrical connectivity, together with inhibitory tone, plays a decisive role in shaping rhythmic coordination within TRN-like networks.

ABSTRACT. Climate change is intensifying its impacts worldwide [1], making in- dividual behavior a key lever for an environmentally friendly future. In [2], we built a georeferenced ABM of Siena (Italy), calibrated on a 2023 resident survey, to study individual behavioral dynamics and the emergence of sustainable practices. Mitigation also requires in- stitutions. In [3], Spain is modeled with a two-layer multiplex ABM: citizens interact via peer pressure on a social network with region and income based homophily, while political endorsement and account- ability link society and elected regional leaders. Here, we extend this framework to Italy using graded homophily driven by geographical, socio-economic, and gender dimensions, im- proving the representation of peer influence and national-level policy support.

[1] Abbass K. et al. Environ Sci Pollut Res, 29:42539–42559, (2022). [2] Socci V. et al. EPJ B Topical Issue: Recent Advances in Complex Systems, (2025). [3] Lipari F. et al. Ecol Econ, 217:108084, (2024).

ABSTRACT. Large Language Models (LLMs) are increasingly deployed in complex systems simulations with agent-based models (ABM). In such settings, agent behavior is highly sensitive to system prompt design, yet prompt design is typically manual and heuristic-based. This creates a methodological issue, especially when developing LLM-enhanced agent-based models. We propose an evolutionary framework that automates system-prompt design for LLM-driven agents. System prompts are modeled as modular discrete objects: a prompt is assembled from short components (instructions, constraints, contextual facts), and a binary genome encodes which components are active. This generates an interpretable combinatorial search space and enables contextual behavior optimization without fine-tuning or access to internal model states. The prompt optimize prompts with a binary Genetic Algorithm (GA) using realized cumulative profit as the only fitness signal. We evaluate the method in an agent-based market simulation, where firms repeatedly choose price and product quality while competing for probabilistic consumers. Each GA generation is evaluated via a co-evolutionary market: the genome population is instantiated as a population of firms competing simultaneously, mirroring competitive selection and exposing strategies to an evolving set of opponents. Across controlled experiments, the GA reliably converges to stable and interpretable prompt configurations in a compact 8-component space and remains robust in an augmented 14-component space containing neutral and adversarial instructions. Harmful components are rapidly eliminated, while neutral content shows transient drift and is pruned as selection concentrates near the optimum.

ABSTRACT. Simulation-based inference (SBI) methods enable Bayesian calibration of stochastic simulators such as agent-based models (ABMs). SBI is useful when the likelihood of a model is not tractable but can be approximated numerically by sampling parameters and generating simulations. This work applies the Sequential Neural Posterior Estimation (SNPE), a state-of-the-art method that employs neural network (NN)- based estimation as a flexible density estimator. The framework applied is SBI4ABM, an NN SBI tailored for ABMs. The target ABM is the INET Oxford macroeconomic model (INET ABM), in which we isolate variables related to labour market dynamics. The methodology consists of replacing the standard fully connected multilayer perceptron with an NN that incorporates temporal convolution (to better represent the ABM's temporal dependencies) and domain-specific information (to incorporate knowledge of the ABM application domain). This setup helps the embedding capture sector-specific labour-market changes, such as differences in hiring rates or wage rigidity across industries. This work aims to contribute to the integration of machine learning and complex systems modelling by emphasising the design of summary representations in likelihood-free inference. By embedding econometric structure directly into NN architectures, the approach enhances the efficiency of SBI for large-scale economic simulation.

ABSTRACT. How do public policies emerge and diffuse across thousands of heterogeneous local governments without central coordination — and when does that coordination become indispensable? This paper addresses this question through an agent-based model (ABM) of policy diffusion in Brazilian federalism, applied to the Unified Health System (SUS). The model integrates three policy process theories — the Multiple Streams Framework (MSF), Punctuated Equilibrium Theory (PET), and Diffusion and Innovation Theory (DIT) — within a complex adaptive systems ontology, treating federalism as a dynamic system of incentives, interdependencies, and arenas of political bargaining. A 40×40 spatial grid hosts 1,600 municipality-agents that are heterogeneous in institutional capacity across three strata, operating under bounded rationality. Each municipality may innovate locally, emulate spatial neighbors, or adopt federal policy. Federal intervention is endogenously triggered by an Institutional Coordination Index (ICI) combining mean policy quality, share of municipalities in active search, and inter-municipal inequality. Across 2,400 simulations, the system exhibits seven qualitatively distinct emergent regimes — three of horizontal diffusion and four of federal coordination — whose emergence is sensitive to initial capacity distributions, the timing of federal intervention, and the degree to which pre-existing local trajectories are consolidated. This sensitivity to initial conditions, combined with sharp discontinuities between regime classes, is characteristic of a complex system operating near multiple equilibria. Horizontal regimes (2,191 simulations) reveal the structural limits of decentralized diffusion. Even in the best-performing regime (incipient coordination by innovation), mean final quality reaches only 0.054 and Gini stays above 0.885. In the absence of federal coordination, 90.3% of municipalities remain inactive throughout — not fragmented, but stagnant. Moran’s I stabilizes around 0.09: good practices diffuse within immediate neighborhoods and attenuate progressively with distance from the innovating core. Federal coordination regimes (209 simulations, 8.7%) produce a discontinuous transition: federal intervention triggers immediate adoption in 94.9% of municipalities, with mean final adoption of 92.8% and ∆Q = 0.410. Four regimes emerge: early induction (95.2% adoption), late convergence (92.8%), accumulated autonomy (87.0%), and systemic paralysis. Intervention timing follows a right-skewed distribution (median step 30, mean 46, range 4–618), consistent with nonlinear threshold dynamics and path dependence. No horizontal regime approaches federal-level outcomes, and no high-coverage federal regime approaches horizontal outcomes. The system operates in two structurally separated performance strata, bridgeable only through vertical coordination mechanisms operating at a different scale. These findings offer a formal, generative account of a core theoretical puzzle: why horizontal diffusion in large federations produces persistent territorial inequality even when local innovation and emulation are active.

ABSTRACT. Agent-based modelling (ABM) has become a cornerstone tool for studying complex urban systems, from epidemic spreading and transportation demand to demographic dynamics and environmental exposure. Large-scale ABMs rely on synthetic populations that faithfully reproduce the social, demographic distributions of the real population. With regard to spatial distributions, most existing methods do not map to geographic locations agents’ activities beyond their household, hindering the study on spatial transmission of diseases, mobility, and inequalities of exposure to hazards. We present a novel population synthesis framework that couples sociodemographic realism with spatial resolution. Our method combines Markov Chain Monte Carlo (MCMC) sampling with Linear Programming (LP) optimization to reconcile fine-grained census-tract-level demographic distributions with national-scale microdata constraints. This is a multi-scale consistency challenge that prior paradigms (e.g. synthetic reconstruction, combinatorial optimization) addressed only partially, sacrificing either scalability or convergence guarantees. The framework assigns each synthetic individual a full sociodemographic profile (age, gender, professional condition, and occupation), embeds them in household structures respecting the distributions from microdata surveys, and georeferences their primary activity location to real areas extracted from official repositories (ISTAT, Italian Ministry of Education). Applying our framework to the Italian population using ISTAT microdata, we obtain a synthetic population in strong agreement with official marginals on sociodemographic and mobility patterns, and age-structured contact matrices collected from representative surveys (Figure 1). The resulting population is a dataset enabling spatially explicit simulations across multiple domains: modeling epidemic transmission within settings to address epidemic inequalities, analysing urban mobility and transport demand, and assessing population-level exposure to climate and environmental risks. Our work opens new avenues for evidence-based interventions from neighbourhood to national level scale.

ABSTRACT. Cities evolve as a result of self-organised processes and planning strategies concurrently happening, giving rise to complex economic and infrastructural patterns. A key element of these patterns is firms (physical establishments such as supermarkets, restaurants, retail stores), as their spatial distribution serves as a proxy of local and global economic prosperity. Understanding how these firms organise, persist, and evolve across cities remains a central challenge in complex urban systems research. Following, we gather a decade of geo-referenced firm-level data from two Mexican (Mexico City and Merida) and two British (London, Newcastle) cities, we study the Abundance, Diversity and Longevity (ADL) to characterise local economic patterns and their temporal trajectories at hexagonal spatial index (H3) level. Abundance is defined as the density of economic units, diversity as the heterogeneity of activities, and longevity as the persistence of establishments over time. By comparing distributions, scaling relations, and spatial patterns of the ADL variables, we investigate whether common structural signatures emerge across cities with different sizes, institutional contexts, and developmental stages. In particular, we examine how ADL configurations reflect cities' monocentric or polycentric nature, allowing the identification of intermediate and evolving structural states. We explored the trade-offs between abundance/diversity and longevity, clustering of high-abundance regimes, and the emergence of distinct urban “regimes” defined by ADL configurations. Our results identified invariant patterns and context-dependent deviations in the organisation of urban economic activity. By comparing cities with markedly different scales and urban trajectories, found common ADL signatures that may reflect underlying urban structures positioned along a continuum between monocentric and polycentric organisation.

ABSTRACT. Internal migration shapes urban systems by redistributing people across cities and rural areas. Although migration toward major cities is well documented, it remains unclear how migration-flow structures differ across demographic groups and respond to societal shocks. Network-based analysis has been used to extract interpretable structure from complex interaction systems, such as corporate ownership, and can similarly characterize migration flows. Here, we analyze age- and gender-specific internal migration patterns in Japan using annual municipality-level origin-estination data covering about 90% of domestic migrants over eight years spanning COVID-19. For each age-gender group, we construct a directed weighted migration network and decompose it into four hotspot-based flow types, interpreted as urban-to-urban, rural-to-urban, urban-to-rural, and rural-to-rural migration. We use entropy to quantify these components and interpret the entropy-maximizing hotspot threshold as each group's effective urban scale. Results reveal demographic heterogeneity. Entropy increased among people in their twenties and peaked in 2021, whereas other age groups declined and reached a minimum in 2021. The optimal hotspot threshold differed more by generation than by gender, with larger thresholds among child-rearing-age groups and people aged 60 and older; most groups showed a gradual decrease in this threshold over time, indicating a narrowing effective urban scale rather than an abrupt pandemic-induced shift. Flow components captured interpretable patterns, including strong rural-to-urban migration among younger groups and relatively high urban-to-rural migration among middle-aged and older men. These results suggest hotspot-based flow entropy provides a systematic framework for comparing demographic migration systems and detecting structural changes under societal disruption.

ABSTRACT. Understanding urban systems, either from a theoretical viewpoint or for practical sustainable planning applications, often implies intricate dynamics, that can be in some cases with circular causalities between populations of agents or objects, be coined as co-evolution [1]. In particular, the co-evolution between transportation networks and land-use remains quite unknown, despite direct usefulness regarding several Sustainable Development Goals (better accessibility, decarbonised transport, limited urban sprawl). A method to characterise co-evolution, and more generally the structure of the causal graph between the considered urban system variables, and to establish causality regimes either in empirical data or across the parameter space of a simulation model, has been proposed [2], based on Granger causalities. This contribution investigates how the method can be extended and refined with Transfer Entropy (TE), which provides a more robust and non-linear measure of causality. Given two or more time-series of urban dynamics, observed on a population of objects, Effective Transfer Entropy between all directional couples of variables are estimated on the ensemble (Kraskov estimator implemented in the jidt Java library). We test the method on a stylised yet rich model of urban morphogenesis [3], which simulates settlement and network growth simultaneously, with weight parameters balancing between the explaining factors of future urbanisation (among local density, distance to centre, distance to road network) and with a network growth following a simple connection rule. We compute causalities on the population of raster cells, and density, distance to centre and distance to roads variables. A first experiment, a grid exploration of the model parameter space, provide a broad set of causality regimes, with some illustrated in Fig. 1. The TE method appears more systematic and robust than the previous method, with some regimes not found significant by lagged correlations before. This confirms that TE is an appropriate tool to quantify co-evolution in urban systems, to be further tested on more cases.

ABSTRACT. Collective hostility in animal and human groups is often described with the language of dominance hierarchies, yet many real cases look less like a full ladder and more like a shared fixation on one easy target. We study this process with an agent-based model of moving individuals of several types who observe nearby aggressive encounters and form impressions of entire types. A witnessed victory makes the winner's type appear more dangerous and avoided, and the loser's type less respected and more vulnerable to attack. Across 810 simulations covering different population sizes, perception ranges, numbers of types, and attraction-repulsion balances, the system robustly converges to the same macroscopic state: one type becomes the loser, disrespected even by its own members, whereas the remaining types become near-peers with almost zero intention to attack one another. This social polarization is accompanied by stable clustered groupings in space (see Fig. 1). For the idealized case of perfect observation, where every aggressive act is seen by the whole population, we reduce the model to a shared type-level dynamics and prove that, once a unique current loser appears, the feedback reinforces that asymmetry and drives the system in finite time toward the one-loser/peer regime under natural assumptions on contact rates. Dugatkin's bystander model is the closest predecessor [1]: observers update estimates of particular protagonists; bystander winner effects can yield one clear omega and unclear relations among the rest, while adding direct winner-loser updating can restore a linear hierarchy. We shift learning from individuals to visible types. After a red agent beats a blue agent, witnesses update expectations toward red and blue strata, so uninvolved members can inherit stigma. Thus, the novelty is not witnessing-induced bottom rank itself, but a category-level route from bystander reputation to collective stigma and a peer plateau. The result suggests that collective witnessing can turn local conflict into group-wide consensus about whom it is safe to attack.

ABSTRACT. Measles resurgence in high-income countries that previously achieved elimination reveals a critical surveillance failure: current systems rely on county-level aggregates that obscure fine-scale spatial clustering where outbreaks originate. We assembled the first nationwide multiscale vaccination database spanning 45 US states (2013–2025), encompassing over 50,000 schools, 13,000 districts, and 3,000 counties, integrating MMR coverage, enrollment, exemption type, and geographic coordinates. Using a gravity-based transmission framework that weights pairwise contact intensity between schools by their enrollment sizes and decaying with inter-school distance, we estimated effective reproduction numbers (Rv) at each spatial scale. This framework captures the ecological landscape of susceptibility – not as a uniform field, but as a spatially structured mosaic in which small clusters of under-vaccinated schools generate disproportionate transmission potential amplified by local connectivity. We demonstrate that school-level Rv crossed the epidemic threshold in 2022–2023, a transition entirely invisible when the same data are aggregated to county or state levels. Average susceptibility among school-age children doubled from approximately 5% to 10% following the COVID-19 pandemic, driven by disruptions to routine immunization and expanding nonmedical exemption uptake. Three independent mechanisms generate this aggregation bias. First, enrollment-weighted averaging systematically dilutes small, highly under-vaccinated schools into seemingly adequate county means. Second, gravity kernels attenuate estimated transmission potential at coarser scales by averaging over distances that exceed realistic contact ranges, smoothing away the sharp spatial gradients in the susceptible landscape that drive outbreak initiation. Third, clustering signal erasure converts heterogeneous risk topographies—where localized pockets of high susceptibility sit embedded within well-vaccinated surroundings—into falsely uniform county averages. School district boundaries frequently misalign with county jurisdictions, creating cross-boundary transmission corridors where spillover from under-vaccinated districts pushes neighboring, ostensibly well-vaccinated counties above the epidemic threshold. State trajectories diverge markedly, shaped by exemption policy regimes, enforcement infrastructure, and geographic connectivity of the susceptible landscape. Preventing measles re-endemicity requires surveillance systems operating at the spatial scale where immunological vulnerability accumulates and transmission initiates.

ABSTRACT. Political fragmentation denotes the differentiation of a political system into multiple groups and the extent of separation among them. It often manifests structurally in online interaction behaviors. To measure and compare political fragmentation across contexts, previous scholarship has often relied on network measures of polarisation such as modularity and the Krackhardt E–I index. Here, we show that these metrics combine two aspects of fragmentation: the strength of separation and the number of fragments. These two aspects have not been clearly distinguished in previous work, making comparisons across varied systems difficult to interpret. In addition, none of them is designed to capture the multiscale fragmentation structures that characterize real-world multi-dimensional political spaces. We compare several network measures and show that the two aspects of network fragmentation are best captured by the pairwise adaptive E-I index and the effective number of communities (ENC), while other measures confound the strength of separation and the number of fragments. Furthermore, we introduce a novel metric for multiscale fragmentation, the effective branching factor (EBF), capturing how political fragments at one level split into smaller fragments at the next level. Applying EBF to two empirical datasets spanning Brazil, Spain, and the United States yields consistent country rankings across datasets. Overall, these results clarify three complementary dimensions of structural political fragmentation: strength of separation, number of fragments, and between-level branching. They support a more holistic characterization of structural political fragmentation.

ABSTRACT. In static networks, the outcome of an epidemic spreading process can be predicted through analysis of the underlying network’s structure. However, real world epidemics take place on temporal networks, for which such solutions are not readily available. Our solution to this problem is to map temporal networks to temporal event graphs, a static graph representation that encodes the relevant temporal information and enables the use of tools from static network theory. The temporal event graph [1,2] is constructed by mapping temporal events into nodes and connecting them via ∆t weighted links, directed forward in time if they represent temporally adjacent event pairs. The result is a causal directed acyclic graph (DAG), which encodes all time-respecting paths in the network. Since these time respecting paths form the structural backbone for any epidemic process, there exists a unique structure for any spreading process on a temporal network, (see Fig 1). In order to recover these structures, one must prune the unused links from the event graph. The challenge - and opportunity - lies in how one can recover these structures in a manner which is cost effective and analytically tractable, without explicitly simulating the epidemic process. For spreading processes with recovery, a weight thresholded event graph can be used to identify the reachable event sets, with waiting-time as a control parameter [2]. We demonstrate that this constrained reachability maps to a reinforcing susceptible-infected-susceptible (rSIS) model, where contact between two infectious nodes will re-infect an already infected node, thus restarting its recovery clock. This model forms an upper bound for a conventional SIS process in the supercritical regime, while the threshold of these processes map directly on to each other in random directed temporal networks. Crucially, this method allows us to measure the outcome of an epidemic process through analysis of the event graph structure. This approach to modelling spreading on temporal networks allows us to uncover the interplay between temporal dynamics and static structure in a single analytical framework, paving the way to greater understanding of spreading process on temporal networks.

ABSTRACT. Temporal hypergraphs capture time-resolved group interactions among nodes. Empirical data support that time-stamped group interactions show bursty event sequences and non-trivial temporal correlations. Here we introduce node-driven temporal hypergraph models in which each node stochastically alternates between low- and high-activity states [1], and a hyperedge produces time-stamped events with a probability that depends on the number of high-state nodes in the hyperedge. We consider two event-generation rules; (i) by the AND rule, the event probability on the hyperedge is lambda_h only when all nodes in the hyperedge are in the high state, otherwise it is lambda_l. (ii) By the LIN rule, the event probability on the hyperedge is a linear function of the number of h-state nodes in the hyperedge. We analytically derive interevent time distributions and autocorrelation functions of event sequences, both for hyperedges and nodes (see Fig. 1 for the analytical and numerical results). Despite Markovian node-state dynamics, the induced event processes become mixtures of Poissonian, short-tailed components, resulting in longer-tailed interevent time distributions and slowly decaying autocorrelation. The theory further shows the dependence of these features on the size of hyperedge, which largely agrees with various empirical data. We expect our models to provide a simple, interpretable framework for connecting individual-level activity fluctuations to the timing patterns observed in real group interactions [2].

ABSTRACT. Fish schools form behavioral patterns through interactions among individuals and responses to external factors, and understanding these collective patterns is important not only for biological research but also for the automation of aquaculture management. Although previous studies have shown that feeding-related states can be extracted, behavioral states beyond these remain insufficiently understood. In this study, we aimed to extract and interpret a broader range of collective behavioral states. Following the approach of prior work, dense optical flow was applied to approximately 5.5 hours of optical camera recordings including a feeding event. Speed- and direction-based histograms were then constructed for one-minute intervals and visualized using UMAP. This visualization revealed that the time-ordered samples formed a small number of distinct and temporally persistent clusters, suggesting that collective fish behavior can be represented as transitions among discrete states rather than as a continuously varying pattern of motion.

ABSTRACT. From the formation of ice in small clusters of water molecules to the mass raids of army ant colonies, the emergent behavior of collectives depends critically on their size. At the same time, common wisdom holds that such behaviors are robust to the loss of individuals. This tension points to the need for a more systematic study of how number influences collective behavior. We initiate this study by focusing on collective behaviors that change abruptly at certain critical numbers of individuals. We show that a subtle modification of standard bifurcation analysis identifies such critical numbers, including those associated with discreteness- and noise-induced transitions. By treating them as instances of the same phenomenon, we show that critical numbers across physical scales and scientific domains commonly arise from competing feedbacks that scale differently with number. We then use this idea to find overlooked critical numbers in past studies of collective behavior and explore the implications for their conclusions. In particular, we highlight how deterministic approximations of stochastic models can fail near critical numbers. We close by distinguishing these qualitative changes from density-dependent phase transitions and by discussing how our approach could generalize to broader classes of collective behaviors.

ABSTRACT. How do animal groups dynamically adjust their collective behavior in response to environmental changes remains an open and challenging question. Here, we investigate the mechanisms that allow fish schools to tune their collective state under stress, testing the hypothesis that these systems operate near criticality, a state that maximizes sensitivity, responsiveness, and adaptability. We combine experiments and data-driven computational modeling to study how group size and stress influence the collective behavior of rummy-nose tetras ({\it Hemigrammus rhodostomus}). We quantify the collective state of fish schools using polarization, milling, and cohesion metrics and use a burst-and-coast model to infer the social interaction parameters that drive these behaviors. Our results indicate that group size modulates stress levels, with smaller groups experiencing higher baseline stress, likely due to a reduced social buffering effect. Under stress, fish adjust the strength of their social interactions in a way that leads the group into a critical state, thus enhancing its sensitivity to perturbations and facilitating rapid adaptation. However, while large groups require an external stressor to enter the critical regime, small groups are already near this state. Our findings suggest that adjusting interaction strength is a simple yet effective mechanism that drives collective state transitions while minimizing individual cognitive load. By revealing how stress and group size drive self-organization toward criticality, our study provides fundamental insights into the adaptability of collective biological systems and the emergent properties in animal groups.

ABSTRACT. Firms are complex systems in which interdependent tasks create the potential for error cascade, i.e., incidents in which one error leads causally to another. While prior research has examined complexity arising from task number and task interdependence, two gaps remain: error cascade as a consequence of complexity is understudied, and task variability, i.e., the degree to which the composition of tasks changes over time, has received limited attention as a source of complexity. We ask: How does task variability influence error cascade within firms? Using proprietary assembly line data from a luxury sports car manufacturer covering approximately 14,000 cars across 47 stations, we find that upstream errors cascade to downstream stations, that task variability amplifies this cascade, and that task-specific experience weakens it. Qualitative evidence reveals that the firm buffers against cascade by delaying customization, modularizing tasks, and separating variability from interdependence.

ABSTRACT. A central and recurring debate in complex systems science concerns the relative cost of competing structural representations, in particular whether dynamical systems on networks are more parsimoniously described using graph-based or hypergraph-based formalisms [1].

We address this question of graph versus hypergraph representation from first principles using Kolmogorov complexity as a unified framework for measuring model cost. We show that, in full generality, all representational choices are informationally equivalent: every valid representation (including both graph and hypergraph formalisms) encodes the same total information content, with apparent differences arising only from how information is distributed between structural and dynamical components of the model.

This equivalence, however, is not universal in practice. For most real networked systems, dynamical rules are constrained by mechanistic assumptions rather than being arbitrary. We demonstrate that these constraints break the representational symmetry between graph and hypergraph descriptions, creating a context-dependent landscape in which one representation can become strictly more or less costly than the other.

Our results provide a principled theoretical resolution to the graph-versus-hypergraph debate: in the unconstrained setting no universally optimal structural representation exists, but under mechanistic realism the relative cost of graph and hypergraph models becomes system-dependent, systematically favoring one description over the other depending on the nature of the dynamics (Figure~1).

ABSTRACT. Accurate long-time simulation of plasma physics is essential for realizing clean nuclear fusion energy, yet remains computationally prohibitive and numerically fragile with standard methods. This work addresses two coupled challenges at the intersection of complex systems simulation and artificial intelligence.

Firstly, we develop and analyze structure-preserving numerical methods for plasma simulation that exactly conserve key geometric and physical invariants (e.g., ∇·B = 0) at the discrete level. Unlike standard integrators such as Runge-Kutta methods, which minimize local truncation error but allow conservation violations to accumulate over millions of timesteps, structure-preserving methods prevent unphysical behavior by design.

Secondly, we develop a hierarchical, multi-agent AI system that automates the generation and verification of C / C++ simulations of plasma physics using structure-preserving methods. The system comprises an orchestrator AI agent coordinating four specialist AI agents (mathematical modeling, numerical analysis, HPC code generation and execution, output visualization / analysis) operating over retrieval-augmented knowledge of the PETSc library.

We present weak scaling studies of the resulting simulation pipeline on Polaris and Aurora supercomputers of Argonne National Laboratory. The interplay between complex plasma dynamics, structure-preserving discretization, and AI-driven automation represents a new paradigm for reliable and scalable scientific discovery in fusion energy research.

ABSTRACT. Complex Systems and Complexity to Analyze Data Results from Science, Education and Technology Research Project and Formative Experience with Interdisciplinary Practices M. A. Lucena1 1 Complex Systems and Complexity Laboratory, Institute of Social Research, Fundação Joaquim Nabuco, Recife, PE, BRAZIL marcos.lucena@fundaj.gov.br This work presents results from the research project “Fundaj Goes to School: Education and Science & Technology in the Pandemic” The study combines extension activities and research actions aimed to science education in several schools from Brazil. The theoretical and methodological tools of Complex Systems and Complexity were used to understand, analyze, and conclude. It was also analyzed the experience of the short course “Interdisciplinary experimentation in the classroom among Chemistry, Physics, Biology, and Mathematics,” held during the 77th Meeting of the Brazilian Society for the Advancement of Science (SBPC). The course aimed to expand the theoretical and practical training of basic education teachers through interdisciplinary experiments that connect different scientific fields and pedagogical practices, integrating concepts from Physics, Chemistry, Biology, and Math with contemporary socio-environmental and technological issues. The activities were developed using low-cost experimental kits, non-toxic materials, and digital resources, including virtual experiments and educational games based on free software. These strategies sought to promote innovative, inclusive, and sustainable teaching practices. The research was conducted using action-research methodology, which integrates scientific investigation with educational intervention in a collaborative way. Data were collected through questionnaires applied to student schools about the pandemic and education, and online to high school teachers and undergraduate students who participated in the short course, aiming to understand their perceptions of science education, current educational challenges, and potential of interdisciplinary methodologies in classroom practice. The analysis of the data revealed important aspects related to educational infrastructure, access to technology, remote teaching, teacher training, and teaching methodologies, highlighting challenges that remain present in the educational context. The results indicate the need for more contextualized continuing teacher education policies, as well as greater investment in teaching materials and professional development opportunities that support teachers’ work. Furthermore, interdisciplinarity emerges as a promising pathway to make science teaching more dynamic, meaningful, and accessible. The authors would like to thank CNPq, SBF, SBPC and Fundação Joaquim Nabuco for their support. See Figure 1 with results. (a) b) c) Figure 1. (a) Factorial correspondence analysis of student responses (research project); (b) Factorial correspondence analysis of teacher responses (research project); (c) Similarity Graph (Short Course)

ABSTRACT. Sarcopenia is a disease characterized by age-related decline in muscle mass and strength. It is caused by the combined effects of impaired glucose metabolism, inflammation, and poor nutrition. Many traditional methods that analyze these factors separately do not capture the complex nonlinear interactions involved in disease progression. We present a Complex Adaptive System (CAS) simulation for sarcopenia modeling, based on a discrete-time, four-variable dimensional approach (Figure 1). The model uses the NHANES database of adults aged 60 years or older, which includes grip strength as a measure of muscle function, glycemic burden represented by HbA1c level, systemic inflammation by Neutrophil-to-Lymphocyte Ratio (NLR), and nutritional status using Prognostic Nutritional Index (PNI). A non-linear Hill function was employed to model the inflammatory muscle tissue damage. If the inflammatory response does not exceed the critical threshold value (θ = 1.30), the effect of inflammation is negligible, while exceeding the threshold leads to an exponential growth in the muscle mass damage rate. Our parameter sweep analysis demonstrates tipping-point dynamics just a slight increase in the glucose-inflammation coupling constant leads to a drastic decline in muscle mass from ~51 kg to ~31.5 kg, as well as higher damage sensitivities leading to sudden muscle mass collapse even in cases when the inflammatory response remains below the threshold level, thus indicating the importance of sensitivity to triggering tipping points. The self-organizing analysis suggests that in all analyzed cases, there was convergence towards a single attractor (muscle mass ~51-52 kg), depending on various values of initial muscle mass, inflammation shock, and nutrition. In all these cases, there was stabilization within 4-5 years and recovery from inflammation shock after 6-7 years, due to the action of homeostatic (r_m, r_g, r_i, r_n) and compensatory nutrition terms. These key findings indicate that the progression of sarcopenia is controlled by non-linear feedback processes in which small disturbances may cause collapse, but built-in adaptability leads to stabilization.

ABSTRACT. Startup ecosystems can be viewed as complex adaptive systems in which investors and startups interact through repeated capital allocation under technological and market uncertainty. While prior studies have focused on predicting individual startup success or describing investor portfolios statically, less is known about how investor strategies evolve over time. We interpret portfolio construction as a dynamic search process balancing exploration across emerging technological domains and exploitation within selected niches. We model global startup investments as a temporal investor–startup bipartite network using large-scale Crunchbase data. Startups are associated with multiple business-category tags, and investment events with funding-round stages. By projecting investors’ evolving neighborhoods onto a portfolio strategy space, we analyze how investment strategies move over time. The results suggest heterogeneous, class-dependent pathways. Accelerators tend to move toward broader category coverage and wider investment-stage scope, consistent with exploration-oriented expansion. In contrast, category-specialized investors tend to narrow their portfolio scope over time, suggesting a shift toward exploitation within selected technological domains. These contrasting trajectories indicate that increasing investment activity does not necessarily imply monotonic diversification; strategic evolution is path-dependent and heterogeneous. This motivates a more granular representation of diversification. A scalar diversity index cannot distinguish whether investors diversify by entering more categories, reallocating investments more evenly across existing categories, or spanning more distant technological domains. We therefore extend the analysis by decomposing portfolio diversification into Variety, Balance, and Disparity. This decomposition allows us to examine which dimension drives each trajectory and whether different investor classes follow distinct modes of strategic adaptation.

ABSTRACT. Agent-Based Models (ABMs) provide a powerful paradigm for simulating the emergent behavior of complex systems. However, they remain challenging to analyze when social interactions generate non-linear dependencies, hidden feedback loops, or cascading coordination failures. While the GRAMS (General Reference model for Agent-based Modeling and Simulation) [1] framework formalizes agent-based architectures with rigorous execution semantics—defining agents, environments, events, and constraints—it does not reveal how variability originates and propagates throughout the system. Conversely, the Functional Resonance Analysis Method (FRAM) [2] offers a powerful lens for tracing performance variability and functional resonance in socio-technical systems, yet it inherently lacks a formal representation of agents, internal states, and event-driven execution. This paper introduces HiFRAM (Hierarchical FRAM), a unified representational framework (see Figure 1) that embeds GRAMS execution semantics within an extended FRAM topology. HiFRAM formalizes agents as hierarchical functional structures by introducing multi-port aspects, nested hexagon notations, and multiple-instance operators, all supported by a standardized template for model construction. By applying the framework to the canonical "Warehouse" case study, we demonstrate how HiFRAM successfully exposes systemic bottlenecks and resonance-prone couplings that remain invisible when using GRAMS or FRAM in isolation

ABSTRACT. Human settlements are largely self-organized and have been for most of our evolutionary history. Settlements, then, are emergent from individual or household decisions that balance access to resources, other group members and safety, whether from the elements, inter-group hostility or animal predators. Despite the immense fitness importance of settlements, why do settlements look so different from one another, even within the same ethnolinguistic populations? How is this variation structured? Is the variation predicted by variation in social and ecological environments? These are some of the questions we answer using the case study of villages among the Hamar, a semi-nomadic pastoralist population in Ethiopia. We collected aerial imagery of Hamar villages using Google Earth Pro, after which we calculated village metrics such as the number of huts, area, nearest-neighbor distances, and so on, using ArcGIS Pro. A principal component analysis showed that 80% of the variation within the Hamar villages was along only three orthogonal axes, two of which captured organization within the village, and one which captured variation at a landscape level. We then ran spatial generalized additive models on the principal components and found that the PCs capturing variation within internal organization of villages (PC1 and PC2) were not meaningfully explained by environmental variables, while PC3 was (70% deviation captured, significant predictors (P < 0.0.5): distance to rivers, distance to hostile neighbors, visibility, and distance to roads) This suggests that PC1 and PC2 may themselves be emergent from more basic environmentally determined local interaction rules, that we are unable to capture. Additionally, PC3 is a higher-order organization axis. Currently, we are building agent-based models to understand what local interaction rules lead to PC1 and PC2.

ABSTRACT. International climate targets rely on the success of the energy transition. However, systematic monitoring of how firms adopt low-carbon energy remains scarce. In recent work [1], we use nationwide supply-chain network data to reconstruct energy portfolios for more than 25,000 Hungarian firms between 2020 and 2024, covering 75% of gas, 70% of electricity, and 50% of oil consumption. Energy-providing firms (electricity, gas, oil) are identified from their NACE 4-digit classification; their monetary transactions with every other firm are converted into kilowatt-hours via semi-annual EUROSTAT prices. We fit a linear decarbonization trend δᵢ and an exponential rate λᵢ to each firm's low-carbon share lᵢ(t). This allows us to quantify the speed of the energy transition, i.e. the transition towards low-carbon electricity, on the firm level. We find substantial heterogeneity in decarbonization progress: half increase their low-carbon share, while the other half reduce it. Using logistic regression models we find that energy cost structures are closely associated with transition behavior: firms with a high fossil-cost share are less likely to transition; those with a high electricity-cost share, more likely — hinting at technology-related lock-in effects. Extrapolating current trends yields an aggregate low-carbon share of only 26.6% by 2050 (Figure 1b), highlighting a substantial gap to net-zero pathways. If firms with negative λᵢ adopted the positive rate λᵢ of the closest-matching frontrunner in their NACE 4-digit sector (nearest neighbour by revenue and employment), an aggregate low-carbon share of 69.9% could be achieved by 2050 (Figure 1d), putting climate targets within reach. Our results indicate that the binding constraint for the energy transition is not technological feasibility but the diffusion of best-practice strategies, pointing to the importance of firm-level monitoring and targeted policies.

ABSTRACT. Natural disasters cause large and growing economic damage, but two questions remain open: how resilient are firms to these shocks, and how much of the total economic damage is transmitted through supply chains rather than felt at the point of impact? We address both questions using Ecuador's administrative tax records, which capture the near-universe of formal firm-to-firm transactions between 2012 and 2022, letting us reconstruct the national production network at the firm level and at monthly frequency. We exploit the April 2016 magnitude-7.8 earthquake as a natural experiment, using a difference-in-differences event study to quantify firm-level resilience and shock propagation through the production network. Our framework builds on the seminal studies of Barrot and Sauvagnat [1] and Carvalho et al. [2], but uses continuous transaction values rather than binary links, and monthly rather than annual resolution. Preliminary results show that direct effects exhibit a clear dose–response across local earthquake severity: sales show no significant change in low-severity parishes, fall by 11% in moderate-severity parishes, and by 18% in high-severity parishes in the month of the earthquake (Figure 1, left). Indirect effects scale with pre-event trade exposure: firms outside the disaster zone with more than 20% pre-event trade exposure to affected parishes see sales fall by 14%, despite never being physically affected (Figure 1, right). These findings show that time-resolved firm-level supply chain data are essential for capturing the temporal dynamics and spatial reach of disaster shocks, and for mitigating their indirect effects by targeting aid to firms exposed only through trade links.

ABSTRACT. From the COVID-19 pandemic to the Iran war, recent events have highlighted the systemic fragility of supply chains. Due to highly specific and mutual buyer-supplier dependencies, even the failure of a single firm can cause system-wide economic disruptions in the form of cascading failures up and down the supply chain network. Only recently has it become possible to quantify the systemic impact of the failure of individual firms on the total supply chain. Here, we demonstrate that the systemic risk contributions of combinations of firm failures can be drastically larger than the sum of the damage caused by the firms individually. Using a unique data set that allows us to reconstruct the national supply chain network of Ecuador at the firm-level, we find that combined failures can produce systemic risk amplifications of up to a factor of 257. However, only 0.14% of pairs exhibit a more than 4-fold amplification of systemic risk. Because of this rarity, the average systemic risk increases by 0.71% for pairwise firm failures. We develop a simple, systematic method to identify those combinations of firms that lead to large amplifications of systemic risk. We understand the origin of these amplifications as a breakdown of the substitutability of defaulted suppliers. We discuss the implications of the existence of rare but strong systemic risk amplification for situations that simultaneously affect multiple firms, such as natural disasters and wars.

ABSTRACT. We extend Structural Cellular Hash Chemistry (SCHC) [1,2], a minimalistic artificial chemistry model designed for open-ended evolution, by incorporating spatial locality and dyadicity of interactions among replicating patterns. While the original model relied on global competition based on individual-level hash values of replicating patterns, the revised version introduces two adjustable parameters: spatial interaction range $D$ that limits the distance between two competing patterns, and dyadic interaction probability $P$ that determines the likelihood for two competing patterns to be evaluated by the hash function {\em together} rather than independently. Systematic numerical simulations revealed that incorporating spatial locality and dyadicity significantly enriches SCHC's evolutionary dynamics. Specifically, spatial locality facilitates the spontaneous growth of self-replicating pattern size, while a small probability of dyadic interactions promotes the exploration of a greater number of replicating pattern types. We identified specific ``sweet spot'' parameter settings for $D$ and $P$ at which locality and dyadicity lead to substantially enhanced size growth and open-ended exploration compared to the original model setting. These results demonstrate the importance of ecological interactions for open-ended evolutionary dynamics of artificial life systems.

ABSTRACT. International crises are often encoded as dyadic event sequences, although escalation often turns on coalition threats, security guarantees, joint sanctions, and mediation. We represent each crisis window as a signed temporal hypergraph whose multi-actor episode is an oriented strategic interaction rather than a clique closure. Linearization at the low-escalation state gives a crisis Jacobian in which every multi-actor episode enters as a signed rank-one perturbation of the pairwise backbone. This yields a Strategic Higher-Order Fragility Index, measuring the marginal effect of an episode on spectral stability through eigenvector-weighted alignment between actors generating escalation pressure and actors receiving it. For an escalatory family added to a stable pairwise crisis matrix, we derive an exact threshold in terms of the spectral radius of a reduced transfer operator, together with a local threshold formula that is directly interpretable from the pairwise stability margin. Controlled ensembles show that aligned strategic episodes can lower the exact escalation threshold relative to matched role-randomized nulls while preserving episode sizes and gains. At pairwise hostility alpha=.75, the median exact threshold falls from 1.065 under matched nulls to .765 under aligned higher-order structure across 500 ensembles. The result shows that pairwise hostility may remain subcritical while aligned multi-actor episodes create the first unstable direction of crisis escalation.

ABSTRACT. Evolutionary game theory typically assumes that payoffs are fixed or shaped by external envi- ronmental variables. We introduce an endogenous-feedback framework in which the game played coevolves directly with the population state: the payoff matrix is a time-dependent function of the level of cooperation. This allows strategic incentives to be continuously reconstructed by the collective behaviour they generate. Even in the simplest case of linear and instantaneous feedback, the model reveals feedback-induced regimes, termed chimera games, in which stable cooperation arises despite being incompatible with the predictions of standard fixed-game dynamics. We further show that delayed feedback can destabilize these equilibria and generate sustained oscillations, while nonlinear feedback reshapes equilibrium structure and introduces path dependence. Our results show how cooperation can be promoted, suppressed, or destabilized by incentives generated endogenously by a population’s collective behavior. We conclude by outlining how the model connects to real-world systems shaped by endogenous feedback.

ABSTRACT. Short-term wind speed forecasting remains a difficult problem due to the nonstationary, nonlinear, and multiscale nature of atmospheric dynamics. We propose a reconstruction-aware multiscale forecasting framework that combines four-level discrete wavelet transform (DWT) decomposition with scale-specialized predictive models and an adaptive Self-Error Stabilization Mechanism (SESM) [1]. Unlike conventional wavelet–machine learning hybrids, which typically assume that reconstruction-stage errors are negligible, the proposed approach explicitly addresses error propagation across spectral scales during inverse reconstruction. SESM introduces a causal, scale-aware attenuation operator that detects and suppresses structurally inconsistent high-frequency predictions before signal recompositing, thereby reducing spurious oscillations and improving forecast stability. The framework models low-frequency dynamics using a Long Short-Term Memory (LSTM) network, while medium- and high-frequency wavelet components are predicted using dedicated machine learning regressors. The method was evaluated using a 22-year daily wind speed dataset from the Brazilian coast. Among the tested configurations, the hybrid LSTM–MLP model achieved the best one-step-ahead forecasting performance, with RMSE = 0.275 m/s, MAE = 0.188 m/s, MAPE = 1.99%, R² = 0.991, and C30 = 99.97%. Structural ablation experiments demonstrate the importance of reconstruction-aware stabilization: removing SESM increased RMSE from 0.275 to 1.350 m/s and reduced R² from 0.991 to 0.784. Comparative benchmarking against related daily t+1 wind forecasting studies indicates that the proposed framework can reduce RMSE by up to 47% relative to existing hybrid approaches.

[1] https://doi.org/10.1109/ACCESS.2026.3683005

ABSTRACT. In spreading processes, the presence of immune compartments fundamentally alters the contagion dynamics. Unlike SIS-like dynamics in which reinfections can sustain an endemic state, SIR-like models do not allow for reinfection, leading to transient outbreaks without stable active states. In the context of higher-order contagion models, SIS-like dynamics have been thoroughly studied, uncovering distinct effects not present in pairwise dynamics, most notably the onset of discontinuous transitions and multiple stable states. Here, we study the effects of immune compartments in higher-order contagion dynamics. We show how hyperedge activation is severely hindered when immunity is at play, and how the critical mass threshold - the number of infected nodes required to activate the hyperedge infection - plays a crucial role in discontinuous phenomena in the higher-order SIR model (see Figure 1). The higher-order network structure is also a paramount factor for inducing discontinuous transitions. We analyze how different group size distributions, degree heterogeneity, and hyperedge embedding affect the transition, and we also observe discontinuous transitions in empirical higher-order networks. These findings allow for a deeper understanding of the phenomena that arise for more realistic modeling of higher-order contagion, be it epidemics, information, etc.

ABSTRACT. This contribution questions the epistemic status of the cost-minimising (or utility-maximising) “laws” that are applied to humans and animals, using pedestrian dynamics as an illustration. Usually, it is assumed that these “laws” hold at best in a statistical sense and the question turns into “Is it reasonable to assume that a human can perform such an optimisation?” [1], at the origin of the notions of partial information, bounded rationality, and satisficing choices. I will contend that the question is more satisfactorily addressed by redefining the notion of cost and grounding it in observations, rather than positing its expression in a heuristic fashion. This conduces to shift the focus away from ontology, to a question of approximation quality. To go further, we will notice that cost approximations that only involve the present and past configurations are too short-sighted to account for decisions, when anticipation is at play [2]: for example, a predator chasing a prey is not interested in where the prey currently is, but in where it will soon be. The cost must therefore be framed in a game-theoretical setting in which each agent appears to predict the evolution of the system in response to their own actions and refreshes these predictions every so often, in light of actual observations. Between two observations, agents are in a state similar to Wigner’s friend’s in Quantum Mechanics, in that they must concurrently evolve different possible evolutions for their neighbours, although only one of these options has actually been chosen. These fairly abstract considerations will prove particularly relevant for the concrete context of pedestrian dynamics, marked by anticipatory behaviour: a pedestrian does not fear collisions with pedestrians at their current positions, but where they should be next. We will show how these considerations naturally prompt a novel branch of models for pedestrian dynamics [3], which is particularly well suited to handle scenarios with conflicting moves between pedestrians.

ABSTRACT. Archipelagic fragmentation breaks the continuous landmass underlying previous universality claims, yet the universality properties hold. The Philippine road network with 1.2M intersections distributed across an ensemble of 49,556 isolated components, provides the natural test case. We find that betweenness centrality remains statistically invariant across the 10 largest island components under a two-sample Kolmogorov–Smirnov test (med pairwise KS=0.144) despite a 7× range in component size. The component-size distribution exponent \hat{\alpha}=2.159 lies within the heavy-tailed regime near the 2D percolation ref \tau \approx 2.055, though distinguishing power-law from log-normal in the bulk requires further analysis. Allometric scaling of nodes |V| and road length L against convex-hull area follows sublinear exponents \beta_v= 0.85, \beta_L= 0.83, consistent with sublinear-scaling predictions for urban infrastructure. Persistent homology under the road-class filtration reveals a fourth invariant: across all components, the hierarchical road spine forms a sparse tree, with residential roads contributing 92% of all topological cycles, nearly all of which close at sub-kilometer scales. Near-critical fragmentation flags structural vulnerabilities for disaster resilience and modeling epidemic spread, while universal betweenness offers a transferable backbone for mobility and tourism planning across geographically isolated islands. Planar-network universality thus appears rooted in local growth rules rather than global connectivity, surviving geographic fragmentation.

ABSTRACT. Percolation has been studied widely in diverse applications. In many physical systems, interactions between sites can either promote or inhibit spreading processes. However, signed percolation processes that capture such phenomena are understudied. Due to the existence of negative edge weights, the growth of the giant-cluster (GC) size is not monotonic in such processes. We study a two-state model where 0 and 1 represent nodes being inactive and active, respectively. We update each node's state based on the input signal from their neighbors. By examining a stochastic block model with different Gaussian weight distributions within and between its blocks, we observe oscillations in the GC size for synchronous state updates but not for asynchronous updates. To gain further insight, we apply our threshold model to tree networks where the top layer is the root and only the bottom layer has outgoing edges with negative weights. This set up successfully reproduce the oscillations. Our theoretical analysis and numerical simulations reveal that these oscillations depend on the number of active nodes in the bottom few layers. For both synchronous and asynchronous state updates, we do not observe these oscillations in a configuration model with a Poisson degree distribution in which we choose the edge weights uniformly at random from 1 and -1. Our results suggest that the presence of inhibitory nodes, rather than just edges with negative weights, is responsible for the periodic GC size. For asynchronous updates, we find that fluctuations of the GC size in the stochastic block model can resemble local field potentials in the power-spectrum density when tuning the mean values of the Gaussian weight distributions. For asynchronous updates, we observe a very sharp phase transition in the GC size as we vary parameters in our configuration model. These findings may provide theoretical guidance on how signals propagate in biological neural networks.

ABSTRACT. To understand the structure of real-world networks, it is standard to use null models: random graphs with a set of prescribed properties against which we can compare the structure of real networks. Classical models such as Erdős-Rényi random graphs are mathematically well-studied but fail to reproduce many characteristics observed in social networks. More realistic approaches, such as the configuration model, preserve the degree distribution of the network, but still do not account for higher-order structures such as clustering, motifs, or communities. Developing unbiased random graph models capable of preserving such properties therefore remains an important challenge in network science. In this way, we aim to identify the essential building blocks of the network: motifs or properties that can explain the other characteristics of the graph. To construct these models, a method called the probabilistic k-swap method has recently been proposed [1]. This method requires an initial graph. Starting from this graph, we randomly select a set of edges and perform edge swaps. Figure 1 illustrates a 3-edge swap: the highlighted pink edges in the graph on the left are selected and then rewired in the graph on the right. After a certain number of swaps, which is itself non-trivial to determine, the generated graph can be considered uncorrelated with the initial one. The generation process can therefore be regarded as unbiased. Figure 1: A graph before and after a 3-edge swap preserving the number of triangles and the degree distribution. The swapped edges are highlighted in pink. Our work focuses on generating ego graphs, that is, graphs representing an individual’s social contacts and the relation- ships between them. In particular, we study ego networks extracted from Facebook, which provide detailed observations of online social structures [2]. To achieve this, we aim to identify which motifs are structurally relevant and preserve their number within the graph. For example, in Figure 1, we preserve the number of triangles, a common indicator of local clustering in social networks. By comparing the original networks with these constrained random models, we hope to better characterize which structures emerge from simple local constraints and which reflect more specific social organization principles. References [1] Lionel Tabourier and Julien Karadayi. Probabilistic k-swap method for uniform graph generation beyond the configuration model. Journal of Complex Networks 12 (2024). [2] Raphaël Charbey and Christophe Prieur. Stars, holes, or paths across your Facebook friends: A graphlet-based characterization of many networks. Network Science 7, 476–497 (2019)