pith. machine review for the scientific record. sign in

arxiv: 2605.01803 · v1 · submitted 2026-05-03 · 💻 cs.MA

Recognition: unknown

Koopman Representations for Early Outbreak Warning and Minimal Counterfactual Intervention in Multi-Agent Epidemic Simulations

Florin Leon

Authors on Pith no claims yet

Pith reviewed 2026-05-09 16:27 UTC · model grok-4.3

classification 💻 cs.MA
keywords Koopman representationsmulti-agent epidemic simulationearly outbreak warningcounterfactual interventionattack rate thresholdlatent space forecastingminimal interventions
0
0 comments X

The pith

Koopman representations of early epidemic trajectories enable accurate prediction of major outbreaks and identification of minimal interventions that can prevent them.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a framework that encodes daily aggregate observations from the initial phases of multi-agent epidemic simulations into a low-dimensional latent space using Koopman theory. This space allows approximately linear forecasting of the system's evolution, which is then used with a classifier to determine if the final attack rate will exceed a major outbreak threshold. Experiments demonstrate effective early warning near tipping points, and counterfactual analysis identifies that isolating a single agent for one day can often reduce spread enough to avoid the threshold. A sympathetic reader would care because this points toward data-efficient methods for anticipating and containing epidemics with minimal disruption.

Core claim

Aggregate daily observables from early trajectory windows are encoded into a low-dimensional Koopman latent space whose approximately linear evolution supports short-horizon forecasting and outbreak risk estimation. These representations combined with a random forest classifier predict whether the final attack rate exceeds a major outbreak threshold, and counterfactual analysis shows that minimal interventions such as keeping a single selected agent at home for one day can reduce attack rates and often shift the trajectory below the outbreak threshold.

What carries the argument

The Koopman latent space derived from aggregate daily observables, which approximates the nonlinear epidemic dynamics with linear evolution to enable forecasting and risk classification.

If this is right

  • Early trajectory data suffices for reliable outbreak risk estimation in near-critical regimes.
  • Single-agent, one-day interventions can frequently alter epidemic outcomes below threshold levels.
  • Koopman-derived features enhance the separation between outbreak and non-outbreak classes in classifiers.
  • Minimal counterfactual interventions become identifiable through analysis in the latent space.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the Koopman approximation holds, similar techniques could extend to other multi-agent systems exhibiting threshold behaviors, such as opinion dynamics or resource allocation.
  • Real-world application would require mapping observable aggregates like case counts or mobility to the simulation variables.
  • The low-dimensional latent space suggests potential for scalable monitoring in large populations without full individual tracking.

Load-bearing premise

Aggregate daily observables from early trajectory windows can be effectively encoded into a Koopman latent space that supports accurate short-horizon forecasting and risk estimation in near-critical regimes.

What would settle it

Conducting additional simulations near the tipping points and finding that the random forest classifier's accuracy in predicting major outbreaks drops significantly below the reported strong performance, or that the identified single-agent interventions fail to shift trajectories below the threshold in a majority of cases.

Figures

Figures reproduced from arXiv: 2605.01803 by Florin Leon.

Figure 1
Figure 1. Figure 1: Viral load curves Viral load affects both transmission potential and behavior. First, higher viral load increases the ability of an infected agent to transmit disease during co‐location. The transmission rule is described in the next subsection. Second, viral load determines whether the agent continues ordinary movement, becomes symptomatic, becomes homebound, recovers, or dies. The behavioral and state‐tr… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed workflow. Epidemic simulations generate early aggregate observations, which are mapped into a Koopman latent representation for early outbreak warn‐ ing. High‐risk trajectories are then evaluated with counterfactual interventions to identify small changes that may reduce the outbreak. 24 view at source ↗
Figure 3
Figure 3. Figure 3: displays the model training curves, while view at source ↗
Figure 4
Figure 4. Figure 4: PCA projection of six‐dimensional Koopman latent states in the test set, colored by final outbreak class. 27 view at source ↗
Figure 5
Figure 5. Figure 5: PCA projection of the same test set Koopman latent states, colored by final attack rate. representation. The separation between the two cases supports the use of Koopman features as early indicators of the final epidemic regime view at source ↗
Figure 6
Figure 6. Figure 6: Representative early window trajectories in the Koopman latent space, shown after PCA projection. The trajectories show different regions for contained and major outbreak runs. The Koopman‐derived probability for a major outbreak also provides a direct early warn‐ ing signal. As shown in view at source ↗
Figure 7
Figure 7. Figure 7: Counterfactual intervention in Case 1. The baseline trajectory reaches a large active infection peak and a final attack rate of 0.650. The one‐day quarantine of the selected agent prevents sustained transmission, leaving a final attack rate of 0.002. 32 view at source ↗
Figure 8
Figure 8. Figure 8: Counterfactual intervention in Case 3. The baseline trajectory reaches a peak of 151 active infected agents and a final attack rate of 0.740. The intervention suppresses the cascade, with a peak of 3 active infected agents and a final attack rate of 0.006 view at source ↗
Figure 9
Figure 9. Figure 9: Counterfactual intervention in Case 4. The baseline trajectory develops into a major outbreak, while the intervention trajectory remains close to extinction and reaches a final attack rate of 0.002 view at source ↗
Figure 10
Figure 10. Figure 10: Counterfactual intervention in Case 6. The baseline trajectory peaks late and reaches a final attack rate of 0.660. The intervention prevents sustained transmission, with a peak of 2 active infected agents and a final attack rate of 0.004. In view at source ↗
Figure 11
Figure 11. Figure 11: Counterfactual intervention in Case 2. The one‐day quarantine delays the epidemic peak, but the final attack rate remains above the major outbreak threshold. 34 view at source ↗
Figure 12
Figure 12. Figure 12: Counterfactual intervention in Case 5. The intervention reduces and delays the active infection peak, but the final attack rate remains high and the trajectory remains a major outbreak view at source ↗
read the original abstract

This paper presents a Koopman-based framework for early outbreak detection and intervention selection in a multi-agent epidemic simulation. Agents exhibit mobility patterns, heterogeneous susceptibility, immunity-dependent viral load progression, and local transmission through co-location. The goal of the simulation is to study near-critical epidemic regimes in which small changes in exposure or timing can alter the final outcome. Aggregate daily observables from early trajectory windows are encoded into a low-dimensional Koopman latent space whose approximately linear evolution supports short-horizon forecasting and outbreak risk estimation. These representations are combined with a random forest classifier trained to predict whether the final attack rate exceeds a major outbreak threshold. Experiments near the system tipping points show strong early warning performance, with Koopman-derived features contributing to class separation. Counterfactual analysis further shows that minimal interventions, such as keeping a single selected agent at home for one day, can reduce attack rates and, often, shift the trajectory below the outbreak threshold.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes a Koopman operator framework to encode early-window aggregate daily observables from multi-agent epidemic simulations (with mobility, heterogeneous susceptibility, immunity-dependent progression, and co-location transmission) into a low-dimensional latent space with approximately linear dynamics. This representation supports short-horizon forecasting and is combined with a random forest classifier trained on simulation trajectories to predict whether the final attack rate exceeds a major-outbreak threshold. Experiments near tipping points are reported to yield strong early-warning performance with Koopman features aiding class separation, while counterfactual analysis indicates that minimal single-agent interventions (e.g., one-day home isolation of a selected agent) can reduce attack rates and often shift trajectories below the outbreak threshold.

Significance. If validated with quantitative evidence, the work would demonstrate a promising data-driven approach for applying Koopman linearization to stochastic multi-agent systems near criticality, potentially enabling interpretable early-warning signals and minimal-intervention strategies in epidemic modeling. It integrates latent-space representations with supervised classification and counterfactual reasoning, which could inform real-time public-health decision tools. The focus on near-tipping-point regimes and single-agent perturbations is a notable strength, though the simulation-only setting and absence of reported metrics limit immediate claims of practical impact.

major comments (2)
  1. [Abstract] Abstract: the central claims of 'strong early warning performance' and effective minimal interventions are asserted without any quantitative metrics (e.g., AUC, precision-recall, comparison to baselines, or error bars), validation details, or error analysis, rendering the support for the claims unverifiable from the provided information.
  2. [Counterfactual analysis] Counterfactual analysis description: the claim that a Koopman latent-space perturbation corresponding to a single-agent one-day home isolation reliably maps to the true post-intervention distribution is load-bearing for the intervention result, yet the paper provides no explicit validation that the operator learned on unperturbed aggregate trajectories generalizes to such local perturbations; near criticality, aggregate observables average away the co-location heterogeneity that the intervention exploits, creating a risk that the linear embedding produces artifacts rather than accurate counterfactuals.
minor comments (2)
  1. The choice of Koopman latent-space dimension and the major-outbreak attack-rate threshold are listed as free parameters but receive no sensitivity analysis or justification for their selected values.
  2. Details on how the Koopman operator is learned (e.g., DMD variant, dictionary functions, training data split) and how the random forest is trained and evaluated (cross-validation, feature importance) are not specified in the abstract-level description.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which have helped us identify areas for improvement in our manuscript. Below, we provide a point-by-point response to the major comments and outline the revisions we intend to make.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claims of 'strong early warning performance' and effective minimal interventions are asserted without any quantitative metrics (e.g., AUC, precision-recall, comparison to baselines, or error bars), validation details, or error analysis, rendering the support for the claims unverifiable from the provided information.

    Authors: We acknowledge that the abstract, as currently written, does not include quantitative metrics to support the claims of strong early-warning performance and effective interventions. This was an oversight in the presentation. In the revised version, we will update the abstract to include key quantitative results, such as AUC values for the classifier, performance comparisons to baseline methods, and error bars where applicable. We will also reference the specific sections in the manuscript where detailed validation and error analysis are provided. revision: yes

  2. Referee: [Counterfactual analysis] Counterfactual analysis description: the claim that a Koopman latent-space perturbation corresponding to a single-agent one-day home isolation reliably maps to the true post-intervention distribution is load-bearing for the intervention result, yet the paper provides no explicit validation that the operator learned on unperturbed aggregate trajectories generalizes to such local perturbations; near criticality, aggregate observables average away the co-location heterogeneity that the intervention exploits, creating a risk that the linear embedding produces artifacts rather than accurate counterfactuals.

    Authors: This is a valid concern, particularly given the stochastic and heterogeneous nature of the multi-agent system near criticality. The current manuscript relies on the assumption that the learned Koopman operator can be applied to perturbed states for counterfactual prediction, but does not include a direct comparison between the predicted post-intervention trajectories and those obtained from re-running the simulation with the intervention applied. To address this, we will add an explicit validation experiment in the revised manuscript. This will involve selecting a subset of trajectories, applying the single-agent intervention in the full simulation, and comparing the resulting attack rates and trajectories to those predicted via the Koopman perturbation. We will report quantitative discrepancies and discuss any limitations arising from the averaging of co-location effects in the aggregate observables. revision: yes

Circularity Check

0 steps flagged

No significant circularity; standard supervised feature extraction and classification on simulation trajectories

full rationale

The derivation proceeds by applying the Koopman operator approximation to early-window aggregate observables to obtain a latent representation, then training a random forest on those features (plus possibly raw observables) to classify whether final attack rate exceeds threshold. This is a conventional supervised pipeline: labels are the simulation outcomes, features are computed from initial segments, and performance is reported on held-out or cross-validated trajectories near tipping points. Counterfactuals are described via minimal single-agent interventions whose effects are measured in the simulator. No equation reduces to its input by definition, no fitted parameter is relabeled as an independent prediction, and no load-bearing claim rests on self-citation. The mapping from early aggregates to late outcomes is learned and empirically testable rather than tautological.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

Limited information from abstract only; no explicit free parameters or invented entities detailed. The framework relies on standard assumptions in Koopman theory and machine learning applications to dynamical systems.

free parameters (2)
  • Koopman latent space dimension
    Chosen to be low-dimensional but specific value not given in abstract
  • Major outbreak threshold for attack rate
    Used to define the classification target for the random forest
axioms (2)
  • domain assumption Epidemic dynamics admit a Koopman operator representation that is approximately linear in the latent space
    Enables the forecasting and risk estimation from early trajectories
  • domain assumption The random forest classifier can effectively separate outbreak classes using Koopman features
    Basis for the prediction performance

pith-pipeline@v0.9.0 · 5454 in / 1467 out tokens · 60280 ms · 2026-05-09T16:27:38.710314+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 12 canonical work pages

  1. [1]

    , title =

    Watts, Duncan J. , title =. Proceedings of the National Academy of Sciences , year =

  2. [2]

    , title =

    Das, Subir K. , title =. 2023 , eprint =

    2023

  3. [3]

    Farahi, Zahra and Kamandi, Ali and Abedian, Rooholah and Rocha, Luis E. C. , title =. 2024 , eprint =

    2024

  4. [4]

    International Journal of Advanced Robotic Systems , year =

    Zhu, Qiuguo and Wu, Jun and Xiong, Rong , title =. International Journal of Advanced Robotic Systems , year =

  5. [5]

    Statistical Analysis of Tipping Pathways in Agent-based Models , journal =

    Helfmann, Luzie and Heitzig, Jobst and Koltai, P. Statistical Analysis of Tipping Pathways in Agent-based Models , journal =. 2021 , volume =

    2021

  6. [6]

    and Martin-Linares, Cristina P

    Fabiani, Gianluca and Evangelou, Nikolaos and Cui, Tianqi and Bello-Rivas, Juan M. and Martin-Linares, Cristina P. and Siettos, Constantinos and Kevrekidis, Ioannis G. , title =. Nature Communications , year =

  7. [7]

    and Makeev, Alexei and Kevrekidis, Ioannis G

    Evangelou, Nikolaos and Cui, Tianqi and Bello-Rivas, Juan M. and Makeev, Alexei and Kevrekidis, Ioannis G. , title =. Chaos: An Interdisciplinary Journal of Nonlinear Science , year =

  8. [8]

    Physica A: Statistical Mechanics and its Applications , year =

    Peng, Xiaoyi and Zhao, Yi and Small, Michael , title =. Physica A: Statistical Mechanics and its Applications , year =

  9. [9]

    Chaos: An Interdisciplinary Journal of Nonlinear Science , year =

    Patel, Dhruvit and Ott, Edward , title =. Chaos: An Interdisciplinary Journal of Nonlinear Science , year =

  10. [10]

    Extrapolating Tipping Points and Simulating Non-stationary Dynamics of Complex Systems Using Efficient Machine Learning , journal =

    K. Extrapolating Tipping Points and Simulating Non-stationary Dynamics of Complex Systems Using Efficient Machine Learning , journal =. 2024 , volume =

    2024

  11. [11]

    Applied Koopmanism , journal =

    Budi. Applied Koopmanism , journal =. 2012 , volume =

    2012

  12. [12]

    and Kevrekidis, Ioannis G

    Williams, Matthew O. and Kevrekidis, Ioannis G. and Rowley, Clarence W. , title =. Journal of Nonlinear Science , year =

  13. [13]

    Nathan and Brunton, Steven L

    Lusch, Bethany and Kutz, J. Nathan and Brunton, Steven L. , title =. Nature Communications , year =

  14. [14]

    and Brunton, Bingni W

    Brunton, Steven L. and Brunton, Bingni W. and Proctor, Joshua L. and Kutz, J. Nathan , title =. PLOS ONE , year =

  15. [15]

    Exploring the Latent Space of Autoencoders with Interventional Assays , booktitle =

    Leeb, Felix and Bauer, Stefan and Besserve, Michel and Sch. Exploring the Latent Space of Autoencoders with Interventional Assays , booktitle =. 2022 , volume =

    2022

  16. [16]

    2024 , eprint =

    Chen, Jianming and Wang, Yawen and Wang, Junjie and Xie, Xiaofei and Hu, Jun and Wang, Qing and Xu, Fanjiang , title =. 2024 , eprint =

    2024

  17. [17]

    Brunton, Bingni W

    Steven L. Brunton, Bingni W. Brunton, Joshua L. Proctor, and J. Nathan Kutz. Koopman invariant subspaces and finite linear representations of nonlinear dynamical systems for control. PLOS ONE, 11 0 (2): 0 e0150171, 2016. doi:10.1371/journal.pone.0150171

    work page doi:10.1371/journal.pone.0150171 2016

  18. [18]

    Budiˇ si´ c, R

    Marko Budi s i \'c , Ryan Mohr, and Igor Mezi \'c . Applied koopmanism. Chaos: An Interdisciplinary Journal of Nonlinear Science, 22 0 (4): 0 047510, 2012. doi:10.1063/1.4772195

    work page doi:10.1063/1.4772195 2012

  19. [19]

    Understanding individual agent importance in multi-agent system via counterfactual reasoning, 2024

    Jianming Chen, Yawen Wang, Junjie Wang, Xiaofei Xie, Jun Hu, Qing Wang, and Fanjiang Xu. Understanding individual agent importance in multi-agent system via counterfactual reasoning, 2024

    2024

  20. [20]

    Subir K. Das. Finite-size behavior in phase transitions and scaling in the progress of an epidemic, 2023

    2023

  21. [21]

    Bello-Rivas, Alexei Makeev, and Ioannis G

    Nikolaos Evangelou, Tianqi Cui, Juan M. Bello-Rivas, Alexei Makeev, and Ioannis G. Kevrekidis. Tipping points of evolving epidemiological networks: Machine learning-assisted, data-driven effective modeling. Chaos: An Interdisciplinary Journal of Nonlinear Science, 34 0 (6): 0 063128, 2024. doi:10.1063/5.0187511

    work page doi:10.1063/5.0187511 2024

  22. [22]

    Bello-Rivas, Cristina P

    Gianluca Fabiani, Nikolaos Evangelou, Tianqi Cui, Juan M. Bello-Rivas, Cristina P. Martin-Linares, Constantinos Siettos, and Ioannis G. Kevrekidis. Task-oriented machine learning surrogates for tipping points of agent-based models. Nature Communications, 15 0 (1): 0 4117, 2024. doi:10.1038/s41467-024-48024-7

    work page doi:10.1038/s41467-024-48024-7 2024

  23. [23]

    Zahra Farahi, Ali Kamandi, Rooholah Abedian, and Luis E. C. Rocha. Critical node detection in temporal social networks, based on global and semi-local centrality measures, 2024

    2024

  24. [24]

    u rgen Kurths, and Christof Sch \

    Luzie Helfmann, Jobst Heitzig, P \'e ter Koltai, J \"u rgen Kurths, and Christof Sch \"u tte. Statistical analysis of tipping pathways in agent-based models. The European Physical Journal Special Topics, 230 0 (16--17): 0 3249--3271, 2021. doi:10.1140/epjs/s11734-021-00191-0

    work page doi:10.1140/epjs/s11734-021-00191-0 2021

  25. [25]

    o glmayr and Christoph R \

    Daniel K \"o glmayr and Christoph R \"a th. Extrapolating tipping points and simulating non-stationary dynamics of complex systems using efficient machine learning. Scientific Reports, 14: 0 507, 2024. doi:10.1038/s41598-023-50726-9

    work page doi:10.1038/s41598-023-50726-9 2024

  26. [26]

    Exploring the latent space of autoencoders with interventional assays

    Felix Leeb, Stefan Bauer, Michel Besserve, and Bernhard Sch \"o lkopf. Exploring the latent space of autoencoders with interventional assays. In Advances in Neural Information Processing Systems, volume 35. Curran Associates, Inc., 2022

    2022

  27. [27]

    Nathan Kutz, and Steven L

    Bethany Lusch, J. Nathan Kutz, and Steven L. Brunton. Deep learning for universal linear embeddings of nonlinear dynamics. Nature Communications, 9: 0 4950, 2018. doi:10.1038/s41467-018-07210-0

    work page doi:10.1038/s41467-018-07210-0 2018

  28. [28]

    Using machine learning to anticipate tipping points and extrapolate to post-tipping dynamics of non-stationary dynamical systems

    Dhruvit Patel and Edward Ott. Using machine learning to anticipate tipping points and extrapolate to post-tipping dynamics of non-stationary dynamical systems. Chaos: An Interdisciplinary Journal of Nonlinear Science, 33 0 (2): 0 023143, 2023. doi:10.1063/5.0131787

    work page doi:10.1063/5.0131787 2023

  29. [29]

    Identification and prediction of bifurcation tipping points using complex networks based on quasi-isometric mapping

    Xiaoyi Peng, Yi Zhao, and Michael Small. Identification and prediction of bifurcation tipping points using complex networks based on quasi-isometric mapping. Physica A: Statistical Mechanics and its Applications, 560: 0 125108, 2020. doi:10.1016/j.physa.2020.125108

    work page doi:10.1016/j.physa.2020.125108 2020

  30. [30]

    Duncan J. Watts. A simple model of global cascades on random networks. Proceedings of the National Academy of Sciences, 99 0 (9): 0 5766--5771, 2002. doi:10.1073/pnas.082090499

    work page doi:10.1073/pnas.082090499 2002

  31. [31]

    Alexander D Wilson, Joshua A Schultz, and Todd D Murphey

    Matthew O. Williams, Ioannis G. Kevrekidis, and Clarence W. Rowley. A data-driven approximation of the koopman operator: Extending dynamic mode decomposition. Journal of Nonlinear Science, 25 0 (6): 0 1307--1346, 2015. doi:10.1007/s00332-015-9258-5

    work page doi:10.1007/s00332-015-9258-5 2015

  32. [32]

    Disturbance rejection of a class of discrete-time multi-agent systems with pinning control

    Qiuguo Zhu, Jun Wu, and Rong Xiong. Disturbance rejection of a class of discrete-time multi-agent systems with pinning control. International Journal of Advanced Robotic Systems, 16 0 (6): 0 1729881419881542, 2019. doi:10.1177/1729881419881542

    work page doi:10.1177/1729881419881542 2019