Source-linked AI summary
Analytical computation of the epidemic threshold on temporal networks
Eugenio Valdano, Luca Ferreri, Chiara Poletto, Vittoria Colizza
TL;DR
Temporal-network epidemic thresholds remain analytically limited when network and spreading timescales overlap or when network models are specific. The paper extends a Markov-chain SIS model with a multilayer matrix formulation for arbitrary temporal networks, and validates the resulting spectral-radius threshold against simulations and diverse networks. It also examines observation-window length to inform empirical data collection.
Problem
Analytical epidemic-threshold calculations remain limited for temporal networks with overlapping timescales or context-specific dynamics, despite the importance of identifying wide-spreading conditions.
Method
The paper extends a discrete-time SIS Markov-chain approach by flattening a multilayer representation that encodes temporal network structure and disease dynamics.
Results
The predicted threshold agrees closely with Markov-chain solutions and stochastic SIS simulations across temporal models and empirical networks.
Takeaways & Limitations
The framework analytically computes thresholds for arbitrary temporal networks and helps assess the observation period needed to characterize spreading properties.
Takeaways & Limitations
The formulation uses discrete time and periodic network boundary conditions; continuous-time extensions are needed when continuous descriptions are more appropriate.
Abstract
from arXiv · showhide
The time variation of contacts in a networked system may fundamentally alter the properties of spreading processes and affect the condition for large-scale propagation, as encoded in the epidemic threshold. Despite the great interest in the problem for the physics, applied mathematics, computer science and epidemiology communities, a full theoretical understanding is still missing and currently limited to the cases where the time-scale separation holds between spreading and network dynamics or to specific temporal network models. We consider a Markov chain description of the Susceptible-Infectious-Susceptible process on an arbitrary temporal network. By adopting a multilayer perspective, we develop a general analytical derivation of the epidemic threshold in terms of the spectral radius of a matrix that encodes both network structure and disease dynamics. The accuracy of the approach is confirmed on a set of temporal models and empirical networks and against numerical results. In addition, we explore how the threshold changes when varying the overall time of observation of the temporal network, so as to provide insights on the optimal time window for data collection of empirical temporal networked systems. Our framework is both of fundamental and practical interest, as it offers novel understanding of the interplay between temporal networks and spreading dynamics.
1 Introduction
The epidemic threshold is difficult to derive for temporal networks whose dynamics overlap with spreading timescales. The paper extends a Markov-chain SIS framework through a multilayer representation to compute thresholds for arbitrary temporal networks.
- Motivation: Epidemic thresholds identify the critical condition for wide spreading and support containment or propagation control.
- Motivation: Existing analytical approaches mainly address timescale-separated dynamics or specific models of time-varying networks.
- Method: The framework represents SIS spreading on a sequence of temporal snapshots using a tensor space indexed by node and time.
- Method: Each node is linked to its future self, while active contacts create cross-time links between connected nodes in successive layers.
- Method: The flattened matrix M jointly encodes temporal topology, causality, and SIS transition probabilities, enabling threshold computation from its spectral radius.
- Limiting case: For static-network limits, the threshold becomes (λ/µ)c,slow = T/ρ(A), with the aggregated matrix retaining the relevant spreading information.
3 Validation and comparison with stochastic simulations
The analytical threshold is validated across model-generated and empirical temporal networks by comparison with Markov-chain solutions and stochastic SIS simulations. Agreement remains strong across varied temporal properties and network sizes, with periodic constraints identified as a technical consideration for real-network data.
- Validation setup: Six test networks combine three temporal-network models with three empirical human-contact datasets.The models include Erdős–Rényi, activity-driven, and BURSTY networks; the empirical datasets include HT09, SEX, and SCHOOL.
- Validation setup: The Markov-chain equation is iterated to its periodic state, and average prevalence is compared with quasi-stationary stochastic SIS simulations.The simulation comparison uses different transmission probabilities after discarding an initial transient.
- Results: The predicted threshold agrees with the Markov-chain solution, while stochastic simulations show a similar transition from zero prevalence to epidemic growth.The two prevalence curves are nearly superimposed, with smoother simulated transitions near the threshold, especially for the small HT09 network.
- Caveat: Periodic boundary conditions may induce temporal paths absent from real contact sequences and thereby alter threshold estimates.The study analyzes this technical assumption separately in relation to data availability and collection.
4 Optimal data collection time
The study examines how observation-window length affects epidemic-threshold estimates in empirical and modeled temporal networks. Thresholds generally saturate once the window captures relevant temporal structure, but network timescales, degree variation, and infection duration influence convergence.
- Motivation: Empirical datasets may incompletely represent contact processes, so observation-window length can affect epidemic-threshold predictions.The analysis seeks a minimum collection period that reliably characterizes spreading potential.
- Threshold saturation: Threshold estimates λc saturate as the period T increases, indicating that sufficiently long collection windows can characterize epidemic dynamics.Saturation and relaxation depend on the network’s typical timescale and temporal structural variability.
- Network dependence: Variation in average degree strongly affects λc, with pronounced oscillations in SCHOOL linked to students’ daily activity.The BURSTY model also reaches a constant λc rapidly, while the other two models saturate even faster.
- Network dependence: When T is shorter than the average infection duration, truncating BURSTY’s inter-contact distribution alters the threshold estimate.Different recovery probabilities generally produce similar saturation behavior, mainly differing by a scaling factor.
- Practical implication: These results support identifying a finite minimum observation window that captures stable properties and patterns of a real contact system.The conclusion is framed as guidance for contact-data collection.
5 Conclusion
The paper presents a multilayer analytical framework for computing epidemic thresholds on arbitrary temporal networks and validates its predictions against numerical SIS dynamics. It also connects observation-period choice with reliable characterization of spreading properties and highlights extensions beyond discrete time.
- Significance: Reliable epidemic-threshold estimates support predicting wide-spreading events and identifying containment or propagation-enhancement strategies.The stated applications include infectious-disease containment and information diffusion.
- Contribution: The framework analytically computes thresholds on arbitrary temporal networks without assumptions about topology or time variation.It extends the static-network Markov-chain model using spectral decomposition of a flattened representation encoding network structure and temporal dynamics.
- Validation: The predicted threshold reproduces Markov-chain and stochastic microscopic simulation behavior with high accuracy.Validation is reported across the temporal networks studied in the paper.
- Data collection: Arbitrary period lengths preserve the method’s applicability while allowing the framework to inform the observation period required for data collection.The paper motivates this use in empirical settings where time is naturally discretized by measurement resolution.
- Perspective: The multilayer formulation supports analysis of how activation rate, temporal correlations, and temporal resolution interact with spreading.The stated aim is to open new theoretical understanding of these coupled temporal processes.
A Proof of Eq. (8)
The proof reduces the spectral calculation for the multilayer matrix from the full tensor-space determinant to a lower-dimensional determinant. This reduction yields Eq. (8).
- Spectral calculation: The eigenvalues of M† are obtained by solving a determinant equation over the R^N T space.The determinant is evaluated for the full multilayer representation.
- Dimensionality reduction: Because x − M† consists of T^2 blocks of size N × N, matrix-determinant results reduce the problem from det_NT to det_N.The block structure and zero blocks simplify the general reduction.
- Conclusion: Equation (8) follows immediately from the reduced determinant expression.
B Networks considered for validation
Validation uses six temporal-network examples spanning synthetic models and empirical contact datasets, with distinct structural and temporal characteristics.
- Validation networks: Six networks are used to validate the epidemic-threshold expression.The set includes ER, ACTIVITY, BURSTY, HT09, SEX, and SCHOOL.
- Synthetic models: ER consists of 500-node Erdős-Rényi snapshots with 750 edges and mean degree 3.
- Synthetic models: ACTIVITY renews links each snapshot and assigns heterogeneous activity potentials sampled from a power-law distribution.The model uses two connections per active node and tunes average degree with η = 10.
- Synthetic models: BURSTY models activity using a time-since-last-activation probability to generate bursty inter-event times.
- Empirical networks: HT09 and SCHOOL record face-to-face proximity using wearable RFID sensors, whereas SEX infers sexual contacts from online escort-forum posts.HT09 measurements use 20-second temporal resolution.
C Estimation of the epidemic threshold from numerical simulations
Numerical simulations estimate prevalence near the epidemic transition using a quasistationary-state method, which addresses the rarity of surviving active configurations.
- Simulation method: The quasistationary-state method constrains simulations to active states near the transition.This improves efficiency when surviving configurations are rare and many realizations would otherwise be needed.
- Simulation method: The numerical procedure measures average simulated prevalence after discarding an initial transient.
- Context: The section’s validation context includes the study’s stated funding and acknowledgments.
S1 Optimal data collection time
The epidemic threshold stabilizes rapidly as the observation period grows, while its short-period oscillations track fluctuations in snapshot average degree and are damped by cumulative averaging.
- Optimal data collection time: For ER and ACTIVITY, λc converges quickly, with no significant oscillations for T > 5.The result is associated with low average-degree variation across snapshots, which is zero for ER by construction.
- Optimal data collection time: A small number of snapshots is sufficient to compute the epidemic threshold correctly.
- Threshold–degree relation: λc oscillations are associated with instantaneous fluctuations in the network’s average degree as T varies.
- Threshold–degree relation: Increasing T damps threshold fluctuations because each added snapshot has less relative influence over a longer cumulative period.
- Threshold–degree relation: Cumulative average degree, computed over all snapshots up to T, shows similarly damped oscillations.
- Higher-order correlations: Two-point correlations are nearly absent in ER but weakly nonzero in HT09, consistent with more accurate threshold computation for ER.
S3 Weighted and directed networks
The methodology extends beyond undirected unweighted networks: directed networks require transposed adjacency matrices, and weighted networks require weight-dependent transmission probabilities.
- Scope and extensions: The main formulation assumes undirected, unweighted temporal networks with symmetric binary adjacency matrices.
- Directed networks: Directed networks are handled by replacing A with A† in the annealed-approximation threshold computation.The proposed approach does not require A(t) to be symmetric.
- Weighted networks: Weighted networks are tractable when transmission probability is defined from the weight in the relevant application context.Integer weights correspond to the probability of at least one infection across repeated trials.
- Weighted networks: After incorporating the weighted transmission contribution into M, the threshold is computed as in the main methodology.
S4 Computational performance
The computational method uses sparse numerical linear algebra to compute the spectral radius of matrix P. Its implementation relies on Python scientific-computing libraries and a modified power-iteration algorithm.
- The algorithm is implemented in Python 2.7 using numpy and scipy libraries.The implementation uses scipy for sparse matrix representation and multiplication.
- Sparse matrix operations support the numerical computation of the spectral radius of matrix P.The implementation uses numpy.dot and scipy.sparse.csr.
- A modified power-iteration method computes the spectral radius of matrix P.