Pith. sign in

REVIEW 2 major objections 5 minor 16 references

A closed drift–diffusion approximation built from approximate master equations predicts the time-dependent variance of SIS epidemics on configuration-model networks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-12 03:24 UTC pith:B7WM4ROY

load-bearing objection Solid, usable AME-Langevin variance tool for SIS on configuration-model graphs; the hypergeometric closure is heuristic but openly tested and the numerics hold where they should. the 2 major comments →

arxiv 2607.03300 v1 pith:B7WM4ROY submitted 2026-07-03 cond-mat.stat-mech math.PRphysics.soc-phq-bio.QM

Variance of the SIS Epidemic on Networks: A Diffusion Approximation

Lucija Nora Farka\v{s} , Sebastian Morel Balbi , Hrvoje \v{S}tefan\v{c}i\'c , Istv\'an Zolt\'an Kiss , Vinko Zlati\'c This is my paper
classification cond-mat.stat-mech math.PRphysics.soc-phq-bio.QM
keywords SIS epidemicconfiguration-model networksdiffusion approximationvan Kampen expansionapproximate master equationfluctuation covariancedegree heterogeneity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Mean-field limits tell us only the average course of an epidemic on a network; they say nothing about how much individual runs wander around that average. For the irreversible SI process a rigorous fluctuation theorem already exists, but recovery in the SIS model returns nodes with partially known neighborhoods and breaks the exact bookkeeping those theorems need. This paper supplies a practical substitute: it closes a three-variable count system (susceptibles, SI edges, SS edges) with a hypergeometric neighborhood approximation drawn from Gleeson’s master equations, builds the corresponding noise matrix, and solves a Lyapunov equation for the covariance. The resulting variance tracks Gillespie simulations closely on Poisson and regular networks, while deviations appear near threshold and on strongly heterogeneous graphs. The method therefore gives a cheap, time-resolved uncertainty estimate that deterministic curves omit and that thousands of stochastic runs would otherwise be needed to obtain.

Core claim

Combining Gleeson’s approximate master equations with a van Kampen system-size expansion yields a closed drift and diffusion matrix for the reduced vector of susceptible counts, SI half-edges and SS half-edges; the associated Langevin/Lyapunov equation then produces a time-dependent covariance whose leading diagonal entry approximates the variance of the susceptible fraction, matching simulation on Poisson and regular configuration-model networks across a range of infection rates.

What carries the argument

The reduced three-dimensional count vector (S, [SI], [SS]) together with a hypergeometric closure for the infected-neighbor distribution of a susceptible vertex; this supplies the Jacobian and diffusion matrix that enter the Lyapunov equation for the fluctuation covariance.

Load-bearing premise

A susceptible node’s neighborhood is treated as a fresh random sample from the current global pool of SI and SS half-edges, as if past infection and recovery history has been fully forgotten.

What would settle it

On a sparse regular network (degree 4) or a strongly heterogeneous power-law network, measure whether the height of the transient variance peak of the susceptible fraction systematically exceeds or falls short of the Lyapunov prediction when compared with a large Gillespie ensemble at fixed distance above threshold.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Forecast uncertainty for network SIS prevalence can be read off a 3×3 ODE system instead of requiring thousands of full stochastic simulations.
  • The third-moment structural factor κ_S^(3) enters the noise matrix explicitly, making the higher-degree dependence of variance computable over the whole epidemic time course.
  • The same reduced construction can be reused for other binary-state processes (SIR, SIRS, adaptive contact models) once the corresponding AME closures are supplied.
  • Near-threshold and hub-dominated regimes are flagged as the regimes where the approximation degrades, guiding when full simulation remains necessary.

Where Pith is reading between the lines

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

  • A fully degree-resolved or neighborhood-resolved Langevin system would remove the hypergeometric closure and should restore accuracy on power-law networks where hubs repeatedly re-infect.
  • The same variance machinery could quantify how incomplete knowledge of the degree distribution inflates forecast error, turning structural uncertainty into a numerical confidence band.
  • Because the method is cheap, it can be embedded inside control or early-warning loops that need real-time estimates of extinction probability or threshold-crossing risk.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper constructs a tractable diffusion approximation for the time-dependent variance of Markovian SIS epidemics on configuration-model networks. Starting from Gleeson’s approximate master equations for the deterministic trajectory, the authors reduce the state to the three-dimensional count vector (S, [SI], [SS]), close the neighborhood distribution of a susceptible degree-k vertex by a hypergeometric draw from the current SI/SS half-edge pool (Eq. 35), obtain the associated drift (Eq. 41), Jacobian (Table 1) and diffusion matrices (Tables 2–3), and integrate the Lyapunov equation for the covariance. The resulting Var(s) is compared with Gillespie ensembles on Poisson, regular and power-law graphs (Figs. 2–5). Agreement is good away from threshold and for moderately homogeneous networks; larger deviations appear near threshold, on sparse regular graphs (k=4) and on strongly heterogeneous power-law networks. The construction is presented as a validated approximation rather than a rigorous FCLT.

Significance. Variance of network epidemics is controlled by higher degree moments that are invisible at the FLLN level, yet closed functional descriptions have been available only for irreversible SI dynamics. The present AME–Langevin reduction supplies a computationally cheap, time-resolved covariance for the recurrent SIS process that matches direct simulation over a useful range of parameters and topologies. The explicit appearance of the structural factors κ^(r)_S and κ^(r)_I makes the third-moment dependence transparent and computable. Because the method is orders of magnitude faster than large Gillespie ensembles, it is immediately useful for uncertainty quantification, threshold-risk estimates and exploratory parameter scans. The failure modes (threshold, sparsity, hubs) are reported rather than concealed, which strengthens the practical value of the approximation.

major comments (2)
  1. The hypergeometric closure P(m|S,k) (Eq. 35) is the single load-bearing approximation. While the paper correctly flags the regimes γ≪β and low degree, the numerical survey does not systematically isolate the two effects: under the constraint βk=2γ used in Fig. 4, increasing k simultaneously lowers the per-edge transmission rate. A modest additional panel (or short table) that holds β fixed while varying k, or that varies γ/β at fixed k, would make the claimed degradation mechanism falsifiable and would strengthen the scoping of the approximation.
  2. Section 3.2 sets all derivatives of κ^(2)_S to zero in the Jacobian after a limited SI test (Fig. 1). On heterogeneous networks the degree-resolved derivative (Eq. 45) is nonzero and can be large for hubs. Because the largest discrepancy already occurs on the power-law example (Fig. 5), the manuscript should either (i) quantify the size of the neglected terms along the AME trajectory for that ensemble or (ii) state clearly that a degree-resolved Langevin system is required before the method can be trusted for strongly heterogeneous graphs.
minor comments (5)
  1. Affiliation numbering for author 3 is duplicated (two “3” labels); the second should be “4”.
  2. In Eq. (24) X_SS is defined as 2[SS]; the subsequent conservation relations (Eq. 26) are consistent, but a one-sentence reminder that half-edge versus edge counting is being used would prevent confusion when readers compare with the SI literature.
  3. Figure 1 caption states “visually very similar”; a quantitative relative-error measure (e.g., integrated absolute difference) would make the claim that the κ derivative is negligible more precise.
  4. The text occasionally switches between “half-edges” and “edges” when referring to [SI] and [SS]; a uniform convention would improve readability.
  5. References [13] and [14] address quasi-stationary distributions; a brief sentence clarifying that the present work targets the full temporal trajectory rather than the endemic quasi-stationary regime would help place the contribution.

Circularity Check

0 steps flagged

No circularity: variance is obtained from an independent Langevin/Lyapunov ODE whose coefficients are supplied by AME trajectories plus an openly approximate hypergeometric closure, then validated against out-of-sample Gillespie statistics.

full rationale

The derivation chain is: (i) Gleeson’s AME supplies the deterministic trajectory ϕ and the structural factors κ_S^(r), κ_I^(r); (ii) a hypergeometric neighborhood closure (Eq. 35) that is exact for SI but only heuristic for SIS is used to close the reduced drift (Eq. 41) and the diffusion matrices B (Tables 2–3); (iii) the Jacobian J is formed while deliberately setting ∂κ_S^(2)/∂ϕ = 0 (justified by an SI test in Fig. 1 and exact on regular graphs); (iv) the Lyapunov equation is integrated to obtain C(t) and hence Var(s). No parameter is fitted to the variance curves themselves; the only free inputs are the epidemiological rates and the degree distribution. The resulting prediction is compared to independent Gillespie ensembles (Figs. 2–5). Ordinary citations to Gleeson, KhudaBukhsh et al. and van Kampen supply background methods, not load-bearing uniqueness claims that force the SIS result. The hypergeometric step is an approximation whose failure modes are stated and observed, not a definitional identity. Consequently the central claim does not reduce to its inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 2 invented entities

The central claim rests on standard configuration-model and continuous-time Markov SIS assumptions, on Gleeson’s AME as the deterministic backbone, and on two paper-specific closures (hypergeometric neighborhood sampling and neglect of ∂κ^(2)_S in the Jacobian). No numerical free parameters are fitted to the variance data; β, γ, degree distributions, and N are experimental inputs. The invented structural factors κ^(r)_S/I are definitions, not new physical entities.

axioms (5)
  • domain assumption Configuration-model random graphs with given degree distribution P(k); large-N mean-field limit exists.
    Standard network-epidemic setting used throughout §§2–3; underpins both AME and the half-edge pool sampling.
  • domain assumption Gleeson’s approximate master equations (12)–(16) correctly supply the deterministic trajectory ϕ(t) and the degree-resolved measures needed for κ factors.
    Taken as given from Gleeson 2013; the Langevin is built on top of AME solutions (§2.2, §3.2).
  • ad hoc to paper Conditional neighborhood of a susceptible degree-k vertex is approximately hypergeometric from the current SI/SS half-edge pool (Eq. 35).
    Exact for SI; for SIS justified only by a re-randomization argument (§3.1). Load-bearing for closing wedges and B.
  • ad hoc to paper Derivatives of the susceptible-centered wedge factor κ^(2)_S with respect to the reduced variables may be set to zero in the Jacobian.
    Motivated by a negligible SI test (Fig. 1) and exactness on regular graphs; not derived for general SIS (§3.2).
  • standard math van Kampen system-size expansion X = Nϕ + √N ζ yields a closed linear Lyapunov equation for Cov(ζ) that captures leading-order fluctuations.
    Standard diffusion approximation; used as in the SI FCLT spirit (§2.3).
invented entities (2)
  • Susceptible-centered wedge factor κ^(r)_S (and infected analogue κ^(r)_I) no independent evidence
    purpose: Close higher-order motif counts (wedges, 3-stars) appearing in drift and diffusion in terms of the reduced (S, SI, SS) variables.
    Defined in Eqs. (38)–(40) from degree-resolved AME measures; they are summary statistics, not new physical objects. independent_evidence false because they are internal definitions.
  • Reduced three-dimensional count vector (X_S, X_SI, X_SS) with associated closed Langevin system no independent evidence
    purpose: Obtain a tractable ODE for time-dependent covariance without tracking the full neighborhood-resolved state.
    Constructed in §2.3 and closed in §3.1–3.2; the reduction itself is the methodological contribution.

pith-pipeline@v1.1.0-grok45 · 19546 in / 3314 out tokens · 25789 ms · 2026-07-12T03:24:51.583656+00:00 · methodology

0 comments
read the original abstract

Functional laws of large numbers (FLLNs) describe the mean-field trajectory of epidemics on networks, but say nothing about the fluctuations around it. These fluctuations are governed by moments of the degree distribution not relevant at the level of the mean. A rigorous functional central limit theorem (FCLT) exists for the susceptible--infected ($SI$) process on configuration-model graphs, but no analogue exists for $SIS$, where recovery reintroduces vertices into the susceptible pool with partially known neighborhoods, breaking the clean neighborhood distribution the $SI$ derivation relies on. We develop a tractable variance approximation for Markovian $SIS$ on configuration-model graphs, combining Gleeson's approximate master equation (AME) framework with a van Kampen system-size expansion in the spirit of the $SI$ FCLT. We derive a closed drift and diffusion matrix for a reduced susceptible/$SI$-edge/$SS$-edge count vector and obtain the time-dependent covariance via the associated Langevin/Lyapunov equation. Validation against Gillespie simulation across Poisson, regular, and power-law networks shows close agreement, with deviations near the epidemic threshold and in strongly heterogeneous networks.

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

16 extracted references · 10 canonical work pages

  1. [1]

    The Annals of Applied Probability22(2), 541–575 (2012) https://doi.org/10.1214/11-AAP773

    Decreusefond, L., Dhersin, J.-S., Moyal, P., Tran, V.C.: Large graph limit for an SIR process in random network with heterogeneous connectivity. The Annals of Applied Probability22(2), 541–575 (2012) https://doi.org/10.1214/11-AAP773

    work page doi:10.1214/11-aap773 2012

  2. [2]

    Random Structures & Algorithms45(4), 726–763 (2014) https://doi.org/10.1002/rsa.20575

    Janson, S., Luczak, M., Windridge, P.: Law of large numbers for the SIR epidemic on a random graph with given degrees. Random Structures & Algorithms45(4), 726–763 (2014) https://doi.org/10.1002/rsa.20575

    work page doi:10.1002/rsa.20575 2014

  3. [3]

    Random Structures & Algorithms41(2), 179–214 (2012) https://doi

    Bohman, T., Picollelli, M.: SIR epidemics on random graphs with a fixed degree sequence. Random Structures & Algorithms41(2), 179–214 (2012) https://doi. org/10.1002/rsa.20401

    work page doi:10.1002/rsa.20401 2012

  4. [4]

    Electronic Journal of Probability18, 1–30 (2013) https://doi.org/10.1214/EJP.v18-2557

    Barbour, A.D., Reinert, G.: Approximating the epidemic curve. Electronic Journal of Probability18, 1–30 (2013) https://doi.org/10.1214/EJP.v18-2557 . Article no. 54 21

    work page doi:10.1214/ejp.v18-2557 2013

  5. [5]

    Journal of Biological Dynamics12(1), 746–788 (2018) https://doi.org/10.1080/17513758

    Jacobsen, K.A., Burch, M.G., Tien, J.H., Rempała, G.A.: The large graph limit of a stochastic epidemic model on a dynamic multilayer network. Journal of Biological Dynamics12(1), 746–788 (2018) https://doi.org/10.1080/17513758. 2018.1515993

    work page doi:10.1080/17513758 2018

  6. [6]

    Journal of Mathematical Biology68(7), 1583–1605 (2014) https://doi

    Graham, M., House, T.: Dynamics of stochastic epidemics on heterogeneous net- works. Journal of Mathematical Biology68(7), 1583–1605 (2014) https://doi. org/10.1007/s00285-013-0679-1 . Erratum: J. Math. Biol. 73 (2016) 257–258, doi:10.1007/s00285-016-1004-6

    work page doi:10.1007/s00285-013-0679-1 2014

  7. [7]

    Journal of Mathematical Biology78(6), 1875–1951 (2019) https://doi.org/10.1007/s00285-019-01329-4

    Ball, F., Britton, T., Leung, K.Y., Sirl, D.: A stochastic SIR network epidemic model with preventive dropping of edges. Journal of Mathematical Biology78(6), 1875–1951 (2019) https://doi.org/10.1007/s00285-019-01329-4

    work page doi:10.1007/s00285-019-01329-4 1951

  8. [8]

    Journal of Mathematical Biology75(3), 577–619 (2017) https://doi.org/10.1007/s00285-016-1092-3

    Ball, F., House, T.: Heterogeneous network epidemics: real-time growth, vari- ance and extinction of infection. Journal of Mathematical Biology75(3), 577–619 (2017) https://doi.org/10.1007/s00285-016-1092-3

    work page doi:10.1007/s00285-016-1092-3 2017

  9. [9]

    arXiv preprintarXiv:1812.03105(2018)

    Ball, F.: Central limit theorems for SIR epidemics and percolation on configura- tion model random graphs. arXiv preprintarXiv:1812.03105(2018). Available at https://arxiv.org/abs/1812.03105

    Pith/arXiv arXiv 2018

  10. [10]

    The Annals of Applied Probability29(2), 1046–1069 (2019) https://doi.org/10

    Barbour, A.D., Röllin, A.: Central limit theorems in the configuration model. The Annals of Applied Probability29(2), 1046–1069 (2019) https://doi.org/10. 1214/18-AAP1425

    2019

  11. [11]

    Advances in Applied Probability54(3), 880–912 (2022) https://doi.org/10.1017/apr.2022.52

    KhudaBukhsh, W.R., Woroszylo, C., Rempała, G.A., Koeppl, H.: A functional central limit theorem for SI processes on configuration model graphs. Advances in Applied Probability54(3), 880–912 (2022) https://doi.org/10.1017/apr.2022.52

    work page doi:10.1017/apr.2022.52 2022

  12. [12]

    Zeitschrift für Wahrscheinlichkeitstheorie und verwandte Gebiete51(3), 269–286 (1980) https: //doi.org/10.1007/BF00587353

    Rebolledo, R.: Central limit theorems for local martingales. Zeitschrift für Wahrscheinlichkeitstheorie und verwandte Gebiete51(3), 269–286 (1980) https: //doi.org/10.1007/BF00587353

    work page doi:10.1007/bf00587353 1980

  13. [13]

    Bulletin of Mathematical Biology84(1), 4 (2022) https://doi.org/10.1007/ s11538-021-00964-7

    Overton, C.E., Wilkinson, R.R., Loyinmi, A., Miller, J.C., Sharkey, K.J.: Approximating quasi-stationary behaviour in network-based SIS dynamics. Bulletin of Mathematical Biology84(1), 4 (2022) https://doi.org/10.1007/ s11538-021-00964-7 . Preprint arXiv:2208.05901

    Pith/arXiv arXiv 2022

  14. [14]

    arXiv preprintarXiv:2509.11706(2025)

    Cantwell, G.T., Moore, C.: The threshold and quasi-stationary distribution for the SIS model on networks. arXiv preprintarXiv:2509.11706(2025). Available at https://arxiv.org/abs/2509.11706

    Pith/arXiv arXiv 2025

  15. [15]

    Physical Review X3(2), 021004 (2013) https://doi.org/10.1103/ PhysRevX.3.021004 22

    Gleeson, J.P.: Binary-state dynamics on complex networks: pair approximation and beyond. Physical Review X3(2), 021004 (2013) https://doi.org/10.1103/ PhysRevX.3.021004 22

    2013

  16. [16]

    American Institute of Physics (1983) 23

    Van Kampen, N.G., Reinhardt, W.P.: Stochastic processes in physics and chemistry. American Institute of Physics (1983) 23

    1983