{"id":"d8c22125-2e3c-4cca-9c7d-5c83f504ec9e","arxiv_id":"2502.04225","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A functional law of large numbers is proved for non-Markovian SIR epidemics with infection-age-dependent infectivity on large heterogeneous random graphs, with a graphon PDE as the limit.","lead":"This paper proves that for a large population, the random spread of an SIR epidemic on a heterogeneous random network converges to a deterministic limit described by equations on a graphon. The result lets scientists replace detailed stochastic simulations of large realistic networks with tractable deterministic equations that keep individual differences and infection-age-dependent infectivity.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Key convergence lemmas contain a false bound: V_N = λ*√Υ_N + ω*γ̄_N does not majorize λ*√(Υ_N + ω*γ̄_N), so the proof of Theorem 3.6 is incomplete as written.","rationale":"The reader's verdict of CONDITIONAL is appropriate, and the concern about the restrictive Assumption 3.1 inequality is legitimate: the condition excludes natural ongoing-epidemic initial states with no recovered individuals but infected individuals of positive age, and this restriction is not flagged in the abstract or introduction. However, the most load-bearing issue for the central claim is internal to the proof: Lemma 5.6 and Lemma 5.7 assert a bound that does not follow from the displayed estimates. The failure is concrete and occurs exactly in the asymptotic regime γ̄_N → 0 required by Assumption 3.5. Because Proposition 5.11 depends on these lemmas, Theorem 3.6 is not established by the argument as written. Replacing V^N by the quantity actually obtained from the Cauchy–Schwarz estimates preserves the conclusion, since both candidate error terms tend to zero under Assumption 3.5, so the paper can likely be repaired with a routine correction. Given that a fix is straightforward and the overall strategy appears sound, the verdict should remain conditional rather than reject: the authors should correct the bound in Lemmas 5.6–5.7 and then the main proof goes through. The agreement with the reader is partial because we identify a different, more immediate correctness concern, while acknowledging the initial-condition restriction they emphasized.","tokens_in":42474,"tokens_out":16026,"duration_ms":176520,"concrete_test":"Set X = {x0}, choose edge weights with zero variance (υ ≡ 0) so Υ_N = 0, and take γ_N(x0,x0) = 1/N, giving γ̄_N = 1/N and ω* = 1. Substitute into Lemma 5.6: (5.34) gives (1/N)Σ_j E[(~S_j)^2] = ω*γ̄_N = 1/N, so the proof's estimate yields U^N ≤ λ*√(1/N), while the stated V^N is 1/N. For N > 1, λ*√(1/N) > 1/N, disproving the claimed U^N ≤ V^N. Then re-derive Proposition 5.11 with V^N := λ*√(Υ_N + ω*γ̄_N) and verify that the Gronwall argument still closes with a vanishing error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Lemma 5.6 (and similarly Lemma 5.7) claims U^N ≤ V^N with V^N defined in (5.27) as λ*√Υ_N + ω*γ̄_N. The proof, via (5.31)–(5.34) and Cauchy–Schwarz, actually establishes U^N ≤ λ*√(Υ_N + ω*γ̄_N). The inequality λ*√(Υ_N + ω*γ̄_N) ≤ λ*√Υ_N + ω*γ̄_N is not true in the regime required by Assumption 3.5. For example, with Υ_N = 0 and γ̄_N = 1/N, the left side is λ*√(ω*/N) while the right side is ω*/N, so the claimed bound fails for large N. Proposition 5.11 uses 2V^N + L^N as a vanishing error in the Gronwall estimate that yields D^N(t) → 0, which is the core of the FLLN proof. With the corrected bound V^N := λ*√(Υ_N + ω*γ̄_N), or equivalently λ*(√Υ_N + √(ω*γ̄_N)), the error still converges to zero under Assumption 3.5, so the theorem is likely salvageable, but the proof as written contains a genuine algebraic gap. This is a more load-bearing concern than the Assumption 3.1 inequality, which only limits the class of initial conditions rather than invalidating the convergence argument.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies an individual-based stochastic SIR model on a large random graph with individual characteristics, random edge weights, and infection-age-dependent infectivity. The main theorem (Theorem 3.6) asserts that, under Assumptions 2.2, 2.4, 2.6, 3.1, 3.3, and 3.5, the normalized measure-valued processes (μ̄S,N, μ̄I,N, μ̄R,N) converge in probability to a deterministic limit characterized by the measure-valued equations (3.6)-(3.9), with an additional PDE-on-graphon representation. The proof constructs an auxiliary mean-field process driven by the limiting force of infection, decomposes the discrepancy between the original and auxiliary processes into seven terms, and controls them through a series of lemmas and a Gronwall argument. Appendix A supplies technical supporting results, including a generic convergence lemma (Proposition 5.9) of independent interest.","tokens_in":1032,"tokens_out":990,"duration_ms":119161,"significance":"If the result holds, this is a substantial extension of functional law of large numbers results for SIR epidemics: it combines non-Markovian infectivity, individual heterogeneity, and non-homogeneous random graph connectivity in a single graphon limit. The paper is methodical and self-contained, and the generic convergence result Proposition 5.9 is of independent interest. There is no parameter fitting or circular prediction; the deterministic limit is derived from explicit probabilistic assumptions. The main proof has a localized algebraic gap in the error bounds of Lemmas 5.6-5.7 that should be corrected before publication.","major_comments":[{"comment":"The proof of Lemma 5.6 establishes the bound U^N ≤ λ* sqrt(Υ_N + ω* γ̄_N), but the displayed V^N in (5.27) is λ* sqrt(Υ_N) + ω* γ̄_N. The inequality λ* sqrt(Υ_N + ω* γ̄_N) ≤ λ* sqrt(Υ_N) + ω* γ̄_N is not true in general; for instance, with Υ_N = 0 and γ̄_N = 1/N, the left side is λ* sqrt(ω*/N) while the right side is ω*/N. Lemma 5.7 contains the same mismatch, and Proposition 5.11 and Proposition 5.14 rely on V^N as a vanishing error term in the Gronwall estimate. The proof as written therefore has a genuine algebraic gap. The gap is local and repairable: replacing V^N by λ* sqrt(Υ_N + ω* γ̄_N), or by λ*(sqrt(Υ_N) + sqrt(ω* γ̄_N)), preserves the required convergence to zero under Assumption 3.5.","section":"Section 5.1, Lemma 5.6 and Lemma 5.7, Eq. (5.27) vs. (5.34)"}],"minor_comments":[{"comment":"The compatibility inequality (3.2) is a substantive restriction on admissible initial conditions; it fails when μ̄R_0 = 0 and μ̄I_0 has positive mass on ages a with F_x(a) > 0. Since this inequality is used in Proposition 4.1 to bound the force of infection and close the uniqueness argument, the authors should state this restriction prominently as a modeling limitation or self-consistency condition, rather than only providing the heuristic in Remark 3.2.","section":"Assumption 3.1 and Remark 3.2"},{"comment":"In the paragraph following (5.60), the claim that the right-hand side converges by 'Proposition 4.1' should refer to Proposition 5.11, since it is Proposition 5.11 that establishes the convergence of D^N(t).","section":"Proof of Proposition 5.16"},{"comment":"The citation 'Pólya and Szegő [?]' contains a placeholder and should be completed. Additionally, reference [23] appears poorly formatted, and the text contains several OCR-like artifacts (e.g., '/BD' in place of indicator functions, doubled letters, and missing accents); these should be cleaned up in the final version.","section":"Appendix A, Lemma A.1 and references"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Pang, Pardoux, and Velleret prove a functional law of large numbers for an individual-based SIR model with infection-age-dependent infectivity on large random graphs converging to a graphon. This is genuinely new: prior graphon FLLNs (Keliger-Horvath-Takacs, Delmas et al.) were Markovian, and prior non-Markovian FLLNs were not on graphon connectivity. The measure-valued limit (3.6)-(3.9), the PDE representation in Proposition 4.3, and the uniqueness proof in Proposition 4.1 are coherent. The auxiliary mean-field process is well executed, and Proposition 5.9 in Appendix A is a useful standalone result on kernel weak convergence. The literature coverage is thorough and the paper is careful about what is new relative to [20], [37], and the Forien-Pang-Pardoux line.\n\nThe soft spots are real but not fatal. Assumption 3.1's compatibility inequality (3.2) excludes natural initial conditions, e.g., an epidemic starting with infected individuals of positive infection age and no recovered individuals. The paper never flags this in the abstract or introduction; the authors should state whether it is necessary or can be relaxed.\n\nMore seriously, the stress test points to a genuine algebraic gap in Lemma 5.6 (and, by inheritance, Lemma 5.7 and Proposition 5.11). The proof of Lemma 5.6, via Cauchy-Schwarz, establishes U_N <= lambda* sqrt(Upsilon_N + omega* gamma_bar_N), not the stated U_N <= V_N = lambda* sqrt(Upsilon_N) + omega* gamma_bar_N. The claimed inequality is false in the relevant regime, e.g., Upsilon_N = 0 and gamma_bar_N = 1/N. The corrected bound still converges to zero under Assumption 3.5, so Theorem 3.6 is very likely salvageable, but the written proof has a load-bearing error that a referee should require the authors to fix.\n\nThere are also minor textual issues: an incomplete citation in the references and some cross-reference errors.\n\nWho is this for? Researchers working on rigorous scaling limits of epidemic models and on graphon dynamics. It deserves a serious referee. The right outcome is conditional acceptance: the main theorem is significant and the proof strategy is sound, but the authors must repair the V_N bound and discuss the scope of Assumption 3.1.","headline":"A substantial FLLN for non-Markovian SIR on graphons, with a load-bearing but fixable gap in Lemma 5.6 and a restrictive initial-condition assumption that should be flagged.","tokens_in":43319,"tokens_out":2989,"would_cite":true,"duration_ms":29474,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60F17","60K35","92D30"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves a functional law of large numbers for stochastic SIR epidemics on large non-homogeneous random graphs, with a limit described by explicit measure-valued equations.","keywords":["stochastic SIR","infection-age dependent infectivity","measure-valued processes","graphon","non-homogeneous random graph","functional law of large numbers","PDE limit","individual heterogeneity"],"falsifier":"Simulate the finite-$N$ model with a fixed a.e.-continuous kernel such as $\\bar\\omega(x,x')=\\mathbf{1}_{\\{|x-x'|\\le\\delta\\}}$ and compare the empirical susceptible measure with $\\bar\\mu^S_t(dx)=\\exp(-\\int_0^t F(s,x)\\,ds)\\,\\bar\\mu^S_0(dx)$; Theorem 3.6 predicts the difference tends to zero in probability. A reproducible simulation showing a nonzero gap that does not shrink with $N$, or an explicit initial distribution satisfying Assumption 3.1 for which (3.6)-(3.7) admits two distinct solutions, would falsify the claim.","tokens_in":42279,"feed_emoji":"🦠","tokens_out":13045,"duration_ms":125424,"temperature":0.7,"pith_summary":"The paper studies a stochastic SIR epidemic on a random graph of $N$ individuals, where each individual carries a type that controls connectivity and infectivity, and infectiousness depends on time since infection. It proves that as $N\\to\\infty$, the empirical distributions of susceptible, infected, and recovered individuals converge in probability to the unique solution of an explicit system of measure-valued equations, equivalently a PDE on a graphon, the continuum limit of the random contact kernel. The susceptible measure takes the closed form $\\bar\\mu^S_t(dx)=\\exp(-\\int_0^t F(s,x)\\,ds)\\,\\bar\\mu^S_0(dx)$, with the force of infection $F$ built from the limiting graph kernel and the average infectivity profile. If the theorem is right, deterministic PDE/graphon models are the correct large-population approximation for a very wide class of heterogeneous, non-Markovian epidemics, including spatial, age-structured, and activity-structured populations.","feed_headline":"Epidemics on huge random graphs become one deterministic PDE","feed_subtitle":"Stochastic variation and network heterogeneity average out to an explicit graphon PDE as the population grows.","key_machinery":"The proof constructs an auxiliary or intermediate process in which each susceptible individual is infected at rate $F(t,X^N_i)$ using the limiting force of infection rather than the sum over actual infected neighbours in the graph. Because the auxiliary infection times are conditionally independent, the dependence among individuals is decoupled, and the difference between the true and auxiliary processes is controlled by decomposing it into seven terms: age mismatch, kernel approximation error, conditional-mean and variance fluctuations of the contact rates, and errors in the initial measures. Each term is bounded under Assumptions 2.2, 3.3, and 3.5, with a generic weak-convergence result for kernels that are continuous only almost everywhere, and a Gronwall inequality on $\\bar D^N(t)=\\mathbb E^N_0[N^{-1}\\sum_{i}\\sup_{r\\le t}|D^N_i(r)-\\tilde D^N_i(r)|]$ closes the comparison.","core_discovery":"The central result, Theorem 3.6, states that under Assumptions 2.2, 2.4, 2.6, 3.1, 3.3, and 3.5, the LLN-scaled measure-valued processes $(\\bar\\mu^{S,N},\\bar\\mu^{I,N},\\bar\\mu^{R,N})$ converge in probability to $(\\bar\\mu^S,\\bar\\mu^I,\\bar\\mu^R)$ in $D_1\\times D_2\\times D_1$. The limit is the unique solution of the system (3.6)-(3.9), where the force of infection $F(t,x)$ is given by (3.7) in terms of the limiting graph kernel $\\bar\\omega$, the average infectivity $\\bar\\lambda$, and the current susceptible measure. In particular, the susceptible population satisfies $\\bar\\mu^S_t(dx)=\\exp(-\\int_0^t F(s,x)\\,ds)\\,\\bar\\mu^S_0(dx)$. When the recovery-time distributions have densities, the infected measure solves a linear transport PDE with boundary condition $\\tilde\\mu^I_t(dx,0)=F(t,x)\\bar\\mu^S_t(dx)$, so the entire epidemic reduces to a deterministic evolution on the space of types and infection ages.","pith_inferences":["Editorial inference: the same intermediate-model comparison suggests a natural route to a central limit theorem for these measure-valued processes, because the variance terms in the seven-term decomposition carry the $N^{-1/2}$ fluctuations; the paper does not pursue that step.","Editorial inference: the aggregate averaging conditions in Assumption 3.5 leave room for localized superspreading on a negligible fraction of edges, so an extension to unbounded kernels would likely require $L^p$ integral conditions rather than uniform boundedness.","Editorial inference: the explicit exponential formula for $\\bar\\mu^S_t(dx)$ gives a directly testable prediction for the fraction of each type remaining susceptible, which could be compared with simulations or contact-tracing data."],"forward_implications":["For large $N$, the random epidemic trajectory is essentially deterministic: fluctuations around the limiting PDE vanish in probability, so a single PDE solve replaces many stochastic simulations.","The limit supports continuous type spaces and kernels that are continuous only almost everywhere, so spatial or age-structured models with discontinuous contact thresholds are included.","The boundary condition $\\tilde\\mu^I_t(dx,0)=F(t,x)\\bar\\mu^S_t(dx)$ links new infections to the current susceptible measure, making the infection-age distribution computable without tracking the whole graph.","Non-Markovian infectivity profiles, including random individual infection durations with non-exponential laws, obey the same equations, extending earlier Markovian graphon limits."],"supporting_citations":[{"why":"Supplies the intermediate-model construction that drives infection by the limiting force of infection, the key decoupling device.","marker":"[27]"},{"why":"Provides the dense-graphon scaling of contact kernels and the measure-valued state description that the present model extends to non-Markovian SIR.","marker":"[20]"},{"why":"Establishes a graphon FLLN for Markovian density-dependent processes, the baseline class this paper generalizes.","marker":"[37]"},{"why":"Introduces infection-age dependent infectivity models and shows boundedness of the infectivity functions suffices for convergence.","marker":"[26]"},{"why":"Gives the measure-valued FLLN and PDE limit for homogeneous non-Markovian SIR models, which the heterogeneous graph setting embeds in a graphon.","marker":"[46]"},{"why":"Supplies the stochastic-block-model SIR law of large numbers whose scaling is compared with the graphon scaling used in the main theorem.","marker":"[10]"},{"why":"Motivates the heterogeneous-population formulation by showing how population heterogeneity shapes herd immunity.","marker":"[11]"},{"why":"Underlies the propagation-of-chaos viewpoint that justifies decoupling finite-size populations through the limiting infection field.","marker":"[52]"}],"fun_headline_variants":["Stochastic SIR on random graphs collapses to one PDE","Heterogeneous epidemics average out to a graphon PDE","Infection-age SIR converges to a deterministic limit","Large random networks yield a single epidemic law","Measure-valued SIR becomes a graphon transport PDE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument needs the initial recovered measure to dominate the remaining infectivity carried by initially infected individuals, a compatibility inequality that fails for a fresh outbreak with no recovered people and infected individuals who already have positive infection age.","fun_headline_variants_meta":{"raw":{"variants":["Stochastic SIR on random graphs collapses to one PDE","Heterogeneous epidemics average out to a graphon PDE","Infection-age SIR converges to a deterministic limit","Large random networks yield a single epidemic law","Measure-valued SIR becomes a graphon transport PDE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000279,"raw_usage":{"total_tokens":1681,"prompt_tokens":992,"completion_tokens":689,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":613}},"tokens_in":608,"tokens_out":689,"duration_ms":6383,"temperature":1.0,"reasoning_tokens":613,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T23:06:17.442056+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the finite-$N$ model with a fixed a.e.-continuous kernel such as $\\bar\\omega(x,x')=\\mathbf{1}_{\\{|x-x'|\\le\\delta\\}}$ and compare the empirical susceptible measure with $\\bar\\mu^S_t(dx)=\\exp(-\\int_0^t F(s,x)\\,ds)\\,\\bar\\mu^S_0(dx)$; Theorem 3.6 predicts the difference tends to zero in probability. A reproducible simulation showing a nonzero gap that does not shrink with $N$, or an explicit initial distribution satisfying Assumption 3.1 for which (3.6)-(3.7) admits two distinct solutions, would falsify the claim.","supporting_citations":[{"cited_title":"Forien, G","cited_arxiv_id":null,"evidence_quote":"Supplies the intermediate-model construction that drives infection by the limiting force of infection, the key decoupling device."},{"cited_title":"Delmas, P","cited_arxiv_id":null,"evidence_quote":"Provides the dense-graphon scaling of contact kernels and the measure-valued state description that the present model extends to non-Markovian SIR."},{"cited_title":"Keliger, I","cited_arxiv_id":null,"evidence_quote":"Establishes a graphon FLLN for Markovian density-dependent processes, the baseline class this paper generalizes."},{"cited_title":"Forien, G","cited_arxiv_id":null,"evidence_quote":"Introduces infection-age dependent infectivity models and shows boundedness of the infectivity functions suffices for convergence."},{"cited_title":"Pang and ´E","cited_arxiv_id":null,"evidence_quote":"Gives the measure-valued FLLN and PDE limit for homogeneous non-Markovian SIR models, which the heterogeneous graph setting embeds in a graphon."},{"cited_title":"A Law of Large Numbers for SIR on the Stochastic Block Model: A Proof via Herd Immunity","cited_arxiv_id":"2410.07097","evidence_quote":"Supplies the stochastic-block-model SIR law of large numbers whose scaling is compared with the graphon scaling used in the main theorem."},{"cited_title":"Britton, F","cited_arxiv_id":null,"evidence_quote":"Motivates the heterogeneous-population formulation by showing how population heterogeneity shapes herd immunity."},{"cited_title":"Sznitman","cited_arxiv_id":null,"evidence_quote":"Underlies the propagation-of-chaos viewpoint that justifies decoupling finite-size populations through the limiting infection field."}],"review_version":1}