{"id":"5f3c33e9-c221-423e-b56b-7a9764f8729d","arxiv_id":"2411.13130","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"In heterogeneous SIR and SEIR models, the condition for herd immunity is finiteness of the mean recovery time E[1/γ], not the mean recovery rate E[γ].","lead":"This paper proves that in a heterogeneous SIR epidemic model, herd immunity, defined as some susceptibles never getting infected, occurs exactly when the average recovery time is finite. The common summary statistic, the average recovery rate, is shown to be unreliable and can be equal across populations that behave very differently.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assumption (9) is not 'technical': without βmin>0 the iff in Proposition 3 is false, and real contact matrices often violate it, so the abstract overstates the result.","rationale":"The reader correctly identified hypothesis (9) as the weakest point. I examined the proof of Proposition 3 in detail: Part I is straightforward; Part II is valid under (9) provided one fills the standard argument that I(t,x)→0 pointwise when γ(x)>0 (using F∈L^1 and convolution with e^{-γt}), and corrects the constant in (25) from C3 to 1−C3. The real issue is that (9) is not merely technical: removing it falsifies the theorem. Since the paper itself acknowledges (9) but downplays it, the central claim's presentation is stronger than the theorem. This does not change the CONDITIONAL verdict: the theorem is correct under its stated assumptions, but the abstract and discussion should prominently qualify the uniform-transmission requirement.","tokens_in":8162,"tokens_out":14916,"duration_ms":147706,"concrete_test":"Compile the two-group counterexample above and verify analytically: since β12=β21=0, group 2's equations reduce to S'(t,2)=0 and I(t,2)=0, so S(∞,2)=S0(2)>0; hence S(∞)>0, while E[1/γ]=∞. This settles that (9) is necessary. To test practical relevance, take a published age-structured contact matrix (e.g., Prem et al. 2017), set one group's recovery rate to a heavy-tailed distribution with E[1/γ]=∞ and all other parameters per Section 2, and simulate (1)-(3); if S(∞)>0 because of zero contact blocks, the abstract's conclusion is inapplicable to that setting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 3's 'only if' direction relies crucially on βmin>0 from hypothesis (9). The step at (22)-(25) uses βmin to force S(∞,x)≤C3<1 for every trait x, leading to the bound S(∞) ≤ S0 exp(−c∫1/γ), which contradicts S(∞)>0 when ∫1/γ=∞. If (9) is relaxed to allow zero transmission between some trait groups, the equivalence is false. Explicit counterexample: Ω={1,2}, β(1,1)=β(2,2)=β>0, β(1,2)=β(2,1)=0, γ(1)=0, γ(2)=1, S0(i)=1/2, I0(1)>0, I0(2)=0. Group 2 is never infected, so S(∞)≥1/2>0 (herd immunity property holds), yet E[1/γ]=∞ because P(γ=0)=1/2. Thus the 'only if' direction fails. The paper's Section 2 calls (9) 'mainly needed for technical reasons' and lumps it with excluding 'totally immune individuals', but this is a structural assumption without which the central theorem collapses. Realistic contact matrices (age/behavior structure) typically have zero or very small cross-group entries, so the main claim's scope is much narrower than the abstract suggests. The proof under (9) itself is basically sound; minor issues (C3 vs 1−C3 in (25); SEIR case I(0)=0) are repairable and do not affect the conclusion when (9) holds.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies a heterogeneous SIR model (and then an SEIR extension) in which the recovery rate is a trait-dependent function γ(x). Herd immunity is defined as S(∞)>0, the aggregate susceptible fraction remaining positive at infinite time. The paper first gives numerical examples showing that the mean recovery rate E[γ] does not predict herd immunity, then states its central result (Proposition 3): under assumptions (8)–(11), the system has the herd immunity property if and only if the mean recovery time E[1/γ] is finite. Proposition 5 extends this characterization to the SEIR model. The paper concludes that the mean recovery time, rather than the mean recovery rate, is the relevant metric for herd immunity in heterogeneous populations.","tokens_in":8460,"tokens_out":5875,"duration_ms":60903,"significance":"The proposed criterion is elegant and, if established, practically relevant: it gives a sharp, parameter-free condition for whether a positive susceptible fraction survives an epidemic in a heterogeneous population. The forward direction is clean, the proof is self-contained, the examples are informative, and the code is made available. The main caveat is that the equivalence relies on the full-contact assumption 0<βmin≤β(x,y), and the paper's abstract and discussion state the result without this qualifier; as written, the theorem is correct under its hypotheses, but the scope claimed outside the assumptions is broader than what is proved.","major_comments":[{"comment":"The assumption 0<βmin≤β(x,y) is load-bearing for the 'only if' direction and is not merely technical. Without it the equivalence is false: take Ω={1,2}, β(1,1)=β(2,2)=β>0, β(1,2)=β(2,1)=0, γ(1)=0, γ(2)=1, S0(1)=S0(2)=1/2, I0(1)>0, I0(2)=0. Group 2 is never exposed, so S(∞)≥1/2>0 (herd immunity holds), yet E[1/γ]=∞ because P(γ=0)=1/2. The paper's statement in §2 that (9) mainly prevents 'totally immune individuals' mischaracterizes the issue; zero cross-group transmission is a structural feature of many realistic contact matrices. Please restrict the abstract and discussion to models satisfying the full-contact condition, or state clearly which parts of the theorem survive when β has zero entries.","section":"§2 (Assumption (9)) and Proposition 3"},{"comment":"The chain after (23)–(24) is not written correctly: from R(∞,y)≥1−C3 one obtains S(∞,x)≤S(0,x)exp(−∫ βmin γ(y)^{-1}(1−C3)dy), not the bound with C3 in the exponent. The displayed inequality with C3 is not justified unless C3≤1/2, which is not established. The argument is easily repaired by replacing C3 with any positive constant bounded away from zero, but the displayed equation should be fixed.","section":"§3, proof of Proposition 3, Eq. (25)"},{"comment":"The statement 'it can be easily checked that lim_{t→∞}I(t,x)=0' is used to pass from S(∞,x)≤C3 to R(∞,x)≥1−C3. Please include the short argument: S(t,x) and R(t,x) are monotone, so I(t,x) has a pointwise limit; since ∫0∞ γ(x)I(t,x)dt=R(∞,x)≤1 and γ(x)>0 a.e., the limit must be zero. Similarly, in the SEIR proof the conclusion E_Z(∞)=0, I_Z(∞)=1 from (35) needs a few lines of justification rather than being asserted.","section":"§3, proof of Proposition 3, Part II"}],"minor_comments":[{"comment":"The formula for I(t) appears to contain a typo: I(t)=(I(t,1)+S(t,2))/2 should presumably be I(t)=(I(t,1)+I(t,2))/2.","section":"§2.1, Lemma 2 proof, Eq. (14)"},{"comment":"The sentence 'since I(0)>0 we have I(0,1)>0' is not automatic; if I(0,1)=0, positivity follows from I′(0,1)=βS(0,1)I(0)>0.","section":"§2.1, Lemma 2 proof"},{"comment":"Notation such as I(0) and I(t) in the proof of Proposition 3 refers to the aggregate variables defined in (4), but this is not restated at the point of use; please clarify to avoid confusion with the trait-dependent I(t,x).","section":"§3, after Eq. (23)"},{"comment":"In the middle panel, the caption says S(∞)=0.0, but for the γ=0 group the susceptible and infected fractions do not vanish individually; the text should specify that this is the aggregate S(∞).","section":"§2.1, Figure 1 and surrounding text"},{"comment":"The GitHub link contains spaces: 'epidemiology heterogeneous gamma' should be a proper URL or percent-encoded, otherwise it may not be clickable.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about Assumption (9) lands: the central 'if and only if' is not a theorem without the full-contact lower bound, and the counterexample should be acknowledged in the paper rather than dismissed under 'technical reasons'. The proof is otherwise mostly sound and the main idea is worthwhile. I would be willing to accept after the claims are narrowed to the stated hypotheses and the proof details in Eqs. (24)–(25) and in the SEIR part are repaired."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper gives a clean criterion for herd immunity in heterogeneous SIR/SEIR: under their assumptions, S(∞)>0 iff E[1/γ]<∞. The forward direction is solid—it is just the standard exponential bound. The reverse direction is mostly right but has two fixable gaps. The bigger issue is that the βmin>0 assumption (9) is doing real work, despite the paper calling it technical.\n\nWhat is genuinely new: the explicit equivalence is not in the cited works, and the counterexamples showing E[γ] is a bad proxy are simple and instructive. The code is available, which is good.\n\nSoft spots. First, the 'only if' proof assumes pointwise convergence I(t,x)→0 with a handwave ('it can be easily checked'). That is plausible given γ>0 a.e. and finite βmax, but it needs an argument. Second, assumption (9) is not just technical. Without it the theorem is false: take two groups, β(1,1)=β(2,2)>0, β(1,2)=β(2,1)=0, γ(1)=0, γ(2)=1, split initial population equally. Group 2 is never exposed, so S∞≥1/2>0 while E[1/γ]=∞. The paper's Section 2 downplays (9) as mainly for technical reasons, which misleads. Realistic contact matrices often have zero cross-group entries, so the result's scope is narrower than the abstract suggests. Third, the definition of herd immunity as S∞>0 is nonstandard—it is not the classical threshold (1−1/R0). It should be flagged for readers. Minor: in (25) the constant should be 1−C3, not C3; harmless since both are positive. The SEIR part inherits the βmin issue and has a sketchier convergence step.\n\nOverall, the mathematical core is sound under (9), and the paper is honest in stating its assumptions, though the abstract overreaches. It deserves a serious referee who will ask for a proof of the convergence step and a discussion of what happens when βmin=0. I would engage with it.\n\nRecommendation: send to peer review; expect heavy revision but not rejection on methodological grounds.","headline":"The criterion E[1/γ]<∞ is genuine and mostly proved, but the uniform transmission bound βmin>0 is load-bearing, not technical, and the abstract oversells the scope.","tokens_in":9006,"tokens_out":9701,"would_cite":true,"duration_ms":90515,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92D30"],"pacs":[],"model":"deepseek-v4-flash","headline":"In heterogeneous SIR populations, herd immunity holds exactly when the average recovery time is finite; the mean recovery rate is not the deciding quantity.","keywords":["herd immunity","heterogeneous recovery rate","mean recovery time","SIR model","SEIR model","final epidemic size","recovery rate heterogeneity"],"falsifier":"Run the trait-structured SIR system with $\\Omega=[0,1]$, uniform $\\beta=1/4$, and $\\gamma(x)=x/3$, so $\\mathbb{E}[1/\\gamma]=\\infty$; the theorem predicts $S(\\infty)=0$, matching the paper's numerics. To disprove the theorem one would need, under assumptions (8)-(11), any instance with $\\mathbb{E}[1/\\gamma]=\\infty$ and $S(\\infty)>0$; a simulation or analytic example producing that would settle the claim negatively.","tokens_in":7937,"feed_emoji":"🦠","tokens_out":7576,"duration_ms":69174,"temperature":0.7,"pith_summary":"This paper asks which single number tells an epidemiologist whether an epidemic in a population with variable recovery speeds will stop before infecting everyone. The answer it defends is the mean recovery time, the average of $1/\\gamma(x)$ across people, not the mean recovery rate $\\mathbb{E}[\\gamma]$. The central result is that, under the model's assumptions, the heterogeneous SIR system has the herd immunity property $S(\\infty)>0$ if and only if $\\mathbb{E}[1/\\gamma(x)]<\\infty$, and the same criterion holds for the SEIR extension. The paper also exhibits populations with identical average recovery rate that reach opposite fates, so the mean rate alone is not a reliable metric. If the result is right, forecasts of final epidemic size and vaccination needs should be built from the distribution of recovery times, especially their tail, rather than from the average recovery rate.","feed_headline":"Herd immunity hinges on mean recovery time, not recovery rate","feed_subtitle":"A proof for heterogeneous SIR/SEIR populations: the average of 1/γ, not γ, decides who escapes infection.","key_machinery":"The load-bearing object is the trait-structured SIR system (1)-(3), in which each individual carries a trait $x$ with recovery rate $\\gamma(x)$, and the key identity obtained from rewriting the infection term as a recovery-driven integral: $\\frac{d}{dt}\\ln S(t,x)=-\\int_\\Omega \\frac{\\beta(x,y)}{\\gamma(y)}\\frac{d}{dt}R(t,y)\\,dy$. This identity converts the infection history into cumulative recoveries, giving the bound $S(\\infty,x)\\ge S(0,x)\\exp(-\\int_\\Omega \\beta_{\\max}/\\gamma(y)\\,dy)$. The named object deciding the outcome is the mean recovery time $\\mathbb{E}[1/\\gamma(x)]$: if it is finite the exponential bound is positive, while if it diverges the same structure forces $S(\\infty)=0$; the proof handles the case $\\gamma=0$ by showing infecteds with that trait grow exponentially and drag all susceptibles to zero.","core_discovery":"Proposition 3 states the main discovery: assume the transmission function is uniformly positive and bounded, there are no initially recovered individuals, and some initial infectives exist; then the trait-structured SIR equations (1)-(3) have the herd immunity property, meaning a positive fraction of the initially susceptible population remains uninfected, if and only if the population mean of the inverse recovery rate $1/\\gamma(x)$ is finite. The direction 'finite mean recovery time implies herd immunity' follows from an exponential lower bound on the susceptible fraction that depends on $\\int \\beta_{\\max}/\\gamma(y)\\,dy$. The reverse direction shows that if $\\mathbb{E}[1/\\gamma]=\\infty$, the susceptible fraction decays to zero, with the special case $\\gamma=0$ on a non-negligible subgroup forcing total infection. Proposition 5 extends the same equivalence to the SEIR model when the exposure rate is bounded away from zero and infinity. The paper's examples with $\\gamma=1/6$ versus a two-point or density distribution with the same mean illustrate that equal average recovery rates can produce herd immunity in one population and complete infection in another.","pith_inferences":["A practical consequence the paper leaves implicit: surveillance should track the tail of long infectious periods, since an infinite-mean tail alone destroys herd immunity even when most patients recover quickly.","In contact networks with disconnected components the criterion likely localizes: each epidemiologically closed component needs finite mean recovery time for its own susceptibles to be protected.","The theorem yields a testable ordering: among populations matched for mean $\\gamma$, final susceptible fraction should increase as $\\mathbb{E}[1/\\gamma]$ decreases; stratified recovery data could check this.","Vaccination strategy could exploit the result by prioritizing subgroups with the longest infectious periods, since shortening their infectious period directly restores herd immunity even if it barely moves the population mean $\\gamma$."],"forward_implications":["In any heterogeneous SIR population meeting the assumptions, a positive fraction of susceptibles escapes infection exactly when the population-averaged infectious period $\\mathbb{E}[1/\\gamma]$ is finite.","The mean recovery rate $\\mathbb{E}[\\gamma]$ is not a reliable predictor: the paper exhibits three settings with the same mean rate $1/6$ whose final susceptible fractions are approximately $0.42$, $0$, and $0$.","Any non-negligible subgroup that never recovers ($\\gamma=0$) eliminates herd immunity for the whole population, no matter how quickly everyone else recovers.","Herd immunity can fail even when every individual has finite recovery time, because distributions like $\\gamma(x)=x/3$ on $[0,1]$ give $\\mathbb{E}[1/\\gamma]=\\infty$ and empty the susceptible class.","The identical finite-mean-recovery-time criterion governs the SEIR model with an exposed class, provided exposure rates are bounded away from $0$ and $\\infty$."],"supporting_citations":[{"why":"Supplies the classical SIR model and the herd-immunity threshold that the paper generalizes to heterogeneous recovery rates.","marker":"[10]"},{"why":"Provides the heterogeneous SIR formulation with trait-dependent transmission and recovery that equations (1)-(3) follow.","marker":"[14]"},{"why":"Gives the heterogeneous SEIR framework, including recovery rates bounded below, that the SEIR extension builds on.","marker":"[1]"},{"why":"Shows population heterogeneity shifts the herd-immunity level, motivating the search for the correct summary statistic.","marker":"[4]"}],"fun_headline_variants":["Mean recovery time, not rate, decides herd immunity","Herd immunity's real driver: mean recovery time","Why mean 1/γ matters more than γ for herd immunity","Herd immunity threshold depends on mean recovery time","Swap recovery rate for recovery time in herd immunity math"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument that finite mean recovery time is also necessary assumes every trait group can both infect and be infected by every other group, through the lower bound $\\beta_{\\min}>0$ on transmission rates; if some subgroup is epidemiologically isolated, it can remain susceptible no matter how long other groups stay infectious.","fun_headline_variants_meta":{"raw":{"variants":["Mean recovery time, not rate, decides herd immunity","Herd immunity's real driver: mean recovery time","Why mean 1/γ matters more than γ for herd immunity","Herd immunity threshold depends on mean recovery time","Swap recovery rate for recovery time in herd immunity math"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000317,"raw_usage":{"total_tokens":1838,"prompt_tokens":1037,"completion_tokens":801,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":723}},"tokens_in":653,"tokens_out":801,"duration_ms":8063,"temperature":1.0,"reasoning_tokens":723,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:47:58.464809+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trait-structured SIR system with $\\Omega=[0,1]$, uniform $\\beta=1/4$, and $\\gamma(x)=x/3$, so $\\mathbb{E}[1/\\gamma]=\\infty$; the theorem predicts $S(\\infty)=0$, matching the paper's numerics. To disprove the theorem one would need, under assumptions (8)-(11), any instance with $\\mathbb{E}[1/\\gamma]=\\infty$ and $S(\\infty)>0$; a simulation or analytic example producing that would settle the claim negatively.","supporting_citations":[{"cited_title":"Kermack and Anderson G","cited_arxiv_id":null,"evidence_quote":"Supplies the classical SIR model and the herd-immunity threshold that the paper generalizes to heterogeneous recovery rates."},{"cited_title":"Novozhilov","cited_arxiv_id":null,"evidence_quote":"Provides the heterogeneous SIR formulation with trait-dependent transmission and recovery that equations (1)-(3) follow."},{"cited_title":"Final size and convergence rate for an epidemic in heterogeneous populations","cited_arxiv_id":null,"evidence_quote":"Gives the heterogeneous SEIR framework, including recovery rates bounded below, that the SEIR extension builds on."},{"cited_title":"A mathematical model re- veals the influence of population heterogeneity on herd immunity to SARS- CoV-2","cited_arxiv_id":null,"evidence_quote":"Shows population heterogeneity shifts the herd-immunity level, motivating the search for the correct summary statistic."}],"review_version":1}