{"id":"48bc4e4d-8a0e-4ea9-bdab-e279324b3ecb","arxiv_id":"2411.18541","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A modified SIRS model with a feedback-driven recovery rate produces endogenous popularity cycles, and an in-sample comparison to Google Trends finds it closer to residuals than random walks in 67% of queries.","lead":"This paper adds a feedback loop to the standard SIRS epidemic model, making the rate at which people lose interest depend on how many already have the idea, so that popularity naturally cycles. The authors analyze the model's mathematics and compare it against Google Trends data, claiming it beats random walk baselines in 67% of cases.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Empirical comparison is asymmetric: β is fitted in-sample to each residual while the random-walk baseline is unfitted; the 67% DTW result may reflect fitting flexibility, not endogenous oscillation.","rationale":"The theoretical contribution is a legitimate extension: the feedback rule is clearly specified, the fixed-point and local-stability analysis is coherent, and the Hopf-bifurcation argument is plausible, though the stability of the bifurcating cycle is only numerically supported. The load-bearing weakness is the empirical validation, which is the part that connects the model to the paper's headline about real-world popularity. The comparison is not fair: the SIRS model has a parameter fitted to each query, while the random-walk baseline is unfitted and independent. DTW's time-warping flexibility compounds the problem, so the 67% statistic likely measures in-sample fitting advantage rather than the presence of the proposed feedback mechanism. The reader's weakest assumption identifies exactly this issue, and my proposed null-model and out-of-sample checks would settle whether the empirical claim survives a fair test. Because the theoretical model may still be valid and could be validated more rigorously, the appropriate verdict remains conditional rather than reject or accept.","tokens_in":16809,"tokens_out":5989,"duration_ms":61011,"concrete_test":"For each of the 1,000 queries, repeat the DTW comparison using a null model with the same fitting protocol: fit a sinusoid A sin(ωt+φ) (or an AR(2) estimated on the residual) by minimizing DTW to the residual, then compare this fitted-null distance against the average unfitted random-walk distance exactly as in Fig. 3. If the fitted sinusoid/AR(2) beats the random-walk baseline in at least 67% of queries, the current evidence cannot distinguish the SIRS feedback mechanism from generic fitted smoothness. A complementary check: fit β on the first half of each residual, evaluate DTW on the second half, and require the model to retain its advantage over random walks out of sample.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4 and Appendix C.3 compare each query's detrended residuals against (i) the SIRS model with β chosen per query to minimize the DTW distance, with ξ=0.1 and α set just beyond the Hopf threshold, and (ii) the average DTW distance to 500 random walks that have no fitted parameters and are independent of the query. This asymmetry is load-bearing: the model's β is optimized on the target series, DTW already permits arbitrary time-axis warping, and the residuals are standardized to [-1,1]. A smooth periodic trajectory with one free parameter can therefore closely match any sufficiently smooth residual, so the reported 67% does not establish that residuals have the specific SIRS feedback structure. It may simply show that a fitted smooth curve is closer than an unfitted random-walk average. No out-of-sample evaluation, error bars, or comparison against an equally flexible null model is provided, leaving the paper's central empirical claim unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a three-dimensional ODE extension of the SIRS model in which the recovery rate Gamma is not constant but evolves according to dGamma/dt = Gamma(alpha I - delta S), encoding interest saturation and influencing enthusiasm. The authors characterize the local stability of the unique interior fixed point (Proposition 1), prove a Hopf bifurcation for the restricted case alpha = delta (Proposition 2), and report numerical evidence of limit cycles. They then apply the model to 1000 Google Trends queries: after detrending and deseasonalizing, they compare the residuals to the model's I(t) (with the transmission rate beta fitted per query) and to unfitted random walks using Dynamic Time Warping, reporting that the model matches better in 67% of cases. The central claims are that the feedback mechanism generates endogenous periodic cycles and that these cycles are visible in real-world search-volume residuals.","tokens_in":16991,"tokens_out":3927,"duration_ms":36950,"significance":"If fully established, the paper would provide a simple and tractable mechanism for endogenous oscillations in idea diffusion, with an explicit empirical implication and public code on GitHub. The analytical derivation of local instability and the Hopf bifurcation is a solid start, and the manuscript is careful to separate trend and seasonality before testing for oscillatory residuals. However, two load-bearing gaps currently limit the contribution: the proof does not establish that the bifurcating cycles are stable (attracting), and the empirical validation compares an in-sample fitted model against an unfitted baseline, so the 67% result does not yet discriminate the proposed feedback mechanism from generic smoothness or flexible curve fitting. These issues are fixable, which is why I recommend major revision rather than rejection.","major_comments":[{"comment":"Proposition 2 proves that the fixed point loses stability via a Hopf bifurcation at alpha = beta + xi + sqrt(xi(beta + xi)), but it does not prove that the resulting limit cycles are stable. The text in Section 3 states that the feedback mechanism leads to 'stable limit cycles', and Section 3.1 says extensive numerics support 'stable limit cycles appear', but a Hopf bifurcation can be subcritical, producing unstable cycles, or supercritical, producing attracting cycles. Please compute the first Lyapunov coefficient (or provide a center-manifold reduction) to determine the criticality, or explicitly label the stability claim as numerical only. This is load-bearing because the paper's theoretical contribution is the existence of persistent, observable cycles, not merely the loss of stability of a fixed point.","section":"Section 3.2, Proposition 2"},{"comment":"The empirical comparison is asymmetric: for each query, beta is chosen by minimizing the in-sample DTW distance between the residual series and the model, while the random-walk baseline is generated unfitted with no parameter adjustment. The model therefore benefits from per-series optimization, and DTW already permits arbitrary time-axis warping. As a result, the reported 67% closer-match rate does not establish that residuals have the specific SIRS feedback structure; it may only reflect that a smooth curve with one free parameter, when optimally warped, is closer to a smooth residual than the average of unfitted random walks. Please add an out-of-sample evaluation (e.g., split the series, fit beta on a training segment, measure DTW on a holdout segment) and compare against at least one equally flexible null model that is fitted in the same way, such as a sinusoid with fitted amplitude, frequency, and phase, or an AR(2) process fitted per query. Also report standard errors or confidence intervals for the 67% statistic.","section":"Section 4 and Appendix C.3"},{"comment":"The definition of the success rate is ambiguous. The text says residuals 'systematically show a closer match' to the model in 67% of cases, but the preceding explanation defines a blue circle as the case where the model DTW distance is smaller than the first quartile of the random-walk distance distribution. These are two different counting rules: the fraction below the diagonal (model distance < random-walk average distance) and the fraction below the first quartile. Please specify exactly which criterion gives 67%, and report both counts if appropriate. This matters for interpreting the strength of the empirical claim.","section":"Section 4, Figure 3"}],"minor_comments":[{"comment":"There is a typo: 'computaions' should be 'computations'.","section":"Section 3.1, after Figure 1"},{"comment":"The abbreviation 'DWT' is used inconsistently; the text refers to 'best value for beta ... in terms of DWT' and Figure 3's caption uses 'DWT distances', while the standard abbreviation defined in the paper is DTW (Dynamic Time Warping).","section":"Section 4"},{"comment":"The parameter delta is not specified in the empirical setup; it would be helpful to state explicitly whether the restricted model alpha = delta is used (as in Proposition 2), since the Hopf bifurcation result applies only in that case.","section":"Appendix C.3"},{"comment":"The derivation of the sufficient condition for Delta > 0 is algebraically dense; adding a brief explanatory sentence about the sign of the numerator and denominator would improve readability.","section":"Section 3.2, Eq. (20)"}],"recommendation":"major_revision","confidential_remarks":"The paper's theoretical core is reasonable but the empirical section overstates its case: the in-sample fitted-vs-unfitted comparison is the main issue, and the 67% claim is likely to be challenged. The stability of the limit cycles is also not proven, only the Hopf bifurcation. Both are fixable with additional analysis, so major revision is appropriate. The manuscript would benefit from a more careful framing of the empirical results as an illustration rather than a definitive validation of the feedback mechanism."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Rough take: the mathematical core is a genuine, small extension that deserves serious attention; the empirical validation is too weak to support the paper's central claim about real-world popularity dynamics.\n\nWhat's actually new: the recovery rate feedback in Eq. (4) — dΓ/dt = Γ(αI − δS) — is a clean way to get endogenous oscillations out of a three-variable SIRS system. The stability condition (Prop 1) and the Hopf bifurcation result (Prop 2, with the β > (11/25)ξ condition) are new and, as far as I can tell, correct. The proof in Appendix A is standard but careful. The paper also does a solid job placing the model in the diffusion and opinion dynamics literature, and it does not oversell the theoretical contribution: it is a tractable extension, not a revolution.\n\nThe soft spots are on the empirical side, and they are load-bearing. The DTW comparison in Section 4 fits β per query in-sample to minimize the distance, then compares against the average of 500 random walks that have no fitted parameters. That asymmetry alone could explain the 67% result: any smooth curve with a free parameter will generally beat an unfitted random walk under DTW. The model is also always tuned to be just past the Hopf threshold (α = β + ξ + √(ξ(β+ξ)) + 0.01), so it is effectively a family of smooth periodic curves, not a sharp test of the specific mechanism. There are no error bars, no out-of-sample splits, and no comparison against an equally flexible null (say, an AR(2) or a fitted sinusoid). So the claim that the model 'captures' Google Trends residuals is not supported by the evidence as presented.\n\nOne more point: the theoretical claim of 'stable limit cycles' is stated without proving supercriticality. Proposition 2 proves a Hopf bifurcation, not that the cycles are attracting. The numerics suggest it, but a referee should ask for the first Lyapunov coefficient or a softer statement of what is proved. Also, the code is referenced as 'this link' without a URL, which is a reproducibility problem.\n\nBottom line: send it to peer review. The model is worth having and the theory section is solid enough to justify referee time, but the empirical validation needs substantial rework before the data claim can stand.","headline":"Tractable new SIRS variant with a honest theory section, but the Google Trends validation is too asymmetric to carry the paper's empirical claim.","tokens_in":17523,"tokens_out":4219,"would_cite":true,"duration_ms":35430,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["34C23","34C25","37G15"],"pacs":[],"model":"deepseek-v4-flash","headline":"A minimal feedback rule that makes the rate of losing interest rise with overexposure and fall with remaining potential turns the SIRS opinion model into a source of stable popularity cycles, and these cycles match detrended Google Trends…","keywords":["Idea diffusion","SIRS model","Opinion dynamics","Endogenous fluctuations","Limit cycles","Hopf bifurcation","Google Trends","Dynamic time warping"],"falsifier":"Run the same per-query $\\beta$-fitting and DTW comparison on control series that have the same length, variance, smoothness, and autocorrelation as the real residuals but are generated by a stationary process with no feedback (for example, shuffled residuals or matched ARMA simulations). If the fitted model beats random walks on those controls at a rate close to 67%, the reported result is an artifact of fitting and series texture rather than evidence for the feedback mechanism.","tokens_in":16583,"feed_emoji":"📈","tokens_out":7584,"duration_ms":60756,"temperature":0.7,"pith_summary":"This paper claims that the rise and fall of an idea's popularity need not be imposed from outside: a minimal internal feedback can make popularity oscillate on its own. The authors take the epidemiological SIRS model, read the three compartments as people open to an idea, actively spreading it, and temporarily tired of it, and let the rate at which spreaders lose interest depend on the current state: it rises with the number of spreaders (interest saturation) and falls with the number of people still open to the idea (influencing enthusiasm). With that feedback, the model's steady state becomes unstable and the system settles into stable limit cycles, so popularity naturally revives and collapses without any external shock. They then compare the model's oscillations with detrended and deseasonalized Google Trends residuals for 1,000 queries and report that the model fits better than a random walk in 67% of cases. If this is right, it gives a one-equation mechanism for fads, virality, and opinion cycles that many current models relegate to exogenous noise.","feed_headline":"A feedback rule makes idea popularity cycle on its own","feed_subtitle":"Modified SIRS model beats random walks on 67% of 1,000 Google Trends queries after detrending.","key_machinery":"The load-bearing object is the endogenous recovery rate $\\Gamma(t)$, whose log evolves through $\\frac{d\\Gamma}{dt} = \\Gamma(\\alpha I - \\delta S)$ alongside the SIRS equations for $S$ and $I$. This single state-dependent rate encodes the two psychological mechanisms: the positive $\\alpha I$ term says widespread promotion accelerates loss of interest (interest saturation), and the negative $\\delta S$ term says a large pool of potential adopters makes current promoters hold on longer (influencing enthusiasm). The rate is what destabilizes the otherwise stable SIRS fixed point; the Hopf bifurcation proof for $\\alpha = \\delta$ shows that the resulting oscillations are genuine stable limit cycles, not transient decay or noise.","core_discovery":"The central claim is that the feedback rule $\\frac{d\\Gamma}{dt} = \\Gamma(\\alpha I - \\delta S)$, appended to the standard SIRS equations, is enough to turn the system into a generator of sustained popularity cycles. Here $\\Gamma$ is the recovery rate, $I$ the fraction actively promoting the idea, $S$ the fraction still open to it, and $\\alpha, \\delta > 0$. The paper proves, as Proposition 1, that the unique interior fixed point is locally unstable exactly when $\\alpha > \\beta + \\xi$ and $\\delta \\geq \\frac{\\alpha^2\\xi}{(\\alpha-\\beta)(\\alpha-\\beta-\\xi)}$, and, as Proposition 2, that in the restricted case $\\alpha = \\delta$ a Hopf bifurcation occurs at $\\alpha = \\beta + \\xi + \\sqrt{\\xi(\\beta+\\xi)}$ provided $\\beta > \\frac{11}{25}\\xi$, so stable limit cycles emerge from the fixed point. The empirical section then claims that these periodic solutions, with the transmission rate $\\beta$ chosen per query to minimize Dynamic Time Warping distance, match the residuals of 1,000 Google Trends series after trend and seasonality removal better than the average of 500 random walks in 67% of cases. The intended upshot is that observed popularity volatility can be an intrinsic property of the diffusion process rather than a sequence of exogenous shocks.","pith_inferences":["The same feedback could plausibly be inserted into simpler or network-based compartmental models, where it would likely produce oscillations as well; the paper does not test this.","A stricter out-of-sample test would split each time series, fit $\\beta$ on the first half, and evaluate the DTW distance on the second half; the paper's in-sample fitting may overstate the 67% figure.","The model predicts that ideas with strong saturation and strong enthusiasm parameters should show more regular, longer-lasting cycles, which could be checked by sorting Google Trends queries by estimated cycle regularity."],"forward_implications":["Popularity cycles can be explained without external shocks: the same SIRS-plus-feedback system produces recurring booms and busts from its own dynamics.","There is a sharp parameter threshold: once $\\alpha$ and $\\delta$ are large enough to satisfy the instability condition, small perturbations grow into limit cycles rather than decaying, so the model predicts a qualitative switch from damping to sustained oscillation.","The Hopf bifurcation locates that switch explicitly in the symmetric case, so for $\\alpha = \\delta$ the period and amplitude of the cycle are determined by the model's parameters and can be computed from the local crossing.","The empirical comparison implies that a substantial share of Google search dynamics after detrending is structured oscillation rather than pure noise, since the model beats a random walk in 67% of 1,000 queries."],"supporting_citations":[{"why":"Provides the standard SIRS model and its steady-state behaviour that the paper modifies with feedback.","marker":"Anderson and May (1991)"},{"why":"Shows nonlinear oscillations in epidemic models, the tradition the paper extends by making the recovery rate state-dependent.","marker":"Hethcote et al. (1981)"},{"why":"Supplies the stability and bifurcation theory used to analyze the fixed point and prove the Hopf bifurcation.","marker":"Guckenheimer and Holmes (2013)"},{"why":"Gives the Hopf bifurcation result invoked for Proposition 2.","marker":"Marsden and McCracken (2012)"},{"why":"Source of Dynamic Time Warping, the distance metric used in the empirical comparison.","marker":"Hastie et al. (1991)"},{"why":"Evidence for interest saturation in groups, motivating the positive $\\alpha I$ term.","marker":"Jetten and Hornsey (2014)"},{"why":"Economic analysis of cultural transmission behind the influencing enthusiasm mechanism, motivating the negative $\\delta S$ term.","marker":"Bisin and Verdier (2001)"}],"fun_headline_variants":["Feedback rule alone can drive idea popularity cycles","SIRS tweak creates intrinsic idea fads and crashes","No shocks needed: feedback rule cycles idea popularity","Modified SIRS beats random walks on 67% of trends","One feedback term self-generates popularity cycles"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The empirical evidence depends on the assumption that detrended and deseasonalized Google Trends residuals stand in for the model's infected population, and that fitting the transmission rate $\\beta$ per query to minimize in-sample Dynamic Time Warping distance, then comparing the fitted model against unfitted random walks, is a valid way to show the residuals' oscillations are endogenous.","fun_headline_variants_meta":{"raw":{"variants":["Feedback rule alone can drive idea popularity cycles","SIRS tweak creates intrinsic idea fads and crashes","No shocks needed: feedback rule cycles idea popularity","Modified SIRS beats random walks on 67% of trends","One feedback term self-generates popularity cycles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000638,"raw_usage":{"total_tokens":2969,"prompt_tokens":1003,"completion_tokens":1966,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":619,"completion_tokens_details":{"reasoning_tokens":1892}},"tokens_in":619,"tokens_out":1966,"duration_ms":15121,"temperature":1.0,"reasoning_tokens":1892,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:06:53.004178+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same per-query $\\beta$-fitting and DTW comparison on control series that have the same length, variance, smoothness, and autocorrelation as the real residuals but are generated by a stationary process with no feedback (for example, shuffled residuals or matched ARMA simulations). If the fitted model beats random walks on those controls at a rate close to 67%, the reported result is an artifact of fitting and series texture rather than evidence for the feedback mechanism.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the standard SIRS model and its steady-state behaviour that the paper modifies with feedback."},{"cited_title":"and Holmes, P","cited_arxiv_id":null,"evidence_quote":"Supplies the stability and bifurcation theory used to analyze the fixed point and prove the Hopf bifurcation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the Hopf bifurcation result invoked for Proposition 2."},{"cited_title":", Kishon, E","cited_arxiv_id":null,"evidence_quote":"Source of Dynamic Time Warping, the distance metric used in the empirical comparison."},{"cited_title":"and Hornsey, M","cited_arxiv_id":null,"evidence_quote":"Evidence for interest saturation in groups, motivating the positive $\\alpha I$ term."},{"cited_title":"and Verdier, T","cited_arxiv_id":null,"evidence_quote":"Economic analysis of cultural transmission behind the influencing enthusiasm mechanism, motivating the negative $\\delta S$ term."}],"review_version":1}