{"id":"5a8e47b1-995e-4d83-8dbb-700aa9de83a4","arxiv_id":"2607.24121","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A spectral-submanifold reduction collapses SIS epidemic dynamics on networks to a single scalar equation that can predict outbreak onset and, at higher truncation orders or with Pade extension, post-outbreak infection levels and per-node trajectories.","lead":"This paper applies spectral submanifold theory to reduce high-dimensional network dynamics—such as SIS epidemic spread—to a single scalar equation while reconstructing every node's trajectory. The method predicts epidemic thresholds even at low order and, with higher-order or rational (gSSM) extensions, reproduces post-outbreak infection levels on heterogeneous networks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unproved heteroclinic capture: the one-dimensional SSM is asserted, not demonstrated, to contain the endemic equilibrium; post-onset predictions hinge on this.","rationale":"The reader's weakest-assumption pick is the same one I would identify. The coefficient derivation in Appendix A is internally consistent, and the Supplementary error diagnostics show systematic improvement with order—this is genuine supporting evidence, but it does not substitute for proving or verifying the global manifold extension. The missing step is precisely the transition from a local invariant manifold to a global heteroclinic orbit containing the endemic equilibrium. This is not a suggestion of misconduct or sloppiness; it is an unproved step in an otherwise plausible argument. Secondary issues—the absence of a mean-field baseline in the comparisons, the inaccessible 'Google Drive Repository' link, and missing ensemble error bars—are real but secondary; they would not change a conditional verdict. A direct membership test of the endemic equilibrium against W(η*) would settle the primary concern. Since the reader already marked the paper CONDITIONAL and this concern supports that same assessment, I recommend no change to the verdict.","tokens_in":26552,"tokens_out":6957,"duration_ms":66204,"concrete_test":"For a fixed SF realization as in Fig. 4j at β/γ=0.5 (and, if desired, one ER, one SW, and the Rural network): (1) compute the true endemic equilibrium x* of Eq. (4) by long-time integration or a Newton solver; (2) for O(10), O(15), O(20), and the [8/7] gSSM, find all real roots η* of the reduced drift R(η) inside the claimed Taylor/Padé domain and compute ε = ||W(η*) − x*||₂ / ||x*||₂; (3) integrate the full system from x(0) = εu₁ + DFE and the reduced system from η(0)=ε, and compare node-level trajectories over [0,T]. If ε and the maximum node-level deviation are below ~1e-2 for all tested networks, the heteroclinic/SSM-extension assumption is supported; if not, the reduction is only local and the post-onset claims require revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the one-dimensional SSM constructed at the disease-free equilibrium contains the endemic equilibrium and the heteroclinic orbit connecting them. Sec. II.B asserts this directly: 'In all examples considered... the resulting one-dimensional SSM recovers the heteroclinic orbit connecting to the other fixed point.' No proof or numerical verification is provided. This is load-bearing because the post-onset amplitude, saturation, and node-level steady states in Figs. 4–5 all depend on the reduced model reaching the true endemic state. For β/γ above the epidemic threshold, the disease-free equilibrium is unstable (the linearization has a positive eigenvalue), so the relevant object is a one-dimensional unstable manifold; local SSM theory guarantees uniqueness of this manifold, but it does not guarantee that the manifold contains another equilibrium or a global heteroclinic connection. The paper's root analysis (Fig. 2d,h,l) only demonstrates that a truncated amplitude polynomial has a positive real zero inside an estimated convergence domain; it never checks whether the lifting map W evaluated at that zero reproduces the full endemic equilibrium x*. This gap is most acute in scale-free and Rural networks, where the root sits near the convergence boundary and gSSM is invoked. If the heteroclinic capture fails, the one-dimensional reduction cannot represent post-onset saturation, and the claimed node-level/global superiority loses its foundation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a model-reduction framework for nonlinear network dynamics based on spectral submanifolds (SSMs) and a Padé-based globalization (gSSM). For SIS epidemic dynamics on synthetic and empirical networks, the authors construct a one-dimensional invariant manifold tangent to the dominant eigenvector at the disease-free equilibrium, derive the reduced dynamics and lifting map via the parameterization method, and test the reduced model's ability to reproduce node-level trajectories, mean prevalence, and steady-state response curves across β/γ sweeps. The paper also extends the approach to higher-order triadic SIS dynamics and to Lotka–Volterra, gene-regulatory, and logistic-diffusion models, reporting that SSM/gSSM consistently outperform spectral and mean-field baselines.","tokens_in":26957,"tokens_out":4307,"duration_ms":42925,"significance":"If the central claims hold, the paper offers a notable contribution: a single scalar ODE plus a lifting map that reconstructs full node-level behavior in N-dimensional epidemic networks, with a rigorous invariance-equation derivation. The homological coefficient recursions in Appendix A are explicit and standard, the reduced coefficients are derived from the model rather than fitted to full-system output, and the authors provide code and reproducibility documentation. The reported agreement at moderate-to-high truncation orders in homogeneous networks, and the improved performance of gSSM on heterogeneous networks, are plausible and potentially useful. However, the broad 'consistently outperform' claim is currently not fully supported because no mean-field baseline is implemented, only one network instance per topology is used, and the post-onset predictions rest on an unproved assertion about global heteroclinic capture by the local one-dimensional SSM.","major_comments":[{"comment":"The load-bearing assumption that the one-dimensional SSM constructed at the disease-free equilibrium contains the endemic equilibrium and the heteroclinic orbit connecting them is asserted without proof. The text states this directly in Sec. II.B: 'the resulting one-dimensional SSM recovers the heteroclinic orbit connecting to the other fixed point.' This is not a consequence of local SSM existence/uniqueness theory, and for β/γ above threshold the disease-free equilibrium is unstable, so the relevant object is an unstable manifold. The root analysis in Fig. 2(d,h,l) only locates a positive real zero of the truncated reduced vector field; it never checks whether the lifting map evaluated at that zero reproduces the full endemic equilibrium x*. I ask the authors to verify, for each network and parameter value, that ||W(ρ*) − x*|| is small, and to state precisely in what sense the heterocl","section":"Sec. II.B and Figs. 2–5"},{"comment":"The abstract and conclusion claim that SSM/gSSM 'consistently outperform classical spectral and mean-field methods,' but no mean-field baseline is implemented or plotted anywhere in the manuscript or SM. The sweeps in Fig. 4 compare the full system, spectral reduction, modified spectral reduction, and SSM/gSSM only. The conclusion repeats the mean-field claim without supporting data. Since this is a central advertised result, the authors must either add the relevant mean-field baselines (e.g., heterogeneous or quenched mean-field) across the same synthetic and empirical networks, or explicitly qualify the claim.","section":"Abstract, Sec. III, Fig. 4"},{"comment":"The comparison is based on a single network instance for each topology. The text states 'we use the same network instance for each synthetic topology—ER, SW, SF—and the same empirical contact networks... across all analyses,' and Table S1 confirms 'Fixed per topology, reused in all tests.' With N=200 and stochastic network models, especially SF and modular SBM, individual realizations can be atypical. The 'consistently outperform' claim would be much better supported by ensemble statistics (e.g., median and spread over at least a few dozen realizations per topology, with the same SSM order). As written, the reader cannot assess whether the reported accuracy and ranking are robust or instance-specific.","section":"Sec. III (comparison protocol)"},{"comment":"On the Rural network, even high-order Taylor SSM reductions underestimate the steady state for β/γ > 0.2, and accuracy is restored only by gSSM. On SF, the positive real root of the truncated reduced dynamics lies near the estimated convergence boundary (Fig. 2l). This is a concrete regime where the Taylor-based SSM alone does not support the claimed superiority, and the gSSM results are presented without a quantitative error metric or convergence study. Please report the gSSM error (e.g., MSE or steady-state error vs. full system) for these cases and clarify the criterion for 'where necessary' in selecting gSSM.","section":"Sec. III, Fig. 4(l) and Sec. II.B"}],"minor_comments":[{"comment":"Typo: 'Rural, srepresenting' should be 'Rural, representing'.","section":"Sec. III.B"},{"comment":"Typo: 'mises-timates' should be 'misestimates'.","section":"Sec. III.B, Fig. 4(l) caption/description"},{"comment":"The text says 'Fig. 2d,h,j' but the SF panel is labeled (l); please correct to (d,h,l).","section":"Sec. III, Fig. 2 panel references"},{"comment":"The main text fixes η0=0.01, while Table S1 lists 'η0 ∈ [0.01,0.05]'; please reconcile or specify when other values are used.","section":"Table S1 vs. Sec. III"},{"comment":"The Padé construction is described as data-driven via [21], but the paper is equation-driven. Clarify whether the Padé coefficients are determined purely from the Taylor coefficients of the reduced dynamics or involve any trajectory information.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The paper has a sound local derivation and a potentially useful workflow, making rejection inappropriate. However, the central comparative claim is currently overstated relative to the evidence, and the global heteroclinic issue is a genuine theoretical gap. Both are addressable with additional benchmarking and verification, which is why I recommend major revision rather than acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Kavya — quick take on 2607.24121. The useful core is real: an established SSM parameterization applied to SIS-type network dynamics, with a one-dimensional reduced coordinate and a lifting map that reconstructs node-level trajectories, plus an extension to triadic interactions. The homological recursions in Appendix A are standard and correctly set up, and the logistic-diffusion case where all higher coefficients vanish is a nice sanity check. If the central claim holds, this would be a genuinely useful capability for epidemic and ecological modeling.\n\nBut the central claim is not yet supported. The paper asserts in Sec. II.B that the local one-dimensional SSM 'recovers the heteroclinic orbit' to the endemic equilibrium, with no proof and no numerical check that the lifting map W evaluated at the reduced-model fixed point matches the full endemic state. The root plots only show a zero of the truncated drift polynomial; they don't compare W(ρ*) with x*. Because the disease-free equilibrium is unstable above threshold, local SSM theory gives uniqueness of the unstable manifold but says nothing about whether it contains the other equilibrium. This is load-bearing: post-onset amplitudes, saturation, and node-level steady states all depend on it.\n\nThere are smaller issues. The abstract promises 'consistently outperform classical spectral and mean-field methods,' but the comparison figures include only spectral and modified spectral baselines; no mean-field baseline appears. Each synthetic topology is one network instance, with no ensemble error bars, yet the abstract says 'across all realizations.' On the Rural network, even O(15) underestimates until gSSM, so 'consistently' is too strong. And the data-availability statement points to a Google Drive repository without a URL — not acceptable for reproducibility.\n\nNone of this is a takedown. The coefficient derivations are sound, the numerical evidence is suggestive, and the framework is a reasonable new application. The unproved heteroclinic capture is the kind of thing that can be fixed: either prove it for these systems or numerically verify W at the endpoint for every example. If the authors add the missing mean-field baseline, ensemble realizations, and accessible code, this would be a solid paper. Send it to peer review, but ask for a serious revision; with the heteroclinic gap closed, I'd be comfortable with it.","headline":"A solid new application of SSM to network dynamics, but the unproved heteroclinic-capture assumption and the missing mean-field baseline keep the core claim from standing without revision.","tokens_in":27332,"tokens_out":3897,"would_cite":false,"duration_ms":37056,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An SIS epidemic on a complex network can be collapsed to a single scalar differential equation on a spectral submanifold—the smoothest invariant curve through the disease-free equilibrium—that still reproduces every node's infection traject","keywords":["Spectral submanifolds","SSM reduction","node-level model reduction","complex networks","tipping-point prediction","Susceptible-Infected-Susceptible model","higher-order interactions","Padé globalization"],"falsifier":"Within the regime the paper assumes (a real, spectrally separated leading eigenvalue), take a network that satisfies that condition but whose endemic equilibrium lies away from the manifold extending the leading eigenvector—detectable by comparing full-system steady states with the O(20) and gSSM lifted steady states across a fine β/γ grid. If, for any β/γ above the threshold, the reduced scalar equation lacks a positive root inside the Padé convergence domain or the lifted state disagrees with the full endemic state, the central claim fails.","tokens_in":26498,"feed_emoji":"🦠","tokens_out":9577,"duration_ms":93469,"temperature":0.7,"pith_summary":"The paper tries to show that complicated networks do not always need N-dimensional simulation: by constructing a spectral submanifold attached to the slowest-decaying mode at the disease-free equilibrium, the full SIS epidemic reduces to one scalar equation plus a lifting map that rebuilds each node's infection level. If that works, one short differential equation and a handful of coefficients replace a large coupled system while retaining the information that matters—who gets infected, when, and how large the endemic state becomes. The strongest claim is that this one-dimensional reduction predicts the epidemic threshold even at quadratic order, with higher-order terms or a rational extension (gSSM) recovering post-onset saturation, and that it consistently beats classical spectral and mean-field reductions across homogeneous, heterogeneous, modular, empirical, and higher-order-interaction networks. A sympathetic reader would care because it promises interpretable, node-resolved forecasting in epidemiology and ecology without discarding the nonlinearity that creates tipping points.","feed_headline":"One scalar equation captures a full epidemic network at every node","feed_subtitle":"The dominant slow manifold yields tipping-point and node-level forecasts that spectral and mean-field methods miss","key_machinery":"The central object is the spectral submanifold (SSM): the smoothest invariant manifold tangent to the spectral subspace of the linearized network at the selected equilibrium, here the one-dimensional subspace of the slowest-decaying mode. Its reduced dynamics are generated by the invariance equation, DW(η)R(η) = f(W(η)), solved with Taylor expansions to produce the scalar drift R(η) and the lifting coefficients W(η) that rebuild the network state. The globalized extension (gSSM) replaces the truncated Taylor drift with a Padé-type rational function, extending the reduced model beyond the local convergence radius—the step that recovers post-onset saturation on hub-localized networks. The mach","core_discovery":"The paper establishes that, for SIS epidemics on networks, the nonlinear dynamics are governed after a short transient by a single latent coordinate on the one-dimensional spectral submanifold tangent to the leading eigenvector of the linearized disease-free equilibrium. Solving the invariance equation order-by-order yields both the scalar reduced dynamics and the lifting map that reconstructs the full N-dimensional state; where the Taylor series converges poorly, a Padé-type rational continuation (gSSM) extends the reduced model to the high-prevalence regime. The reduced system reproduces the mean prevalence, node-level trajectories, community-level curves, and the tipping onset on Erdős–Ré","pith_inferences":["The reduction's reach depends on a single slow curve persisting all the way from the healthy state to the endemic state; the most natural stress test is a bistable higher-order regime with a saddle-node where no single heteroclinic connection exists, which would require a two-branch or higher-dimensional manifold construction.","The convergence pattern suggests a practical pre-analysis: compute the inverse participation ratio of the leading eigenvector to predict whether low-order or high-order/gSSM reduction will be needed, since hub localization marks the regime where polynomial truncations shrink.","Because the reduced scalar equation is polynomial or rational, it may yield closed-form expressions for effective reproduction numbers and final-size relations in heterogeneous networks, connecting this reduction to standard epidemic theory.","A data-driven version that learns the SSM from node-level time series would make the reduction usable when the network topology is unknown; the paper cites such methods but does not develop that step here."],"forward_implications":["SIS epidemics on homogeneous and small-world networks are effectively one-dimensional after a short transient; a root of the reduced scalar drift gives the endemic steady state, so final prevalence can be computed from a polynomial.","The epidemic threshold is captured even by the quadratic truncation, meaning a low-order SSM computation can estimate the onset of sustained activity without a full network simulation.","Node-level forecasts are available from the same scalar coordinate via the lifting map, giving per-node infection curves that mean-field and classical spectral projections do not provide.","On scale-free networks, modular-bottleneck networks, and higher-order triadic models, low-order truncations are biased; higher-order Taylor terms or gSSM restore agreement with the full system.","The same workflow extends beyond SIS to generalized Lotka–Volterra, gene-regulatory, and logistic–diffusion networked equations, each collapsed to a one-dimensional reduced model, with the logistic–diffusion case exact at quadratic order."],"fun_headline_variants":["Single equation maps entire epidemic network","One coordinate predicts network tipping onset","Scalar model beats spectral methods on contagion","One-dimensional reduction captures SIS dynamics","SSM reduces network epidemics to one scalar law"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The reduction is built only at the healthy equilibrium, and the whole argument rests on the stated but unproved premise that the one-dimensional curve found there continues all the way to the infected steady state, carrying the transition between them; if that connection does not exist for some network, the reduced equation can only describe the early near-healthy phase.","fun_headline_variants_meta":{"raw":{"variants":["Single equation maps entire epidemic network","One coordinate predicts network tipping onset","Scalar model beats spectral methods on contagion","One-dimensional reduction captures SIS dynamics","SSM reduces network epidemics to one scalar law"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000126,"raw_usage":{"total_tokens":941,"prompt_tokens":731,"completion_tokens":210,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":475,"completion_tokens_details":{"reasoning_tokens":147}},"tokens_in":475,"tokens_out":210,"duration_ms":3242,"temperature":1.0,"reasoning_tokens":147,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T22:57:11.773842+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Within the regime the paper assumes (a real, spectrally separated leading eigenvalue), take a network that satisfies that condition but whose endemic equilibrium lies away from the manifold extending the leading eigenvector—detectable by comparing full-system steady states with the O(20) and gSSM lifted steady states across a fine β/γ grid. If, for any β/γ above the threshold, the reduced scalar equation lacks a positive root inside the Padé convergence domain or the lifted state disagrees with the full endemic state, the central claim fails.","supporting_citations":[],"review_version":1}