{"id":"debe85cf-7d10-4714-a5d7-9dadb5da0c27","arxiv_id":"2511.12701","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new L1-trajectory dissimilarity measure for generalized Lotka-Volterra systems on networks detects dynamical differences missed by edge-count statistics and tracks a stability transition as negative interactions grow.","lead":"This paper proposes a new metric for comparing generalized Lotka-Volterra systems—a standard model of interacting species—by tracking the accumulated population difference between two versions of the system over time. It shows the metric can rank network graphs by dynamical similarity and can flag when rising fractions of negative interactions drive the system toward instability, although the findings are based on selected synthetic networks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claim that sign-flip fraction controls stability transitions rests on a single 12-node clique network; no evidence the observed thresholds generalize across topology, size, or weight statistics.","rationale":"The reader's weakest_assumption already identifies the single-topology generalization as the main soft spot. My analysis agrees and sharpens it: the instability-prediction claim is also under-supported because D_max is a symptom of divergence rather than an independent predictor, and the observed thresholds could be artifacts of the specific clique-ring topology and uniform weight distribution. These concerns do not invalidate the methodological framework—the dissimilarity measure is well-defined and clearly demonstrated on several synthetic examples—but they do require the authors to qualify the abstract and provide broader numerical evidence. The appropriate verdict remains CONDITIONAL, as the reader recommended, so no change is needed.","tokens_in":16370,"tokens_out":4915,"duration_ms":43407,"concrete_test":"Run the Sec. 3.3 protocol for at least three additional network ensembles: (a) same N=12 but with two cliques of six and four cliques of three; (b) N=24 and N=48 using three cliques of N/3, keeping the same mean degree as the clique-ring graph; (c) Erdős–Rényi directed graphs with N=12 and matched mean degree, with weights drawn from U(0,1]. For each ensemble and each p, compute ⟨D_max⟩ over ≥1000 realizations and record (i) the p at which ⟨D_max⟩ first exceeds 10% of its maximum and (ii) the p at which more than 50% of realizations diverge. If these p-values vary by more than ±0.05 across ensembles, the claimed universal transition is not established. Additionally, test whether the sharp rise at 1−p≈0.2 can be used as a predictor by checking if it precedes the divergence threshold in at least two of these ensembles.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's conclusion that \"the fraction and distribution of negative interactions control the transition from stable to unstable dynamics\" (and the related claim that dissimilarity measures can \"predict instabilities\") is supported almost entirely by Fig. 5(e) (and Fig. 6(e) for localized perturbations). These simulations use a single network family: N=12 nodes arranged as three cliques of four nodes connected by a ring, with weights Λ_ij independently drawn from U(0,1] and signs flipped with probability 1−p. Only 200 realizations per p are reported. No variation in N, community structure, degree sequence, or weight distribution is presented. In random Lotka–Volterra systems, stability thresholds are known to depend on N, interaction strength, and connectivity (May 1972; Allesina & Tang 2012); without a scaling analysis or a sweep over these parameters, the broad generality claimed is unsupported. Moreover, the \"prediction\" of instability is not established as a forward-looking diagnostic: D_max is computed after simulating both the reference and the sign-flipped system, so it is a post-hoc measure of trajectory divergence. When the modified system diverges, D_max necessarily diverges because the L1 distance grows with the unbounded populations; the sharp rise at 1−p≈0.2 could in principle serve as an early warning, but the paper does not demonstrate its predictive value or robustness. Thus the central claim about the role of negative interactions in controlling stability needs qualification or additional evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a dissimilarity measure for two generalized Lotka-Volterra (gLV) processes that share initial conditions but differ in interaction parameters, network structure, or governing equations. The measure D(t) is defined in Eq. (13) as the normalized L1 distance between the two trajectories, with summary statistics D_max (Eq. 14) and the long-time limit D_inf. The authors apply the framework to two-species predator-prey systems, to all non-isomorphic connected directed graphs with N=3, and to a 12-node modular network with randomly sign-flipped interactions. They conclude that the fraction and distribution of negative interactions control the transition from stable to unstable dynamics and that dissimilarity measures can predict instabilities. The core definition is simple and well posed, and the N=3 census is instructive, but several broad interpretive claims go beyond the evidence presented.","tokens_in":16619,"tokens_out":6787,"duration_ms":59930,"significance":"If the full claims were supported, the framework would supply a practical, systematic method for comparing gLV systems across parameter, topology, and functional-form changes, with potential applications in ecology and microbiome research. The core measure is mathematically clean and parameter-free, and the three-node graph census (10^4 realizations per pair) convincingly shows that simple aggregate structural statistics such as edge count do not order dynamical similarity. The extension to different nonlinearities in Section 3.5 is a useful generalization. However, the alleged stability-transition thresholds are derived from one network topology and from a post-hoc divergence measure; the advertised predictive power is not demonstrated. The strengths are real but the conclusions need to be scaled back or supported by additional experiments.","major_comments":[{"comment":"The abstract and conclusions claim that D_max 'predicts instabilities' and serves as an early-warning diagnostic. As defined, D_max is computed after both the reference and modified systems have been integrated; when the modified system diverges, the L1 distance in Eq. (13) grows without bound, so D_max necessarily diverges. The sharp rise at 1-p≈0.2 in Fig. 5(e) is therefore a post-hoc description of the stability boundary, not a prediction. To support 'predict instabilities,' the authors would need an out-of-sample or pre-divergence test, e.g., examining whether D(t) exhibits a characteristic signature before divergence occurs. As written, the predictive claim is unsupported.","section":"Abstract; §4; Eq. (14)"},{"comment":"The transition near 1-p≈0.2 and the localized-vs-global perturbation contrast are demonstrated for a single network configuration: N=12, three cliques of four connected by a ring, with Λ_ij iid U(0,1] and 200 realizations per p. No variation in N, community structure, degree sequence, weight distribution, or self-regulation strength is reported. Since random-matrix stability thresholds for Lotka-Volterra systems are known to depend on N, connectivity, and interaction variance (May 1972; Allesina & Tang 2012, both cited in §2.3), the broad conclusion that 'the fraction and distribution of negative interactions control the transition from stable to unstable dynamics' is not supported by the evidence. A scaling analysis or a sweep over these parameters is required.","section":"§3.3–3.4; Figs. 5(e), 6(e)"},{"comment":"In Sections 3.3 and 3.4 the reference interaction matrix Λ always has all positive entries drawn from (0,1], and the modified matrix Λ* is obtained by sign-flipping each entry with probability 1-p. Thus 1-p is not the fraction of negative interactions in the ecosystem; it is the probability that a previously positive interaction becomes negative. The reference state contains no negative interactions, so the experiments do not address how mixtures of positive and negative interactions in both systems affect stability. This conflates perturbation size with interaction-type composition and should be stated explicitly and discussed.","section":"§3.3; construction of Λ and Λ*"},{"comment":"The text states that the differences between graphs G1 and G2 are 'invisible at the adjacency-matrix level.' This is an overstatement: G1 and G2 are non-isomorphic, so their adjacency matrices differ, and standard adjacency-based metrics such as Hamming distance distinguish them. What actually coincides are aggregate quantities: the number of edges and the sum of matrix differences with G5. The legitimate point is that these simple aggregate statistics do not order the systems by dynamical similarity. The abstract's phrase 'differences that are invisible at the adjacency-matrix level' should be revised accordingly.","section":"§3.2; Fig. 3 discussion"}],"minor_comments":[{"comment":"The caption labels the panels as '(a) p=0.8, (b) p=0.7, (c) p=0.6, (b) p=0.5'; the last label should be (d).","section":"Fig. 7 caption"},{"comment":"The double bars ||·|| are used for the absolute value of a scalar quantity. Using single vertical bars would avoid confusion with a norm.","section":"Eq. (13)"},{"comment":"No details are given for the numerical integration (solver, step size, divergence criterion, simulation horizon, or handling of unbounded growth). The location of the threshold in Fig. 5(e) may depend on these choices; please report them.","section":"§3.3; numerical methods"},{"comment":"The text says the reference process follows Eq. (12) and the modified process follows Eq. (15) with the same Λ. It would be helpful to state explicitly that the same rescaling and initial-condition ensemble are used for both, to make the comparison fully parallel.","section":"§3.5; Eq. (15)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the journal's scope as a statistical-physics treatment of dynamical dissimilarity, and the core definition is sound. The main risk is that the advertised claims about stability prediction and generality substantially exceed the evidence. I believe these issues are fixable within the manuscript's scope by rewording the conclusions and adding targeted numerical experiments (varying N, topology, and weight statistics) or by explicitly presenting the results as a case study. I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should read this if you care about comparing dynamical processes on networks. The core measure is a normalized L1 distance between two gLV trajectories with the same initial conditions, and it is well-defined and easy to compute. The two-species analysis is clear, and the exhaustive N=3 graph comparison in Fig. 3 is the best part: it genuinely shows that pairwise dynamical similarity does not track simple structural counts. For instance, G2 is closer to G5 than G1 is, even though both have the same number of edges and identical aggregate entry differences. That is a real, reproducible observation and it makes the case for dynamical comparison.\n\nThe gamma asymmetry in Section 3.5 is also interesting—the dissimilarity is much larger for gamma > 1 than for gamma < 1, even at the same absolute deviation from the reference. That is not obvious a priori and suggests the measure can compare across functional forms.\n\nNow the soft spots, in proportion. The biggest problem is the stability claim. The abstract says the fraction of negative interactions 'control the transition from stable to unstable dynamics,' and the conclusion calls this a way to 'predict instabilities.' What the paper actually shows is one 12-node network (three cliques of four in a ring), one weight distribution U(0,1], 200 realizations per p, and no scaling with N or topology. That is a numerical observation about a single ensemble, not a general result. The stress-test note is right: in random Lotka-Volterra systems, thresholds depend on N and interaction strength, so the generality claim is unsupported.\n\nAlso, D_max is not a predictive diagnostic in the way the abstract implies. It is computed after simulating both systems, so when the modified system diverges, the L1 distance diverges by construction. The sharp rise at 1-p ≈ 0.2 could be an early-warning candidate, but the paper does not test whether it predicts instability before it happens. That needs to be softened.\n\nThe abstract's 'invisible at the adjacency-matrix level' is an overstatement. The three-node example compares graphs with equal edge counts, not identical adjacency matrices—so the difference is visible if you look at the full matrix. The point is still valid (edge count is not enough), but the wording oversells.\n\nMinor issues: no code or data provided, integration details are thin, and there are small notation inconsistencies. None of that is fatal.\n\nBottom line: the framework is useful and the N=3 and gamma results deserve airtime, but the paper overclaims on stability and generality. I would send it to peer review, not desk reject, but with a clear request to either add a multi-topology scaling analysis or limit the claims. A serious referee could make this into a solid methodological note.\n\nFor my own work, I would not cite it in the next year—the measure is too simple to need this reference, and the stability claims are not yet reliable. But I would bring it to a reading group if we were discussing network-comparison methods.","headline":"A clean but conventional trajectory-distance measure for gLV systems, with a few genuinely nice demonstrations and one overblown stability claim that rests on a single network topology.","tokens_in":17205,"tokens_out":2193,"would_cite":false,"duration_ms":22786,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper introduces a dissimilarity measure that compares two generalized Lotka-Volterra dynamical processes sharing initial conditions but differing in interaction parameters, network topology, or governing equations, and claims it can de","keywords":["generalized Lotka–Volterra","dissimilarity measures","network dynamics","ecological stability","modular networks","nonlinear dynamics","predator–prey","structural sensitivity"],"falsifier":"Repeat the sign-flip protocol on a different network geometry (e.g., two cliques, a scale-free modular graph, or a larger number of cliques) or with weights drawn from an exponential or Gaussian distribution, and check whether ⟨D_max⟩ still rises sharply near 1-p≈0.2 and diverges for 1-p>0.7. If the thresholds shift or vanish, the claimed transition is not generic.","tokens_in":16169,"feed_emoji":"🕸️","tokens_out":5056,"duration_ms":43202,"temperature":0.7,"pith_summary":"This paper aims to establish a general, quantitative way to compare two generalized Lotka-Volterra (gLV) dynamical processes that start from the same initial conditions but differ in interaction strengths, network structure, or even the mathematical form of their dynamics. The proposed dissimilarity measure, D(t), tracks the normalized sum of absolute abundance differences over time, with a maximum D_max and a long-time limit D_inf that together capture both transient and stationary behavior. Using this measure, the authors show that structural differences invisible at the adjacency-matrix level can still produce markedly divergent dynamics in small directed networks, and that in modular networks the fraction of negative interactions acts as a control parameter for a sharp transition from stable to unstable behavior. A sympathetic reader would care because this offers a practical tool for robustness analysis and instability prediction in ecology, microbiome science, and other fields that model complex systems with nonlinear interactions.","feed_headline":"A single metric pinpoints when interaction networks turn unstable","feed_subtitle":"Comparing two nearly identical Lotka-Volterra systems reveals hidden divergence before collapse.","key_machinery":"The central object is the dissimilarity D(t) = (1/S0) Σ_i |x_i(t) − x*_i(t)|, comparing a reference process x_i(t) and a modified process x*_i(t) that share initial conditions and total initial abundance S0, evolving under gLV dynamics with possibly different interaction matrices Λ vs Λ*, different network topology, or different governing equations. Its maximum over time, D_max, condenses the comparison into a single scalar, and its long-time limit D_inf indicates whether the two processes converge to the same steady state. The sign-flip protocol — taking a reference interaction matrix Λ and constructing Λ* with the same absolute weights but each sign kept positive with probability p — is th","core_discovery":"The paper claims that a scalar, time-dependent dissimilarity D(t) — defined as the normalized sum of absolute abundance differences between two gLV processes that share initial conditions — can serve as a universal comparator for generalized Lotka-Volterra systems. Defined for two-species predator–prey models (Eq. 4) and for arbitrary N-species networks (Eq. 13), with a maximum D_max (Eq. 14) and long-time limit D_inf, the measure captures both transient divergence and stationary differences. The authors demonstrate that in directed three-node graphs, D(t) distinguishes dynamics that are indistinguishable at the adjacency-matrix level; in a 12-node modular network, increasing the fraction of","pith_inferences":["Because the dissimilarity is normalized by initial abundances and uses identical initial conditions, it isolates the effect of the interaction structure; one could invert the perspective and use it to quantify sensitivity to initial-condition uncertainty by sampling initial states instead of perturbing Λ.","The sharp threshold near 1-p≈0.2 and divergence near 1-p>0.7 in a modular network may be a finite-size signature of a broader phase transition; testing other N, other modular connectivities, and other weight distributions could reveal whether these values scale or shift.","The localized-vs-global contrast suggests a design principle for interventions in microbial communities: treatments aimed at one community may be safer than global network changes, assuming the buffer is not an artifact of the specific clique-ring topology.","Extending D(t) to stochastic gLV dynamics (with demographic or environmental noise) could produce a robustness index that flags systems on the verge of instability earlier than deterministic trajectories."],"forward_implications":["Dynamical comparison supersedes purely structural network comparison: two graphs with identical link counts and identical sum of matrix differences can still exhibit very different dissimilarity, as shown for three-node graphs.","The fraction of negative interactions in a modular network acts as a control parameter: below ~20% sign-flip probability the reference and modified dynamics stay close, above it they separate sharply, and for more than ~70% the modified process diverges.","Localized structural changes confined to a single community are buffered: even complete sign inversion inside one clique does not destabilize the whole 12-node system, whereas the same fraction of global sign flips does.","The measure works across functional forms: comparing a reference gLV equation to a modified one with interaction term x_j^γ detects an asymmetry around γ=1, with larger dissimilarity for γ>1 than γ<1.","The framework offers a practical tool for microbiome and ecology studies, where inferred interaction matrices are uncertain and alternative functional forms are plausible."],"fun_headline_variants":["Dissimilarity metric flags hidden instability in networks","One scalar predicts divergent dynamics in Lotka-Volterra","Measure captures transient and steady-state divergence","Network structure secrets revealed by new dissimilarity","Compare any two species communities with one metric"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim that a threshold fraction of negative interactions (above ~20%) drives instability and divergence (above ~70%) is based on simulations of a single 12-node clique-ring network with uniformly drawn weights; if these thresholds depend on network size, topology, or weight distribution, the broad conclusion that the fraction and distribution of negative interactions control the stability transition would be unsupported.","fun_headline_variants_meta":{"raw":{"variants":["Dissimilarity metric flags hidden instability in networks","One scalar predicts divergent dynamics in Lotka-Volterra","Measure captures transient and steady-state divergence","Network structure secrets revealed by new dissimilarity","Compare any two species communities with one metric"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000172,"raw_usage":{"total_tokens":1102,"prompt_tokens":722,"completion_tokens":380,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":466,"completion_tokens_details":{"reasoning_tokens":311}},"tokens_in":466,"tokens_out":380,"duration_ms":4692,"temperature":1.0,"reasoning_tokens":311,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T21:57:53.783820+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the sign-flip protocol on a different network geometry (e.g., two cliques, a scale-free modular graph, or a larger number of cliques) or with weights drawn from an exponential or Gaussian distribution, and check whether ⟨D_max⟩ still rises sharply near 1-p≈0.2 and diverges for 1-p>0.7. If the thresholds shift or vanish, the claimed transition is not generic.","supporting_citations":[],"review_version":1}