{"id":"38154992-89a1-44fe-861c-6966ca4b9b52","arxiv_id":"1908.04731","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new network entropy measure, the average binary entropy of normalized node degrees, is shown to characterize chaotic attractors and quantify structural information loss through link density changes.","lead":"This paper proposes an entropy measure for recurrence networks built from chaotic time series, defined as the average uncertainty of each node's connections. The authors show it stabilizes with data length and embedding dimension, distinguishes chaotic attractors from white noise, and links changes in the measure to structural information loss on the attractor.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's ΔEm and ρh values are inconsistent with the paper's own defining relation Eh_m(ρh)=Ec_m, undermining the central information-loss quantification.","rationale":"The reader's weakest_assumption concerns the generality of the recurrence-threshold scheme (Ref. [37]). That is a relevant external validity concern, but it is not the most load-bearing issue: the paper does provide some robustness evidence (Fig. 3) and the threshold scheme is the authors' prior contribution, so a failure is speculative. By contrast, Table 1 is the paper's own quantitative evidence for the central claim, and it violates the paper's defining equations. This is an internal inconsistency, not a matter of consensus or external assumptions. If the table cannot be reproduced under the paper's own definitions, then the claimed relationship between ΔEm and Δρ is not established, and the central claim lacks its main support. The correct response is to require the authors to recompute and correct the table (and ideally provide code/data), which is consistent with a CONDITIONAL verdict. The reader's verdict already calls for corrections, so I leave the verdict unchanged, though I flag a specific error that must be fixed. I disagree with the reader's identification of the threshold as the weakest assumption because the table error is more concrete and directly falsifiable from the paper's text.","tokens_in":11333,"tokens_out":12978,"duration_ms":118387,"concrete_test":"Recompute Table 1 from the authors' definitions: for each system, compute Ec_m directly from the RN degree distribution via Eq. (2), compute Eh_m(ρc) from Eq. (7), then solve Eh_m(ρh) = Ec_m for ρh. Check whether the reported ΔEm and Δρ satisfy |Eh_m(ρc) − Ec_m| = ΔEm and |ρc − ρh| = Δρ to within rounding error. If the identity fails, the table is incorrect and the claimed information-loss quantification is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 4, the paper defines ΔEm = |Eh_m(ρc) − Ec_m| and ρh as the link density satisfying Eh_m(ρh) = Ec_m, where Eh_m(ρ) = −[ρ log ρ + (1−ρ) log(1−ρ)] (Eq. 7). Table 1 reports ΔEm, ρc, ρh, and Δρ for four systems. These numbers do not satisfy the defining relation. For Lorenz, ρc = 0.0184 gives Eh_m = 0.0918 nats; subtracting reported ΔEm = 0.0034 leaves Ec_m ≈ 0.0884. Solving Eh_m(ρh) = 0.0884 yields ρh ≈ 0.0175, but the table reports ρh = 0.0162, for which Eh_m = 0.0829. Similar discrepancies occur for Rössler (reported ρh = 0.0194 vs. computed ≈ 0.0205) and Hénon (0.0421 vs. ≈ 0.0436); only the white-noise row is approximately consistent. Thus the central quantitative demonstration linking entropy difference to link-density difference is internally inconsistent. Since the paper's headline claim is that Em can uniquely characterize structural information loss in terms of link-density changes, and Table 1 is the primary numerical evidence for that linkage, this inconsistency is load-bearing.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces an entropy measure Em for unweighted undirected networks, defined through per-node probabilities p_i=k_i/(N-1) in Eq. (2), with an extra term accounting for absent links. For homogeneous networks the measure reduces to the closed form Eh_m(ρ) in Eq. (7), a one-hump function of link density with maximum log 2 at ρ=1/2. Recurrence networks are constructed from chaotic time series using the threshold scheme of the authors' prior work, and numerical experiments on Lorenz, Rössler, Hénon, and white noise show saturation of Em with the number of nodes N and embedding dimension M. The paper proposes that the difference ΔEm=|Eh_m(ρc)−Ec_m|, together with the associated link-density difference Δρ=|ρc−ρh|, quantifies structural information loss on chaotic attractors, and applies the approach to a Kepler RR Lyrae light curve, claiming that the measure rules out pure stochastic variability.","tokens_in":11609,"tokens_out":11185,"duration_ms":107340,"significance":"If the central claims hold, the measure would be a simple, directly computable scalar that is stable with respect to N and M and could quantify noise-induced structural change on attractors in terms of link density. The homogeneous-network formula is exact, and the saturation plots for standard attractors are a useful check. The potential applicability to short real-world time series is also attractive. However, the quantitative connection between entropy difference and link-density difference currently rests on Table 1, which is internally inconsistent, and several statements (uniqueness, the heterogeneous-network minimum, and the claim of ruling out stochastic processes) exceed what the presented evidence supports. With those points corrected, the contribution would be a modest but useful addition to recurrence-network analysis.","major_comments":[{"comment":"The numbers in Table 1 do not satisfy the defining relation Eh_m(ρh)=Ec_m stated in the text. For Lorenz, with ρc=0.0184 and ΔEm=0.0034, Eq. (7) gives Eh_m(ρc)=0.0917 nats, so Ec_m=0.0883; solving Eh_m(ρ)=0.0883 gives ρ≈0.0175, not the reported ρh=0.0162. The same discrepancy appears for Rössler (reported 0.0194 vs computed about 0.0204) and Hénon (0.0421 vs about 0.0434); only the white-noise row is roughly consistent. Because the paper's headline claim is that Δρ measures information loss through the relation between Eh_m and Ec_m, this inconsistency is load-bearing. Please recompute Table 1 or clarify the definitions of ρh and Δρ.","section":"Section 4, Table 1"},{"comment":"The assertion that 'the entropy measure proposed here is minimum for the completely heterogeneous network' with degree sequence {1,2,...,N−1} is not proved and is not correct as stated. For N nodes there are only N−1 positive degree values available, so no network with no isolated nodes can realize a distinct degree for every node; moreover, for a fixed link density ρ, the minimum of the per-node binary entropy sum occurs at extremal distributions (as many p_i as possible at 0 or 1), not at the uniformly spread sequence {1,...,N−1}, which has mean ρ=1/2. The maximum property Eh_m≥Em used in Section 4 follows from concavity and is sufficient; the minimum claim should be removed or replaced by a correct statement.","section":"Section 2, paragraph after Eq. (7)"},{"comment":"The generality of the threshold scheme is assumed rather than demonstrated. Fig. 3 tests robustness to ε only for Lorenz and white noise, yet the paper concludes that Em is a characteristic property of 'every chaotic attractor' and applies the measure unconditionally to real data. In particular, the statement in Section 5 that the average values ΔEm=0.0066 and Δρ=0.0031 for the light curve 'rules out pure stochastic process' is not justified without a null model or surrogate test; white noise is only one member of the stochastic class, and colored or measurement noise could give similar values. Please add a surrogate analysis or soften the claim.","section":"Sections 3 and 5"}],"minor_comments":[{"comment":"The text says 'as Δt becomes large, the attractor gets slightly over sampled'; a larger sampling interval corresponds to undersampling, not oversampling. Please correct the wording.","section":"Section 3, Fig. 4"},{"comment":"Eq. (3) contains a redundant expression and can be simplified to the form in Eq. (7); please use the consistent form −[ρ log ρ + (1−ρ) log(1−ρ)] throughout to avoid sign-convention misreading.","section":"Eq. (3)"},{"comment":"The values ΔEm=0.0066 and Δρ=0.0031 for the light-curve segments are reported without segment-to-segment variability; given that the data are divided into several segments, please report mean ± standard deviation or a range.","section":"Section 5"},{"comment":"The conclusion states that 'the finiteness of the data will not affect the accuracy of the analysis since the measure is independent of N', but Fig. 1 shows saturation rather than exact independence for all N; please rephrase to 'approximately independent after saturation'.","section":"Section 6"},{"comment":"Minor typographical issues include 'un-weighted' and 'un-directed' in the Abstract and an encoding artifact in 'R¨ossler' in the text.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The Table 1 inconsistency is the main obstacle; if the numbers cannot be reproduced, the central quantitative claim fails. The authors should be asked to provide the underlying data or a corrected table. The paper is also somewhat borderline for physics.soc-ph, being primarily a nonlinear time series analysis contribution, but this is not a decisive issue if the technical content is fixed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the proposed measure is simple and genuinely new in this specific form: adding the absence-of-link term to degree-based Shannon entropy gives a per-node binary entropy. The homogeneous-network closed form in Eq. (7) is correct, and the hump curve with maximum log 2 at rho = 1/2 is correct. The saturation with N and M and the clear separation from white noise are reasonable numerical findings. The idea of using |E_h^m(rho_c) - E_c^m| as a measure of structural information loss is worth taking seriously.\n\nSecond, the stress-test note is right, and it is worse than a rounding issue. In Section 4, rho_h is defined as the link density of a homogeneous network with the same entropy as the recurrence network, i.e. E_h^m(rho_h) = E_c^m. Table 1 reports values that do not satisfy this. For Lorenz, rho_c = 0.0184 and Delta_E = 0.0034 give E_c^m approximately 0.0884, which implies rho_h approximately 0.0175, not the reported 0.0162. Rössler and Hénon show similar discrepancies; only white noise is close. Since the paper's headline claim is that the measure characterizes information loss in terms of link-density changes, and Table 1 is the primary numerical evidence, this inconsistency is load-bearing.\n\nThe assertion that E_m is minimum for the completely heterogeneous network is also asserted without proof and is questionable. The real-data claim that the light curve 'rules out pure stochastic process' is made without surrogates or statistical testing. The reliance on the authors' own threshold scheme is a domain assumption rather than an internal flaw, but the stability tests only cover Lorenz and white noise, so the generality is not demonstrated. No code or data are provided, which makes reproduction harder.\n\nWhat holds up: the core definition is clear, the homogeneous-network calculation is correct, and the measure does appear to saturate and to distinguish chaotic attractors from white noise in the tests shown. The soft spots are concentrated in the quantitative link between entropy difference and link-density difference, plus the extremal claim and the real-data inference.\n\nI would send this to peer review rather than desk reject it, because the underlying idea is coherent and the Table 1 problem is fixable. But the revision needs a corrected Table 1, a proof or careful numerical test of the extremal claim, and statistical support for the real-data conclusion. The right reader is someone working on recurrence-network measures who can tolerate a modest incremental contribution. Do not cite it until Table 1 is fixed.","headline":"The proposed entropy measure for recurrence networks is plausible and partly useful, but Table 1's numbers do not satisfy the paper's own defining relation, so the central quantitative claim currently does not hold.","tokens_in":12122,"tokens_out":3329,"would_cite":false,"duration_ms":30776,"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 shows that a new recurrence-network entropy measure saturates as a chaotic-attractor fingerprint and quantifies noise-driven information loss through link-density gaps.","keywords":["recurrence networks","network entropy","chaotic attractors","information loss","link density","white noise","time series analysis","nonlinear dynamics"],"falsifier":"Compute $E_m$ for recurrence networks built from a chaotic system with strongly non-uniform phase-space density, such as a hyperchaotic or intermittent system, using the same $\\epsilon$ values ($0.06, 0.10, 0.14, 0.18$ for $M=2$ to $5$): if $E_m$ fails to saturate with $N$ or $M$, or drifts toward the white-noise value, the claim that $E_m$ is a characteristic attractor property is falsified. A second check is to rebuild the same attractors with a different accepted threshold rule and see whether the ordering of $\\Delta\\rho$ across attractors is preserved.","tokens_in":11109,"feed_emoji":"🌐","tokens_out":7224,"duration_ms":65908,"temperature":0.7,"pith_summary":"The paper proposes an entropy measure $E_m$ for recurrence networks built from chaotic time series and aims to show that it is a characteristic property of the underlying attractor. Concretely, $E_m$ saturates with the number of data points $N$ and the embedding dimension $M$, and it clearly separates the recurrence networks of chaotic systems from those of white noise. The key proposal is that structural information loss on an attractor, such as that produced by noise contamination, can be quantified by how far the network's entropy lies below the maximum entropy of a homogeneous network with the same link density. A sympathetic reader would care because this offers a finite-data-friendly way to detect and monitor structural degradation without relying on dimension estimates.","feed_headline":"Entropy of recurrence networks pins down chaos lost to noise","feed_subtitle":"Saturating attractor entropy makes noise damage measurable from a few thousand data points.","key_machinery":"The load-bearing object is the network entropy $$E_m = -\\frac{1}{N}\\sum_{i=1}^N \\left[p_i\\log p_i + (1-p_i)\\log(1-p_i)\\right],\\qquad p_i = \\frac{k_i}{N-1},$$ with $k_i$ the degree of node $i$. The second term, counting the information carried by absent links, is the paper's addition to the usual Shannon entropy of the degree distribution. For a homogeneous network the formula collapses to $E_m^h(\\rho) = -[\\rho\\log\\rho - \\rho\\log(1-\\rho)+\\log(1-\\rho)]$, where $\\rho = 2L/N(N-1)$ is the link density; this one-hump function peaks at $\\rho=1/2$ with value $\\log 2$. The recurrence network itself is constructed by time-delay embedding of the time series and placing an edge between two nodes when their embedded points lie within a threshold $\\epsilon$, with $\\epsilon$ fixed by the authors' earlier scheme. The machinery works by comparing the measured $E_m$ of a recurrence network with the homogeneous value at the same link density: the deficit $\\Delta E_m$, or equivalently the link-density shift $\\Delta\\rho$, is interpreted as structural information loss.","core_discovery":"The paper's central claim is that the recurrence-network entropy $E_m$ defined in Eq. (2) is a characteristic property of a chaotic attractor: under the authors' recurrence-threshold scheme, $E_m$ converges to a constant as the number of nodes $N$ and embedding dimension $M$ increase, and it separates chaotic attractors from white noise. For an attractor's recurrence network with link density $\\rho_c$ and entropy $E_m^c$, the gap $\\Delta E_m = |E_m^h(\\rho_c) - E_m^c|$ to the homogeneous network of the same link density, equivalently expressed as the link-density difference $\\Delta \\rho = |\\rho_c - \\rho_h|$, is taken as the information loss caused by structural change. The paper demonstrates the measure on noise-contaminated attractors and on a variable-star light curve, where the computed $\\Delta E_m$ and $\\Delta \\rho$ indicate nonlinear, non-stochastic variability.","pith_inferences":["Beyond the paper, the same entropy gap could serve as a general structural-complexity index for any unweighted undirected network, not only recurrence networks, letting practitioners compare networks of different sizes on a common scale.","The near-universal curve for $R(\\rho)$ suggests a possible noise-amplitude estimator, but only if colored noise and dynamical noise are tested; the paper itself tests only additive white noise.","If the threshold scheme is replaced by a fixed link-density rule, the ordering of attractors by $\\Delta \\rho$ may change; this is a testable extension that would separate the measure's content from the threshold convention."],"forward_implications":["For fixed embedding dimensions $M=2$ to $5$, $E_m$ can be treated as a stable fingerprint of a chaotic attractor, independent of time-series length once $N$ exceeds a few thousand points.","Noise-induced structural loss can be tracked by the ratio $R(\\rho)=(\\rho_c-\\rho)/(\\rho_c-\\rho_w)$, which rises from $0$ for the clean attractor to $1$ for white noise and behaves approximately identically for different chaotic systems.","The gap measures $\\Delta E_m$ and $\\Delta \\rho$ provide a discriminating statistic for real data: the variable-star light curve KIC 4484128 yields values close to those of chaotic systems, arguing against a purely stochastic origin.","Because the measure is normalized by $N$, it avoids the finite-data problems of correlation-dimension estimates for short experimental records."],"supporting_citations":[{"why":"Supplies the automated recurrence-threshold scheme (epsilon per embedding dimension M) used to construct every recurrence network in the paper.","marker":"[37]"},{"why":"Documents how additive noise depletes the range of node degrees in recurrence networks, the mechanism behind the proposed noise analysis.","marker":"[38]"},{"why":"Gives the network-ensemble structural entropy concept the paper adapts into its homogeneous-network maximum entropy $E_m^h$.","marker":"[22]"},{"why":"Provides the degree-heterogeneity measure supporting the claim that the completely heterogeneous network has minimal $E_m$.","marker":"[25]"},{"why":"Reviews alternative recurrence-threshold choices and motivates the paper's stability check of $E_m$ with respect to $\\epsilon$.","marker":"[23]"}],"fun_headline_variants":["Recurrence network entropy measures chaos lost to noise","Attractor entropy gap quantifies chaos degradation","Noise-damaged chaos measured via network entropy","Link density gap reveals chaos information loss","Recurrence nets quantify chaos information loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the fixed recurrence threshold $\\epsilon$ chosen per embedding dimension by the authors' earlier scheme faithfully captures attractor structure for every chaotic system and for white noise; the paper tests stability around those values mainly for the Lorenz system and white noise, so if the threshold rule fails elsewhere the saturation and information-loss claims could be artifacts of that choice.","fun_headline_variants_meta":{"raw":{"variants":["Recurrence network entropy measures chaos lost to noise","Attractor entropy gap quantifies chaos degradation","Noise-damaged chaos measured via network entropy","Link density gap reveals chaos information loss","Recurrence nets quantify chaos information loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000476,"raw_usage":{"total_tokens":2329,"prompt_tokens":885,"completion_tokens":1444,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":1375}},"tokens_in":501,"tokens_out":1444,"duration_ms":12330,"temperature":1.0,"reasoning_tokens":1375,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:28:13.470980+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute $E_m$ for recurrence networks built from a chaotic system with strongly non-uniform phase-space density, such as a hyperchaotic or intermittent system, using the same $\\epsilon$ values ($0.06, 0.10, 0.14, 0.18$ for $M=2$ to $5$): if $E_m$ fails to saturate with $N$ or $M$, or drifts toward the white-noise value, the claim that $E_m$ is a characteristic attractor property is falsified. A second check is to rebuild the same attractors with a different accepted threshold rule and see whether the ordering of $\\Delta\\rho$ across attractors is preserved.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the automated recurrence-threshold scheme (epsilon per embedding dimension M) used to construct every recurrence network in the paper."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents how additive noise depletes the range of node degrees in recurrence networks, the mechanism behind the proposed noise analysis."},{"cited_title":"Anand and G","cited_arxiv_id":null,"evidence_quote":"Gives the network-ensemble structural entropy concept the paper adapts into its homogeneous-network maximum entropy $E_m^h$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the degree-heterogeneity measure supporting the claim that the completely heterogeneous network has minimal $E_m$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Reviews alternative recurrence-threshold choices and motivates the paper's stability check of $E_m$ with respect to $\\epsilon$."}],"review_version":1}