{"id":"0373c635-5eb1-4bb0-ad67-93f7e2bf3fe2","arxiv_id":"2412.14864","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"For unlabelled temporal networks, invariant-based pseudo-distances recover periodicity and memory and qualitatively detect chaotic instability, but they cannot yield model-independent Lyapunov exponents.","lead":"This paper shows how to measure a changing network's periodicity, memory, and chaotic instability even when the identities of the nodes are unknown or hidden. Using label-independent measurements built from standard graph features, it finds that these dynamical fingerprints can be partially recovered, and that a universal chaos measure for unlabelled networks is out of reach.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim leans on an unproven monotone bridge between pseudo-distances and labelled distance; if this bridge has inversions or plateaus, the chaos-expansion fingerprints in Sec. 3.1 are not diagnostic.","rationale":"The strongest claim is explicitly partial, and the paper's own disclosure that invariants do not capture full graph information makes the equivalence-class limitation an acknowledged scope condition rather than a hidden defect. A trajectory that stays inside an equivalence class is a genuine counterexample, but the claim never promises universal recovery; the synthetic demonstrations already require the chosen invariants to vary. The more insidious assumption is the numerical monotonicity of d_X as a function of d_lab, which is used as the bridge from invariant dissimilarity to dynamical instability. It is only tested on two perturbation families, with fixed N and similar edge density, and the paper admits the mapping is not universal. The chaos experiments in Figs. 2-3 interpret an expansion phase as evidence of sensitivity to initial conditions; if f_X is non-monotonic or has plateaus outside the tested families, that interpretation can fail. I would not reject the paper: the synthetic controls (shuffled and permuted trajectories) and the consistent results across three invariants and several models support the qualitative proof-of-concept. But the missing monotonicity analysis is exactly the kind of gap that a revision should close, and it is the load-bearing premise behind the most ambitious part of the conclusion. Secondary issues (Eq. 5 vs Eq. 15 factor of two; the hand-chosen z=4 threshold; absence of code and data links) do not change this assessment, though they reinforce the conditional status.","tokens_in":20426,"tokens_out":6650,"duration_ms":48337,"concrete_test":"For N=20 and p=0.2, generate many pairs of labelled graphs (G,H) separated by t random edge flips, t=1..50, and compute d_lab, d_deg, d_spec, and d_eig for each pair. Test monotonicity by counting the fraction of realizations where d_X decreases while d_lab increases, or where d_X is flat while d_lab changes substantially. If any inversion or plateau occurs with non-negligible frequency, construct a labelled network trajectory that follows the violating sequence and run the Wolf recurrence protocol of Sec. 3.1.1; if d_X(t) shows expansion while d_lab(t) does not, the chaotic-fingerprint claim fails. Repeat the exercise with degree-preserving rewires to specifically probe d_deg and d_spec.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's conclusion that unlabelled trajectories reveal 'sensitive dependence on initial conditions' depends on treating d_X as a faithful proxy for the labelled ground-truth distance d_lab. Section 2.3 only establishes this by numerical simulation on two perturbation families (ER latent-noise and edge displacement), and the text itself concedes that 'we cannot establish a simple and universal quantitative mapping d_X = f_X(d_lab)' and that the relation is model-dependent. That is not just a quantitative caveat: the inference in Sec. 3.1 that an expansion phase in d_X(t) indicates chaotic instability requires d_X to be a non-decreasing function of d_lab along the sampled trajectory. If d_X has inversions or flat plateaus outside the two tested families, then d_X can grow while d_lab shrinks, or stay constant while d_lab grows, so an expansion phase is neither necessary nor sufficient for sensitivity to initial conditions. The same bridge underlies the autocorrelation-like measures in Sec. 3.2, since c_X normalizes by the average invariant distance J; if d_X saturates, c_X can acquire spurious decay or periodicity. The factor-of-two discrepancy between Eq. (5) and the appendix derivation Eq. (15) is a separate correctness warning, but the monotonicity bridge is the more load-bearing gap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for characterizing the dynamics of temporal networks when node labels are unavailable. It defines three permutation-invariant pseudo-distances based on the degree sequence, the adjacency spectrum, and the eigenvector centrality, and uses them to construct autocorrelation-like functions and expansion measures for unlabelled network trajectories. The approach is validated on synthetic models of noisy periodic dynamics, DARN processes with memory, and low- and high-dimensional chaotic graph dynamics, as well as on three empirical temporal networks with removed labels. The central conclusion is that periodicity, memory, and, qualitatively, sensitivity to initial conditions can be partially recovered from unlabelled trajectories, while a model-independent Lyapunov exponent cannot be defined.","tokens_in":20581,"tokens_out":14316,"duration_ms":93090,"significance":"The problem addressed is timely and relevant: unlabelled temporal networks arise in privacy-constrained or technically difficult tracking settings, and few tools exist for their dynamical analysis. The paper's strengths are its clear formulation of invariant-based pseudo-distances, the use of labelled ground-truth distances as benchmarks, the explicit comparison against shuffled and permuted baselines, and a candid discussion of limitations, including the non-universal and sublinear relation between pseudo-distances and the labelled distance. If the results hold, the paper provides a useful proof of concept for extending network time-series analysis to unlabelled trajectories. However, the analytical inconsistency in the ground-truth derivation and the unproven monotonicity bridge between pseudo-distances and labelled distance prevent the claims from being fully established as stated.","major_comments":[{"comment":"Equation (5) is inconsistent with the appendix derivation in Eq. (15). Eq. (15) contains a factor 1/2 multiplying the two exponential terms, whereas Eq. (5) does not; for example, with p=0.5 and sigma=1, Eq. (5) predicts a negative value of d_lab, which is impossible. In addition, the appendix states that Z in Eq. (4) is chosen as N(N-1)/2, while the main text sets Z=N(N-1); with the appendix's normalization, the right-hand side of Eq. (14) would need a factor of 2. Since d_lab is the ground-truth distance used to calibrate the pseudo-distances in Fig. 1, the theoretical curve in Fig. 1(a) and the horizontal axis in Fig. 1(b) should be recomputed with the corrected formula before the results in Sec. 2.3 are taken as quantitative.","section":"Sec. 2.3.1 / Appendix"},{"comment":"The interpretation of the expansion phases in Figs. 2 and 3 as evidence of sensitive dependence on initial conditions rests on the assumption that the pseudo-distance d_X is a non-decreasing function of the labelled distance d_lab along the sampled trajectory. The paper only provides numerical evidence for monotonicity of ensemble averages on two perturbation families (ER latent noise and edge displacement) and explicitly concedes that no universal quantitative mapping d_X = f_X(d_lab) exists. This is not sufficient for the inference made: pseudo-distances can vanish for non-isomorphic graphs (e.g., d_deg=0 for any two graphs with the same degree sequence), so inversions or plateaus are possible outside the tested families. The authors should either prove monotonicity for a relevant class of latent graph dynamics, or explicitly restrict the chaotic-fingerprint claims to the tested models and state that the expansion phase is not necessarily diagnostic in general.","section":"Sec. 2.3 and Sec. 3.1"}],"minor_comments":[{"comment":"The reference 'Appendix 4' should be updated to the actual appendix label (e.g., 'Appendix A').","section":"Sec. 2.3.1"},{"comment":"The word 'Autorregressive' in the section title should be spelled 'Autoregressive'.","section":"Sec. 3.2.3"},{"comment":"The misspelling 'egdes' should be corrected to 'edges'.","section":"Appendix"},{"comment":"The phrase 'ground true' is used several times; it should be 'ground-truth'.","section":"Throughout"},{"comment":"The description of the 'permuted' curve in panel (a) is ambiguous; please clarify whether the node-label permutation is applied independently to each snapshot, and why the corresponding curves are not shown in panels (b)-(d).","section":"Figure 2 caption"},{"comment":"The definition of J includes the t=t' term where d_X=0; state this explicitly, as it affects the interpretation of c_X(tau) for small tau.","section":"Eq. (11)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the scope of the journal. The analytical normalization error in the d_lab derivation and the unproven monotonicity bridge are both fixable but need to be addressed before publication. I see no concerns about attribution or overlap with the authors' prior work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid proof-of-concept: unlabelled temporal network trajectories, compared via permutation-invariant pseudo-distances, retain recoverable dynamical fingerprints like periodicity, memory decay, and an expansion phase. The extension of network autocorrelation and instability quantifiers to the unlabelled setting is genuinely new, and the synthetic tests are clean: noisy periodic trajectories show peaks at the right period, DARN decays reproduce the memory slope, and both low- and high-dimensional chaotic dynamics show an expansion phase that is flat under shuffling or label permutation. The empirical validation on flights, Malawi village, and SFHH is a nice touch, and the authors are candid about the crucial limitation—no model-independent Lyapunov exponent exists, and pseudo-distance growth is sub-linear and non-universal.\n\nSoft spots, in proportion. First, Eq. (5) in the main text disagrees with the appendix derivation Eq. (15) by a factor of two on the exponential correction terms. My own calculation of the integrals in Eq. (14) matches the appendix, so the main-text formula appears to be a typo rather than a conceptual error, but it must be fixed because Fig. 1(a) claims theory–simulation agreement.\n\nSecond, the monotonicity bridge. The paper establishes by simulation on two perturbation families that d_X is continuous and non-decreasing in d_lab, and it explicitly disclaims a universal mapping. That is honest, but the chaos detection in Sec. 3.1 rests on interpreting an expansion of d_X as sensitive dependence on initial conditions. If d_X can plateau or invert outside those families, the expansion is not diagnostic on its own. The shuffled and permuted controls mitigate this, but a broader numerical check or a more guarded interpretation would close the gap.\n\nThird, the empirical validation likely confounds topology with total edge count. The daily periodicity in US flights shows up in number-based invariants very easily; a control preserving the edge-count sequence while randomizing topology would strengthen the claim that structure, not just busyness, is being captured. This is moderate, not fatal; the Malawi and SFHH results show slow decay that is not obviously edge-count-driven.\n\nMinor: the z=4 detection threshold is hand-chosen (the authors say so), and the data/code availability line is a placeholder. Both are acceptable for a proof-of-concept but should be addressed before publication.\n\nOverall, the central qualitative claim holds up. This is a useful methods paper for anyone working with temporal networks where node identities are hidden or untracked—animal tracking, privacy-restricted contact data, or molecular systems. It deserves a serious referee, and with a few revisions it will be a solid contribution. I would bring it to a reading group and would cite it if I published in this area. Recommendation: send to peer review; the issues are addressable and the core message is sound.","headline":"A solid proof-of-concept that unlabelled temporal networks retain fingerprints of periodicity, memory, and instability; the main fixes are an analytic typo, a missing topology-only control, and a slightly overreaching monotonicity claim.","tokens_in":21194,"tokens_out":6534,"would_cite":true,"duration_ms":47774,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["89.75.Fb"],"model":"deepseek-v4-flash","headline":"This paper claims that the dynamical fingerprints of a temporal network—periodicity, memory decay, and sensitive dependence on initial conditions—can be partially recovered from a sequence of unlabelled snapshots, using…","keywords":["temporal networks","unlabelled graphs","graph invariants","permutation-invariant pseudo-distances","network autocorrelation","Lyapunov exponents","network trajectories","network dynamics"],"falsifier":"Construct a labelled temporal trajectory whose snapshots are pairwise non-isomorphic but all share the same sorted degree sequence, the same adjacency spectrum, and the same sorted eigenvector centrality; if every pseudo-distance stays identically zero and $c_X(\\tau)$ shows no structure, then the claimed recovery fails for that dynamical class. Regular graphs with rearranged edges are a natural place to search for such a counterexample.","tokens_in":20119,"feed_emoji":"🕸️","tokens_out":7082,"duration_ms":49687,"temperature":0.7,"pith_summary":"Temporal networks are usually studied as sequences of adjacency matrices that require nodes to be tracked over time, but there are many settings—animal swarms, privacy-restricted contact data, molecular graphs—where labels are missing or cannot be matched. This paper asks whether the dynamics of the underlying network can still be characterized after the labels have been removed. The answer it argues for is a qualified yes: quantitative fingerprints such as the period of noisy periodic trajectories and the memory decay of autoregressive network processes remain detectable from unlabelled snapshots, and chaotic expansion of nearby network trajectories stays visible qualitatively, although not with a well-defined Lyapunov exponent. The authors make the case by defining three permutation-invariant pseudo-distances and showing, on synthetic models and three empirical datasets, that autocorrelation-like functions built from them recover the same qualitative behaviour as the labelled ground truth.","feed_headline":"Network dynamics stay readable after node labels are stripped","feed_subtitle":"Degree lists, spectra and centralities reveal periodicity, memory decay and chaos in temporal networks where node identity is lost.","key_machinery":"The machinery is the permutation-invariant pseudo-distance: a normalized dissimilarity between two unlabelled graphs computed from a graph invariant, so that it does not depend on which labelling of the nodes is chosen. The paper uses three: the $L^1$ distance between sorted degree sequences ($d_{\\mathrm{deg}}$), the $L^1$ distance between ordered adjacency spectra ($d_{\\mathrm{spec}}$), and the $L^2$ distance between sorted, normalized eigenvector centrality vectors ($d_{\\mathrm{eig}}$). These feed an autocorrelation-like function $c_X(\\tau)=1-\\frac{1}{T-\\tau}\\sum_{t=1}^{T-\\tau} d_X(G_t,G_{t+\\tau})/J$, where $J$ is the mean pairwise pseudo-distance over snapshots, and a recurrence-based expansion analysis for chaos. The load-bearing numerical fact is that, in all tests, each $d_X$ is a continuous, monotonically increasing, nonlinear function of the labelled ground-truth distance $d_{\\mathrm{lab}}$; that monotonicity is what makes period, memory, and expansion visible in the unlabelled signal.","core_discovery":"The central discovery is that removing node labels does not erase all dynamical information: graph properties that are invariant under relabelling can stand in for the adjacency matrix. More precisely, the paper defines pseudo-distances $d_{\\mathrm{deg}}$, $d_{\\mathrm{spec}}$, and $d_{\\mathrm{eig}}$ from the sorted degree sequence, the ordered spectrum of the adjacency matrix, and the sorted eigenvector centrality vector, and uses them to build unlabelled autocorrelation functions $c_X(\\tau)$. These recover the period of noisy periodic network trajectories, reproduce the flat-then-decaying autocorrelation shape of $DARN(\\rho)$ processes, and reveal an expansion phase for initially close trajectories in both low- and high-dimensional chaotic network dynamics. The paper states clearly that the expansion measured by the pseudo-distances is not strictly exponential and that, because the relation between pseudo-distance and labelled distance is monotonic but model-dependent, a model-independent Lyapunov exponent for unlabelled networks cannot currently be defined.","pith_inferences":["Combining several graph invariants into a single pseudo-distance should shrink the set of graphs that look identical to the measure, likely improving period and memory detection; this is a direct testable extension of the paper's single-invariant comparison.","The non-exponential shape of the chaotic expansion observed in the pseudo-distances could itself be a usable signature of instability, for instance through the time-dependent local slope of $\\log d_X(t)$, even without a universal Lyapunov exponent.","If the invariance-blindness limitation can be controlled in practice, these methods open a privacy-friendly route: behavioural rhythms and correlation scales could be measured from anonymized contact or proximity data without tracking individual identities.","A model-specific calibration of the monotonic function relating each pseudo-distance to the labelled distance could, in principle, turn the qualitative chaos detection into a quantitative exponent estimate for known generative models, although the paper does not pursue this."],"forward_implications":["The period of a noisy periodic network trajectory can be read off as a peak in $c_X(\\tau)$, though it remains detectable to higher noise levels in the labelled case than in the unlabelled one.","The memory order $\\rho$ of a $DARN(\\rho)$ process is visible in unlabelled trajectories: the autocorrelation-like functions stay roughly flat for $\\tau \\le \\rho$ and then decay, with slopes close to the labelled version for the degree- and spectrum-based measures.","Sensitive dependence on initial conditions produces an expansion phase in all three pseudo-distances for both low-dimensional and high-dimensional chaotic network dynamics, so chaos is qualitatively detectable without labels.","No model-independent Lyapunov exponent can be extracted from unlabelled trajectories, because the relation between each pseudo-distance and the labelled distance is nonlinear and varies with the generative model.","Empirical temporal networks with labels deliberately removed—US domestic flights and two face-to-face contact datasets—still show their main dynamical features, including the daily periodicity of the flight network.","The pseudo-distances are blind to dynamics that leave the chosen invariants unchanged; for example, any trajectory preserving the total number of edges yields a vanishing degree-sequence pseudo-distance."],"supporting_citations":[{"why":"Supplies the labelled network autocorrelation function and the dictionary trick used to generate periodic and chaotic trajectories, serving as the ground truth for unlabelled comparisons.","marker":"[22]"},{"why":"Provides the network version of Wolf's algorithm and the globally coupled chaotic map model used to measure sensitivity to initial conditions.","marker":"[23]"},{"why":"Defines the DARN($\\rho$) process whose memory structure is used to test whether unlabelled autocorrelation functions reproduce correlation decay.","marker":"[20]"},{"why":"Basis for the spectral pseudo-distance between graphs used as one of the three invariants.","marker":"[47]"},{"why":"Supplies the eigenvalue bounds that justify the normalization of the spectral pseudo-distance.","marker":"[48]"},{"why":"Defines eigenvector centrality, the invariant behind the third pseudo-distance.","marker":"[49]"},{"why":"Provides the US domestic flight temporal network used for empirical validation of periodicity detection.","marker":"[55]"},{"why":"Supplies the Malawi village face-to-face contact network used to test memory decay on empirical data.","marker":"[56]"},{"why":"Supplies the SFHH conference contact network used as a second empirical face-to-face validation dataset.","marker":"[57]"}],"fun_headline_variants":["Unlabelled networks still yield their dynamical secrets","Graph invariants recover periodicity and chaos sans labels","Node labels stripped, network dynamics remain readable","Dynamics of unlabelled networks cracked via invariants"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach only works when changes in the network show up in its degree list, its spectrum, or its centrality scores; a trajectory in which all three stay the same while the network genuinely changes would look perfectly static to every measure.","fun_headline_variants_meta":{"raw":{"variants":["Unlabelled networks still yield their dynamical secrets","Graph invariants recover periodicity and chaos sans labels","Node labels stripped, network dynamics remain readable","Dynamics of unlabelled networks cracked via invariants"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000215,"raw_usage":{"total_tokens":1427,"prompt_tokens":941,"completion_tokens":486,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":425}},"tokens_in":557,"tokens_out":486,"duration_ms":4478,"temperature":1.0,"reasoning_tokens":425,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:52:43.691632+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a labelled temporal trajectory whose snapshots are pairwise non-isomorphic but all share the same sorted degree sequence, the same adjacency spectrum, and the same sorted eigenvector centrality; if every pseudo-distance stays identically zero and $c_X(\\tau)$ shows no structure, then the claimed recovery fails for that dynamical class. Regular graphs with rearranged edges are a natural place to search for such a counterexample.","supporting_citations":[{"cited_title":"P., Eguiluz, V","cited_arxiv_id":null,"evidence_quote":"Supplies the labelled network autocorrelation function and the dictionary trick used to generate periodic and chaotic trajectories, serving as the ground truth for unlabelled comparisons."},{"cited_title":"M., Di Gaetano, L., Galla, T., Lacasa, L.: Lya- punov exponents for temporal networks","cited_arxiv_id":null,"evidence_quote":"Provides the network version of Wolf's algorithm and the globally coupled chaotic map model used to measure sensitivity to initial conditions."},{"cited_title":"Linear Algebra and its Applications 436(5), 1425 (2012)","cited_arxiv_id":null,"evidence_quote":"Basis for the spectral pseudo-distance between graphs used as one of the three invariants."},{"cited_title":"E., Haemers, W","cited_arxiv_id":null,"evidence_quote":"Supplies the eigenvalue bounds that justify the normalization of the spectral pseudo-distance."},{"cited_title":"Social Networks 29(4), 555 (2007)","cited_arxiv_id":null,"evidence_quote":"Defines eigenvector centrality, the invariant behind the third pseudo-distance."},{"cited_title":"J., Musolesi, M.: Spatio-temporal networks: reachability, centrality and robustness","cited_arxiv_id":null,"evidence_quote":"Provides the US domestic flight temporal network used for empirical validation of periodicity detection."},{"cited_title":"P., Haenni, S., Phuka, J., Leal-Neto, O","cited_arxiv_id":null,"evidence_quote":"Supplies the Malawi village face-to-face contact network used to test memory decay on empirical data."},{"cited_title":"https://doi.org/10.1140/ epjds/s13688-018-0140-1","cited_arxiv_id":null,"evidence_quote":"Supplies the SFHH conference contact network used as a second empirical face-to-face validation dataset."}],"review_version":1}