{"id":"ee6c6215-6df8-41ab-a279-998ca3f530eb","arxiv_id":"2412.02715","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Pairwise graph distances compressed by PCA or MDS yield scalar time series that inherit periodicity, memory, and chaos from temporal network trajectories.","lead":"This paper shows that a temporal network, a sequence of changing graphs, can be squeezed into a single time series by applying PCA or multidimensional scaling to the matrix of distances between network snapshots. This lets standard time series tools measure periodicity, memory, and chaos in networks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The scalar embeddings are one-dimensional projections (Eqs. 4, 5, 8), so they can discard dynamical signals that are not the dominant variance mode; the preservation claim is thus conditional, not general.","rationale":"The reader's weakest assumption is that the pairwise distance matrix is a sufficient statistic for the intrinsic dynamics. My concern is downstream: even if the distance matrix is sufficient, the scalar embedding is obtained by truncating the spectral decomposition to the first eigenvector, which is a one-dimensional projection. This projection can discard dynamical structure that is not aligned with the direction of maximum variance. That is a concrete, internal limitation of the method rather than an external consensus disagreement, and it is directly testable with the synthetic experiment described above. The paper's own discussion in Section 4 acknowledges that there should be a lower bound on embedding dimension below which dynamics cannot be recovered, but it asserts that this bound 'appears to saturate to one' for the analyzed range; that assertion is not supported by experiments that separate signal variance from noise variance. Appendix A is a genuine piece of independent support, but it covers scalar inputs only, not network trajectories. The proposed test would settle whether the preservation claim extends to a low-amplitude signal embedded in a high-variance noise background. Because the paper is framed as a proof of concept and the reader already assigned a conditional verdict, I do not recommend changing the verdict; the condition should explicitly include the requirement that the dynamical fingerprint of interest dominate the leading variance direction of the trajectory.","tokens_in":21772,"tokens_out":5397,"duration_ms":51920,"concrete_test":"Simulate T = 500 snapshots of N = 20 weighted networks with A_ij(t) = ξ_ij(t) + ε cos(2πt/P + φ_ij), where ξ_ij(t) are i.i.d. N(0,1), P = 50, φ_ij are i.i.d. uniform on [0,2π), and ε takes values such as 0.01, 0.05, 0.1, 0.2, 0.5, and 1. Compute D^(2) using Eq. (12), then the PCA-embedding and classical-MDS scalar embeddings via Eqs. (5) and (8), and evaluate the periodicity z-score of Eq. (13) on the scalar ACF. If for small ε the z-score drops below the paper's threshold of 3 while periodicity remains visible in, for example, the average link-level ACF or the nACF of the full trajectory, then the scalar embedding does not preserve the planted dynamics in a simple regime, and the abstract's claim must be restricted to dynamics that dominate the leading variance mode.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the truncation to the first principal component or eigenvector. Equations (4), (5), and (8) define the scalar embeddings as z_t = v_t · e_1, z_t = sqrt(Λ_1) e_t^1, and z_t = sqrt(λ_1) u_t^1, respectively. All three are one-dimensional projections of the network trajectory in graph space. Even granting that the pairwise squared-distance matrix D^(2) of Eq. (1) is a sufficient statistic—and Appendix A only proves recovery for scalar data, not for general network trajectories—a scalar projection can preserve only the component of the dynamics that aligns with the direction of maximum variance of the embedded points. The paper's synthetic validations are all cases where the planted dynamics is the dominant variance mode: the pulsating model is a collective oscillation across many links, the DARN memory is planted in every link, and the change-point examples switch the global generative process. No experiment independently varies the relative variance of the 'signal' and 'noise' directions. A network trajectory with a strong white-noise component and a weaker periodic or memory-bearing component is a direct counterexample to the abstract's unrestricted claim: the first principal component will track the high-variance noise, so the scalar ACF will lose the periodicity even though the full network trajectory contains it. This is not a failure of the distance matrix but of the dimension-truncation step, and it is not covered by the provided validations.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes to embed a temporal network trajectory—a sequence of network snapshots—into a low-dimensional, typically scalar, time series by applying PCA or MDS to the matrix of pairwise squared distances between snapshots. Four variants are defined: PCA-projection, PCA-embedding, classical-MDS, and metric-MDS. The central claim is that nontrivial dynamical properties of the original trajectory—periodicity, memory, change points, and maximum Lyapunov exponent—are preserved in the scalar embedding. The method is validated on one-dimensional time series, several synthetic network trajectory models (white, pulsating periodic, DARN memory, Type-1/2/3 periodic, dictionary-based chaotic), and two empirical networks (email and primary school contacts). Appendix A proves that classical-MDS recovers a one-dimensional scalar time series exactly up to a shift and a sign flip.","tokens_in":21984,"tokens_out":8582,"duration_ms":69796,"significance":"The paper is clearly written and proposes a simple, computationally accessible methodology that could be useful for temporal network analysis. The key idea of using inter-snapshot distances as features is well motivated and connects to existing work on network trajectories. The numerical experiments cover a broad range of models, and the Appendix A proof is elegant. The method demonstrably works on the tested models, and the empirical results are suggestive. However, the paper's central claim as stated in the abstract is too strong: the scalar embeddings are one-dimensional projections, and the evidence is partly circular in the chaotic case. With appropriate qualifications and additional control experiments, the contribution would be a useful proof of concept.","major_comments":[{"comment":"The chaotic validation is circular. The dictionary is constructed so that any two networks G[s] and G[t] are exactly |s−t| rewirings apart, hence ∥G[s]−G[t]∥ = |t−s|; after assigning each cell of the logistic map to a dictionary entry, the squared distance matrix of the network trajectory is (up to discretization) the squared distance matrix of the scalar logistic time series. Appendix A proves that classical-MDS recovers a one-dimensional signal up to shift and sign, and the PCA-embedding is numerically equivalent to classical-MDS on this input. The recovery of λ = ln 2 is therefore a corollary of the scalar proof, not an independent validation for network trajectories. Please replace or supplement this experiment with a network trajectory whose distance matrix is not that of a scalar time series, for example a chaotic latent variable driving a subset of links, and test whether the scalar embedding still recovers the correct Lyapunov exponent.","section":"3.2.5 and Appendix A"},{"comment":"The unrestricted claim that 'nontrivial dynamical properties of the network trajectories are preserved already in their scalar embeddings' is not supported for general network trajectories. All three successful scalar embeddings are one-dimensional projections of the trajectory in the feature space, so they can only capture the component of the dynamics aligned with the first principal component (or first eigenvector). In the experiments, the planted signal is global (pulsating model, DARN memory in every link, global change points) so it dominates the variance; no experiment varies the relative variance of a signal and an independent noise component. Provide a control experiment of the form G_t = A_noise(t) + ε A_signal(t), where A_noise(t) is a white-noise adjacency matrix and A_signal(t) is periodic or memory-bearing, and report the range of ε for which the scalar ACF recovers the signal. Until then, the abstract and Section 3.2.4 should be qualified to state that preservation holds when the relevant dynamics dominate the variance of the distance matrix.","section":"2.1, Eqs. (4), (5), (8)"},{"comment":"The paper states that the relative graph distance between snapshots is the property that needs to be preserved, but this sufficient-statistic assumption is not proven for network trajectories; Appendix A proves recovery only for scalar data. Moreover, even when the distance matrix is a sufficient statistic, the scalar embedding is a further lossy projection, so the conditions under which the first eigenvector preserves a given dynamical invariant (e.g., the maximum Lyapunov exponent or the memory order) remain unspecified. Please either provide a formal statement of when scalar recovery is exact (as in Appendix A for one-dimensional data) or explicitly frame the method as heuristic with empirical support, and discuss the classes of dynamics for which the first principal component is expected to suffice.","section":"Section 2 and Discussion (fourth paragraph)"}],"minor_comments":[{"comment":"There is an extra closing parenthesis in the denominator: σ(ACF(τ ))τ <P) should be σ(ACF(τ))_{τ<P}.","section":"Eq. (13)"},{"comment":"The caption says 'N = 500 snapshots'; the symbol N is used elsewhere for the number of nodes. Please use T = 500 snapshots.","section":"Fig. 3 caption"},{"comment":"The text refers to Fig. 8(C) for both the scalar autocorrelation function and the nACF; the scalar ACF is in panel (D) and the nACF in panel (B). Please correct the figure references.","section":"Section 3.3.1"},{"comment":"Typo: 'autorreggressive' should be 'autoregressive'.","section":"3.1, list of processes"},{"comment":"The statement ∥G[s]−G[t]∥ = |t−s| is inconsistent with the normalized Frobenius norm in Eq. (12) (which would give sqrt(|t−s|)/N) and also with the standard unnormalized Frobenius norm (which would give sqrt(2|t−s|) for a single edge rewiring). Please specify the norm used in the dictionary construction and how it relates to Eq. (12).","section":"3.2.5"},{"comment":"Please state explicitly whether the antiphase correction was applied to the metric-MDS embedding in the chaotic experiment and, if so, whether the smoothness assumption is justified for that signal; otherwise the reader cannot assess the impact of the correction on the reported Lyapunov exponent.","section":"Appendix B and 3.2.5"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid proof of concept with a clear presentation, but the circular chaotic validation and the unqualified preservation claim will likely draw criticism from referees. The requested control experiments (varying signal-to-noise in the variance sense, and a non-circular chaotic network model) are feasible and should be manageable within a revision. The paper fits the scope of the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one. It extends the earlier MDS work of Thongprayoon et al. to scalar embeddings via PCA and MDS on the snapshot distance matrix, and argues that nontrivial dynamics—periodicity, memory, change points, even Lyapunov exponents—survive in one dimension. The core idea is plausible and the paper is a good synthesis of four closely related methods.\n\nThe good bits first. The linear algebra is standard, and Appendix A gives a clean proof that classical-MDS exactly recovers any 1D scalar signal up to shift and reflection. That is a legitimate formal result and explains why the method works when the network trajectory is effectively low-dimensional. The numerical work is careful in places: they check periodicity under fixed number of links, varying network size and edge density, and even when only 10% of edges carry the periodic signal. The empirical examples (email, primary school) include a link-contamination control to rule out density-driven artifacts. Those details make the method credible.\n\nThe soft spots are real but not fatal. First, the chaotic validation is circular. The dictionary trick builds snapshots that are literally metrical: ||G[s] − G[t]|| = |s − t|, so the distance matrix equals that of the logistic time series. Recovering λ = ln 2 is then a direct corollary of the Appendix A proof for scalar data, not evidence about general network trajectories. The paper should reframe that as a consistency check, not independent validation. Second, most comparisons are qualitative and on single realizations, with no error bars across realizations. The z-score thresholds help for periodicity detection, but the memory and change-point results are one-shot illustrations. Third, the abstract says the results \"certify\" preservation; that overreaches. Fourth, no code is currently available.\n\nThere is also a structural limitation the paper never addresses head-on: all three successful embeddings are one-dimensional projections. If the dynamical signal is not the dominant variance mode of the distance matrix, the first principal component can track high-variance noise and lose the dynamics. The paper's validations all plant the signal as the dominant mode; the 10%-edge experiment helps but does not systematically vary the relative variance of signal versus noise directions. The stress-test note I saw makes exactly this point, and it holds. A more cautious claim—preservation holds when the dynamics aligns with the dominant variance direction—would be accurate.\n\nWho is this for? Anyone working on temporal network analysis who wants a quick way to turn network snapshots into a scalar time series. It deserves serious refereeing: the method is useful, the proof is sound, and the weaknesses are addressable with code, error bars, a rewritten abstract, and a reframed chaotic test. I would accept it for peer review and require those revisions.","headline":"A useful, clearly written toolbox for reducing temporal network trajectories to scalar signals, but the preservation claim is more conditional than the abstract suggests, and the chaotic validation is circular.","tokens_in":22613,"tokens_out":3924,"would_cite":true,"duration_ms":35627,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["89.75.-k","05.45.-a"],"model":"deepseek-v4-flash","headline":"A temporal network trajectory can be compressed into a one-dimensional time series—by projecting the matrix of pairwise distances between snapshots—without losing its main dynamical fingerprints: autocorrelation, periodicity, memory…","keywords":["temporal networks","network trajectory","graph distance matrix","multidimensional scaling","principal component analysis","scalar embedding","autocorrelation function","Lyapunov exponent"],"falsifier":"Generate a network trajectory made of two independent edge sets oscillating at two incommensurate periods with comparable noise levels, so that no single scalar coordinate can track both oscillations, and check whether the scalar embedding's autocorrelation function shows both periods. If either period is missing, the claim that scalar embeddings inherit multiscale dynamics is falsified; the paper reports such multiscale results only as 'results not shown'.","tokens_in":21467,"feed_emoji":"🕸️","tokens_out":13275,"duration_ms":104459,"temperature":0.7,"pith_summary":"A temporal network, viewed as a sequence of snapshots, can be treated as a trajectory in graph space, and the paper argues that the information needed to characterize its dynamics lies in the relative distances between snapshots rather than in the internal structure of any single snapshot. From the matrix of squared pairwise graph distances it builds scalar embeddings via PCA and multidimensional scaling, yielding one time-series value per snapshot. Across synthetic models—uncorrelated, noisy periodic, memory-carrying, and chaotic network trajectories—and two empirical datasets, the three spectral strategies (PCA-projection, PCA-embedding, and classical MDS) reproduce the network autocorrelation function, the period, the memory time scale, change points, and the maximum Lyapunov exponent. Metric MDS is the exception, introducing occasional sign flips that corrupt the signal. The practical upshot is that standard time series analysis becomes directly applicable to temporal networks after this distance-based reduction.","feed_headline":"A single number per snapshot preserves temporal network dynamics","feed_subtitle":"Distance-based PCA and MDS embeddings recover periodicity, memory, change points, and chaos from a network trajectory.","key_machinery":"The central object is the matrix of pairwise squared graph distances between snapshots, $D^{(2)}_{t\\ell}=\\|G_t-G_\\ell\\|^2$, evaluated with a fixed node labeling and the normalized Frobenius norm of Eq. (12). The PCA-based and classical-MDS strategies are both spectral decompositions of a centered version of this matrix: PCA centers columns and diagonalizes the covariance, while classical MDS double-centers to reconstruct the Gram matrix $B=-\\tfrac12 J D^{(2)} J$ and takes its leading scaled eigenvector as the scalar coordinate. The algebra that makes this work is the identity $d^2_{t\\ell}=\\|x_t\\|^2+\\|x_\\ell\\|^2-2x_t\\cdot x_\\ell$, which connects squared distances to an inner-product space so that the leading eigenvector becomes a faithful one-dimensional coordinate for the relative geometry of the snapshots. What carries the argument is the deliberate choice to preserve inter-snapshot distances rather than intra-snapshot topology.","core_discovery":"The central claim is that nontrivial dynamical properties of a network trajectory are preserved already in its scalar embedding. Given a trajectory $S=(G_1,\\ldots,G_T)$ with adjacency matrices $A(t)$, the paper forms the squared-distance matrix $D^{(2)}_{t\\ell}=\\|G_t-G_\\ell\\|^2$ using the normalized Frobenius norm of Eq. (12), and extracts a scalar time series $z_t$ either by projecting the column-centered distance matrix onto the first principal component, by taking the first scaled eigenvector, or by classical MDS double-centering. The finding is that these $z_t$ inherit the network autocorrelation function—including the flat segment for lags below the memory order and the exponential tail—along with periodicity, the location of change points, and the largest Lyapunov exponent of chaotic trajectories. The one rigorous guarantee, proven in Appendix A, is that for one-dimensional input the classical-MDS embedding recovers the signal up to a mean shift and a global sign flip; for full network trajectories the preservation is demonstrated across the synthetic suite and two real datasets.","pith_inferences":["Because the embedding is a deterministic function of the pairwise distance matrix, any two network trajectories sharing the same distance structure will be indistinguishable after reduction; this suggests a characterization of when distance matrices are sufficient statistics (e.g., for isometry-invariant dynamics) as a natural theoretical follow-up.","The success of the scalar embeddings across very different models suggests a universality: for dynamics that are themselves low-dimensional in graph space, a single coordinate may suffice, and the first principal component's explained variance could be used as an a priori diagnostic of whether a scalar reduction is faithful.","The same distance-matrix recipe can be extended to unlabeled networks by replacing the Frobenius norm with a graph-invariant distance (e.g., spectral distance), which the paper explicitly identifies as an open direction; a testable prediction is that the preservation results would carry over.","The 'antiphase' failure of metric-MDS is analogous to sign ambiguities in eigenvector-based embeddings in other spectral methods; an online or incremental variant of the successful PCA/MDS strategies could stabilize the embedding when new snapshots are appended, an issue the paper only partially explores."],"forward_implications":["A temporal network can be reduced to an ordinary scalar time series, so the full toolkit of time series analysis—spectral estimation, change-point detection, forecasting—can be applied to network trajectories.","Periodicity in a network trajectory survives embedding even when the average degree is constant and when only a small fraction (about 10%) of edges carry the periodic activity, so embeddings isolate genuine dynamical structure rather than trivial density fluctuations.","Memory in DARN(p) processes is inherited: the flat-then-decaying shape of the network autocorrelation function appears in the autocorrelation of the scalar embedding, and change points between different memory orders are visible as variance changes in the embedding.","Chaotic network trajectories built by the dictionary trick yield scalar embeddings whose recurrence expansion rate reproduces the theoretical Lyapunov exponent ($\\ln 2$ for the fully chaotic logistic map), so dynamical instability is measurable after reduction.","The scalar embedding behaves as a noise filter: it enhances the signal-to-noise ratio of periodic and memory-laden trajectories, improving the visibility of periodicity in empirical networks."],"supporting_citations":[{"why":"Defines the network autocorrelation function that serves as the ground truth for memory and periodicity comparisons.","marker":"[18]"},{"why":"Defines the network maximum Lyapunov exponent used as the ground truth for chaotic trajectories.","marker":"[22]"},{"why":"Introduced MDS embedding of tie-decay temporal networks, the starting point the present work expands and systematically validates.","marker":"[39]"},{"why":"Supplies the PCA formulation, including column centering and spectral decomposition, for the two PCA-based strategies.","marker":"[45]"},{"why":"Supplies the classical and metric multidimensional scaling framework, including double centering and the stress function.","marker":"[46]"},{"why":"Provides the recurrence-based algorithm used to estimate the maximum Lyapunov exponent from the scalar embedding.","marker":"[47]"},{"why":"Introduces the DARN model of temporal-network memory that serves as a testbed for memory inheritance.","marker":"[25]"}],"fun_headline_variants":["Scalar embedding preserves temporal network dynamics","One scalar per snapshot retains network trajectory dynamics","Distance-based scalar embedding keeps dynamics of network trajectories","Single-number time series captures temporal network dynamics","Scalar projection of snapshot distances preserves dynamics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the matrix of squared pairwise distances between snapshots, defined for a fixed labeling of the nodes and a chosen matrix norm, contains enough information about the intrinsic dynamics of the temporal network; if two trajectories with different dynamics ever share the same distance matrix, no embedding built on those distances can tell them apart.","fun_headline_variants_meta":{"raw":{"variants":["Scalar embedding preserves temporal network dynamics","One scalar per snapshot retains network trajectory dynamics","Distance-based scalar embedding keeps dynamics of network trajectories","Single-number time series captures temporal network dynamics","Scalar projection of snapshot distances preserves dynamics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000268,"raw_usage":{"total_tokens":1617,"prompt_tokens":942,"completion_tokens":675,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":609}},"tokens_in":558,"tokens_out":675,"duration_ms":5555,"temperature":1.0,"reasoning_tokens":609,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:59:51.051345+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a network trajectory made of two independent edge sets oscillating at two incommensurate periods with comparable noise levels, so that no single scalar coordinate can track both oscillations, and check whether the scalar embedding's autocorrelation function shows both periods. If either period is missing, the claim that scalar embeddings inherit multiscale dynamics is falsified; the paper reports such multiscale results only as 'results not shown'.","supporting_citations":[{"cited_title":"Egu ´ ıluz, Leonardo Di Gaetano, Tobias Galla, and Lucas Lacasa","cited_arxiv_id":null,"evidence_quote":"Defines the network maximum Lyapunov exponent used as the ground truth for chaotic trajectories."},{"cited_title":"Rodriguez, and Victor M","cited_arxiv_id":null,"evidence_quote":"Defines the network autocorrelation function that serves as the ground truth for memory and periodicity comparisons."},{"cited_title":"Embedding and trajectories of temporal networks","cited_arxiv_id":null,"evidence_quote":"Introduced MDS embedding of tie-decay temporal networks, the starting point the present work expands and systematically validates."},{"cited_title":"Jolliffe","cited_arxiv_id":null,"evidence_quote":"Supplies the PCA formulation, including column centering and spectral decomposition, for the two PCA-based strategies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the classical and metric multidimensional scaling framework, including double centering and the stress function."},{"cited_title":"Determining lyapunov exponents from a time series","cited_arxiv_id":null,"evidence_quote":"Provides the recurrence-based algorithm used to estimate the maximum Lyapunov exponent from the scalar embedding."},{"cited_title":"The shape of memory in temporal networks","cited_arxiv_id":null,"evidence_quote":"Introduces the DARN model of temporal-network memory that serves as a testbed for memory inheritance."}],"review_version":1}