Pith. sign in

REVIEW 3 major objections 6 minor 158 references

Every node-exchangeable discrete-time network process can be represented by a 2N-graphon, and a two-stage least-squares estimator recovers it at explicit rates that separate the network's spatial approximation from the quality of edge-level

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 03:27 UTC pith:SKPD46XR

load-bearing objection New two-stage rates for temporal graphon estimation; core is solid, but Theorem 3.2 overstates the α>1 case and the hospital analysis sits outside the Markov-chain assumptions. the 3 major comments →

arxiv 2607.23145 v1 pith:SKPD46XR submitted 2026-07-25 stat.ME

Decorated graphons for temporal network estimation

Charles Dufour , Sofia C. Olhede This is my paper
classification stat.ME MSC 62G0562H3005C80
keywords decorated graphonstemporal networksgraphon estimationnode exchangeabilitystochastic block modelsMarkov chainsnonparametric estimationnetwork histograms
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper's central claim is that every node-exchangeable discrete-time network process can be represented by a single nonparametric object, a 2N-graphon (probability-graphon): a symmetric function assigning to each latent node pair a probability distribution over binary edge histories. This makes dynamic stochastic block models, autoregressive edge models, and Markov edge models special cases of one framework that preserves exchangeability. The paper proposes a two-stage estimator—estimate each edge's temporal law separately, then cluster nodes by blockwise least squares on those summaries—and proves that its error is a sum of a spatial approximation floor, fixed by the smoothness of the graphon, and an edge-estimation term governed by ψ² and β. In the Hölder-smooth regime the rate is n^{-2α/(α+1)} plus ψ² log n/n plus β², and in the block-model regime ψ²(k²/n² + log k/n)+β²; ψ² is 1/T for Bernoulli edges and roughly d√τ/(Tπ_*^{3/2}) for Markov chains. If correct, the field gains a common nonparametric baseline for dynamic networks, with guarantees that transfer to any edge model whose summaries concentrate sub-Gaussianly.

Core claim

The discovery is that time-evolving networks can be viewed as decorated graphs whose edge decorations are infinite binary sequences, and that the law of any jointly exchangeable graph-valued process equals a 2N-graphon W*: W(x,y) is a probability measure on {0,1}^N. Estimation then decouples: first fit each observed edge series by a parametric or summary map Λ, obtaining λij; then fit a k-block model to the λij by least squares, obtaining θ̂. Theorems 3.1 and 3.2 bound the mean squared error by C(ψ²(k²/n² + log k/n) + β²) and C((1+ψ²)n^{-2α/(α+1)} + ψ² log n/n + β²), where ψ is the sub-Gaussian concentration of the edge summaries and β their worst-case bias. In the Markov case, Λ is the empi

What carries the argument

The central object is the 2N-graphon (probability-graphon): a Borel-measurable symmetric function W:[0,1]^2 → P({0,1}^N), assigning to each pair of latent positions a probability distribution over infinite binary edge time series. The carrying mechanism is the two-stage least-squares estimator: stage one maps each observed binary edge series through Λ to a summary λij in a Hilbert parameter space X; stage two solves a k-block least-squares problem over assignments z and block parameter matrices Q. Assumption 3—λij sub-Gaussian with uniform norm ψ and bias β—is the single interface that lets any edge model plug in. A version of the argument uses a De Bruijn-state expansion to turn M-th order

Load-bearing premise

The load-bearing premise is Assumption 3: every edge's summary estimator must be sub-Gaussian with a uniform bound ψ and bias β across all edges—and, for the Markov rate, each edge chain must be irreducible, aperiodic, with stationary probabilities bounded below by π* and uniformly bounded mixing time; sparse or near-degenerate edges violate this and are exactly the edges the hospital analysis drops.

What would settle it

Take a block model with k=O(1) and π* → 0 for a positive fraction of edges (e.g., edge activation probability of order 1/T), run the empirical-transition-matrix estimator, and check whether the blockwise least-squares error decays at ψ²(k²/n² + log k/n) with ψ² ≈ d√τ/(Tπ_*^{3/2}); if low-activity edges make the concentration fail or the bias dominate, the stated rate collapses. A second check: verify whether the additive n^{-2α/(α+1)} floor is genuinely independent of T by increasing T at fixed n in the smooth regime and observing the error plateau.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • The error of any estimator in this framework is transparently additive: a spatial floor from block approximation plus an edge-estimation term ψ² log n/n (Hölder) or ψ²(k²/n² + log k/n) (block model) plus β².
  • For Bernoulli edge processes ψ²=1/T, so T=1 recovers the known minimax graphon rate; for α≥1 temporal replication reduces the clustering penalty from n^{-1} log n to (nT)^{-1} log n.
  • For Markov edge processes satisfying Assumption 4, ψ² ≈ d√τ/(Tπ_*^{3/2}) log(2d/π_*) and β decays exponentially in T, so after a moderate T the temporal error is dominated by ψ².
  • Any edge model with an identifiable parameter and a sub-Gaussian, near-unbiased estimator can be substituted in stage one without changing the network-level theorems.
  • Existing dynamic models—Pensky's dynamic graphon, BALARM, autoregressive edge models—are special cases of the decorated-graphon representation, so the framework provides a common nonparametric baseline.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that the framework assigns information to never-observed edges: pairs that never interact still encode a low activation probability and can be modeled by a degenerate law, so discarding such edges in sparse data may bias block recovery; testing a missing-at-random or cemetery-point variant on the hospital data would be a direct extension.
  • A hybrid estimator is worth exploring: the paper notes pooling and two-stage estimation are asymptotically equivalent in the block regime but heterogeneous in the smooth regime, so a method that pools only within well-identified blocks could reduce variance in exactly the sparse regimes where the two-stage alternative struggles.
  • If the representation claim holds at the discrete-time level, the natural next test is continuous time: the paper itself flags that no extension of the decorated-graphon representation to continuous-time edge processes is known, so establishing one would transfer the same two-stage recipe to point-process networks.
  • The n^{-2α/(α+1)} floor being independent of T suggests that, in smooth regimes, longer edge histories improve clustering but cannot fix the nonparametric graphon approximation; this points to investing in more nodes or in post-hoc smoothing rather than longer observation windows.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes a unified nonparametric framework for temporal networks using decorated graphons (2N-graphons), in which each node pair is assigned a distribution over a binary edge time series. Under global node exchangeability, such processes admit a Kallenberg-type representation. The authors develop a two-stage estimator: each edge’s time series is first summarized by a parametric estimator Lambda, and the resulting parameter matrix is then approximated by a blockwise least-squares graphon estimator. They prove convergence rates for block-model and Hölder-smooth graphons, with explicit dependence on the edge-wise estimation error psi and bias beta, and specialize to Bernoulli and Markov-chain edge processes. The methodology is illustrated on simulations and a hospital contact network.

Significance. If the results hold, the framework provides a valuable unified baseline for dynamic network estimation, reconciling node exchangeability with nontrivial temporal edge dynamics. The modular two-stage construction is conceptually clean, and the explicit rates separating spatial approximation from temporal estimation are useful. The supplement is thorough, includes the main proofs, and the authors provide code and discuss limitations candidly. The block-model and alpha<=1 Hölder rates are plausible and build on established techniques; the Markov-chain concentration bounds, if correct, are a useful technical contribution.

major comments (3)
  1. [Theorem 3.2, §A.2.2, Proposition 4.1] Theorem 3.2 states the rate n^{-2α/(α+1)} for all α>0, but the proof (Eq. (A.24)) with k=⌈n^{1/(α∧1+1)}⌉ yields (1/k²)^{α∧1} + ψ²(k²/n² + log k/n). For α>1 this is n^{-1} + ψ² log n/n up to constants, not n^{-2α/(α+1)}. Proposition 4.1 correctly gives the saturated (nT)^{-1} log n rate for α≥1. The theorem statement and the surrounding text therefore overclaim for α>1 and must be corrected.
  2. [§6, §7, Assumption 4, Lemma 4.1] The hospital analysis restricts attention to pairs with at least one observed contact, while the Markov-chain theory (Assumption 4 and Lemma 4.1) requires every edge chain to be irreducible and aperiodic, with a uniform lower bound π_*>0 on stationary probabilities and enough visits to every state. Zero-contact pairs have stationary mass near zero and violate these conditions; conditioning on the edge-specific event also changes the retained edge laws. Section 7’s assertion that degenerate edges are covered because ‘our bounds hold uniformly over edges’ is not supported by Lemma 4.1, and no boundary-case verification is supplied. The data illustration is outside the theorem’s regime, and the coverage claim should be either removed or backed by an appropriate extension.
  3. [§3.2, Eq. (5) and following paragraph] The claim that the rate n^{-2α/(α+1)} is minimax optimal for all T is not established by the argument given. Eq. (5) is a lower bound for the T=1 Bernoulli model; the sentence ‘additional temporal observations can only help, so this floor persists’ would only imply R(T) ≤ R(1), which is the wrong direction for a lower bound. The conclusion can be rescued by embedding the T=1 model through perfectly correlated edge time series (a valid 2N-graphon), but that argument is absent. Please provide a correct lower-bound argument.
minor comments (6)
  1. [Theorem 3.1 statement] The display has a small typo: θ_{ij} = Q_{z(i)z(j)} should be typeset with a subscript separator; consider fixing throughout.
  2. [§4.1, Proposition 4.1] The two-case display for α<1 and α≥1 should be reconciled with the corrected Theorem 3.2; it may be clearer to state the saturated rate directly after the theorem.
  3. [Definition 2.1] The notation W(2^N) is introduced but the term ‘2N-graphon’ is used throughout; consider consistent formatting.
  4. [§4.2, Figure 3] The statement that all transition probabilities bounded away from 0 is sufficient for irreducibility and aperiodicity is correct, but the De Bruijn example in Figure 3 may not satisfy it; please clarify the range of allowed transition matrices.
  5. [Abstract and §2] The abstract mentions autoregressive edge processes as special cases, but the verification in the main text covers only Bernoulli and Markov chains; either add a verification for AR-type processes or soften the claim.
  6. [§D.1] The missing-data extension is described only informally; the statement that rates remain valid ‘under appropriate regularity conditions’ should be marked as a conjecture or provided with proof in the supplement.

Circularity Check

0 steps flagged

No significant circularity: the central representation theorem and convergence rates rest on external results (Kallenberg; Gao et al.; Wolfer–Kontorovich/Paulin), and the hospital comparison is retrospective, not a fitted prediction.

full rationale

The paper's load-bearing claims are not circular. The representation of node-exchangeable graph-valued processes as 2N-graphons is attributed to Kallenberg (2005), an external theorem, and the decorated-graphon formalism is a definition rather than a derived prediction. The main estimation rates (Theorems 3.1 and 3.2) are proved in the supplement by standard least-squares oracle inequalities, with the only input being Assumption 3 on edge summaries (sub-Gaussian concentration with variance proxy ψ² and bias β). The Bernoulli example derives ψ²=1/T by direct Hoeffding-type concentration, and the Markov-chain example proves Lemma 4.1 using external concentration results of Wolfer and Kontorovich (2019) and Paulin (2015); no target rate is assumed in the proofs. The minimax optimality argument for α<1 invokes Gao et al. (2015, Theorem 1.2) for the static T=1 submodel, which is a valid lower-bound argument, not a circular one. Self-citations (Dufour & Olhede 2024; Süveges & Olhede 2023; Verdeyme & Olhede 2024) are present but not load-bearing: they supply background, a comparison model, and optional smoothing, and they are not used to force the central claims. The hospital analysis compares the fitted 3-block decorated SBM with BALARM clusters from prior work on the same dataset; this is a descriptive comparison, not a prediction from a fitted parameter to a closely related held-out quantity, so it is not fitted-input-called-prediction. One non-circular weakness should be flagged: Assumption 4 and Lemma 4.1 require every edge's Markov chain to be irreducible, aperiodic, with stationary mass bounded below by π*>0, and Supplement D.1 explicitly states that Section 6 excluded zero-contact edges 'since our estimator assumes that every edge follows a well-defined Markov process'. Section 7's claim that 'Pairs that never interact remain informative' and that 'our bounds hold uniformly over edges and so cover this case' is therefore unsupported by the paper's own Markov-chain argument. This is a correctness/assumption gap, not a circularity, and it does not raise the circularity score.

Axiom & Free-Parameter Ledger

2 free parameters · 6 axioms · 0 invented entities

The central rates are purchased with exchangeability, conditional edge independence, sub-Gaussian edge summaries, and (for Markov examples) uniform positive stationary mass. The hyperparameters k, M, P are selected by BIC or hand rather than derived from the theory. No new physical entities are introduced; 2N-graphons are a mathematical representation inherited from the decorated-graphon literature.

free parameters (2)
  • number of blocks k = selected by BIC/HBIC in applications; k≈√n in simulations
    Theorems 3.1/3.2 take k as an input (e.g., k=⌈n^{1/(α∧1+1)}⌉); in practice the paper chooses k by BIC/HBIC without a guarantee that the selected k achieves the theoretical rate.
  • Markov order M and period P = M=1 or 2, P=96 in hospital analysis; P=96 in simulation
    The edge-process family and Theorem 4.1 rates depend on M and P through d=2^MP; the paper selects these by BIC or prior knowledge, but the theorem's constants are not restated for d=2^MP.
axioms (6)
  • domain assumption Global exchangeability of the graph-valued process (Assumption 1).
    Underpins the 2N-graphon representation and prevents node labels from carrying information; if real networks have role- or attribute-driven non-exchangeability, the model is misspecified. Invoked in Section 2 and throughout.
  • domain assumption Conditional independence of edge processes given latent variables.
    Required for sub-Gaussian concentration of sums of λ_ij in Lemmas A.1–A.3 and explicitly excludes triadic/feedback effects (Section 7). This is the main structural restriction of the framework.
  • standard math Kallenberg (2005) representation theorem for jointly exchangeable arrays.
    Used to assert that every exchangeable discrete graph-valued process is a 2N-graphon (Section 2). Paper relies on it without proof.
  • domain assumption Edge estimator Λ is sub-Gaussian with uniform constants ψ and β (Assumption 3).
    Theorems 3.1 and 3.2 depend on this concentration/bias condition; verifying it is the main work for each edge-process model.
  • domain assumption Irreducible, aperiodic Markov chains with stationary mass bounded below by π* and bounded mixing time (Assumption 4).
    Needed for Lemma 4.1/Theorem 4.1; sparse zero-contact edges violate it, which is why the hospital analysis drops such pairs (Section 6) without a fully worked correction.
  • domain assumption Hölder smoothness W∈H^α(M) and iid Uniform[0,1] latent variables.
    The Theorem 3.2 rate uses these; the estimator is not adaptive to α, and the k choice in the theorem depends on it.

pith-pipeline@v1.3.0-alltime-deepseek · 31057 in / 21332 out tokens · 196187 ms · 2026-08-01T03:27:57.323118+00:00 · methodology

0 comments
read the original abstract

We propose a unified nonparametric framework for modeling time-evolving networks using decorated graphons (also known as probability-graphons): symmetric functions that assign to each node pair a probability distribution over binary edge time series. This generalizes the static decorated-graphon construction to dynamic graphs while preserving node exchangeability and allowing temporal dynamics such as memory and periodicity. Models in which edges evolve independently given the latent variables, such as autoregressive and Markov edge processes, arise as special cases. We develop a two-stage estimation procedure that separates temporal modeling from network structure. Because the network stage requires only mild regularity conditions on the edge-process estimator, a broad class of temporal edge models can be used in the first stage. We establish nonparametric convergence rates in both block-model and H\"older-smooth regimes, and make explicit how the rate depends on the number of observed time steps and on the quality of the edge-level estimation. We illustrate the method on simulated data and a hospital contact network, recovering latent community structure and time-varying interaction patterns. The framework gives a nonparametric baseline for dynamic network analysis with explicit convergence guarantees.

Figures

Figures reproduced from arXiv: 2607.23145 by Charles Dufour, Sofia C. Olhede.

Figure 1
Figure 1. Figure 1: Illustration of changing from an infinite sequence of simple graphs to a 2 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Flowchart of our estimation procedure. The observed edge series [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Representation of the decorated graphon where edge dynamics are governed by a [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Estimated transition heatmaps (as depicted in Figure [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Normalized estimation error ∑i, j ∥θˆ i j −θi j∥ 2 2 plotted against n on the left (curves colored by T) and against T on the right (curves colored by n). Error bands denote one standard deviation. We used an ordered initialization of the node for the top row and a random initialization for the bottom one; even though the random initialization yields higher variability, the overall trends remain the same, … view at source ↗
Figure 6
Figure 6. Figure 6: On the left, fitted time varying probability of edge activation for the model picked via the right-hand side plot. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Role-specific contact rates and overall contact volume over time in the hospital ward after the 15-minutes [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Result of fitting a decorated SBM with 3 blocks to the hospital ward contact data with daily periods. Panel A represents each different type of connectivity, where black marks indicate that no contact occurred. Panel B shows the composition of the 3 node communities in terms of roles. Panel C shows the fitted time varying probability of edge activation and prediction interval obtained via bootstrap with co… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

158 extracted references · 7 canonical work pages

  1. [1]

    and Hall, Georgina , date =

    Abbe, Emmanuel and Bandeira, Afonso S. and Hall, Georgina , date =. Exact

  2. [2]

    Probability-Graphons:

    Abraham, Romain and Delmas, Jean-François and Weibel, Julien , date =. Probability-Graphons:

  3. [3]

    doi:10/hbth6m , eventtitle =

    Large-Scale Analysis of Disease Pathways in the Human Interactome , author =. doi:10/hbth6m , eventtitle =

  4. [4]

    Representations for Partially Exchangeable Arrays of Random Variables , author =

  5. [5]

    On Semidefinite Relaxations for the Block Model , author =

  6. [6]

    Anderson, T. W. and Goodman, Leo A. , date =. Statistical. 2237025 , eprinttype =

  7. [7]

    Advances in Neural Information Processing Systems , volume =

    Arastuie, Makan and Paul, Subhadeep and Xu, Kevin , date =. Advances in Neural Information Processing Systems , volume =

  8. [8]

    and Vogelstein, Joshua T

    Arroyo, Jesús and Athreya, Avanti and Cape, Joshua and Chen, Guodong and Priebe, Carey E. and Vogelstein, Joshua T. , date =. Inference for

  9. [9]

    doi:10/gp2s23 , eventtitle =

    Compressing Data on Graphs with Clusters , author =. doi:10/gp2s23 , eventtitle =

  10. [10]

    Distributional Reduction:

    Assel, Hugues Van and Vincent-Cuaz, Cédric and Courty, Nicolas and Flamary, Rémi and Frossard, Pascal and Vayer, Titouan , date =. Distributional Reduction:

  11. [11]

    2008.04790 , eprinttype =

    Community Recovery in Non-Binary and Temporal Stochastic Block Models , author =. 2008.04790 , eprinttype =

    Pith/arXiv arXiv 2008

  12. [12]

    Stochastic

    Barbillon, Pierre and Donnet, Sophie and Lazega, Emmanuel and Bar-Hen, Avner , date =. Stochastic

  13. [13]

    Bauer, Martin and Mémoli, Facundo and Needham, Tom and Nishino, Mao , date =. The. 2408.08233 , eprinttype =

    arXiv

  14. [14]

    doi:10/hbth6p , howpublished =

    Begleiter, Henri , date =. doi:10/hbth6p , howpublished =

  15. [15]

    and Bevacqua, D

    Bellingeri, M. and Bevacqua, D. and Sartori, F. and Turchetto, M. and Scotognella, F. and Alfieri, R. and Nguyen, N. K. K. and Le, T. T. and Nguyen, Q. and Cassi, D. , date =. Considering Weights in Real Social Networks:

  16. [16]

    , date =

    Bezanson, Jeff and Edelman, Alan and Karpinski, Stefan and Shah, Viral B. , date =. Julia:

  17. [17]

    Multilayer Networks:

    Bianconi, Ginestra , date =. Multilayer Networks:

  18. [18]

    Asymptotic Normality of Maximum Likelihood and Its Variational Approximation for Stochastic Blockmodels , author =

  19. [19]

    A Nonparametric View of Network Models and

    Bickel, Peter and Chen, Aiyou , date =. A Nonparametric View of Network Models and

  20. [20]

    Statistical

    Billingsley, Patrick , date =. Statistical. 2237603 , eprinttype =

  21. [21]

    and Chayes, J

    Borgs, C. and Chayes, J. T. and Lovász, L. and Sós, V. T. and Vesztergombi, K. , date =. Convergent Sequences of Dense Graphs

  22. [22]

    Borgs, Christian and Chayes, Jennifer and Cohn, Henry and Zhao, Yufei , date =. An

  23. [23]

    and Cohn, Henry and Holden, Nina , date =

    Borgs, Christian and Chayes, Jennifer T. and Cohn, Henry and Holden, Nina , date =. Sparse

  24. [24]

    Brémaud, Pierre , date =. Markov. doi:10.1007/978-3-030-45982-6 , isbn =

    work page doi:10.1007/978-3-030-45982-6

  25. [25]

    Disturbed Connectivity of

    Cao, Rui and Wu, Zheng and Li, Haifang and Xiang, Jie and Chen, Junjie , date =. Disturbed Connectivity of

  26. [26]

    Emergence of Network Features from Multiplexity , author =

  27. [27]

    doi:10.1214/12-EJS729 , issue =

    Consistency of Maximum-Likelihood and Variational Estimators in the Stochastic Block Model , author =. doi:10.1214/12-EJS729 , issue =

    doi:10.1214/12-ejs729

  28. [28]

    Structured Matrix Learning under Arbitrary Entrywise Dependence and Estimation of

    Chai, Jinhang and Fan, Jianqing , date =. Structured Matrix Learning under Arbitrary Entrywise Dependence and Estimation of

  29. [29]

    Chan, Stanley and Airoldi, Edoardo , date =. A

  30. [30]

    2202.09326 , eprinttype =

    Edge Coherence in Multiplex Networks , author =. 2202.09326 , eprinttype =

    Pith/arXiv arXiv

  31. [31]

    Local Linear Graphon Estimation Using Covariates , author =

  32. [32]

    2001.04938 , eprinttype =

    Nonparametric Regression for Multiple Heterogeneous Networks , author =. 2001.04938 , eprinttype =

    Pith/arXiv arXiv 2001

  33. [33]

    and MacDonald, Peter W

    Chang, Jinyuan and Fang, Qin and Kolaczyk, Eric D. and MacDonald, Peter W. and Yao, Qiwei , date =. Autoregressive. 2404.15654 , eprinttype =

    arXiv

  34. [34]

    Classification and Estimation in the

    Channarond, Antoine and Daudin, Jean-Jacques and Robin, Stéphane , date =. Classification and Estimation in the. doi:10.1214/12-ejs753 , issue =

    doi:10.1214/12-ejs753

  35. [35]

    Remapping the Brain to Compensate for Impairment in Recovering Alcoholics , author =

  36. [36]

    Matrix Estimation by

    Chatterjee, Sourav , date =. Matrix Estimation by

  37. [37]

    Chowdhury, Samir and Mémoli, Facundo , date =. The

  38. [38]

    Generalized Dimension Reduction Using Semi-Relaxed

    Clark, Ranthony A and Needham, Tom and Weighill, Thomas , date =. Generalized Dimension Reduction Using Semi-Relaxed

  39. [39]

    Dynamic Random Networks and Their Graph Limits , author =

  40. [40]

    doi:10.1201/9781315209661 , isbn =

    Probabilistic foundations of Statistical Network Analysis , author =. doi:10.1201/9781315209661 , isbn =

    work page doi:10.1201/9781315209661

  41. [41]

    Relative Exchangeability with Equivalence Relations , author =

  42. [42]

    Sugli Estremi Dei Momenti Delle Funzioni Di Ripartizione Doppia , author =

  43. [43]

    A Combinatorial Problem , author =

  44. [44]

    and Demaine, Martin L

    Demaine, Erik D. and Demaine, Martin L. , date =. Jigsaw

  45. [45]

    0712.2749 , eprinttype =

    Graph Limits and Exchangeable Random Graphs , author =. 0712.2749 , eprinttype =

    Pith/arXiv arXiv

  46. [46]

    https://doi.org/10.1137/1118101 , pages =

    Calculation of the Wasserstein Distance between Probability Distributions on the Line , author =. https://doi.org/10.1137/1118101 , pages =

    doi:10.1137/1118101

  47. [47]

    and Kramarz, Francis and Choné, Philippe and D'Haultfoeuille, Xavier , date =

    Donier-Meroz, Etienne and Dalalyan, Arnak S. and Kramarz, Francis and Choné, Philippe and D'Haultfoeuille, Xavier , date =. Graphon. 2304.03590 , eprinttype =

    Pith/arXiv arXiv

  48. [48]

    2408.12339 , eprinttype =

    Inference for Decorated Graphs and Application to Multiplex Networks , author =. 2408.12339 , eprinttype =

    Pith/arXiv arXiv

  49. [49]

    doi:10.5281/zenodo.10212972 , organization =

    Dufour, Charles and Grainger, Jake , date =. doi:10.5281/zenodo.10212972 , organization =

    work page doi:10.5281/zenodo.10212972

  50. [50]

    Moghaddam, Amin and Wattenhofer, Roger , date =

    Faber, Lukas and K. Moghaddam, Amin and Wattenhofer, Roger , date =. When. doi:10.1145/3447548.3467283 , isbn =

    arXiv

  51. [51]

    and Athreya, Avanti and Meng, Lingyao and Lyzinski, Vince and Priebe, Carey E

    Fishkind, Donniell E. and Athreya, Avanti and Meng, Lingyao and Lyzinski, Vince and Priebe, Carey E. , date =. On a Complete and Sufficient Statistic for the Correlated

  52. [52]

    and Boisbunon, Aurélie and Chambon, Stanislas and Chapel, Laetitia and Corenflos, Adrien and Fatras, Kilian and Fournier, Nemo and Gautheron, Léo and Gayraud, Nathalie T

    Flamary, Rémi and Courty, Nicolas and Gramfort, Alexandre and Alaya, Mokhtar Z. and Boisbunon, Aurélie and Chambon, Stanislas and Chapel, Laetitia and Corenflos, Adrien and Fatras, Kilian and Fournier, Nemo and Gautheron, Léo and Gayraud, Nathalie T. H. and Janati, Hicham and Rakotomamonjy, Alain and Redko, Ievgen and Rolet, Antoine and Schutz, Antony and...

  53. [53]

    Fournet, Julie and Barrat, Alain , date =. Contact

  54. [54]

    On the Histogram as a Density Estimator:

    Freedman, David and Diaconis, Persi , date =. On the Histogram as a Density Estimator:

  55. [55]

    Hierarchical Graph Learning for Protein–Protein Interaction , author =

  56. [56]

    , date =

    Gao, Chao and Lu, Yu and Zhou, Harrison H. , date =. Rate-. 43818863 , eprinttype =

  57. [57]

    Maximum Likelihood Estimation of Sparse Networks with Missing Observations , author =

  58. [58]

    Optimality of Variational Inference for Stochasticblock Model with Missing Links , author =

  59. [59]

    Glover, Fred and Laguna, Manuel , date =. Tabu. doi:10.1007/978-1-4615-6089-0 , isbn =

    doi:10.1007/978-1-4615-6089-0

  60. [60]

    Bootstrapping Exchangeable Random Graphs , author =

  61. [61]

    The Multiplex Network of Human Diseases , author =

  62. [62]

    Inductive Representation Learning on Large Graphs , author =

  63. [63]

    Han, Xiaotian and Jiang, Zhimeng and Liu, Ninghao and Hu, Xia , date =. G-

  64. [64]

    Hao, Yi and Orlitsky, Alon and Pichapati, Venkatadheeraj , date =. On. Advances in Neural Information Processing Systems , volume =

  65. [65]

    and Laskey, Kathryn Blackmond and Leinhardt, Samuel , date =

    Holland, Paul W. and Laskey, Kathryn Blackmond and Leinhardt, Samuel , date =. Stochastic Blockmodels:

  66. [66]

    Temporal Networks , author =

  67. [67]

    Relations on

    Hoover, Douglas N , date =. Relations on

  68. [68]

    Hu, Wenya and Wu, Jia and Qian, Quan , date =

  69. [69]

    and Pernoud, Agathe , date =

    Jackson, Matthew O. and Pernoud, Agathe , date =. Systemic. doi:10.1146/annurev-economics-083120-111540 , issue =

    doi:10.1146/annurev-economics-083120-111540

  70. [70]

    Revisiting

    Jaggi, Martin , date =. Revisiting

  71. [71]

    Can Smooth Graphons in Several Dimensions Be Represented by Smooth Graphons on [0,1] , author =

  72. [72]

    1009.2376 , eprinttype =

    Graphons, Cut Norm and Distance, Couplings and Rearrangements , author =. 1009.2376 , eprinttype =

    Pith/arXiv arXiv

  73. [73]

    Autoregressive

    Jiang, Binyan and Li, Jialiang and Yao, Qiwei , date =. Autoregressive

  74. [74]

    Foundations of

    Kallenberg, Olav , date =. Foundations of. doi:10.1007/978-3-030-61871-1 , isbn =

    doi:10.1007/978-3-030-61871-1

  75. [75]

    Probabilistic

    Kallenberg, Olav , date =. Probabilistic. doi:10.1007/0-387-28861-9 , isbn =

    work page doi:10.1007/0-387-28861-9

  76. [76]

    On the Representation Theorem for Exchangeable Arrays , author =

  77. [77]

    Karypis, George and Kumar, Vipin , year =

  78. [78]

    Multilevel K-Way

    Karypis, George and Kumar, Vipin , date =. Multilevel K-Way

  79. [79]

    Modelling

    Kimpton, Louise and Challenor, Peter and Wynn, Henry , date =. Modelling. 2211.16921 , eprinttype =

    Pith/arXiv arXiv

  80. [80]

    Multilayer Networks , author =

Showing first 80 references.