pith. sign in

arxiv: 2606.08786 · v1 · pith:YDVDIJ6Dnew · submitted 2026-06-07 · 📊 stat.ME

Inference for Balance in Dynamic Signed Networks

Ergan Shang , Yuan Zhang , Weijing Tang This is my paper

Pith reviewed 2026-06-27 17:43 UTC · model grok-4.3

classification 📊 stat.ME
keywords dynamic signed networksstructural balancegraphon modelkernel smoothingnonparametric inferenceEdgeworth expansionstudentized inferencetemporal smoothing
0
0 comments X

The pith

Kernel smoothing enables inference on structural balance at any time in dynamic signed networks.

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

The paper develops nonparametric inference for the population degree of structural balance in signed networks that evolve over time. It models both edge presence and sign with smoothly varying graphon functions and builds a kernel-smoothed estimator that pools information from nearby observed snapshots. This construction supports studentized confidence intervals whose accuracy is refined by an Edgeworth expansion. The analysis shows that the smoothing step lowers estimator variance in sparse regimes, subject to controlled bias and discretization error.

Core claim

In the dynamic signed graphon model, the kernel-smoothed estimator of the structural balance measure at a target time admits a studentized inference procedure whose finite-sample distribution is approximated to higher order by an Edgeworth expansion; temporal smoothing reduces the variance contribution from observation noise up to smoothing bias and time-discretization error.

What carries the argument

The kernel-smoothed estimator of the structural balance measure under the dynamic signed graphon model, which aggregates signed edges across a local time window.

If this is right

  • Temporal smoothing reduces variance of the balance estimator in sparse networks.
  • The procedure yields valid inference at target times that do not coincide with any observed snapshot.
  • The Edgeworth expansion supplies a higher-order correction to the normal approximation for the studentized statistic.
  • The method applies directly to observed dynamic networks such as international relations data.

Where Pith is reading between the lines

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

  • The same smoothing construction could be used for other time-varying network functionals such as clustering coefficients.
  • When the smoothness assumption fails, the bias term would dominate and the reported variance reduction would disappear.
  • Node-level heterogeneity could be incorporated by replacing the global graphon with a latent-space version while retaining the temporal kernel.

Load-bearing premise

Both edge formation and sign generation follow smoothly time-varying graphon functions.

What would settle it

A simulation study in which the underlying graphon functions change abruptly would produce coverage rates for the studentized intervals that deviate systematically from nominal levels.

Figures

Figures reproduced from arXiv: 2606.08786 by Ergan Shang, Weijing Tang, Yuan Zhang.

Figure 1
Figure 1. Figure 1: Coverage proportion across varying network sizes under Setting (1), where the target inference time point is t ∗ = 0. 50 100 200 400 n 0.88 0.90 0.91 0.93 0.95 0.96 0.98 Coverage Two-Sided 50 100 200 400 n Upper 50 100 200 400 n Lower our_method our_method_norm one_snap one_snap_norm Expected [PITH_FULL_IMAGE:figures/full_fig_p016_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Coverage proportion across varying network sizes under Setting (2), where the target inference time point is t ∗ = −0.1. 4.2 Impact of Network Sparsity To investigate how network sparsity affects the inference validity, we modify Setting (1) such that the probability of edge formation is entirely determined by the sparsity parameter ρn: (3) F(x, y, t) ≡ 1 and G(x, y, t) = 0.4(1 − t 2 )(x 2 + y 2 ). We vary… view at source ↗
Figure 3
Figure 3. Figure 3: Length of CIs across varying network sizes. The left panel corresponds to Set￾ting (1) with t ∗ = 0 and the right panel corresponds to Setting (2) with t ∗ = −0.1. n 0.4 n 0.25 n 0.2 n 0.1 n 0.87 0.89 0.91 0.93 0.94 0.96 0.98 Coverage Two-Sided n 0.4 n 0.25 n 0.2 n 0.1 n Upper n 0.4 n 0.25 n 0.2 n 0.1 n Lower our_method our_method_norm one_snap one_snap_norm Expected [PITH_FULL_IMAGE:figures/full_fig_p017… view at source ↗
Figure 4
Figure 4. Figure 4: Coverage proportion across varying network sparsity levels under Setting (3), where the target inference time point is t ∗ = 0. consistently short and stable CIs across all sparsity regimes, whereas the one-snapshot methods lead to longer CIs as ρn decreases. These results indicate that by incorporating information from neighboring timestamps, our method effectively mitigates the instability that arises in… view at source ↗
Figure 5
Figure 5. Figure 5: Length of CIs (normalized by ρ 3 n ) across varying network sparsity levels under Setting (3), where the target inference time point is t ∗ = 0. methods improves steadily with larger T and converges toward the nominal level. When T is small, coverage proportion may deviate due to bias introduced by kernel smoothing. In contrast, under this sparse regime, the one-snapshot methods deviate from nominal level … view at source ↗
Figure 6
Figure 6. Figure 6: Coverage proportion across varying number of timestamps under Setting (3), where ρn = n −0.4 and t ∗ = 0. 4.4 Impact of the Degree of Balance Finally, we assess the performance of our method under varying degrees of balance using the graphon model defined in Setting (2). In this model, the degree of balance evolves over time. When t = −0.5, each edge is equally likely to be positive or negative, representi… view at source ↗
Figure 7
Figure 7. Figure 7: Portion of balanced triangles among all triangles and 90% two-sided confidence interval constructed by our method under Setting (2), where the portion is defined as (1 + µ(t ∗ )/ρ3 n )/2 and µ(t ∗ ) is the inference target. 0.4 6 0.4 0.3 0.1 0.3 t 0.85 0.87 0.89 0.92 0.94 0.96 0.98 Coverage Two-Sided 0.4 6 0.4 0.3 0.1 0.3 t Upper 0.4 6 0.4 0.3 0.1 0.3 t Lower our_method our_method_norm one_snap one_snap_no… view at source ↗
Figure 8
Figure 8. Figure 8: Coverage proportion across varying degrees of balance under Setting (2). [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Length of CIs across varying degrees of balance under Setting (2). [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The 95% confidence intervals for µ(t ∗ ) at seven inference time points obtained by three methods. The x-axis denotes the year and month of each inferred time point (e.g., 200511 corresponds to November 2005). In addition, the temporal pattern in [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Length comparison of the 95% confidence intervals for µ(t ∗ ) at seven inference time points obtained by different methods. The x-axis denotes the year and month of each inferred time point (e.g., 200511 corresponds to November 2005). thermore, our established approximation error bounds for the Edgeworth expansion of the studentized statistic suggests the trade-off inherent to kernel methods: local averag… view at source ↗
read the original abstract

Signed networks consist of both positive and negative relations, and structural balance theory provides an important conceptural framework for understanding their global tension structure. While existing statistical methods mainly focus on assessing empirical evidence of balance in a single observed network, many real-world signed relations evolve over time. This paper develops nonparametric inference for the population degree of structural balance at specified time points in dynamic signed networks, where the target time may or may not coincide with an observed snapshot. We consider a dynamic signed graphon model in which both edge formation and sign generation are governed by smoothly time-varying graphon functions. To exploit temporal smoothness, we construct a kernel-smoothed estimator that borrows information from snapshots near the target time point. Our theoretical analysis establishes a studentized inference procedure and a higher-order distributional approximation based on Edgeworth expansion, showing that temporal smoothing improves inference in sparse networks by reducing variance of observation noise, up to smoothing bias and time-discretization errors. We demonstrate the finite-sample performance and practical usefulness of the proposed method through extensive simulation studies and an application to a dynamic international relation network in political science.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper develops nonparametric inference for the population degree of structural balance at target times in dynamic signed networks under a dynamic signed graphon model where edge and sign probabilities are governed by smoothly time-varying graphon functions. It constructs a kernel-smoothed estimator that borrows information across nearby snapshots, derives a studentized inference procedure, and provides a higher-order Edgeworth expansion to show that temporal smoothing reduces the variance of observation noise in sparse networks (up to smoothing bias and time-discretization errors). Finite-sample performance is assessed via simulations, and the method is applied to a dynamic international relations network.

Significance. If the theoretical results hold under the stated assumptions, the work offers a useful extension of structural balance analysis to temporal settings with explicit higher-order distributional approximations. The Edgeworth expansion and variance-reduction claim via smoothing in sparse regimes represent a technical contribution that could improve inference for evolving signed networks in applications such as political science.

major comments (2)
  1. [Abstract; model description paragraph] Abstract and model section: the claim that temporal smoothing improves inference by reducing variance (up to bias and discretization errors) rests on the assumption that both the edge-probability and sign-probability graphons are smooth functions of time, yet no explicit regularity index (Hölder, Sobolev, or otherwise) is supplied. Without this, it is impossible to verify that the bias term is o_p of the studentized standard deviation at the sparsity levels implicit in the signed-graphon construction, which is load-bearing for the central variance-reduction guarantee in the Edgeworth expansion.
  2. [Theoretical analysis (Edgeworth expansion)] Theoretical analysis section (on studentized statistic and Edgeworth expansion): the higher-order approximation is asserted to demonstrate improved inference, but the manuscript supplies no explicit verification that the smoothing bias remains negligible relative to the Edgeworth remainder under the sparsity regime. This needs to be shown with concrete rates before the studentized procedure can be guaranteed to inherit the claimed accuracy.
minor comments (1)
  1. [Model section] Notation for the balance measure and the two graphons (edge and sign) should be introduced with consistent symbols early in the model section to avoid ambiguity when stating the target parameter.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful and constructive report. Both major comments correctly highlight the need for explicit regularity conditions and rate verifications to support the bias control underlying the variance-reduction claim. We address each point below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract; model description paragraph] Abstract and model section: the claim that temporal smoothing improves inference by reducing variance (up to bias and discretization errors) rests on the assumption that both the edge-probability and sign-probability graphons are smooth functions of time, yet no explicit regularity index (Hölder, Sobolev, or otherwise) is supplied. Without this, it is impossible to verify that the bias term is o_p of the studentized standard deviation at the sparsity levels implicit in the signed-graphon construction, which is load-bearing for the central variance-reduction guarantee in the Edgeworth expansion.

    Authors: We agree that an explicit regularity index is required to make the bias control rigorous. In the revised manuscript we will add the assumption that the time-varying edge- and sign-probability graphons belong to a Hölder class of smoothness index α > 1/2. We will then state the resulting uniform bias bound of order O(h^α) and verify, under the sparsity regime of the signed-graphon model, that this bias is o_p(σ_n) relative to the studentized scale σ_n, thereby justifying the claimed variance reduction in the Edgeworth expansion. revision: yes

  2. Referee: [Theoretical analysis (Edgeworth expansion)] Theoretical analysis section (on studentized statistic and Edgeworth expansion): the higher-order approximation is asserted to demonstrate improved inference, but the manuscript supplies no explicit verification that the smoothing bias remains negligible relative to the Edgeworth remainder under the sparsity regime. This needs to be shown with concrete rates before the studentized procedure can be guaranteed to inherit the claimed accuracy.

    Authors: The observation is accurate: the current text does not supply the explicit comparison of smoothing bias to the Edgeworth remainder. We will insert a new proposition (or appendix lemma) that derives the concrete rates, showing that for bandwidth sequences satisfying the usual conditions the bias term is of strictly smaller order than the Edgeworth remainder under the model’s sparsity and smoothness assumptions. This will be used to confirm that the studentized statistic inherits the stated higher-order accuracy. revision: yes

Circularity Check

0 steps flagged

No circularity detected; derivation relies on external model assumptions

full rationale

The paper develops a kernel-smoothed estimator and studentized inference procedure with Edgeworth expansion under an externally posited dynamic signed graphon model where edge and sign probabilities are governed by smoothly time-varying functions. No quoted step reduces a claimed prediction or uniqueness result to a fitted quantity, self-citation chain, or definitional tautology within the paper's own equations. The variance-reduction claim is derived from the smoothing bias-variance tradeoff under the stated Hölder-type smoothness and sparsity conditions, which are model inputs rather than outputs of the derivation. The analysis is therefore self-contained against the external graphon framework and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the modeling assumption of a dynamic signed graphon with smooth time variation; no free parameters, additional axioms, or invented entities are specified in the abstract.

axioms (1)
  • domain assumption Edge formation and sign generation are governed by smoothly time-varying graphon functions.
    Explicitly stated as the model considered for the dynamic signed networks.

pith-pipeline@v0.9.1-grok · 5714 in / 1349 out tokens · 24469 ms · 2026-06-27T17:43:00.924461+00:00 · methodology

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

14 extracted references · 4 canonical work pages

  1. [1]

    doi: 10.1214/15-AOS1338. Peter J. Bickel, Aiyou Chen, and Elizaveta Levina. The method of moments and degree distributions for network models.The Annals of Statistics, 39(5):2280 – 2301,

    work page doi:10.1214/15-aos1338

  2. [2]

    PJ Bickel, Friedrich G¨ otze, and WR Van Zwet

    doi: 10.1214/11-AOS904. PJ Bickel, Friedrich G¨ otze, and WR Van Zwet. The edgeworth expansion for u-statistics of degree two.The Annals of Statistics, pages 1463–1484,

    work page doi:10.1214/11-aos904

  3. [3]

    Applications of structural balance in signed social networks.arXiv preprint arXiv:1402.6865,

    J´ erˆ ome Kunegis. Applications of structural balance in signed social networks.arXiv preprint arXiv:1402.6865,

    Pith/arXiv arXiv

  4. [4]

    Catherine Matias and Vincent Miele

    doi: 10.1214/EJP.v12-430. Catherine Matias and Vincent Miele. Statistical clustering of temporal networks through a dynamic stochastic block model.Journal of the Royal Statistical Society Series B: Statistical Methodology, 79(4):1119–1141,

    work page doi:10.1214/ejp.v12-430

  5. [5]

    Bernstein-like concentration and moment inequal- ities for polynomials of independent random variables: multilinear case.arXiv preprint arXiv:1109.5193,

    Warren Schudy and Maxim Sviridenko. Bernstein-like concentration and moment inequal- ities for polynomials of independent random variables: multilinear case.arXiv preprint arXiv:1109.5193,

    Pith/arXiv arXiv

  6. [6]

    Journal of the American Statistical Association , author =

    doi: 10.1080/01621459.2025.2520459. Jiliang Tang, Yi Chang, Charu Aggarwal, and Huan Liu. A survey of signed network mining in social media.ACM Computing Surveys (CSUR), 49(3):1–37,

    work page doi:10.1080/01621459.2025.2520459 2025

  7. [7]

    SinceY 1(t) = 3n−1P i g1h(Xi, t) is a sum of iid mean-zero terms, var(Y1(t)) = 9 n var(g1h(X1, t)) = 9ξ2 1h(t) n

    +O(ρ 6 nn−2). SinceY 1(t) = 3n−1P i g1h(Xi, t) is a sum of iid mean-zero terms, var(Y1(t)) = 9 n var(g1h(X1, t)) = 9ξ2 1h(t) n . Lemma B.3 gives the lower boundξ 2 1h(t)≳ρ 6 n, sovar(Y 1(t))≳ρ 6 nn−1. Moreover, by Lemma B.2, it follows that var( ˜Unh(t)−U nh(t)) =O n−2(T h)−1ρ5 n +n −3(T h)−2ρ4 n +n −3(T h)−3ρ3 n . Finally, it is sufficient to compare the...

    2007

  8. [8]

    By Lemma D.2,S K(t)≍Tand PT ℓ=1 K2 h(t−t ℓ) = O(T /h)

    B.2.1 Proof of Lemma B.4 Proof.LetS K(t) = PT ℓ=1 Kh(t−t ℓ). By Lemma D.2,S K(t)≍Tand PT ℓ=1 K2 h(t−t ℓ) = O(T /h). Conditional onX, the variables{η ij(tℓ) : 1≤i < j≤n,1≤ℓ≤T}are independent, centered, and bounded. By Lemma 7.1 in Schudy and Sviridenko (2011), each ˜ηij is central moment bounded because of its boundedness. Then, Lemma D.4 could be applied ...

    2011

  9. [9]

    The factorn −2 inµ 1q comes from the normalization n 3 −1 and theO(n) possible third vertices containing a fixed noise variable

    Using| ˜Wij(t)| ≤Cρ n andE{|η ij(tℓ)| |X}= O(ρn), its hypergraph parameters satisfy µ1q =O{n −2ρ2 n(T h)−1}, µ 2q =O{n −3ρn(T h)−2}. The factorn −2 inµ 1q comes from the normalization n 3 −1 and theO(n) possible third vertices containing a fixed noise variable. By the variance bound forR q(t) and Lemma D.4, with probability at least 1−O(n −2), |Rq(t)| ≤C ...

    2022

  10. [10]

    eisJn exp ( − s2 2nρn E[σ2 X] + 1 n nX i=1 gσ;1(Xi) !)# ds =O (nρnT h)−1 logn+n −1 logn .(S8) 32 Inference on Dynamic Signed Network It remains to bound Ψn(s) :=E

    = E[g3 1h(X1, t)]√n ξ3 1h(t) + ˜Op,1(n−1 log1/2 n). Thus, the first part contributes a deterministic shift, while the off-diagonal part is a sym- metric degree-two term. For the second summand in ˇδn(t), Hoeffding decomposition gives 4 n(n−1) X i̸=j gigij ξ2 = 4 nξ2 nX i=1 ζh(Xi, t) + ˜Op,1(n−1 logn), where the first Hoeffding projection is defined as ζh(...

    2001

  11. [11]

    The last step usesϵ ′ =ϵ/2≤1/7 and Assumption 3.4

    forz=o(1), we obtain Eeis(Jn+ ˜∆n+δT ) =E[e isJn]− s2 2nρn E[eisJnσ2 X]− 1 2 cδn−1 logn s 2E[eisJn] +R G(s),(S14) where Z nϵ′ 0 RG(s) s ds≤C Z nϵ′ 0 s3E " σ2 X nρn +c δn−1 logn 2# ds =O n n4ϵ′ (nρnT h)−2 +n 4ϵ′−2 log2 n o =O{(nρ nT h)−1 +n −1 logn}. The last step usesϵ ′ =ϵ/2≤1/7 and Assumption 3.4. Next, we analyze the leading termE[e isJn] and the remai...

    1972

  12. [12]

    Similarly, we can derive E[eisL(1) n ·isL (3) n ] 38 Inference on Dynamic Signed Network =E  eisL(1) n is n3/2ξ3 1h(t) X 1≤i̸=j≤n h g1h(Xi, t){g2 1h(Xj, t)−ξ 2 1h(t)} i   = is(n−1)√nξ3 1h(t) φn−2 n (s)E exp is g1h(X1, t) +g 1h(X2, t)√nξ1h(t) h g1h(X1, t){g2 1h(X2, t)−ξ 2 1h(t)} i = is(n−1)√nξ3 1h(t) φn−2 n (s))E " g1h(X1, t){g2 1h(X2, t)−ξ 2 1h(t)} +i...

    2007

  13. [13]

    By combining the above result with (S24) and (S25), we obtain: Z nϵ′ 0 E[eis(Jn+ ˜∆n+δT )]−Ch.f.(G nh;s) s ds=O(n −1 logn+ (nρ n)−1(T h)−1)

    Finally, we also have the following bound Z nϵ′ 0 E eisJn s2 2 cδn−1 logn ·s −1ds=n −1 logn·O Z nϵ′ 0 |sE[eisJn]|ds ! =O(n −1 logn), since the leading term inE[exp(isJ n)] is of the forme −t2/2P ≥0(s). By combining the above result with (S24) and (S25), we obtain: Z nϵ′ 0 E[eis(Jn+ ˜∆n+δT )]−Ch.f.(G nh;s) s ds=O(n −1 logn+ (nρ n)−1(T h)−1). Appendix D. Te...

    1992

  14. [14]

    For the second term, sincet∈[δ,L −δ], both lower and upper limits satisfyt/h≥δ/h and (L −t)/h≥δ/h

    Thus, by Lemma D.1 and Assumption 3.1, 1 h Z L 0 Kj u−t h du− 1 h X ℓ Kj tℓ −t h ∆tℓ ≤ 1 h ·T V(g; [0,L])· C T =O( 1 hT ). For the second term, sincet∈[δ,L −δ], both lower and upper limits satisfyt/h≥δ/h and (L −t)/h≥δ/h. By (K3) and|K j(u)| ≤M j−1|K(u)|, we have Z L−t h − t h Kj(u)du− Z ∞ −∞ Kj(u)du ≤ Z ∞ L−t h Kj(u) du+ Z − t h −∞ Kj(u) du ≤M j−1 Z ∞ L−...

    2011