pith. sign in

arxiv: 2410.09810 · v3 · submitted 2024-10-13 · 📊 stat.ME

Doubly unfolded adjacency spectral embedding of dynamic multiplex graphs

Maximilian Baum , Francesco Sanna Passino , Axel Gandy This is my paper

Pith reviewed 2026-05-23 19:18 UTC · model grok-4.3

classification 📊 stat.ME
keywords dynamic networksmultiplex graphsspectral embeddinglatent position modelconsistencyasymptotic normalitygraph clusteringchangepoint detection
0
0 comments X

The pith

DUASE estimates are consistent and asymptotically normal for parameters of the dynamic multiplex random dot product graph model.

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

The paper introduces the dynamic multiplex random dot product graph model in which edge probabilities arise from inner products of separate layer-specific and time-specific latent position vectors for each node. It then defines doubly unfolded adjacency spectral embedding as a method that unfolds the observed adjacency array along both the layer and time modes before performing a single spectral decomposition to recover the latent positions. The authors prove that these estimates converge to the true positions and are asymptotically normal as the number of nodes grows. A reader would care because the separation of time and layer effects keeps the representation low-dimensional while still allowing the model to track evolving multi-type relations, and the embedding supports direct downstream tasks such as clustering and changepoint detection.

Core claim

Under the DMPRDPG, the doubly unfolded adjacency spectral embedding recovers node representations that live in two distinct latent spaces—one indexed by time and one indexed by layer—such that the estimates are consistent for the true latent positions and asymptotically normal; the same embedding can be fed directly into ISOMAP to recover temporal trends or into standard clustering algorithms to group nodes.

What carries the argument

Doubly unfolded adjacency spectral embedding (DUASE), which performs spectral decomposition on the adjacency tensor after unfolding it first along the time dimension and then along the layer dimension to obtain separate time-specific and layer-specific latent spaces.

If this is right

  • The recovered embeddings can be used with the ISOMAP algorithm to identify global changepoints and long-term trends in the network.
  • The same embeddings support standard graph clustering procedures that group nodes according to their joint time-and-layer behavior.
  • Because time and layer effects occupy separate low-dimensional spaces, the model can represent complex dynamic multiplex structure without a rapid increase in total dimension.
  • The asymptotic normality result supplies a theoretical basis for using the embeddings in subsequent statistical inference tasks on the observed networks.

Where Pith is reading between the lines

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

  • The separation of time and layer spaces may make it easier to test whether a detected change affects all layers equally or only selected relation types.
  • The same unfolding technique could be applied to other inner-product models that include additional discrete factors beyond time and layer.
  • Because the method is computationally linear in the total number of edges across layers and times, it could scale to networks with hundreds of layers or time points where joint tensor decompositions become prohibitive.
  • Real-data applications suggest the embeddings preserve enough signal to distinguish geopolitical blocs or financial sectors even when the underlying graphs are sparse.

Load-bearing premise

The observed networks must be generated exactly from the DMPRDPG model in which every edge probability equals an inner product of a time-specific vector and a layer-specific vector.

What would settle it

Generate networks from the DMPRDPG with known latent positions, compute DUASE as the number of nodes tends to infinity, and check whether the estimates fail to converge in probability to the true positions or whether their scaled fluctuations fail to match a normal distribution.

Figures

Figures reproduced from arXiv: 2410.09810 by Axel Gandy, Francesco Sanna Passino, Maximilian Baum.

Figure 1
Figure 1. Figure 1: Scatterplots of the first two dimensions of the left and right DUASE under a simulated DMPSBM with n = 1000, G = 4 groups of equal size, K = 3 and T = 3. Note that only the first two dimensions of the five-dimensional embedding are displayed. error, are depicted in [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Scale of error in latent position recovery for increasing graph sizes. (a) Iso-mirror for T = 3 time indices 1 2 3 Time −0.10 −0.05 0.00 0.05 0.10 0.15 0.20 Iso-mirror (b) Iso-mirror for K = 3 layers 1 2 3 Layer −0.05 0.00 0.05 0.10 0.15 Iso-mirror [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Iso-mirrors calculated from DUASE on the DMPSBM in [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Iso-mirror across time and event types on the POLECAT data. further confirm this, Figure 4d displays a scatterplot between the average PLOVER intensity, calculated from the POLECAT Data Dictionary v5.8, and the iso-mirror scores, confirming a similar structure to Figure 4c. 6.2 Financial news In a second example, we apply the DUASE algorithm to the FinDKG dataset Li and Sanna Passino (2024), which contains… view at source ↗
Figure 5
Figure 5. Figure 5: Iso-mirror across time and event types on the FinDKG data. returns a left and right embedding suitable for running the iso-mirror routine simultaneously, which is com￾putationally convenient compared to the two separate UASE procedures needed to obtain an iso-mirror across time, and an iso-mirror across layers. This is a relevant advantage of DUASE over alternative embedding methods. The results are presen… view at source ↗
read the original abstract

Many real-world networks evolve dynamically over time and present different types of connections between nodes, often called layers. In this work, we propose a latent position model for these objects, called the dynamic multiplex random dot product graph (DMPRDPG), which uses an inner product between layer-specific and time-specific latent representations of the nodes to obtain edge probabilities. We further introduce a computationally efficient spectral embedding method for estimation of DMPRDPG parameters, called doubly unfolded adjacency spectral embedding (DUASE). The DUASE estimates are proved to be both consistent and asymptotically normally distributed. A key strength of our method is the encoding of time-specific node representations and layer-specific effects in separate latent spaces, which allows the model to capture complex behaviors while maintaining relatively low dimensionality. The embedding method we propose can also be efficiently used for subsequent inference tasks. In particular, we highlight the use of the ISOMAP algorithm in conjunction with DUASE as a way to efficiently capture trends and global changepoints within a network, and the use of DUASE for graph clustering. Applications on real-world networks describing geopolitical interactions between countries and financial news reporting demonstrate practical uses of our method.

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

0 major / 3 minor

Summary. The paper introduces the dynamic multiplex random dot product graph (DMPRDPG) model in which edge probabilities arise from inner products of layer-specific and time-specific latent positions. It proposes doubly unfolded adjacency spectral embedding (DUASE) as a spectral method to estimate these parameters and states that the resulting estimates are consistent and asymptotically normal. The method is shown to support downstream tasks including ISOMAP-based changepoint detection and clustering, with illustrations on geopolitical and financial networks.

Significance. If the stated consistency and asymptotic normality results hold under the DMPRDPG generative model, the work supplies a theoretically supported, computationally efficient embedding procedure that maintains separate latent spaces for temporal and layer effects. This separation permits relatively low-dimensional representations while accommodating complex dynamic multiplex structure. Explicit proofs of consistency and asymptotic normality constitute a clear strength, as does the demonstration that DUASE embeddings can be directly fed into standard algorithms such as ISOMAP for trend and changepoint recovery.

minor comments (3)
  1. [Abstract] The abstract asserts consistency and asymptotic normality but provides no indication of the regularity conditions (e.g., eigenvalue gaps, sparsity regime, or growth rates of the number of layers and time points) required for the results; a one-sentence summary of the main assumptions would improve accessibility.
  2. [Section 3] Notation for the unfolded adjacency tensor and the precise definition of the double unfolding operation should be introduced with an explicit display equation early in the methods section to avoid ambiguity when the embedding is later compared with single-unfolding alternatives.
  3. [Section 5] The real-data examples would benefit from a brief statement of the chosen embedding dimensions and the criterion used to select them, as this choice directly affects the reported changepoint and clustering performance.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive assessment of the manuscript, accurate summary of the DMPRDPG model and DUASE procedure, and recommendation for minor revision. We are pleased that the theoretical guarantees and downstream applications were viewed favorably.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper defines the DMPRDPG generative model (inner-product edge probabilities from layer- and time-specific latent positions) and introduces the DUASE spectral embedding procedure. It then states that consistency and asymptotic normality of the DUASE estimates are proved under this model. No equations or steps in the provided text reduce a claimed prediction or theorem to a fitted parameter, self-citation chain, or definitional equivalence. The central result is a model-specific theoretical guarantee derived from standard concentration arguments rather than by construction from the inputs themselves.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract only; no information available on free parameters, background axioms, or invented entities beyond the high-level model description.

pith-pipeline@v0.9.0 · 5734 in / 1000 out tokens · 36183 ms · 2026-05-23T19:18:01.447207+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Multiscale Euclidean Network Trajectories: Second-Moment Geometry, Attribution, and Change Points

    stat.ML 2026-05 unverdicted novelty 7.0

    MENT imposes isotropic normalization on latent positions to reduce ambiguity to orthogonal transformations, defines trace and mode-wise variation distances, applies MDS for low-dimensional trajectories, and proves con...

  2. Multiscale Euclidean Network Trajectories: Second-Moment Geometry, Attribution, and Change Points

    stat.ML 2026-05 unverdicted novelty 6.0

    MENT imposes isotropic normalization on anchor latent positions in unfolded spectral embeddings to preserve second-moment geometry under orthogonal transformations, yielding consistent multiscale trajectories for dyna...

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages · cited by 1 Pith paper

  1. [1]

    (a) ∥UAU⊺ A − UPU⊺ P∥ = OP n T −1/2 n ρ−1/2 n n−1/2 log1/2(n) o , (b) ∥VAV⊺ A − VPV⊺ P∥ = OP n T −1/2 n ρ−1/2 n n−1/2 log1/2(n) o

    work page

  2. [2]

    (a) ∥UA − UPU⊺ PUA∥F = OP n T −1/2 n ρ−1/2 n n−1/2 log1/2(n) o , (b) ∥VA − VPV⊺ PVA∥F = OP n T −1/2 n ρ−1/2 n n−1/2 log1/2(n) o

    work page

  3. [3]

    (a) ∥U⊺ PUADA − DPV⊺ PVA∥F = OP n K1/2 n T −1/2 n log(n) o , (b) ∥DPU⊺ PUA − V⊺ PVADA∥F = OP n K1/2 n T −1/2 n log(n) o

    work page

  4. [4]

    ∥U⊺ PUA − V⊺ PVA∥F = OP T −1 n ρ−1 n n−1 log(n) . Proof. This proof follows the steps outlined in Proposition 13 in Jones and Rubin-Delanchy (2021) and is divided into four parts, corresponding to the four statements in the result. Without loss of generality, we assume max(Kn, Tn) = Kn, for notational simplicity

    work page 2021

  5. [5]

    σd to be the singular values of the matrixU⊺ PUA and θℓ = cos−1(σℓ) to be the principal angles

    Define σ1, . . . σd to be the singular values of the matrixU⊺ PUA and θℓ = cos−1(σℓ) to be the principal angles. From Lemma 2.4 in Chen et al. (2021), we know that the non-zero eigenvalues of UAU⊺ A − UPU⊺ P are equal to sin(θℓ). By invoking a variant of the Davis-Kahan theorem (see Yu et al., 2014; Jones and Rubin-Delanchy, 2021, Theorem 12) we find that...

    work page 2021

  6. [6]

    Again, the same argument can be used to show that ∥VA − VPV⊺ AVA∥F

    Using the rate derived above in Part 1, and the fact the fact that UA is a truncated unitary matrix, and hence multiplication by UA does not increase the norm, we see that ∥UA − UPU⊺ PUA∥F = ∥(UAU⊺ A − UPU⊺ P)UA∥F ≤ ∥UAU⊺ A − UPU⊺ P∥F = OP ( log1/2(n) T 1/2 n ρ1/2 n n1/2 ) . Again, the same argument can be used to show that ∥VA − VPV⊺ AVA∥F

    work page

  7. [7]

    Algebraic manipulation shows that U⊺ PUADA − DPV⊺ PVA = U⊺ P(A − P)VA = U⊺ P(A − P)(VA − VPV⊺ PVA) + U⊺ P(A − P)VPV⊺ PVA. Analyzing these terms separately and applying the bound from Part 2, combined with the results in Propositions 4 and 7, we find that ∥U⊺ P(A − P)(VA − VPV⊺ PVA)∥F = OP n K1/2 n T −1/2 n log(n) o , and ∥U⊺ P(A − P)VPV⊺ PVA∥F = OP n log1...

    work page

  8. [8]

    Therefore, the following identity holds: U⊺ PUA − V⊺ PVA + DP(U⊺ PUA − V⊺ PVA)D−1 A = [(U⊺ PUADA − DPV⊺ PVA) + (DPU⊺ PUA − V⊺ PVADA)]D−1 A

    Via simple algebraic manipulation, we can write: U⊺ PUA − V⊺ PVA = [(U⊺ PUADA − DPV⊺ PVA) + (DPU⊺ PUA − V⊺ PVADA)]D−1 A − DP(U⊺ PUA − V⊺ PVA)D−1 A . Therefore, the following identity holds: U⊺ PUA − V⊺ PVA + DP(U⊺ PUA − V⊺ PVA)D−1 A = [(U⊺ PUADA − DPV⊺ PVA) + (DPU⊺ PUA − V⊺ PVADA)]D−1 A . (10) From the definition of DA and DP, the absolute value of the (ℓ...

    work page

  9. [9]

    Then: max ∥U⊺ PUA − W∥F , ∥V⊺ PVA − W∥F = OP log(n) Tnρnn . Proof. In this proof we follow the logic of Jones and Rubin-Delanchy (2021), Proposition 14. From Schöne- mann (1966) we have that: W = W1W⊺ 2 = min Q∈O(d) ||U⊺ PUA − Q||2 F + ||V⊺ PVA − Q||2 F . (11) Next, denote the SVD of U⊺ PUA by U⊺ PUA = WU,1DUW⊺ U,2 and define the d × d orthogonal matrix W...

    work page 2021

  10. [10]

    ∥WDA − DPW∥F = OP n K1/2 n T −1/2 n log(n) o ,

    work page

  11. [11]

    ∥WD1/2 A − D1/2 P W∥F = OP n ρ−1/2 n n−1/2T −3/4 n K1/4 n log(n) o ,

    work page

  12. [12]

    ∥WD−1/2 A − D−1/2 P W∥F = OP n K−1/4 n T −5/4 n ρ−3/2 n n−3/2 log(n) o . Proof. The proof is divided in three parts, corresponding to the three rates stated in the proposition. It is a modified version of Jones and Rubin-Delanchy (2021), Proposition 15, which is based on Lyzinski et al. (2016), Lemma 17

    work page 2021

  13. [13]

    Via algebraic manipulation, we get: WDA − DPW = (W − U⊺ PUA)DA + U⊺ PUADA − DPW = (W − U⊺ PUA)DA + (U⊺ PUADA − DPV⊺ PVA) + DP(V⊺ PVA − W). Applying the Frobenius norm, using the triangle inequality, and applying the rates from Propositions 3, 8 and 9 on each component of the right-hand side of the summation above, we find that ||(W − U⊺ PUA)DA + (U⊺ PUADA...

    work page

  14. [14]

    From the definition of DA and DP, we can write the (ℓ, h)-th entry of WD1/2 A − D1/2 P W as: h WD1/2 A − D1/2 P W i ℓ,h = Wℓ,h h σh(A)1/2 − σℓ(P)1/2 i = Wℓ,h[σh(A) − σℓ(P)] σh(A)1/2 − σℓ(P)1/2 = [WDA − DPW]ℓ,h σh(A)1/2 − σℓ(P)1/2 , for ℓ, h ∈ [d]. By taking the Frobenius norm of the right-hand side and applying the result from Part 1 as well as Propositio...

    work page

  15. [15]

    Using a similar approach as the previous part, we get: h WD−1/2 A − D−1/2 P W i ℓ,h = Wℓ,h[σℓ(P)1/2 − σh(A)1/2] σℓ(P)1/2σh(A)1/2 = [WD1/2 A − D1/2 P W]ℓ,h σℓ(P)1/2σh(A)1/2 . Once again, taking the Frobenius norm of the right-hand side and applying the result from Part 2 combined with Proposition 3, we get: ∥WD−1/2 A − D−1/2 P W∥F = OP ( log(n) K1/4 n T 5/...

    work page 2021

  16. [16]

    Applying the relation∥AB∥2→∞ ≤ ∥A∥2→∞ ∥B∥ yields ∥UP∥2→∞ ≤ ∥X∥2→∞ ∥˜L∥ ∥D−1/2 P ]∥

    Recall that UPD1/2 P = X˜L. Applying the relation∥AB∥2→∞ ≤ ∥A∥2→∞ ∥B∥ yields ∥UP∥2→∞ ≤ ∥X∥2→∞ ∥˜L∥ ∥D−1/2 P ]∥. By applying Propositions 3 and 12 and using the fact that the rows of X are by definition OP(ρ1/2 n ), we find that ∥UP∥2→∞ = OP(K−1/2 n n−1/2). Hence: ∥R1,1∥2→∞ ≤ ∥UP∥2→∞ ∥U⊺ PUAD1/2 A − D1/2 P W∥ ≤ ∥UP∥2→∞ h ∥(U⊺ PUA − W)D1/2 A ∥F + ∥WD1/2 A −...

    work page

  17. [17]

    Define M1 = (UPU⊺ P)(A − P)(VA − VPW)D−1/2 A and M2 = (A − P)(VA − VPW)D−1/2 A . M. Baum, F. Sanna Passino, and A. Gandy 30 Doubly unfolded adjacency spectral embedding of dynamic multiplex graphs Hence, R1,2 = M2 − M1, which implies that ∥R1,2∥2→∞ ≤ ∥M2∥2→∞ + ∥M1∥2→∞. Therefore, we bound these terms individually. For M1: ∥M1∥2→∞ ≤ ∥UP∥2→∞ ∥A − P∥ ∥VA − V...

    work page

  18. [18]

    |MKn], and claim that the Frobenius norms of each of the rows of each Mj are exchangeable

    Next, we divide M into Kn equally sized chunks, written M = [M1| . . . |MKn], and claim that the Frobenius norms of each of the rows of each Mj are exchangeable. Define Mj i to be the i-th row of Mj, and recall that ∥Mj∥F ≤ ∥ M∥F = OP{K1/2 n T −1/2 n log(n)}. By treating each ∥M j i ∥2 F as a random variable and making use of Lemma 1, we therefore have th...

    work page 2010

  19. [19]

    Using the rate for ∥UP∥2→∞ from Part 1, as well as Propositions 6 and 7, we see that ∥R1,3∥2→∞ ≤ ∥UP∥2→∞ ∥ − UPU⊺ P(A − P)VPWD−1/2 A ∥ ≤ ∥UP∥2→∞ ∥U⊺ P(A − P)VP∥F ∥WD−1/2 A ∥F = OP ( log1/2(n) ρ1/2 n nK3/4 n T 1/4 n )

    work page

  20. [20]

    B Proof of Theorem 1 Proof

    By Propositions 4 and 10, we get: ∥R1,4∥2→∞ ≤ ∥R1,4∥F ≤ ∥A−P∥ ∥VP ∥F ∥WD−1/2 A −D−1/2 P W∥F = OP ( log3/2(n)K1/4 n ρnnT 5/4 n ) . B Proof of Theorem 1 Proof. This proof is an adaptation of the proof of Jones and Rubin-Delanchy (2021), Theorem 2. For the left embedding, we write ˆX − XPW = UAD1/2 A − UPD1/2 P W = UAD1/2 A − UPU⊺ PUAD1/2 A + UP(U⊺ PUAD1/2 A...

    work page arXiv 2021