acceptodds
Under review as a conference paper at ICLR 2027

Inference for sender-receiver alignment in directed networks

Abstract

In a directed network, does understanding a node's sending behaviour give us any information about its receiving behaviour? To address this question, we propose the effective shared dimension, a real-valued count of how many directions the two roles share, alongside a degree-free counterpart which removes the effect of a node's overall activity level. We define both for general latent position models, and demonstrate that under a directed random dot product graph model they are explicit functions of moments of the underlying node positions. For each quantity we give an estimator, for which we prove consistency and, provided the average node degree grows faster than , asymptotic normality, before discussing a bias correction to potentially extend the latter to sparser regimes. Finally, we apply this methodology to estimate the effective shared dimension across 10 real directed networks spanning social, information, infrastructure and biological data, and ask whether the answer differs systematically between these domains.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Directed graphs, latent position models, principal angles, statistical inference

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_hXVfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose the effective shared dimension, the sum of squared cosines of the principal angles between the sending and receiving subspaces of a latent position model. It reflects a real-valued count of how many directions the two roles share. A a proposed degree-free variant controls for node activity. Under a directed random dot product graph, the quantities have closed form in second moments of the latent positions. The authors show that the specttral estimator are consistent and asymptotically normal at rate \sqrt(n) when the average degree grows faster than n^{1/2}. A second-order bias correction is derived, which the authors claim potentially extend to sparser networks. The experiments evaluate the methodology in real networks from several domains and study the behaviors of the model across them.

I recommend rejection due to 1) The title promises inference, but the paper doesn't provide a way to perform it, and both the sharpness of the n^{1/2} threshold and the bias correction rest on simulation only (see below). 2) The experiments show comparisons without calibration, outside the theory. I also find the contribution delta small.

Critical strengths
  1. Clean, well-defined quantity that answers a natural question. It is invariant to the non-identifiability of the latent positions, and has closed-form formulas. The degree-free version handles its own floor of 1. The correct characterizations (zero alignment iff each node has at most one active role). As I commented below, closing the inference gap would make the paper useful.

  2. Correct assymptotic. Consistency and CLTs hold for both estimators, and the simulations show the n^{-1/4} bias decay that the bias expansion predicts. The explicit second order bias is useful.

Critical weaknesses
  1. Claim validity:
  • The standardization in Theorems 4 and 5 depend on functionals of the unkown latent distribution, but the paper does not estimates them. Tables 1 and 2 normalize by the true \tau_F based on the closed form of the simulated co-blockmodel in B.2. So, every coverage figure is oracle coverage, and even with the true variance the nominal 95% interval covers only 71%, 84% and 90% at n = 500, 2,500 and 10,000, inside the regime where the theorem applies (Table 1).

  • the CLT is degenerate at the two natural cases, \Psi_F = 0 (fully separated roles) and \Psi_F = d (fully aligned roles, y = Gx), where \tau_F^2 = 0. So, the paper doesn't offers a test for fully aligned or fully separated roles. OpenFlights falls within this boundary, with \hat\Psi/d = 1.00 and a corrected estimate of 1.04, above the maximum of 1.

  • the paper doesn't study coverage for the degree-free estimator. The Intro's line "we demonstrate that this threshold is sharp" should be interpreted as an empirical observation, since the paper doens't shows a nonvanishing bias at n\rho_n \asymp n^{1/2}. The bias correction in Section 6 is also unproved. No result covers the corrected estimator, its remainder of order \sqrt{n}/(n\rho_n)^2 and the error Q(\hat P) - Q(P) are not proved, and the degree-free estimator has no correction, although Table 3 reports it next to the corrected \tilde\Psi. So claims that come after Theorems 4 and 5 rest on simulated co-blockmodels only.

  1. Experiments outside the theory:
  • Section 7 applies the estimators to networks that likely lie outside the proved resutl, and its main comparison is uncalibrated. The average degree is not reported for any network. HEP-Ph, for example, has n = 11,962, so the proved redsult calls for an average degree well above \sqrt{n} \approx 109, but the full citation graph has 421,578 edges, which caps the average degree of the retained subgraph at about 35.

  • below the n^{1/2} threshold, Table 2 puts the uncorrected estimate 8 to 44 standard deviations off, and the correction has no theorem. Reciprocity up to 0.98 put a strain on the assumption that the two edge directions are independent, as the paper itself notes for OpenFlights. The (1,1)-core filter removes nodes that only send or only receive, which are the nodes that pull \Psi_F toward zero, and the retained fractions are not reported (wiki-Vote keeps 1,300 of its 7,115 nodes). Finally, the claim that each network shares more than its reciprocity alone would suggest compares quantities on different scales, since under the model reciprocity scales with \rho_n while \Psi_F does not. This matters because the cross-domain conclusions cannot be checked against the theory, so I read Section 7 as descriptive only.

  1. Novelty:
  • the paper credits the index to Krzanowski (1979), but does not note that its closed form under the DRDPG (Proposition 3), \Psi_F = tr(\Delta_x^{-1}\Delta_{xy}\Delta_y^{-1}\Delta_{yx}), is the sum of squared uncentered canonical correlations between sending and receiving positions, a Pillai trace (CCA).

  • because \tau_F^2 has no network term, the CLT reduces to the delta method on n latent pairs plus a bias bound from known expansions (Xia, 2021, through the Hermitian dilation, and Lei and Rinaldo, 2015, for concentration). The degree-free theorem does not follow from these and involves real additional work, which the paper supplies. Section 6 should also be positioned against Xia (2021), which proves normal approximation with explicit bias corrections for singular subspaces, whereas here the corrected estimator's behaviour is only conjectured.

  • A suggestion, (and not part of my evaluation): Liu, Xu and Zhang (February, 2026), "Unfolding and Fusion: Debiased Inference for Generalized Multilayer Latent Space Models", derive a second-order bias from fusing two separately estimated components in a directed sender-receiver latent space model, and prove asymptotic normality for a plug-in debiased estimator. Could you contrast Section 6 against it?

Questions for authors
  1. Could you provide consistent estimators of \tau_F^2 and \tau_F^{\circ 2}, coverage in Tables 1 and 2 using the estimated variance, and intervals in Table 3?

  2. For each network in Table 3, what is the average degree relative to n^{1/2}, what fraction of nodes does the core filter keep, and how do the estimates change with d - 1 and d + 1?

  3. Is the n^{1/2} threshold proved sharp, or only observed?

  4. How can a user test fully aligned and fully separated roles, given that \tau_F^2 = 0 in these cases?

2 comments

Thank you for your feedback. In response to your points:

  • As in point 2 of our general response, we propose a variance estimate for the effective shared dimension with an associated consistency result, with tables being adjusted to incorporate this. Given our proposed restructuring, we cannot promise to have an equivalent result for the degree-free measure, but will remove it from Table 3, leaving it as an extension to the original measure and noting that the theory is not yet fully developed.

  • As mentioned in the general response, we will no longer restrict to the (1,1) core for the networks, and instead consider the largest weakly-connected component of each network (which removes fewer than 1%1\% of nodes in each case). The original restriction was for the benefit of the degree-free estimate, which we will now no longer mention. Table 3 then becomes:

networknnremovedmean degreen1/2n^{1/2}ddΨ^/d\widehat\Psi/d [95% interval]Ψ~/d\widetilde\Psi/d [95% interval]
email-Eu-core986025.331.470.84 [0.82, 0.86]0.91 [0.89, 0.93]
CollegeMsg1,893610.743.520.71 [0.67, 0.74]0.79 [0.75, 0.82]
Epinions75,87726.7275.570.47 [0.45, 0.49]0.53 [0.51, 0.54]
wiki-Vote7,0664914.784.150.13 [0.11, 0.15]0.14 [0.12, 0.16]
political blogs1,222215.635.020.34 [0.29, 0.39]0.36 [0.31, 0.41]
HEP-Ph citations34,40114512.3185.5110.15 [0.13, 0.16]0.17 [0.15, 0.18]
OpenFlights routes3,3972811.158.341.00 [1.00, 1.00]1.04 [1.04, 1.04]
US faculty hiring3,284018.857.330.46 [0.44, 0.49]0.49 [0.46, 0.52]
C. elegans connectome446010.921.190.40 [0.36, 0.43]0.46 [0.42, 0.49]
Florida Bay food web128016.711.350.30 [0.24, 0.35]0.31 [0.26, 0.36]

In all cases but the Florida Bay food web, the average degree falls below the threshold of our theoretical results, so we will make this clear.

In regards to the sensitivity of our results to dimension, the following table shows the estimates computed for d−1d-1, dd and d+1d+1. In most cases the estimate is fairly stable, but CollegeMsg shows a large difference between dimensions.

networkddΨ^d−1/(d−1)\widehat\Psi_{d-1}/(d-1)Ψ^d/d\widehat\Psi_d/dΨ^d+1/(d+1)\widehat\Psi_{d+1}/(d+1)Ψ~d−1/(d−1)\widetilde\Psi_{d-1}/(d-1)Ψ~d/d\widetilde\Psi_d/dΨ~d+1/(d+1)\widetilde\Psi_{d+1}/(d+1)
email-Eu-core70.820.840.840.880.910.91
CollegeMsg20.560.710.640.620.790.75
Epinions70.490.470.480.530.530.54
wiki-Vote50.150.130.120.150.140.13
political blogs20.370.340.330.380.360.35
HEP-Ph citations110.150.150.150.170.170.17
OpenFlights routes41.001.001.001.031.041.05
US faculty hiring30.480.460.440.500.490.48
C. elegans connectome90.410.400.390.470.460.45
Florida Bay food web50.280.300.270.290.310.28
  • Yes, we now have a result showing that the threshold is sharp, and will add this to Theorem 4.

  • Our theory currently doesn't cover the boundary cases, as whenever τF=0\tau_F = 0 we can only say that n1/2(Ψ^−ΨF)→0n^{1/2}(\widehat\Psi-\Psi_F)\to 0. We can check full separation directly simply by determining whether every node either only sends or only receives, but we do not currently have the tools to test for full alignment, which we will state as a limitation.

On a couple of other points you raised:

  • We agree that we should mention Xia (2021) in the main paper, rather than just the Appendix as we do currently. Our work is not a direct translation of his results however, as he considers the distance between an estimated singular subspace and the true one, while we compare two estimated subspaces of the same matrix, so our errors come from the same noise. Also, his bias corrections are derived for noise with independent standard Gaussian entries, where they depend only on the singular values and the dimensions of the matrix, so they can be computed from the data at once, whereas our noise is Bernoulli with different variance from entry to entry, and so our correction depends on the whole matrix of edge probabilities and on the singular vectors.

  • Thank you for alerting us to Liu, Xu and Zhang (2026). As far as we can see the mechanism is similar to that of our Section 6: their connection-matrix estimator and our U^⊤V^\widehat{\mathbf{U}}^\top\widehat{\mathbf{V}} are both products of two separately estimated factors, whose second-order errors survive as a bias. Their setting is different, as they have many layers and no sparsity, while we have a single sparse network, but it is closely related enough to deserve a mention in our discussion of related work.

Thank you for the response. Table 3 and the boundary limitation answer my questions 2 and 4.

I am keeping my rating. The proposed changes narrow the claims, but the new results come without proofs or a revised PDF, and coverage with τ^\hat\tau is not shown. Even if they hold, novelty, already raised in my review, decides.

  1. The index is classical. The paper credits Krzanowski (1979), and you now add that ΨF\Psi_F is Pillai's trace from CCA without centring.

  2. In the proved regime the network adds nothing to the limit. Theorem 4 has the limit of Proposition 8 for observed positions, with the remainder controlled by existing expansions (Xia, 2021; Lei and Rinaldo, 2015).

  3. What is new is the explicit bias term of Section 6 and the degree-free CLT. The correction is proved only where Ψ^\widehat\Psi is already asymptotically valid. There it improves finite-sample coverage; in sparser networks it remains, in your words, a conjecture. The degree-free estimator has no variance estimate.

I find the proved contribution small relative to prior work and insufficient for ICLR.

The Table 3 intervals should be dropped below the threshold: the corrected OpenFlights interval, [1.04,1.04][1.04, 1.04], lies outside [0,1][0, 1].

Reviewer_1iVUOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces the effective shared dimension, a count of how many directions a node's sending and receiving roles share in a directed network, together with a degree-free variant that removes node activity. The estimation on 10 real directed networks show the effectiveness of the proposed concepts.

Critical strengths

S1. This paper gives a closed-form expressions for the proposed concept of effective shared dimension and its degree-free version.

S2. The values of effective shared dimensions and related concepts are calculated on 10 real-world directed network datasets.

Critical weaknesses

W1. In this paper, it is assumed that the sending and receiving subspaces S_{out} and S_{in} have finite dimension d. However, for a large number of real-world directed networks, the assumption may be strong. For instance, for teacher-student relationships, both teachers and students can make progress though the sending and receiving behaviors, whose dimensions clearly may be different for teachers and students. Then, this assumption greatly restricts the applicable scenarios of the proposed effective shared dimension metrics. How about the presented concepts and conclusions of the paper when the dimension of S_{out} is different from that of S_{in}?

W2. Some related work could be reviewed and discussed, such as [1][2].

W3. Case studies should be provided for the potential applications.

[1] A Latent Space Approach to Inferring Distance-Dependent Reciprocity in Directed Networks. Stat, 14(1): e70042, 2025.

[2] Optimal estimators and tests for reciprocal effects. Journal of the American Statistical Association, 2026.

Questions for authors

W1-W3. The authors are encouraged to provide their rebuttal, and I will be glad to increase my rating base on the potential discussion.

2 comments

Thank you for your feedback. In response to the points you have raised:

  • W1. For the network as a whole the two subspaces must necessarily have the same dimension, which is equal to the rank of the integral operator associated to ff (we make the assumption that this rank is finite, the fact that the dimensions are equal is then a direct consequence of this). However, within this framework the dimensions for each type of node are allowed to differ (e.g. in a 44-dimensional example, teachers could have sending and receiving subspaces of dimensions 33 and 11 respectively, while students have these dimensions swapped). We note that the estimate Ψ^\widehat\Psi can be written as ∑i=1nui⊤M^v^i\sum_{i=1}^n{\mathbf{u}}_i^\top\widehat{\mathbf{M}}\hat{\mathbf{v}}_i where M^=U^⊤V^\widehat{\mathbf{M}}=\widehat{\mathbf{U}}^\top\widehat{\mathbf{V}}, and restricting the sum to a specific type gives an estimate for that type's contribution to ΨF\Psi_F, and so a potential extension of our measure is to develop the asymptotic theory to give a per-type decomposition of the effective-shared dimension.

  • W2. Thank you; we will cite and discuss both papers. Both concern (in the case of [1] an extension of) the p1p_1 model of Holland and Leinhardt, in which reciprocity arises as a dependence between the two directions of a dyad, and the sending and receiving roles of a node are either scalars or (in the case of [1]) share a single latent position. Our model assumes two distinct sending and receiving positions for each node, and we then measure how these roles align across the whole network.

    What both models do allow for that ours lacks is dependence within dyads. Our assumption of independence in this case simplifies the theory, but isn't entirely necessary: the proof of the CLT of Theorem 4 doesn't depend on this assumption, but the bias correction term does, so would need to be changed to account for it. An extension of our model would be to allow for within-dyad dependence, either by simply removing the independence assumption or explicitly modelling it. Given that many of the real data networks display some degree of reciprocity, we should at least make mention of our current model's limitation in this regard.

  • W3. As a potential application we can consider the ramifications of the additional theoretical result proposed in point 3 of our general response. Given that as ΨF\Psi_F decreases our ability to accurately estimate the underlying rank dd probability matrix also reduces, we propose a demonstration that certain typical downstream analysis tasks that use an embedding (such as clustering or link prediction) display reduced performance when we consider a single dd-dimensional embedding as the effective shared dimension decreases.

Thank the authors for their rebuttal. Considering the rebuttal and the revised paper address most of my concerns, I will raise my scores.

Reviewer_zWtkOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how closely sending and receiving roles are related in directed networks: whether nodes with similar outgoing connection patterns also exhibit similar incoming connection patterns. It quantifies this relationship through the effective shared dimension, a measure of overlap between sending and receiving subspaces based on their principal angles. It also introduces a degree-normalised version to adjust for differences in node activity.

The main contribution is a statistical analysis of these quantities under a directed random dot product graph model. The authors derive population expressions in terms of latent-position moments and construct estimators from the singular vectors of the observed adjacency matrix. They establish consistency and asymptotic normality under suitable sparsity conditions. They further derive a leading bias correction and empirically investigate its ability to support inference in sparser networks. The theoretical guarantee for this extension remains conjectural. Simulations and applications to ten real networks illustrate the estimators' behavior, variation in role alignment across domains, and their distinction from reciprocity.

Critical strengths
  1. Strong theoretical guarantees for a simple estimator. The paper measures sender–receiver alignment using the singular vectors of the adjacency matrix. It proves that the estimate approaches the true value as the network grows and has an asymptotically normal error when the network is sufficiently dense. These results cover both the standard and degree-normalized scores and include explicit convergence rates and variance formulas. This is a meaningful contribution because both roles must be estimated from the same noisy network.

  2. A useful analysis and correction of estimation bias. The paper investigates why the normal approximation becomes inaccurate in sparser networks. It derives the leading second-order bias term and proposes a way to estimate and subtract it. Simulations show that the correction reduces bias and improves confidence-interval coverage, particularly in denser settings. Although its guarantee for sparser networks remains open, the correction provides a concrete improvement supported by the experiments.

The main contribution is theoretical, and the authors frame it that way.

Critical weaknesses
  1. The limiting variance depends on the unknown distribution of the latent positions, but the paper does not provide a way to estimate it from the observed graph. The simulations use the true variance, and Table 3 reports only point estimates. It is therefore unclear how a practitioner should construct confidence intervals using the proposed method. Furthermore, the uncorrected intervals have lower coverage than intended at the tested network sizes, and the proposed guarantee for the bias correction in sparser networks remains unproved. A variance estimator, together with simulations checking the resulting confidence intervals, would strengthen the inference contribution.

  2. Section 7 describes the score as measuring whether a node sends to and receives from the same kinds of nodes. However, consider three groups connected in a cycle: group 1 sends to group 2, group 2 to group 3, and group 3 to group 1. The score is maximal, even though every node sends to one group and receives from a different group. This happens because sending and receiving patterns identify the same underlying groups. Thus, the score measures correspondence between sending and receiving roles across nodes; it does not require a node’s outgoing and incoming partners to be of the same kind. The authors should clarify this distinction when interpreting the real-network results.

Questions for authors

Refer Weakness.

2 comments

Thank you for your feedback. In response to your points, we agree that the lack of an estimate for the variance is unsatisfactory, and so plan to remedy this by including a consistency result for such an estimate, as described in our general response (point 2). Using this result will allow us to provide confidence intervals rather than the point estimates that we currently use. We are also able to provide a guarantee for the corrected estimate Ψ~\widetilde{\Psi} in the dense regime, although unfortunately we are as yet still unable to extend this proof to sparser regimes, and so its use there remains a conjecture. However, given that in the dense regime the corrected estimate addresses the lower than expected coverage provided by our initial estimate, we would recommend that a practitioner use Ψ~\widetilde{\Psi} and the new variance estimate τ^\widehat{\tau} to report confidence intervals.

Finally, your cycle example highlights some lax wording on our part in Section 7, which we thank you for drawing to our attention. Since this wording is inconsistent with our message throughout the rest of the paper we should have noticed it ourselves, and we will correct it.

Thank you for the rebuttal. The authors response sufficiently addressed my concerns. I’ll keep my positive rating.

Reviewer_sMRrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

To address the question whether understanding a node’s sending behavior gives any information about its receiving behavior, the paper proposes to use a measure proposed by Krzanowski (1979) - the effective shared dimension and a degree-free counterpart. The paper provides close-form expressions for these two measures under the assumption that the network can be modeled by a directed random dot product graph. It applied the estimators to estimate the effective shard dimension across 10 real directed networks, and analyzed whether the answer differs systematically across these domains.

Critical strengths

The paper provides close-form estimators for the effective shared dimension and its degree-free variant. It proves consistency of the estimators, as well as asymptotic normality given the average degree grows faster than n^½.

Critical weaknesses

It is not clear how the proposed measures (effective shared dimension and its variants) provide any new insights on understanding the relationship between a node’s sending behavior and receiving behavior in a directed network, compared with other measures.

Questions for authors

Are there any other standard measures that can be used to understand the relationship between a node’s sending behavior and receiving behavior in a directed network? How do the proposed measures differ from them?

1 comment

Thank you for your feedback. In response to your question, some other similar standard measures include:

  • Reciprocity is the proportion of links in a network that are returned. In the p1p_1 model of Holland and Leinhardt (1981) it is presented a parameter that determines the probabilitiy of a link being returned, while Garlaschelli and Loffredo (2004) present it as the correlation between the presence of a link and of its reverse. Regardless of the description, it is a quantity that describes the behaviour of pairs of nodes, while ours describes the roles that nodes play. The two measures can disagree completely: in the example given by reviewer zWtk of a cyclical network comprising three communities in which community 1 sends to community 2, 2 to 3 and 3 to 1, no link is returned, yet a node's sending role determines its receiving role, and so reciprocity is equal to zero while ΨF=d\Psi_F=d.
  • Degree correlation (Newman, 2018) asks whether nodes that send a lot also receive a lot. This is essentially the effective shared dimension when d=1d = 1, as the first dimension is typically associated with degree of nodes. More precisely, when d=1d = 1 a node's expected out-degree is a linear functional of its sending position (and likewise its expected in-degree is a linear functional of its receiving position), and ΨF\Psi_F is exactly the squared uncentred correlation between these positions. When d>1d>1 however, the degrees still only see a single direction of each role (the overall activity of a node) while ΨF\Psi_F counts all dd directions: two networks can have the same degrees and entirely different role alignment. The degree-free effective shared dimension is designed to remove exactly the direction captured by node degree.
  • Canonical correlation analysis (Hotelling, 1936) is the closest relative of our measure. Given two sets of observed variables, the canonical correlations are the cosines of the principal angles between the two sets (Björck and Golub, 1973), and the sum of their squares is Pillai's trace (Pillai, 1955). As mentioned in the paper, Krzanowski (1979) used this sum to compare the principal components of two groups, and ΨF\Psi_F is this sum for a node's sending and receiving positions (although in CCA the data is centred first, which we do not do). However, in the classical setting the observed variables come from two independent datasets, while in our case the underlying latent positions are unobserved, and instead estimated from a single noisy, potentially sparse matrix, making the underlying statistical estimation theory fundamentally different.

References.

  • Björck, Å. and Golub, G. H. (1973). Numerical methods for computing angles between linear subspaces. Mathematics of Computation 27(123), 579–594.
  • Garlaschelli, D. and Loffredo, M. I. (2004). Patterns of link reciprocity in directed networks. Physical Review Letters 93, 268701.
  • Holland, P. W. and Leinhardt, S. (1981). An exponential family of probability distributions for directed graphs. Journal of the American Statistical Association 76(373), 33–50.
  • Hotelling, H. (1936). Relations between two sets of variates. Biometrika 28(3/4), 321–377.
  • Krzanowski, W. J. (1979). Between-groups comparison of principal components. Journal of the American Statistical Association 74(367), 703–707.
  • Newman, M. E. J. (2018). Networks, 2nd edition. Oxford University Press.
  • Pillai, K. C. S. (1955). Some new test criteria in multivariate analysis. Annals of Mathematical Statistics 26(1), 117–121.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to all reviewers

We thank the reviewers for their feedback. The following response details our main proposed changes to our submission, with individual responses to follow.

  1. We currently introduce the effective shared dimension ΨF\Psi_F and its degree-free variant ΨF∘\Psi^\circ_F and establish a CLT for both, but then focus only on ΨF\Psi_F for the remainder of the paper (the simulated examples and the discussion of the bias correction term). The additional theoretical results we propose below also follow this pattern, and so we propose to postpone (and condense) the discussion of ΨF∘\Psi^\circ_F until the end of the paper. This retains the content of the paper, but presents ΨF∘\Psi^\circ_F as an extension of the ΨF\Psi_F with a less-developed theory. It's debatable whether this improves the paper, but we feel it draws a more honest distinction between the two.

  2. We can define an estimate τ^2\widehat\tau^2 for the variance term τF2\tau_F^2 by setting τ^2=1n∑i=1ng^i2,g^i=n(2 u^i⊤M^v^i−u^i⊤M^M^⊤u^i−v^i⊤M^⊤M^v^i)\widehat\tau^2=\frac1n\sum_{i=1}^n\widehat g_i^2,\quad \widehat g_i=n\bigl(2\,\widehat{\mathbf{u}}_i^\top\widehat{\mathbf{M}}\widehat{\mathbf{v}}_i-\widehat{\mathbf{u}}_i^\top\widehat{\mathbf{M}}\widehat{\mathbf{M}}^\top\widehat{\mathbf{u}}_i-\widehat{\mathbf{v}}_i^\top\widehat{\mathbf{M}}^\top\widehat{\mathbf{M}}\widehat{\mathbf{v}}_i\bigr) where M^=U^⊤V^\widehat{\mathbf{M}}=\widehat{\mathbf{U}}^\top\widehat{\mathbf{V}}, and u^i\widehat{\mathbf{u}}_i and v^i\widehat{\mathbf{v}}_i are the ii th rows of U^\widehat{\mathbf{U}} and V^\widehat{\mathbf{V}} respectively.

    Under the assumptions of Theorem 4, we can show that if nρn=ω(n1/2)n\rho_n=\omega(n^{1/2}) then τ^2→τF2\widehat\tau^2\to\tau_F^2 in probability, and so n1/2(Ψ^−ΨF)/τ^→N(0,1)n^{1/2}(\widehat\Psi-\Psi_F)/\widehat\tau\to\mathcal{N}(0,1).

    We can also prove that the corrected estimate Ψ~\widetilde\Psi satisfies n1/2(Ψ~−ΨF)→N(0,τF2)n^{1/2}(\widetilde\Psi-\Psi_F)\to\mathcal{N}(0,\tau_F^2), and thus n1/2(Ψ^−ΨF)/τ^→N(0,1)n^{1/2}(\widehat\Psi-\Psi_F)/\widehat\tau\to\mathcal{N}(0,1) also (again, above the nρn=ω(n1/2)n\rho_n = \omega(n^{1/2}) threshold).

    As a result, we will replace the coverage values appearing in Tables 1 and 2 of the original paper (which used the true variance τF2\tau_F^2) with those obtained using the estimated variance τ^2\widehat{\tau}^2.

  3. At the time of submission, we decided for space considerations to simply present the definitions and theoretical results around ΨF\Psi_F and eschew our reasoning for looking at it in the first place, which we realise now leaves the treatment somewhat lacking.

    Our interest arose from the question of much information we lose if we consider only a single dd-dimensional embedding for a DRPDG of rank dd. In the undirected case (where ΨF=d\Psi_F = d) a single embedding captures everything, so it stands to reason that as the shared dimension decreases, so too does the effectiveness of a single embedding. We propose to add a short section, presenting ΨF\Psi_F as a descriptive parameter which informs us about the loss of information from considering a single embedding, supported by the following theoretical result:

    If ΨF<d\Psi_F<d, there is a constant 0<c≤C0<c\le C such that, with probability tending to one,

    c(d−ΨF)1/2≤∥P−ZBZ⊤∥F∥P∥Fc(d-\Psi_F)^{1/2} \le \frac{\|\mathbf{P}-\mathbf{Z}\mathbf{B}\mathbf{Z}^\top\|_F}{\|\mathbf{P}\|_F}

    for every Z∈Rn×d\mathbf{Z} \in \mathbb{R}^{n \times d}, where B\mathbf{B} minimises ∥A−ZBZ⊤∥F\|\mathbf{A} - \mathbf{Z}\mathbf{B}\mathbf{Z}^\top\|_F. In other words, the relative error of the best rank dd approximation to A\mathbf{A} we can obtain using Z\mathbf{Z} compared to P\mathbf{P} (the best rank dd approximation) is always bounded below by a non-zero constant if the effective shared dimension is less than dd, so we cannot fully recover the structure.

  4. We also propose the following minor edits:

  • The threshold of n1/2n^{1/2} appearing in Theorem 4 is sharp, and one can demonstrate that at this threshold the limiting distribution of Ψ^−ΨF\widehat{\Psi}-\Psi_F has non-zero mean, so we will add this to the Theorem's statement, with the proof presented in the appendix.
  • As reviewer xHVf rightly pointed out, it was remiss of us to not mention the link to CCA, and we will make note of this in the revised edition.
  • For the real data examples of Section 7, we will present the average degree of each network as well as confidence intervals for the effective shared dimension. We will also consider the largest weakly-connected component of each network, instead of restricting to a core, which we had done previously to ensure that each node had at least one in and one out edge, allowing us to calculate the degree-free estimate. All but one of the networks (the Florida Bay food web) are then sparser than the regime covered by Theorem 4, so we will make it clear that this section is purely descriptive.

We have uploaded a revised version of our paper, roughly in line with the proposed changes of our earlier comment. Since this involves a fair amount of restructuring of the paper, we summarise the main changes below (note that due to renumbering, Theorem 5 in this revision was Theorem 4 in the original submission):

  1. We have introduced a variance estimate τ^\hat\tau, which we prove to be consistent (Theorem 5.ii).
  2. The estimate Ψ^\widehat\Psi is now shown to converge to a Gaussian distribution with non-zero mean at the sparsity threshold nρn∼cn1/2n\rho_n \sim cn^{1/2} (Theorem 5.iv).
  3. We prove that a CLT holds for the corrected estimate Ψ~\widetilde{\Psi} when the sparsity satisfies nρn=ω(n1/4)n\rho_n = \omega(n^{1/4}) (Theorem 6), thus proving that it does allow us to reach sparser regimes than the standard estimate. In particular all but two of the real data examples have average degree covered by this regime.
  4. We expand Section 4 slightly to include a simulation demonstrating how performance at a standard community recovery task for a co-blockmodel varies with ΨF\Psi_F.
  5. The paper's focus is directed towards the standard effective shared dimension ΨF\Psi_F, with the degree-free estimate now moved to a separate, final section.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.