- 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
-
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.
-
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
- 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.
- 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.
- 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
-
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?
-
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?
-
Is the n^{1/2} threshold proved sharp, or only observed?
-
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 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 is not shown. Even if they hold, novelty, already raised in my review, decides.
-
The index is classical. The paper credits Krzanowski (1979), and you now add that is Pillai's trace from CCA without centring.
-
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).
-
What is new is the explicit bias term of Section 6 and the degree-free CLT. The correction is proved only where 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, , lies outside .
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% 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:
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−1, d and d+1. In most cases the estimate is fairly stable, but CollegeMsg shows a large difference between dimensions.
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 we can only say that n1/2(Ψ−ΨF)→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 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.