acceptodds
Under review as a conference paper at ICLR 2027

Incidence-Structured Spatial Factors for Latent Causal Discovery

Abstract

Latent causal discovery compresses many spatial sites into a few latent time series and learns the causal graph among them. Existing methods link latents to sites through kernels over site coordinates, learned site partitions or principal components. Spatial data often carry memberships such as regions, and the hypergraph lift of a kernel sees them only through their pairwise operator. We show that this operator does not in general determine the latent coordinates that the memberships determine. Two incidences with the same pairwise operator admit models with the same observed law and inequivalent latent laws. We therefore propose incidence-structured spatial factors, which build each latent mode as a weighted union of structural units. For the unsmoothed factor, we prove the latents identifiable up to permutation, scaling and translation under site-specific smooth invertible observation maps when every mode owns an anchor site and bridge sites connect the modes. On a public grid benchmark, the factor recovers the latent causal graph more accurately than kernel, single-parent, two-stage and sequence-model baselines. On random committee hypergraphs with membership-following footprints, hyperedge atoms, given or reconstructed from the pairwise operator, recover the latents more accurately than pair atoms and heat kernels built on that operator.

Primary area:
causal reasoning
Keywords:
causal discovery, causal representation learning, spatiotemporal data, latent variable models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_dP8YOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes incidence-structured spatial factors (ISF) for latent causal discovery. ISF constructs spatial loadings from whole membership units, with each unit assigned to at most one latent variable while allowing observed sites to have multiple latent parents. The paper proves that a pairwise hypergraph projection can fail to determine latent coordinates that are identifiable from the full incidence structure. It also establishes identifiability for unsmoothed factors under anchor and bridge conditions. Synthetic experiments evaluate latent recovery and causal discovery against kernel-based and other baselines.

Critical strengths
  1. The separation result connects ambiguity in hypergraph projections to ambiguity in latent coordinates, providing a substantive reason to preserve incidence information.

  2. The anchor and bridge conditions clarify how membership structure constrains latent representations, including with dependent latents and site-specific nonlinear observation maps.

  3. Shared-training comparisons, structure-preserving rewiring, reconstructed memberships, and repeated independent draws help distinguish structural benefits from training effects.

Critical weaknesses

I'm not familiar with the topic of this paper, and I find only one notable weakness: the paper combines two distinct choices: constructing loadings from whole hyperedges and restricting each hyperedge to at most one latent variable. However, Appendix G, Table 32 shows that continuous nonnegative loadings on the same hyperedge dictionary outperform the proposed assignments in both settings, which support the value of the hyperedge dictionary more strongly than the single-parent restriction used in the identifiability theory.

Questions for authors

In Section 2, Identifiability, the authors claim that "pure-measurement conditions identify linear latent causal structure". In fact, some studies [1,2] do not rely on this condition. I recommend the authors to discuss this point.

[1] Learning linear bayesian networks with latent variables. ICML 2013.
[2] Recovery of causal graph involving latent variables via homologous surrogates. ICLR 2025.

3 comments
AuthorsOpenReview ↗

Response to Reviewer dP8Y

We thank the reviewer for the careful reading of the appendix and for the two references. Below, we address each point. All new experiments use the code from the submission and will be added to the revised paper.


W1 – Hyperedge dictionary and single-parent restriction – We agree. The two choices have different roles, and Theorem 1 now needs the restriction only on anchor-bearing units.

Thank you for this observation. It led to a stronger result.

  • The hyperedge dictionary carries the empirical gain. On the committee hypergraphs, hyperedge atoms raise the latent MCC over pair atoms by 0.089 at noise 2.0 and by 0.044 at noise 1.5, on 10 of 10 structures (Table 9). Continuous loadings on the same dictionary belong to the ISF family of Definition 1. They add a further 0.028 and 0.014 (Table 32).
  • The assignment carries the guarantee. It is the hypothesis class of Theorem 1. This matters for latent causal discovery because mixing the latent coordinates changes the causal graph.

New in this response: the guarantee needs the assignment only on anchor-bearing units. Call a unit anchor-bearing when it contains a site that belongs to no other unit. Beyond the anchors of the true factor, the proof of Theorem 1 uses the assignment of a candidate factor only on these units. Theorem 1 therefore holds, with the same proof, when anchor-bearing units load on at most one mode and all other units carry continuous nonnegative loadings. We will state this as a corollary.

New experiment. We compare three ISF variants on one hyperedge dictionary. Table 32 covers only the committee hypergraphs. The new runs add the anchored structures of Appendix F.4, where Theorem 1 applies, and twin incidences built from Theorem 2.

Table R3: Mean latent MCC of three ISF variants, over ten structures or draws with three model seeds each. Bold marks the best value in each row. The committee rows rerun Table 32 on different hardware.

SettingAssignment on every unit (paper)Assignment on anchor-bearing units (corollary)Continuous loadings
Anchored structures, noise 0.50.9760.9880.987
Committee hypergraphs, noise 2.00.7900.8090.818
Committee hypergraphs, noise 1.50.8850.8840.894
Twin incidences, one unit per mode0.825—0.872
  • The new runs support your reading in every setting. The hyperedge dictionary carries the gain, with or without the assignment on every unit.
  • Where Theorem 1 applies, the corollary variant reaches 0.988, with at least 0.987 on each of the ten anchored structures. It carries the guarantee and has the highest mean MCC of the three.
  • On the committee hypergraphs, every ISF variant exceeds pair atoms (0.702 and 0.839) and the heat kernel (0.464 and 0.570) in Table 3.

In the revision, we will present Definition 1 as the factor family and the single-mode restriction as the hypothesis class of the theory. We will also report the three variants side by side in the main text.


Q1 – Identification without pure measurements – Thank you, we agree. We will revise the sentence in Section 2 and cite both works.

Both works identify latent structure without pure children.

  • Anandkumar et al. (ICML 2013) [1] use expansion conditions on the graph between latent and observed variables of a linear model.
  • Li et al. (ICLR 2025) [2] replace pure children with homologous surrogates, which allow far more flexible parents.

The revised sentence will say that pure-measurement conditions are one route to identifying linear latent causal structure, and that these works remove the need for pure children.

The same distinction appears in our paper. Theorem 1 uses anchors, a condition of the pure-measurement kind. The fixed-incidence part of Theorem 2 and the cyclic family of Section 5.1 identify the latents in their stated classes without any anchor. Every site belongs to two units in the first, and to three units in the second.

We hope these clarifications and the new results strengthen your positive assessment.


[1] Anandkumar, Hsu, Javanmard and Kakade. Learning Linear Bayesian Networks with Latent Variables. ICML 2013.

[2] Li, Wang and Liu. Recovery of Causal Graph Involving Latent Variables via Homologous Surrogates. ICLR 2025.

Thanks for your response. I have no other concern. Because I'm not familiar with the topic of this paper, I'd like to maintain my positive score with a low confidence.

Thank you for your kind reply and for maintaining your positive assessment of our work. We are delighted to hear that all of your concerns have been resolved. Your feedback from a broader perspective has been very helpful in improving the clarity and accessibility of our manuscript.

Reviewer_RMc7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper brings known hypergraph memberships (the incidence matrix H) directly into the spatial factor used for spatiotemporal latent causal discovery. It replaces coordinate kernel factors with incidence structured factors (ISF), F=HˉAF = \bar{H}A, where the spatial loading of each latent mode is a normalized combination of structural units (hyperedges or edges), with optional heat smoothing. On the theory side, the paper shows that (i) the weighted clique expansion operator may fail to determine the latent coordinates that the hypergraph incidence determines (Thm 2), and (ii) under anchor and bridge conditions and a known, shared noise law, the latents of the unsmoothed hard assignment class are identifiable up to permutation, scaling and translation (Thm 1, via the affine reduction in Thm 3). Empirically, on the SPACY grid benchmark with baselines trained under the same recipe, ISF recovers the causal graph better than RBF, heat kernel, single parent and two stage baselines. On synthetic committee hypergraphs, hyperedge atoms recover the latents better than pair atoms, a heat kernel on the same pairwise operator, and rewired memberships.

Critical strengths

The core novelty is using hypergraph memberships as whole units to build the spatial factor, rather than accessing them only indirectly through a pairwise operator, as heat kernels and spectral hypergraph convolutions do. Thm 2 gives a theoretical reason for this choice, and on the committee hypergraphs hyperedge atoms achieve higher MCC than both pair atoms and the heat kernel on every structure.

The identifiability result is clearly stated, with explicit assumptions. Its conditions can be read off the incidence and the unit assignment, and it allows dependent latents and nonlinear observation maps that vary by site.

The experiments include fairly extensive controls and repeated validation, and ISF wins on a benchmark whose true factors come from the RBF family used by SPACY.

Critical weaknesses

The benefit of hypergraph memberships is shown only on synthetic data; a real dataset with known memberships is still missing.

Thm 2 is an existence result. On the committee structures, the hyperedges can be recovered exactly from the pairwise operator (App. F.3), so it remains unclear whether the information the pairwise operator lacks matters in practice.

The grid benchmark uses only edges between pairs of sites as its incidence and contains no higher order memberships. It is not yet clear whether the advantage of ISF on the grid comes from the incidence structure or from smoothing (at D = 20, unsmoothed ISF falls below RBF).

Questions for authors

In the D = 20 grid experiment, if Free Footprint and Free Edges are given the same heat smoothing as ISF, does ISF still have an advantage?

Can Thm 1 be extended to continuous nonnegative A with anchor rows? If not, what concrete guidance would you give for choosing hard assignment over continuous loadings?

The identifiability results assume a known, shared noise law, while the estimator fixes the Gaussian variance at 1/2. How sensitive is latent recovery to a misspecified noise scale?

5 comments
AuthorsOpenReview ↗

Response to Reviewer RMc7 (1/3)

We thank the reviewer for the precise summary and the constructive questions. Below, we address each point. We ran four new experiments and extended Theorem 1. All new experiments use the code from the submission and will be added to the revised paper.


W2 – Does the missing information matter in practice – Yes. We added an experiment in which two different memberships share one pairwise operator.

We agree that the committee hypergraphs cannot answer this question, because their operator determines their hyperedges (Appendix F.3). We therefore ran the test on structures where it does not.

Setup. We tile five copies of the six-site twin of Theorem 2 and join them by shared pair units. This gives 30 sites, 24 units and 24 modes. Incidences A and B have the same weighted clique expansion, Laplacian, degrees and unit sizes. The data are generated from A. The same observations also follow from a model on B with different latent coordinates. We score every estimate against both coordinate systems, over ten draws with three model seeds each.

Table R1: Latent MCC on twin incidences with an identical pairwise operator. Bold marks the best value in each column.

MethodInputCoordinates of ACoordinates of B
ISF on incidence Aincidence0.8260.646
ISF on incidence Bincidence0.6410.828
ISF on pair atomspairwise operator0.7220.713
Heat kernelpairwise operator0.6550.637
ISF on reconstructed unitspairwise operator0.7420.723

As shown in Table R1, ISF scores highest against the coordinates of the incidence it receives, on 10 of 10 draws in both directions. The three operator-only methods receive the same input under A and B, and they do not separate the two coordinate systems. Against the generating coordinates, ISF exceeds pair atoms by 0.104 MCC and the heat kernel by 0.171, each on 10 of 10 draws.

The reconstruction in Appendix F.3 is also no longer unique here. The operator is the same in all ten draws, and the factorization recovers only 13.3 of the 24 units on average. The missing information therefore matters in practice on such structures. We will add this experiment to Section 6.3, next to the reconstruction result.


W3, Q1 – Incidence structure or smoothing – Yes, ISF keeps its advantage when the free variants receive the same heat smoothing.

We ran the requested experiment on the twenty D = 20 draws of Table 1. Both free variants receive the smoothing of Equation (5), with the same parameterization, prior, initial scale and learned per-mode scale as ISF. ISF is rerun on the same machines.

Table R2: Edge F1 and AUPRC at D = 20, and the number of draws won by ISF. ISF is rerun here (Table 1 reports 0.657 and 0.662). Bold marks the best value in each column.

VariantF1AUPRCDraws won by ISF (F1, AUPRC)
Free footprint (Table 2)0.5620.56420/20, 20/20
Free weights on grid edges (Table 2)0.5720.58619/20, 20/20
Free footprint with heat smoothing (new)0.5590.56519/20, 18/20
Free weights on grid edges with heat smoothing (new)0.5720.58720/20, 18/20
ISF0.6540.657—

As shown in Table R2, ISF keeps a mean F1 gain of 0.094 over the smoothed free footprint and 0.082 over the smoothed free weights on grid edges. Smoothing leaves the free variants essentially unchanged.

The learned scales explain why. The free variants keep their smoothing scale near the initial value of 0.5, because a free footprint can already take any shape. ISF learns a large scale, with a mean of 91.5, because smoothing turns hard-edged unions of grid edges into connected regions. Smoothing therefore complements the atoms.

Three results in the paper agree with this.

  • Smoothing without atoms is the heat kernel. It reaches an F1 of 0.486 at D = 20. This is below the coordinate kernel (0.574), and below ISF (0.657) on 20 of 20 draws (Tables 1 and 29).
  • Smoothing on the wrong incidence does not help. When the grid adjacency is rewired and the smoothing is recomputed on it, ISF falls to an F1 of 0.275 (Table 17).
  • No smoothing is involved at D = 10. Unsmoothed ISF exceeds the coordinate kernel there: F1 0.666 against 0.631, and AUPRC 0.699 against 0.619.

The grid advantage therefore needs the atoms and the correct incidence. At D = 20 it also needs the smoothing, which complements the atoms. Higher-order units are tested in Section 6.3 and in Table R1. We will state this division of roles at the start of Section 6.

AuthorsOpenReview ↗

Response to Reviewer RMc7 (2/3)

Q2 – Theorem 1 with continuous loadings – Yes, Theorem 1 extends to continuous loadings in two forms.

All statements below hold under (A0) to (A3) with D≥2D \ge 2 and a fixed incidence.

Form 1: the assignment is needed only on anchor-bearing units. Call a unit anchor-bearing when it contains a site that belongs to no other unit. This is a property of HH alone. Consider the factors F=HˉAF = \bar H A with A≥0A \ge 0 and rank F=DF = D in which every anchor-bearing unit loads on at most one mode. All other units carry continuous nonnegative loadings on any number of modes. Let the true factor satisfy (C1) and (C2) in this form:

  • (C1) every mode has an anchor-bearing unit that loads on that mode alone,
  • (C2) the mode co-occurrence graph is connected, where two modes are adjacent when some site loads on both.

Theorem 1 then holds for this larger class with the same proof. Beyond the anchor rows of the true factor, the proof uses the assignment of a candidate factor only at an anchor site vv of a unit ee. There the row F^v=HˉveA^e\hat F_v = \bar H_{ve} \hat A_e must load on at most one mode. The class of Definition 2 is contained in this class. We will state this result as a corollary.

Form 2: fully continuous AA with anchor rows. Suppose every compared factor F=HˉAF = \bar H A with A≥0A \ge 0 has, for each mode, a row that loads on that mode alone. Suppose also that the true factor satisfies (C2). Suppose further that the observation maps of the compared models have the same orientation at every site, for example all increasing. Then the latents are again identified up to permutation, scaling and translation. The proof is short. For observationally equivalent models, Theorem 3 gives CF^=FJ0C \hat F = F J_0 with a positive diagonal CC. The anchor rows of FF give J0≥0J_0 \ge 0, and the anchor rows of F^\hat F give J0−1≥0J_0^{-1} \ge 0. An invertible nonnegative matrix with a nonnegative inverse is monomial.

What cannot be dropped. Two examples show this. Both use identity maps in the true model and independent standard Gaussian latents ZZ.

  1. Anchors in the true factor alone do not suffice. Take A≥0A \ge 0 with (C1) and (C2), and J=I+ε11⊤J = I + \varepsilon \mathbf{1}\mathbf{1}^\top with ε>0\varepsilon > 0. The factor HˉAJ\bar H A J is nonnegative with rank DD. The models (HˉA,Z)(\bar H A, Z) and (HˉAJ,J−1Z)(\bar H A J, J^{-1} Z) have the same observed law. The latents J−1ZJ^{-1} Z are correlated, so no permutation, scaling and translation maps their law to that of ZZ.
  2. Form 2 can fail without the orientation condition. Take D=2D = 2, a true factor with rows (1, 0), (0, 1), (2, 1), (1, 1), and a second factor with rows (1, 1), (1, 2), (1, 0), (0, 1). Both are nonnegative and have an anchor row for each mode. Both are realized on one incidence: four single-site units and one unit of all four sites with zero loading. Let J0J_0 have rows (1, 1) and (−1, −2). With latents J0−1ZJ_0^{-1} Z and maps (y,−y,y,−y)(y, -y, y, -y) at the four sites, the second model has the same observed law as the first. Its latents are again correlated.

Form 1 needs no orientation condition, because its argument uses supports rather than signs.

AuthorsOpenReview ↗

Response to Reviewer RMc7 (3/3)

Q2 – Hard assignment or continuous loadings – We recommend the assignment on anchor-bearing units and continuous loadings on all other units.

This rule follows from Form 1 above. It keeps the guarantee of Theorem 1 and adds flexibility. We tested it in a new experiment that compares three ISF variants on one hyperedge dictionary.

Table R3: Mean latent MCC of three ISF variants, over ten structures or draws with three model seeds each. Bold marks the best value in each row. The committee rows rerun Table 32 and the twin row reruns Table R1 on different hardware.

SettingAssignment on every unit (paper)Assignment on anchor-bearing units (corollary)Continuous loadings
Anchored structures, noise 0.50.9760.9880.987
Committee hypergraphs, noise 2.00.7900.8090.818
Committee hypergraphs, noise 1.50.8850.8840.894
Twin incidences of Table R10.825—0.872
  • Anchored structures. Theorem 1 applies here. The corollary variant reaches 0.988, with at least 0.987 on each of the ten structures. It carries the guarantee and has the highest mean MCC of the three.
  • Committee and twin structures. Few units are anchor-bearing here: 2 to 7 out of 60 on the committee hypergraphs, and none on the twin incidences. Continuous loadings are the most accurate variant.
  • All variants. On the committee hypergraphs, every ISF variant exceeds pair atoms (0.702 and 0.839) and the heat kernel (0.464 and 0.570) of Table 3.

Our guidance is therefore simple. Keep the assignment on the anchor-bearing units when the latent coordinates are the target, as in latent causal discovery. Use continuous loadings on all other units. The assignment on every unit is the right choice when a partition of the units into modes is itself of interest. We will report the three variants side by side in the main text.


Q3 – Misspecified noise scale – Recovery is stable when the declared variance is at or below the paper's value. We added a sensitivity study.

Assumption (A0) concerns the generating class. In the estimator, the variance of the Gaussian reconstruction term weights reconstruction against the prior terms on observations standardized to unit global variance (Appendix D.2). We varied this variance by a factor of four in both directions.

Table R4: Mean latent MCC for each declared variance, over ten structures with three model seeds each. Bold marks the best value in each row.

Declared variance1/81/41/2 (paper)12
Anchored structures (noise share 0.47)0.9820.9780.9760.9710.941
Committee hypergraphs, noise 1.5 (noise share 0.81)0.9080.9120.8830.7840.633

As shown in Table R4, recovery is stable for every declared value at or below 1/2. It declines only when the variance is overstated, because the prior terms then outweigh the reconstruction. In both settings, every declared value at or below the noise share of the standardized variance is safe. The paper's value was not tuned in favor of ISF. On the committee hypergraphs, a declared variance of 1/4 gives a higher MCC than 1/2 on 10 of 10 structures. We will add this study to Appendix H.


W1 – Real dataset with known memberships – We follow the standard protocol of this literature, which scores recovery on simulated data, and our controls isolate the effect of the memberships.

Recovery of the latents and the latent causal graph can be scored only against known latents and a known graph. Real spatiotemporal data provide neither. This is the standard quantitative protocol in this literature. Wang et al. (2025) and Brouillard et al. (2026) score recovery on simulated data. Tibau et al. (2022) introduce their synthetic benchmark for the same reason. On real climate data, the first two works inspect the learned maps qualitatively and report no recovery score. Our grid benchmark uses the public generator of Wang et al. (2025).

Simulation also gives clean controls. Rewired, reconstructed and twin memberships (Table R1) each change one property of the incidence and nothing else. The abstract and the conclusion state the claims within this scope.

We hope these answers and the new results strengthen your positive assessment.

Reviewer_na3POpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The primary contribution of the paper is an interesting separation result for factor models. This leads to a nice theoretical case against pairwise and heat-kernel lifts of hypergraphs. This is a significant result given their prevalence. Th epaper provides (synthetic) data that demonstrates the case.

Critical strengths

The core result of the paper (that a factor should read indicence not its projection) is significant, and of value to the research community. The authors achieve this by building two 6-site, 4-triple incidences with the same operator, degrees and unit sizes. Each incidence identifies its own latents, but the two admit models with the same observed distribution whose latents are not related by a permutation-and-scaling map.

Critical weaknesses

The core result is demonstrated on their synthetic data only. The real data tests do not demonstrate teh same result.

The grid benchmark has no hyper-edges, so ti seems that what it really demostrates is that learned unions of edges beat an RBF.

Given the above it seems to me that the paper makes a valid contribution but overstates it.

Questions for authors

It seems to me that, on the committee hypergraphs, symmetric nonnegative factorisation of the clique expansion recovers every hyper-edge exactly. ISF on those reconstructed memberships matches ISF on the true ones, so the measured gain comes from using unit-shaped atoms, not from information that the pairwise operator lacks. Maybe that's what line 462 is saying? Tell me if I've got that wrong.

2 comments
AuthorsOpenReview ↗

Response to Reviewer na3P (1/2)

We thank the reviewer for the careful reading and for recognizing the value of the separation result. Below, we address each point. All new experiments use the code of the submission and will be added to the revised paper.


Q1 – Reading of line 462 – Yes, your reading is correct. We now also test the case where the pairwise operator lacks information.

Your reading matches L459 - 463 and Appendix F.3. On the committee hypergraphs, the pairwise operator determines the hyperedges, so the gain there comes from unit-shaped atoms. This is a result about how a factor should use memberships. Given the same input, the heat kernel reaches a latent MCC of 0.464 and 0.570 at noise 2.0 and 1.5. ISF on the reconstructed units reaches 0.801 and 0.882 (Table 3).

The committee hypergraphs cannot show what happens when the operator lacks information. We therefore added an experiment in which two different memberships share one pairwise operator.

Setup. We tile five copies of the six-site twin of Theorem 2 and join them by shared pair units. This gives 30 sites, 24 units and 24 modes. Incidences A and B have the same weighted clique expansion, Laplacian, degrees and unit sizes. The data are generated from A. The same observations also follow from a model on B with different latent coordinates. We score every estimate against both coordinate systems, over ten draws with three model seeds each.

Table R1: Latent MCC on twin incidences with an identical pairwise operator. Bold marks the best value in each column.

MethodInputCoordinates of ACoordinates of B
ISF on incidence Aincidence0.8260.646
ISF on incidence Bincidence0.6410.828
ISF on pair atomspairwise operator0.7220.713
Heat kernelpairwise operator0.6550.637
ISF on reconstructed unitspairwise operator0.7420.723

As shown in Table R1, ISF scores highest against the coordinates of the incidence it receives, on 10 of 10 draws in both directions. The three operator-only methods receive the same input under A and B, and they do not separate the two coordinate systems. Against the generating coordinates, ISF exceeds pair atoms by 0.104 MCC and the heat kernel by 0.171, each on 10 of 10 draws. The reconstruction of Appendix F.3 is also no longer unique here. It recovers 13.3 of the 24 units on average.

Theorem 2 proves the separation for the six-site twin. This experiment shows the same pattern in finite samples on the tiled structure. We will add it to Section 6.3, next to the reconstruction result.

AuthorsOpenReview ↗

Response to Reviewer na3P (2/2)

W1 – Synthetic data only – We follow the standard protocol of this literature: recovery is scored on simulated data, where the latents and the graph are known.

All experiments in the paper use simulated signals. This includes the grid benchmark, which is the public synthetic benchmark of Wang et al. (2025). The reason is the evaluation target. Recovery of the latents and the latent causal graph can be scored only against known latents and a known graph. Real spatiotemporal data provide neither.

This is the standard quantitative protocol in latent causal discovery from spatiotemporal data.

  • Coordinate kernels. Wang et al. (2025) score graph and latent recovery on synthetic data, for the same reason. Our grid benchmark uses their public generator with their grid size, sample count and sequence length.
  • Single-parent decoding. Brouillard et al. (2026) also score recovery on simulated data.
  • Two-stage pipelines. Tibau et al. (2022) introduce their synthetic benchmark because causal discovery methods cannot be benchmarked on real climate data without ground truth.

Wang et al. (2025) and Brouillard et al. (2026) also apply their models to real climate data. There they inspect the learned maps qualitatively and report no recovery score.

Simulation also gives clean controls. Each control changes one property of the incidence and nothing else:

  • rewired memberships that keep degrees and unit sizes,
  • memberships reconstructed from the pairwise operator,
  • twin memberships with an identical pairwise operator (Table R1).

The abstract and the conclusion state the scope as "a public grid benchmark" and "random committee hypergraphs with membership-following footprints".


W2 – Grid benchmark has no hyperedges – True. The grid benchmark tests incidence atoms on a graph, and Section 6.3 and Table R1 test hyperedges.

The two experiments answer different questions.

  • Grid benchmark. This is the public benchmark in this literature. Its true factors are radial basis kernels, the family of the strongest baseline. ISF still exceeds the same-recipe kernel in edge F1: 0.666 against 0.631 at D = 10, and 0.657 against 0.574 at D = 20 (Table 1). Incidence atoms are therefore competitive even where the coordinate kernel is correctly specified.
  • Correct incidence. The gain depends on the correct incidence. When the grid adjacency is rewired with degrees, atom sizes and atom count preserved, ISF falls below the kernel on every tested draw (Appendix F.8).
  • Hyperedges. Higher-order units are tested in Section 6.3 and in Table R1.

We will state this division of roles at the start of Section 6.


W3 – Scope of the claims – We agree, and we will state each claim within the scope of its evidence.

We will make three changes in the revision.

  1. Committee result. The introduction and the contribution list will state this as in L459 - 463: given the same information, unit-level atoms outperform pair-level atoms and heat kernels.
  2. Missing information. The claim about information that the operator lacks will rest on Theorem 2 and on the twin experiment of Table R1.
  3. Grid result. We will describe it as a result for incidence atoms on a graph.

We hope these clarifications and the new experiment resolve your concerns. If they do, we would be grateful if you would reconsider your score and confidence.

Comments on the paper

1 comment
AuthorsOpenReview ↗

To the Area Chair and all Reviewers: summary of the discussion and the revised paper. Thank you!

Dear Area Chair and Reviewers,

As the discussion period closes, we sincerely thank the organizers, the area chair and all reviewers. The reviews were insightful, and the discussion has made the paper more complete. Below we summarize the discussion, the contributions of the paper, our view on the remaining question, and where each change appears in the revised PDF. All new and changed text is marked in blue.


Summary of the discussion

  • We answered every comment. We ran four additional experiments and extended Theorem 1. All new experiments build on the code submitted with the paper. The revised paper includes every change described in our responses.
  • Reviewer RMc7 wrote that the new experiments and theoretical clarification address the questions on where the gains come from and when the identifiability result applies. The reviewer added that the rebuttal "strengthens my assessment of the technical contribution" and maintained a positive rating.
  • Reviewer dP8Y wrote "I have no other concern" and maintained a positive score.
  • Reviewer na3P recognized the separation result as significant and asked us to match each claim to its evidence. The reviewer's reading was correct. We revised the introduction and contribution list accordingly, and added an experiment for the case raised by this question.

What the paper contributes

  1. A separation result. The pairwise operator of a hypergraph does not in general determine the latent coordinates that its incidence determines (Theorem 2). Heat kernels and spectral hypergraph convolutions act through this operator, so the result gives a concrete reason to keep the incidence itself.
  2. Identifiability read off the structure. Anchor and bridge sites identify the latents under site-specific nonlinear observation maps and dependent latents (Theorem 1). The discussion led to Corollary 1, which requires the assignment only on anchor-bearing units and allows continuous loadings on all other units.
  3. Evidence with controls. On a public benchmark where the coordinate kernel is correctly specified, ISF still recovers the latent causal graph more accurately than every compared baseline. On twin incidences that share one pairwise operator, ISF scores highest on every draw against the coordinates of the incidence it receives. Rewired and reconstructed memberships each isolate one property of the incidence.

Our view on practical applicability

This is the one reservation that remains, and we agree that it is the right question. Scoring recovery on simulated data is the standard protocol in this literature because real spatiotemporal data provide no ground-truth latents or graph. Our controls show that the method helps when the memberships match the footprints, and the paper states its claims for this setting. Looking ahead, digital environments such as those used for the post-training of AI agents record their structure and allow low-cost interventions. We believe that they may become a domain where latent causal discovery finds wider application. This is our outlook.


Where each item is in the revised paper

In the responsesIn the revised paper
Table R1 (twin incidences)Table 4 in Section 6.3, Appendix K.2 (Table 37)
Table R2 (free variants with smoothing)Ablations paragraph of Section 6.2, Appendix K.3 (Table 38)
Table R3 (three members of the factor family)Table 5 in Section 6.3, Appendix K.4 (Table 39)
Table R4 (declared variance)Appendix H (Table 36)
  • Corollary 1 is in Section 5, with its proof in Appendix B. The second form with fully continuous loadings and the two counterexamples follow it there.
  • Section 4 presents Definition 1 as the factor family and the single-mode restriction as the hypothesis class of the theory. Section 6 opens with the division of roles between the grid benchmark and the hypergraph experiments. Section 2 cites the two suggested works.
  • Numbering. Theorems, definitions, sections, equations, figures and Tables 1 to 3 keep their numbers. The appendix tables of the submission are numbered two higher: Table 17 is now Table 19, Table 29 is now Table 31, and Table 32 is now Table 34. The sentence at L459 - 463 of the submission is now at L483 - 486.

Thank you again for your time and constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.